Codebase intelligence for AI and humans: code health scores, auto-generated docs, git analytics, dead code detection, and architectural decisions via MCP.
Python aicode-complexitycode-healthcode-intelligencecode-qualitydead-codedeveloper-toolsdocumentationgit-analyticsmcpopen-sourcepythonrefactoringstatic-analysistechnical-debt
docs: rewrite the README, benchmarks and user docs around what each reader needs (#3002)
Raghav Chamadiya6100a2c2ff57
diff --git a/.github/CONTRIBUTING.md b/.github/CONTRIBUTING.mdindex 274950d15..45fce51f5 100644--- a/.github/CONTRIBUTING.md+++ b/.github/CONTRIBUTING.md@@ -65,7 +65,7 @@ uv run repowise serve # dashboard + MCP server on localhost```Then point your coding agent at the MCP server (see the-[Start in minutes](../README.md#start-in-minutes-no-api-key) for Claude Code, Codex+[Start in minutes](../README.md#quickstart) for Claude Code, Codexand others) and ask it questions directly:```diff --git a/.github/SECURITY.md b/.github/SECURITY.mdindex c666e8259..cce46b851 100644--- a/.github/SECURITY.md+++ b/.github/SECURITY.md@@ -2,9 +2,13 @@## Supported Versions+Security fixes go into the latest minor release line only. Older lines do not+receive backports; upgrade to the latest release to pick up a fix.+| Version | Supported ||---------|-----------|-| 0.1.x | Yes |+| 0.54.x (latest minor) | Yes |+| Older than 0.54 | No |## Reporting a Vulnerabilitydiff --git a/.github/assets/banner-v2.png b/.github/assets/banner-v2.pngindex 8f4045a81..1eba14600 100644Binary files a/.github/assets/banner-v2.png and b/.github/assets/banner-v2.png differdiff --git a/.github/assets/demo.gif b/.github/assets/demo.gifindex e34ecd6ba..5e8319d2d 100644Binary files a/.github/assets/demo.gif and b/.github/assets/demo.gif differdiff --git a/.github/assets/product-map-dark.png b/.github/assets/product-map-dark.pngdeleted file mode 100644index e386d1daf..000000000Binary files a/.github/assets/product-map-dark.png and /dev/null differdiff --git a/.github/assets/product-map-dark.svg b/.github/assets/product-map-dark.svgnew file mode 100644index 000000000..428230364--- /dev/null+++ b/.github/assets/product-map-dark.svg@@ -0,0 +1,18 @@+<svg xmlns="http://www.w3.org/2000/svg" width="1600" height="900" viewBox="0 0 1600 900" role="img" aria-label="Repowise reads your code, git history, tests and coverage, docs and decisions, and the contracts between repositories into one local index. From it you and your coding agent get answers to understand the code, change it safely and improve it, in coding agents, the editor, CI and pull requests, a local dashboard and multi-repo workspaces."><defs><style>@font-face{font-family:'Fraunces';src:url(data:font/woff2;base64,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) format('woff2');font-weight:100 900}@font-face{font-family:'Geist';src:url(data:font/woff2;base64,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) format('woff2');font-weight:100 900}@font-face{font-family:'Geist Mono';src:url(data:font/woff2;base64,d09GMgABAAAAAFpEABMAAAAApvwAAFnSAAEAAAAAAAAAAAAAAAAAAAAAAAAAAAAAGoItG5AUHIQqP0hWQVKEPj9NVkFSPAZgP1NUQVSBOACFEC9+EQgK8mTbBQuEEgAwgZVSATYCJAOIEAQgBYkGB4hDG7yXNWybVgO7HSDv5+ttNztQux1FQBX+JsyMChsHQEh8fvT/f05QYwwfrHtQc2sLRIwcqe7ZU1RTKtrdbtZ+Vs/AFLLMs7vVgfNUaR0EEUGESPmxai/NsmoNfETLtlhalfjbl5wbdf6On7vAG2D4F8n/0W9w9FWfi5r0SzaFSSZF+jK+Bsg2+abuQnq42xngTo6IVDwJFX3T+sjq0Qjw5DGJzPIa8ICJ/HdA/ABlQg3xc/oXIAkEDyl4iCgJUSJKRF+0yYsSLAQtCRTtoKq0nWl9K1XZaJ3W13lt++3Mul8zIJ4/RO/u/YztoYAs3haBWDZQgCgkq632/89pffc+lK5kScZgQXcnVdWUzgegfFysZrazWc0pS57hxXIWO4J84m5OAaWIU45Dji0L3yP+4zD3/fNCN36gd3oUbH28FfB0D0csIG084xFs+qdStWokAmyE1Gg0IhsgmGVQoiiOLGetVxvD722yufFe9/hc/fe+OXtycJIzJWcqWCLnX53VB8dhQmJStjfkVF1R3eu3aW7w9801fcgboveeNzkPZkAGWYCEYvpSRggRdKgzT5ItGSObkzyQkwfADlvH8U75FHDZAXFqlwXRf96ur3sLg3vqjRg7Ywal0EOpWf1//TF2tyt9nQ6gbGEFwKHOf2UI+QNy241gR9i22CnQMneZgOABJmyKLEsnWTiw3dg7Evbf2pBZn937J4aIVF4nNEJhef8yWZP725XyQ82p00Jtog/1WpapAF4ugPvk/1TNWnx9gcFJgrnviXKUHLl2LjpTduWKz0Xlphz8ATgcjEYAKckWCNHkbqbgoOgQBcGBonPucxcqySlcSLmLqbrrrryivqIpr+jv/2xZpvVnV/taa2oZv7nM39yHsZFDZ44dJKWRZPXMSLrW7O27njlq9VJrVrpXI+0+18yipAOAyED8R+Ar7RlKOmrzAUbeIDVBZI7tKHEWO3IQ2hlgEKaGIsZWZWEOs0QkRa3B9BANncQQmthr89g+9UeYHFBkHdtyzoKAId/2e7+cREKfUzWXIlKMyPzv1/4/9qvvYursp/kcT7ms4xgRIRoi4U3JMat57w8q4JZ0mAw/4vDf5vg4CGdCmyqICEFDRMv6DluNcGS95rZz3VMkEUP0IOxjjtamWqxDdFpHGZI/vCwEv1oIKl3HGOOwppkjgRbUqcxKIkTOsfQ7KWJqzTAbwYMEboDzvuMBAYAB+kHX9MnCvXxWeHL1ReslMC0OHbWP9h0L1TlptNWuKp4Ep/M5rJXWXwTL99X2Na7wlkwwjwseCGL2Bzwv1S1RcSWVVqFakJV2HO5WB0+PzM6+OjpD4wAGEPp99GGN5vmxFFFAMLdgi9KPxAIQ2AjHAGAdGkBAjDKDpA3u7TDIOuKxLwGy8ACq1QBhBuiA/KspuhPIAqzloPCDvZBxkQgMPhkIYvSNyPsZNYR3Ahk4lsFLxam7h2Ri7xxuWZuu52T2vTeyKNjhite7ixrAnxVoqBg+YCuWD+8SCq877v3/OaAosnNQ+DMTEeQdMBdBuDkSnDI4PKPGI/1uGi6dIdAq7Jjwm3EMjXwqcvQo4+LiyRlgFQTdkvX5+jAwDF+xYykxZTvtbJPZ6r25CpYFuZbbCAdqUYat/ZkSRysnoCnY+v4Z/9ZNXdQJ7Xadr7qYOsviTNgCfS7ZwpiwBuXy2Iam5Yb5ZvBbJzHmXiNZK9S5FusYlmP1cF/Y6XFTJ14HV293sWDcGj3Ysbaj6958h6AwVN84dnVPoepmhhTi9WdPZ0uAcHv5Px/dwd6uA1KVOtTK49gV6bzaiBlG0JY1NA6GVy+SPZwLmY8Z7wqlG/mR4pxwEXZal3dGrlVjj9ysM8LbGxAPbMfNYDtsh00dC8SQKsvMXo0R1c0FBYzp3uICuJ0HugoPxu/KIsFcpnOkYFFlg6Wrhp3tjsuJha5DgVYnaOK8pEjy6XQfAMyeggj7JUX79SrWz68Jen5tFvuXf+ZGwNlAiQIc8HsIJszegojGa83KWaWy3wQ4e7A02hnXXQDz+8VNVECd2DsjgH4tAM1K7FES4Y4bAbLQhDFsx4FwEErkbgab6dVo6J7KGTTPZlin+rKwGtajQTe24Q4n5gA86HWwBiIbdv0Ce4o2YrPz9sKTnVTb3yBGqQ/D4w1VMViUcDIFafxF7+vPDSHn13badAp4fm0vbExtALvs/yGv+pCUjjjThHXpUgl1LquSQlyicLCQwv/yNe4cTVgOJOeNvJ5nAUZHT49kGbDR7yhGP3sIDpn9BQhB0riNDCI40WUHsMZfLYup7IowqNR7sGZ2GALG2VUxl62HaEKqDkEVsCE7uDleXIBwdl1M+wWCfdmeft3ykOxa6GdXh/DORgSRj0tYSpwGmWjAKLZin9UP0fSRJDGwIZGt4xzzCUz/tltArftgJCBd8sDJWtaGh9fshiZjR5tj7Xhs3G8Y3Hg+HqGf4FgLMdc/G8Lf/zneDSyHia440P3fV24ApJddTZ1AsF7Nvu6P9wCa1jmgA7vCESYyzFE+fCYQV0kUM7zmE6vMlSB7GApdJ93NWp0wE/7nYUKRNd3Pz/2TOhcPKVAEhfKsutYGpdB7u9DIryp55335cQEMQ5FAUpEzLRT5oVGQq4rCUIwkR7UrRYXdIASewfaI9nLKGgT3FQdAIs/J0YySINiWQWVEywwG59Rdd/+kN71LPrSa2ZTBLgNsGwO27Gf1td2QVFOT8bVvoP9fxoiTIEmKNBm5sPLgTJKvQKEixUqUKoNXjoCIhIyCioaOgakCC1slDi4ePgGhKiJiElIycgpKKmrVNLR09AyMTMwsrCA2dg5OLm6TeXjBfPwCgkLCakTWTW1dfUNjtCnW3NLa1t4xpTOe6Oqe2tPb1z8wOO2VoeHpM2bOmg2AEIygGE6QFM2wHC+Ikqyomm6Ylu24nh+EUZykWV6UVd20XT+M07ys236c1/2837+5tb2zu3e/ul97UG9gB80Wfki0O12S6tEMy/GCKMmKqumGadnIcT0/CKO4n1jEN5uNxpMpzKsKJJmcPIXkK4zAklTCOnysdhL7r/5juOm33aectvYVOKx8DXVaoxWw9xf4MTzbxVUgwS4r9gLYPCrRXgiijk1EeR07BzjY4Yf5oey0pbeDYwhTR/hn6RoA3jouPQHg5q9vAZZtHECzu5BiCgeWcEnSye9tfQjFnz6QR1A/nwABm5QFrEGnaeZHa1ENMYAeqeWsqkdZZ+R0TdM1AHujqJ33fjBSLNjbxDLSMoRW28zHp5NNZQDgSFBBEaVWip3xy7YGghQKuGRm2mpCTEN6+6xxu243HcJ5LuraThANaUpz2qTrdIteJTzRQsr+vOW1Vuwyh2Y8cmts5856xoFzLnKiSwmNL3pAz05v6fW7UJW9BVjafKFcYi4xFu/WE2DxG+FhAODXZ2GxsDH9hwO3BP/bw9f+L/LDFRCsAbCLWx4D6CZgWlfA8GWOpMNHPWKHS/b73j+uuOwib7XPbVscMuqAU6nfLJ0h+hwCqdpJ0ujkl2IS0x4WAgnkYfrWxx10wl8+C2CNiKlvRRo/iCxMdRuO2CrQ43s55KhrjvnDZjvd8D833XLdlyb4NOycP30dJF/7HRd+iEHwH1/YExT3mO68D7zvQ7uhMDwihAgTJUENzuoSfI2zaEXHVMmM8JMWoehKObY4DXpaOtVMWyYOzNQBpIwt6sWQUbZE+Fn3IuHSadq9okKi/5jFG3USeeJTENDfLBeBNGIulAMxsn1uvw3kLBLEECh4eYAUVWcBwV41WMRcVM4ScpFlKlgGiIoqWU5RUc0SKSqqVTJFebXHquZHIUaMTE7JktR+gGBS4l0CcBs4Drm5+ro7HD/lH6HbRwqLzXKGqcnbQERueNzXeAR7P2NYXAxRAiDOnIxc/nTJuNzoQbEVxUn1tCXaiUrCrchzYqGBhAi3cdcxOsz8p/Yg+eZAQt9JtXqcT0E5QnnwwWkXOymoKnTTFZK6EAKTb5wktfuZAAaCwrlFcAxHYY551AjlK2F0FSB56962sELnj2z6YNq0iJ8xQs7OEOvYxozEsRM2SUzItbJ9/7wStYnM/T647ycSsHmG8X6T0oJ8B5O41AFnkV+dw3i5y8duzhB8zSUWkAKvZyyKw8Fa8xkppb9645N8MpF5JclZEOtH1hYtnsr3vXaGciuPRZ7dbtdU26YaIzU0kcCmGE9QoSOfyU9F/Q8NlW100JJ26Zwyt1iREJIwiaW1NQgxRn8Ywg9xpjAg+2MMY4OoCsWeR391/N8ZSP0OyHcHMJzs40Ab4QKcUOy6CrGihBJO1zcFoTfY04bhW/y7ESMQu+S+wG2CjWp9iEjCDVcXgRELY7dEoH1MK7dLgKGWDkFzjYlupPJ9YxeNKKSqHeIYLemksvwd7p8Tl+0D86MyuCrXlFIoytKQ/mNaEqSFl8bQ4fqaXyZBGPtObz5//MZ1jKiXI7PS+8yhaI4g3HIKpJked/f4tHen+upB2LnG5rrb9N8xVXdDhtqhdCuw3IiFp5PUm+tr3ZGb5raGdJly2hBWv32RIyuQZpKnKoS3DcAHE1Y0ochU4YbaNYj8Yq8OfW0gJBh0Wb+v+m458u1Fb7dg/U0Ajg06bQJ5ZX+N2En+Aybf7Slz9vRRIU54Jwo1KvQvNHj05WilJUaw6FZsOgrcnwnDNl3Ew3wF1Z1zk/c71hc2Txut21bDzi2naF5arqpbXa1KmxbG5vtCKjB5QFqKJa/WQDCApyuVsIBjb45BxfIArJwVbV901HfGV8DIr4G4GwFojOCn8nmJflQScoD5effyvc8HEh6UToR6uCEsDkAiht0gxtpIdkdDoF2YvI3OxGqq4iZqf6UQdVkRIjPvkfsq2XCRxvV+9sYL3+4B7u1S7ZCoatpxSDctzxXsXhzWzoIXhFWjevJw/ttMihr5/oFr1Evk5CYXWw2f5ew2csqwR8J/ZkKrcYjZIWUSP5uM9drbCzPv/6Y5XbMB34ijXPWDf6dxnzInfQXv8+sgbkZP5AcYOo8Re++yDDMcfmiXkVj0bbmqWOlnqKk2c2l10r2hkEHjrOgk7EbCc/+50ymmyg3/1pQ8dcGccWRMMvNezy5MtgtuD+8G165oDOYKP024t2eJ6NXZ1upR63MN514fUkezeo/tezttqkRQoxwz8slHJNW1GZiOsWrEVWv81hGR9BLOtuif520116cz5xOWMxK/WA3KurJuR+1fa0b2WxLYY7/9rOCOaoETrmHZz6e+Ck3CS/FO/lQoStJ5t+ZnTuHS7imfv7s1f8RSC05HCK4ylV7+M2vgfpxe+FO9HyR7sY33VtG6S/D0DymYv7M+9l8vaX1a4o7zXbf1rov85Jh1w7nFudvYm5qhUeIRjh/TZ+Be1LVdt81BDyzRkAlFV7qHulU1a2TYCkHRaNdxec+8Sd/47g3sMx/t8mFLdPRjH3yq/eaYdeOztYsXnR82bufo2zrznrE2A/dbm9/dxsK+wiISsuBrQ8iVmCV5e0Vfe6njqGaHg6ND+k3X35s4T/4Kpko5uivWU2FVzw2TiBV9AT+w6uRRc8fkHPhI333XUwoRbqwA/OiRpxy15KXMIOMkXU9Xd9EfMrRXtxfvqMYdLCIVAgU23O68j+PZY+S0Uzdcev6IXvCn4IprCVvfvuOOb4ghZH8QTFF1eHPdnzoSrny31VF7ftbbzavNh6qdB//s7C7jdmiQmO1WdVD2ndLdDPnbmos415MLcKwoCaeTge711y2NfxBu3sj/baWt8K9dK3SrDo/67UKcSchZgBR1KGyMNzO14Rz/V21C+jh3ebT/u9EOUUMyeeA0aXyoxqSD6s6zk42buVHlLAzpBG5UiWUH7aBb5yDdCJHWf/tJ9z6wdOEyN/ognzny7cBi6f19n1Vrcw1rh2rq5MJVb29RH6dG3naTzjwyec1c5I4Hil5et+x6TphX4XrmPvhL2fjcypt+6pdof9XH8Xqr2C+sGiFuka83I1HHAXdk30XDI+MqzOOOksAJA61oFsBRInqYRB8ynYFHZ2ymxntlAHWbBoq9uni/g7g7DTe8uf8zmhxY70lOVFpXoUen2GopVojRvj9A6hyZFEiTEEEeZZYlrH+tDpWX903bjFUOujcZk+o+mdoLNx2Ubjh0J03CCKVpRWeEbV7FXrenZ/eZPEOXhPexvueExwmzbhUr3bW7occPfyv/jrFhLYlxV6OkcDBshBspDkwyhOzTSPZtajAopXeiboj03+vqPYN839juykfYxDwKAYNKmoZBXsFOrNflivrco6P9jdH1sOdmiGDvb175b+CSO0f+DF5jyXMBq07iuNBs4mQvr18gkxypic/k3mvsfqwtFB1zb2kP+0LKD2KiV/ShveMDrscGMaKTNBjnsdFLt8MgGKU2GwctTMdBQ8ETaznokSvf+fDo58T3odq0YbN3mudfUrtI1R77es6ZZc2eHh3Lj8ewwb0D5lLHig7Kq9NZwdZ1M3bcxWNUkV1hfXLMOwP17SlduQ3rzwBT/Ad5EBPHr7Quu9YWphaHVVeefXs1q7zxsT1sUPzd6b41juwbnLKcc2XeLBwcAHP7v93w1l/WeNb8thfG9GNP9nHOFXk/gEF4/Gz/QGuYzTTa6AKUOsjk2Kois9qnLvUud1gUYpmtGloOD0B6mUgcWGgAVGb7+kPKjT295Nwf9tXNG5oUr0Hm2D0Oqcyuo8+O1NBn5ZtNpvb5suBwCrZvwXwQHv8kstgDfdjTA32weHJkstfR1rNw0aKexW0O76DeCYd/cXqcv0VhEBs/6Fk6rcjtcQsAM9vW59FvGOpVb+wPe7Uxmb4lSxdC5mid3mqJQ8eYHalhznLqZEJLFbOudNTztNWd/PHglzyXnQ2U4/7QruW7Qn5A7a61wA3vwu+CD0Z2gM02tSrkCQTkRSf5rF5otsO+q/6rAzMWgUW3+4RdHX+cHjPHqx61jlE3gtiW+3DLChggXQuoqQO/k7nLzIw1nPuIBlJ/tAPT+BBm8Y/L/Rv2pxWAlQ33+ZQbB6aZyhWYXj9twOL1Suwa+qzaWvpsu1ZCoRa9CGe1nMFX2Y1tY3v3th1t5JpE3KaCRR5PwUJuFBrHy+JpJl0J60zjaVkVfC17dbUH4Gs+0I+/hD/55X3P+7/tNJ5xFYoLFQ0GA6ucC5qLa5sGuwwarUsPTwM9r8PHJ7v1eq1bM/n4fdzdJVDUnbRFu4Tua6+P+8fB8MAct3+9eBtb1P5QmPdZlvx2U/hXyTR+tXtOd5Og0m5meRSmZJMCqc+94M85Rlis47yO75kLOXQqtUMLzQXibFuvytiJNIeQMY4VogoE+kICmXQLrvDkW3FhAcMiUkrD7QKrtVWubEZrQsgwVael8nQOtZB43krQYf0FLQr2ZIVU5Y9zAaLnGpwKX+sBrOzuWSuGLvsvB501NpuzJrgNhxoGuP/N6cjYisI1jGw40Ok4Hda+hnHoIwgxCMP7ZK6VW8gmEJyMHW2IvVo6BaJHG082xd4kBEAS6dr66Fpbx+h360lA4F9WMgRf//wlfB0Gv6QLXORKDcvghvw5u8gGpVYmuTIWRGVB3ZFwQ3tP803sixwS5F6VQfqlG5XTaHLhqVaxmULWsTlry0NvE6kMsC7dE1i5YGFgxFNlltH6QvQnyZPZQWpZ6kcwChswWWHKR0S9qNrm16b60t0EI5m0N+hFZnghrwayJrqo+1KNZlellUt123MMmIXFREb+6adrzUU3EKmaC9jTK4wluJOXUwbt4zBt/uI7xvIKU7NePOzzkXrfNZBLb2+vidjbtHwjhxJQKylBAxfI0qG4XjEdhhXDcQOkqKlQTB4NIPeapEKrpy0SQWQVNsNPGx93VOWv5QJM+h7dTCxD+Z3F82nwuD1l8PIeirGE4uLFu4gSVKopASOHr7W3RWrs7eBF5iz+lFkNPbOVG2Z469F51rq0eXb3MHlN/f8x1k6VotflVvRkZ6skyKmGiD46rBjsVFtcVnekzuEOWSmu8fnsn1HAT7d16uXTYVg+3GGwGts1iuFQ0dwU2O0zVul4Z0IonKkv/K+BhrQ7zAqVEwJfV5ktkG0UGi1HEatun4FOQICxVlZDVRm3BZB7z6ssdpncz/Tyq/RtWnmvzS7radHqpEFhCez+Vy0JtnKgTMPkOjAZRXEaKjgme53H46wDsXGYufq9R8ZipswpYIQVPljlEFulGrsDUgI39M3Hi+9D5WcPQAdgEH8D+oTeyPtr8SRIg7Wa7W4QHoe/QJR2X7l1ndey7VFrfCnIKSsmCAkzoDlWCLLiGGXQLQgwsI6k/tVL5QYOV6BHajEr1GaeRhZoaYBh81XCpKfMSg8MGy4QV6MA7v6qF779od1axlS5Rcyw1AerIJldydSSaDq+CrDSrZ0GFj6VS9ZSQXWoWnHxGMdtkYgNW8kvENiS7gm+OXN28HWPyKKk9PtpLXRcYw++MevYxS606ujDfjd9htUkcBWpsUGVZAF1UWF1jFQq5Ko+en+tBzGsZBLdClxvfOJEDO5TRapVrpiDeyW0QmWbi2kqQcwZS33jmtHZWluLjkaOjkUJaLXkgJ41slTET17z+YQ+YDkaF4rtDpa+8J87u5JXXpPoJGVhgYgZnVQi1svw4So+fWd6KUC/AaeOzUGYSpiHsqsj1ZX7EIPOPy5xdSxyQKu9KRnmyP2bAoShPYs+X7wHuoJe0Y826HWyuL2T7jzMxg9egYAo3djElbr5AWS3SVKt0BApBbV737Fn+hkqFqfSHGJqdUGO0D4jgKzBWCjyGaSSj159NDkjxFCwuWJdHRm4H+pFjQ1V2mi82WPw+U0pYCvMF3jFdfOmTNO4NS5YpanosOPlbBdbbawWcQ+rlWaL9iEYctjM0HP4ORVLHoG+OQN9CwFUuqu2o84Hqx1ii0xtdpuVMNY9bzxbEzKBR6dyIx98uoQRjsGMtjIzHHJ0eVb11SKtuR1Gx8MFiE9kH0REl55ZdCZ2IPmYHn31G7KCRW9AqtI1UWLZKXRkzaE3uzt/C/95QAQ8sKvXBK1Ld30QAg8Gq16yKntkiUkkRFnpE7Mveb5TWm2xmvQzMpempSsJiHRcGbE8mrgppxdTCzy5bwzmYr2ANeisqXO7NZfTEal3uSN1dg3zXJXwPLPiU4HwMBhsLnjhSqZg01RHd0ixsadX6w47nF3hd9TbA64r5PAKTEba7ECIFGoCvrn0QgGF/1GA0qzDUAPtFxfDNC6Gb8FAYONMG4i3S9I+3UWhPrzDtfkdGHMImd66aLjX2ioUR1RishPWW2n/9xvZmuKlMKBm9w4nUgJBJFIDSzQsW+HvPO77llKOFJZSv9EwJEapTm8QJULINNdgtNXYzBH7FDqP252F9ac1tzoN2vpePpj/bsTp+z5y1Q/73w39k9IAmNnGFr4yBIWQw06lyuiBbQKyRUfjVtgZyAhnZAhmlSMnrOTnckOLWho3m1WJThWozHZ3W+QrmoOl72j6OnSu6gBTbPsihD60VqHw6jgLELG89Z1uitDUqpB1Wp3q7oTS7DX5gi7I7dPzbDSWmc+n2cwV4FhCeX3qJWd4LnzLqbM1vKF6u2xdGLCyTVO52lqjoyZmy807Btl1UkmgQVBd3SgQB6Q6q/1Knj03HHMYe+zdXJPXFNLo1S7I7XJDaj0vY4zIDFKZUKdQyPUKIdUB85vyvujlLa/FLrkC1lX2dClNrxlwW93+r4a9VOusCarwMYNSqvxoxze+X/D0Z9TV3xHkjv+UoMV8jf4elTLIKCRqC1Pe636eXcr8p5dOONdw+le85F8Cltzkm1JXB3eYzdabuSkmBbvSbDJwLCz6/8ho1vAd4MdVf98GqbqX9UBQyL7I4M8HG3UVl147oKzwvJ1XrCikCPR3gyaDRq5XCXji/Qv+9wnuiByw5M4xzt2CSga1obTMwyqSFZUWFRV8TyjPyLtXo1XIq59ySdQnJUKixK0GIl14r/DAA/UdsVwc3vFaYiWgy2WmKs3dAA1PXkAlpJNLKbtUBbeuy9dkNI2tKFQrbNJquxFUyDWuqq/3//wq1n5XvvoCPh+dsgXc0aqKpwo6XIfkKexGSTt+wFh9Z0guIr+l2vyQwtWIVOoHXMrDzf2byCKrY8poy8vKFtKobybjfcqhsj9L8n8vbrHf80v+BP8OTR8Cd9aK1oqHpnUrtoCcKyR0fUoT41f/r4yUJlQ9KXSF4e1yPBISKl+i68vvTb9lwYuzFjZc3dp3FV+l16sguw1tDiGRWnVzh8yqjQrFNcqKbU1b1nx9lNx2dj9R6drXgpVir4y9o3nLksNVfm7Sf4PA14se5LChqqjj25us82Wlx8vwx0rNlnEsLTueAW5LDjQueh96P0w90Ly6ZfWB6G8iqjQDzHmnLatUn0o3Htuw4tTLPSYfaAZT4nnJEbel4K+4O7CAwWeuNhdoeu5c5QR5b8FDl8N5kGmKOwzwCGGV1crq3VsakMm7kjhisbDhHYHM4F4cjqoYyAlXumeA2V3A6sEAVG9vYwkHxB7xgJDVZt+vKBB0BjgOTsQm8VvRKCtKErBFwkkBJzg9+NERpQELCmVBS/1nEdzP/OBCXBPRmP678h9YFt+Epr1qjdww3MjyX04E8etxUmDbfGZJWhevg8LpuOQ4sN4Wha2N84M9LBYQbEiO1SUKHv5Q8DBRJ/8TAeiGzKeHxToyn7LF2Avl0aEei+b9GX2cbd6+bpnOk/BKX218hTr/HoN0fSZmduamIoEk4tLRu3XGikIRBf0uJmJCGdOnoBzF1Kos7Jas5sBkar+7Wsv3cTWN5NdiMfZMr1JWdjdp1uqcy80ZBppgTXE+b5KzUrwD3Mlu6Y/WcIpMSq0hfh9Z7oj4J7t9brNMH3ZVrb1cmGsTetHplK4TcEuz/UKMoHIJ9QWIHd/DlUauQKeQTg1cVRRR+imMQzTcm0tAlkDs4/HDarUwHOSJb4+7h0xDQFRtiHJkbp0fOcso1XL4jfnlmkpntj7LSw/Q2BNf7VfJXKEhQlUpw3SJrdmHnGHmy8oRR3HlOmueIy/IClcy3lz12jCHJdJFqOUSbgeZWAhhaSVf787PP28qARXpgppqzZs64+aNc2oCgbm1lo3xTsumuTV+lanJLl5S1hpWNpMRRUtaEdXijLUCuwhSVtSr1RV1vqugwkzh6JyGlDCLwRc6tBqhk8+sTtTWVEudD7B3y6IVEMSMymSishdlHt/ItBIFprFigaZOIWpSKmWHOrlG0yCntQYF21jR9MOdyyov+c0+XG3o4/SvF5Y9noR7UCYEC1v6T6+0TUPo0MgH6qJ1irrlNOktaSmtXlE/JjbAeodlZlaYKisrzGYGS3oYFeaSqXUzWMlZVFo5iVROo+LBn7crf/KJfwLJ8Yg7AlKWNjmbHI//NHA7XX3bpD94GSEoblAMw7Bielw/IjFZG+X1Umm9XG4qn15O8zIj7UEXnd71gEZ/UA++/sGWMwYWLBgsn79g/iD5CwiD3sBKb8HApwRQoSFoqMg6TrWQSD/ANO8ka15YxLDyqCfQzr2fRBmt1WtJ0zNIe8RsB/iytPlQL6YzfgeqTybxqa87JNnY06tn0j1l8mCpRUeKSLPyGaRwN3I8U5XR6rS8oqlQw3i2vlwqDtYLlbZ6ntCVaEIdv8pj02b/Ti24eWA3tCnE7rzzbgA4RpxHnGOv/V0wjXQZqysnnN84J52x9AfgmkCyLLg0kWh5lMXxAi2AJcRjS0qw+PD8Ze2/ZckfOC86R2DK6IHLogMzwb9kR4Y2+8+c7wDZhLtCfw4PzNB5KL39vPN8+piTvxqw8C/y6FfSFmR68T6Ry2NultMurV69Lq94U1XGAvZod4ahDU9BvG+xu0PQH8/LOACIvPEJrLSLbBKNZbLJt3ULVkIiLnXtxXPgnbjSG/b5vSGl0iK/r5w+RpwiFk8hrjLzls2n7beLq0GzGECuzM7KzOwwJitL9HqPswfIl+7CrCg7MjOpshYn+CQWDUb9XAUTFAtEvtCsOu0scMOnVRHKPvJF5tetaHI05dxgwAqrbEamXLuDRN7BlzMhI2BV0w1097t4RCN2EHuQg9bGvJyMTEU2diivqijMKCRmPt8tI09JqWKowcZuS4YFGB1K4T5s9t0AG37/19zX5nr5F39dJkVg4sZ8lzCo/Oa9f0/bgi+Kiv+qJu/vBsd2X6dMtzpL/78AuEZdp5vAsWzu1DjBrNdfYHjA/Gm8M9P6E6j8afneso2fDoAyBBBmnG57M5WTUDlRzSsrUDlWJSunwJyG9K7PvEMx01q8ROlsye017COVEyxXWSFbGKtavQUKuyBULC8oL9ClfPjJZiJsJossJ5eTy8lEFnyWIEoKLOeX88v55fzNMgLKq3zGc9BDE/+voFIZGnu3gqkKTqPxkHmhdWS8b2lowiqouEPGfkDPyGiRofFNKpgoPnmlBPshY1MYmhndG5xaWpcBWKemARuGxrsrmKgOVahg7NuRBcwRM9jlt0EFEzNDlVnmwGc4eJ/xHiitCMLcXV96M5v3mPea95n3mw+YD8JD8JZ6OOuhAjvhPXgv3of34wP4ID4ED8ed1nRTbeA9kg0DG+YW7hfbsa5fGRzCTHaXH0XrfToPejyM4nEiu9sAYBnsqc9YZxmHhbItL9DGodG5pN8P7KCdyVia76es8AJ8wCr/3Sc7DSwOqYATOIXTOAPzJMJ48FXdXVaJ3xpR0m8r69fO5L+RwsSfPgZZGYrfwGyytDZrc4CtNUtXE63sN1xsbt1uBFovxe5bcDQ+3DjhVnaWgbnKn54P7lbEXQbm1f7p0GW03+ZQJnYSz+5h2AnMlayS+Yf3LROUUHACcyWzGnQZgA3ymcUdV+5b80a/M2A+3u46pqZUGp5qajX2NZjX/NPgucXPdoaY3/o3yIbFBxPdoYNSUDdULWG+tmffI++4km5g3ug9QyA0CcPXCeecccH8wX8Hrqrqz40CzA9e1Y99V8vYDGhz57MJ8rlafsTwKsztPuU81kO+JPCefJNx1J9PPnhVFJ61y2yOXltn2MRyN+RHAodgbnd746yH3C/wnhzjwmh0VaIz9FINoFCuJca/HR3mcw/tOjkMq8wQ3LPi/CfbarmcBy7mMjpq4/wJ0e3d7JWqH9meH7UFv16bq9kdnWFN+T3YDa6UxsU95qqq8bxdZ1NNdXeHONQUPwTDXxm1tnEQgDCwsqkA1g14VO3cAtVMPOrlGwM/PapsI7Y52+W+7QZ3esZK0xWutb6GW9TujnbT7rrCKMcyq+aDGZ1dM7b8Pp7yiMuzlqybjECiIFcg1yAvSFelX6S/pcd+NkqKCqC6UTNQb6CT0FnoYjQRzUBXoRXoUwouNZgaS92bhkurSWtL60ubnbY8bT+mGOPEhDH9mOmYxZjXMWswH2P2YK7BX1RkOid9XvrB9KPpzxOpGTjCkv92b2f8rqszRzLPGdys5Vn7s3417qZeppHZudn0bHt2Ivvt7A3ZO7PPWryc1pxDOd/nVuSuyt2GBdg27EheUl7HAqaez9mvC1eOE+F0OBcugnsPdwA3MSmd1D/pcH5uvij/w8V3f2v+ifzL+dfyfy3IKVAULCn4tuB2YVqhpLC2MF44o3BZ4TuFHxfuK7xeOFFkirYUjRV9Tr+fN4qTinOLK4pFxdrixuKB4mXFHxePlXSX/FOaV7qjjFa2ruwiPgNfjmfhZXgb/l38y/Ke8pMEPMFMqCV0EhYRjhO+JtwhYohUYgvxFeJq4ikSj+QkxUmzSSOktaS9pM9JD8kccpw8TN5O/oz8iIKmTKKQqVJEMf/CYU1XNbABrhlkgHxQAIUAgIdbn1vLgD560HuV7Ms1cOuxfYt5kz+e0TZhI6h/rcP9r3bsvPCv393/dXgBOYem4GaMEnaxCJLBw3+cIBUgAYDuOZBCtkG8gIhZJBBH728sZ9BegAkoMei545riXwGXMN9cgO+FL22Z7/HNQd/79L3L18H6Orj1+JbmMRrsG2Be8Un4ndZ3MNO3iF1n7G5Wdz4DP/Jv//tCCQ8eWD0r0EYg5YIXrXHpelFKAulq0S+pzV0l5o23FjKm/8COpwQ8SxeSi378LvID7omvMVTJJmXOHSarI9qyxvUggkROH3n56xdNF+PPbz3/eZPng/DzrOP5SpOGpEXDZd2Pv23+7hn0ektxHaX99vqAlkDsqUDWx/sByYc+42Nyt0f4aoxjV8bMafeeY50cbXGpF4eQQod7dJ2RB5kd9AksAV55Yoc4f/8Uzda2YZy3a+Lj9St+2EgIre2Dk+QF1thch7xvAbyxM7nVadlTUzw5AsymgF1gdmACACmk/f6MCWcl4Z20lcYBATvCWkCu77wftIA/CrD0DhNjlaiLk0eW9Be12/nWyfsiYrghNri7axeWgK2soa4VyJrIAkBKDTsmKf1n6XTe+vTuxetAmke/pOp566WfHVJESonxB38HLCEHaT48eHYzuP/s0me6A5nlTHUErMFSGZoq4pPB3i0tvcHT+sL/vV3+OgKXcGrQy1tObfGU0QxXtxrOgu3i5OfCqkWKYqzG5HqXRLaYkNAOPJ/EuXaZWBlCqAL3mfryQjEI7sgV8cagrg/aqgayNI55Tdf9m0fqbHIFJve36tr4lx+vVVUEj4rM0XRCqHQfkGXjDbL9E+Q9C+A3vrdw73hXPG2Pxc4ZFw4UhXIo5zRC33CFxeV8hg7RwrtGaLhM5pig+xgO9ML4DDr332s3Op2K/vx69R5jsNow3p3xOh0OTBYuTadOEnn9JfAShH+q96Hc2ASbnkzcOe2RBigeBbDRqa8QbZN+fWSlxVyahFzmm6x+4CAas6gyiFhBGbwR6AmttUTo7dFUtyjWdoyahlca8IFSawO4M4i+kN/wb4mvvtJk9lr8EpLEmxfhYMrA39j4EzqCxt/kf4FJ2NufP924NRUJ2V42+MYkpKvEIIXEsFGttipTeAXeK8HEMNzYYIVyYPgaf7DCEvzFn8ZOmtxdKR9BxO9+l3GPJJhzT83sgcuaeXJz4X9Wgr/baC4yOHllwm49yRUCuDlRa3V3NFzXIBRAScOH9NvqGK3+6ZsEAQ1e7UD2Rmfqp3D4P9duEO2/aXLua1RwGx7okuU1ZjPAMzugA22KJUWvD1HwNJEShfWS1E34P4tLNkhIZPBsFOeHEiqMi3W2D359GLauwQO2DvFk3QCkMYgrRBwOv21+QX6vg/DSL1svhcxwC9am3sD009/+dt+Au631/Pm8FELQ35lna3MygQ+CHPvSWukd0p369m+hf6x3aEZFRGVoqlWLWtR1WmGbNxL73/RZifSk5ErA88ygDz/OEEC9YVl6CNjt+or69GhXuUj+ylYnjbeQeAtPlIpKgf7RXzsED/zWDrzbTus58fmzrcQ5ITbztLgseYO0IsG006ti9t+6X/B6itXXd+bKmzdWVr5iPoOQAIp251BwTEWU/JyrI3qdBdgtO+uUpSP23SgBerGDl0YjcG6GipwkUCLb8I9bX7CXso15rDkrHaQweKB5OXnvEOuBBzyt5DCNSgMw9biUMqgzCvafwLT9tdgHry72siBD2nrk/66cRle+MLoCBuMd6BFZWbtTwtzp44XJYrgjjmXSOCg/GXpXGtjB/kFWBb7tgxVeiZ+0w0lEorgtriuu2FHT8IOsfs3CD4A9Xsm99i+h0nutO8MqRebEM0YZeqJVAeGw1k20ActSatI2qKcK7Trodr1IXe0W/rzXdfQqt2BHka8lN7oPfKwdtxngARn7frFJ2cCYdHXfFrqVjjBGf3X3l79r9P6AD8ivrNvgp4y5tZWCg9ACFsl77W5LD0SX4140TQRNELXMLcKKE1zDyMD4pCtphK7naBYHK6NpvtTN1GpBQHqSHqaExTShLtqsX2/XcSSdEbOCkDhmq4pxVpqu2m4OyZNobFq6yIlKaQdzszKo9CveCwRIv5zhif7nygjqAW9aMLK0/7n2J2W68jPIDzYboSrmUZiYjtsjxVl/xa3uKvmcoUjBiuo4kya3xdq6i94OTeSQOAprTsWuAaNjALAEZBpQMoanMyMwDoB3FXhL2IRBBG+jU1CqIgrLJwLRVR1HP8pnKDvvlGinXjfAawCYyaDEK8d1QPeMIZzmYKI9budiIORBgavoH3MPB9mQu2rVhFupM5ivQYN3mCS5aOHnisf53/ZTpVgdrtqd7rHnKHQs2757Zezh8p+YrMF/N59Jef7pUO2Hn881lVXJL6eJf6c0cR03YSkjMezXHMgREu7dVW7Wyd26VakDNYOGSkO/cfu2HF4W7+dWIubi3V2Sz+Wbe8C9j//6mqJDHCuYit+o6s6fDeJcVPecjiANtVwkdwwvI/j/WTnfjuefFq98WrGDvWNOOSAwb0HnDcBocqbXIQQSK4gPxbsFqbhBHCc26joMB3qHqA/7XkPkncJyHeE0NYtFs3VEsUWUzi5hASA0Ij9Jsa1T5TkU/6T+i9chBdZMgWJ6cgAXCFbG0oXnEoG5mhjfCBoKhlYhKC4QT86mBFxIOSMup1MM81m4eeVvlJkLBfl5Ci2aZQuBZ0UvmzvZwPDPr3+XXWFmTvp+w6e9HtLIgYJSQ9Vq6Y/zSm+bQUACtt/FtFIMkY9wubQVMY3PAz+eGeQXSsBZPP39MzCdhaeQPxtbZILZsCDZFXW7pe5INJZN2prFq5uCMIwpH8daHkXceIY1CstZmCpKLpls0MF/NTtwZpgXNzNljLYhsQLxSWDL9Jj3AMMmxXkwCZmBUta84gDXhLd/E3w6AQQRwQnsCumE0PXs45AlXLsQS5kWHBF3tlDTKfGgpOdiXDy7uKkdIYBmSvedAXAjyEh69MVO/OlcsRzq/P8+3/cAhVFLS/bQGx7fc05c0taJXD0eq42vi6B1sIb0a+1BJ6k+VEFZLuV+2kTGyyWk0QFDyARgAVtbdUAyXWO4waChGG48Y6SKJEwf7EjrmM64PHo/IR6cbsZMPGopacysTZJbmAm8AXkyCPdYDFavF42V1eHgM/KpstNqtHo5QUiSP9/6Rd7zVVZeUy4nxP050mDNoTOlOjDS1uDr27G0V3UHKpZobVs9UIEq6+MAg+QhIE7r3H2IWtx4A/cQrayY2WGdkHBceDSMk8p21BMDEAML4/Qkt960N6MLAd6MM9dKw7cOzGxuE7t/nNar50N7SotCSaEJYgpN6VbRmMu33/kgjT3FbxRKjUaMrY6Y9ZdUvSDs6YHn+i0SX5aYRRh4pvAWcmWgqcP/c/jJPlobu2zckmCy3KQGXMbYCCirsYEHzCm5xhkobgbXpmiqi+OPtN+pNw9p5sTRNo7kR1F+Ncyd7tVOeg0qc7MFk1XiaxhfQ9JkNcbOXGwFhOCHiDa52Cum55q1F5RVuWEMrSKosnm/tAyJ4rxHsZov8FvBhOEKyF85FB1agjhUPgP7Nx41DZjox7cwBLB+r5ONwcl7LMX8envLTdu6kkqZoYamDXDbjuTwkvwGt6erP8zRYM499UTFHlBYvSKqxnbHTZslVYEY8zNf6dbUF/2n/37lluvJ3EX1QQParH85ndMkvHbKTtJ/WiifVsio/ufpwbtjDKwFGFnKu553/FsF5Tg36LkOBlRK+guJZ8TLTNsJruOFC7Wlju2f1+RXf7eljHtbjJaRtwKd1f6L+ty2wZXI5wDrPKwOOAukLISioAAQ4DFgoHOafRxeewmYnTnmw4tn9+KZHgNZqdrTsWekPi21qAQHRoK4Yk3kmsyZNnhqDGtMD6GBieDHNy6rD0aPopgNR7H7bnZmDPBBCV5WDepELQ/Tr42ndb8mvJGbSIhhbAsGFLKwFwT3F7LjIkqT2BWLzAqmUJtzZk6jbDDX8k5z7aRCjI/ZE6IStQ0JFx4YAOa2YDMewGkbmJb5vipbalXJfu+Q5KUAT7bxjZ3qdRZJcme/er8dyHSLYMwMWRDBgG2RpyLuNca0VMIWBkTrYQV4qybLpyTb+8xcZwBs4FaKTqrBSW4/G8qFN2tTuHS4zwDYJ5KVxqw2ie2+oVXHQXGHeW6wOlzrtIrAQvwDXnRPDQLMuMESNijg+05f4ou9qxTigauvk5AjcjRHRobM2wFRdyJQvovUK+LlsyXAV9SoHYweR72PZzf38dLmpAcC0KXnFj0zvyC7xJzs+wt91qKg1zlQDgSthtDiUN4oE42IGyeoLzMoa5flBOo/157FWKcnJo/lZ8jrfyB6n423NfvfvNES/PrLMgq4jqAIijifkNvSa89Xxl8+vbhCAx9VoK6wFWfqb2FLxfWrjfA/fNQLyf+SiG+8OIusQAJSpKOmvo2QOp2oszFD0B4tdUu5iR01Z+ADfFVfgfZYhHKvo1TfqAmhywfVavWkUacEpL4e2GFy1oZ0k/xhF1YyXRd58aVUeLdlYsijibgiq0W1TZjQsAdCHIvdYau3wow/Vv5SX1MgsVg2XqsiIhu0R2ipVXRbeTOhJKK0VwvfaslHthmLXZmQ1xq2MIyIkIeoe0ptGuHLY+W0r1o7ps/aYb+lAROEfInQHUmXZBQeLVGL25u9BR626anTPCjes/2S25S/KUfotfOjjbK32ktRxL/IW8NwdbuNXfOIZkShuhGE1Fd2Z883wPX+wF66ahB2y2mSE0FHYgh4HuUyqlBUoAV8Zgc/MWD/vsvvsoSVJU0O+brC4oG3PxDN1yQPy5bRMeTDX+49vHHVULsWoXKJXI6gS/eqVef6tm83t8/37HC4VBsGvEGQbPAy6rIyNa3k/2+rRbuzdmKaquki8yrj2oSzg7JrEKespiR+Q0mt81Q/lg8ISOwI6UlglDrmPRy4sbxAatIrBDxM0SML5QK7NRgMOi5wu9gGjmlysnG01Et6wV4f+ZMRv8c+2CqIh/vBSyQK1E6NILUl7HLEnnHoaNs2ovTftqhAJMw0pUNcPjKFt5YN789FUM1R0hrwtVZIku0cqJkzknxl29JReSaBn8aOQLreoSgtfCbpxcp6vMt1KYIY+zWG1PY8KkbR/l5VXLHIh/tJnPSGF8+BWmPLRIJnMjuyTT6zPlhvlxlG1Qln5craercs55uPB99NKF+gDhP0iYSQarfKSsZJ7UFXenzpWJD3wgNwgStuS3JjdLuCrjKc0jrHqKZKdx8lWX2gJjFrqTOSsceJ8bZrTFhT7aqltGTuqdUMNBq4BnCnp1NhA9RpwPYd5pnVtUbeLNo5QU+aZ1Y7jX4li/H+hmRfmfZWz7YPI2a8F5PUOF3ZLemjMr3Q7qRNWbLJlowdy8rqenBGZT2AuPndlVYn97SFJGg+DY3FVqvbnYfaqRVkZqp092uLE+JfuDjEEyulyIcxf1jdr4RbmDwCmXhtwmZHzn2Ad7CDNi5YyI6WgJN1rNF8BE43Hw/mEipEsRB3Ceu4WeiWed0lEYJIb4BhxY+k37eEKhC51SGZU8Eigrs6SmYqR/FxKnc4Gdghyss4OluWY+4qjJ01k6TqjJ5lq/VHknsO3RE7q0pW70nTMM7bFaun6T+bep/djLr0ZOjnUcTFhFTQlYf9WAT10MDWOCZdulQ3C+FKnqjrj/t0MFXbPK6q2w0ThaSPooxxBpxTCbUrTskYY1VKUvU84rua44g4E9RodJqXBs4K3LdN242yxgxmslTRtefW7ePJlqShT2epZNEiDLqt4hNKHYh4C7LWbmdRyEVGQq8jY7i8Cd8w8HCu2W5UmRDisaK81PCi+pj8ADsucnRZF/BBIJqQIu0snMmk08C6vEi9/LGdawtxdYqvun1ITsjwwL5rTaNx7IKWpYzWNH6cfihFv9iDxR1MdDxA9zcTW7JtFx1atJsScSPHjq8/3wkQg45tw709ZFO3sDBkv2sHsWcwpUvvqHGh4uz7yp4FpKa4gC82Gt9d2W/fWfX/4tmcp7iRagZCvXYrJJxsVeQt6j4QxNnwsMSjm1ZKhMCkRrPQG5/f2nQGAA3YSaZ6pLiKaBwUHffdpSzIfNZV6fwn6+TMgq42yK7RmuINLOQnNftN+nGzzTKQ/LSJoL96le9xw4HU52dX/Ol4/aw0pp9y36VQrhOQCbmZzNPbYH/DOGjg7XvygvEW2Lq4xlAQ/RVNtgwXaa1zUIdbk8V7M4P5xm0w4MrfnhXfhvCTqbabfDk56N3ZKqFHAP8tMPkq/F1vi+tJoj78syAogRLej2vBF3nzHq/c4e58n0FRgct7Uf1OX5n697rMU0v8did/Ly8ZKkH0ZUiqDx96NEBxiERxMPVcv/C/LDYeipymf10Wl3F54wTerjBu/yZzh/GQ5xkt4Tlu62a/reQJNj7b3w8bm+02pL8meknTJag5/UPgc08ysc7Hley3VF5/Crz0Afqjf38u/zpfGh5X+ugVgCg04iwnaTmPzxOYTvGDUtXAgpvbZENRVH+msCAjwwKiv6ccsAqkBhpBcH24RQciSxb3srqQr8RFXugolXIVfM+G1RGjIfOMhbPPeZKIN551MsIykLYpKRQ67GHo+05JOclsZFrjPZY+b6bMJOjIa2RnfmIupZpjywvI87rxGEe2II1CTVuyx6n0K14TQ4EPvATraF6E8MjAAYQUTbgr1MPEnO6c0Pbj26SXO22D7tIlIfamKcbQwEPatv27WRNiCBODJEJRUIVHejan4ATSRmy985vQzr1cmbovZdNEpsF9vvFbxGS/Nu5bsqNlHOeVREqSTUwMXoAq3GcOMdLad8+AUfpQoZA4SlxLWhOEwWJr72l6ao48ARwnV8UrJw+G8IDAq32ggQBw33opBUcUSOIRkXHmBGcrSojt8uyuOjMHAujdi5vrxJh+n9uoYY/MysK/vNesCtsAMWCKTKCDQ5AClbbiNp+qpiSai9AxTY2ZOpMMur67M2ZlozRlWl11wIwzQ3PFnPEmvnEMIUmR6WattjqUQj1n7n9/ns6/eTBc/hn8MqxMV/fIMcLZpV/5ICUYteuUaIaWwv46X/aGP9Gxb6u0Lh03ap+bb4R6qF9gBrZ1uoi7pWDOxDDsJHatZj8sqvKnwsUYVzHjpIVte+YxA0bLBad5XcKnKLWpeJSmrr5kjK8M2WrsM+VqK5itJ74wmGEphxfd7crxj5P5RyWa8Xq8+23Ya5NTVy37XmylHyKbhGDIBH5Cy7rrlzuQS1201FhxXKrJ+PAXyxYvRrHP4Pmw0dYuRuTcLWCPwYIXgb8PzsG150BWHpG7t8CNc4D/CP3zr38tO/7m36vRk+CpGSj/Yh6Jk13Po4uH3gIpezUtJE7WDhclz7tcw1s9O61l3YXf1LRZUo9DpgA1WwubWNY4/PRol5EiS/EalWUyqFhWIvSZJeIfrun0tg7Ms1d/F5cbdAV5ltEhPoJz1YSa2vfC13aQVJmt17K2HvXDH+quYjDJpeZYP2gQ0RRmpdbOzp4tzXKRFy5hMOjP6J4kW9GUVXlkbHmnkM5rNmPsepqQyJi1btDGpTs35Imzw/86alxw2no8XKxr0+gMbN2uzDbqENfSEuEcZEZ3bdv8//v3xt2NHUscuxb2MaJuK1mdsgex7m67ilJPSArX810ZVznB+eFp1BgSFFiBSDljHaDvUyBt1CfErOnc2e/PA/clTAm73f6g/4lNwxcKjPupr2TljdFobH38wojKg/B6p7m9/j+NYKfoj0bHstreWLsx+bOnXJ8FEBeJOlAci8fmKPW+x4p+cKfMWCqKFjLVLxbKRgt4OOrAKmsesMqvyU/IE8IgxWklyRVxJVySUqOCl3OEiZm5WXatFou65+XaiVkuwx5A/OQ4Q8qokPWYGDH4JJa6Un5HejrcWMProPN9bADxN7wP6LZZY+RZawrCsaUqLBHjCuO0ylmGpitKmkaxVUwWNHfIdeWAWued3PQW68wGktdLE+lseyMX1TJ638qr3nyjmSZY97ywg9TgBWKkO6cuFKOhAukQY5jWiqdu27TY7Uaa4kwwa8Ek5A8OZTkvo2LIcM1rIdxMh4ts6QXfN+Tibm391hPV8YcIJL2BQ78vM63477cDX44TkhVnlKu40kGEXvQVOQXtjKxRQhq0+dJ9kF5mpwHmX9xfm1k/bApAgPnYBHCDu9p+3WwmoWYP1E/N5LxQFHpVJojIVr2SW+H8roeEhLiAZ+O4z07HhdA43yYy48MLcJqaLJ/BhKBsytrYTunk8DnfkU+qruGSxJO5LQa2MS9NeOxoNj8hjimGuE8KrxuCYmpDRsIEPJJCcDF/eu2KIUNlIQ/SY0Wtt9f9pPHAT4DGguBhMlCzRYnM50d8dpwVUICK+FwtOcSmNDrNVtBzLOWr3ovDHlOnDyy5AgnIE8wxt5HBywEaRBFq8EDtEAzTnpJ0ACdR1/e4GIJcXoV/ueGd+Zzh4r4MesBOhu1BRESBrkpIY54Qs8AmKYMJWI41eKSwbwxubDQOPOjV3TOfBhG5PBnbrjtOM0naZgmCqi8jrT0eok9dsz5s9Cs3a1/QmTzzuh7fSmty9gkdC0WBVUahtU+7+1Tvwifk15znsfH5KDWP2Qc+cQWjkCL+aCitDkyZ8sDDMDfgRtVeEjzxhuUhC8cwQEGjIesEo06WBJYmCzKXuOYg1+HEzaFppLRnut1pAF7wcGNHKB+uUOqQfKmCxnryEjHqUhYLfq0EUlLEQZvQgi3g0RC9iXDFS9QHSS0PVhw35pqoLxi8LEzHqqaurjW8k3kXqkwLafE25+R2lu05Ub0fgXZVhsEuIz/rwVaKllIVZMXNcSH/3LiuzyYOE5hJpwA17Y8tnWq0n+7LZVnugLtza+srckp2nihWyLzzsATSHlL5CkAN1Vw/j0o+lXzwk7JW4ZGhFCvGTljaTI8VQZsQiLE3GeIAoihgJO7J9Lad2wQgp/2467q3UwsEDJmZuSTWvJb4ceSYQdOgllDo+uFCio6Ch7xQqi8eCItCzF31+ywuvkESAxMrn0Jzxlf5dKJZidFQ6InufEtV4SS10ZixQlKqIxOj3eR4xRyGoUiAW0JgOgGZCU89IbFFkg5uLrunk2fA/WoPIc2f90ATj+cyAAeFJRNoKYJZqVmq+eAxCVUQIcCnBjyyJr2MGScB8FD27Qx8n3Y6ss4bLTlPbITMKE3HVmTxuceCXdvSpeh4nx7M1+sMRZpynU5FbBFaNhjZQDTBfi2CXkACa8ROEAhoFMinqMJ9ASPwKHrAFsK+253C9qTB8oZ4aNOuXCPRIwmN7bUNVrnUn6qWBZmOVLIE13b0Rme0L6MBSIKu1AmQgFDGEBoJ6nAPYQKEbHpCDrcKAcFzHg8dLAwhSbIc8KlnyaVtHctLOUzLCZJ3tNwEGTr9tDiHDHuPCp8Abvcqt35/XhXDKfRSCv96Z/rEo+SRdD1FV7a+Whm3p91zwA/P5N5cVgZd/Onhwu6tYmLPArH7DQt4xpRFYjkkanF3BYbjncePLGhV7ZMMVdTI3pjoemdV32dpSvGw3sjprnI1/kSeIibRxycK52vgmdXpHZF9ImtPHEPHr0QRcYLTeWYLUYQeGouOJxZ99lqXTWs38wbUtkbR7DPc/YY8FI0gBUyhmm+82KwZWMBsDEcRMedS2rxR4y3hMjFFrfCU6ZYl67qqlr49CtkUPLWJR2WpL1h+fV/zeiSJnKpaZnWM4UcS2Wvt78nbw97W+XuaGeZ7SIhM6606tT40b2VMd9NeM3Eq1c4Wpgp4LgmthekUQdRoeVUeHFcVzIkgm2Gl2E98XaZmMJQuQh25ldF7RlCn1nnMaFHwkuOqSTvnXexBNGr80rAmeGp0ok2H9HSuJZFYiv9Zv562m/3NLW8SLA93dy6HvjHn3+AhLeiwaRsO4kGnQwuqm7leeZHvNALTvO+cbXKR2UOyj2aH0z8dTxaoxae+1xxOY2Uw0I9/wyKn4S/Gvz14Ml0Fd1OH6xiE9iFsDBZomIn+wxYmR/h5+QR3PMK2pAAd6Hn5s0P5WL5sP0PBPlfh99wT05jSM9ETrwxogM5qaMyjPuOtw/Izup+2gLPg+C157xdoN7+vzYkjQACalZ7p+eND77xdu+5jG5wA4RrFtMY6vLHys0x095xzTI+3bw5vGzfDG/XF7KZ2nRzjzUvB51ueLBW7oiAd0L9AIka7CCJWeiDuEEsr7PAPf7ccAz0uESsfMbM47/eSUu/rE1wacatfFa39Zs2Xw2ixyWEX3NuE7BzdXl7OhQhX15ndhXVuVL4GP8NRwGdxTlzU1CX5TGJ5F1ZCEBXnCBP19OP4YocioVSn+cja1oXa1lrb7WZMmDMBQdk35bSehdZT9I3hIBfHgMeUMQ5Q4Em8/llzbxu61ySf6bba1VaXgD2OZwkBF7jHWnEr5CmuSuS1qeZWPIcz8hCpOqmnabQQPeESrRSKzHfiGCOWBlOxWA9v9oPaMaFIqeI/nmSetcR4bZOWpNNfmkMB6fXBEI/GeY6HYa9cktaEhtZrkfL8nboP9djuV7zNV3Gnb4IlD7qra+e5kJeJqk27Jfh02EMaCHwxbT6alOATSoP6L02BFZRGwVCp62lkqkJqwiCpiHDdlU9qigVyklpXY4fPQk4lrQrg30adUpPjrDTC2k/qen+TbIogLf1SNlQNH0vHtVXP4JYbVuDVV3jCL2JIB87V6zd4Ky0PigGkjaTdO2jhnZmCKuy1T2Y+e+R8e3CR/3WQhcZoVvjksXG4M31D7if/Ne4ZYoZmoxhH5WM/AQIQGPGJRJIVPrGR2rl2zMJTHX/PmtkfuzD7XDoYzKFvpxHcwkgYmqjLxtSQASuA1OFUKUUcL77sh9WABWxdPkrmemtKIj/J96GcDmgHnkxvSyi0/UFdOBjaBJyC1+DizLp0L/wLHe+ft15dSHzyOUJU2UUfG0YbdnB4fppE4U7LG8Nh7a2xGUKEA2/nYi/PmGAtqreiRzHuPLIhYe/RPXnmSa/nPdSsrmz+wT3GHbGHP7Ws7iFCuYp6p8cluxoUCXlQ6Oou8aMGvby+viwkBmgM3m/d2jzY4v/ZGSGBGxVR/N/N+oOjYK8POQ0bxfEuKrWBGoOa2g++70EZuxH22ZgVdsCosSu4mpkA9hM5WqkKeBf0giRPaejahoKHCRks5lmWvIJGIn2tSWJqVRw7NhWm/CgB1zar/ek3SqkXKU0SHXqeqy2tJFXyJt4tM/oIliBOyzE1SutJeMlW26ScpFwHlck2BLls3GvTyIxSczKKFOWhq88gC1Yrq43IcLNtuOjCo4krE2njpKM08V2XNyDbuicEq+jICCgtJcV7zRa5Vx1m4xBK0hzjTIqqJXm+rugSORLaPCWYefthamGghNBxwGy5lIssannDk/HK1LRlXATy5mxRzTqYINUIEJTW3FIRMqY5SSyF2iDh9YyFgLCZahhI1p0olrTXetTHqGQz2VHiT9kkfqQarnLLJM9vnF1X0zFy/qSHI1SykosUf3GLU629qoPNwMcwDhxyXe/y4f3/37voZ/3mggWaTSV5rVaphk3NL0Pj5HSliW0lTdZ3RBQSTjGXAvfWv+VwvPL9B4KkPJSZmn775kCMb3x/IVjfQdbLuMqxnQ2+HV0ZrW56MkJ7Cyg0RXFWgsNcYvQ/uvtOxevZ4+bUoh/+6PLmdakQrv8oHxYLDtv07XHFHZPfDoRxXZY1BBc3Wf2MXFSMTYwwBnF+wN7oU/1fl37+hx3bN344+/GPz1CSXklZQkjBYGKg322E8Aa8ye/yqBv+ja5VPDzUj0yEnq/hqbOFbnFsluSqchHbp2QLRzppzin+sUEnkjAVNWhLmV/Gco9TooQWEOGl3SRHRxvUB61qLRGD8KygaRaPlxJxd5WogZ47LrxBiZwj0FbrGtAITZTotXXSWYEzBHdlNiLJV0bxYvRgB7/RxEm7W/v9ecvS8wrVxVbnOobmnfYGsi2W+wrsnjnTkXXnJrFPhUWxF7VW52uKSU9tpPal7sw3gzV2hMvIOtWZJ0t3YOLxipau2E4OWrXCKkK+LrSSwokkY06HOk+nAjsYUe6i2PqgBYYzXF26JHElPB8A4RpJJySOBM887dHeSHI2Qal5aWfqrnS8okskUdFc5io5HT4QrCFoj3J4T4YnlnAvtl6ziQb8n3qj3l/wpS8YxM8SIgGFBAwMEkzwuODFSCXluw3vWOd75u5oh79W5ySVkKg14kK8wDk8n6Ae1hm8iFIaUY2YfimZMXg5wJKqBhi2EQG3P1E3ghaqLTi2D1DwUGMguh05R3ZzSt+yby24wIFl33Y1RDPfxIFbq7lCXYqiO1FybSQ75j80F8zF0NTc0ZLF/VzqYVBHOE3xmlZpCbqep9ahCOe5lDYbPKLg1pGh5A2I4WAweFDXeVUmrNoymQUBJsCMLToqO5F58qZzFyTR6yq1gvLaHd2dQO3qouMEE2SNS6evjoa/Z0XHw6hwcy5TlHMAFYguJXe85kyjg/S4fohDMRMAn9zup+2XNDF44N2DpTEs/Vk5c21lX+ybLW5BnlCMC+eY9sJcVJLjcKsCtybLQFGR3BlU1W3PLJXhqTvASTJQThczSucskZ8elXmI5DjWnTAvR5DaIe2yMw1dmMTRXw9YNEOF0Xmjw4UgmEd86gM3A8dxlOjfqqSRQNqthg/qdVzHDOVDRiXDu/bIHQ6yCmRQ3D5Y6pSYnOmkT+9297ovPDvTRx4XhDwwpjfYJh9p++WQ3lC+/9BMnhQe1ze6j3ZeN1UIFrDEQ2k6PiV3HGdQVfqrJoFU1VQTIQN5oUTnZFqzyuFUGp0M+X4PnzalCu8O69Cv194n3Dy+gGu89FqNKzUsJXja5pAPtTMTxyGzKc82N1GcLp3bqed7Bua1BnVzHQXXE06gghbXb+MzyVd8QvBq1vWSxFNGsoxGJxfr83UgzzTlJ5MTx2vX3DB4vOqhRgmS8/za6408Bmy2wj2Zb0iPi8KHuRW/U8Y4F3DTvy5BRW/pKh33lKPNA1pDpAWNwKfGMeVwbm/WtrWCFMvRUqby5wklXaJnPqDvkOP78xvvbMnzHppf23npiqAGtr+evu/hHXbMjR7xNRVISGpaNKEcro7SQDH1Dz+u1Pal2vSqX+AxnrGtm1WK/RbyJ3bkVMohCfqvHlzxbe+CExuF70L4OEt2Tloi+TUeFqn22BBZPSm8Sqa4PGVZ1+BXnwknzps0oq5KyYgenFkFqekf4sivjEpkZq2KTuGBflyx+wIHcPNKb6U/pwN18jk5T07GF6XjVgrviiTxl3ncfONYC1t/+9gqxpc9/Cq/HPMP/3+JGl6iipZUOCL76HS/RQEcwSByTBR0gsaRKM40MwtRVV6a+mP9Jl/iBauxfFUZ9p/2FpUEgw8Ie8Ctz0g6BkfI2jhQ2rJeaLSxJibTE/Nod+/fXx17rlt6IDRPkSxqYJw3F/7DUyMLume5Kb5rSpJt9fmfF4cMg4+3zVIKxjwOng2US4LxJlv9iewdZXDE/DhTduH896xCMgoBaiOqqEwk+30pp13NPzAuvvxmJ348J4weqDfFZL/HdPZ1H/iHO+YwXzyuMzgm6s6hS0OSwPps4f6MeTdJL5E+Xhp9tST03hnvYIyHoZ0OO3GVmGUckT3ux/XeARzgXFQAnycL81mJqRELTPA9OnkVeODcdG+epXsfSLZu4PwZ6Q6dCQ4FZ2ynK64H4s/DmubJ1DwaSsHi2tL8kMXtkGSM1klB7OwIzvjqrNHoCo7mjByONdJR4qXqz2ZU1n9vj6tsUz+d5/b9Ux99zSzmlO4CwdfHvzkh5J13Ipx7Kr+qT6CiXqlXVCEDvdtFceXp7nKg0oFQ6VA8wovH6EP+RufGIFf5+Km6yQJt3PJXRPaoLKBaG3hq/tHaDfNmdyRjdqNohArG+1RQXEVo7sOvaBcbQx2mS58GT20HycuIfEYmZ6NW1CAsYaaWSdUAr8GBQ8LHepI0lxpdI/IeZWEsPfNdD92U/tAaLzFv3AhfCbtBvd8fezKqaJbjGYfjFShhU6k6F7ttpJUvscDQqffeHnFo2hNCXiZlWFn+DvmbZ5Y5xeb7OOxOQvzwItBJIdshTe5I58UqhUBji2ithOzySL6e7oeh+4BIv01ChcEBAHIj2G40ZnDbfE6uF229XPIPNGiUnRLafJtoUY5Zs5jmN6mnwVQ9+tN8uPVxm049f8WYIemf2/rul9loQ+qzp94KK+UAPEVoxoW5ipISyFAMU09DTHMhgErFfKNknkSTVIpt2KXoQbgfDawrW2hJsZZRSBc2tOLc1tgSzdt5gSeXU9Uv0EgmqvwEKzzm4GLFAJkOeRRVQaIz9mhAZc+mi2Jt0YQaoS+mO9gXbtvFZcRypCZawRMwT2R4YvBuDEGT37y/39fEoWC6K8CLCeL9vaPzxUvnT1ENAmPgVkiI8FzTLof5xH/47cksSAEXHnxVsJiPty/sVcsFZv9pBiT/12aFjUmCvrBXw3tt8uUTVL21/aY+tcHArfSPn4TvzLw4sASE/womLyS392afsIuX7hb8D4K78SgX7vlAPSS/J3zhH3mbfKrNxOQoGmR6ZJTd9brxR97o9Jj3yhED7eIrsijUNuBz7vstRws8Mi+k+5lyAtPj+BXTms2c5uEw4Xatp5VCwvhG+QB6GLqcF8UcZhy7YeKbWY7KDoQHDiIkQx1k04ZO4RySnN8UwiaEFJw0PikewbdrHzgdBBAaclHwWmuz3xKOpxG3Dj4p0gVH52OzfMYBuAirfV75MReBQWlsx+ilFBvaHLQM4YtgKDI9SEF6UygekErDVazvLUamUoLBy6nr2zNtkDQ6yMWYoDdFlQLrBYgm6iDPzmotnRVuEExora0njtGM1WSv4WmHeOnvGizt4GgPRuJH6YccInSEk+ndw9w/yx0e5Lp3gS3Y45Dun3IgzaBNMBpOlmKQpOzBdSg7PeqRkhhJ1W4mJx6bPFCU7UBibvaY4Rz02ljBSAbpOEcUHUJ9lSYB2zrMbcB9c5vrzaEKqrjytGs5B5SbQOUAIEz92NRtzs+SJLIEQzmhBkdacSZUHFifc4pYqdis3qtqYSQCzp+VZ1Oaph5Mag+VlJnisWTgLdEdS0fU6NhGCgFapDPZgUBk5pfXu5dl5ybMFwYK79KI27nYo27LT3lD6DydCb9q5dCn/1PE7Y9IiGjsQDwhb7e3v/n+gCW7+8gKVEzE8mccdEJjhwA2g9fYOoDeXPlHIEDPf/zr+fQ/1aeVf1Od+wcA/joYDQDgbzPpX86+tDTAjywSAB8DQPBXj89sFze6+FmT6+h1T2iCYf4IxQkYwqIRBlct7BgLypfcIIaVrAhsghppPRUUxjRlizpDhEcgBoJaxWAJy2QGvFQNj6Y6vNYNpQmHeWAYhlrlYORG8Gw4ew4QgENcWoVHcxGglGDEgDP5rGIzm/iIuaznPY+m7/UeGwxnJO9RyGcA/Mt/+vdRQKOT7vGYftxJmEbMuCKvASW03P+lXeAhf8qtvWzkqXByg6GULEb4K7A9ZsoAKWghwgZjjk6s4idkE21LghkXYRLUggmWFQAejYoUP+oePuUqHvLPLLzKNlwxSHLDzc4uck0vU7nJDX+BwN8V0Qzd2mUgXpXLNouJUg0vKnq2ZtIXiKA6GCnbg2OVovBoLyBELr/yac2pTLFSRoogofPUkKrTBJSmTi24QOpghKQejBo0gPf0GlUbuJzcAXky+J8CBJcVyXFECYp3NZHKdZopNaAFrJhWaNTawLCCdvDytwMR2/Lrg+10iWuUYNWnvVO7JmbNxfB/CETo0OWNd6mkg8aeKUE6W2g0rR7r0Uhnv+8ktkrlE0W3mkV6uRaPgojsreUBB8RkvyToVFZpb4St2qKgbksdBVtqkM4bwtrgJ9J2QDM7JUxLu/WS4zK2S9nco1q4dZ9xmELCasZ58h6tUqTnFbXqN1TYw6/Qo9qjkIA/enEJHr5PiCcgCOYuj4Lfe7F6dgqpyUy6JKWvYfsVPRoICQQcC1QhaOjxzQicoVsPyw5VTdlIq19vqqt+jPszdChbmRhW9oviaLL3iGtAI6RLr5NFBH6lrjWHcw9TEnPnWPUOErgw8SNUBGh3Lkm8I2qbJtutQkUTQ/cHhmZnnXeB+yHAGu3KJLMbohYCwAa0qtUYjscK/UMBuebXvtHmWxKI/CtGBu3shpUdEuLWZNkV5Z9Rdb+IP5UI8um/02fAYFI4kSNI+ZCHeBqPV5gf7zRkrRl28foXEaKWRDWT/uHP4oQs/O2wWtRHKnbkrzpoRRRVTHEllFRKaWXMFWueOCeZb4GFFlnMZ+4zyZTyllom3nIJEiVJJlnt5MqmCG+LFO5Hss5Ji2Tik1Utx2576KSTYC8N1DROOU0HiTYbJYP8L0ISo00al1homaWWmxeqNOr9br4xaHWywJuy+L//eAUBUbnXNNpAGwap6rJClmwr5ciVJ1+BQqsUKVaiVBkj9Fa46nvXXPdDPoxcaaM9zQPN6f2Jdh6vSssT8vhp7ZzOrqZO2GNDD+cFQsl0UEx7cpJPmL2pXYnmaqZvcKAwG2hS9NJoWm4z2z6wtqO3fRqHeDmJ++py0o5Ee8K5gzzHkJyKRHJF2RGZIxKJmyvKrkiOyOhpJwp4Eh56VicrpkAL57pKpsgUm5JJ6fq8lofih02iPHh5Kch796qOCN3Y0+0Z7O9OC+SPi09aQp5USyRprT2NAxeDeHQlaurva1yZAi0ur6JN62vvjB0rqjLburqmsL5ouCcj1tXn5mtdmhFIwg6hNMMT/eyL26HU+eqGTg8=) format('woff2');font-weight:100 900}+.d{font-family:Fraunces,Georgia,serif;font-weight:350;font-size:48px;fill:#F2F2F3;letter-spacing:-0.5px}+.s{font-family:Geist,'Segoe UI',Helvetica,Arial,sans-serif}+.m{font-family:'Geist Mono',Consolas,monospace}+.lede{font-size:19px;fill:#B4B4B9}+.lab{font-size:12.5px;fill:#7C7C82;letter-spacing:2.2px;font-weight:500}+.kick{font-size:12.5px;fill:#F59520;letter-spacing:2.6px;font-weight:600}+.rt{font-size:20px;fill:#F2F2F3;font-weight:600}+.rs{font-size:15.5px;fill:#B4B4B9}+.gh{font-size:13.5px;fill:#F2F2F3;letter-spacing:1.8px;font-weight:600}+.gi{font-size:17px;fill:#B4B4B9}+.ix{font-size:15px;fill:#F2F2F3;letter-spacing:3px;font-weight:600}+.il{font-size:15px;fill:#B4B4B9}+.it{font-size:11.5px;fill:#7C7C82;letter-spacing:1.4px}+.wh{font-size:17.5px;fill:#F2F2F3}+.sep{fill:#F59520}+.ft{font-size:13px;fill:#7C7C82;letter-spacing:.6px}+</style></defs><rect width="1600" height="900" fill="#0E0E0F"/><text x="72" y="84" class="m kick">REPOWISE</text><text x="72" y="142" class="d">Your codebase, indexed once.</text><text x="72" y="182" class="s lede">Code, history, tests and decisions in one local index. You and your coding agent ask it questions.</text><text x="72" y="252" class="m lab">WHAT IT READS</text><line x1="72" y1="268" x2="472" y2="268" stroke="rgba(255,255,255,0.10)"/><text x="72" y="314" class="s rt">Code</text><text x="72" y="340" class="s rs">Every file, symbol and call, 26 languages</text><line x1="72" y1="366" x2="472" y2="366" stroke="rgba(255,255,255,0.10)"/><text x="72" y="402" class="s rt">Git history</text><text x="72" y="428" class="s rs">Churn, ownership, co-change, bug fixes</text><line x1="72" y1="454" x2="472" y2="454" stroke="rgba(255,255,255,0.10)"/><text x="72" y="490" class="s rt">Tests and coverage</text><text x="72" y="516" class="s rs">Which tests reach which code</text><line x1="72" y1="542" x2="472" y2="542" stroke="rgba(255,255,255,0.10)"/><text x="72" y="578" class="s rt">Docs and decisions</text><text x="72" y="604" class="s rs">ADRs, comments, PRs, agent sessions</text><line x1="72" y1="630" x2="472" y2="630" stroke="rgba(255,255,255,0.10)"/><text x="72" y="666" class="s rt">Other repositories</text><text x="72" y="692" class="s rs">The API contracts between services</text><path d="M 492 318 C 590 318, 560 432.8, 640 432.8" fill="none" stroke="rgba(255,255,255,0.24)" stroke-width="1.4"/><circle cx="492" cy="318" r="3" fill="rgba(255,255,255,0.24)"/><path d="M 492 406 C 590 406, 560 459.2, 640 459.2" fill="none" stroke="rgba(255,255,255,0.24)" stroke-width="1.4"/><circle cx="492" cy="406" r="3" fill="rgba(255,255,255,0.24)"/><path d="M 492 494 C 590 494, 560 485.6, 640 485.6" fill="none" stroke="rgba(255,255,255,0.24)" stroke-width="1.4"/><circle cx="492" cy="494" r="3" fill="rgba(255,255,255,0.24)"/><path d="M 492 582 C 590 582, 560 512.0, 640 512.0" fill="none" stroke="rgba(255,255,255,0.24)" stroke-width="1.4"/><circle cx="492" cy="582" r="3" fill="rgba(255,255,255,0.24)"/><path d="M 492 670 C 590 670, 560 538.4, 640 538.4" fill="none" stroke="rgba(255,255,255,0.24)" stroke-width="1.4"/><circle cx="492" cy="670" r="3" fill="rgba(255,255,255,0.24)"/><circle cx="800" cy="482" r="148" fill="#141416" stroke="rgba(255,255,255,0.10)" stroke-width="1.5"/><circle cx="800" cy="482" r="159" fill="none" stroke="#F59520" stroke-width="1.8" stroke-dasharray="1 8" stroke-linecap="round"/><image x="741.0" y="356" width="118" height="108" href="data:image/png;base64,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"/><text x="800" y="498" text-anchor="middle" class="m ix">ONE LOCAL INDEX</text><text x="800" y="526" text-anchor="middle" class="s il">graph · git history · docs</text><text x="800" y="548" text-anchor="middle" class="s il">decisions · health</text><text x="800" y="576" text-anchor="middle" class="m it">NO API KEY · RUNS LOCALLY</text><text x="1128" y="252" class="m lab">WHAT YOU GET</text><line x1="1128" y1="268" x2="1528" y2="268" stroke="rgba(255,255,255,0.10)"/><path d="M 960 431.3 C 1060 431.3, 1008 313, 1102 313" fill="none" stroke="rgba(255,255,255,0.24)" stroke-width="1.4"/><path d="M 1102 308 L 1113 313 L 1102 318 Z" fill="#F59520"/><text x="1128" y="318" class="m gh">UNDERSTAND THE CODE</text><text x="1128" y="351" class="s gi"><tspan>Cited answers</tspan><tspan class="sep" dx="7">·</tspan><tspan dx="7">What calls what</tspan></text><text x="1128" y="378" class="s gi"><tspan>Why it was built this way</tspan><tspan class="sep" dx="7">·</tspan><tspan dx="7">Who knows it</tspan></text><line x1="1128" y1="418" x2="1528" y2="418" stroke="rgba(255,255,255,0.10)"/><path d="M 960 475.7 C 1060 475.7, 1008 461, 1102 461" fill="none" stroke="rgba(255,255,255,0.24)" stroke-width="1.4"/><path d="M 1102 456 L 1113 461 L 1102 466 Z" fill="#F59520"/><text x="1128" y="466" class="m gh">CHANGE IT SAFELY</text><text x="1128" y="499" class="s gi"><tspan>What breaks</tspan><tspan class="sep" dx="7">·</tspan><tspan dx="7">Which tests to run</tspan></text><text x="1128" y="526" class="s gi"><tspan>Change risk</tspan><tspan class="sep" dx="7">·</tspan><tspan dx="7">Cross-repo breaking changes</tspan></text><line x1="1128" y1="566" x2="1528" y2="566" stroke="rgba(255,255,255,0.10)"/><path d="M 960 520.1 C 1060 520.1, 1008 609, 1102 609" fill="none" stroke="rgba(255,255,255,0.24)" stroke-width="1.4"/><path d="M 1102 604 L 1113 609 L 1102 614 Z" fill="#F59520"/><text x="1128" y="614" class="m gh">IMPROVE IT</text><text x="1128" y="647" class="s gi"><tspan>Fix first</tspan><tspan class="sep" dx="7">·</tspan><tspan dx="7">Code health</tspan><tspan class="sep" dx="7">·</tspan><tspan dx="7">Dead code</tspan></text><text x="1128" y="674" class="s gi"><tspan>Refactoring plans your agent can follow</tspan></text><line x1="72" y1="754" x2="1528" y2="754" stroke="rgba(255,255,255,0.10)"/><text x="72" y="790" class="m lab">WHERE IT SHOWS UP</text><text x="332" y="791" class="s wh"><tspan>Coding agents (MCP)</tspan><tspan class="sep" dx="14">·</tspan><tspan dx="14">Editor</tspan><tspan class="sep" dx="14">·</tspan><tspan dx="14">CI and pull requests</tspan><tspan class="sep" dx="14">·</tspan><tspan dx="14">Local dashboard</tspan><tspan class="sep" dx="14">·</tspan><tspan dx="14">Multi-repo workspaces</tspan></text><text x="72" y="862" class="m ft">github.com/repowise-dev/repowise</text><text x="1528" y="862" text-anchor="end" class="m ft">docs.repowise.dev</text></svg>\ No newline at end of filediff --git a/.github/assets/product-map.svg b/.github/assets/product-map.svgnew file mode 100644index 000000000..2f9b28d8a--- /dev/null+++ b/.github/assets/product-map.svg@@ -0,0 +1,18 @@+<svg xmlns="http://www.w3.org/2000/svg" width="1600" height="900" viewBox="0 0 1600 900" role="img" aria-label="Repowise reads your code, git history, tests and coverage, docs and decisions, and the contracts between repositories into one local index. From it you and your coding agent get answers to understand the code, change it safely and improve it, in coding agents, the editor, CI and pull requests, a local dashboard and multi-repo workspaces."><defs><style>@font-face{font-family:'Fraunces';src:url(data:font/woff2;base64,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) format('woff2');font-weight:100 900}@font-face{font-family:'Geist';src:url(data:font/woff2;base64,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) format('woff2');font-weight:100 900}@font-face{font-family:'Geist Mono';src:url(data:font/woff2;base64,d09GMgABAAAAAFpEABMAAAAApvwAAFnSAAEAAAAAAAAAAAAAAAAAAAAAAAAAAAAAGoItG5AUHIQqP0hWQVKEPj9NVkFSPAZgP1NUQVSBOACFEC9+EQgK8mTbBQuEEgAwgZVSATYCJAOIEAQgBYkGB4hDG7yXNWybVgO7HSDv5+ttNztQux1FQBX+JsyMChsHQEh8fvT/f05QYwwfrHtQc2sLRIwcqe7ZU1RTKtrdbtZ+Vs/AFLLMs7vVgfNUaR0EEUGESPmxai/NsmoNfETLtlhalfjbl5wbdf6On7vAG2D4F8n/0W9w9FWfi5r0SzaFSSZF+jK+Bsg2+abuQnq42xngTo6IVDwJFX3T+sjq0Qjw5DGJzPIa8ICJ/HdA/ABlQg3xc/oXIAkEDyl4iCgJUSJKRF+0yYsSLAQtCRTtoKq0nWl9K1XZaJ3W13lt++3Mul8zIJ4/RO/u/YztoYAs3haBWDZQgCgkq632/89pffc+lK5kScZgQXcnVdWUzgegfFysZrazWc0pS57hxXIWO4J84m5OAaWIU45Dji0L3yP+4zD3/fNCN36gd3oUbH28FfB0D0csIG084xFs+qdStWokAmyE1Gg0IhsgmGVQoiiOLGetVxvD722yufFe9/hc/fe+OXtycJIzJWcqWCLnX53VB8dhQmJStjfkVF1R3eu3aW7w9801fcgboveeNzkPZkAGWYCEYvpSRggRdKgzT5ItGSObkzyQkwfADlvH8U75FHDZAXFqlwXRf96ur3sLg3vqjRg7Ywal0EOpWf1//TF2tyt9nQ6gbGEFwKHOf2UI+QNy241gR9i22CnQMneZgOABJmyKLEsnWTiw3dg7Evbf2pBZn937J4aIVF4nNEJhef8yWZP725XyQ82p00Jtog/1WpapAF4ugPvk/1TNWnx9gcFJgrnviXKUHLl2LjpTduWKz0Xlphz8ATgcjEYAKckWCNHkbqbgoOgQBcGBonPucxcqySlcSLmLqbrrrryivqIpr+jv/2xZpvVnV/taa2oZv7nM39yHsZFDZ44dJKWRZPXMSLrW7O27njlq9VJrVrpXI+0+18yipAOAyED8R+Ar7RlKOmrzAUbeIDVBZI7tKHEWO3IQ2hlgEKaGIsZWZWEOs0QkRa3B9BANncQQmthr89g+9UeYHFBkHdtyzoKAId/2e7+cREKfUzWXIlKMyPzv1/4/9qvvYursp/kcT7ms4xgRIRoi4U3JMat57w8q4JZ0mAw/4vDf5vg4CGdCmyqICEFDRMv6DluNcGS95rZz3VMkEUP0IOxjjtamWqxDdFpHGZI/vCwEv1oIKl3HGOOwppkjgRbUqcxKIkTOsfQ7KWJqzTAbwYMEboDzvuMBAYAB+kHX9MnCvXxWeHL1ReslMC0OHbWP9h0L1TlptNWuKp4Ep/M5rJXWXwTL99X2Na7wlkwwjwseCGL2Bzwv1S1RcSWVVqFakJV2HO5WB0+PzM6+OjpD4wAGEPp99GGN5vmxFFFAMLdgi9KPxAIQ2AjHAGAdGkBAjDKDpA3u7TDIOuKxLwGy8ACq1QBhBuiA/KspuhPIAqzloPCDvZBxkQgMPhkIYvSNyPsZNYR3Ahk4lsFLxam7h2Ri7xxuWZuu52T2vTeyKNjhite7ixrAnxVoqBg+YCuWD+8SCq877v3/OaAosnNQ+DMTEeQdMBdBuDkSnDI4PKPGI/1uGi6dIdAq7Jjwm3EMjXwqcvQo4+LiyRlgFQTdkvX5+jAwDF+xYykxZTvtbJPZ6r25CpYFuZbbCAdqUYat/ZkSRysnoCnY+v4Z/9ZNXdQJ7Xadr7qYOsviTNgCfS7ZwpiwBuXy2Iam5Yb5ZvBbJzHmXiNZK9S5FusYlmP1cF/Y6XFTJ14HV293sWDcGj3Ysbaj6958h6AwVN84dnVPoepmhhTi9WdPZ0uAcHv5Px/dwd6uA1KVOtTK49gV6bzaiBlG0JY1NA6GVy+SPZwLmY8Z7wqlG/mR4pxwEXZal3dGrlVjj9ysM8LbGxAPbMfNYDtsh00dC8SQKsvMXo0R1c0FBYzp3uICuJ0HugoPxu/KIsFcpnOkYFFlg6Wrhp3tjsuJha5DgVYnaOK8pEjy6XQfAMyeggj7JUX79SrWz68Jen5tFvuXf+ZGwNlAiQIc8HsIJszegojGa83KWaWy3wQ4e7A02hnXXQDz+8VNVECd2DsjgH4tAM1K7FES4Y4bAbLQhDFsx4FwEErkbgab6dVo6J7KGTTPZlin+rKwGtajQTe24Q4n5gA86HWwBiIbdv0Ce4o2YrPz9sKTnVTb3yBGqQ/D4w1VMViUcDIFafxF7+vPDSHn13badAp4fm0vbExtALvs/yGv+pCUjjjThHXpUgl1LquSQlyicLCQwv/yNe4cTVgOJOeNvJ5nAUZHT49kGbDR7yhGP3sIDpn9BQhB0riNDCI40WUHsMZfLYup7IowqNR7sGZ2GALG2VUxl62HaEKqDkEVsCE7uDleXIBwdl1M+wWCfdmeft3ykOxa6GdXh/DORgSRj0tYSpwGmWjAKLZin9UP0fSRJDGwIZGt4xzzCUz/tltArftgJCBd8sDJWtaGh9fshiZjR5tj7Xhs3G8Y3Hg+HqGf4FgLMdc/G8Lf/zneDSyHia440P3fV24ApJddTZ1AsF7Nvu6P9wCa1jmgA7vCESYyzFE+fCYQV0kUM7zmE6vMlSB7GApdJ93NWp0wE/7nYUKRNd3Pz/2TOhcPKVAEhfKsutYGpdB7u9DIryp55335cQEMQ5FAUpEzLRT5oVGQq4rCUIwkR7UrRYXdIASewfaI9nLKGgT3FQdAIs/J0YySINiWQWVEywwG59Rdd/+kN71LPrSa2ZTBLgNsGwO27Gf1td2QVFOT8bVvoP9fxoiTIEmKNBm5sPLgTJKvQKEixUqUKoNXjoCIhIyCioaOgakCC1slDi4ePgGhKiJiElIycgpKKmrVNLR09AyMTMwsrCA2dg5OLm6TeXjBfPwCgkLCakTWTW1dfUNjtCnW3NLa1t4xpTOe6Oqe2tPb1z8wOO2VoeHpM2bOmg2AEIygGE6QFM2wHC+Ikqyomm6Ylu24nh+EUZykWV6UVd20XT+M07ys236c1/2837+5tb2zu3e/ul97UG9gB80Wfki0O12S6tEMy/GCKMmKqumGadnIcT0/CKO4n1jEN5uNxpMpzKsKJJmcPIXkK4zAklTCOnysdhL7r/5juOm33aectvYVOKx8DXVaoxWw9xf4MTzbxVUgwS4r9gLYPCrRXgiijk1EeR07BzjY4Yf5oey0pbeDYwhTR/hn6RoA3jouPQHg5q9vAZZtHECzu5BiCgeWcEnSye9tfQjFnz6QR1A/nwABm5QFrEGnaeZHa1ENMYAeqeWsqkdZZ+R0TdM1AHujqJ33fjBSLNjbxDLSMoRW28zHp5NNZQDgSFBBEaVWip3xy7YGghQKuGRm2mpCTEN6+6xxu243HcJ5LuraThANaUpz2qTrdIteJTzRQsr+vOW1Vuwyh2Y8cmts5856xoFzLnKiSwmNL3pAz05v6fW7UJW9BVjafKFcYi4xFu/WE2DxG+FhAODXZ2GxsDH9hwO3BP/bw9f+L/LDFRCsAbCLWx4D6CZgWlfA8GWOpMNHPWKHS/b73j+uuOwib7XPbVscMuqAU6nfLJ0h+hwCqdpJ0ujkl2IS0x4WAgnkYfrWxx10wl8+C2CNiKlvRRo/iCxMdRuO2CrQ43s55KhrjvnDZjvd8D833XLdlyb4NOycP30dJF/7HRd+iEHwH1/YExT3mO68D7zvQ7uhMDwihAgTJUENzuoSfI2zaEXHVMmM8JMWoehKObY4DXpaOtVMWyYOzNQBpIwt6sWQUbZE+Fn3IuHSadq9okKi/5jFG3USeeJTENDfLBeBNGIulAMxsn1uvw3kLBLEECh4eYAUVWcBwV41WMRcVM4ScpFlKlgGiIoqWU5RUc0SKSqqVTJFebXHquZHIUaMTE7JktR+gGBS4l0CcBs4Drm5+ro7HD/lH6HbRwqLzXKGqcnbQERueNzXeAR7P2NYXAxRAiDOnIxc/nTJuNzoQbEVxUn1tCXaiUrCrchzYqGBhAi3cdcxOsz8p/Yg+eZAQt9JtXqcT0E5QnnwwWkXOymoKnTTFZK6EAKTb5wktfuZAAaCwrlFcAxHYY551AjlK2F0FSB56962sELnj2z6YNq0iJ8xQs7OEOvYxozEsRM2SUzItbJ9/7wStYnM/T647ycSsHmG8X6T0oJ8B5O41AFnkV+dw3i5y8duzhB8zSUWkAKvZyyKw8Fa8xkppb9645N8MpF5JclZEOtH1hYtnsr3vXaGciuPRZ7dbtdU26YaIzU0kcCmGE9QoSOfyU9F/Q8NlW100JJ26Zwyt1iREJIwiaW1NQgxRn8Ywg9xpjAg+2MMY4OoCsWeR391/N8ZSP0OyHcHMJzs40Ab4QKcUOy6CrGihBJO1zcFoTfY04bhW/y7ESMQu+S+wG2CjWp9iEjCDVcXgRELY7dEoH1MK7dLgKGWDkFzjYlupPJ9YxeNKKSqHeIYLemksvwd7p8Tl+0D86MyuCrXlFIoytKQ/mNaEqSFl8bQ4fqaXyZBGPtObz5//MZ1jKiXI7PS+8yhaI4g3HIKpJked/f4tHen+upB2LnG5rrb9N8xVXdDhtqhdCuw3IiFp5PUm+tr3ZGb5raGdJly2hBWv32RIyuQZpKnKoS3DcAHE1Y0ochU4YbaNYj8Yq8OfW0gJBh0Wb+v+m458u1Fb7dg/U0Ajg06bQJ5ZX+N2En+Aybf7Slz9vRRIU54Jwo1KvQvNHj05WilJUaw6FZsOgrcnwnDNl3Ew3wF1Z1zk/c71hc2Txut21bDzi2naF5arqpbXa1KmxbG5vtCKjB5QFqKJa/WQDCApyuVsIBjb45BxfIArJwVbV901HfGV8DIr4G4GwFojOCn8nmJflQScoD5effyvc8HEh6UToR6uCEsDkAiht0gxtpIdkdDoF2YvI3OxGqq4iZqf6UQdVkRIjPvkfsq2XCRxvV+9sYL3+4B7u1S7ZCoatpxSDctzxXsXhzWzoIXhFWjevJw/ttMihr5/oFr1Evk5CYXWw2f5ew2csqwR8J/ZkKrcYjZIWUSP5uM9drbCzPv/6Y5XbMB34ijXPWDf6dxnzInfQXv8+sgbkZP5AcYOo8Re++yDDMcfmiXkVj0bbmqWOlnqKk2c2l10r2hkEHjrOgk7EbCc/+50ymmyg3/1pQ8dcGccWRMMvNezy5MtgtuD+8G165oDOYKP024t2eJ6NXZ1upR63MN514fUkezeo/tezttqkRQoxwz8slHJNW1GZiOsWrEVWv81hGR9BLOtuif520116cz5xOWMxK/WA3KurJuR+1fa0b2WxLYY7/9rOCOaoETrmHZz6e+Ck3CS/FO/lQoStJ5t+ZnTuHS7imfv7s1f8RSC05HCK4ylV7+M2vgfpxe+FO9HyR7sY33VtG6S/D0DymYv7M+9l8vaX1a4o7zXbf1rov85Jh1w7nFudvYm5qhUeIRjh/TZ+Be1LVdt81BDyzRkAlFV7qHulU1a2TYCkHRaNdxec+8Sd/47g3sMx/t8mFLdPRjH3yq/eaYdeOztYsXnR82bufo2zrznrE2A/dbm9/dxsK+wiISsuBrQ8iVmCV5e0Vfe6njqGaHg6ND+k3X35s4T/4Kpko5uivWU2FVzw2TiBV9AT+w6uRRc8fkHPhI333XUwoRbqwA/OiRpxy15KXMIOMkXU9Xd9EfMrRXtxfvqMYdLCIVAgU23O68j+PZY+S0Uzdcev6IXvCn4IprCVvfvuOOb4ghZH8QTFF1eHPdnzoSrny31VF7ftbbzavNh6qdB//s7C7jdmiQmO1WdVD2ndLdDPnbmos415MLcKwoCaeTge711y2NfxBu3sj/baWt8K9dK3SrDo/67UKcSchZgBR1KGyMNzO14Rz/V21C+jh3ebT/u9EOUUMyeeA0aXyoxqSD6s6zk42buVHlLAzpBG5UiWUH7aBb5yDdCJHWf/tJ9z6wdOEyN/ognzny7cBi6f19n1Vrcw1rh2rq5MJVb29RH6dG3naTzjwyec1c5I4Hil5et+x6TphX4XrmPvhL2fjcypt+6pdof9XH8Xqr2C+sGiFuka83I1HHAXdk30XDI+MqzOOOksAJA61oFsBRInqYRB8ynYFHZ2ymxntlAHWbBoq9uni/g7g7DTe8uf8zmhxY70lOVFpXoUen2GopVojRvj9A6hyZFEiTEEEeZZYlrH+tDpWX903bjFUOujcZk+o+mdoLNx2Ubjh0J03CCKVpRWeEbV7FXrenZ/eZPEOXhPexvueExwmzbhUr3bW7occPfyv/jrFhLYlxV6OkcDBshBspDkwyhOzTSPZtajAopXeiboj03+vqPYN839juykfYxDwKAYNKmoZBXsFOrNflivrco6P9jdH1sOdmiGDvb175b+CSO0f+DF5jyXMBq07iuNBs4mQvr18gkxypic/k3mvsfqwtFB1zb2kP+0LKD2KiV/ShveMDrscGMaKTNBjnsdFLt8MgGKU2GwctTMdBQ8ETaznokSvf+fDo58T3odq0YbN3mudfUrtI1R77es6ZZc2eHh3Lj8ewwb0D5lLHig7Kq9NZwdZ1M3bcxWNUkV1hfXLMOwP17SlduQ3rzwBT/Ad5EBPHr7Quu9YWphaHVVeefXs1q7zxsT1sUPzd6b41juwbnLKcc2XeLBwcAHP7v93w1l/WeNb8thfG9GNP9nHOFXk/gEF4/Gz/QGuYzTTa6AKUOsjk2Kois9qnLvUud1gUYpmtGloOD0B6mUgcWGgAVGb7+kPKjT295Nwf9tXNG5oUr0Hm2D0Oqcyuo8+O1NBn5ZtNpvb5suBwCrZvwXwQHv8kstgDfdjTA32weHJkstfR1rNw0aKexW0O76DeCYd/cXqcv0VhEBs/6Fk6rcjtcQsAM9vW59FvGOpVb+wPe7Uxmb4lSxdC5mid3mqJQ8eYHalhznLqZEJLFbOudNTztNWd/PHglzyXnQ2U4/7QruW7Qn5A7a61wA3vwu+CD0Z2gM02tSrkCQTkRSf5rF5otsO+q/6rAzMWgUW3+4RdHX+cHjPHqx61jlE3gtiW+3DLChggXQuoqQO/k7nLzIw1nPuIBlJ/tAPT+BBm8Y/L/Rv2pxWAlQ33+ZQbB6aZyhWYXj9twOL1Suwa+qzaWvpsu1ZCoRa9CGe1nMFX2Y1tY3v3th1t5JpE3KaCRR5PwUJuFBrHy+JpJl0J60zjaVkVfC17dbUH4Gs+0I+/hD/55X3P+7/tNJ5xFYoLFQ0GA6ucC5qLa5sGuwwarUsPTwM9r8PHJ7v1eq1bM/n4fdzdJVDUnbRFu4Tua6+P+8fB8MAct3+9eBtb1P5QmPdZlvx2U/hXyTR+tXtOd5Og0m5meRSmZJMCqc+94M85Rlis47yO75kLOXQqtUMLzQXibFuvytiJNIeQMY4VogoE+kICmXQLrvDkW3FhAcMiUkrD7QKrtVWubEZrQsgwVael8nQOtZB43krQYf0FLQr2ZIVU5Y9zAaLnGpwKX+sBrOzuWSuGLvsvB501NpuzJrgNhxoGuP/N6cjYisI1jGw40Ok4Hda+hnHoIwgxCMP7ZK6VW8gmEJyMHW2IvVo6BaJHG082xd4kBEAS6dr66Fpbx+h360lA4F9WMgRf//wlfB0Gv6QLXORKDcvghvw5u8gGpVYmuTIWRGVB3ZFwQ3tP803sixwS5F6VQfqlG5XTaHLhqVaxmULWsTlry0NvE6kMsC7dE1i5YGFgxFNlltH6QvQnyZPZQWpZ6kcwChswWWHKR0S9qNrm16b60t0EI5m0N+hFZnghrwayJrqo+1KNZlellUt123MMmIXFREb+6adrzUU3EKmaC9jTK4wluJOXUwbt4zBt/uI7xvIKU7NePOzzkXrfNZBLb2+vidjbtHwjhxJQKylBAxfI0qG4XjEdhhXDcQOkqKlQTB4NIPeapEKrpy0SQWQVNsNPGx93VOWv5QJM+h7dTCxD+Z3F82nwuD1l8PIeirGE4uLFu4gSVKopASOHr7W3RWrs7eBF5iz+lFkNPbOVG2Z469F51rq0eXb3MHlN/f8x1k6VotflVvRkZ6skyKmGiD46rBjsVFtcVnekzuEOWSmu8fnsn1HAT7d16uXTYVg+3GGwGts1iuFQ0dwU2O0zVul4Z0IonKkv/K+BhrQ7zAqVEwJfV5ktkG0UGi1HEatun4FOQICxVlZDVRm3BZB7z6ssdpncz/Tyq/RtWnmvzS7radHqpEFhCez+Vy0JtnKgTMPkOjAZRXEaKjgme53H46wDsXGYufq9R8ZipswpYIQVPljlEFulGrsDUgI39M3Hi+9D5WcPQAdgEH8D+oTeyPtr8SRIg7Wa7W4QHoe/QJR2X7l1ndey7VFrfCnIKSsmCAkzoDlWCLLiGGXQLQgwsI6k/tVL5QYOV6BHajEr1GaeRhZoaYBh81XCpKfMSg8MGy4QV6MA7v6qF779od1axlS5Rcyw1AerIJldydSSaDq+CrDSrZ0GFj6VS9ZSQXWoWnHxGMdtkYgNW8kvENiS7gm+OXN28HWPyKKk9PtpLXRcYw++MevYxS606ujDfjd9htUkcBWpsUGVZAF1UWF1jFQq5Ko+en+tBzGsZBLdClxvfOJEDO5TRapVrpiDeyW0QmWbi2kqQcwZS33jmtHZWluLjkaOjkUJaLXkgJ41slTET17z+YQ+YDkaF4rtDpa+8J87u5JXXpPoJGVhgYgZnVQi1svw4So+fWd6KUC/AaeOzUGYSpiHsqsj1ZX7EIPOPy5xdSxyQKu9KRnmyP2bAoShPYs+X7wHuoJe0Y826HWyuL2T7jzMxg9egYAo3djElbr5AWS3SVKt0BApBbV737Fn+hkqFqfSHGJqdUGO0D4jgKzBWCjyGaSSj159NDkjxFCwuWJdHRm4H+pFjQ1V2mi82WPw+U0pYCvMF3jFdfOmTNO4NS5YpanosOPlbBdbbawWcQ+rlWaL9iEYctjM0HP4ORVLHoG+OQN9CwFUuqu2o84Hqx1ii0xtdpuVMNY9bzxbEzKBR6dyIx98uoQRjsGMtjIzHHJ0eVb11SKtuR1Gx8MFiE9kH0REl55ZdCZ2IPmYHn31G7KCRW9AqtI1UWLZKXRkzaE3uzt/C/95QAQ8sKvXBK1Ld30QAg8Gq16yKntkiUkkRFnpE7Mveb5TWm2xmvQzMpempSsJiHRcGbE8mrgppxdTCzy5bwzmYr2ANeisqXO7NZfTEal3uSN1dg3zXJXwPLPiU4HwMBhsLnjhSqZg01RHd0ixsadX6w47nF3hd9TbA64r5PAKTEba7ECIFGoCvrn0QgGF/1GA0qzDUAPtFxfDNC6Gb8FAYONMG4i3S9I+3UWhPrzDtfkdGHMImd66aLjX2ioUR1RishPWW2n/9xvZmuKlMKBm9w4nUgJBJFIDSzQsW+HvPO77llKOFJZSv9EwJEapTm8QJULINNdgtNXYzBH7FDqP252F9ac1tzoN2vpePpj/bsTp+z5y1Q/73w39k9IAmNnGFr4yBIWQw06lyuiBbQKyRUfjVtgZyAhnZAhmlSMnrOTnckOLWho3m1WJThWozHZ3W+QrmoOl72j6OnSu6gBTbPsihD60VqHw6jgLELG89Z1uitDUqpB1Wp3q7oTS7DX5gi7I7dPzbDSWmc+n2cwV4FhCeX3qJWd4LnzLqbM1vKF6u2xdGLCyTVO52lqjoyZmy807Btl1UkmgQVBd3SgQB6Q6q/1Knj03HHMYe+zdXJPXFNLo1S7I7XJDaj0vY4zIDFKZUKdQyPUKIdUB85vyvujlLa/FLrkC1lX2dClNrxlwW93+r4a9VOusCarwMYNSqvxoxze+X/D0Z9TV3xHkjv+UoMV8jf4elTLIKCRqC1Pe636eXcr8p5dOONdw+le85F8Cltzkm1JXB3eYzdabuSkmBbvSbDJwLCz6/8ho1vAd4MdVf98GqbqX9UBQyL7I4M8HG3UVl147oKzwvJ1XrCikCPR3gyaDRq5XCXji/Qv+9wnuiByw5M4xzt2CSga1obTMwyqSFZUWFRV8TyjPyLtXo1XIq59ySdQnJUKixK0GIl14r/DAA/UdsVwc3vFaYiWgy2WmKs3dAA1PXkAlpJNLKbtUBbeuy9dkNI2tKFQrbNJquxFUyDWuqq/3//wq1n5XvvoCPh+dsgXc0aqKpwo6XIfkKexGSTt+wFh9Z0guIr+l2vyQwtWIVOoHXMrDzf2byCKrY8poy8vKFtKobybjfcqhsj9L8n8vbrHf80v+BP8OTR8Cd9aK1oqHpnUrtoCcKyR0fUoT41f/r4yUJlQ9KXSF4e1yPBISKl+i68vvTb9lwYuzFjZc3dp3FV+l16sguw1tDiGRWnVzh8yqjQrFNcqKbU1b1nx9lNx2dj9R6drXgpVir4y9o3nLksNVfm7Sf4PA14se5LChqqjj25us82Wlx8vwx0rNlnEsLTueAW5LDjQueh96P0w90Ly6ZfWB6G8iqjQDzHmnLatUn0o3Htuw4tTLPSYfaAZT4nnJEbel4K+4O7CAwWeuNhdoeu5c5QR5b8FDl8N5kGmKOwzwCGGV1crq3VsakMm7kjhisbDhHYHM4F4cjqoYyAlXumeA2V3A6sEAVG9vYwkHxB7xgJDVZt+vKBB0BjgOTsQm8VvRKCtKErBFwkkBJzg9+NERpQELCmVBS/1nEdzP/OBCXBPRmP678h9YFt+Epr1qjdww3MjyX04E8etxUmDbfGZJWhevg8LpuOQ4sN4Wha2N84M9LBYQbEiO1SUKHv5Q8DBRJ/8TAeiGzKeHxToyn7LF2Avl0aEei+b9GX2cbd6+bpnOk/BKX218hTr/HoN0fSZmduamIoEk4tLRu3XGikIRBf0uJmJCGdOnoBzF1Kos7Jas5sBkar+7Wsv3cTWN5NdiMfZMr1JWdjdp1uqcy80ZBppgTXE+b5KzUrwD3Mlu6Y/WcIpMSq0hfh9Z7oj4J7t9brNMH3ZVrb1cmGsTetHplK4TcEuz/UKMoHIJ9QWIHd/DlUauQKeQTg1cVRRR+imMQzTcm0tAlkDs4/HDarUwHOSJb4+7h0xDQFRtiHJkbp0fOcso1XL4jfnlmkpntj7LSw/Q2BNf7VfJXKEhQlUpw3SJrdmHnGHmy8oRR3HlOmueIy/IClcy3lz12jCHJdJFqOUSbgeZWAhhaSVf787PP28qARXpgppqzZs64+aNc2oCgbm1lo3xTsumuTV+lanJLl5S1hpWNpMRRUtaEdXijLUCuwhSVtSr1RV1vqugwkzh6JyGlDCLwRc6tBqhk8+sTtTWVEudD7B3y6IVEMSMymSishdlHt/ItBIFprFigaZOIWpSKmWHOrlG0yCntQYF21jR9MOdyyov+c0+XG3o4/SvF5Y9noR7UCYEC1v6T6+0TUPo0MgH6qJ1irrlNOktaSmtXlE/JjbAeodlZlaYKisrzGYGS3oYFeaSqXUzWMlZVFo5iVROo+LBn7crf/KJfwLJ8Yg7AlKWNjmbHI//NHA7XX3bpD94GSEoblAMw7Bielw/IjFZG+X1Umm9XG4qn15O8zIj7UEXnd71gEZ/UA++/sGWMwYWLBgsn79g/iD5CwiD3sBKb8HApwRQoSFoqMg6TrWQSD/ANO8ka15YxLDyqCfQzr2fRBmt1WtJ0zNIe8RsB/iytPlQL6YzfgeqTybxqa87JNnY06tn0j1l8mCpRUeKSLPyGaRwN3I8U5XR6rS8oqlQw3i2vlwqDtYLlbZ6ntCVaEIdv8pj02b/Ti24eWA3tCnE7rzzbgA4RpxHnGOv/V0wjXQZqysnnN84J52x9AfgmkCyLLg0kWh5lMXxAi2AJcRjS0qw+PD8Ze2/ZckfOC86R2DK6IHLogMzwb9kR4Y2+8+c7wDZhLtCfw4PzNB5KL39vPN8+piTvxqw8C/y6FfSFmR68T6Ry2NultMurV69Lq94U1XGAvZod4ahDU9BvG+xu0PQH8/LOACIvPEJrLSLbBKNZbLJt3ULVkIiLnXtxXPgnbjSG/b5vSGl0iK/r5w+RpwiFk8hrjLzls2n7beLq0GzGECuzM7KzOwwJitL9HqPswfIl+7CrCg7MjOpshYn+CQWDUb9XAUTFAtEvtCsOu0scMOnVRHKPvJF5tetaHI05dxgwAqrbEamXLuDRN7BlzMhI2BV0w1097t4RCN2EHuQg9bGvJyMTEU2diivqijMKCRmPt8tI09JqWKowcZuS4YFGB1K4T5s9t0AG37/19zX5nr5F39dJkVg4sZ8lzCo/Oa9f0/bgi+Kiv+qJu/vBsd2X6dMtzpL/78AuEZdp5vAsWzu1DjBrNdfYHjA/Gm8M9P6E6j8afneso2fDoAyBBBmnG57M5WTUDlRzSsrUDlWJSunwJyG9K7PvEMx01q8ROlsye017COVEyxXWSFbGKtavQUKuyBULC8oL9ClfPjJZiJsJossJ5eTy8lEFnyWIEoKLOeX88v55fzNMgLKq3zGc9BDE/+voFIZGnu3gqkKTqPxkHmhdWS8b2lowiqouEPGfkDPyGiRofFNKpgoPnmlBPshY1MYmhndG5xaWpcBWKemARuGxrsrmKgOVahg7NuRBcwRM9jlt0EFEzNDlVnmwGc4eJ/xHiitCMLcXV96M5v3mPea95n3mw+YD8JD8JZ6OOuhAjvhPXgv3of34wP4ID4ED8ed1nRTbeA9kg0DG+YW7hfbsa5fGRzCTHaXH0XrfToPejyM4nEiu9sAYBnsqc9YZxmHhbItL9DGodG5pN8P7KCdyVia76es8AJ8wCr/3Sc7DSwOqYATOIXTOAPzJMJ48FXdXVaJ3xpR0m8r69fO5L+RwsSfPgZZGYrfwGyytDZrc4CtNUtXE63sN1xsbt1uBFovxe5bcDQ+3DjhVnaWgbnKn54P7lbEXQbm1f7p0GW03+ZQJnYSz+5h2AnMlayS+Yf3LROUUHACcyWzGnQZgA3ymcUdV+5b80a/M2A+3u46pqZUGp5qajX2NZjX/NPgucXPdoaY3/o3yIbFBxPdoYNSUDdULWG+tmffI++4km5g3ug9QyA0CcPXCeecccH8wX8Hrqrqz40CzA9e1Y99V8vYDGhz57MJ8rlafsTwKsztPuU81kO+JPCefJNx1J9PPnhVFJ61y2yOXltn2MRyN+RHAodgbnd746yH3C/wnhzjwmh0VaIz9FINoFCuJca/HR3mcw/tOjkMq8wQ3LPi/CfbarmcBy7mMjpq4/wJ0e3d7JWqH9meH7UFv16bq9kdnWFN+T3YDa6UxsU95qqq8bxdZ1NNdXeHONQUPwTDXxm1tnEQgDCwsqkA1g14VO3cAtVMPOrlGwM/PapsI7Y52+W+7QZ3esZK0xWutb6GW9TujnbT7rrCKMcyq+aDGZ1dM7b8Pp7yiMuzlqybjECiIFcg1yAvSFelX6S/pcd+NkqKCqC6UTNQb6CT0FnoYjQRzUBXoRXoUwouNZgaS92bhkurSWtL60ubnbY8bT+mGOPEhDH9mOmYxZjXMWswH2P2YK7BX1RkOid9XvrB9KPpzxOpGTjCkv92b2f8rqszRzLPGdys5Vn7s3417qZeppHZudn0bHt2Ivvt7A3ZO7PPWryc1pxDOd/nVuSuyt2GBdg27EheUl7HAqaez9mvC1eOE+F0OBcugnsPdwA3MSmd1D/pcH5uvij/w8V3f2v+ifzL+dfyfy3IKVAULCn4tuB2YVqhpLC2MF44o3BZ4TuFHxfuK7xeOFFkirYUjRV9Tr+fN4qTinOLK4pFxdrixuKB4mXFHxePlXSX/FOaV7qjjFa2ruwiPgNfjmfhZXgb/l38y/Ke8pMEPMFMqCV0EhYRjhO+JtwhYohUYgvxFeJq4ikSj+QkxUmzSSOktaS9pM9JD8kccpw8TN5O/oz8iIKmTKKQqVJEMf/CYU1XNbABrhlkgHxQAIUAgIdbn1vLgD560HuV7Ms1cOuxfYt5kz+e0TZhI6h/rcP9r3bsvPCv393/dXgBOYem4GaMEnaxCJLBw3+cIBUgAYDuOZBCtkG8gIhZJBBH728sZ9BegAkoMei545riXwGXMN9cgO+FL22Z7/HNQd/79L3L18H6Orj1+JbmMRrsG2Be8Un4ndZ3MNO3iF1n7G5Wdz4DP/Jv//tCCQ8eWD0r0EYg5YIXrXHpelFKAulq0S+pzV0l5o23FjKm/8COpwQ8SxeSi378LvID7omvMVTJJmXOHSarI9qyxvUggkROH3n56xdNF+PPbz3/eZPng/DzrOP5SpOGpEXDZd2Pv23+7hn0ektxHaX99vqAlkDsqUDWx/sByYc+42Nyt0f4aoxjV8bMafeeY50cbXGpF4eQQod7dJ2RB5kd9AksAV55Yoc4f/8Uzda2YZy3a+Lj9St+2EgIre2Dk+QF1thch7xvAbyxM7nVadlTUzw5AsymgF1gdmACACmk/f6MCWcl4Z20lcYBATvCWkCu77wftIA/CrD0DhNjlaiLk0eW9Be12/nWyfsiYrghNri7axeWgK2soa4VyJrIAkBKDTsmKf1n6XTe+vTuxetAmke/pOp566WfHVJESonxB38HLCEHaT48eHYzuP/s0me6A5nlTHUErMFSGZoq4pPB3i0tvcHT+sL/vV3+OgKXcGrQy1tObfGU0QxXtxrOgu3i5OfCqkWKYqzG5HqXRLaYkNAOPJ/EuXaZWBlCqAL3mfryQjEI7sgV8cagrg/aqgayNI55Tdf9m0fqbHIFJve36tr4lx+vVVUEj4rM0XRCqHQfkGXjDbL9E+Q9C+A3vrdw73hXPG2Pxc4ZFw4UhXIo5zRC33CFxeV8hg7RwrtGaLhM5pig+xgO9ML4DDr332s3Op2K/vx69R5jsNow3p3xOh0OTBYuTadOEnn9JfAShH+q96Hc2ASbnkzcOe2RBigeBbDRqa8QbZN+fWSlxVyahFzmm6x+4CAas6gyiFhBGbwR6AmttUTo7dFUtyjWdoyahlca8IFSawO4M4i+kN/wb4mvvtJk9lr8EpLEmxfhYMrA39j4EzqCxt/kf4FJ2NufP924NRUJ2V42+MYkpKvEIIXEsFGttipTeAXeK8HEMNzYYIVyYPgaf7DCEvzFn8ZOmtxdKR9BxO9+l3GPJJhzT83sgcuaeXJz4X9Wgr/baC4yOHllwm49yRUCuDlRa3V3NFzXIBRAScOH9NvqGK3+6ZsEAQ1e7UD2Rmfqp3D4P9duEO2/aXLua1RwGx7okuU1ZjPAMzugA22KJUWvD1HwNJEShfWS1E34P4tLNkhIZPBsFOeHEiqMi3W2D359GLauwQO2DvFk3QCkMYgrRBwOv21+QX6vg/DSL1svhcxwC9am3sD009/+dt+Au631/Pm8FELQ35lna3MygQ+CHPvSWukd0p369m+hf6x3aEZFRGVoqlWLWtR1WmGbNxL73/RZifSk5ErA88ygDz/OEEC9YVl6CNjt+or69GhXuUj+ylYnjbeQeAtPlIpKgf7RXzsED/zWDrzbTus58fmzrcQ5ITbztLgseYO0IsG006ti9t+6X/B6itXXd+bKmzdWVr5iPoOQAIp251BwTEWU/JyrI3qdBdgtO+uUpSP23SgBerGDl0YjcG6GipwkUCLb8I9bX7CXso15rDkrHaQweKB5OXnvEOuBBzyt5DCNSgMw9biUMqgzCvafwLT9tdgHry72siBD2nrk/66cRle+MLoCBuMd6BFZWbtTwtzp44XJYrgjjmXSOCg/GXpXGtjB/kFWBb7tgxVeiZ+0w0lEorgtriuu2FHT8IOsfs3CD4A9Xsm99i+h0nutO8MqRebEM0YZeqJVAeGw1k20ActSatI2qKcK7Trodr1IXe0W/rzXdfQqt2BHka8lN7oPfKwdtxngARn7frFJ2cCYdHXfFrqVjjBGf3X3l79r9P6AD8ivrNvgp4y5tZWCg9ACFsl77W5LD0SX4140TQRNELXMLcKKE1zDyMD4pCtphK7naBYHK6NpvtTN1GpBQHqSHqaExTShLtqsX2/XcSSdEbOCkDhmq4pxVpqu2m4OyZNobFq6yIlKaQdzszKo9CveCwRIv5zhif7nygjqAW9aMLK0/7n2J2W68jPIDzYboSrmUZiYjtsjxVl/xa3uKvmcoUjBiuo4kya3xdq6i94OTeSQOAprTsWuAaNjALAEZBpQMoanMyMwDoB3FXhL2IRBBG+jU1CqIgrLJwLRVR1HP8pnKDvvlGinXjfAawCYyaDEK8d1QPeMIZzmYKI9budiIORBgavoH3MPB9mQu2rVhFupM5ivQYN3mCS5aOHnisf53/ZTpVgdrtqd7rHnKHQs2757Zezh8p+YrMF/N59Jef7pUO2Hn881lVXJL6eJf6c0cR03YSkjMezXHMgREu7dVW7Wyd26VakDNYOGSkO/cfu2HF4W7+dWIubi3V2Sz+Wbe8C9j//6mqJDHCuYit+o6s6fDeJcVPecjiANtVwkdwwvI/j/WTnfjuefFq98WrGDvWNOOSAwb0HnDcBocqbXIQQSK4gPxbsFqbhBHCc26joMB3qHqA/7XkPkncJyHeE0NYtFs3VEsUWUzi5hASA0Ij9Jsa1T5TkU/6T+i9chBdZMgWJ6cgAXCFbG0oXnEoG5mhjfCBoKhlYhKC4QT86mBFxIOSMup1MM81m4eeVvlJkLBfl5Ci2aZQuBZ0UvmzvZwPDPr3+XXWFmTvp+w6e9HtLIgYJSQ9Vq6Y/zSm+bQUACtt/FtFIMkY9wubQVMY3PAz+eGeQXSsBZPP39MzCdhaeQPxtbZILZsCDZFXW7pe5INJZN2prFq5uCMIwpH8daHkXceIY1CstZmCpKLpls0MF/NTtwZpgXNzNljLYhsQLxSWDL9Jj3AMMmxXkwCZmBUta84gDXhLd/E3w6AQQRwQnsCumE0PXs45AlXLsQS5kWHBF3tlDTKfGgpOdiXDy7uKkdIYBmSvedAXAjyEh69MVO/OlcsRzq/P8+3/cAhVFLS/bQGx7fc05c0taJXD0eq42vi6B1sIb0a+1BJ6k+VEFZLuV+2kTGyyWk0QFDyARgAVtbdUAyXWO4waChGG48Y6SKJEwf7EjrmM64PHo/IR6cbsZMPGopacysTZJbmAm8AXkyCPdYDFavF42V1eHgM/KpstNqtHo5QUiSP9/6Rd7zVVZeUy4nxP050mDNoTOlOjDS1uDr27G0V3UHKpZobVs9UIEq6+MAg+QhIE7r3H2IWtx4A/cQrayY2WGdkHBceDSMk8p21BMDEAML4/Qkt960N6MLAd6MM9dKw7cOzGxuE7t/nNar50N7SotCSaEJYgpN6VbRmMu33/kgjT3FbxRKjUaMrY6Y9ZdUvSDs6YHn+i0SX5aYRRh4pvAWcmWgqcP/c/jJPlobu2zckmCy3KQGXMbYCCirsYEHzCm5xhkobgbXpmiqi+OPtN+pNw9p5sTRNo7kR1F+Ncyd7tVOeg0qc7MFk1XiaxhfQ9JkNcbOXGwFhOCHiDa52Cum55q1F5RVuWEMrSKosnm/tAyJ4rxHsZov8FvBhOEKyF85FB1agjhUPgP7Nx41DZjox7cwBLB+r5ONwcl7LMX8envLTdu6kkqZoYamDXDbjuTwkvwGt6erP8zRYM499UTFHlBYvSKqxnbHTZslVYEY8zNf6dbUF/2n/37lluvJ3EX1QQParH85ndMkvHbKTtJ/WiifVsio/ufpwbtjDKwFGFnKu553/FsF5Tg36LkOBlRK+guJZ8TLTNsJruOFC7Wlju2f1+RXf7eljHtbjJaRtwKd1f6L+ty2wZXI5wDrPKwOOAukLISioAAQ4DFgoHOafRxeewmYnTnmw4tn9+KZHgNZqdrTsWekPi21qAQHRoK4Yk3kmsyZNnhqDGtMD6GBieDHNy6rD0aPopgNR7H7bnZmDPBBCV5WDepELQ/Tr42ndb8mvJGbSIhhbAsGFLKwFwT3F7LjIkqT2BWLzAqmUJtzZk6jbDDX8k5z7aRCjI/ZE6IStQ0JFx4YAOa2YDMewGkbmJb5vipbalXJfu+Q5KUAT7bxjZ3qdRZJcme/er8dyHSLYMwMWRDBgG2RpyLuNca0VMIWBkTrYQV4qybLpyTb+8xcZwBs4FaKTqrBSW4/G8qFN2tTuHS4zwDYJ5KVxqw2ie2+oVXHQXGHeW6wOlzrtIrAQvwDXnRPDQLMuMESNijg+05f4ou9qxTigauvk5AjcjRHRobM2wFRdyJQvovUK+LlsyXAV9SoHYweR72PZzf38dLmpAcC0KXnFj0zvyC7xJzs+wt91qKg1zlQDgSthtDiUN4oE42IGyeoLzMoa5flBOo/157FWKcnJo/lZ8jrfyB6n423NfvfvNES/PrLMgq4jqAIijifkNvSa89Xxl8+vbhCAx9VoK6wFWfqb2FLxfWrjfA/fNQLyf+SiG+8OIusQAJSpKOmvo2QOp2oszFD0B4tdUu5iR01Z+ADfFVfgfZYhHKvo1TfqAmhywfVavWkUacEpL4e2GFy1oZ0k/xhF1YyXRd58aVUeLdlYsijibgiq0W1TZjQsAdCHIvdYau3wow/Vv5SX1MgsVg2XqsiIhu0R2ipVXRbeTOhJKK0VwvfaslHthmLXZmQ1xq2MIyIkIeoe0ptGuHLY+W0r1o7ps/aYb+lAROEfInQHUmXZBQeLVGL25u9BR626anTPCjes/2S25S/KUfotfOjjbK32ktRxL/IW8NwdbuNXfOIZkShuhGE1Fd2Z883wPX+wF66ahB2y2mSE0FHYgh4HuUyqlBUoAV8Zgc/MWD/vsvvsoSVJU0O+brC4oG3PxDN1yQPy5bRMeTDX+49vHHVULsWoXKJXI6gS/eqVef6tm83t8/37HC4VBsGvEGQbPAy6rIyNa3k/2+rRbuzdmKaquki8yrj2oSzg7JrEKespiR+Q0mt81Q/lg8ISOwI6UlglDrmPRy4sbxAatIrBDxM0SML5QK7NRgMOi5wu9gGjmlysnG01Et6wV4f+ZMRv8c+2CqIh/vBSyQK1E6NILUl7HLEnnHoaNs2ovTftqhAJMw0pUNcPjKFt5YN789FUM1R0hrwtVZIku0cqJkzknxl29JReSaBn8aOQLreoSgtfCbpxcp6vMt1KYIY+zWG1PY8KkbR/l5VXLHIh/tJnPSGF8+BWmPLRIJnMjuyTT6zPlhvlxlG1Qln5craercs55uPB99NKF+gDhP0iYSQarfKSsZJ7UFXenzpWJD3wgNwgStuS3JjdLuCrjKc0jrHqKZKdx8lWX2gJjFrqTOSsceJ8bZrTFhT7aqltGTuqdUMNBq4BnCnp1NhA9RpwPYd5pnVtUbeLNo5QU+aZ1Y7jX4li/H+hmRfmfZWz7YPI2a8F5PUOF3ZLemjMr3Q7qRNWbLJlowdy8rqenBGZT2AuPndlVYn97SFJGg+DY3FVqvbnYfaqRVkZqp092uLE+JfuDjEEyulyIcxf1jdr4RbmDwCmXhtwmZHzn2Ad7CDNi5YyI6WgJN1rNF8BE43Hw/mEipEsRB3Ceu4WeiWed0lEYJIb4BhxY+k37eEKhC51SGZU8Eigrs6SmYqR/FxKnc4Gdghyss4OluWY+4qjJ01k6TqjJ5lq/VHknsO3RE7q0pW70nTMM7bFaun6T+bep/djLr0ZOjnUcTFhFTQlYf9WAT10MDWOCZdulQ3C+FKnqjrj/t0MFXbPK6q2w0ThaSPooxxBpxTCbUrTskYY1VKUvU84rua44g4E9RodJqXBs4K3LdN242yxgxmslTRtefW7ePJlqShT2epZNEiDLqt4hNKHYh4C7LWbmdRyEVGQq8jY7i8Cd8w8HCu2W5UmRDisaK81PCi+pj8ADsucnRZF/BBIJqQIu0snMmk08C6vEi9/LGdawtxdYqvun1ITsjwwL5rTaNx7IKWpYzWNH6cfihFv9iDxR1MdDxA9zcTW7JtFx1atJsScSPHjq8/3wkQg45tw709ZFO3sDBkv2sHsWcwpUvvqHGh4uz7yp4FpKa4gC82Gt9d2W/fWfX/4tmcp7iRagZCvXYrJJxsVeQt6j4QxNnwsMSjm1ZKhMCkRrPQG5/f2nQGAA3YSaZ6pLiKaBwUHffdpSzIfNZV6fwn6+TMgq42yK7RmuINLOQnNftN+nGzzTKQ/LSJoL96le9xw4HU52dX/Ol4/aw0pp9y36VQrhOQCbmZzNPbYH/DOGjg7XvygvEW2Lq4xlAQ/RVNtgwXaa1zUIdbk8V7M4P5xm0w4MrfnhXfhvCTqbabfDk56N3ZKqFHAP8tMPkq/F1vi+tJoj78syAogRLej2vBF3nzHq/c4e58n0FRgct7Uf1OX5n697rMU0v8did/Ly8ZKkH0ZUiqDx96NEBxiERxMPVcv/C/LDYeipymf10Wl3F54wTerjBu/yZzh/GQ5xkt4Tlu62a/reQJNj7b3w8bm+02pL8meknTJag5/UPgc08ysc7Hley3VF5/Crz0Afqjf38u/zpfGh5X+ugVgCg04iwnaTmPzxOYTvGDUtXAgpvbZENRVH+msCAjwwKiv6ccsAqkBhpBcH24RQciSxb3srqQr8RFXugolXIVfM+G1RGjIfOMhbPPeZKIN551MsIykLYpKRQ67GHo+05JOclsZFrjPZY+b6bMJOjIa2RnfmIupZpjywvI87rxGEe2II1CTVuyx6n0K14TQ4EPvATraF6E8MjAAYQUTbgr1MPEnO6c0Pbj26SXO22D7tIlIfamKcbQwEPatv27WRNiCBODJEJRUIVHejan4ATSRmy985vQzr1cmbovZdNEpsF9vvFbxGS/Nu5bsqNlHOeVREqSTUwMXoAq3GcOMdLad8+AUfpQoZA4SlxLWhOEwWJr72l6ao48ARwnV8UrJw+G8IDAq32ggQBw33opBUcUSOIRkXHmBGcrSojt8uyuOjMHAujdi5vrxJh+n9uoYY/MysK/vNesCtsAMWCKTKCDQ5AClbbiNp+qpiSai9AxTY2ZOpMMur67M2ZlozRlWl11wIwzQ3PFnPEmvnEMIUmR6WattjqUQj1n7n9/ns6/eTBc/hn8MqxMV/fIMcLZpV/5ICUYteuUaIaWwv46X/aGP9Gxb6u0Lh03ap+bb4R6qF9gBrZ1uoi7pWDOxDDsJHatZj8sqvKnwsUYVzHjpIVte+YxA0bLBad5XcKnKLWpeJSmrr5kjK8M2WrsM+VqK5itJ74wmGEphxfd7crxj5P5RyWa8Xq8+23Ya5NTVy37XmylHyKbhGDIBH5Cy7rrlzuQS1201FhxXKrJ+PAXyxYvRrHP4Pmw0dYuRuTcLWCPwYIXgb8PzsG150BWHpG7t8CNc4D/CP3zr38tO/7m36vRk+CpGSj/Yh6Jk13Po4uH3gIpezUtJE7WDhclz7tcw1s9O61l3YXf1LRZUo9DpgA1WwubWNY4/PRol5EiS/EalWUyqFhWIvSZJeIfrun0tg7Ms1d/F5cbdAV5ltEhPoJz1YSa2vfC13aQVJmt17K2HvXDH+quYjDJpeZYP2gQ0RRmpdbOzp4tzXKRFy5hMOjP6J4kW9GUVXlkbHmnkM5rNmPsepqQyJi1btDGpTs35Imzw/86alxw2no8XKxr0+gMbN2uzDbqENfSEuEcZEZ3bdv8//v3xt2NHUscuxb2MaJuK1mdsgex7m67ilJPSArX810ZVznB+eFp1BgSFFiBSDljHaDvUyBt1CfErOnc2e/PA/clTAm73f6g/4lNwxcKjPupr2TljdFobH38wojKg/B6p7m9/j+NYKfoj0bHstreWLsx+bOnXJ8FEBeJOlAci8fmKPW+x4p+cKfMWCqKFjLVLxbKRgt4OOrAKmsesMqvyU/IE8IgxWklyRVxJVySUqOCl3OEiZm5WXatFou65+XaiVkuwx5A/OQ4Q8qokPWYGDH4JJa6Un5HejrcWMProPN9bADxN7wP6LZZY+RZawrCsaUqLBHjCuO0ylmGpitKmkaxVUwWNHfIdeWAWued3PQW68wGktdLE+lseyMX1TJ638qr3nyjmSZY97ywg9TgBWKkO6cuFKOhAukQY5jWiqdu27TY7Uaa4kwwa8Ek5A8OZTkvo2LIcM1rIdxMh4ts6QXfN+Tibm391hPV8YcIJL2BQ78vM63477cDX44TkhVnlKu40kGEXvQVOQXtjKxRQhq0+dJ9kF5mpwHmX9xfm1k/bApAgPnYBHCDu9p+3WwmoWYP1E/N5LxQFHpVJojIVr2SW+H8roeEhLiAZ+O4z07HhdA43yYy48MLcJqaLJ/BhKBsytrYTunk8DnfkU+qruGSxJO5LQa2MS9NeOxoNj8hjimGuE8KrxuCYmpDRsIEPJJCcDF/eu2KIUNlIQ/SY0Wtt9f9pPHAT4DGguBhMlCzRYnM50d8dpwVUICK+FwtOcSmNDrNVtBzLOWr3ovDHlOnDyy5AgnIE8wxt5HBywEaRBFq8EDtEAzTnpJ0ACdR1/e4GIJcXoV/ueGd+Zzh4r4MesBOhu1BRESBrkpIY54Qs8AmKYMJWI41eKSwbwxubDQOPOjV3TOfBhG5PBnbrjtOM0naZgmCqi8jrT0eok9dsz5s9Cs3a1/QmTzzuh7fSmty9gkdC0WBVUahtU+7+1Tvwifk15znsfH5KDWP2Qc+cQWjkCL+aCitDkyZ8sDDMDfgRtVeEjzxhuUhC8cwQEGjIesEo06WBJYmCzKXuOYg1+HEzaFppLRnut1pAF7wcGNHKB+uUOqQfKmCxnryEjHqUhYLfq0EUlLEQZvQgi3g0RC9iXDFS9QHSS0PVhw35pqoLxi8LEzHqqaurjW8k3kXqkwLafE25+R2lu05Ub0fgXZVhsEuIz/rwVaKllIVZMXNcSH/3LiuzyYOE5hJpwA17Y8tnWq0n+7LZVnugLtza+srckp2nihWyLzzsATSHlL5CkAN1Vw/j0o+lXzwk7JW4ZGhFCvGTljaTI8VQZsQiLE3GeIAoihgJO7J9Lad2wQgp/2467q3UwsEDJmZuSTWvJb4ceSYQdOgllDo+uFCio6Ch7xQqi8eCItCzF31+ywuvkESAxMrn0Jzxlf5dKJZidFQ6InufEtV4SS10ZixQlKqIxOj3eR4xRyGoUiAW0JgOgGZCU89IbFFkg5uLrunk2fA/WoPIc2f90ATj+cyAAeFJRNoKYJZqVmq+eAxCVUQIcCnBjyyJr2MGScB8FD27Qx8n3Y6ss4bLTlPbITMKE3HVmTxuceCXdvSpeh4nx7M1+sMRZpynU5FbBFaNhjZQDTBfi2CXkACa8ROEAhoFMinqMJ9ASPwKHrAFsK+253C9qTB8oZ4aNOuXCPRIwmN7bUNVrnUn6qWBZmOVLIE13b0Rme0L6MBSIKu1AmQgFDGEBoJ6nAPYQKEbHpCDrcKAcFzHg8dLAwhSbIc8KlnyaVtHctLOUzLCZJ3tNwEGTr9tDiHDHuPCp8Abvcqt35/XhXDKfRSCv96Z/rEo+SRdD1FV7a+Whm3p91zwA/P5N5cVgZd/Onhwu6tYmLPArH7DQt4xpRFYjkkanF3BYbjncePLGhV7ZMMVdTI3pjoemdV32dpSvGw3sjprnI1/kSeIibRxycK52vgmdXpHZF9ImtPHEPHr0QRcYLTeWYLUYQeGouOJxZ99lqXTWs38wbUtkbR7DPc/YY8FI0gBUyhmm+82KwZWMBsDEcRMedS2rxR4y3hMjFFrfCU6ZYl67qqlr49CtkUPLWJR2WpL1h+fV/zeiSJnKpaZnWM4UcS2Wvt78nbw97W+XuaGeZ7SIhM6606tT40b2VMd9NeM3Eq1c4Wpgp4LgmthekUQdRoeVUeHFcVzIkgm2Gl2E98XaZmMJQuQh25ldF7RlCn1nnMaFHwkuOqSTvnXexBNGr80rAmeGp0ok2H9HSuJZFYiv9Zv562m/3NLW8SLA93dy6HvjHn3+AhLeiwaRsO4kGnQwuqm7leeZHvNALTvO+cbXKR2UOyj2aH0z8dTxaoxae+1xxOY2Uw0I9/wyKn4S/Gvz14Ml0Fd1OH6xiE9iFsDBZomIn+wxYmR/h5+QR3PMK2pAAd6Hn5s0P5WL5sP0PBPlfh99wT05jSM9ETrwxogM5qaMyjPuOtw/Izup+2gLPg+C157xdoN7+vzYkjQACalZ7p+eND77xdu+5jG5wA4RrFtMY6vLHys0x095xzTI+3bw5vGzfDG/XF7KZ2nRzjzUvB51ueLBW7oiAd0L9AIka7CCJWeiDuEEsr7PAPf7ccAz0uESsfMbM47/eSUu/rE1wacatfFa39Zs2Xw2ixyWEX3NuE7BzdXl7OhQhX15ndhXVuVL4GP8NRwGdxTlzU1CX5TGJ5F1ZCEBXnCBP19OP4YocioVSn+cja1oXa1lrb7WZMmDMBQdk35bSehdZT9I3hIBfHgMeUMQ5Q4Em8/llzbxu61ySf6bba1VaXgD2OZwkBF7jHWnEr5CmuSuS1qeZWPIcz8hCpOqmnabQQPeESrRSKzHfiGCOWBlOxWA9v9oPaMaFIqeI/nmSetcR4bZOWpNNfmkMB6fXBEI/GeY6HYa9cktaEhtZrkfL8nboP9djuV7zNV3Gnb4IlD7qra+e5kJeJqk27Jfh02EMaCHwxbT6alOATSoP6L02BFZRGwVCp62lkqkJqwiCpiHDdlU9qigVyklpXY4fPQk4lrQrg30adUpPjrDTC2k/qen+TbIogLf1SNlQNH0vHtVXP4JYbVuDVV3jCL2JIB87V6zd4Ky0PigGkjaTdO2jhnZmCKuy1T2Y+e+R8e3CR/3WQhcZoVvjksXG4M31D7if/Ne4ZYoZmoxhH5WM/AQIQGPGJRJIVPrGR2rl2zMJTHX/PmtkfuzD7XDoYzKFvpxHcwkgYmqjLxtSQASuA1OFUKUUcL77sh9WABWxdPkrmemtKIj/J96GcDmgHnkxvSyi0/UFdOBjaBJyC1+DizLp0L/wLHe+ft15dSHzyOUJU2UUfG0YbdnB4fppE4U7LG8Nh7a2xGUKEA2/nYi/PmGAtqreiRzHuPLIhYe/RPXnmSa/nPdSsrmz+wT3GHbGHP7Ws7iFCuYp6p8cluxoUCXlQ6Oou8aMGvby+viwkBmgM3m/d2jzY4v/ZGSGBGxVR/N/N+oOjYK8POQ0bxfEuKrWBGoOa2g++70EZuxH22ZgVdsCosSu4mpkA9hM5WqkKeBf0giRPaejahoKHCRks5lmWvIJGIn2tSWJqVRw7NhWm/CgB1zar/ek3SqkXKU0SHXqeqy2tJFXyJt4tM/oIliBOyzE1SutJeMlW26ScpFwHlck2BLls3GvTyIxSczKKFOWhq88gC1Yrq43IcLNtuOjCo4krE2njpKM08V2XNyDbuicEq+jICCgtJcV7zRa5Vx1m4xBK0hzjTIqqJXm+rugSORLaPCWYefthamGghNBxwGy5lIssannDk/HK1LRlXATy5mxRzTqYINUIEJTW3FIRMqY5SSyF2iDh9YyFgLCZahhI1p0olrTXetTHqGQz2VHiT9kkfqQarnLLJM9vnF1X0zFy/qSHI1SykosUf3GLU629qoPNwMcwDhxyXe/y4f3/37voZ/3mggWaTSV5rVaphk3NL0Pj5HSliW0lTdZ3RBQSTjGXAvfWv+VwvPL9B4KkPJSZmn775kCMb3x/IVjfQdbLuMqxnQ2+HV0ZrW56MkJ7Cyg0RXFWgsNcYvQ/uvtOxevZ4+bUoh/+6PLmdakQrv8oHxYLDtv07XHFHZPfDoRxXZY1BBc3Wf2MXFSMTYwwBnF+wN7oU/1fl37+hx3bN344+/GPz1CSXklZQkjBYGKg322E8Aa8ye/yqBv+ja5VPDzUj0yEnq/hqbOFbnFsluSqchHbp2QLRzppzin+sUEnkjAVNWhLmV/Gco9TooQWEOGl3SRHRxvUB61qLRGD8KygaRaPlxJxd5WogZ47LrxBiZwj0FbrGtAITZTotXXSWYEzBHdlNiLJV0bxYvRgB7/RxEm7W/v9ecvS8wrVxVbnOobmnfYGsi2W+wrsnjnTkXXnJrFPhUWxF7VW52uKSU9tpPal7sw3gzV2hMvIOtWZJ0t3YOLxipau2E4OWrXCKkK+LrSSwokkY06HOk+nAjsYUe6i2PqgBYYzXF26JHElPB8A4RpJJySOBM887dHeSHI2Qal5aWfqrnS8okskUdFc5io5HT4QrCFoj3J4T4YnlnAvtl6ziQb8n3qj3l/wpS8YxM8SIgGFBAwMEkzwuODFSCXluw3vWOd75u5oh79W5ySVkKg14kK8wDk8n6Ae1hm8iFIaUY2YfimZMXg5wJKqBhi2EQG3P1E3ghaqLTi2D1DwUGMguh05R3ZzSt+yby24wIFl33Y1RDPfxIFbq7lCXYqiO1FybSQ75j80F8zF0NTc0ZLF/VzqYVBHOE3xmlZpCbqep9ahCOe5lDYbPKLg1pGh5A2I4WAweFDXeVUmrNoymQUBJsCMLToqO5F58qZzFyTR6yq1gvLaHd2dQO3qouMEE2SNS6evjoa/Z0XHw6hwcy5TlHMAFYguJXe85kyjg/S4fohDMRMAn9zup+2XNDF44N2DpTEs/Vk5c21lX+ybLW5BnlCMC+eY9sJcVJLjcKsCtybLQFGR3BlU1W3PLJXhqTvASTJQThczSucskZ8elXmI5DjWnTAvR5DaIe2yMw1dmMTRXw9YNEOF0Xmjw4UgmEd86gM3A8dxlOjfqqSRQNqthg/qdVzHDOVDRiXDu/bIHQ6yCmRQ3D5Y6pSYnOmkT+9297ovPDvTRx4XhDwwpjfYJh9p++WQ3lC+/9BMnhQe1ze6j3ZeN1UIFrDEQ2k6PiV3HGdQVfqrJoFU1VQTIQN5oUTnZFqzyuFUGp0M+X4PnzalCu8O69Cv194n3Dy+gGu89FqNKzUsJXja5pAPtTMTxyGzKc82N1GcLp3bqed7Bua1BnVzHQXXE06gghbXb+MzyVd8QvBq1vWSxFNGsoxGJxfr83UgzzTlJ5MTx2vX3DB4vOqhRgmS8/za6408Bmy2wj2Zb0iPi8KHuRW/U8Y4F3DTvy5BRW/pKh33lKPNA1pDpAWNwKfGMeVwbm/WtrWCFMvRUqby5wklXaJnPqDvkOP78xvvbMnzHppf23npiqAGtr+evu/hHXbMjR7xNRVISGpaNKEcro7SQDH1Dz+u1Pal2vSqX+AxnrGtm1WK/RbyJ3bkVMohCfqvHlzxbe+CExuF70L4OEt2Tloi+TUeFqn22BBZPSm8Sqa4PGVZ1+BXnwknzps0oq5KyYgenFkFqekf4sivjEpkZq2KTuGBflyx+wIHcPNKb6U/pwN18jk5T07GF6XjVgrviiTxl3ncfONYC1t/+9gqxpc9/Cq/HPMP/3+JGl6iipZUOCL76HS/RQEcwSByTBR0gsaRKM40MwtRVV6a+mP9Jl/iBauxfFUZ9p/2FpUEgw8Ie8Ctz0g6BkfI2jhQ2rJeaLSxJibTE/Nod+/fXx17rlt6IDRPkSxqYJw3F/7DUyMLume5Kb5rSpJt9fmfF4cMg4+3zVIKxjwOng2US4LxJlv9iewdZXDE/DhTduH896xCMgoBaiOqqEwk+30pp13NPzAuvvxmJ348J4weqDfFZL/HdPZ1H/iHO+YwXzyuMzgm6s6hS0OSwPps4f6MeTdJL5E+Xhp9tST03hnvYIyHoZ0OO3GVmGUckT3ux/XeARzgXFQAnycL81mJqRELTPA9OnkVeODcdG+epXsfSLZu4PwZ6Q6dCQ4FZ2ynK64H4s/DmubJ1DwaSsHi2tL8kMXtkGSM1klB7OwIzvjqrNHoCo7mjByONdJR4qXqz2ZU1n9vj6tsUz+d5/b9Ux99zSzmlO4CwdfHvzkh5J13Ipx7Kr+qT6CiXqlXVCEDvdtFceXp7nKg0oFQ6VA8wovH6EP+RufGIFf5+Km6yQJt3PJXRPaoLKBaG3hq/tHaDfNmdyRjdqNohArG+1RQXEVo7sOvaBcbQx2mS58GT20HycuIfEYmZ6NW1CAsYaaWSdUAr8GBQ8LHepI0lxpdI/IeZWEsPfNdD92U/tAaLzFv3AhfCbtBvd8fezKqaJbjGYfjFShhU6k6F7ttpJUvscDQqffeHnFo2hNCXiZlWFn+DvmbZ5Y5xeb7OOxOQvzwItBJIdshTe5I58UqhUBji2ithOzySL6e7oeh+4BIv01ChcEBAHIj2G40ZnDbfE6uF229XPIPNGiUnRLafJtoUY5Zs5jmN6mnwVQ9+tN8uPVxm049f8WYIemf2/rul9loQ+qzp94KK+UAPEVoxoW5ipISyFAMU09DTHMhgErFfKNknkSTVIpt2KXoQbgfDawrW2hJsZZRSBc2tOLc1tgSzdt5gSeXU9Uv0EgmqvwEKzzm4GLFAJkOeRRVQaIz9mhAZc+mi2Jt0YQaoS+mO9gXbtvFZcRypCZawRMwT2R4YvBuDEGT37y/39fEoWC6K8CLCeL9vaPzxUvnT1ENAmPgVkiI8FzTLof5xH/47cksSAEXHnxVsJiPty/sVcsFZv9pBiT/12aFjUmCvrBXw3tt8uUTVL21/aY+tcHArfSPn4TvzLw4sASE/womLyS392afsIuX7hb8D4K78SgX7vlAPSS/J3zhH3mbfKrNxOQoGmR6ZJTd9brxR97o9Jj3yhED7eIrsijUNuBz7vstRws8Mi+k+5lyAtPj+BXTms2c5uEw4Xatp5VCwvhG+QB6GLqcF8UcZhy7YeKbWY7KDoQHDiIkQx1k04ZO4RySnN8UwiaEFJw0PikewbdrHzgdBBAaclHwWmuz3xKOpxG3Dj4p0gVH52OzfMYBuAirfV75MReBQWlsx+ilFBvaHLQM4YtgKDI9SEF6UygekErDVazvLUamUoLBy6nr2zNtkDQ6yMWYoDdFlQLrBYgm6iDPzmotnRVuEExora0njtGM1WSv4WmHeOnvGizt4GgPRuJH6YccInSEk+ndw9w/yx0e5Lp3gS3Y45Dun3IgzaBNMBpOlmKQpOzBdSg7PeqRkhhJ1W4mJx6bPFCU7UBibvaY4Rz02ljBSAbpOEcUHUJ9lSYB2zrMbcB9c5vrzaEKqrjytGs5B5SbQOUAIEz92NRtzs+SJLIEQzmhBkdacSZUHFifc4pYqdis3qtqYSQCzp+VZ1Oaph5Mag+VlJnisWTgLdEdS0fU6NhGCgFapDPZgUBk5pfXu5dl5ybMFwYK79KI27nYo27LT3lD6DydCb9q5dCn/1PE7Y9IiGjsQDwhb7e3v/n+gCW7+8gKVEzE8mccdEJjhwA2g9fYOoDeXPlHIEDPf/zr+fQ/1aeVf1Od+wcA/joYDQDgbzPpX86+tDTAjywSAB8DQPBXj89sFze6+FmT6+h1T2iCYf4IxQkYwqIRBlct7BgLypfcIIaVrAhsghppPRUUxjRlizpDhEcgBoJaxWAJy2QGvFQNj6Y6vNYNpQmHeWAYhlrlYORG8Gw4ew4QgENcWoVHcxGglGDEgDP5rGIzm/iIuaznPY+m7/UeGwxnJO9RyGcA/Mt/+vdRQKOT7vGYftxJmEbMuCKvASW03P+lXeAhf8qtvWzkqXByg6GULEb4K7A9ZsoAKWghwgZjjk6s4idkE21LghkXYRLUggmWFQAejYoUP+oePuUqHvLPLLzKNlwxSHLDzc4uck0vU7nJDX+BwN8V0Qzd2mUgXpXLNouJUg0vKnq2ZtIXiKA6GCnbg2OVovBoLyBELr/yac2pTLFSRoogofPUkKrTBJSmTi24QOpghKQejBo0gPf0GlUbuJzcAXky+J8CBJcVyXFECYp3NZHKdZopNaAFrJhWaNTawLCCdvDytwMR2/Lrg+10iWuUYNWnvVO7JmbNxfB/CETo0OWNd6mkg8aeKUE6W2g0rR7r0Uhnv+8ktkrlE0W3mkV6uRaPgojsreUBB8RkvyToVFZpb4St2qKgbksdBVtqkM4bwtrgJ9J2QDM7JUxLu/WS4zK2S9nco1q4dZ9xmELCasZ58h6tUqTnFbXqN1TYw6/Qo9qjkIA/enEJHr5PiCcgCOYuj4Lfe7F6dgqpyUy6JKWvYfsVPRoICQQcC1QhaOjxzQicoVsPyw5VTdlIq19vqqt+jPszdChbmRhW9oviaLL3iGtAI6RLr5NFBH6lrjWHcw9TEnPnWPUOErgw8SNUBGh3Lkm8I2qbJtutQkUTQ/cHhmZnnXeB+yHAGu3KJLMbohYCwAa0qtUYjscK/UMBuebXvtHmWxKI/CtGBu3shpUdEuLWZNkV5Z9Rdb+IP5UI8um/02fAYFI4kSNI+ZCHeBqPV5gf7zRkrRl28foXEaKWRDWT/uHP4oQs/O2wWtRHKnbkrzpoRRRVTHEllFRKaWXMFWueOCeZb4GFFlnMZ+4zyZTyllom3nIJEiVJJlnt5MqmCG+LFO5Hss5Ji2Tik1Utx2576KSTYC8N1DROOU0HiTYbJYP8L0ISo00al1homaWWmxeqNOr9br4xaHWywJuy+L//eAUBUbnXNNpAGwap6rJClmwr5ciVJ1+BQqsUKVaiVBkj9Fa46nvXXPdDPoxcaaM9zQPN6f2Jdh6vSssT8vhp7ZzOrqZO2GNDD+cFQsl0UEx7cpJPmL2pXYnmaqZvcKAwG2hS9NJoWm4z2z6wtqO3fRqHeDmJ++py0o5Ee8K5gzzHkJyKRHJF2RGZIxKJmyvKrkiOyOhpJwp4Eh56VicrpkAL57pKpsgUm5JJ6fq8lofih02iPHh5Kch796qOCN3Y0+0Z7O9OC+SPi09aQp5USyRprT2NAxeDeHQlaurva1yZAi0ur6JN62vvjB0rqjLburqmsL5ouCcj1tXn5mtdmhFIwg6hNMMT/eyL26HU+eqGTg8=) format('woff2');font-weight:100 900}+.d{font-family:Fraunces,Georgia,serif;font-weight:350;font-size:48px;fill:#241B2C;letter-spacing:-0.5px}+.s{font-family:Geist,'Segoe UI',Helvetica,Arial,sans-serif}+.m{font-family:'Geist Mono',Consolas,monospace}+.lede{font-size:19px;fill:#5E5360}+.lab{font-size:12.5px;fill:#8C7F88;letter-spacing:2.2px;font-weight:500}+.kick{font-size:12.5px;fill:#A16215;letter-spacing:2.6px;font-weight:600}+.rt{font-size:20px;fill:#241B2C;font-weight:600}+.rs{font-size:15.5px;fill:#5E5360}+.gh{font-size:13.5px;fill:#241B2C;letter-spacing:1.8px;font-weight:600}+.gi{font-size:17px;fill:#5E5360}+.ix{font-size:15px;fill:#241B2C;letter-spacing:3px;font-weight:600}+.il{font-size:15px;fill:#5E5360}+.it{font-size:11.5px;fill:#8C7F88;letter-spacing:1.4px}+.wh{font-size:17.5px;fill:#241B2C}+.sep{fill:#F59520}+.ft{font-size:13px;fill:#8C7F88;letter-spacing:.6px}+</style></defs><rect width="1600" height="900" fill="#FBF6F1"/><text x="72" y="84" class="m kick">REPOWISE</text><text x="72" y="142" class="d">Your codebase, indexed once.</text><text x="72" y="182" class="s lede">Code, history, tests and decisions in one local index. You and your coding agent ask it questions.</text><text x="72" y="252" class="m lab">WHAT IT READS</text><line x1="72" y1="268" x2="472" y2="268" stroke="rgba(88,67,108,0.16)"/><text x="72" y="314" class="s rt">Code</text><text x="72" y="340" class="s rs">Every file, symbol and call, 26 languages</text><line x1="72" y1="366" x2="472" y2="366" stroke="rgba(88,67,108,0.16)"/><text x="72" y="402" class="s rt">Git history</text><text x="72" y="428" class="s rs">Churn, ownership, co-change, bug fixes</text><line x1="72" y1="454" x2="472" y2="454" stroke="rgba(88,67,108,0.16)"/><text x="72" y="490" class="s rt">Tests and coverage</text><text x="72" y="516" class="s rs">Which tests reach which code</text><line x1="72" y1="542" x2="472" y2="542" stroke="rgba(88,67,108,0.16)"/><text x="72" y="578" class="s rt">Docs and decisions</text><text x="72" y="604" class="s rs">ADRs, comments, PRs, agent sessions</text><line x1="72" y1="630" x2="472" y2="630" stroke="rgba(88,67,108,0.16)"/><text x="72" y="666" class="s rt">Other repositories</text><text x="72" y="692" class="s rs">The API contracts between services</text><path d="M 492 318 C 590 318, 560 432.8, 640 432.8" fill="none" stroke="rgba(88,67,108,0.32)" stroke-width="1.4"/><circle cx="492" cy="318" r="3" fill="rgba(88,67,108,0.32)"/><path d="M 492 406 C 590 406, 560 459.2, 640 459.2" fill="none" stroke="rgba(88,67,108,0.32)" stroke-width="1.4"/><circle cx="492" cy="406" r="3" fill="rgba(88,67,108,0.32)"/><path d="M 492 494 C 590 494, 560 485.6, 640 485.6" fill="none" stroke="rgba(88,67,108,0.32)" stroke-width="1.4"/><circle cx="492" cy="494" r="3" fill="rgba(88,67,108,0.32)"/><path d="M 492 582 C 590 582, 560 512.0, 640 512.0" fill="none" stroke="rgba(88,67,108,0.32)" stroke-width="1.4"/><circle cx="492" cy="582" r="3" fill="rgba(88,67,108,0.32)"/><path d="M 492 670 C 590 670, 560 538.4, 640 538.4" fill="none" stroke="rgba(88,67,108,0.32)" stroke-width="1.4"/><circle cx="492" cy="670" r="3" fill="rgba(88,67,108,0.32)"/><circle cx="800" cy="482" r="148" fill="#FFFFFF" stroke="rgba(88,67,108,0.16)" stroke-width="1.5"/><circle cx="800" cy="482" r="159" fill="none" stroke="#F59520" stroke-width="1.8" stroke-dasharray="1 8" stroke-linecap="round"/><image x="741.0" y="356" width="118" height="108" href="data:image/png;base64,iVBORw0KGgoAAAANSUhEUgAAAPAAAADbCAYAAABTPhxnAACHx0lEQVR42ux9eXxdVbX/d629z03T0pYWWqBlLCBSZtIkZTIFmgEEUfQ2SVvUn2IVh+f4UJ8+L1d9T33OIqh1RCFJe2UukKRluMjQJo0MQpSZMlPoPCX37L3W749zbnKbDkkHtYW73wfhNU1y7j577TV91/cLFFdxFVdxFVdxFVdxFVdxFVdxFVdxFVdxFVdxFVdxFVdxFVdxFVdxFVdxFVdxFVdxFVdxFVdxFVdxFVdxFVdxFVdxFVdxFVdxFVdxFVdxFVdxFVdxFVdxFVdxFVdxFVdxFVdxFVdxFVdxFdc/eVEymTQAqLgVxbUnrVQqxalUios7UVzFtfeZb9FwB/O8AFB92uzxdeWzPlR94uwR8Z8VPXFx7RHGW1Mx8+zaKTOnF57Xf9fa426TZDLJAMDefbR02PA/UIn7KQCN/7xoxMX1b1lVVVUWSEtNWUO1YW4FydV1R9WVANB/57ncY8MBVRqZc71ijPloTXnjVzKZjM8bd3EV17/YqZhsNutqKmacRIb/wMQJAKEdbc2/+9n2OINYvnw5xYFJDwAWkZCZv1M7peFTsRGb4pEqrn+l8WYyGV8zdfZRBHM9E00Q8QpQb1lXWU/RgLftgjdE/wapirIxP68pb/hMgREXw+ni+ucaLyLjPWdqciKJv8GwPdKrD4mYQBSmkZaiAQ9Y48ePVwAgwiuqCiVlQCEqymx+VlNef1kmk/FVVVVFT1xc/1zPi4w/p7LhACv2RmPsCc47DyWO/Iv2or+IpUUDHrAEeF1UQCACiACFqqphe1VteeOl2WzWRYWF4iquf07YXH3i7PEJ5RsDY8udc54IJnIuBII+BwAppIpV6MI1OTM58sCwK0Ql13/DEQGiogJmnps34mI4XVz/DOOdXpY8lIf5BcaY08IC4+3L8IheAIDuZHfRgAtXGmkFgCAhLxDwOhMXRCjEGoXTIKZf1ZY3fCKTyfhUKlXsExfX7jPeyvpjjQlajbHlod/SeCPD0ZVAQdG1aMCbr9vua1qt0JeIGKok/SgPIo3CaWJjf1Fd3vD1dDotADRVRMkU1y4WrGqm1J9ulO8wzMc65zxhc+MlAot4eKGnC2s2RQMuiE5iQ1QArxLRVqBaUU4s4iUwwbdqy2deWXdUXUkaaSm2mYoLO4G5zyDja6Y0vp8M38FsDnPeb83zKgDyKjli/0Jhylc04IJVkFc8up19JwDkxXlr7ad1vzG3nF+ePDCuUBeLW8WFIUIjNZPJ+OqKhk8R0zwCj/LeydbCZkCViUHAa6W9pcsKU76iAW/VF3OXqACkW3lOzYfVJvShtxzUeLZ31k1pOKVY3CquoYTMQFqSk5OJ2vLGH1u2PwfBiHolItbofOkAdKAyM6D4x82PXLP6391C2mMNOBOHJdbhb+L9SiYzcKOU2TIRkUKFABO60Bsyk8G8qHpKY0Mmk/HFvLi4thcyTz+lYcLafewCa4PPiTiBqsY1FjFsmYlpa7M2Cl1aiNsvGvAWKy0A6PaHmpYB+ggzQxWiUGE2gOpfnbgvQHWDYcuq8EQwzjsBYaw13Fxb0fCdqqoqW8yLi6swZFYookpz45k24DutsdWhCz1AnI/sAptgL/4eL9LMbPK5L4jAXpwqaXZP+US8B5f0OUZSPkyg/rktVShwQK70mCvhUacqLwbWGigcEbGqqKioNSVfGbZx4s21lfWHZzIZH3niojd+O7eIgLQQSGvKG75olNqJ+Z39Pd4oJTMm4NDl/qA95gJAlzNFziPOf0lFX5WRia4oUswUoZSDxjtEWYn2lgAiUa+G7cHDep44v7Wr+T4HPdeL/4u1gY3yFgIACl3OG2vPh5r7assbZkS41aI3fnsWqlKcyWT8+afMPKy2ouHPgQ1+QIRSL1GxSgFPxEzE5J3/7/aOlv+3MhieI9C06OwpqZIwMxTovPPOP62IbUeLBrzNPDi63SgRPuCde81E1QOFkmc2gOAsIMWLljQ/Zd3aOufDX1sOGARSVSGCcS70TDSR2cyrrWj8U82U+kPi3JiKufHbhZIpurhryxtm+ITeZ03i/aF3XqN8l1XhLRsDxWoRN6uts+nbAGh/XjeFiI4T8aC+XiaBFHfHHp2KITQG7we33pd5gwj3MBuokiopq3oo6Nzk5G6bTCbNgq4FG9s6mud4H36WwDlj+vNirz4OqYPZZPiB6vKGDwHQNNISVSKLleq3ZrgctYdqptQfUlvecA2zmcfgg0MX9oEzVOEDGxgv8ph6d25757ym5ORkAoAScI4x1qrCAwAzGedzG5DgVuDf3//Nrz06nByfHM/d3d161MHHJ4j4/aLRbaiqYKJxvQl75y0LW5ZVVVXZDy/7MH77ylWLj5ow+UEAp1sb7O/VC8Whtxfxhnhfw+a9R044bsrRE45/6uZX5r+cf+Hd3d3Fk/8WCJdTmEZXd18tyWTSHDrypMtAfI21wVlevAhEiciowjMTWxOw8+56yoUz2h7KPFNVlbK3L73a1dXVlcimYT9gooOiEJrUGkuimm1b3PRTAJRFVoseeIhh9LDekju8hC8aNgyoqkKMsRaQSwDotGnTJO9RWzvn3am+513Ou+utCZiIqcAbixcnxtgLlPje2srG719Q1rh/vuX0Fs6PCf1FPHorh8tppKW2bNa0tcuCOw3bq5hoYlRlBhOIVOGttQbAeufdl9s6mpKtD2feSCaTZvz4bgWgfuW+ZzCbU714jcJnzU/E3VxA74SiAQ8ljE6l+OZHrllNoBvzYXRczgdUP3DulJmT0um0ppDiDDI+iaRp67rx1baOpg945y8DsCKw1mgf8IM4dM4DKLUcfCm0tKS2vOETZWVzgnx+/BYy5Pxn0XwuWHBR0Vvp88WtoaNryxt/CyN3GmOqQue8qI8KVVEoTIENjHduMbw7p62j6f/ys0WZTMZPnhyFxQx81LCJq88AM7PzuTUgva0Qp1A04B15U0y/cT7sYYqeWUS9tYkxlmUOAM3DLzPI+Lynaets+qV6OsOLu9lywAU9Y1aohlGRa5IxwS/253X3nlfR+N78YXgLGDLlP0symTTTz0geWl0x+4gLyi4Yno849uK22maGe2FlwwG1FQ3ftorF1gYfAcD5+V0FoFCJvC71Ou++t8a4c1qXzu/sf7+kKaQ4nU5rXWXDiQC/3/lQicCqEMMWAG5rWzLv+aj4mZY9isJ172gFpKW2vOFGaxPvDV3oicBEBFVdaQyV3f5g0wtAigo3Nz8eBgC15fUfBpn/tsZOct5BVT0RopAcgOWAFQIVXeThv7+wY157wWHhOJzXvcl4q6qq7LCeCZ+C0myFHAmFIcIrIFrQI+6H2c7Ma9jDDuR2TwFS3J3spvw7Pbf8kv0shR8F6FPWBId6cRDJv1cgSrXYEBii7m7x8vX2pfMeyP+sQkqcVCrF6XRaasobfhnYxMfzZyxuZSq8ntO6tDlbeKaKBryjc5rl9edasm0KIQCcryKGrvfH7Z3zvrC1zY1flALQupOT4ySwX2LCp4wJRjgfos8jqyoRYE3AXjxE9DYl+VmBIcfPMVn38ANPAFBVlRwxbJNpCmzJhSIecS8dRAzDBs6FTwv4A+0d1z6yhxsxJZNJnpyZrHmDe/fUD050kvsglD5mbXBENN4nmxkuMxnDFs67Z6H6nVGd7vf5FCuDzS/jvDFXT5l9vGG/GKDh0cgqxFprnAv/MrrTnx1/H/aki3yvyYMUSpS6gmrueOLugIN3xYPWDJASYZOoO6O9Y/4jA2/WrXrjshmnwgSXE2SGMQE5H6oqJB8yEcFYE8CLg6reS6S/tm7UDQu65m4c4An+3V6Z8rQu+RTi2WfHcFfX3LCmov4XCVv6idD1hKow+V6mQhWAD0wiCF34aO9wdwawvAeYhmnTIOk0gP4JG/13G27hhVxXMWuyqnwMhEZrggP6DTcqMhW+O+fdKgV+nvD6swVdzW9uzetu6X3rmwJb0hh7X5PHRIuT97ctbbphT/O+e5UB9/EUTWm80Bi+WdUrQKyqYm3Azod39pa+Wjdt2jRJpyOPO9ihqK5oOMeA/5OI6ogZ3ruBhszGWIICXtzfSXEdW7r29geblhU+V6F32P3vR5HCFTSQumV7l0dtecNxRPywRt+/9cqzqrM2YUMXfrK9s/kX29MB6u7u/93RZ71C88DW3f15U6kUdXf3h8kAuLqiYZpRuhSk7zEmMSKKkLZluGEOij+pxQ/aH2z+R8HZ2ep+9Z2r8hnvMhzcqSoMgFRVo2jMLX3Tjzy9q2uu29O8L/a+SmSKgbTWlDfcFtjEeX03ZRxKO5e7rK2z5ZeD3ZR5FFbe6GqnNLwHxJ8lpnN4K4YMUoqnU+B8uFIJbRD940sb/F3d3Zlc/tmSO++VKe9Fly9fTuPHj9ehXArVJ84eYUb1DqfeYWMUuUOEeCIp9lfSDxgyp0l0yW3rHSsRQ0VeFdIfMPRZKC93Sq+DzBru1Z6Fj167YbD3kdzymXXHP3/f3vW9s+rTZo8nkYsg+hEmmmqMxdZy3ALDVQL9GXA/au2Yv3gwwy28MO655/lEyabeey3bcidOCMSACpNhFX9xa2fLjXui993rDDgfAp1/+syTxeEBQEvyriAqaGGFFVTetrTp2aHkdQNfcG1Fw7sB+hQRnRfnT3GxS0mjI6/53MqLh4o8oowWkPtz++LM00PwylQY7k6ePFljOqCtrrKyOcH+wzfszz1upCM6mJUmgfQogA4EdDyUjlbCPlAdwcwjiRgEgqpAVIYk+UEgRO05gRfxgK4joo0K3UTAS6r0CkFXKOEZAp6C4nkfYJVdvnpF69OtvdsLf/s/41Y99ha5LZDi2op/nKFKs0F0oTX2oDj6QWH3oPA9OB+GBNwIxtWti5uz/eekb6oN25NLyWazrqa8/ivWJr7TV7mOe8XOh23tHcecn0oB23tPRQPeQWnHdDottRUzv2uN/XKhF7bGGufdXbzf/udXVo4NtxNKbzM/BiLxKoLMAehia4JEPtcCNGrrx71ow4Yo8srroGgTgz+WhCPvzOfKhQc5Dnu3eoOfU9lwgPEyjoEDQTxZSY+F8gQiPRSgQwAdBtAIw7b/jSmgBR8tP7EV1Z8VXvxQDFiJmOIh9f5qwwADz4+tiwpEZBMBmxR4iUhfUMXLTPwEEf6hSss29dpXstGw+3bD8cK9qDqp/vBhw3ARwA2keW8bhcnRnkdh8oBIaA2A+ST869al13X2I7H6I6uhvPO6KY1VYGpVaAKb7S5yonTaws6mh/NnrmjAuytPQooeOPGZUh7mHjRsT3DeSTRKGFelfe5H7R0tX4xvWD/UkG4Ljzx1xqkq/EGAZgQmOEhV4SUfXitHNtTvDUQFKvIPUb1WWG5YtGTe3wc++7lTZh6RYDncAycSUA7Q4QpMAjCGgJI+I80bqPZHwYX/q1B47z0Bm0BYqaovAPQigOcBPYLZNIiK0DZ6/QoVQ4ZF/d9E6U6CHAzQwQAOip6FhkcFgP7fF9W/+hmNCi8NEQ8V7QHwqhKeBfC4Ch4kxbMEfrGt67pXC3//BWUXDM/xiHNA5hJSVFtrx8T7G3tb5Yg/nZQIxhgbXSLinoDiT2LcvP6oZ8sQfFACO2T82afMPCwR4F5mOtSLl/xwQ2ADkwtz3124tOWre2rovDcbcF8oXVveeBox3QVFQqFxjEZiiI2Iu6yts+WXO2rEW2sX1ZbNOoisvE8FlxDT1CiEjvKxqAieL6Qocd5DuHATgPsUciPAINJyBZ0A1Xcwm1Gcnx+HQlRjGy0wivjPVf0GBW0k1RUAPafQFyLvRy9A5Gl4Wc66z5rWh3//RmFOWVP+xP2BTUx1PhcCFGxev4qwwKraS6SVrUta8txjdEFZ436biEYRu/EkfJQhOlhBRyjpkQAOBzAK0BFMZjgTbXahxLcOiDgqn4lE7SvV5SA8r0A3KR4CY5QKNRpjJjNx/17mvW1fmMzGsIHzoYPiLjX4o4RuwaKuzJq+WsYOhrd9FeepybHw9jZr7FTn3WZRnPfh4kBGnVt2wUE9Q43iiga8s1XpiobLEybxvXwoHW+2ErF6cZcu7Jz3h50x4q0BB5LJpFn7XHA2SGaC8F5rEmOisNL3F73yh4/IGGP6qQxBUFWI+shM48Mf/5lC8aoSXobiaRA9zfDLRPE6gZ4LGW+aTXb9YEWlFFL8atmrZm7X3LB6yozjjbG3G7aHOB9uzjRGTEQE591/tHe2XFlWNifo6jrID5Yzvuf0j4x04boROROMI6eHE+mBRJgE4GgFTSLFASBMYOLIiAGoChQKJoNChtFo35yqkgwoSrFhS0SE0OdeI6IW8tzUHybvfOU/f2befcLMMX64Xm/Ynh32IbZUIuomXeO9O2vh0vmPbavtVDTg3dsnlNryxvnWBh9wLnQgWECFwAyCqMin2jpbfplvyQC0E7epUjI5Y7OeZG1l/eFQmqHAJYbs8cwUF72wWZV04G4TwKr6NICniOjvIvowLD9nesInRx6NlYOEa5vl1JMnT1akN2NG1AKIrEyvvPhoqyXfFtD7o1Ba42kuPKuq32jvbL5uwCGNz0OKUimgsH00WBhZVVVlS9eMH8OWjw4ZE0n5FEBPAnAiEU3cci+U4mqvRsPyhcVB30VMv9sk7oYILbbriLgqVNkssm76KQ0TOEEtAQdnhb5vtFABEmY2Im5mW0dLcz7M3hsgd9jL8b646KQPje4pyS20xk6JjJhslLAxmJi8yrfaO5pSAHQXX8wWldOqqg8NK9kYTifSj0NRZ21g+9sdSv1cSzEmly078T9r72j+7Pa8/uZ9150BVvRX4Wsr6/+bKfgmAIj37Q6bZt7ZedOKnUBgbQEc6Qf3b/3n1FQ03mzZvsdtoXCgooVhsgs3gXALgX/3hv/H3V1dXeFu6rNTVVVVn74vw1zLbI93shlpuzMmsC4MU+1Lm7+5p+e9byUD7m8tnTLzMB+g3bB5R8FhUQBqTcChdxnx4ccWdWXW7J4XtGXhpKZyZhkpPgpIgzWJMQOQQtxf+SUC5FdC/r/aF2dWJicnE5nuyQ7YvfnWnLI5wdyuua62ov49RPYmwwb5Al8ymUxkMvke9u6LiPA4TKY7kzuvvP4dSvxDIr6gsB+tUIGSGjaGmeF8+BqA60T8HxYunf/YtgqKu4IbAKDTKxvPt4rfM9vxm10mCmdtwrow9+u2pc1zds/vRXEaCUPXUpIUUnz7Q03LIPQeL/4Za6yBqosvKHIudIGxSWOC9urymSfnSe52jVYnLQOmlqh9SVNXW0fTJwW+LPRhSlReCGxgmAxH42yqeZSPMYmPswbtteWNp2W6d6sh9a1Vk1ZJXA8oLbixh8X74nczzzJnMhmf6c7kaqfMvFjI3GVM0Ge8CpU8NNHawAj8c973/qcPtayto/lLkfGmOL+X/RNTu0ZiB0CrKxr/04JuJOLxzof9pO2qztrAOtd7/aiN7tMA9rahFbwleKHy1LFtS697AhK+W8T/w9pEng6FQLChC71hU2EY99SU11+Wjoe/YxWHXYlEtJCDOplMmoUdmefaO5q/uSHnpoQSfrXAkCk2ZIRh6JlMGRHfWVvZ+NkBs7rYvVrpShRXihUId/cBzfMsV5100b41FY0/JYPrmaNBeo2S3MhwOWAR/7Tzuc8ruSmtHfN+sOihlleSyaTJj+ntquGmCkjspp/SMKGuYua8BAf/p6oJr14oqq6pAj6Gkt48yvhLoks0hb3JeIE9nFJnR1Z3d7cmk0lzy6LMG5MOmnyLAuWBDQ6Pp1QoEqUST0Slhu27Jx103PFHHnryQ3fdf+ubBbQ6u/Tysshq9DNSnEyO59a7Muufefmx+w4+8IRrSWUVVN8Z2GBfhZIqVFSUGAnDtm7ShOOPnrTfUffdevf166uqquyyZct2+SAdd9xx3N3drUdOOO5EJn5/jELpePrlx1rzX9sdDqC7u1uqy2e8KzAl8xJB4iIvXkVEAYIx1hg2pOKfUC8pIv50W0dz9pmXujdFe/44urtnyG6gqKFkMmmu7r5agKzWVM6sN0zXGWPPDCX0EaVGVDQDSAMTGOfddb3DEx+8/f6mnhRSnN1Lxirfch64sFKaTCZN+9J5L5bY9eeHPpwf2CDOhVWjXp+oEyfWBh9gkfurK2Z+vu6oupLdy1a5eXid7Wp+s72z+bsGrtz53PdUdG1gAxNdKuqdd94aO8sMK83WVMw8O5vNun+WqoTuprpHTAgoALS2YublhmybMfakMMy5aJyPObCBEZFnvISfXms3lLcubb66teO6tZuHyaS7c7i/trzhuLqKmX82xC3ENKmAxC5S8SAmZsPO577b1tE0O5u9pmdvaBe95T1woSdOIcVzX5zb+8zLj/35iAnHGSaexsQUgQ2IKCa5I6aR1phaKR123lETJ7/+zMuP/yOLrKaQ4mmYtluIy2IvR8lk0ty8MLPumZcfW3TkQSfcBNL9iPgEw8wiqipejDHjoVp/5MHH86mnH3t/npxtZz1lnweeePwJeQ8s3i155pXHd8kDJ5NJk+nO+LqTk+MmHXrS762xnxOoVREH0shwVTeKyI96vH7kzqUt2RdffCKXJw/s7u7eLcYS/zzp7u7WuopZo46ccPzlIJ5rjT3VS6iiqkRk8omEYctQ3STQT7d3NP9f3FqkLM7WvfW8vyW5kaPbVAkAtXc0fUPF16vKciLSApkMo6rqnPOGuIzJ3lRb0XBLbXnjafn8eNcLXVvkyZRMJk17V/M/Wpc0zRSV94nI44ENDDFbJy5U6DDLwTfXvmBvryurPzIfVew+onyWXdYVymT89PKGKUgk7gxMMMM55yHimNkaDth738YiVW0dTV/JdjW/ubsKUwNoY/M/j2oqGj+oJA9aG3yTCGNiqZQ4ZM5X/g1E5UlHWtO+pGluHEFgN0QARQP+56zoxZSVzQnaOlvmC3SJYUtQ9TGzpY/dsXHeScwdfSEB99ZUNv7xvLKZJ+QNub/IsnsMOX8xtHc035Sww05zPvc/gG6wJgigcM7lnOWgBsbcWz2l/rx+aZidC3+V+41WSXfSgLVvCKGmvHGWZV7ETCeEPhcSKVkbWFF9w4v/dGtH03l3dLYs3d2Gm0qlOE9AB0BrymdeUFvRcLdlcw2TmRy60MfMKiYW4ZE4b/CGDYnK0kVLmu8rK5sTxFgA3dtPOb/FCb65q2tuWD2lYbYhc57zoSdmy2TIGhu9ZNV8ZZJC57wSbMDBJWJ0cW1F49zplfXHZjIZn6et3R2GXHgx3PLA79a1dbR8XUBVIu5ua60lZhv6MEdEEwybW+oqZn4pztF2Ki/meLYQAFh3vH0UeavoQqwrb/yeYXMtgNFOXGjIBIYDds792ZOe0dbRdFW+Grx7DRcUYZ5J68pnnVtXMbPVMG41Jqjy4iQvkxK16eCJiCwHTCAQk3U+J4bMB2orGt7d1TU3fKswj75lDTiVilsJlY1HG0M/UagxbI2o3uMUSS9uAROTNQGrwkcFjujmjnDVNNya4GNGeUlNReMvplfWH5tBbMi7ySNvFlYvaeoaucRVO+e+CMUaa2zCqw8Vaowx36+pmDm/ZmpybEHrawfCZkjejmQHQ8Y+9fqpybG1FY3XGxtc7tV7FXWBSQSqeMX73P9r62xOLlrS/FS+oLQ7ikJ5Dx4PK2hdZeOZNZWNNynJImNsrcQpEEBMMTsLAA1sYFR1o5fwuwr/XhVdScREQAJEV15Y+dEDor3f++V1+K1LZB7BHFn110Rmv4iK1r+ocB9Z2HHdn9s6mi8M4S/y4pZYa02ecjbCC0f5cWzIIwMTfMKAF9dWzLy6dsqsY/IeOYUUx95ht4TVGWR8W2fzjxzjLO/dnYFJBAAQ+jAMrE2S2jurp8w4fkcFzHcWsVFVVWUzmYw/pyI5mTVotyZ4X+jDXDSoYa2T8EYjelZb57w/5Cv4uwGC2Efnm/fgtRWzzqqpbLxJFdmA7UUAEJO1U34QQaFiTcBMTM6760XpjLaOlq+2dcy7WSGfQQSMDy0HR/TK+p9Hv6ebiga8B66qqiqTTqelZGPvNwOTqBLvHQAnKnMWdmSemzw5mQCUFi1puYVWrqryzn1MxD8ZtXbyYAtsbsigUdbYy0DSUVsx8+q6ilmT00hL7B1oV8nSY49FyWTSLFrc9Lc3/ZPnOR9+DaCcZRuELtdryJzMxt5VU9n4/tiDUD43HUThUbf239s1IiRNNpt11RX1NQkEi5hNWehzvZZtAsAaUfeptiXNF9+2tOnZ3eR1N2sHAUBteWNtTWXjAkDuCTi4CECelD/Oc6NahmHL1gTs1S92JO9p62j6wMLOpoeTyaQpK5sTtHfOa1KRn1oTBKEPc9YkPjB9Sv1/7O4CYdGAd1NrIZvNutrK+ouNMf/pfC60NrDi5bvtnS2tVVVVNuKxIk0mk6b16dbets7m3/T0bqx03v0XVF8JbMLEUztbGjJHhqyQztrymddMr5g1tRCNtYuG3OeNu7q6wraO5v8FfLUXeSgwiRLvnSNgnCH+c2154xUReou0v6I6BLJoHczgUwxAM8j4mvLGywyZW0E4yLvQBSZRIuof8IJprUuar45msHfZ625muFVVVba2cubFNRWNdxNza8D23ZHhhgWGi3jqy1DUa/ZPiveXjloXVi1c3HwroH3RQFfXXJdCikcZ91/O57KWTULEOWvM/9ZW1J+1txvxW8qA8y/tvPL6d5CaX0QQviBwLnf36I3+26lUiuO54C1y0OwjN69u62j6DkCV3uW+A9WVEQgkRiHmw7X+0Hq4tfaDBnpfXWXjgtryxtrCsG9XDLnQG7d1zPtLiV1fFbrwKjbWxqmAszZI1ZY3NE8vS47OYPuHkGVolec8nLGqKmXrKmb+0BpztapaVYDZWu/Dn/lNXJP3bgTaFY7szQx36tRkaW3FzEuGbZrwABNfb42dpiq6uceNaAKImAIbGIW85CX8qrI7rbWj6beZ7kwu+plUGA0oAGQWZzYp+0u9+OUgYgINB/g3sTaW7K358FsJyEHTMI1K60oT0jP8BsPmWI3Wq07pfQsemfdmNpulrVVFC8EWtyxsWfP0K4/dedSBx10vhGEEnMTMRrUwCo3Enb1InhXxHQBmH3nwcVVHTTxx9eyXL34qgvTtmvJhHh56Y9uNPc+88tjtR0yY/DgRnW6M3dc5lwts4iQCnX3YhMn3Llg4/80YgikDgRxHTDz+WCaqZ2KIyP3PvPLYooFAjjwMsaqscf8St6LF2uBDTlzOsLFQXSnqLm3rbPn+s68/GiYRATl2xXALARhHTTxuVomYXxtjL2OiiSJORVQi9Q0Uts/UcMAislLV/9CH+vH2pS139MMytw4SySKryWTS3NqeWTHpwOOfN8wzRL2zJjHeiZ/4zCuP3ZCMlTCLHvjf2DJKIy2yYnQqMMEZXlwIIhK4T93Zn6fJ4GALpWQyaVq75j3TtqTp4zl2ld7Lo8wG/QJpm2ETnfO5UFTUcHC2Yb5pcfkT2eqKhg8kJycTu6p8WBglLOxo+TPgz3LO3RbYIOF8LmQ2lZbNndVlM87IZrNuaxVqu1kOrNvkRq4rqz9ymKE7Ahu8O3RhLjBBQlUfDkWmt3fOa+rr6+7cPPVmHrdmanJsTXnDZwDpYA5+a4w5yUko3nunSnkgjhYMZIhhS96H1xuLU1s7mv87PwgxlF5zPlRe2NXyZyf+h9YkAudzobHBrJopDR/ZW0NpfqvkvZlMxtdV1Ncw2684nwuNSQTeu5+3d8y/acfmfynOxVIWALzXVwiwKHDBcbuCrLXG2sAy2wBQOO9yzntvTHCmZZtZt499sLq84UOT+w0ZO3lI+kTK2pbMe769s/k9oQ9TTCbw4pSAg42xd0yvaGjMZrMu3zfNf7OI9L1nGXDI85Xm6sqGSliz0DBPcS501tpE6N38TU6q71za8lD+7+0M53Wh4dZVzBpVU97wGRLbEdjEz4jNMc4778WFkYgd28AGxpogT7Aum/FlEic2vPLy64XPPtRnymQykkKKc8NL/jt0ufsMB4EX74n5h+eVzTyhADBTDKH/lXnv1d1XS23ZrIOI6SYQ7RtTozw8LBhxScOL73FXd1+9w1xY1yxL++qK5BEJsousscc6ccgzX1oTGFVxov4mVf9XUR3BZMZG4BBlrz6EqLIxE40x7x2ZoPOOmnjc+mdefuzx7u5uiSZfphF2EGudx3lnkdVnXn4sO2nC5KcM8dlEVKrQwJD5wJETj9v026ar7gOA0tIz7auvdsnRh5xQRuCLY/rIxc+88tjCcePGmcMPP5wjXuSGOgO+AUQTJGbBFJUr2juaP7Ps1cc2JpE0ty+73e9MLz6bjSa0qg6rGvbOd5x2KQi/DmzigyCM9eJCVRFmttYEholIRF73Knep+tcAPtwaQ6IiRMSiIoEJ3mlKR1RNPOiEW7MPLlgf5a5D38dpmEbXLPtpeMQhJ9xHKh8g0D7MXOrJV5561LFN45aNC7PIFj3wv1gbiJXdr5nN4aoqorIWbD56ywO/WxeNeO4YFU0aaZl+esMEhr2ZjT0m0mGKCCUCa40Tf58wTWvvaHl/W0fL7PW29BRVf0How/kisjYwiYCZrYgLvQudIZ5i2DbVls9cVF3RcE461urdmUJXvjiTTCZNe+e8ppzIdFH5W9THltCY4Hs1lY0/r6qqsnnEkYAC6qOmpRAAHbP+GMpms662ov4SJnMDIlpbEPA6vL+4vaP5mzHT586EzJRE0qTTEXqtrrKxoWT8hPutDX7BbI51PheK947ZBtYkAlVd71zuJu9dPaTnlPaO5ovaOlqmQbXOi19qOWAAQiAKXeiNDc4aZnDzBWWN+8dFNNrR2fFFS5qf8qqfIiIWcWFgEqeu2cjfjb/Oe5UB7P3MlI3/mTDB/0U5oQ288x9pX9r8+52gziFAMb1sxijDts3aoNK5nAORAYEIDBH/v7xq3Ddbn76yd6BECwDUldUfKcbMBvBha8zhqgoRF0KJjLXWiwdUmyD8zbal1z2xNWJ57ADQIpvNunPL37tfwKW/NRxc5HwujPO7W8XYSxc+eO3y6vLGLwTG/jAO81vaO1saAaC6ovE/DdH3Ig1cY7y4x0R848Kl8x/bWWqZws8S6Q2ZK5js2Yh4rB2gYGNtzIr5DAR/FIPmRUuanxpIkwQAyanJ0rViv8VsvijiI2NV+IhJI7yrJCh97ynVh2zYUfrX/HPWlDf8JLCJz+bPjnhX39bZMn9v4cWivZ0La3rFrKkWuEshgTXWhmH4m/alLR/bmReQ5wyurmi4tsSWzArDXgciE1PAOqhe1trR9NutGN0WZHc1U5NjIXYmAZ+xxr5DIrWEHKA2MAmO1PP0J9pjfrjw0Ws3bE32ckcOYqwF/H1D5nNOxFtjjHfu7wp0EbSGmMfHabxA9TYlcob5fSIixlj2zrVD+MNtXde9upOHl1JIURppOb88eaBS8A0FPsbGWOecB6kYMgFFPM9PENHVNsCfbruvadXmHGPzJY+7LnyOuoqZnyeiH4lKbMTqg6DE5lzvb9o7Wj62EzO9BKQoObW7ZI3YrGFTriqiihUgX9G2ZN7ze4N28l5qwClOpYDue7qHr91kH2AyJ0RE6PLYsNLEWTdnr1mzoyyOfYWwyhkNzCXNsU4OR2M4JF7k/y1c2nLtYBzTA7mk6ypmjQL8R0D8GcN2khcHUc0xUcKwhffhXwX+S+0d8+/engTmUBkoayoaPsdkvqcqlomZ2eRlUvsLH8aCAITeecvGiPo/YcWqj7U+3dq7M6ydAwwtqcB3A2MnhT6EQkMCBbHk5+uq+PFqol90dFy3doisk32KkjUVDZ8zbH+c98SqEGuMCcUlF3a0/HlHL578hV13SsOJSFBWlUYaNsZJuGj0ev/uycnJrkjs/k8MnWvLG660NvFp50IHZu+91C7aKRV1JYBwbvl7x1oa9lcmc4gXr0Skhq0Jfe6LCzvn/aiqKmWz2bTbGX3bc8vfu5+h0k8x6LPG2LERGCQCmnjxXiE/Tfj1/72ga8HGnfWASSQ5QlDVzzRkfuehlhSKPhBEvzIDSMVyEDh1t51W9473pNMRtnuHydLRz4dVUjLi+4b5UkAh3jsFkTXWeHUhKV2j5P8n8mw7zDrZt5fVFQ3fDjj4mhPngUgryXv/PIwra1+cWbWzF3d1ecOHAmP/4MWFxiQCF/Z+u33pvP/e00Nps5doIfH45Hg+7rjj+MzSM+2fFv3J1VQ2VDPxT8V7b4y1qu5rCztbmnZmw5PJbtPd3S3vmHjKl61NXOS8FxAoMIFxEv5hYee8rySTSXP77VfLzrJxLFh4w4ZnXn4se/jBx98I0QnW2OMAGK8+JMAGJnGap+CcQ8cf13nbnfNf2xkmjuNwHCeRpNcOeTP0oI8RqCTWaqEB00kMJRhjWL3c/tumn7dWVVXZa5Zd43fGeKsrGyoTdtifrbHne3GqIp6IrTWWnXd/YcIHWzuar3rm5cdX7ywrR3d3N5LJpLl14fy7jphw7ImBCSaLiqqoBDYYK0Kbnnn5b/fs6L4V/NyHjpgw+fDAJMq8D0MmPuuoCZMfuGXh/GfKyuYEZ545iY477jiOATBFD7yjoVnhqpmaHEs+WEyMo5gMOcndeFrHOz/QvXP6vARAzy9PHuhgH2Xm/VXEszFWRP7hfTh1UVdm7Y7c7JpKMY6LJ12SkxWZbrrn8eV0dfd4znMx11TWz2Hw/zLb/bwLXRTaBlbErxTROe1Lm6/f0UJSX2RS0XC53VxuZitMlbGao4TZ0Yf5c3dw3ygapEhLdXnDhwzxlcxmpPPOR5/DGBXpUeCbb/p9ftDVNTdMVVXZCcespznTJwmSk6Pfk+kmPD5ZachhapQqTD+lYYKxtISYJ4qIRLRE/g2IOTkSUVPaQaYNUihqp84YQ2LvYeYTVAEVfcrBnHZn559WDPVs/quXxR6c52YyaV9WNifYHxtOgNFDSIUV9Jp6fNkYc3SkaeueLiG6LI20IBMB8XcUwZXJZLyD/XBgg3GhCz0iZTxA8LUdIYLXVIrNN9NC2xHbmj8/aR6/ajmls/PmVk+Z8YCCfmdtojx0oXc+9IbNWDKYX1Pe8LlMpuXKOKzdsVaYYtygf580jq4x+vHHYYCYgnfw35OXqJGaisZvGOK0RtVtDwBRyOyfU+Cj7R1Nd6dS4Dllc4KPz50bIgt8fG7X1uaVMe8DSTNj0D2OWkCZTMsr1eUNXw+I/yDw5EUksInxoc/VA/hJJIOzQ1OUOiM5w7RnMivrKmbMUaG7FFpijDlaJPxDTXnDVQQ6GFBLME+B0JnJXLd2iPv1dvTA0U1bV1afhDFfUdUTmdnmRbGYGSIiBO71wHsXdlzXvrN5IwCtq6srkRX7LrXGHu/FhYaDIO+ZhirroZpioujv/WhO1ZRJB4dnjRi2fvKwQPfJCa/udcP/9uJrI/7y8R8v+hsApJLJRDqTyU0vS442JrjKGjsr9KGHRgp9xlgSH17R2tGSLlQYGGJt4OvWJr41mAcObMA5l7t5YWfL+1IpcDoNGeJ50dqKxp9ZE3zG+VAAJVWSwAbGSXhvLtf7wbsfumFZ/jMCQKph+oRTJtt3jSpdVQasPtL5xLre3H5/fWl54sHLft7eGW1himlwpUFKIUULyl41+/GaB60JyqJ3Zq3z/mFetf9prU9f2bszxtWfD9d/wZrgh168N2wMVMEcgdu8OEDpKSH9QfuSprn/biPmPZFJA0hLdUXD5ykI5jPzqSBY0ej/Yn0fZ9iyqOta2HFde8yo73cGPw0AfsXYSiae7MWr5usCynMzmYzvHnzomyLyw7Rc9ZnTTs/+5Mzbayp6H5x69OofTT101aXlB6xpmHrQuk9UTnzjqqrJL3cs+cXUm6/775qydCaTS6Wq7KKuzJq2jqbZocv9NODAEEWyw947MSa4orai4btRhTlFg164mbxHM4siBQIaxHMrGTK3RrlgkoZiOLHe0S+sCT4T+mioPm+8oQtvJl11YaHxfq3+rEOyPzjjqsZzVj009fDXmisnrv7SGRM2vu9dh2784OlHrPjJOce/8cCDV56WnXv5WbWUTks87ri9Z9HuZDd1dc0NCfh51BaG8eKJCCfJvsvLCt8tdgx3LgDo9M53/sSLe9iyMeKdk3gqKuLcEiWmoy3bX9WWN6T/WfS/e6UBJ5MReue8ioZzLJkfiYg4732fQmf+tiMyIl4JfGJN5cwyALIrzBgEfbcxlgF1hgx7cS96mLaCl7o9z0tEpH/+xhmfrTklvOvUCc+dd6B9wtr1y33vujV+3YYVftO6DZ42rPEH8MvDjh3z8nsqDl9zzz3fO+fL6XTW3Z2qshGqquVzTsIfWmNMBOYHhd55YxJfrp7S8B30I4Ro2/YbDT60dly3GEAmsAmjUDfQiBUaBtYa59xjI0tz8+NhABmsChxrMv8wMMEnnAsdAUYVYm1gnAv/zKtW1bd2tK791Zw5QTqTyV2bftd5jdN7/lJx8CufPKjk6fE294z0rHvFr1+/1m9Y/4bHpuf9fsHT5rixL5919jG9d9z87bPSRKSq27+s8s+qG+QW58OXDBkGNLQmYBC9exewBQRAl5z+jyMAmigqsdhxNEoaRTNEXpx47zwb+426ysbz0/2oure3AUcqd4BX+iyzAVSEBrRA8oYsqhIEiVGq+qG4mrij6UCBMJm+S1WhIGU2UNE77+z804r4ZtXBwuam/552ecUx8pNx/EyJW/2mtzmvBmKIjAEHRowab8R4dbppwxt+lH12n3cc+uZ3b/tx1VfPTmfd5MnLCUiato7mLzkf/j6w1qhCCGDvQp8Igq/UlNd/OR5oGMo7o8Dj06HrXZiwJQEVCvMCCExJ4MS/RMofyWQz6/MHd7A6QU1Fw+eMtZ8PfehBsKrw1lrjfHhnz/DEJa1Pt/amUlX243Pnhpn/Pv+i0w7ecMMRw7sP6133kvM9TlmJlckIs1FiAyLDnlTXr/Dj+B869aj137jpW2d/lygtmkpt1wunkOL27sxKgNqiSTHSaNgBZ8TMlTsMwOiLtkJuSJjEuEh0fEsbifm3wERQxeeGcAG+HQw4Cp3POSk5kYEzvXhFHyn31sM/EVGCTinIf3c4pz+3/JL9CHSIqoAUrJEq6V8A4J6qe3h7BSuitMz97Gm1p0zy3ymVJ0TCXlEaZZyxJCYEq8J6gNAL5RDKRKRs2PdIkHvev3P/ld++4dvnnBd54uUEKAV+/adDFy62xhpVVRA4nnD67vSKhsYhjL0pACzoan6TVq66MHS5y1X19TwntgK93oV/JOenRaLZ2+/95kkSaipmvJfJ/MB7J7HGsRhjjPfuaasuUjhIVdlvfjPrfvzZdx37zgmrfjchWDbMbQi9wz4WFBDFB87AgeHBkWwVhcxGvNCwDc/5Uw7Z8OU/XH7GDEqnZf52PmdsbATgblUFKVhEAKXJtWWzD4zVOGhHQ+gIO661AtFBiALZiwegp5wzNTkh3nd+2xpwKv63SQTjFBilKjTIBkJVCEoHrXoW++xs/ptAeARID4rCJVgnzoP4EQCYlp22TXQQrkjr1IMPLj35cL1iwrAX2PesVZBhT4CQQNRAQVDyfVyQ5AnWW9jQstlEOJCX8+Gj5X8/mZy8z7Qrsj6ZnMELuhZsDJg+6MW/wcwEqCqEVEUN0S/PmxqPvW0/ZVAA1Pp0a66to/n7AFIEppgX+w+tnU0fau2a98xgUME+uGpl49EEOxeA0dhRMzFUdBN7+vDtnZnXqqqq7BVXZL0q6KRDc98+Yr/lY33PJkc00oix8CxxhYxiXAlDoBBycCbKizgnNNa8hnce0fO/7z7z0DH127mUJ0fRmoLMY05cDpFsDkC6n0IOieopV9COFjRXv1SyH4iOFpE8ccP2d1lphBXaP96vYhEL6pxuZeh8q24myl9DCkvcTv86lvFMhiNaWQJB1xH8imhIYZutICaCfnzG4adPGL1ham7jOmUaZVQDkHqwJ7AIFB4KAiuBBSAwlAFhDzViNm3okYNGrT35zGNLLySCJpNRHeC2Jc1PCfxXYzVDEIhFVAzZUV70t9Unzh6RTg/aRdCqqiqTTCYNEdbk/yYBrwMpnlM2JxgE50tIAVVVKWsEV1ljx4mIj8JHRPhp9d+/o6vp/qqqKjttWlaIoN/98AnHHzgyvCC3cb1CA6PohaIH3CfKpNBIUqmvqGFUQVAwEff2bpQJozcdOXvaIecpAJ2/9ZQhL3oeJOQFKJYzMUHVWRMwmI7aybQKpL2OIgXHwVt3FNUT2NP67Z2Xt4UB51+Iqr5Iipc4qsbqdiLoKF8lemzho9duGCxf3dbywuOIGKQkFL2RN72TGJKX1q0jkKJ18AF09j4lOaiKGEMADay3DfL22emoYWv1kHGJmigqmKz5ofN9D5M/OB/el+etjhUkfGCCckq4L2MIY2/jx4/XTCbjVWRY/s+EYIC0xNrB22c4SaclsfEf/2FtUJ1vSSmi0Nm58BHsa/8vhYhn7Irjoir25KNGnjFueC4hvlc8g5QAFh50LwQEz4Cq0304p/uP1Hf3AWG29/5eW7mRSNeDCCAogUCQ/XdO9y3F7YszK6HaYdhAdTv1D6gSsRLouZGb5MXtnZe3iwfWZDJpFnVl1oBxvTG2jxVyq+ZLYFEPqPxuswLEDlegaTgRRc2QaOpowyubsGn78Xd0qEYN23RIYHJQCER6gR1QLYl8jgHl1tAodkcDYObII/YNQxj8j6hoPpQjAjtxwsxfqD69X6h88M9otR87Szo0EE1GqitmH2HAX/PilPLglj52S/rOwoXXboj3ve9n7j/CH7mPWQuVHKJ0AuDYrLaDKYGC4ElBpDTMbaIxwza+o2BPtvnNlbMqQwWtL3wIYhq+c12Q6AyJ6m9FPPLDLNu6OwwZAmlTP5ke3tYGnM9ryIj/YRiGTwU2EUDhBnJRRfOrlpzz17V3zrtNVWlH+8DLly/PT7ibgaZVWjpGhwYk31RK6gClzaZ9Cj2wFvyzRfNKmcj3IhA//oADqktVo7+X70f2lry6SMR1GLZ5ahlSUTXGjmCn3xlq/1HJ9e+fDKUXn3cy/uvGBmMl7qcAKpYte3GPrbXhLZtVX+NLTf3G8VY3gQAIEZQA1sHa0fF+wIBAYOnFPjYcBkwtGSyZisNkHiBoLjvLP6aqtHBpS6sX3xLtOwaePQ+FS5iSIHS5JSV2+C+KVegBdKq3d2ZeE3UXOwmXGmutYbvZTUiARiKh2ABAZ8yYscOfY/z48fGxkk2qCihF8qNEI8Zg1ZBucCdBr8ZdLiJbCFSKc73tnD4SAD0QEjg4cm4DD4hGOJvNOiLzJ9q8eGdCnxMmrltS+WTdUHqQJP1eV3n7ldl8L/788pknM2Gm82FfBKBKSsSA0vWLF2c2xSF8P64ZAHHJBqUg0gbU+GOShyrF/2yZGDEiLx3FwAIhQGiYAG/K4Bfx8IBUR2rh7iut29kzGJ0lUgK/GdVENn+J1ljDxtrQ5+4VMo23PPC7dTs6AfVWp9QRIMULl85/LHAjq0LvLxXxS5k4HyVJJFYVCrP5aG15Y+2uMAoaYG1sahQBvfQgC96/oLG/las6Oqwbeoe9rloCAgNioH0MqPlCjWxZtMyXb1RhSJV4H4SSeHXFigc2MPUfwjgagQhnQ3G5WIBN+4JOYojoV2NityHTyhQa83aiIHjSz1kTDIs9EOXTFuedF9Bd2/p9G3sTL/XqCCgxWBmsJn5o3SI6KRR/ISg4itLVWaNre4atBp7O0SCfqnTNxpFK2EdVoUosKiCl13dFT6tmSv3pIHzMeafEZAtyI3Xe3abiZ9DKVTULO659rgil3AZoHUjxgq65Gxd2NP2WwO8T1TcDGxAUd4qXbzEbJpAB4cdnnjlzzPzMfNk5bLd9UcQLETguTIwA28O3l1fHiEW8vkLvXdsTgDgKBwgcP4D0+d+th9ASt5YCiBmLVRv4YQDipb8Qly/qrbO9zwD6XExrGwebMF6cMvPpJT0HnZ332Nu8EWmokzlR26hm6uyjVPG+KPftPyNETFBZAcv/iFgm50t/+Bn9+7lXNv11VU8CIGZSC4WBsEM+ElXSeH/ye0Jxt1ZA8Z55k6A16/0DAFTmbT23TMVAj9DS4QQar1GtwHhxIRl5uvAyGnI5JA1Unzh7BJh/yswlzEyi8iNVPBCYBAC8KuJmtXY0ZVqfbu3FThZO3yakdmkBwMlk0rywPrdcVV4lMBQI2pe2fMN73w4CjLHHlvbI/1AklcI7iHuFSegTqvpK5OHVGbYQ+JM3y5O3CLMyogr6Vcc721/dOOIRHjaMBL2elUBqQSRQ1njYiyAU/RPleAyFg6JHA4Be3TTCP/ZaMD96pi0vjMWLM5sIeG1gMBfXARhC/y9qbc0fUg4mRDxYEYec+2hgg1FR7hvdSaoQJoaCntv34N4VA8WxZ2SiPbn+jgPufn3j8KUlw0pJvHiBRWgdlDXqj5ODNyE8O0g8D+WJ4CiE0iZNmIBeXzl+w0tv2D8BwBWPb90I49yXDONYw9ZA4aNuAlZtEr98R9s6eagoStzllu2UiFLY3d9bWvI1xCkboMvtuHU9cbS3R9Dt7Onse5LJZHx3dyYkoo0aKQdOSk5NlorYj4nIGxGnk7mspqLxvUMAOGwBdrjtvqZVIHqCmKHxgSTgTAAolGHZkqc5ydnsNT0vv0E/d3wQMURVBcKAMzlACUYYVkNw7H2E8vHoMKiO8LzPwfzKqpLrP/ez++9XVZoxY7NCXD9IXvFqXCHXzZFADkQ4v7oieQQRDQlUTyRme9DSqqrkPsp0saiASAcQARBAWLaN6rdmMklufbq19/nXS7692o1BkFiHhKqWhCPB6mDgYBQwEsB4glEBowfGJxDIMBiz0VPJKH59w35XffTK+7pVI8qb7cBulZTPpahYJhy9w+5sZ+b16DUOzcD66JSmzCpn4i+JehHxGwn4eLzXhyoUKrSitbW1N778pUgrO3RleIXqytj1DF/rEocu6vrTC6rytQLeqh/WnZwcF3EYDc2I+z22tBIIpJFRgHDW9NMbJuR7g9vywqlUit+TPuT3/3htxPzSUeOt5Y2OfKjQUoA46gtrAJIArASjObD2glTdiH0Osn9fs9+Tf1my7gtE0K1F/+n+HTBbTWdFxJpgNJG5eHshP5FKvnKkuvV3nt+LYRvMOYb4HSJOow+xxbXXu+0CUMarpjj5rb/c/ORro39GIyYaog2eXU5IDaAm2gtvwGoBNQACGCVY5MKgdJx98s1xD99wb+93NNU3QrmNM5GWiFaWpot4EDQqOUEfGCylGLg9kzOTNTk1Warkf8bMw5mYVeWbbZ0tj5ds2nSMgsapKojz7cUUFXmhhxzaRBVmJTwX/3u4Gt0HALV3tvzaeX8TE5MxZpIE5lsRTHmH2laA4QXOhZuIyIqotyaxnwnxnsKwcmun6Ior0qqakV8teu4Tf3t1/AIdPskmjKEgZE/OSAhoSKoqqsZ7Mc55kNGSUQfaZ1fv88I9j69r+HrmkZe/8Y0IF78l0iOdTxwPjNpUAyvIpKqqELw3BvD7IeAFtw5NnBzvBeO9zAbxRNSA4Q0FSI9KJpMmvU0+5rSqpvjcy8MvLX3xkD/k9jnS0rAEe4hXsFdlUTWiakRA4qDemx61o8cFT6095omOp8c0/PTm7OorkMZW9mSzM5Ez+h5j7AQv4gGyzjsVpjt2RpJntZivGRNMBQDn/W29w1/7YXT3mQOZ4r5y7ET2JF3hPd6A83koKd6Msf2lDD2gD/JE/gte/WsiIkTm0prKhup0Oj2k8cL8IWx/sPkJAH9hY0CkcdsTn7qgbM7wfG96G3hsBYCm215YddpnT774r68e/o1X9bBVwT7DTenwBCcS+xAHpWRLLJWUJjgYub9ZSUfTX1+bMP/PD4bVX7ry0Yc0ldruIH3dyclxAI5UFQysyMZgFiKiU+sqG0/oVxjcRhEr2smtsbBQOp2Wuoq6UQpUiUjUSdnydwGKY9c8xxO25YmiPUkrUVc47T/v+cgDz+9/+Su5Y16zIyeaxMjRJjF8GNsSYlNiODF8BJfsM8asxqF4dPnB8/78QFj3yZ+2PaHbJxeIPObkZAKKT0YoapXo0pF/DHMj/zqUMdDCQY3qyoZKA/qiqoiIf0PJfCabzToAKqQjiTguStKTKCoz7HRvaYVCYZhJRCdG2IFksLAj85yK/yKBmImYFD+qPnH2iKEyjuR7mUL662ikEMZ7J9Ykju/l9ZcOxtRPEYaDiOaGZ3/pjm+1PrLP6Q+9dsC3/rHisAde7jni5Vd6Jqx7ofeI5X9fe+SjXcsP/fWDzwyvedfn7qpP/+nRJzWFbTJQ5Gd/NbCnGuYJEZnBFiEtRRFDMFxEz9tZRFr+85GMPYPJTBL1Slv9XSLWBPsy0Tnb80T9ewK94Kt3f/8PC3OVXc/v+6W/Lz9owTMbJvz9hdyBL76waeLzT606pPPRlw/8eeczI8+t/Mz9Delrljwf7cm2L7Q+j7lPcIk1QZn3TlSJoyIkX7Oga+7GIaKiCCnggrI5w1lwFbEZBiJW6FcWdlz7XHJyMgEARjAsT4ugoJf3NLuwe7rh5gEXDHpCvIcxFkrYHwCWj1secyTNa6opb7jQGttgODheS8L/SqfTXxsKzU6+h1ri198e6si/W2Pf6b33Xhwx6zemVzbekck0P7U9ytXYE9P8+UmeMSPzDwDfADT1ifedPm6/ERtGbOwdm/txBiuBxZviPJSuuAJEg1PYqEI/zMzwrm82emBTNYaW8jQA3xuUgIC2jVIS1nMtWzgnHgQb//4tjFmBOVVVqeuQ6dbt9EFVtW9PXvheBj8E8MMTD6gecdZpMmxVOEyabhu+FjEH9dD2JOJJm356wwR2mlYVVZAaZg597tVe+GviYpwOgVGT0+m0r6lo+IrloCxWrWg/rfOYP5yWSvE999wjcfQykkEQ9SD4N4seeAdXX25m/Buq6uJi0z4DqpEUmMSXvPrXNJoD+2LdlFnlQ6xKazKZ5AVdCzYSyQ8BigBE0fTPfqz4RVnZnCCaFNv+CN+MGRmfSoF1ftIQkf7yxgeX/8+1Tz7348zil4HFm1RTPD+aDtL09r2MyWQy/ryymWcQm4ud36wfq/GAPhXMpRJUTz2//EODzsGSbulZM5mMLyubE0B1mqpAY9wzgShGweWr0OwkFGOC0xMbnmjIIOOrqqrMUPdEFfTo6ws3XHXTnSuabrttFSHjdX7SDGVPIlaQyOMbRz83xk704qPLnQ0R+CfZzsxrcTQhgwI2EAE2CHy5qBevfq0j//k00oJ0v+Mg0sOICCLqWfX1zc5k0YAxFHxlfOsn1irpuggwERUVok2OQtzbFv/xZVV8GUTMbEqU/C8uKLtgOAYfu0MmkxGF0ht+9B+9hPcG1pqomOF8YOy5+/P6H0V59eA/K52G0IyM14giglIpcCqFmJYlLYMxLyYRa/ZMTY4Vo79gokQM4IgNlkhVViiQyxNtqIoym/FewsrCIg+2Qv6tA4w7jzbbz64/EcQnePEgEBERFLLBi3ui8CMTiERFmfn7NWWN79yWJvHW9iQisIj2RaOeGNGMjB+cjbKf2L22vCFt2L7POecBUmuscS58hEC/zA9hDAWw8Z7TPzKSmK5iohImwxCk7+rIdBcU6PJhxMQYtrvJGbMa/665wb3VgPOIpESJXw1gXexGxhdWkfNMCu0dzX/y3t9GIBgTlIU84vIhqs3pFakrqKtrbqiw/+G9X89ETKQxL5X5dE35zP+Je5JDJTHTvFeJPYsOqR+JqBdLElxnjD0hYsAgBjQCUqg8a4UrSPWzESgk4sBjNlD21UOYvtrsgOdZR1i02hqbgMJFYE9SJfpMb6mfotC/Fgick6hXw3wgGcw/vzx54FCMeHMcO5S2NeOxlUIToIhVGT7Pxn7Di/dEERxeVHLE+HRrx3VrkwMmo7aXQ/eEG9KGg5NBBO/dXT3DX/lZquAC6LsIiMbFF9grcG554ZksGvAOLFe6ah0pNsRg9UMAUEErQ+OwRlnks17DN0VFQObyc+NQejCsdDqdji+Bax/x8J9jNvEgHNhLKNaY/6opj2Q700hLfGBpd9EJ5cPm6ac3TBi2KbjJGlsX6TNR3yBBZES467alTc8eJqN+K+IesWwp1oUCiE9P9ouJ9wt8+356mIFkCXmwipBOi9tEymzJq3T3vvZKczabWQ/oXzgaYtA8J5TzTgzbE4TsbdMrG4/OZrMuhRTvRoZGynvDFK6guorGb0VEh14jSAypMZbVu8tblzTfl49choJ1nl5efy6z+XQE2JB1xPr5uOqMgoFuraqqGgbQwXHksqKA4L/ogbFDg9ag1tbWXoBejeJIjK2q+lDJ1gywtWveMwr9GhMzE5da9j+fOjVZur120EBOpIUd834buvC7kZB3BJ7yEkpg7aeGbZpwa83U2Ufl2wzRxbBTh5ZSqchwgbRkMhlfVz7rXOv4TmvsuaEPPVHEPdVHxK4CCHcCKZ7bNTcE0fyYNBERSyeduGY4TxlIq0rbwELnSRBqps4+CqDT4zYRMTEIemt2WbYnGq/l7jx/TH6kLsqHnWe2pxrF3dXlDe9LIy1pRJrAO6N9nH+mfBU5k8n4mrLGdy4uf+JWY+zXo0q8QpU0YRLGO/eTtqXzfhpHLoOGzul0JDZnwFcCFDAZVvX/27qk5dGBoTMAJHJjR5FiX0QYvZ5C4y4a8E60OZT02Ri+MDKxyo/a2jxnMpk0ow+V33pxC4gIhoOK0RKkhhpK5xkxFi5t+Wroeq8ObGAAEkWk5meMrSPv76ub0vjJC8ouGB7d+lEkkExGBzcO+6hg9IYQe6f83wGg6XRkuLWV9YfXljdcqaRtxPzO0IUeUAZA1gQcjbURiSoY+kQe4EGhLnA+7AGRQSyVQkznbwWJpf1l635jjsNnIudqAxOMFFEPIut8GHrF9fk9Ee8fc5HIORMRrIlqBATliH+aJxo2N9RUNjadXz7z5AwyPvaGuvmepHiQPUE6vsyqT5s9vqa88Wsw9IA1wfmhDyWqBZAkbGByYe9vRx3uvlSQ9+rgZygtXl3aGHssEcGLe6B3+LCfxJ5ZBtYFyAcHgXRfVYVSfkQxRcU2EnYOzAHFqjhx2pcTOgbA8liTVgvRVWmk5d1TZn7Ww1WAeH8QPl9dNvPWTKbp/iEo8GksZULtnS2fqp7SsDawwVdEPUTUO+c8Mx/AzFeFbuSc6vKG34H8rQs7Ms9tHsKlt0AopQtGmZKTk4m1I2yZQuuhmG1tYj/nQ3jvBCAYsiTqxXn3U4JeYExwtPPhWjH0SoE05t9qyhvus8ZO9xJqbKbVVVVVV8SawWbg0AEVDLvfk73Hx7nuebF1e8PWOB92jDncP4RO5Cu6L0Kxxhg71ol7UiRcRMSfZDbkvZdomosoMEGj0/B9teUNzSD9Pa0c35HJXNm7ncRlsz0BgOllM08wLB+A9x+2NjjUi0PonQcIzGSYDELX+5PTlr7zi+mlfe9dhyAj6qsr6msM7Ke9OCEgR8Kfy2av6Zk2bfOpou5kNyEDkNhxYMSYAnk+3/vOZIoGvLMBw4qIeZD2IZZxAJ5AanNbSffp5zQ9WxMp0F9DoATY/bSu7jNnoBIh0oOGQRoLXhEtpa/WVDQ8xcQ/CqwdHXoXnVjxZI09yRD91Hv6Zm1F41+g6FBop3CwzDhZXVJSsn65rHPjuddKb8mYnCkZQ+qOAWjqWuAcEE4KTAJePCKvG7VqrAnYe/8KQJ9s72y6uaa8sY6JAaIX7ZhVLwDAPffcwwAciG4moulQIi9OCXRqouegMgBL3nhjfKQ8UVCJVo5K188+O4YJ5M+dcvEkIpzhxYFUOR5YWJgXDc9ms9K7UVYNG0EriGgsKYJRxn9ptcciEvwksMGhzodQhQ9dCCYaZmzw/7y4/6dj33y0tnzmvar6V7L8iA/5TQ561rMkXEmp5TAXjg4dxrD6SQCVg3G6Qqdak0hIwZ6AlCwHLOLXewn/s61z3i/bNmdP2D5dcRp49wkzx4TQn4FgDQVwkvtu+9LrOmMCg63mzsy6T18LXPnV7U2oFUNoDAHMQXhWVGDIsMq2mQcLqtJ/9N61RFVpWyYr3vxSnCsPrYoMQgopbu9o+Z1zOMt712rZGmOMAUi9uDAuNI22xl5gbfBNw+YOFvcQWB7qdRsfHiXmoV43/OEcm4ehvssYO8/axOeNMScBgHM5p+IdEdhaa4iInA/nedCZbR1NN08vS55A0EOjarO+GtUCQHnK20DodufDNcRkoeqNCSwp1QLAobl1Ud8UTJuRcQGYtGkVAYDlxHnWBPuqqAOxdT7sYcKNADA+O15TSHG2O7NegMfjYs6Bq5QPX9jZcmOOw9O9uN8A5ANrDRGMqjrnck5VwWxOtDb4tLX2dxDfadg9RGIfUpWHN23sfdh7eZhIutgG1wdByVcMB+8iUMK50Il4B1IyxpjIeN09yq6qrbPllwWFMh0Kz1UaaQlL5euW7TFEBOdz9/eWlnxvYOi8lTbCeCKKC4S6vPAsFg14Z5bochGvzAyQjN/OjagxGod64T/vxb2ogBLTf51X3jBlqGRwcTgtEdle099aO5rPC334YRH/N2usMRwERDAq4kLvcs7nQlEvRCgh5gOYzCTD9hhmM4mZx0TMkt6HPpfz4kJVeGK21iYsE5OIu9d7nN/W0dywsOPa51KpFFuyY0EUh3H6xGazq0jxbUubnlXIXYYtojRZAUgdkOKRp4x0EQbS91uwxs74OPhUKsWqdHG++myMUVVdWrnkmMcAUAYZyUMzGfS8RnzQw1jMkclk0ty1OPNy65Kmj4m6c7x3rQRSawNLbKwqvIgLQ5/Lee9cVAujsUx8KDMfYdgcRsT7Eoi9OAl9b+gkDKHqiMhaG1jLlkXkUfHhh0d2uOlti+f/NYm+YpMOeUxwamMVgT8t6kVFVjryc7LZa3riyG2bP0dEDouYT7wy9KU9MibdG+w23+9V8Guquo6IoEr7DXIjSjKZ5Gxn5jUhXK6qZIiHe+DKurq6kh1RZyw0+IWdLdcEft1Up+4jXvx9UHXWJqw1NsFkAkS8yV7Vh6I+FHGhiA9V1AEqzGQsm4Q1iSBWX1jpvbte1F/QuqT57IVLm+4AQGVlc4J0Oi3CemQM1AcR/X1rA/hQ/DmilImZOoinVJc9UZ7PyX3BOKJSxPGayWT8g3f8/QQiOtOLgyoxgUiJbiko+BVwaeFRVYlRTzKpD72FFC/snH9va0fzeRA9WyT8naq8Zo011iQCyyYR0dJAVdRJfl802hOFSoRjN4E1QWBtYBXS45xrD9U3ruFwamtnyzUZRO8g1n8a8phg1eTkPiL4MTMnmJgV+l9bA2xs40eMjgAtCL2a5TvB8lE04M04o23uDSJaHQ3F0RGDbWi+Kr1wSXOL974JIFhjp/qV+35hiFXpLSaXksmkWdC1YGP7kubft3U0naXMU10Yfs07t9CrXxYJZ5vo8Bb+YwPLZFhE1nrvHwkl/FMo/lJj6NTWjqYPtHW03BZfOgaATop5m0npWCZGJCNjntr880UsHEGJucOLe5EjQSlnOBGwwXsLXjNtpmXdhzLiCyPwhjpmMqHLrU0IXV+4r3nYIHv/mBefiw40TwaA6Bn7Jr+odWlz9o4lzR8NSU8WdY3Oh7924v+qqquJiIy1duCeGDKsgPciz3vvWp0PL2fmirbO5tqFS5pbYvI8U9D336ExwZIR5quW7SkRCaG7tbWjaW6BcPp2HQaR7hv/0UpDm97cw0BYe18Ra/RarF+zD9bE/IUTCw1rW6FQvv+bCBOXhyZ8l4GZyOCv10xtvDOTae6ICjXTJNV/UeggbSZfCO1rX9LUBaALwP/WTE2OZUdHOi9Hk/I4JexDLKygjQCvg+AVGPuPnpIXny8ADsRAjm7KZDKS95p9CnzQo4kIAllhDT/Rb7gRQ0dcYV1VU95wuzXm46H3EPVQ0IXVJ87+9sJHr92grJSfRaQYjFF94uwRBJ+MRyfVsIVK7q7bljY9W1ipjxUgYOywV1RyawHsD+jkFFKczkR7n2fNSCaTZvLkyZpOp18H0AKgJYUUP1DxzGHs5Uin4aFQjCFgFIh6SbBagNeFdFmY2/R09pGbV28FPilDpAymfIunquoezmQyrqassYKIPqcqIiqr2OvnCaSpyantveeCi4IOjOUXVlkZt/LfRd7+VjBgBUCZ7kyuprzxBQAnKDC6unr2iIULr90wmOeMbtw/vlxXUf8FJZpnmId7766cOjU5LZvNbAKyfTdrwe08JENOIUUxEbu0L86sBLASQOdQAAvR903WCMixpU5PFKKuOyT+gzc3JILXB0b++SIesWnx4j5GIOu9E8PmOF8Sng3gNlLiAVcc0TCpYKITI2QTjKhAma4tbKMUHtiJ4bDlyzi3DKD9ifSwpWc+NRr3YVXhwxQYGuWjm3Qm7dGB54CIkGEomOd8K3Cohpu/TPPPms1Cpk5NlpLolcxmeERxLf/V1jXvme1VnQtXXd1nSmTlmwfGFPwbF3TN3bQnGsZe44H7X1LUCwZ0rFmjYwBsiG7e9KChdCYzL1NTUf9uw+ZDxtiK0Z6+WX3a7O8jzI0Xolywavyygr7lUBA3mkZaC/qYlEqlqLu7m5YvX06F+Xn+/48Pp6aRFmS2rVObRlonmFUH52CPjArQeCKbvaZnoCBZfhxy1LrcA2uGm0ettSd78SGzSXj15wFYEItm5x/RRhV2qTGcgPM5x2wCL24ZwyzcyjC8ppDidFc6rK1oeB7QMijGO0dHAFg1sA8/4ILr84z5fL1wX/r2JPLaOuD7hixIlslkfHJqsnSjCw7yTMOcz6033nzGmKAiz7DRvvSdvxnKeGn+LPmVqyYwNGJCIXpjT0Rh7YV9YAAkK+O9HA1gVF7ZMD007WGyKl9x3p3JxJOU9Avk/WyQGW1JRca++XxtReM1h/mRP5nbNTfciReWP4i7tPIe0MEcAWBkxN2oD28DSKDxwczVljfeTMQnQx1H0iB08QVljamQ9IC+ySVoaV3FrFEC3xjJ3wJMBgJ/e2vHdWvjkFy29jyq9BCg72djh3kNjwbw18299baip7T+E8APebyyLdk44ctrBbNAcrCCjDE2BDBaVFRUXiOWTwNpmZwZvPOQ319WfyCIxsZkEk8NcCLFItbOvzXuVhUQUCrqDx06C0VUtLq9M/OaQK9iNgQlZeIDmaiUiEcYNscFNvF/z5u119Ud9ZmSGFL3b2vce+BEYyzHyiZPDDYzTcTznQ83gsiKemXwgaHRvyjo8yIeEfkbLlD4hYbMYVH4TNZ754ioaTBFP1U86qOLAaR85L/zGCgUVVUfGlayYcJ1CZv4NhMfS8wjmWg4EY2Gwhm2REBT25J5zw+t6lxIQ6SliAsHJLpiTwRx7JUGDMhrIgLDlkCYsDPtKAKdmHesEauDqqqqFyeh6w0TtiSp+775iR2tVO/uthkEJxEI4l0PvDyxrap7Op0WhVJrx3XdAO61JlBVkmhu176TKSLFi5gV6QBmWyEqqoAaNiSqfzvUjVyyLS6pvn1j/4KI5ggEFT25ILL5l6dTBNKSjT2fTSQSM0LfG3r1ooUrRo0q9PjBqs5bD/R4HMfcQCDaI0Ece6kBY6VX8fEw6Mih34xRZbWuYtbBUL3I96vP9YHrY+1b48ULWD8xdWqyNLMdkWn8c9tmRIQjYs7j5bbUPre9WdQZ8RC/iF4X5bzRMIUXJ1qgvJa/qGJJhFhjSlvmds0Nt8Ullf+djOA5gr5AIBDj6KqqlAXS8i/eH8rDPAkmKeIl7mFv/i7zvNmKc9c+b965A3Pc+VDvcGaDmNv+xT3VGPYaA+73SnY5KTZFBztiRxzKzRgXUcirTmLmMdqnUjJwcodIVRigg0a44KDC9sS/LEsAtGZqcgyIJsXx4jNT7jt6zfZy8ryHSQzj25wPXzJsGDFIYsDnpPjPlJmMc24deb1+kJ56NNLZcd1aJVoWc4OMHYkn9/13nYdRPQftp5ADRZULpU8H7iOzsUI4FgDtCOEfQffvZzvR1UUD3k1gjiCik1kTv6FJO4COibwL+WGxTpgMUnpJkFEDQOeUvWr+dZ80uiyM5/2hOiZ+S08NIZzXVCrFt93XtIqAm2JuZ91OPiuGLRS4t7Vr3jNIYbtTWn3MldAnIzVHndC7Xo8cOHv8Tw+fkeQUUhyq2TQYlC7WuoLJiwMAZsiOgjCBKOoBw8mreyIKa28LoRUAyi54x0oifSbi6sX+A5g5tjW0TplMJld10of2ZdAHI3Ut2ibvsIh6Yi4xoC9Vn1g9Ym7X3LBgnvWfnN/lRabpncQmUt0TDImPOK8XJIZbovnd7b1fJY1kGpoBINmdHJJ3UsVDogI2xhqjh/4rI5M85VAaaRHIOUQ8VlRkm2aspApVITM9Todyce+eBwNxqOKQ+Ies2eR6VuxpVDp7ZQ6cQtziUKyMb8lxF5RdULqtl1GAKNLqipnJYcNyS4jMLFVhZraq8DHP02YvxkSQRGbiOTxs3F9qyhvqop/Tp8X7Tw+plXGKYYYXpyB6BkMXbNN914adCuk2bEi3HmmoYcNOwtfscLQNlQg9GrGjv4t4ZTIQxSn/KsPN93unVzYeXVvROJeAZiKUEm2RIqhCRRWemYyqEANzRot9oGZK4/sLB1S2dSFXVaUsUVRfUUKutzTRWwyhd8OKcxgCsCrOYfcT3m/UgBfR98LTSEtNWWNFbUXDrYZoviHzDsMMFX1ZVF4MbGBi2tTNcksReUFEVkWMHnwKEd1eV9n4x5qpyaMKmSb+mTS6ChxNxBCVTTbcdgV6KxS5JtOdyRHhRirgsRoYPjMbQLBgQbb5zdQQZDLz1eYcuecAvE4gEOikHTH+nag29xtuWXJ0XXnj16xisTH2Y9bYYaqqovLmwFzdsOXABkZFNojIKmYDw+ZkNvzn2oqGW6rLGirzTCoDL+QUUjxy/ZP7aowxAGjdIYuRKxrwblhjxoyJVRTo7zFV/oiQNpYCaZlTNscWvvC6ilkH15bPvJIM3WM4uICJ4UXWO+9+COHyHnUVoQv/z4t/BgpHEZcqAF3ulM5m0XLnw994EW/YEJO9hCR4sLZi5uXVJ84e0QelTO3WsDqSODmqroRAx0W5Jq1k5jeGGsL19YTVzHM+7CGiLSrLRMoiHmr4hh3oo0c35iF4DdDXIogwHTF1arK0AKW0u2ItzsueAEBN+cyZ1gT3sbHfJjZjVRXOu3tJuRqGzoLidYpfoAI5L/4JL+FVojIVxlZ473/vvXcxxdKFbPieuopZv6iumH3EZtQ/SPbXAZRGxgwm3RFTaMruaSgs/DtBCjsVPiMtkycnE4fuY3/FbD8s6lVFf5qz9JO7H2xaFh3gZOKQEcFHwPo1w/ZgqMKLVyiuJ9ZvtS5pebTw59ZVzBqlKseB9FrDdpLzbq2SP3lhR+Y5AKitrJ9Gwt8kY86Ke8jw4h4G0bfaljTd0MfljIzs+gtWAkjPqWw4wAo9kggSB4Rh7o62zubz81/bkfdaW964yFh7TkQ6EBVwVFWMsey8e7o02HjqLQ/csm6oiLM8SqumvPGPgU1cEvrejaSoaOtseXwIVEXYMVwzUFtRfxaUv07MNbGBQsS/oCLfeVNG/bara25YM6X+EGJ6yHCwnxf3d7CfLRsTTyx8dHOMfN2UxiolfIOZz0H8s7z411Tpe2s596vFizObAKCusuFEUXyEQJ82bIwTWcQr97ug9ekre3fTZ3z7eWCFUhppqZ4y4/hD9rF3gfjDXpyqKrExn0sIllaX119RM6W+/uB9zD3GmF8YtgerKpz4v5LgorbO5mTrkpZHk+gLmSiZTJrWjuvWtnU2P0hE80AMY8woI7YyhRQnJycTbUvm3bNp+CvniHeXifiXoxzZnEzA9XWVjTedXzm7Mp5R1V3Nj/OE7CWCQ0AYExP7PLI1svah6D0BlNkaiImJQcS33PLALeuGqCO02eAEEz2kKmDi4Xna1e5dVOwrjJ7ePWXmpJqKhl8DfBcbU0NEUPHOe/8Ll5PT2jpbfjlp0qoo/CWcxWz3iyICnde2eP5fFz567cY8kV7+n9alzdm2zubpCn+JivwDqmDiA63hH48Sm60tb7y0tqLhWlXutBx8Fohmqw3RdIxdeXd1+cyT04gAM0UD3sFwikA4d8rMScz2dmuCM0R9X9U5Iiij/QObSFkbtFgTnKZQeO+fEvGfGM3uzNalzbfmi1p5Y+sDwecZEz3f6HzYy2QgpNPTSAuOg08mkyabzbq2zpZfhuwqnbhfevG9TAwmc5FXydZUzPxJbdmsg3ZXfuyh7zBsEvFY4N939Pvz+aqxuMO5cC1zfxjdB3Dw/tadvWyU5Jl4uB8AjtvVyEpjadSqky7at6a88Wue9UHLwaXMbCNaTHlAxZ/X1tH0yUUPtbyS399MJuNBdDYTIyLA17tSqRRXVVWZTCaqVvdR3cYXROuSlmtLeoPTROQKUb8CqrBsyo2xvzYmmAUg4bzz+WTDixdjzGmGccf0ysajCaSpf0E34i1jwLGciRrWbwQ2OCT0ubBQbCtCT4k657wX78W7HhH/Q+VwaltH868y/QPhurXwJ5PJSBpp2fTGS38D8HSsoHtmVVVyn0wm4/PTPnkKmfaO5ssAqRaVv8RwqRJrzGdh5MG6yplz8gTjBZ5+Z17LSYYMnPdeVLt3vAcZ0e3c/mDTMoDuNmyhClGocASdfELCoDNPo7vDUFQ1T4vI+kgZAlN2skfaR9xOIK2rmPWB0pLh91lrvw3m8YBCRJ6BypyR68Oz25bOX4SYrjc/I1xV9aFhAJ0GEFTxXInd9HA6nZY8Wf3AibS8p7/5kWtWt3Y0pT1wmvO+2asPvTgXybWoFojIgYg49LnQcnCgUXyrWMTamaLOmclxpHqe86ESyGxdMUTJsDGqeKKto/lL7YszK/PqCYNMkGgymTQRibm2xbIik4Iee3IBR3A85qaUQorbOub9ZdOwl89R9Z9SyEtRe4UPI+Jfra2wd9VW1k8bEFbvkOcEcFxs+ishfvmu9JOV0RIxySqRkjAxWPWWhY9eu2FHwufCIpr1I54H9NVIYk0PKSBW2OFwubZsxqm1FY03gDRDzMdBARVZ7USuUHYVdyxp+nWmO5PLE+ADpHngSGJ970kEvCPKFvQvtzxwy7rBKur54mNVVZVdtKT5KSb+BBSvMbEl0q0OrxDIuoi2t6auYtbB6X89fHTvNOA+QvdeexSIxsdzrbQt5fkIeI4xNVOTYwFQfBMP+YCS11udD8WaICDRaVvmdqTpmEImm8261iXNV5PyaeLdL0V0E0X58bug3F5b0XDVOVOTE/OXxxDCLgLSUlY2JyDCYZGMGJaNmYSXdwZEkPesHnq38/4NZsMKZSdOPHjBrhArLOiau0mJXogf+9BYhBwYJD8sANX4C8oa96+taPgO2NxjjH1fTBskXtw8Q3J6e0dTun1xZmW+rrC1S5iZzjPGBiICFWnbgVxcI1bPFDuSAxS6X9SWpO1MYikRMFK8TCgkfy8a8BCWRIhyGQpSS1WZJeF2xHDzh52M/StUnyEiMNF525piiedl80Wwl1o7mi8zxGd7lXaFgtkExgSfDCR4oLai8eN5hM/20VzR4x7Aq/bTPIQS+mQBod6OhqiaSqX4riUtrxPpAiZWw4ZV9e+54UHHzvZv8wUygj4Zt1nGI5EYFx3qK2ib4XIBm2RNReMHQ0MPGA6+wsaMVBWIyD3QiJHz9iXz/j7AcHWrcqiEdxMIXt0KAt0ffyYduuhlWkiNo22j8gbSzAqM7y2G0DtoWAr/DwAvM7EAqlu/ISPxLxA9MPKw3IYdHMbXFFIcg/Xvj+3pxDUv2aO3M8XSz4+FpLl9ybVL2pY01XqRD4n4v0MBZj6UyfyytmLmwtqK+rO2h+bKV5md8iSADowFFB7blQpvX9VY+QZRIWJWQDPZ7DU9Oxo+b3mp0mOxoFqJwh+xjefsR1Eh42vKGivqyhsXGOJrmPnoCDIqT3qVj4zsCKe3dTa3DegB69bYSgBgLK0+RlWPiy1rcfvSeS/GeypDrxMAudfNq0p4kpmh2zhbgCoTC1Rf7e3tXbYnwSr3dA+sESdzZg1ATcYErAo/YBBBNZrw97GSz5tRgSNldgblRUK3eu9gjN2HnVYPwYA0T3kKgGIy+dNEJKXer4zAA+ZcgBfVljf+tray/vCtVavHjBnDyWTSqKFJxhgj4qHKf9u1CzBireQS3C/il0OVGCazO0Sq2dKT3juxxhKUjok/i9l6njvroJqKhp+wwV1s7LuJGaJ+raj7jpONp7cvaf59BnkFyahqPMh7ArOtssaWRqGvtu3MUEUKKc4uu6YHqhvyTY2tXBqqgDcmYICaso/cvHpXL7+3GZBDCSBML0uOMmxuDIJhZzufQwylzEPnENGGes9AKKqXtnc2XxdLg7gdGeOrLZt1kLJ/JLAl45zLLWzrbK4dqgpAIZk4AEyvrD+Whf+bmRqJGASCqHtVlX4yisMrM4szmyKP018Eqq1o+HZgS76Wc70bjeepFV1HP76jRaItpEWQlpop9fUgOri9s+WHu8btFEUj55/21CHOyUOJoGRMLsz9tr2z+dKBv7P6xNkjeJj/JECfNoYPjVUUAZXrQfhmHlQzRCLBzd9TecMt1iYudC7XQ4zK6Gdtzhe2vZU/G9XlDR8yzL9WUZNnP4nblPFwMbO1CeRcz4OlpSXn35y9Zs2OnIeiARe8tOllydHG2hSUZhJwQITMEafQpQCFhs1ZAlEChU78JYs6W+bvmBFHaKea8oabAltyUeh6V3vlsjsHUK3uMKKosuE9AH/DkCnT+L2L+M3QXNNPqT/WBjRDlS5jwweIyAryUtnaNe8ZIGmAXeJi2o1kbNGznF1Wf2TC8ENMPFJUVgD6a0Bvb+uY9xcAqJvSeKEyvslsTiYQFAoV+StBv3VHR/NNO2G4fRfD9FMaJrDFw4FNjHMuXPKmjDyrq2uuG6ph9RlvWX2NMXwDiEeoClTxFwKONMZMQP8zr1Ho/MDjvxZ0Nb9ZlBfdhernoq7MmrYlzV9IkJykRGeL+gsY9rTTOo8547Tz3zHNi/uVYUOqagzxH6dPqb94R9Tj83koEd2hKjDG7huQnLUTeaj2F59S3Lak5ZaedeE0p+6rIn4FgRAf7OtrKxqvq6ls/LkJuMOYxBXEdIAXr0S0H6y5aXp5wxT0h+g7vX8Furu7eJFmfN3p9UcmrLmWiEeKRs9qTeIrRObemoqGRbXlDTeCcYthezIRwXv/uoi/3Pfwu2Ljpe3luYO1xmzAZzCb/ePCUntXxCbCO2K8NeUNdcaY+SAeEbOefLW9s/ldRFzpRd7rXK5BvbuY1FS0dTTP2RONd29jpdQ8D3N6Sfp1AK/nv3BHFGMSgMtqKmYaa8yl4j0ZY/5UXVm/aWF23h1D8cR9Ei4eWQe3PjDBPk7pIgDX7EzFNu+xY0+zHsB3zyuvv8F5dwUxNUZoLjsz1qpF6CJR70gLWCQwwfECvbFmSv3p6aXpl3blAG2PxnZHorULyhr3Dx1usiY4PvS9EgFpVEMXChGM5eDcGP4KL24DEf7gtPdHd3be8GxhirEzRaC+SS2VC6xJkIs0g+8Z6vf3ed6KhnOYuAXA6BgTnWrvaP5uCilOd6RfAvDSVnD4Whxm2L0zolyoBxyHV5pCipaUP3E1m+DjIl5BWONV3r+wo+WuoYbTClBtecNfErbkjJzLvcwl7pTW+zJv7OINvFlYXVPeUAei/2Hik1S9xHzNNAADHiZsSeDC3u+3drZcvjXKV/zrevKRUFhF47esTXw953pDAgVbwtbhiZlEfYco/mNRZ8vSnQuXt0Uje9G+JRtLHw5scJhz4dNEpqy147q1g72b/LufXpY801p7E0D7MTGcd//b3tnytcKWYWExrP987bm55Vtp9WnG1pQ3/twa+ylRD1VdLU7fv/CvLXcNRu6d/3pNReM3rLHpqBosF7R1tNw2NGLwwfO4qqp7OPIE9R8NTMlvXOR5zVaGOMSwJe/dY73LSyqyy67p+TeFcZSvmq993ixhY0/x4pW2QlGjCh/YwIQSfqR9SfPv58yZE8yde5DHLhpBfu+rK2eex8ACy5ZDH/6yvaP5sqG+0+kVM6ZamJtBPJ6JIN5FF+Me7GHfiqyUg4M5oNTe2fxpJ+5nTAZEtC9bvr52yszp+QEGDN7oXCTeu0gziGp2R+sl338cP368JpNJQwITtR51G+gyIlUlIhyQmOBH/bs3d92yxAgAo6Nzs33UFamOSyHFq1ZFAmi7jxBaplu27MUDQNtgTqjvQi5LVhiyN0bGy3DiftQX1eylxvtWNOA+YW4gxe0dzZ8NxV0ZGTH2JYNMdUXDOXlaUmyP3VE3PCzQJ2La1bOTU5OleQQWdg9IxQNYo9DtAvgiQDN6xNG/XRHgDT9iE0AbhkaehZ7dGHr2o6+Uzo4nlF7Rnp4HooLh1i/Wqqoqm8lkfF1lw4lk7A0MOjAidgh/0d7R8kVEFE17rfG+VQ24T84DAC/saP4PL+HV0UQd9jXgzPQpjVXbqU5rXkKUVNpj9tlj1/pouCGJXWdgnJyZHKG7bNDlvd9ExFutoCoi3mYAr93Z+acVqX/fGJsmkTRdXXNDQP9OxNgG/FCJwCIeTHhsd112efTV/sHqE4j02JhV5P6Fj964PDrDW14U+THQ6afUHwuhm5h4IhHDSfibto6WT8bY873aeN/KBpwPp1UV1NbR8mkn4W8iqCXGWkPX11bUnzVYi8mzLvISSiQ6TRfEvKa7vO6puifqKUvueCJiUfUKSJ5kL0/KRgr1Ip6Yj6kun/muNNKyowiz3cWGkkHGR0MifLSIl0j5AL7vmaP/Fqh6EMGpVAHQmHSfdnW/ou6AqTEmGBZB49EaFwZpW2HzeeX17zAB30TGHEHE8D784+jD/CcKayV7+yF/Kxtw3iMAUPSWvnqZ9/662BPvR2RurKtsPHNrRpwPo5l0saq+TEQg6DlVVSm7q0WsqqpUvpXxAYb9IxESho2xbE2eZC9PymZtYGNOq9GGcGPtqfVnZbPpIfe1sZuqz2mk5ZzKhgMg9hY2XKYAG2YT2OiZbfy8gQ2MMYEV8WI4+EbNlPovZbNZF1d1d9qI8/O9BJxDIHjv1gnbewHowDnkvPGeUzFrspC5nYnfEdEghfN6hr/60YJ2oL4VDrjB22PxsmXXyKmj37mgx+JYY+xkVQxXxYVHHjT5/rsevG1ZVVWVXbZsmRR6nd+9dPXGow4+4Xhmc4qI3zdwq258+uW/vZFCirPI6o4bb5XNZq9xtVPqL2bia0E0gonJi/+lqFwpKoCIFZXXVfydIvJ7hYw1bCcqtJSY33PkxMkP3vnAbc8PfN5/Zuuoqjx5YAnsLZbNaQTAizyqiksB+ZuIhKq60ov8VUWuE8gThkx5NBZmaiZNnLzh1oXz708mk6a7u3snoZtZrZlSfwgIKcN2hIrc397R9FMAVPge+gpWU5NHGeVbmfkoIoaIz4xa7y65fentLiLOz+pb5mC/Pew3LSmkKNOdydGqVbPF+0zEC0XjmM2N1WUzzhjoifuRV3SbSDTcoCLVOzsdlO9D1k6pv5iM+ROIhkU5mftWe0fzZe0dzX9s72i+eI3xJ/F+q09t62xpbOts/lHC0/mi7jYmBgj7Ecwt506ZMX1HEGY7qYAQebLKhgOGwd5omCs0Mt4lEuh57Z1NC1qXNH+3raP5/Kkd7zi9vbP5orbO5m+1dzRfJuK/qARRKAzZ79dWNHw1rv7vsCfOo6/YUJVhM04hUFB7XKvggcZ77pSZk1iCW5n56HjU8M8Yu/KSTHcmF+XSe25Pt2jAg6KiUtz6dGsvVq68xPvwemYGCOPZ2ptqy5OnxeGeAYD58SRPCLlfRF5hZihH7aT813YcAVSfJObrAAwncIwAavlG1BuO2EMWL85sam1t7Y3+LGUXdDW/iRWr3u/FX8tkAMLogIMbq6c0XvjPMuK8AsL0UxomJJRvMcZMjQYRZElIctGiB1peqapK5Wl8YzUDpaqqKptMJk1bZ/OPVPwHodik0eTQ/9ZMafhmlH7smGRrvnUnoBpmA+dcLxiLCr+Wv2zqKmYdbBk3E/M7o1A7vIlW7D+7tbV1j2SURBHIsfPTOXVH1ZVg7JgmZnuxqAcUr0Pcha1L53fmDS7/0msq6zMJU/qBMOxZDjEnt3Vd9+pQJ1/6PG95wwwi/gMIpVFY5/6rraPlO1tBKG1WYMk/QwopXlz+xFxjg4+K9wpCr3id3b60+fodnLoaUtgc0bWamw3zKQDgVB7UTfzehY9euzym0fXbGJagqqoqk81mXfWUGReysdcSaBQTwYt8t62j6at51NwQ8lACoBeUzRkemrVdgS15Z871dq7wo86IKuL9+1N92uzx7OR2NlwWU/8uCPy6+gVdCzbuyJRS0QPvBeF0nyceu/9M78ObIs9GByjbG2sqZpyU98R9ShAeC0UciM14GD+tMLQbzBiy2ayrqWj8IIj/BEJp7Bm+sg3j7aueD8BTUxppbetsmeN97tdsDKkiYQxfVz2lsWF3eeI+tFNF8ghiexsznwIQnPq/cG940TaMd2BBSPPPs3Dp/Fvh3fuhulJVYdh8paai8acFxkRDoVTyZu1UgI+OiO5xTzy8YPLGO70sOZq9m2/yxuvdwp5S1/hWN963qQH35cTc2npl7yjjZ4qPckwmmkgwN59z8qzJmUzGP/tspAQBG9zlvFtnjAWUzh8KKqsfO9zwEQb9lggBQPDqP9/eOe97O4gN1jywo61j3hwXup8ZNiwKawz9sbq84UOF4f8ued6pyaMMgtuZ6YSoYBVmvWx6X+vDmTdSqT5a3qFUjl1VVZVtWzp/EUgvEtXlqgJr7H/Uljf8rmB6aNAz6BV11ljjvRMlXQQAzz47htNIS0XFrFHWmHmGbVVMM3zvhhKqz2Yz61NvceN9m4bQW6o9VJ84ewSX+hsM2xoVgao+Hap/952d856Mw1NfW9nYGpiSmjDseZpodVlrR+u2APR9Qws15Q2fIeKfACAmIhH5Yltn8492AdgfaxWnpaa88XvW2su990IEqOjH2zqbf7MzP7sPK1xZf6xRcxMzvSPK0d29yu597YszK3c2h+yfvW2oZMM3EdGBzAzv3bW0ctWlrU9H+f5WDI0AaFnZnGB/XntfEJRUhGHv00SmrLLjqPVppOU9p39kZG+46c/GmBoA8N4vDgQXLuiK9J7Sb3HjfRt74M0LWwsfvXZDT0+i3ou7m6Lq9FEB8S3nTpk5KZ9bqspCjTigjvDYd8o2KFz6jbey4YvGmJ/FSgYkKpcXGK/fyT5kjDBLcXtn85e9D78bE9WBmObWlNdftqPV3v7hjRknGeU7mPLG67NO7cW7YryFnnhhV8sSVrlQVV9WERhjZ+uYfVsibuc84d9mMi4EAOMSa48H6ARVhRLuae24bm3eeHvcxowxtiaujj/qA33/gq7mNwsI9FA04LdJOJ195JrVPT0bL/bi7iViEJtjLOkNtWWzDopQjeZu58NN1ljDwDnbGxesrWi43MD8QEWVCCTiP9vW0fz93TPNFBlxxE/d8lXn3deZDEckevaq2vLGL2QyGR8bwJCA/udOaTiFyC5g4sOi1pZftCGB993Z+acVu8MY8uH9HZ0tS4nkfBV9TkVhTPDeYZvC+dUnzh6RHmDEea1jeEwzxpZGPXJ6AIj0rHJu03zLQa1CId53g/xFix6IVBsyyPi3y+k1RQMGsshqCim+5vVfbTrykGNvUcFpTOYwYjpQScqPOPDYG/ftlBd6J9KF1iQmevH21NOPvWb58uVUsayCupGkZHJ83ni/ymS+G3N2eafuEws75/1iWxS1u/DMSCaT5taF87JHHnzcaiKqi4zY1B454Vj6bdNVd8eAE+RVFMePH8/J7iRlkUVVVZW9/fbb/blTZpQHbG5hookR0N8tWG83zLj//j+vSSHFV+Pq3fLM3d3dmkwmzS3t81+bdNDxbURaDaL9mfgYsJ9y9PCTbvntyit7qqqqbMWyChpTOoa7Xu3ykyac8EVr7WTnwtXM5ktHlU7slWHD5htjz48ob/wTTumChZ0tz+2mC7KYA++1OXFefW9qciyJudlwcGYs8dE6cn140eoR5vKSoORbYZhbKeSm5BUM86umvP7LxtjvRhOC2iPqP9reOa/pn3iw+r1+eeOlRPQLQA2zIe/9D9o6m/9ze9xf506ZUW7Z3sJEB1JkvE09pYmPZrPX9Pyzcsj8XtRW1h8ONTcYNqdE+au7uyeXuDj7yDWr83/3Pad/ZGRPuOnRRFByeOh65vWUvjp72MYJzcbYD6gKBPqkh16waEnzU29H4y0a8HYKW+eWX7KfJX+TYXMmERD6MAPCEgPzAwAQ9bdC9QYQnhETPMHOf5INp1QBVV0H8R9uWzrvht3Zo93WO8z3XWsr6i8hsr+BimW27CT3M1655nIeM3KMwhzsGEad37ioa/Lj1ZVPlDPoJgYfSETw6n836lA3J8/l9c/MIfuMuGzWQWT0emZzGgA4CR8UxX8QYRQDU6B0LgjVhi05H/6JiBLG2HpVgYo+rRJe0LY088Tb1XiLBrw9GCEyvu7k5DgkggVsTIWIh4o6UMQjZjjiZvfiVKGvMfFBRAyoXyviZ7R1zm/7FxjvlkZR0dBIZP4AaCIS/ZLHVHUMEU2ME+gcqT6sREcY5nEEhvd+bltn0yfiK+xfAjfsy7/L37tfQMNbiHk6VCEijoisMRaqCi8uLzIWqzMDAnkJonVtnS2Pv52Nt2jAQwHyV7zv4GEovYWITlKVvsKQQgUaqdcREWJy+fXweknb0pZbysrKgq6uLod/IeFfVVXKZLNpV1NeP5PZzFXFMGY2kTRhP3bCsIn4mQFR1T/0DH/l49Oy0wQpIB5w/5dpX+VlRUtLht9GxFMl3uOYwJ+I+pQoFSAhwmoHfe+iJc33vd2Nt2jAg6yysjlBpALf8LkgSPzYubDPAw9g7hdDbLz4l3qHl5wwfvzGdbtI4Lazi5PJJGUy86W2vOEJNsHRIs5F3YZ+3g9VCEiJyZBXP31hR8tdqiCif/2IXX8bq+HXgUlcGu/xFtIzqvDWWuNcuKi9s6UabwOQRtGAd1M+XFve8Gdrg/eHzm2VfC7vka0JOAzdpxYubb46lUox0sCCslfNpk2r6NDcOhq5/0jeFI6h3nBj377nXO9230HClmxmVCXBcC0NVum6Nw+UTRMf9QMUGAmA1kxpfL8xnIm9Lm2drAfeGmu8hFl409izKbeudxR8yavDdeT+K3jdPqU8OhxJOddLzuciPmaT0C2f7U2UBCXaG/ZSzo3s+13578l/38j1OVltS3Td8F7Z980D5Y3RvTJp0yqa3D3Z3V/+xKmGKEtEwzWWCdwqvRAZEvWvkTenRnj0PY+nuWjAe5gaRM3U5Fh484hhc7Co1+1IUCoRkYi+QcDDSroPQAFBbcSNAwaISfO99y1J4XTA+6BY+FYHcvYRKRQegCMgzOtCqcITYQMUJxDzRFXRwd5xJPciLwO0BgSl6JsMCCZ+Vur7cKoycHhBY36bmMiOBii+xp+KJDJMEhAECg+KYmWAQgUmMNFEVR0CQQOTV3/uwo7BGUbfDssW7XSb/peAtLIkhiv50kgSZTAiRoVhHmdMUL2FY1BsgwZCh3Sf0sBxJSo0wcK/SHA+hBcXKQwPmjwrApuYSMQTt/5MW9OD3Hxwirb4ypafkrZ4lPwnivSSvITbuxz7WcKIQMLDi+ezaMCDrKiYc6gf8foys/ZFIh4r6pQG8cBe/FoR/3p0GFU18k4KaOxRt04k3+dxtZ/QLnZu2wqBJe+JCRoC5CJCTnJQPYqZx287HC1058iFPvdXgDZCtQSgIGLDJFKCgYKJlFWJicB9UULe25IqgbRPm5OI+79GWhhNKFRJYy9MJBAtAbQXoLHEfMQgHjhy5d47ML0K7C6a32IIjbd6JbqmYuZPEib4bOh6HYjstnJgw5bVuUusrr+hZ0xpQMtLXEmwWkuDUgWA5SXDFQDG926k4SXD9fn4v/M/I/91AMDzAA6P/rO3dyOVxF/L/3dp6SgFnsLIF0bq8nHLZdq0abJgwaumq2tuWFPeUMfMt6tAQLp1XLSqszZhQ5+7ur2j5dP9bjOCM6YQMY8sX76c1q8/JsIlr3mBcdTmP2bkCyO3MKJ1uXVb/L6RiejvPVs6pu/v7/Nmrzlm/xL/ctAz3vnwL8x8hFcvBOKt5ezGGBLvunuGl5Rns/82kvuiB95bVkyYRmz9laGTWWzs/iIujGVQ8oarUMp7IQbhxAVLF1y7Wx5gGXYEbwwAolCiTmqtqWi4MRGUXBy63nAL2RaFYzbW+XAjPF0JQPtlW6LKbhpAv5ZStv8XPb1bi4Q+vSwtVeVJP4wCp4CQQhRQIuV+Zw8hQA0b48X/Ki9Q/nbPf4EiFnpwjHQqxb+99qqVRx98wuMA3m1NMCIOI4mIiNn8//bOPjiu8jrjzznvvcJ2zEcCEQklhJBvuxknI2tVEkC1sbQiJUOd6bV2DQFmMhgS8kEGKEwa2Aia6UwJbRnApA5/xFC0km8TKIF4d2UpUUIatJJjYFJKMh4gJBmglE9TDN77ntM/7l1LEBywLa2NfX4z/s+jke+9j9/3ec97zkOBC5iIWdQLu+Ck97978Xs+uODY6tZntvps+JzujJttwZ9x/AQfPmbDeKJ6chgccpw2z4uaLz0IWVVeAuSC2tTQWCbelv6O3d3dwfrfrvd9SwufCDm4nYk+CigFLmSXzjoiUPqMHQfsnONGsmPd03LY1WsePx1rH1yr9oXaFnq3ykm9uVVLCMElqroMhKMAeFL9gwIxlI4JAneuF4FzDuJ9JWnI5zdtyTpkWrtapEFgS8454pB5jUtJ9VwiOibzxArgR4n4q0anNky2vm+2xNn6Ln25/gjENxG5I9MY0sZDINykSp8i4BQoFirpyyD+b2hyc62+4ZbZzzs2AR9UIgaA7o7iUW1Mxzpq7Nj+Ih4bfzB+EQDyueL5AF1DRIcyEUTl117liyP1obF9EKC18yPPJmH+O5OjRHxcqw+ueu2/qZVXVAEg39l/BVFQAqljdkh8cid5viCr76LnxLPaaYd/R5gk2+6+P/7D680LM0zAuy3imbm/Mw+7mnlHvUv7P0nsbnHM75e0s+FliF5cmSqvnTlUrxW/b3d3d9De3q7bHglOUtaxwLVx4hvXV+vlr65ZsyZYty4dDNdK8facuLKd/fzrHLuCqE/r0JBvVyfKlwGQ5gSU14rUPC+soR+zMMGjOWSuhBJngqY4jn0z9bA2NfyfKn6ZF7+RiaHQNgrcjb254k2nd6xZAAxIhKglZw/t7e0ax7FXUNissWp6OUPT1MCWZTm7GLHvPbH/k+znjzvmgqpAoS8k4s+tTpQvTQVb4qz5oxnmzpjxjO0LNAHPFjpDzDozcTAT8e/+1x96hpfkWmZmEdGQgwsabtvGFR3RcTHeZMTp3Oy5fIt3LJrlLZ9N3lWY6CNZaNVDEPSNTA6tn34Wr9qZZMHaA2JbZhNwy2j2027evK5RrZcvES/nEfCiVw9md0rgwtG+ruJJ2dgbboWNUTfdRUUtsk3NTKVoUdSW7ypew8TrARxKzPDe362ellcny7/YyxlhJmB7BHM1LA8UIXLVyfLNnnEaFFtTBdEHoFTp7Sycl+UN77w8MWcCnnFHWaHSCu8dx7H/dGf0rm0Lwzsdh5dARZkYiffXPS2Hrqxuvu1x87Um4P16mx0jDRIfuXfw54lvnCoqY8wMhc4PXLAunyv8Qzakbm59sUzX+4moMad+F+kw++UdUc5TOMbs8iIJAHpFvb+wVh+8KE1VKLGJ1wS839OcyLhpc/zYc9Q4PZHkRmbHPo3gvLw3Vxz+q4+tfvtc+mKnPN0ZpJrMXX23RDFin+8srApdWKX0cgag2Nogf1ql2WaZhmuLfR0m4LeUL7733nh7baL8Je+TrwDUEPUIOFjlF2CkN7dqSfMQbC58cbMorDMaDGbT7yI7cMrnil8n5jIBRzAzvPgRgJaNTgz/JELkMttgftcE/Bb1xVHkapND14v4lQo8oVAQqIMpGO3LrY6mt5Wz54vVKemcz7b63JH5XPFWx+5bAJg49bv0zLOfqdRv+/3BNq/ZBHyg+uI49hEiNzI1vDER360qv8jGPh0JwoZ8Z2EgHcreunrxXtV349j3dEYfD+FHHQdnpZdX8LJq6ncrW9NoT/O7JuADZ0ud+d3RyeHfQKnPS7KeyUFVxbnwyonKb36wvKtw9Gz5YlUlmvUduSJNMiz8DVM4QkxL0vRAfSxRf3plorz21fnBhgn4APPFESJXqd/2QrVePtdL4zICiagHU3BGqPzjnq5CV/PvzZovFtXZuJwBEPpyq69yxBsIOIqYofA/82j85abJ4dFm6Lb5XRPwAb0Sp7OxSlytD/0jRD+rwBOqAib6qFPe1NdVOCtG7Evpqrf3IuY9X4yblzO6u6OFvZ3F25jdFYCCmCHe3+y3u9NG6vEj5ndNwAcRlNaBo8hVpso/1ESXqcovKb1HvYDAt+Zzxa8PAAKQ7o0vJtCrasJ7dljV/6F524NK4IKiqABKr0jiL6rUB88beeDf/s/8rgn44N1SR5GrbS4/1NDtvYkkGxwHLKriOPhWvnP1+r7cmYftiS8mor3ZN1OplIqyr7P/1JDcGJP7lEKh0N97+DOqU+XrmvVd87sm4IPeF49O3vF0rV7uT8RfTUTsxcM5d7ZCR/OdhcW7Wy+mPaz9Nv3uwMCA9CwtfhHs7gTRnxERROQeD10+Uh+udaM7sPquCdjY6YtLDIBr9cErE/FnEvCMqoCZlhLx6Km5/jOybSrt7uEWQXm3mhH+Iprf21W8IXB8I4AFRATRxr+G/oV8MxVwHOOJvTkTsIHp0HEAGkWR2zQ5POgVp6rqryh9XUcHcN/v6Sx8A4AA0NIbXPrwnnR3FsedzQifWP3ebRrcHXBwoUIBJS/i/7YyMXTBXZvvesn8rgnYwBtc+ogiNzI5eN8OkhWJT/7DOQdAKXTh1flc8Xvdi6KFA9kh2K5+kGPyM0ajyxtdzsiiSk/2IcaI3DIVART/IyL91Xr5muYgA/O7JmDjTfrisYmhJ0+c/NBnvU/+nohZJFHHwTnzFoYb80v6j5/hi/8IIa9vrhkh/U8jnyueT+Q2EuGEdLaXPoCG9NSmyt9vbq3N75qAjd30xQMY0Eq9fIVIcqYCz6e+mE+ieW4sn+s/eVeXPnhmZhFpsKtmhGhR1JbvLP4TE38HwNuYHLz4mHY0VlS2DD3Q3FrbGzEBG3vmi5E2QwwPKtEKhd5PIBDofQTemM/1f276AsW0L/avCfl7Pb/blzvz2G0Lw7tcEHwty1xJvCR/V5kYXFW5L36qVNo5r8qATaU0sHcTL8bHx5PujuJR81i/64Lwr8V7gKAKubo6MfTN7BCsDYB/4dFgBQiVwIVoJI1/Pvz45NKHH347n3DCsxLHse/pKHSx41sd8wcVChF9SqFravXyHek3UrL+XROwgbmZsUx9ueJVIP4GoEi3vcnt1JacX7knfgoAenKF5Q486pxD0kiurU6VL2n+nL6uVQUg+A6BDiciePH3UUPPqWwZeiC7fWV+1wRszKEFUgDa07Wq4DS4kYjeQcRIJPkvePc15R2vENFFjoKVCkDFP+SZLhbo5gD0VQZdDoCYHESSHzR0+5rRyTuetnlVJmCjdX25HMex7+1a3cGK7zG7P1cViCqIACaGl1SL2axqiOiTzvHRmgU9icq1h703uax5mm3iNQEb+8AX5ztWvpvc/FtAWC4qoDREWKcjO1PJEjGpqifCy6J6ea1evsH8rp1CG9h3w/O6u0tBdfPtjwvzF1T0laZoX523SwSAVCQJXOBU5Y5avXzDLoarGyZgo1W0tz+oAIhEFjPzvGZECXYRd6+qUODotJS0qJXBa4YJ2MCuJujAL2V2pArZdYeSkqqAlBYdgeMXZiuvWSoTsLHvDzYofBOrKcFOQUzAxn64BBM/rNA/OYJHQULEqopH2sZf2mZSNgEb2OeND5oFENa8T14k5uzU+XUH+SgRkarenU35YPPAJmBj3yKlUomrE8OPqmJt6NoYQKKqzdtUCqgA2nDOBQ3feELbsD7L37XT57cgzh4BDrRyEoASv/O4J+8JE1kchocsJkpLR0RETEzOhU5UnmfyZ9fuHd6SNSrY6gu7yGHsP+9VFy2K2o59m/syMRch+j4QmIDnoPTThsq/jE4NbUm7lqz2awI29ksRA2kL4vOP8DEeLkwa4TPj969/DjtbDk28hrFf5xfhdaZNluY4VNwwjFldjXeK1nZdhmEYhmEYhmEYhmEYhmEYhmEYhmEYhmEYhjHb/D84SPJoqmwsrwAAAABJRU5ErkJggg=="/><text x="800" y="498" text-anchor="middle" class="m ix">ONE LOCAL INDEX</text><text x="800" y="526" text-anchor="middle" class="s il">graph · git history · docs</text><text x="800" y="548" text-anchor="middle" class="s il">decisions · health</text><text x="800" y="576" text-anchor="middle" class="m it">NO API KEY · RUNS LOCALLY</text><text x="1128" y="252" class="m lab">WHAT YOU GET</text><line x1="1128" y1="268" x2="1528" y2="268" stroke="rgba(88,67,108,0.16)"/><path d="M 960 431.3 C 1060 431.3, 1008 313, 1102 313" fill="none" stroke="rgba(88,67,108,0.32)" stroke-width="1.4"/><path d="M 1102 308 L 1113 313 L 1102 318 Z" fill="#F59520"/><text x="1128" y="318" class="m gh">UNDERSTAND THE CODE</text><text x="1128" y="351" class="s gi"><tspan>Cited answers</tspan><tspan class="sep" dx="7">·</tspan><tspan dx="7">What calls what</tspan></text><text x="1128" y="378" class="s gi"><tspan>Why it was built this way</tspan><tspan class="sep" dx="7">·</tspan><tspan dx="7">Who knows it</tspan></text><line x1="1128" y1="418" x2="1528" y2="418" stroke="rgba(88,67,108,0.16)"/><path d="M 960 475.7 C 1060 475.7, 1008 461, 1102 461" fill="none" stroke="rgba(88,67,108,0.32)" stroke-width="1.4"/><path d="M 1102 456 L 1113 461 L 1102 466 Z" fill="#F59520"/><text x="1128" y="466" class="m gh">CHANGE IT SAFELY</text><text x="1128" y="499" class="s gi"><tspan>What breaks</tspan><tspan class="sep" dx="7">·</tspan><tspan dx="7">Which tests to run</tspan></text><text x="1128" y="526" class="s gi"><tspan>Change risk</tspan><tspan class="sep" dx="7">·</tspan><tspan dx="7">Cross-repo breaking changes</tspan></text><line x1="1128" y1="566" x2="1528" y2="566" stroke="rgba(88,67,108,0.16)"/><path d="M 960 520.1 C 1060 520.1, 1008 609, 1102 609" fill="none" stroke="rgba(88,67,108,0.32)" stroke-width="1.4"/><path d="M 1102 604 L 1113 609 L 1102 614 Z" fill="#F59520"/><text x="1128" y="614" class="m gh">IMPROVE IT</text><text x="1128" y="647" class="s gi"><tspan>Fix first</tspan><tspan class="sep" dx="7">·</tspan><tspan dx="7">Code health</tspan><tspan class="sep" dx="7">·</tspan><tspan dx="7">Dead code</tspan></text><text x="1128" y="674" class="s gi"><tspan>Refactoring plans your agent can follow</tspan></text><line x1="72" y1="754" x2="1528" y2="754" stroke="rgba(88,67,108,0.16)"/><text x="72" y="790" class="m lab">WHERE IT SHOWS UP</text><text x="332" y="791" class="s wh"><tspan>Coding agents (MCP)</tspan><tspan class="sep" dx="14">·</tspan><tspan dx="14">Editor</tspan><tspan class="sep" dx="14">·</tspan><tspan dx="14">CI and pull requests</tspan><tspan class="sep" dx="14">·</tspan><tspan dx="14">Local dashboard</tspan><tspan class="sep" dx="14">·</tspan><tspan dx="14">Multi-repo workspaces</tspan></text><text x="72" y="862" class="m ft">github.com/repowise-dev/repowise</text><text x="1528" y="862" text-anchor="end" class="m ft">docs.repowise.dev</text></svg>\ No newline at end of filediff --git a/README.md b/README.mdindex 8b7dd633f..00d123e88 100644--- a/README.md+++ b/README.md@@ -6,40 +6,35 @@<h1 align="center">Understand your codebase without paying your agent to rediscover it.</h1>-<p align="center">Repowise indexes your code, dependency graph, git history, tests,<br />-documentation, and decisions once, then gives agents and developers cited answers,<br />-change impact, and concrete code-health fixes.</p>+<p align="center">Repowise indexes your code, call graph, git history, tests, docs and design<br />+decisions once, on your machine. Then you and your coding agent ask it things:<br />+what calls this, what breaks if I change it, what is dead, what to fix first, and why it was built this way.</p><p align="center">- <a href="https://www.repowise.dev"><img src="https://img.shields.io/badge/LIVE_DEMO-repowise.dev-F59520?style=for-the-badge&labelColor=0A0A0A" alt="Open the live Repowise demo" /></a>+ <a href="#quickstart"><img src="https://img.shields.io/badge/SET_UP_WITH_YOUR_AGENT-one_line-F59520?style=for-the-badge&labelColor=0A0A0A" alt="Set up Repowise with your coding agent" /></a>+ <a href="https://www.repowise.dev"><img src="https://img.shields.io/badge/LIVE_DEMO-repowise.dev-0A0A0A?style=for-the-badge&labelColor=F59520" alt="Open the live Repowise demo" /></a></p>-<img src=".github/assets/product-map-dark.png" alt="Repowise connects code and dependency data, git history, tests and contracts, documentation, and architectural decisions in one continuously updated local index that gives developers and AI agents cited understanding, change impact, and concrete code-health improvements across editors, pull requests, dashboards, and multi-repository workspaces" width="100%" />-<table align="center"><tr><td align="center" width="250"><h2>−31.6%</h2></td>-<td align="center" width="250"><h2>97.2%</h2></td><td align="center" width="250"><h2>2.3×</h2></td>+<td align="center" width="250"><h2>58,924</h2></td></tr><tr>-<td align="center" valign="top"><sub><strong>less agent output</strong><br />3.8 vs 7.2 tool calls<br /><em>n=43 · p<0.0001</em></sub></td>-<td align="center" valign="top"><sub><strong>smaller context payload</strong><br />393 vs 13,984 tokens<br /><em>30 Flask commits</em></sub></td>-<td align="center" valign="top"><sub><strong>more defects surfaced</strong><br />same 20%-of-lines budget<br /><em>2,770 files · p=0.003</em></sub></td>+<td align="center" valign="top"><sub><strong>less agent output</strong><br />3.8 vs 7.2 tool calls<br /><em>43 questions · p<0.0001</em></sub></td>+<td align="center" valign="top"><sub><strong>more defects surfaced than CodeScene</strong><br />same 20% review budget<br /><em>2,770 files · p=0.003</em></sub></td>+<td align="center" valign="top"><sub><strong>files indexed in one run</strong><br />dotnet/runtime, one laptop<br /><em>120 min · 11.7 GiB peak</em></sub></td></tr></table>-<p align="center"><sub><strong>Graph accuracy leader at matched coverage.</strong><br />-No tool finding as much was more precise in all 7 compiler-graded cells.<br />-<em>5 tools · 37,853 oracle edges</em></sub></p>--<p align="center"><sub><strong>Zero LLM calls</strong> for graph, risk, health, tests,-dead code, and PR review. Generated prose is optional. Every benchmark publishes its-sample, method, limitations, and losing rows.</sub></p>+<p align="center"><sub><strong>Call graph checked against the compiler:</strong> 0.976 to 0.995 precision on Go and TypeScript,<br />+and no tool that finds as much of the graph gets more of it right, in 7 of 7 comparisons.<br />+<em>5 tools · 37,853 compiler edges · every losing row published</em></sub></p><p align="center"><sub>-Free and self-hosted · core analysis stays on your infrastructure · no API key needed ·-AGPL-3.0 or commercial+Graph, risk, health, tests and dead code make <strong>zero LLM calls</strong> · no API key needed ·+free and self-hosted · AGPL-3.0 or commercial</sub></p><p align="center">@@ -50,7 +45,8 @@ AGPL-3.0 or commercial</p><p align="center">- <a href="#why-repowise"><strong>Why Repowise</strong></a> ·+ <a href="#what-it-does"><strong>What it does</strong></a> ·+ <a href="#quickstart"><strong>Quickstart</strong></a> ·<a href="#your-agent-stops-guessing"><strong>Agents</strong></a> ·<a href="#know-whats-dangerous-before-you-merge"><strong>Changes</strong></a> ·<a href="#code-health"><strong>Code health</strong></a> ·@@ -64,53 +60,130 @@ AGPL-3.0 or commercial----<a id="why-repowise"></a>--## One index. Three ways to use it.--| **Understand the code** | **Change it safely** | **Improve it continuously** |-|---|---|---|-| Ask cited questions · explore architecture and execution flows · read always-current docs · recover the decisions behind the code | See symbol-level blast radius · run only the tests a diff exercises · catch missing companion files · detect breaking contracts before merge | Find defect-prone files · separate maintainability from performance risk · remove dead code · hand concrete, graph-aware refactoring plans to an agent |+Repowise is an ambitious project. We want every engineer, and every agent working+beside them, to understand a codebase the way the person who has maintained it for+five years does: what calls what, what tends to break, what nobody uses anymore, and+why it was built this way. Cutting tokens was never the goal. It happens anyway,+because an agent that can ask the index stops searching, opening and re-reading+files to find out. Measured against the other context tools on the same agent+tasks, it is also the largest saving.-These are not disconnected scanners. The graph locates what git history flags; code-health measures it; tests show what guards it; decisions explain why it exists; and-the same evidence reaches your agent, editor, pull request, local dashboard, and-cross-repository system map.--<div align="center">-<img src=".github/assets/demo.gif" alt="The Repowise dashboard running locally: health scores, the code-health map, a graph-aware refactoring plan, change coupling, and the generated documentation" width="100%" />-<p><sub>A dashboard tour recorded on this repository. The same local index powers the UI,-MCP tools, editor views, and PR analysis. No API key and nothing uploaded.</sub></p>-</div>+<a id="quickstart"></a>-### Pick your front door+## Set it up in one line-| If you care about… | Start here |-|---|---|-| **A coding agent that understands the repository** | Repowise finds the right files, returns task-shaped context in fewer calls, and proactively supplies decisions and risk. [For agents ↓](#your-agent-stops-guessing) |-| **Safer pull requests and faster test feedback** | Get change risk, symbol-level callers, co-change partners, and a measured or graph-inferred test run list before merge. [Change intelligence ↓](#know-whats-dangerous-before-you-merge) |-| **Finding and fixing the code most likely to hurt you** | A defect-validated 1–10 health score across defect risk, maintainability, and performance, followed by the concrete refactoring plan. [Code health ↓](#code-health) |-| **Understanding an estate, not one repository** | Match backend and frontend contracts, catch breaking providers, map downstream services, enforce architecture rules, and query every repo through one MCP endpoint. [Workspaces ↓](#past-one-repo) |-| **Rolling this out across an engineering organization** | Keep analysis on your infrastructure, give agents and reviewers the same evidence, and add commercial licensing, security controls, custom extensions, and SLA-backed support. [Teams and enterprise ↓](#for-teams-and-enterprises) |+Open Claude Code, Codex, Cursor or any MCP-capable agent in your repository and paste:-<a id="quickstart"></a>+```text+Read https://docs.repowise.dev/setup.md and set up Repowise in this repository.+```-## Start in minutes (no API key)+The agent installs Repowise, indexes the repo with no API key, wires itself to the+index, and asks you before anything costs money. Prefer to do it yourself:```bash-pip install repowise+uv tool install repowise # or: pipx install repowise / pip install repowisecd /path/to/your/repo-repowise init --no-prose -y-repowise serve+repowise init --yes --no-prose # graph, git, health, dead code, docs. No key, no spend.+repowise serve # local dashboard + MCP server```-That builds the graph, git, decisions, health, dead-code and structural-wiki layers-locally. Connect Claude Code, Codex, Cursor or any MCP host, or open the dashboard.-`init` wires Claude Code automatically. Then ask your agent: *"Use Repowise+`init` wires Claude Code automatically. Then ask your agent *"Use Repowise`get_overview` to summarize this repository"* or *"What breaks if I change`src/auth.py`?"*+[Full setup, every agent, optional model-written docs →](docs/start/QUICKSTART.md)-[Full setup, every agent, and optional model-written prose →](docs/start/QUICKSTART.md)+---++<a id="why-repowise"></a>+<a id="what-it-does"></a>++## What it does++One index, three ways to use it. Find the question you came with; each one links to+the page that answers it.++<picture>+ <source media="(prefers-color-scheme: dark)" srcset=".github/assets/product-map-dark.svg" />+ <img src=".github/assets/product-map.svg" alt="Repowise reads your code, git history, tests and coverage, docs and decisions, and the contracts between repositories into one local index. From it you and your coding agent get answers to understand the code, change it safely and improve it, in coding agents, the editor, CI and pull requests, a local dashboard and multi-repo workspaces." width="100%" />+</picture>++### Understand the code++| You ask | Repowise gives you |+|---|---|+| *How does checkout work in this repo?* | A cited answer built from the call graph and the generated docs, in one call. [Search and answers](docs/layers/WIKI.md) |+| *What calls this function, and what does it call?* | A call graph across 26 parsed languages, every edge stamped with how it was resolved and how far to trust it, plus traced execution flows from each entry point. [The graph](docs/layers/GRAPH.md) |+| *Can I get docs for this codebase?* | A wiki for every module and file, rendered from the code's structure with no key, or written by a model when you choose. It updates incrementally after each commit. [Docs](docs/layers/WIKI.md) |+| *Which of our docs are wrong?* | Markdown checked against the tree: every reference to a file or symbol the code no longer has, with the line to edit. [Doc drift](docs/layers/DOC_DRIFT.md) |+| *Why is it built this way?* | Decisions mined from ADRs, `# WHY:` comments, commit and PR history and your agent sessions, each tied to the code it governs and flagged when it goes stale. [Decisions](docs/layers/DECISIONS.md) |+| *Who knows this code?* | Owners, bus factor, knowledge-loss risk when the main author goes quiet, and suggested reviewers. [Ownership](docs/layers/OWNERSHIP.md) |+| *Can I see the architecture?* | An explorable dependency map, C4 views, and a Structurizr export, no model involved. [Dashboard](docs/start/DASHBOARD.md) |++### Change it safely++| You ask | Repowise gives you |+|---|---|+| *What breaks if I change this?* | Symbol-level blast radius: the callers of what you changed, the files that historically change with it but are missing from your diff, and the tests that reach it. [Change risk](docs/layers/CHANGE_RISK.md) |+| *How risky is this PR?* | Where the change ranks against your repository's own recent commits, with the reasons, as a directive your agent can act on. [Change risk](docs/layers/CHANGE_RISK.md) |+| *Which tests should run?* | The tests a diff actually exercises, from a coverage report if you have one and from the call graph if you do not. [Test intelligence](docs/layers/TEST_INTELLIGENCE.md) |+| *Did this PR add untested lines?* | Patch coverage, branch coverage on changed lines and path-scoped gates in CI, on GitHub, GitLab or any runner. [CI gates](docs/start/CI.md) |+| *Will this API change break another repo?* | HTTP, gRPC, topic and OpenAPI contracts matched across repositories, with a breaking-change guard and the consumer files it affects. [Workspaces](docs/scale/WORKSPACES.md) |+| *Is anyone else editing these files?* | Other open branches touching the same files or their co-change partners, each with the reason it is listed. [Branch overlap](docs/layers/CHANGE_RISK.md#branch-overlap) |+| *Did we just commit a secret?* | Keys, tokens and risky calls found in the working tree and in full git history, with a pre-commit check and a CI gate. [Security signals](docs/layers/SECURITY.md) |++### Improve it continuously++| You ask | Repowise gives you |+|---|---|+| *What should we fix first?* | A ranked queue weighing impact against effort, using churn, fan-in, coverage and bug history. [Fix first](docs/layers/CODE_HEALTH.md#fix-first) |+| *Where is the debt?* | A 1 to 10 score for every file from 53 deterministic detectors, split into defect risk, maintainability and performance, validated against real bug history. [Code health](docs/layers/CODE_HEALTH.md) |+| *Why is this slow?* | N+1 queries, I/O in loops, blocking calls inside async code and quadratic loops, traced across function and file boundaries. [Performance](docs/layers/CODE_HEALTH.md#performance-findings) |+| *How do I break this up safely?* | Concrete refactoring plans: Extract Method, Extract Class, Move Method, Split File, Break Cycle, with the exact symbols that move and what moves with them. Ready to hand to an agent. [Refactoring](docs/layers/REFACTORING.md) |+| *What can we delete?* | Unreachable files, unused exports and unused packages, each with a confidence tier and the evidence behind it. [Dead code](docs/layers/DEAD_CODE.md) |+| *Where do bugs keep landing?* | Bug-fix commits traced to files and symbols, and a warning when your agent edits a bug magnet. [Bug history](docs/layers/BUG_HISTORY.md) |+| *Are our tests testing anything?* | Tests with no assertions, tests that only check their own mocks, and untested hotspots. [Test-quality smells](docs/layers/CODE_HEALTH.md#test-quality-smells) |++<details>+<summary><strong>Things people do not expect it to do</strong></summary>++- **Test coverage without running coverage.** Most repositories never produce a+ coverage report. Repowise answers "is this tested, and by what" from the import and+ call graph, and labels every answer measured or inferred.+- **It checks its own health score on your repository.** After each index it reports+ how many of the lowest-scoring files actually had bug fixes in your recent history,+ so a bad score on your codebase is visible to you.+- **It learns from your agent sessions.** With transcript capture on, the corrections+ you keep repeating ("use the shared HTTP client") become tracked decisions it+ feeds back to the agent later. Transcripts never leave your machine.+- **It writes your `CLAUDE.md` and `AGENTS.md`** from the real index and keeps them+ current, so even an agent with no MCP support starts informed.+- **It shrinks command output before your agent reads it.** `repowise distill pytest`+ keeps every failure and drops the noise, and nothing is lost: an `expand` command+ restores any cut.+- **New git worktrees start indexed.** A linked worktree seeds its index from the base+ checkout instead of starting cold.+- **It knows which external systems you depend on.** Package manifests across PyPI,+ npm, Cargo, Go, NuGet, Maven and CMake feed a map of the services and libraries the+ code reaches.++</details>++<div align="center">+<img src=".github/assets/demo.gif" alt="A real Claude Code session asks Repowise what breaks before editing a file, then the dashboard shows the ranked Fix first queue, a refactoring plan ready to hand to an agent, dead code with confidence tiers, a commit ranked against the repo's own history, and the HTTP contract between a frontend and backend on the workspace system map" width="100%" />+<p><sub>Recorded on this repository and its workspace: an agent asking the index, then the+dashboard and CLI on the same local index. No API key and nothing uploaded.</sub></p>+</div>++### Pick your front door++| If you care about... | Start here |+|---|---|+| **A coding agent that knows the repository** | Task-shaped context in fewer calls, with decisions and risk delivered before the agent asks. [For agents ↓](#your-agent-stops-guessing) |+| **Safer pull requests and faster CI** | Change risk, symbol-level callers, missing co-changes and the tests a diff needs. [Change intelligence ↓](#know-whats-dangerous-before-you-merge) |+| **Paying down the code most likely to hurt you** | A defect-validated health score, then the concrete fix. [Code health ↓](#code-health) |+| **An estate of many repositories** | Contracts matched across repos, breaking-change guards, architecture rules in CI, one MCP endpoint for everything. [Workspaces ↓](#past-one-repo) |+| **Rolling it out across a company** | Self-hosted with nothing leaving your network, per-language accuracy, sizing, compliance status and licensing in one place. [Teams and enterprise ↓](#for-teams-and-enterprises) |---@@ -121,55 +194,54 @@ computed ahead of time. *Who calls this function? What breaks if I change it? Whit written this way? Which files are actually dangerous?* Without an index, the agentrediscovers that answer on every task: grep, read, re-read, forget.-Repowise exposes **ten task-shaped MCP tools** to Claude Code, Codex, Cursor, VS Code-and anything else that speaks MCP: graph, git, docs, decisions, and ten MCP tools-behind one index. See [the canonical surface](#the-ten-mcp-tools). Most tools are built around data entities (one-file, one symbol), which forces agents into long chains of sequential calls. These are-built around **tasks**: pass several targets in one call, get complete context back.--Because the exploration work is already done, that phase mostly disappears. In a-measured agent loop across 43 questions on `django/django`, Repowise cut the agent's-own output by **31.6%** (p<0.0001) and reached the answer in **3.8 tool calls-instead of 7.2**. That is the end-to-end result.--One mechanism is much larger but narrower: loading a commit's context through-`get_context` costs **393 tokens instead of 13,984**, or 97.2% less. That is one-retrieval payload, not a claim of 97.2% total agent savings. Both measurements and-every competitor row are published in [the benchmark report](docs/BENCHMARKS.md).--**And it arrives without being asked.** Optional [hooks](docs/agent/HOOKS.md) push-context into the session at the moment it matters: the governing architectural-decision when your agent edits a file that decision covers, a warning when it touches-a file with a run of recent bug fixes, a compact briefing at session start. Repowise-also generates your `CLAUDE.md` and `AGENTS.md` from the real index, so even an agent-with no MCP support starts informed.--**It learns from how you actually work.** Switch on transcript capture+Repowise gives Claude Code, Codex, Cursor, VS Code and any other MCP host+**ten task-shaped MCP tools** backed by one index of graph, git, docs and decisions.+Most code tools are built around data entities, one file or one symbol at a time,+which pushes agents into long chains of sequential calls. These are built around+tasks: pass several targets in one call and get the whole picture back.+[The tool list ↓](#the-ten-mcp-tools)++**About tokens.** Every tool in this category promises to cut your token bill, and+there are a lot of tools in this category. We think tokens are a symptom. An agent+burns them because it does not know the codebase, so it searches, opens files, opens+more files, and searches again. Give it an index that already knows, and the savings+show up on their own. They also happen to be the best we have measured: in a paired+agent loop over 43 questions on `django/django`, Repowise cut the agent's own output+by **31.6%** (p<0.0001) and got there in **3.8 tool calls instead of 7.2**, ahead+of every other context tool in the same run. On 42 sealed retrieval tasks it found+**0.876** of the files a fix needed, against **0.610** for the next tool.+[Method and every row we lose →](docs/BENCHMARKS.md)++**Context arrives before the agent asks.** Optional [hooks](docs/agent/HOOKS.md) push+it into the session when it matters: the governing decision when your agent edits a+file that decision covers, a warning when it touches a file with a run of recent bug+fixes, a short briefing at session start, and a correction when it reaches for a path+that does not exist.++**It learns from how you work.** Switch on transcript capture(`repowise decision source set session --on`) and Repowise reads your own agent-transcripts for the corrections you keep making ("use the shared HTTP client,-not raw requests"), turning the durable ones into tracked decisions it delivers-back later. Transcripts never leave your machine; one batched model call per-update turns the candidates that clear the deterministic gates into records, and-`--no-llm` keeps the gates and drops that call. The wiki generation budget tilts-toward the modules you and your agent ask about most, with no switch needed.+transcripts for the corrections you keep making, turning the durable ones into+tracked decisions it delivers back later. Transcripts never leave your machine; one+batched model call per update turns the candidates that clear the deterministic+gates into records, and `--no-llm` keeps the gates and drops that call.<details><summary><strong>What the index builds</strong></summary>-| Foundation | What it contributes |+Five layers, one index:++| Layer | What it contributes ||---|---|-| **Graph** | File + symbol dependencies across 26 AST-parsed languages, confidence-stamped call resolution, communities, centrality, cycles, and execution flows |-| **Git** | Hotspots, ownership, co-change, bus factor, and bug-fix history: behavioral signals static analysis cannot see |-| **Docs** | A wiki for every module and file, rebuilt incrementally with freshness and confidence scoring plus hybrid search |-| **Documentation drift** | Your own markdown checked against the tree: the claims each document makes, resolved against the graph, and the ones the code refutes |-| **Decisions** | Architectural rationale mined from five index-time sources plus human and agent capture, each claim traced to evidence |-| **Code health** | 53 deterministic detectors across defect risk, maintainability, and performance, followed by concrete refactoring plans |+| **1. Graph** | File and symbol dependencies across 26 AST-parsed languages, confidence-stamped call resolution, communities, centrality, cycles and execution flows |+| **2. Git history** | Hotspots, ownership, co-change, bus factor and bug-fix history: behavioural signals static analysis cannot see |+| **3. Docs** | A wiki for every module and file, hybrid search, and your own markdown checked against the tree for claims the code no longer supports |+| **4. Decisions** | Architectural rationale from ADRs, inline markers, commits, PRs and agent sessions, each claim traced to evidence |+| **5. Health and change** | 53 deterministic detectors across defect risk, maintainability and performance, change risk, test impact, dead code and concrete refactoring plans |The structural wiki needs no model. Model-written prose is an optional upgrade, one-page or directory at a time. Six of the seven decision sources are deterministic too;-only comment archaeology needs a provider.+page or directory at a time.-[The intelligence layers →](docs/layers/INTELLIGENCE_LAYERS.md) ·+[How the layers fit together →](docs/layers/INTELLIGENCE_LAYERS.md) ·[How the graph earns trust →](docs/layers/GRAPH.md)</details>@@ -177,8 +249,8 @@ only comment archaeology needs a provider.### Also: stop paying for output nobody readsMost of what an agent reads back from a shell command is noise: 300 lines of passing-tests wrapped around 4 failures, full commit bodies when it asked "what changed-recently". `repowise distill <cmd>` compresses command output **before the agent reads+tests wrapped around 4 failures, full commit bodies when it asked what changed+recently. `repowise distill <cmd>` compresses command output **before the agent readsit**, errors first, exit code preserved.```bash@@ -187,14 +259,14 @@ repowise distill git log -50 # 89% fewer tokensrepowise saved # what distillation saved you, in tokens and dollars```-Nothing is lost. Every omission leaves an inline `[repowise#<ref>]` marker that-`repowise expand <ref>` reverses in full, so the agent can always pull the detail back-without re-running the command. Small outputs pass through untouched. An opt-in hook-rewrites noisy commands automatically, shown to you for approval first.+Every omission leaves an inline `[repowise#<ref>]` marker that `repowise expand <ref>`+reverses in full, so the agent can pull the detail back without re-running the+command. Small outputs pass through untouched. An opt-in hook rewrites noisy commands+for the agent automatically.<div align="center"><img src=".github/assets/savings.png" alt="repowise Costs dashboard: tokens and dollars saved across distill and the MCP tools" width="100%" />-<p align="center"><sub>The <strong>Costs</strong> dashboard tallies both savings surfaces. Every event is priced from the model that produced it, never repriced at whatever you are running today, and the total is a deliberate floor: where the evidence is ambiguous it declines to claim a saving. Example from a week of heavy local use.</sub></p>+<p align="center"><sub>The <strong>Costs</strong> dashboard tallies both savings surfaces. Every event is priced from the model that produced it, and where the evidence is ambiguous it declines to claim a saving. Example from a week of heavy local use.</sub></p></div>Full guide: **[docs/agent/DISTILL.md →](docs/agent/DISTILL.md)**@@ -206,100 +278,63 @@ Full guide: **[docs/agent/DISTILL.md →](docs/agent/DISTILL.md)**Four deterministic signals, all computed from the graph and git history, no LLM:- **Change risk.** Score any commit or `base..HEAD` range **0-10** from the shape of- the diff, ranked against your repo's own recent commits. PR mode returns directives- rather than vibes: `may_break`, `missing_cochanges`, `missing_tests`, `tests_to_run`.- One command: `repowise risk main..HEAD`. ([reference →](docs/layers/CHANGE_RISK.md))+ the diff, ranked against your repository's own recent commits. PR mode returns+ directives an agent can act on: `may_break`, `missing_cochanges`, `missing_tests`,+ `tests_to_run`. One command: `repowise risk main..HEAD`.+ ([reference →](docs/layers/CHANGE_RISK.md))- **Bug history.** Which files and symbols actually get bug-fixed, and how recently.Doc, test and config commits are filtered out so the count means what it says, and a- file with a run of recent fixes gets flagged as a bug magnet while you edit it.+ file with a run of recent fixes is flagged as a bug magnet while you edit it.([reference →](docs/layers/BUG_HISTORY.md))-- **[Test intelligence](#which-tests-cover-this-file-without-a-coverage-report).** Which- tests reach a file and which ones a diff actually exercises, from the call graph,- with or without a coverage report.- ([reference →](docs/layers/TEST_INTELLIGENCE.md))+- **Test intelligence.** Which tests reach a file and which ones a diff exercises, from+ a coverage report or from the call graph. ([reference →](docs/layers/TEST_INTELLIGENCE.md))- **Change coordination.** Which other open branches edit the files you are editing,- every row saying why it is listed (`same file`, or a co-change pair with the commit- counts behind it), and whether the diff in front of you is one change or several- groups the index links nothing between. Both stay quiet when there is nothing to- report. `repowise overlap` and `repowise risk`.+ every row saying why it is listed, and whether the diff in front of you is one+ change or several unrelated ones. `repowise overlap` and `repowise risk`.([reference →](docs/layers/CHANGE_RISK.md#branch-overlap))-Plus the free **[Repowise PR Bot](#the-pr-bot)**, which puts all of it on every pull-request. Zero LLM calls.+### Which tests cover this file, without a coverage report------## Which tests cover this file, without a coverage report--Ingest LCOV, Cobertura, Clover, JaCoCo or a Go coverprofile and you get the-measured answer. **Most repositories never produce one**, so the graph answers-instead: a test file that imports a source file *reaches* it, which is a-recorded edge rather than the name-shaped guess everything else falls back to.--That fallback fails in both directions, and this repo is the proof. Five of its-six worst bug-magnet files have no test named for them and read as untested while-the graph names 3 to 23 test files each. The sixth is worse: matching on basename-paired the *health* engine with the *distill* engine's tests and called it tested.+Ingest LCOV, Cobertura, Clover, JaCoCo or a Go coverprofile and you get the measured+answer. Most repositories never produce one, so the graph answers instead: a test file+that imports a source file *reaches* it, which is a recorded edge, where most tools+fall back to matching file names.```bash-repowise impacted-tests main..HEAD # only the tests this diff actually exercises-repowise health # untested hotspots, now graph-aware+repowise impacted-tests main..HEAD # only the tests this diff exercises+repowise health # untested hotspots, graph-aware```-<sub>Dogfooded against a real <code>coverage run --contexts=test</code>:-<strong>95.7% precision</strong> on what reaches a file and <strong>97.5% on the-run list</strong>, at a 100% hit rate, against 72.1% and 94.8% for the one-hop-import walk this replaced. The two tiers are never averaged: rows are stamped-<code>basis: "measured"</code> or <code>"inferred"</code>, measured wins outright-where both can answer, and the inferred tier may never produce a percentage.-Sound as a floor, unsound as a quantity, and labelled so.+<sub>Checked against a real <code>coverage run --contexts=test</code> on this repository:+<strong>95.7% precision</strong> on what reaches a file and <strong>97.5% on the run+list</strong>. Every row is stamped <code>basis: "measured"</code> or+<code>"inferred"</code>, measured wins where both can answer, and an empty answer means+unknown, never "no tests".<a href="docs/layers/TEST_INTELLIGENCE.md"><strong>Test intelligence →</strong></a></sub>------## The PR bot+### In CI, and on every pull request-Install the [GitHub App](https://github.com/apps/repowise-bot) and the index shows up-where the decision actually gets made. One comment per pull request, edited in place on-every push rather than reposted, and **a green PR gets no comment at all**.--<sub>See a real comment on a real PR, not a mockup:-[repowise-dev/repowise#1204](https://github.com/repowise-dev/repowise/pull/1204).</sub>--What decides a review is inline. What is context sits behind one fold, so the comment-stays about seventeen rows whatever it finds.--- **Blast radius, at symbol level.** The contracts this PR changed and every caller of- them in a file the PR does not touch. Importing a module says nothing about whether- the function you changed is the one being called, so file-level impact is the wrong- altitude for the question a reviewer actually has.-- **Before you merge.** The tests that import your changed files, and the files that- changed alongside them in past commits but are missing here.-- **A Check Run that can gate the merge**, with annotations on the specific lines the- PR added. Advisory by default.-- **Change risk**, scored against the repository's own commit distribution rather than- an absolute scale, so it stays meaningful on a repo whose typical commit is large.-- **AI vs human authorship** of the changed files, with the average health of each.-- Then hotspots, hidden coupling, declining health, dead code and the change map, one- fold down.+Patch coverage, doc drift, security and change risk run as gates in your own+pipeline through the [GitHub Action](docs/start/CI.md#github-actions), a+[GitLab template](docs/start/CI.md#gitlab) or plain CLI commands anywhere else, with+annotations, SARIF and GitLab Code Quality output. The gates need no API key.+[Repowise in CI →](docs/start/CI.md)-### And a page the comment links to+<a id="the-pr-bot"></a>-Markdown runs out. The comment shows three callers and says "+6 more"; the page shows-all nine. Public, no sign-in, on a repository the reader has never seen.+On GitHub you can also install the free **[Repowise PR Bot](https://github.com/apps/repowise-bot)**,+a hosted GitHub App that puts the same analysis on every pull request. One comment,+edited in place on every push, and **a green PR gets no comment at all**. It shows+symbol-level blast radius (the contracts the PR changed and every caller outside the+PR), the tests and co-change partners missing from the change, change risk against+the repository's own history, and a public analysis page per PR. Zero LLM calls, so+the same diff always gets the same review.<img src=".github/assets/pr-bot/pr-page-blast-map-dark.png" alt="The dark Repowise per-PR analysis page showing change risk, repository health, changed contracts, outside callers, newly added findings, and a blast-radius treemap of the repository" width="100%" />-<sub>The page leads with change risk and newly introduced findings, then maps every-changed file and outside caller across the repository.-[See it live →](https://repowise.dev/pr/repowise-dev/repowise/1204)</sub>--**[Install the PR bot →](https://github.com/apps/repowise-bot)** ·-[how it works →](https://www.repowise.dev/bot)--Patch coverage, doc drift, security and change risk also run as CI gates in your own-pipeline, with a GitHub Action and a GitLab template, and need no index or API key.-[Repowise in CI →](docs/start/CI.md)+<sub>A real comment on a real PR: [repowise-dev/repowise#1204](https://github.com/repowise-dev/repowise/pull/1204) ·+[its analysis page →](https://repowise.dev/pr/repowise-dev/repowise/1204) ·+[install the PR bot →](https://github.com/apps/repowise-bot)</sub>---@@ -308,66 +343,60 @@ pipeline, with a GitHub Action and a GitLab template, and need no index or API k## ★ Know exactly what to fixA score that says *"this file is risky"* is where most tools stop. Repowise scores-every file, locates where the risk concentrates, and then names the specific fix.+every file, finds where the risk concentrates, and names the specific fix.<div align="center"><img src=".github/assets/health-loop.svg" alt="repowise code-health loop: deterministic markers fan into three signals, the graph and git history locate where risk concentrates, and refactoring intelligence emits concrete plans your agent executes" width="100%" /></div>Every file is scored 1-10 by **53 deterministic detectors** (McCabe complexity, brain-methods, LCOM4 cohesion, god classes, native Rabin-Karp clone detection, untested-hotspots, change entropy, prior-defect history and more), split into three lenses:-**defect risk**, **maintainability**, and **performance**: static N+1 and I/O-in-loop-risk traced *across* files through the call graph, where file-local linters found **0**-of the cross-function cases and repowise surfaced ~90. Only **26** of the 49 are-permitted to move the defect number, because that is the number carrying published-accuracy claims.--> **Zero LLM calls, zero cloud, zero new runtime dependencies.** Pure Python over-> tree-sitter and git data, **under 30 seconds** on a 3,000-file repo, a budget-> enforced by a CI test, not an estimate. Marker weights are **calibrated against a-> real defect corpus, not hand-tuned**: every file scored at a commit preceding the-> bug window so nothing leaks backward, and an L2-logistic fit with file size as an-> explicit control, so a marker only earns weight for defect lift *beyond* being big.-> Only the learned constants ship.--**It proves itself on your repo, not just on a benchmark.** After every index,-Repowise checks its own flags against your git history and reports what it found:-*"16 of the 20 lowest-health files had a bug fix in the last 6 months, 3.3x the 24%-baseline."* If that number is bad on your codebase, you will see it. (It is an-association on your indexed history, not a forward prediction, the leakage-free-version is [in the benchmarks](docs/BENCHMARKS.md#5-code-health-predicts-defects).)--Then it names the fix. Not "this class is too big", but **Extract Class**, **Extract-Helper**, **Move Method**, **Break Cycle**, **Split File**, or **Extract Method**, with-the exact methods, edges and symbols that move, the **blast radius** of callers and-co-changing files that have to move with them, and a graph-aware ranking so a fix on a-central hub outranks the same fix on a leaf. Extract Method goes down to an-intra-procedural dataflow pass that lifts the exact span and infers a-behavior-preserving signature.+methods, LCOM4 cohesion, god classes, clone detection, untested hotspots, change+entropy, prior-defect history and more), read through three lenses: **defect risk**,+**maintainability** and **performance**. Performance findings such as N+1 queries and+I/O in loops are traced *across* functions and files through the call graph, which is+where file-local linters lose them. Only 25 of the 53 detectors move the defect score,+because that is the number carrying published accuracy claims.++> **Zero LLM calls, zero cloud.** Detector weights are **calibrated against a real+> defect corpus, not hand-tuned**: every file scored at a commit before the bug+> window so nothing leaks backward, with file size as an explicit control, so a+> detector only earns weight for defect lift beyond a file being big.++**It checks itself on your repository.** After every index, Repowise compares its own+flags with your git history and tells you what it found: *"16 of the 20 lowest-health+files had a bug fix in the last 6 months, 3.3x the 24% baseline."* If that number is+bad on your codebase, you will see it.++**Then it names the fix.** **Extract Class**, **Extract Helper**, **Move Method**,+**Break Cycle**, **Split File** or **Extract Method**, with the exact methods, edges+and symbols that move, the callers and co-changing files that move with them, and a+ranking that puts a fix on a central hub above the same fix on a leaf. Extract Method+runs a dataflow pass over the function to lift the exact span and infer a+behaviour-preserving signature.```bash+repowise next # what to fix first, ranked by impact and effortrepowise health # KPIs and lowest-scoring filesrepowise health --refactoring-targets # ranked, concrete plansrepowise health --trend # snapshots plus declining-health alerts+repowise dead-code # what can go, by confidence tier```-The dashboard renders each plan as a card with a copy-to-agent button. An optional LLM-step, never in the indexing path and only on request, expands any plan into generated-code and a unified diff.--<sub>Validated on <strong>21 open-source repos across 9 languages</strong> (2,826 files,-scored at a fixed point and checked against the following 6 months of bug fixes,-keyword-labelled): <strong>ROC AUC 0.737</strong> [0.683, 0.787]. The signal is-correlated with file size and weakens sharply within a fixed size band, which we report-rather than bury. Independently recomputed from the raw data.</sub>+The dashboard renders each plan as a card with a copy-to-agent button. An optional+model step, never in the indexing path, expands a plan into generated code and a+unified diff.-<sub>Against <strong>CodeScene</strong>, the leading commercial code-health tool, on the-same 2,770 files and the same defect labels, ranking by repowise health surfaces-<strong>2.3x the defects under a fixed review budget</strong> (paired, p = 0.003).-<a href="docs/BENCHMARKS.md">Full head-to-head, methodology and limitations →</a></sub>+<sub>Validated on <strong>21 open-source repositories across 9 languages</strong>+(2,826 files scored at a fixed point and checked against the following 6 months of bug+fixes): <strong>ROC AUC 0.737</strong> [0.683, 0.787]. Against <strong>CodeScene</strong>+on the 2,770 files both tools scored, ranking by Repowise surfaces <strong>2.3x the+defects under a fixed review budget</strong> (paired, p = 0.003). CodeScene keeps a+shorter, slightly more precise list.+<a href="docs/BENCHMARKS.md#5-code-health-predicts-defects">Full head-to-head and its limits →</a></sub>-Guides: **[code health](docs/layers/CODE_HEALTH.md)** · **[refactoring](docs/layers/REFACTORING.md)**+Guides: **[code health](docs/layers/CODE_HEALTH.md)** ·+**[refactoring](docs/layers/REFACTORING.md)** ·+**[dead code](docs/layers/DEAD_CODE.md)**---@@ -383,18 +412,15 @@ setup, all local.</tr><tr><td width="50%"><img src=".github/assets/dashboard/chat-page.png" alt="Chat view: ask questions against the indexed repo, with answers that cite the files and pages they came from" width="100%" /><br/><sub><b>Chat</b> · ask the codebase a question, answers cite the files and pages they came from</sub></td>-<td width="50%"><img src=".github/assets/dashboard/docs-page.png" alt="Docs view: auto-generated wiki pages with a tree, mermaid diagrams, and freshness badges" width="100%" /><br/><sub><b>Docs</b> · auto-generated wiki pages for the whole codebase, with confidence and freshness badges</sub></td>+<td width="50%"><img src=".github/assets/dashboard/docs-page.png" alt="Docs view: auto-generated wiki pages with a tree, mermaid diagrams, and freshness badges" width="100%" /><br/><sub><b>Docs</b> · generated wiki pages for the whole codebase, with confidence and freshness badges</sub></td></tr></table>-Also in there: **Chat** (ask the codebase in natural language) · **Docs** (the-generated wiki, with Mermaid and a graph sidebar) · **Architecture** and **C4**-(Context → Containers → Components) · **Knowledge Graph** plus a zoomable canvas map ·-**Risk**, **Hotspots**, **Coupling** and **Blast radius** · **Contributors** ·-**Decisions** (evidence drawer and evolution timeline) · **Symbols** · **Security** ·-**Dead code** · **Stats** · **Costs** · **Workspace**.--Every view and what each one answers: **[docs/start/DASHBOARD.md →](docs/start/DASHBOARD.md)**+Also in there: **Architecture** and **C4** views, the **Knowledge Graph** and a+zoomable map, **Risk**, **Hotspots**, **Coupling** and **Blast radius**,+**Contributors** and **Ownership**, **Decisions** with an evidence drawer and+timeline, **Symbols**, **Security**, **Dead code**, **Costs** and **Workspace**.+Every view and what it answers: **[docs/start/DASHBOARD.md →](docs/start/DASHBOARD.md)**---@@ -404,102 +430,39 @@ Every view and what each one answers: **[docs/start/DASHBOARD.md →](docs/startReal systems are not one repository, and the expensive failures live in the gapsbetween them. Change a backend contract and Repowise can name the frontend calls that-consume it, the services downstream, the historical companion files missing from the-change, and the architecture rule the new dependency violates before it ships.+consume it, the services downstream, the companion files missing from the change, and+the architecture rule the new dependency violates, before it ships.| Workspace intelligence | What it answers ||---|---|-| **Contract map** | Which services provide and consume each HTTP, gRPC, event, socket, and data contract? Links retain exact/candidate confidence and the source evidence. |-| **Cross-repo blast radius** | If this provider changes, which downstream services are in structural reach, and which ones may drift through historical co-change? |-| **Breaking-change guard** | Was an endpoint removed or a supported OpenAPI / proto / signature shape changed incompatibly, and which consumer files are linked to that contract? |-| **Test impact** | Which tests in the consumer repos should run for this provider change, measured from coverage or inferred from the call graph, and which links could not be determined? |+| **Contract map** | Which services provide and consume each HTTP, gRPC, event, socket and data contract? Links keep their exact or candidate confidence and the source evidence. |+| **Cross-repo blast radius** | If this provider changes, which downstream services are in structural reach, and which may drift through historical co-change? |+| **Breaking-change guard** | Was an endpoint removed or an OpenAPI, proto or signature shape changed incompatibly, and which consumer files are linked to that contract? |+| **Test impact** | Which tests in the consumer repos should run for this provider change, measured from coverage or inferred from the call graph? || **Architecture as code** | Does the live system graph violate declared dependency rules or contain cycles? `repowise workspace check` gates CI. |-| **Architecture health** | How coupled is the estate? Track propagation cost, the cyclic core, service roles, and a deterministic 1–10 architecture score. |-| **Federated context** | One dashboard and one MCP server answer across every repository while preserving repo-level evidence. |--The system map models **services**, not merely repository boxes, and never conflates a-real contract with “these files often changed together.” HTTP field-level comparison-supports the bounded OpenAPI 3.x JSON subset documented in the workspace guide;-matched consumers prove endpoint exposure, not field use or runtime failure.+| **Architecture health** | How coupled is the estate? Propagation cost, the cyclic core, service roles and a deterministic 1-10 architecture score. |+| **Federated context** | One dashboard and one MCP server answer across every repository while keeping repo-level evidence. |+The system map models **services**, not just repository boxes, and never conflates a+real contract with "these files often changed together". HTTP field-level comparison+covers a bounded OpenAPI 3.x subset; matched consumers prove endpoint exposure, not+field use or runtime failure.**[Workspace guide and exact support matrix →](docs/scale/WORKSPACES.md)**-Worktrees and updates stay lightweight: a linked worktree seeds its index from the base-checkout automatically, and post-commit hooks, file watching, webhooks, or polling keep-each repository and the cross-repo graph current.-------## In your editor--The **Repowise** VS Code extension puts the index where code actually gets written:-know what your change breaks before you push (riskiest files ranked, what is-downstream, forgotten companion files, missing tests, suggested reviewers), health in-the gutter and status bar, callers and ownership on hover, refactoring plans as-CodeLens, and the full dashboards inside the editor. One install also registers the MCP-server with VS Code, so the same local index serves both you and your agent, and-exposes six tools to GitHub Copilot. Quiet by default, everything toggleable, nothing-leaves your machine.--Install from the Marketplace (search **Repowise**) or Open VSX, then run **Repowise:-Set Up This Repository**. Guide: **[docs/agent/VSCODE.md →](docs/agent/VSCODE.md)**-------## Supported agents--**Six agents wired end to end · two at the Full tier · every other MCP host one-paste away.**--<details>-<summary><strong>See integration tiers and supported agents</strong></summary>--<p>- <strong>Full tier </strong>- <img src="https://img.shields.io/badge/Claude_Code-D97757?style=flat-square&logo=claude&logoColor=white" alt="Claude Code" />- <img src="https://img.shields.io/badge/Codex_CLI-000000?style=flat-square&logo=openai&logoColor=white" alt="Codex CLI" />-</p>-<p>- <strong>Good tier </strong>- <img src="https://img.shields.io/badge/VS_Code-007ACC?style=flat-square&logo=visualstudiocode&logoColor=white" alt="VS Code" />- <img src="https://img.shields.io/badge/Cursor-000000?style=flat-square&logo=cursor&logoColor=white" alt="Cursor" />- <img src="https://img.shields.io/badge/OpenCode-000000?style=flat-square&logo=opencode&logoColor=white" alt="OpenCode" />- <img src="https://img.shields.io/badge/Hermes-000000?style=flat-square&logoColor=white" alt="Hermes" />-</p>--**Full** is every surface repowise has: MCP tools, skills, slash commands, a managed-instructions file, hook-level interception of tool calls, and transcript mining after-the session. **Good** is the honest half of that: MCP tools and the config to reach-them, but no hook-level interception and no transcript mining. A Good-tier agent can-ask repowise anything; repowise never sees the tool calls in between. The tier is-computed from what each integration actually wires, so this list cannot claim a depth-the code does not have.--Everything else that speaks MCP is one snippet away. `repowise agents print-config-claude-code` prints a stdio server entry to paste into Cline, Windsurf, Zed, Gemini-CLI or any other host that keys on `mcpServers`, and repowise writes nothing.--Adding an agent takes **one descriptor file and one registry line**, with no changes to-the orchestrators. Full matrix and the contributor recipe:-**[docs/agent/INTEGRATIONS.md →](docs/agent/INTEGRATIONS.md)**--</details>+Worktrees and updates stay light: a linked worktree seeds its index from the base+checkout, and post-commit hooks, file watching, webhooks or polling keep each+repository and the cross-repo graph current.+[Keeping the index fresh →](docs/scale/AUTO_SYNC.md)---## Supported languages-**26 languages parsed to AST · 40 on a five-rung ladder · framework-aware where-an ecosystem handler exists.**--"Do you support X" has five useful answers, not two, so languages land on a-ladder and every rung says what it buys you.--<details>-<summary><strong>See the complete language ladder</strong></summary>+**26 languages parsed to an AST, 40 on a five-rung ladder, framework-aware where an+ecosystem handler exists.** Every language ships in the open-source distribution.<p>- <strong>Full tier </strong>+ <strong>Full </strong><img src="https://img.shields.io/badge/Python-3776AB?style=flat-square&logo=python&logoColor=white" alt="Python" /><img src="https://img.shields.io/badge/TypeScript-3178C6?style=flat-square&logo=typescript&logoColor=white" alt="TypeScript" /><img src="https://img.shields.io/badge/JavaScript-F7DF1E?style=flat-square&logo=javascript&logoColor=black" alt="JavaScript" />@@ -515,7 +478,7 @@ ladder and every rung says what it buys you.<img src="https://img.shields.io/badge/Ruby-CC342D?style=flat-square&logo=ruby&logoColor=white" alt="Ruby" /></p><p>- <strong>Good tier </strong>+ <strong>Good </strong><img src="https://img.shields.io/badge/C-A8B9CC?style=flat-square&logo=c&logoColor=black" alt="C" /><img src="https://img.shields.io/badge/Swift-F05138?style=flat-square&logo=swift&logoColor=white" alt="Swift" /><img src="https://img.shields.io/badge/PHP-777BB4?style=flat-square&logo=php&logoColor=white" alt="PHP" />@@ -532,202 +495,98 @@ ladder and every rung says what it buys you.<img src="https://img.shields.io/badge/Razor-512BD4?style=flat-square&logo=blazor&logoColor=white" alt="Razor / Blazor" /></p>-Below those two rungs the ladder keeps going, and a language on a lower rung is-still doing real work rather than being ignored:+<details>+<summary><strong>The full ladder, and what each rung gives you</strong></summary>| Rung | Languages | What you get ||---|---|---|-| **Full** (13) | Python · TypeScript · JavaScript · Svelte · Vue · Java · Kotlin · Go · Rust · C++ · C# · Scala · Ruby | The whole pipeline: AST symbols, import resolution, a resolved call graph, heritage, docstrings, framework edges, **and code-health markers** |-| **Good** (11) | C · Swift · PHP · Dart · Object Pascal · COBOL · GDScript · VB.NET · Elixir · F# · Objective-C | All of the above except the full health suite, subject to the language-specific ceilings in the full matrix |+| **Full** (13) | Python · TypeScript · JavaScript · Svelte · Vue · Java · Kotlin · Go · Rust · C++ · C# · Scala · Ruby | The whole pipeline: AST symbols, import resolution, a resolved call graph, heritage, docstrings, framework edges and code-health markers |+| **Good** (11) | C · Swift · PHP · Dart · Object Pascal · COBOL · GDScript · VB.NET · Elixir · F# · Objective-C | All of the above except the full health suite, within the language-specific ceilings in the full matrix || **Partial** (2) | Luau / Roblox · Razor / Blazor | Luau: AST symbols and `require()` resolution, Rojo and `.luaurc` aware. Razor: component symbols, `@code` and component-tag call edges, C# health markers; no import resolution yet |-| | | ⎯⎯ *tree-sitter parsing stops here; the rungs below come from git and imports* ⎯⎯ || **Lightweight** (6) | Clojure · Haskell · Lean 4 · Erlang · HTML · QML | A real file-to-file import graph, and no symbol-level claims || **Structural** (8) | R · Zig · Julia · Elm · OCaml · Crystal · Nim · D | Git history: blame, hotspots, co-change, ownership, bug history |-**Every language ships in the open-source distribution.** None is gated behind-the commercial licence, and none will be. COBOL now ships at the Good tier;-languages still moving up the ladder are tracked on the-**[roadmap →](ROADMAP.md#languages)**.--SQL and dbt projects get real `ref()` / `source()` lineage, shell scripts get-function-level symbols, HTML pages contribute their `<script src>` / `<link href>`-dependencies (including `index.html` → `src/main.ts`), and OpenAPI, Protobuf,-GraphQL, Dockerfile, Terraform and friends get dedicated handlers. Anything else is-still tracked through git history: blame, hotspots, co-change.+SQL and dbt projects get `ref()` / `source()` lineage, shell scripts get function-level+symbols, HTML pages contribute their `<script src>` and `<link href>` dependencies, and+OpenAPI, Protobuf, GraphQL, Dockerfile, Terraform and similar formats get dedicated+handlers. Anything else is still tracked through git history.Every call edge is stamped with **how it was resolved and how much to trust it**, from-`same_file` at 0.95 down to a repo-wide name match at 0.50, labelled as the guess it is-([how that works](docs/layers/GRAPH.md)).-Adding a language takes five small steps and **no changes to the parser core**.+`same_file` at 0.95 down to a repo-wide name match at 0.50, labelled as the guess it+is. Accuracy per language, graded against each language's own compiler:+[docs/BENCHMARKS.md#accuracy-by-language](docs/BENCHMARKS.md#accuracy-by-language).Full matrix: **[docs/layers/LANGUAGE_SUPPORT.md →](docs/layers/LANGUAGE_SUPPORT.md)** ·-The graph itself: **[docs/layers/GRAPH.md →](docs/layers/GRAPH.md)** ·-Contributor recipe and internals:-**[docs/architecture/language-support.md →](docs/architecture/language-support.md)**+adding a language takes five small steps and no changes to the parser core:+**[docs/architecture/language-support.md →](docs/architecture/language-support.md)** ·+languages moving up the ladder: **[roadmap →](ROADMAP.md#languages)**</details>----<details>-<summary><strong>Agent setup and optional model-written prose</strong></summary>--**1. Install**--```bash-pip install repowise # Windows: python -m pip install repowise-repowise --version-```--**2. Index your repo**--```bash-cd /path/to/your/repo-repowise init-```--Bare `init` asks. It scans the repo first, then offers three ways to index it:-everything (the wiki written by a model), no prose (the same wiki rendered from-your code's structure, no key and no spend), or advanced, which walks through the-indexing and generation knobs. Nothing is spent before you see an estimate and-confirm it.--If you would rather not answer questions, or you are scripting this, name the-mode and add `-y`:--```bash-repowise init --no-prose -y # free, no key, no questions-repowise init --prose -y # model-written subsystem pages, cost pre-approved-```--Either way you get the dependency graph, git history, code-health scores and-dead-code findings in seconds, plus a complete wiki: file, module, layer and cycle-pages, the architecture diagram, the repo overview, API and infra pages, and the-onboarding collection. On the keyless path every page carries a footer saying it-was derived from structure, and the repo overview describes composition, entry-points, clusters and dependencies rather than what the project does end to end,-because no template can derive that. Full-text search works on this index;-semantic search needs an embedder configured (Ollama is the keyless option).--Went keyless and want the wiki written by a model later? You do not have to decide-now. Upgrade it whenever you like with `repowise generate`, a page, a directory,-or the whole thing at a time, each behind a cost estimate:--```bash-export ANTHROPIC_API_KEY="sk-ant-..." # or OPENAI_API_KEY / GEMINI_API_KEY-repowise generate # write the unwritten subsystem pages, behind one cost estimate-repowise generate --path src/api # or just one area first-repowise generate --all # or rewrite the prose on every subsystem page-```--Bare `repowise generate` prints the wiki's state and writes the unwritten-subsystem (concept) pages behind a single cost estimate. Every other page was-already rendered from structure at index time.--Or pick the provider for the first index directly with `repowise init --provider-gemini|anthropic|openai`.--**Resuming an interrupted index.** If `init` is interrupted (timeout, crash,-Ctrl+C), re-run it with `--resume` and it continues from where it stopped —-pages already written to the vector store are skipped, and only the missing-ones are generated:--```bash-repowise init . --resume-```--`--resume` is a safe no-op on a fully indexed repo, so it is the right thing to-reach for whenever a long run is cut short. It works because pages are written-to LanceDB incrementally, while the SQL `generation_jobs` row only finalizes at-the end — a hard interrupt can leave LanceDB ahead of SQL, and `--resume` is-the supported recovery path (`repowise doctor` flags the drift).--**3. Connect your agent.** Step 2 already did this for Claude Code: `init`-writes a repo-root `.mcp.json` unconditionally and, unless you passed-`--no-editor-setup`, also registers repowise with `~/.claude/settings.json`.-Open a session in this repo and it is already wired; check with `repowise-agents`.--<details><summary><b>Claude Code</b></summary>--Skipped editor setup, or setting up another machine?--```bash-repowise agents add --target=claude-code-```--The plugin additionally adds slash commands and skills, which `init` does not-install:--```bash-/plugin marketplace add repowise-dev/repowise-/plugin install repowise@repowise-```--Or wire the MCP server by hand:--```bash-claude mcp add repowise -- repowise mcp-```-Or edit the project `.mcp.json` `init` already wrote:-```json-{ "mcpServers": { "repowise": { "command": "repowise", "args": ["mcp"] } } }-```-</details>--<details><summary><b>Codex CLI</b></summary>--Add to `~/.codex/config.toml`:-```toml-[mcp_servers.repowise]-command = "repowise"-args = ["mcp"]-```-Or: `codex mcp add repowise -- repowise mcp`-</details>--**4. First real call.** Ask your agent: *"Use repowise `get_overview` to summarize this-repo"*, or *"`get_context` for `src/auth.py`"*. You get graph-grounded architecture and-per-file triage instead of a flurry of greps.+## Supported agents and editors-> `get_overview` and `get_context` work in index-only mode with no key, synthesized-> from the graph, git and health layers. `search_codebase` and `get_answer` read the-> wiki, which index-only mode does build, but they answer from pages rendered from-> structure rather than model-written prose, and `search_codebase` is full-text only-> until you configure an embedder.+**Six agents wired end to end · two at the Full tier · every other MCP host one+paste away.**-Full walkthrough: **[docs/start/QUICKSTART.md →](docs/start/QUICKSTART.md)**+<p>+ <strong>Full tier </strong>+ <img src="https://img.shields.io/badge/Claude_Code-D97757?style=flat-square&logo=claude&logoColor=white" alt="Claude Code" />+ <img src="https://img.shields.io/badge/Codex_CLI-000000?style=flat-square&logo=openai&logoColor=white" alt="Codex CLI" />+</p>+<p>+ <strong>Good tier </strong>+ <img src="https://img.shields.io/badge/VS_Code-007ACC?style=flat-square&logo=visualstudiocode&logoColor=white" alt="VS Code" />+ <img src="https://img.shields.io/badge/Cursor-000000?style=flat-square&logo=cursor&logoColor=white" alt="Cursor" />+ <img src="https://img.shields.io/badge/OpenCode-000000?style=flat-square&logo=opencode&logoColor=white" alt="OpenCode" />+ <img src="https://img.shields.io/badge/Hermes-000000?style=flat-square&logoColor=white" alt="Hermes" />+</p>-</details>+**Full** is every surface Repowise has: MCP tools, skills, slash commands, a managed+instructions file, hook-level interception of tool calls, and transcript mining after+the session. **Good** is MCP tools and the config to reach them, without hooks or+transcript mining. Anything else that speaks MCP is one snippet away:+`repowise agents print-config claude-code` prints a server entry for Cline, Windsurf,+Zed, Gemini CLI or any host that reads `mcpServers`.+[Integration matrix →](docs/agent/INTEGRATIONS.md)++**In VS Code**, the Repowise extension shows what your change breaks before you push+(riskiest files, what is downstream, forgotten companion files, missing tests,+suggested reviewers), health in the gutter and status bar, callers and ownership on+hover, and refactoring plans as CodeLens. One install also registers the MCP server,+so the same index serves you and your agent. Install from the Marketplace or Open VSX+and run **Repowise: Set Up This Repository**. [VS Code guide →](docs/agent/VSCODE.md)---## The ten MCP tools-Every response carries an `_meta` envelope with `index_age_days`, `indexed_commit`, and-a `stale_warning` that fires only when the indexed HEAD diverges from live `.git/HEAD`,-so your agent always knows how much to trust what it just read.+Every response carries a `_meta` envelope with the indexed commit, the index age and a+stale warning when the index has fallen behind your checkout, so your agent always+knows how much to trust what it just read.<details>-<summary><strong>See the complete MCP tool surface</strong></summary>+<summary><strong>See the MCP tool surface</strong></summary>| Tool | What only this tool answers ||---|---|-| `get_overview()` | Architecture summary, module map, entry points, git health. The first call on any unfamiliar codebase. |-| `get_answer(question)` | Hybrid retrieval (full-text plus vector via RRF), PageRank bias and 1-hop graph expansion into one cited answer with a calibrated `retrieval_quality`. Collapses search → read → reason into a single round-trip. |-| `get_context(targets, include?)` | Triage card for files, modules or symbols: summary, signatures, `hotspot` bit, governing decisions, `symbol_id`s. `include` opens callers, callees, ownership and metrics. Batch many targets in one call. |-| `get_symbol("file.py::Name")` | Source for one indexed symbol with exact line bounds. Cheaper and safer than `Read` plus offset math. |-| `search_codebase(query, kind?)` | Semantic search over the wiki, filterable by kind (implementation / test / config / doc), tagging each result's `search_method`. |-| `get_risk(targets?, changed_files?)` | Hotspots, dependents, co-change partners, ownership, test gaps, bug history. Pass `changed_files` for PR mode and get a `directive` block back. |-| `get_change_risk(revspec)` | What a commit, range or uncommitted change newly made worse across defect, maintainability and performance, why each finding is attributable to it, the tests coverage proves it touches, and how the diff's shape ranks against recent commits. |-| `get_why(query?, targets?)` | Architectural decisions and their verbatim evidence spans, stamped exact / fuzzy / unverified. Falls back to git archaeology when no decisions exist. |-| `get_dead_code(...)` | Unreachable code by confidence tier with cleanup-impact estimates, and cross-repo consumer detection in workspace mode. |-| `get_health(targets?, include?)` | Per-file marker scores across all three signals. `include` opens coverage, trends, per-file signals, the accuracy self-check, and structured refactoring plans. |--Ten is a deliberate ceiling rather than a limit we ran into: a small, task-shaped-surface is easier for an agent to choose from than a large one. Worked example (*"add-rate limiting to all API endpoints"* in 5 calls instead of ~30 greps and reads), the-opt-in tools, and the full reference: **[docs/agent/MCP_TOOLS.md →](docs/agent/MCP_TOOLS.md)**+| `get_overview()` | Architecture summary, module map, entry points, git health. The first call on an unfamiliar codebase. |+| `get_answer(question)` | Hybrid retrieval (full-text plus vector), graph expansion and one cited answer with a calibrated `retrieval_quality`. Search, read and reason in a single round-trip. |+| `get_context(targets, include?)` | Triage card for files, modules or symbols: summary, signatures, hotspot flag, governing decisions, symbol ids. `include` opens callers, callees, ownership and metrics. Batch many targets in one call. |+| `get_symbol("file.py::Name")` | One indexed symbol's source with exact line bounds. |+| `search_codebase(query)` | Hybrid search over code and docs, by symbol, path or concept. |+| `get_risk(targets?, changed_files?)` | Hotspots, dependents, co-change partners, ownership, test gaps, bug history. Pass `changed_files` for PR mode and get a `directive` back. |+| `get_change_risk(revspec)` | What a commit, range or uncommitted change made worse across defect risk, maintainability and performance, the tests that touch it, and how the diff ranks against recent commits. |+| `get_why(query?, targets?)` | Architectural decisions with their verbatim evidence. Falls back to git archaeology when no decision exists. |+| `get_dead_code(...)` | Unreachable code by confidence tier, with cross-repo consumers in workspace mode. |+| `get_health(targets?, include?)` | Health scores and findings across all three lenses, Fix first, coverage, trends, doc drift and refactoring plans. |++Ten is a deliberate ceiling: a small, task-shaped surface is easier for an agent to+choose from than a large one. Seven more tools (dependency paths, execution flows,+refactoring code generation, finding triage, and three workspace architecture tools)+are opt-in. Parameters, examples and when to use which:+**[docs/agent/MCP_TOOLS.md →](docs/agent/MCP_TOOLS.md)**</details>@@ -735,20 +594,9 @@ opt-in tools, and the full reference: **[docs/agent/MCP_TOOLS.md →](docs/agent## Measured against the field-Six open-source agent-context tools, the same repositories, the same pinned-commits, the same questions, each one given its own full advertised tool surface.-The full page carries the rows we lose beside the rows we win.--**Token reduction needs a denominator.** If the comparison is one context payload,-Repowise reduces 13,984 naive-read tokens to 393, a **97.2% reduction**. If the-comparison is the agent's complete output across a real task loop, the reduction is-**31.6%**. Competitor pages often publish the first kind as "token savings"; we-publish both and call only the second one agent savings.--The same rule applies to graphs: coverage without correctness rewards fake edges,-while precision without recall rewards drawing almost nothing. Our compiler-graded-claim is therefore the pair: in all seven comparisons, no tool that recovers as much-of the call graph gets more of it right.+Open-source agent-context tools, the same repositories, the same pinned commits, the+same questions, each tool given its full advertised surface. The full page carries the+rows we lose beside the rows we win.<div align="center"><picture>@@ -762,152 +610,120 @@ of the call graph gets more of it right.</div>- **Finds the right files.** 0.876 file coverage against CodeGraph's 0.610 on a- **sealed** 42-instance split, held out from every improvement round. 19 wins,- 1 loss per instance. Deterministic grading, no LLM judge. *n=42, sign test- p=0.00004.* CodeGraph scores the same on both halves to three decimals, so- neither half is the easy one.-- **Less work in a real agent loop.** -31.6% output tokens against a bare agent,- leaner on 37 of 44 questions. *n=43, p<0.0001.* CodeGraph is a genuine second- at -24.4%: more than one tool here works, and we lead the field rather than- being alone in it.-- **Fewer steps to get there.** 3.8 tool calls where the bare agent needed 7.2,- and 3.0 files opened instead of 7.2, the mechanism behind the token saving,- visible directly rather than inferred.--**[The full results, the methodology, and the rows we lose →](docs/BENCHMARKS.md)**-------## How it compares on capability--No single product competes with all of this, so there is no single table. Three-axes, three sets of real peers. Rows marked *measured* are head-to-head numbers,-and they link to **[docs/BENCHMARKS.md](docs/BENCHMARKS.md)** where the sample-sizes, the tests and the rows we lose all live.+ **sealed** 42-instance split held out from every improvement round. 19 wins, 1 loss.+ Deterministic grading, no LLM judge. *n=42, sign test p=0.00004.*+- **Less work in a real agent loop.** -31.6% output tokens against a bare agent, leaner+ on 37 of 44 questions. *n=43, p<0.0001.* CodeGraph is a genuine second at -24.4%.+- **Fewer steps.** 3.8 tool calls where the bare agent needed 7.2, and 3.0 files opened+ instead of 7.2: the mechanism behind the saving, visible directly.+- **A call graph the compiler agrees with.** Against the Go team's own call graph and+ the TypeScript checker's own resolution, Repowise main reads 0.976 to 0.995+ precision, with recall up in all 7 cells since August. In all seven, no tool that+ recovers as much of the graph gets more of it right. The tool with the highest+ recall in most Go cells is still not us, and the page says so.+- **Accuracy by language.** Graded against each language's own toolchain: C 0.95 to+ 0.98, Python 0.96 to 0.98 on three of four repositories, Rust 0.85 to 0.90, C# 0.73+ to 0.92, C++ 0.75, Java 0.73 to 0.78, each with its recall and its weak spots.+- **Scale.** dotnet/runtime, 58,924 files, indexed in 120 minutes at 11.7 GiB peak+ memory on one laptop. Building just the call graph uses 75 MB at the median, the+ lowest of five tools on 35 of 35 repositories.++**[The full results, the methodology, and the rows we lose →](docs/BENCHMARKS.md)** ·+**[The research it is built on →](docs/LINEAGE.md)**<details>-<summary><strong>Open the complete capability comparisons</strong></summary>+<summary><strong>How it compares on capability</strong></summary>-### As an agent context layer+No single product competes with all of this, so there is no single table. Rows marked+*measured* are head-to-head numbers that link to+**[docs/BENCHMARKS.md](docs/BENCHMARKS.md)**, where the sample sizes, the tests and the+rows we lose live. Unmarked rows are capability presence, not measurements.-Against the tools doing the same job: index a repository, serve it to a coding-agent over MCP.+### As an agent context layer| | repowise | CodeGraph | Serena | DeepWiki ||---|---|---|---|---|| Self-hostable, open source | ✅ AGPL-3.0 | ✅ | ✅ | ❌ cloud only || Private repo, no cloud | ✅ | ✅ | ✅ | ❌ OSS forks only |-| MCP tools served | 10 core + workspace tools | 1 | 29 | 3 |+| MCP tools served | 10 default + 7 opt-in | 1 | 29 | 3 || **Finds the gold files** *([measured](docs/BENCHMARKS.md#1-finding-the-right-files), n=42 sealed)* | ✅ **0.876** | 0.610 | not in this run | not measured || **Output tokens vs a bare agent** *([measured](docs/BENCHMARKS.md#2-what-changes-in-a-real-agent-loop), n=43)* | ✅ **-31.6%** | -24.4% | -14.8% | not measured || **Memory to build the graph** *([measured](docs/BENCHMARKS.md#what-it-costs-to-run), 5 tools, 35 repos)* | ✅ **75 MB**, lowest on 35 of 35 | 757 MB | not measured | n/a, cloud || **Time to build the graph** *([measured](docs/BENCHMARKS.md#what-it-costs-to-run), same run)* | **2.77s**, fastest on 14 of 35 | **3.65s**, fastest on 16 | not measured | n/a, cloud || **Time to build the full index, django** *([measured](docs/BENCHMARKS.md#what-it-costs-to-run))* | ⚠️ **366.8s**, slowest here | ✅ **16.4s** | not measured | n/a, cloud |-| | *five layers against their one; one-time, updates after it are incremental* | | | |-| **Call-edge precision** *([measured](docs/BENCHMARKS.md#7-edge-precision), 540 rows hand-graded from source)* | ✅ **84.8%** | 57.0% | not measured | not measured |-| **Call-edge precision, judged by a compiler** *([measured](docs/BENCHMARKS.md#8-the-same-question-against-an-answer-key-we-do-not-control), 5 tools, 7 cells, 37,853 edges)* | ✅ **nothing that finds as much gets more of it right**, 7 of 7 | lower precision in 7, and lower recall in 5 | not measured | not measured |+| **Call-edge precision, hand-graded** *([measured](docs/BENCHMARKS.md#7-edge-precision), 560 rows, 9 languages)* | ✅ **85.7%** | 58.6% | not measured | not measured |+| **Call-edge precision, judged by a compiler** *([measured](docs/BENCHMARKS.md#8-the-same-question-against-an-answer-key-we-do-not-control), 5 tools, 7 cells)* | ✅ **nothing that finds as much gets more of it right**, 7 of 7 | lower precision in 7 | not measured | not measured || Generated documentation | ✅ | ❌ | ❌ | ✅ || Proactive agent hooks | ✅ Claude + Codex | ❌ | ❌ | ❌ |-| Auto-generated AI instructions (`CLAUDE.md`, `AGENTS.md`) | ✅ | ❌ | ❌ | ❌ |+| Generated `CLAUDE.md` / `AGENTS.md` | ✅ | ❌ | ❌ | ❌ || Command-output distillation | ✅ reversible | ❌ | ❌ | ❌ |-| Learns from your usage (session-mined decisions, demand-weighted docs) | ✅ | ❌ | ❌ | ❌ || Architectural decision records | ✅ | ❌ | ❌ | ❌ || Multi-repo workspace intelligence | ✅ contracts, co-change, federated MCP | ❌ | ❌ | ❌ |-**The two cost rows answer different questions.** Building the call graph, we are-the lightest tool measured, about ten times lighter than the next, and roughly as-fast as the fastest. Building the *whole* index, CodeGraph is **22x faster than we-are**, because by then we have also built the git-history layer, the wiki, the-decisions and the health pass. If a call graph is all you need, that is the right-trade and you should take it. With prose generation on, which is what a default-`repowise init` costs, it is **135x**. Graphify and-code-review-graph were in the same measured field and are on the benchmarks page.--The precision row cuts the other way and is worth stating as plainly: of the call-edges we draw, **about fifteen percent are wrong**, and on `seastar` CodeGraph-grades better than we do. Nine languages were read on both sides, four separate,-five are statistical ties.--The compiler row exists because we graded the hand-read one ourselves. On Go and-TypeScript the answer key is the Go team's own RTA call graph and the `tsc`-checker's own resolution, which we neither wrote nor can tune.--**Read that row carefully, because it is a claim about two numbers.** Precision-alone is easy to win by drawing almost nothing, and two of the five tools score-above us that way, one of them at 0.997 from a graph holding 17% of the calls in-the repository. Recall alone is easy to win by drawing everything, and the tool-that leads it emits, on the largest repository measured, more than a third of its-edges as calls that do not exist. What we claim is the pair: **in all seven cells,-no tool that recovers as much of the call graph as we do gets more of it right.**-The column we lose is still there and is still ours to lose: **the tool with the-highest recall in every Go cell is not us.**--<sub>Measured against CodeGraph 1.5.0, Graphify 0.9.31, Serena 1.6.2.dev0,-code-review-graph 2.3.7, on repowise `081a59fa` (between v0.37.0 and v0.38.0),-August 2026. Unmarked rows are capability presence, not measurements.</sub>+**The two cost rows answer different questions.** Building the call graph, we are the+lightest tool measured, about ten times lighter than the next, and roughly as fast as+the fastest. Building the *whole* index, CodeGraph is **22x faster than we are**,+because by then we have also built the git-history layer, the wiki, the decisions and+the health pass. If a call graph is all you need, that is the right trade and you+should take it. With model-written prose on, it is **135x**.++**The hand-graded precision row cuts both ways.** About fourteen percent of the call+edges we draw are wrong, and on `seastar` CodeGraph grades better than we do. The+compiler row exists because we graded the hand-read one ourselves: on Go and+TypeScript the answer key is the Go team's own call graph and the `tsc` checker's own+resolution, which we neither wrote nor can tune. Precision alone is easy to win by+drawing almost nothing, and recall alone by drawing everything, so the claim is the+pair.++<sub>Competitors measured at CodeGraph 1.5.0, Graphify 0.9.31, Serena 1.6.2.dev0 and+code-review-graph 2.3.7 in August 2026. Repowise compiler-graded figures are from main+on 2026-10-03; the other Repowise rows are from `081a59fa`, August 2026.</sub>### As a code health tool| | repowise | CodeScene ||---|---|---|| Self-hostable, open source | ✅ AGPL-3.0 | ⚠️ on-prem Docker, proprietary |-| Code health score (1-10) | ✅ 53 detectors, 26 scoring | ✅ 25-30 |+| Code health score (1-10) | ✅ 53 detectors, 25 scoring | ✅ 25-30 || Brain Method / LCOM4 / god class | ✅ | ✅ || **Defects found at a 20% review budget** *([measured](docs/BENCHMARKS.md#5-code-health-predicts-defects), 2,770 files)* | ✅ **0.173** | 0.074 || **Effort-aware ranking, Popt** *(measured, p=0.003)* | ✅ **0.607** | 0.462 || **Precision at that budget** *(measured)* | 0.580 | ✅ **0.636**, a shorter list || **Discrimination, ROC AUC** *(measured, paired)* | 0.731 | 0.705, *p=0.054, not significant* |-| Defect-prediction AUC, published and reproducible | ✅ 0.737 over 21 repos, held-out 0.76-0.78 | ✅ Code Red study || Business impact (resolution time) | ❌ *we could not replicate this on open data* | ✅ Code Red study || Git intelligence (hotspots, ownership, co-change) | ✅ | ✅ || Pre-merge change-risk scoring | ✅ 0-10 + directives | ✅ |-| Health trend + declining alerts | ✅ rolling snapshots | ✅ |-| Bus factor analysis | ✅ | ✅ || Concrete cross-file refactoring plans | ✅ graph-aware + blast radius | ⚠️ within-function only |-| Dataflow-verified within-function plans | ✅ CFG + reaching definitions | ⚠️ LLM-generated, unverified || Test-coverage intelligence | ✅ LCOV/Cobertura/Clover/JaCoCo/Go | ❌ |-| Untested-hotspot detection | ✅ coverage × hotspot | ❌ || Dead code detection | ✅ | ❌ || Serves it to an AI agent over MCP | ✅ | ✅ |-| Local dashboard | ✅ | ✅ |-CodeScene is the only other vendor in this category with a published empirical-defect study, which is why it is the one we ran head to head against. It flags-about 27 files where we flag 132, so if you want a short list to action rather-than the ranking that catches the most defects, its threshold is the better fit.+CodeScene is the only other vendor in this category with a published empirical defect+study, which is why it is the one we ran head to head. It flags about 27 files where+we flag 132, so if you want a short list to act on, its threshold is the better fit.### Documentation generators-DeepWiki, Google Code Wiki and Swimm generate documentation from a repository,-which overlaps one of our five layers. **We have not measured against them**, so-there is no table here rather than a table of checkmarks. DeepWiki appears above-because it also serves an agent over MCP, which is a job we can be measured on.+DeepWiki, Google Code Wiki and Swimm generate documentation from a repository, which+overlaps one of our layers. **We have not measured against them**, so there is no+table here.-### The PR bot, against the LLM review bots+### The PR bot, against LLM review bots| | Repowise PR Bot | CodeRabbit | Greptile ||---|---|---|---|| LLM calls per PR | ✅ **zero** | ❌ every review | ❌ every review || Same diff, same review | ✅ deterministic | ❌ sampled output | ❌ sampled output || Your code sent to a model provider | ✅ never | ❌ yes | ❌ yes |-| Symbol-level blast radius (changed contracts → their callers) | ✅ call graph | ❌ | ⚠️ prose, from context |+| Symbol-level blast radius | ✅ call graph | ❌ | ⚠️ prose, from context || Co-change partners missing from the PR | ✅ git history | ❌ | ❌ |-| Change risk vs the repo's own distribution | ✅ 0-10 + percentile | ❌ | ❌ |-| Public analysis page per PR, no sign-in | ✅ | ❌ | ❌ |+| Change risk vs the repo's own history | ✅ 0-10 + percentile | ❌ | ❌ || Silent on a clean PR | ✅ by default | ⚠️ configurable | ⚠️ configurable |-| Cost on public repos | ✅ free, uncapped | ⚠️ free tier | ⚠️ free tier |-| Self-hostable | ✅ AGPL-3.0 | ❌ | ❌ |-The axis where this is not close is the first two rows. An LLM reviewer is a different-product with a different failure mode: it can read intent, and it can also be wrong in a-new way on every run. This one does set arithmetic over a call graph and a git history,-so there is nothing to hallucinate and nothing to prompt-inject, and pushing the same-diff twice produces the same review twice.--**Repowise is the intersection:** an agent-native context layer *and* behavioral git-intelligence *and* a defect-validated health score with the fix attached, all out of-one index, self-hostable and open source. Full side-by-side comparisons:-**[repowise.dev/compare →](https://www.repowise.dev/compare)**+An LLM reviewer is a different product: it can read intent, and it can be wrong in a+new way on every run. This one does set arithmetic over a call graph and a git+history, so pushing the same diff twice gives the same review twice. Full side-by-side+comparisons: **[repowise.dev/compare →](https://www.repowise.dev/compare)**</details>@@ -917,74 +733,112 @@ one index, self-hostable and open source. Full side-by-side comparisons:## For teams and enterprises-AI makes producing a change cheaper; it does not make understanding its consequences-cheaper. In a large estate, the answer crosses repositories, ownership boundaries,-service contracts, test suites, and years of architectural history. Repowise gives-developers, agents, reviewers, and platform teams the same evidence about what exists,-what depends on it, what is risky, and what will break.+AI makes producing a change cheaper. It does not make understanding its consequences+cheaper. In a large estate that answer crosses repositories, ownership boundaries,+service contracts, test suites and years of architectural history. Repowise gives+developers, agents, reviewers and platform teams the same evidence about what exists,+what depends on it, what is risky and what will break, from one index, in place of a+separate health tool, dead-code tool, code-search layer for agents, docs generator and+test-impact service.-That is the engineering reason to deploy it. The security reason is structural:-**graph, git, health, change risk, tests, dead code, and PR review make zero LLM-calls.** Documentation prose is optional and can use your provider contract or run-fully offline through Ollama.+### Where it runs, and what leaves your network-| Status | Enterprise capability |+| Deployment | Where source is read | Where the index lives | What leaves your network |+|---|---|---|---|+| **Self-hosted, open source** (`pip install repowise`) | your machine | `.repowise/` next to the repo | anonymous CLI telemetry you can switch off, and your own LLM provider only if you turn prose on |+| **Self-hosted, commercial** | your VPC or an air-gapped network | Postgres plus LanceDB or pgvector, inside your network | the same, plus integrations you configure. Nothing at all in air-gapped mode |+| **Hosted** ([repowise.dev](https://www.repowise.dev)) | the hosted indexer | infrastructure we operate | your code goes to the platform |++Graph, git, health, change risk, tests, dead code and PR review make **zero LLM calls**.+Raw source is parsed in memory and never persisted. Prose is optional and runs on your+own provider contract or fully offline through Ollama, chosen per repository. The+threat model and data flows are in the+[security review pack](docs/business/SECURITY_COMPLIANCE.md). The published research each layer is built on, and how we+checked it, is in [LINEAGE.md](docs/LINEAGE.md).++### How accurate, and how big++Call-graph accuracy is graded against each language's own compiler toolchain, and we+publish precision and recall per language, with the misses, in the+[accuracy table](docs/BENCHMARKS.md#accuracy-by-language). On current main, precision+runs 0.95 to 0.99 for Go, TypeScript, C and most Python, 0.85 to 0.90 for Rust, and+0.73 to 0.92 for Java and C#, where overloaded methods are the main gap.++The largest repository indexed so far is dotnet/runtime: 58,924 files in 120 minutes+at 11.7 GiB peak memory, on one 31 GB laptop with no API key. Time and memory by+repository size: [scale](docs/BENCHMARKS.md#scale).++### Source control and CI++Any git host works, because indexing reads a local checkout. On top of that:++- **GitHub**: the [PR bot](#the-pr-bot) (GitHub App), a GitHub Action for the CI+ gates, and webhook auto-sync.+- **GitLab**: a CI template for the same gates and webhook auto-sync.+- **Bitbucket Pipelines and others**: the gates run as plain CLI commands with the+ pipeline's own branch variables. ([Repowise in CI](docs/start/CI.md))+- **Managed integrations** for GitHub Enterprise, Azure DevOps, GitLab self-managed+ and Bitbucket are rolling out.+- **Not on git?** Point `repowise init` at a plain directory, an export, or a Perforce+ or SVN workspace. The graph, docs, decisions and health layers build normally; the+ history layer needs a commit log, and Perforce and SVN adapters are+ [on the roadmap](ROADMAP.md#source-control-beyond-git).++### What is shipping, and what is not yet++| Status | Capability ||---|---|-| **Shipping now** | Five deterministic intelligence layers, ten MCP tools, multi-repo workspaces, contract extraction and blast radius, test intelligence, architecture conformance, local dashboard, auto-sync, and full-history secret scanning. |-| **GA commercially** | Hosted graph-aware security, CVE prioritization, CycloneDX SBOM and VEX, PCI-DSS and SOC 2 evidence reports, audit exports and webhook stream, Jira and Confluence, customer-infrastructure HA topology, custom extensions, SLA support, and IP indemnification. |-| **Rolling out** | GitHub Enterprise, Azure DevOps, GitLab and Bitbucket integrations; SAML/OIDC SSO and SCIM; engineering-leader dashboards. |-| **Planned** | RBAC and multi-tenancy, packaged air-gap install bundle, and the Helm chart. |--Self-host with `pip install` or run the API, workers, dashboard, Postgres, and-LanceDB/pgvector containers on your infrastructure. Deterministic analysis needs no-provider. When optional prose is enabled, provider choice is per repository. Stored-data includes the graph, embeddings, wiki pages, and git metadata; raw source is-processed transiently and is not persisted.--**Past one repository.** Workspaces index an estate as one unit: API contracts-matched producer to consumer so a breaking change is caught before it ships,-cross-repo co-change, and one federated MCP endpoint that answers across all of-it. *(Estate-scale dashboards: [in development](ROADMAP.md#multi-repo-and-workspace).)*--**Not on git?** Only the history layer needs a commit log. Point `repowise init`-at a plain directory, an export, or a Perforce or SVN workspace and the graph,-documentation, decisions and code-health layers all build normally; what is-missing is hotspots, ownership, co-change and bug history until the history layer-learns to read your system.-*([Perforce, SVN, Endevor and ChangeMan on the roadmap →](ROADMAP.md#source-control-beyond-git))*--The complete capability matrix is maintained in-[COMMERCIAL.md](docs/business/COMMERCIAL.md#4-commercial-capabilities-at-a-glance),-with every item labelled GA, rolling out, in development, or planned.--[**repowise.dev**](https://www.repowise.dev) runs the same engine fully managed, at-feature parity with self-hosted. We run it on our own codebase in the open:-[live snapshot →](https://www.repowise.dev/s/5a6b93fa9a69) ·-[explore public repos →](https://www.repowise.dev/explore).--**[Commercial detail and pricing models →](docs/business/COMMERCIAL.md)** ·-**[Security review pack →](docs/business/SECURITY_COMPLIANCE.md)** ·-**[Roadmap →](ROADMAP.md)** ·-[[email protected]](mailto:[email protected]) ·-[[email protected]](mailto:[email protected])+| **Shipping in open source** | Every deterministic layer, ten MCP tools, multi-repo workspaces, contract extraction and cross-repo blast radius, test intelligence, [architecture conformance](docs/scale/WORKSPACES.md#architecture-conformance), local dashboard, auto-sync, full-history secret scanning. |+| **GA commercially** | Hosted graph-aware security, CVE prioritization, CycloneDX SBOM and VEX, PCI-DSS and SOC 2 control-coverage reports, audit export and webhook stream, Jira and Confluence, reference HA topology on your infrastructure, custom language extensions, SLA support, IP indemnification. |+| **Rolling out** | Managed GitHub Enterprise, Azure DevOps, GitLab and Bitbucket integrations. SAML/OIDC SSO and SCIM. Engineering-leader dashboards. |+| **Planned** | RBAC and multi-tenancy, a packaged air-gap install bundle, the Helm chart. |++Repowise holds no SOC 2, ISO 27001 or other audited certification today. The SOC 2 and+PCI-DSS reports above are control-coverage signals from your own findings, and every+export says so. Every item is tracked with its status in the+[capability matrix](docs/business/COMMERCIAL.md#4-commercial-capabilities-at-a-glance).++### Pricing, licence and support++Commercial licences are priced **per indexed repository**, with unlimited seats inside+the licensed set. More and more of the code in a repository is written and read by+agents and CI, so seat counts stop tracking the value.+[Details →](docs/business/COMMERCIAL.md#7-licensing--pricing)++**Do we have to open-source our code because of the AGPL?** No. Using Repowise inside+your company, including on internal servers, creates no obligation to publish your+code. The obligations apply if you modify Repowise and offer it to others over a+network, or ship it inside your own product. A commercial licence removes them.++Commercial support comes with a named contact, a response-time SLA and a quarterly+architecture review. Security fixes land on the latest minor release; reports go to+[[email protected]](mailto:[email protected]) under the+[disclosure policy](.github/SECURITY.md).++**[Commercial detail](docs/business/COMMERCIAL.md)** ·+**[Security review pack](docs/business/SECURITY_COMPLIANCE.md)** ·+**[Roadmap](ROADMAP.md)** ·+[[email protected]](mailto:[email protected])++[repowise.dev](https://www.repowise.dev) runs the same engine fully managed. We run it+on our own codebase in the open:+[live snapshot](https://www.repowise.dev/s/5a6b93fa9a69) ·+[explore public repos](https://www.repowise.dev/explore).---## Privacy- **Deterministic or offline mode:** with `--no-prose`, code-derived content stays on- your infrastructure. The CLI reports **anonymous, opt-out** usage telemetry- (command names and coarse environment only); disable it with `repowise telemetry- disable`, `DO_NOT_TRACK=1`, or by running fully offline.+ your infrastructure. The CLI reports **anonymous, opt-out** usage telemetry (command+ names and coarse environment only); turn it off with `repowise telemetry disable`,+ `DO_NOT_TRACK=1`, or by running fully offline.[What's collected →](docs/reference/TELEMETRY.md)-- **Optional LLM features:** generated prose, decision extraction and code-generating- refactoring can send code-derived prompts directly to the provider configured with- your own key. Repowise does not proxy those calls; provider handling and retention- follow your account and provider terms.-- **What's stored:** the graph, embeddings, generated wiki pages, and git metadata.- Raw source is processed transiently and never persisted. See the- [security review pack](docs/business/SECURITY_COMPLIANCE.md) for the threat model- and data-flow boundaries.+- **Optional LLM features:** generated prose, decision extraction and refactoring code+ generation send code-derived prompts directly to the provider configured with your+ own key. Repowise does not proxy those calls.+- **What's stored:** the graph, embeddings, generated wiki pages and git metadata. Raw+ source is processed transiently and never persisted.- **Fully offline:** Ollama plus a local embedding model means zero external calls.Doing a security review? **[docs/business/SECURITY_COMPLIANCE.md →](docs/business/SECURITY_COMPLIANCE.md)**@@ -994,27 +848,28 @@ Doing a security review? **[docs/business/SECURITY_COMPLIANCE.md →](docs/busin## CLI```bash-repowise init [PATH] # index a codebase (one-time; asks, or --no-prose -y needs no LLM)-repowise generate [PATH] # write wiki pages with a model, on demand (upgrade a keyless wiki)+repowise init [PATH] # index a codebase (asks; --yes --no-prose needs no key)+repowise generate [PATH] # write wiki pages with a model, on demandrepowise serve [PATH] # MCP server + local dashboard-repowise update [PATH] # incremental update (seconds; --workspace for every repo)-repowise watch # auto-sync daemon, re-index on file change+repowise update [PATH] # incremental update (--workspace for every repo)+repowise watch # re-index on file changerepowise search "<q>" # hybrid search (fulltext / semantic / symbol / path)repowise ask "<q>" # a synthesized answer with citationsrepowise context <files> # triage card: layer, hotspot, fix history, freshnessrepowise symbol <id> # one symbol's body, with verified line boundsrepowise why <q|path> # decisions, rationale, git archaeology+repowise next # what to fix firstrepowise health # code-health KPIs and lowest-scoring files-repowise risk main..HEAD # score a branch or PR range for defect risk-repowise risk -t <file> # what history says about touching a file-repowise impacted-tests # only the tests a diff actually exercises-repowise dead-code # unreachable-code report+repowise risk main..HEAD # score a branch or PR range+repowise overlap # other branches editing the same files+repowise impacted-tests # only the tests a diff exercises+repowise dead-code # unreachable code by confidence tierrepowise doc-drift # documentation the code no longer supports+repowise security # secrets and risky patterns, tree or full historyrepowise decision list # architectural decisions-repowise export --format structurizr # the architecture as Structurizr DSL, no LLM+repowise export --format structurizr # the architecture as Structurizr DSLrepowise distill pytest # compact, errors-first, reversible command outputrepowise saved # tokens and dollars saved by distillation-repowise savings # the full agent savings ledger, priced per eventrepowise workspace add # multi-repo workspace managementrepowise doctor # check setup, API keys, index driftrepowise uninstall # remove what repowise wrote, and say what it left@@ -1022,7 +877,8 @@ repowise uninstall # remove what repowise wrote, and say what it leftEvery command and flag: **[docs/reference/CLI_REFERENCE.md](docs/reference/CLI_REFERENCE.md)** ·config: **[docs/reference/CONFIG.md](docs/reference/CONFIG.md)** ·-examples: **[examples/](examples/)**+something not working: **[docs/start/TROUBLESHOOTING.md](docs/start/TROUBLESHOOTING.md)** ·+all docs: **[docs/README.md](docs/README.md)**---@@ -1037,7 +893,7 @@ uv run pytest tests/unit/```New here? You do not have to read 3,000 files to start. We keep a public index of this-repo built by repowise itself, re-indexed on every push:+repository built by Repowise itself, re-indexed on every push:[**explore repowise with repowise →**](https://repowise.dev/repo/repowise-dev/repowise)(architecture, hotspots, ownership, decisions, and a ranked[refactoring backlog](https://repowise.dev/repo/repowise-dev/repowise/refactoring) you@@ -1051,10 +907,10 @@ Full guide, including how to add languages and LLM providers:## License-AGPL-3.0. Free for individuals, teams and companies using repowise internally.+AGPL-3.0. Free for individuals, teams and companies using Repowise internally.For commercial licensing (the enterprise security and compliance layer, SSO/SCIM, RBAC,-workflow integrations, priority support and SLA, or embedding repowise in a product+workflow integrations, priority support and SLA, or embedding Repowise in a productwithout AGPL obligations), see**[docs/business/COMMERCIAL.md](docs/business/COMMERCIAL.md)** or contact[[email protected]](mailto:[email protected]).@@ -1065,7 +921,7 @@ without AGPL obligations), see<em>Built for engineers who got tired of watching their AI agent <code>cat</code> the same file for the fourth time.</em>-<p align="center"><sub>⭐ If repowise earns a place in your workflow, <strong>give it a star</strong>. It costs you nothing, and it's the signal that keeps a small team building this in the open.</sub></p>+<p align="center"><sub>⭐ If Repowise earns a place in your workflow, <strong>give it a star</strong>. It costs you nothing, and it's the signal that keeps a small team building this in the open.</sub></p><p align="center"><a href="https://repowise.dev"><strong>repowise.dev</strong></a> ·diff --git a/docs/BENCHMARKS.md b/docs/BENCHMARKS.mdindex bbbcc3e15..067d75fe7 100644--- a/docs/BENCHMARKS.md+++ b/docs/BENCHMARKS.md@@ -1,189 +1,222 @@# Benchmarks-Every number repowise publishes, with its sample size, its test, and a link to the-raw data. Nothing here is measured on a private corpus. The harnesses, the-pre-registrations and every graded cell live in-**[repowise-bench](https://github.com/repowise-dev/repowise-bench)**, which is also-where the depth is: this page is the summary, that repository is the evidence.+Repowise indexes a repository once, on your machine, into a call graph, git+history, code health, test links, decisions and documentation, and serves that+index to developers and their coding agents. Every number below was measured on+public repositories at pinned commits, against answer keys and samples we+publish in **[repowise-bench](https://github.com/repowise-dev/repowise-bench)**.+Where another tool beats us, the row is on this page too. The research each method comes from is listed in+[LINEAGE.md](LINEAGE.md).-Each section leads with the result; method and caveats are one fold down, and-nothing put there changes a headline. Where we have not run a comparison, the row-says **not measured** rather than carrying a checkmark.+## At a glance-## The scoreboard+### The headline numbers-**Six of the thirteen rows below are wins against a named competitor. Two are-losses, one is level, three are not measured, and one is a ratio against no tool-at all.** They sit in one table on purpose: a page that reports only its wins is-marketing with a sample size attached.+| Claim | Number | Sample | Measured on |+|---|---|---|---|+| [Call graph checked by the compiler](#8-the-same-question-against-an-answer-key-we-do-not-control) (Go, TypeScript) | No tool that finds as much of the graph gets more of it right, in **7 of 7** cells. Our precision **0.976 to 0.995** | 7 cells, 37,853 compiler edges, 5 tools | repowise main `4fccb31ec`, 2026-10-03; competitors August 2026 |+| [Call graph accuracy by language](#accuracy-by-language) | Precision: C **0.95 to 0.98**, Python **0.96 to 0.98** (attrs 0.50), Rust **0.85 to 0.90**, C# **0.73 to 0.92**, C++ **0.75**, Java **0.73 to 0.78** | 16 repositories, 6 languages, recall beside every figure | repowise main `4fccb31ec`, 2026-10-03 |+| [Finding the right files](#1-finding-the-right-files) | **0.876** of the needed files found, against **0.610** for the next tool | 42 sealed instances, p = 0.00004 | repowise `081a59fa`, August 2026 |+| [Work saved in an agent loop](#2-what-changes-in-a-real-agent-loop) | **31.6% less** agent output than no tool, 3.8 tool calls against 7.2 | 43 django questions on Codex, p < 0.0001 | repowise `081a59fa`, August 2026 |+| [Health score against CodeScene](#5-code-health-predicts-defects) | **2.3x** the later-defective files inside a 20% review budget | 2,770 files, 21 repositories, p = 0.003 | May 2026 |+| [Which tests cover a file, with no coverage run](#test-intelligence) | **95.7%** precision on what reaches a file, **97.5%** on the list of tests to run | 46 claims and 47 targets, checked against per-test coverage | this repository, August 2026 |+| [Largest repository indexed](#scale) | **58,924 files** (dotnet/runtime) in 120 minutes at 11.7 GiB peak | one keyless run on a 31 GB laptop | pre-release build, 2026-10-02; its fixes are on main |+| [Memory to build the call graph](#what-it-costs-to-run) | **75 MB** median against 757 MB for the next tool, lowest on **35 of 35** repositories | 35 repositories, 5 tools | August 2026 |++The rows we lose are below with the rest: two tools draw a bigger call graph+than we do, CodeScene keeps a shorter and slightly more precise review list,+and a full index takes much longer than a graph-only tool because it builds far+more. Each one sits in its section with its numbers.++### One index, the work of several tools++Teams usually assemble these from separate products. Repowise builds them from+one local index, and most of it needs no model key.++| Category | Tools teams use today | What repowise does | Evidence |+|---|---|---|---|+| Code health | CodeScene, SonarQube maintainability rating | 53 detectors score every file on defect risk, maintainability and performance, then rank what to fix first | [Measured](#5-code-health-predicts-defects): predicts later defects, 2.3x CodeScene's recall at a fixed review budget |+| Dead code | Knip, vulture, ts-prune | Unreachable files, unused exports and unused packages, each with a confidence tier and its evidence, across languages | Capability, not measured on this page |+| Code graph and search for agents | Sourcegraph, CodeGraph, Serena | Call graph, hybrid search and answers served to agents over MCP | [Measured](#1-finding-the-right-files) against CodeGraph and Serena: retrieval, agent loop, graph accuracy. Sourcegraph not measured |+| Documentation | DeepWiki, Swimm | A generated wiki per module and file (needs a model key), plus checks that flag docs the code no longer supports (no key) | Capability, not measured |+| Change risk and test impact | Codecov gates, test-impact tools | Where a change ranks among the repository's recent commits, which tests to run, coverage gates, all in the GitHub Action | [Measured](#test-intelligence): test selection precision. Risk ranking: capability, not measured here |+| Decision records | ADR files kept by hand | Finds decisions in ADRs, comments, commits and PR bodies, ties each to the code it governs, flags stale ones | Capability, not measured |+| Multi-repo contracts | Contract tests, hand-kept API catalogs | Matches HTTP, gRPC, message-topic and data contracts between repositories in a workspace, and flags breaking changes | Capability, not measured |++Repowise does not replace a security scanner such as Semgrep or an AI code+reviewer such as CodeRabbit. In open source, the pull-request surface is the+GitHub Action (gates, annotations, SARIF, GitLab Code Quality). The PR comment+bot is a separate free hosted GitHub App.-| Question | Measured against | Result |-|---|---|---|-| **Are the call edges true, judged by a compiler** | CodeGraph, codebase-memory-mcp, Graphify, code-review-graph | [**we win**](#is-the-call-graph-correct). No tool that finds as much of the graph gets more of it right, in all 7 cells |-| **Are the call edges true, hand-graded from source** | CodeGraph | [**we win**](#7-edge-precision), 85.7% against 58.6%, 560 rows read on both sides |-| **Finding the right files** | CodeGraph, Graphify, code-review-graph, cocoindex | [**we win**](#1-finding-the-right-files), n=42 held out, p=0.00004 |-| **Work saved in a real agent loop** | the same field plus Serena and a bare agent | [**we win**](#2-what-changes-in-a-real-agent-loop) on all three agent harnesses we tried, n=43 at p<0.0001 |-| **Does the health score predict real bugs** | CodeScene | [**we win**](#5-code-health-predicts-defects) on recall and effort-aware ranking, p=0.003, and lose the business-impact axis outright |-| **Memory to build the graph** | the same four graph tools | [**we win**](#what-it-costs-to-run), lowest of five tools on **35 of 35** repositories, about 10x lower than the next |-| **Loading one commit's context** | naive file reads, `git diff` | [**35.6x**](#loading-one-commits-context-the-easy-number) fewer tokens than naive |-| **Speed to build the graph** | the same four graph tools | [**level**](#what-it-costs-to-run). Fastest on 14 of 35 repositories, CodeGraph on 16 |-| **How much of the call graph we find** | the same four graph tools | [**we lose**](#is-the-call-graph-correct). Two tools draw a bigger map, and we publish what their extra edges cost them |-| **Time to build the full index** | CodeGraph, Graphify, code-review-graph | [**we lose**](#what-it-costs-to-run), 22x, because we build five layers where they build one |-| **Command-output compression** | RTK | [**not measured**](#command-output-compression) head to head |-| **Documentation generation** | DeepWiki, Google Code Wiki, Swimm | **not measured** |-| **PR review** | CodeRabbit, Greptile | **not measured** |--**The two cost rows are different questions and are the easiest thing here to-misread.** Building the call graph, we are the lightest tool measured and about as-fast as the fastest. Building the *whole index*, we are 22x slower than a tool-that only builds a call graph, because by then we have also built the git history,-the wiki, the decisions and the health pass. The **not measured** rows are-capability comparisons rather than measurements, and live in the-[README's feature table](../README.md#how-it-compares-on-capability); we would-rather write that than let a checkmark do a number's job.+----<div align="center">-<picture>- <source media="(prefers-color-scheme: dark)" srcset="../.github/assets/bench/file-coverage-dark.svg" />- <img src="../.github/assets/bench/file-coverage.svg" alt="File coverage on 42 sealed ContextBench instances: repowise get_answer 0.876, repowise search_codebase 0.742, CodeGraph 0.610, Graphify 0.546, code-review-graph 0.445, cocoindex 0.361" width="100%" />-</picture>-<picture>- <source media="(prefers-color-scheme: dark)" srcset="../.github/assets/bench/agent-output-tokens-dark.svg" />- <img src="../.github/assets/bench/agent-output-tokens.svg" alt="Output tokens an agent writes to reach an answer across 43 django questions on Codex: repowise 1,250, CodeGraph 1,383, Serena 1,550, Graphify 1,658, code-review-graph 1,710, bare agent 1,828 baseline" width="100%" />-</picture>-</div>+<a id="by-language"></a>+<a id="accuracy-by-language"></a>++## Accuracy by language++A call graph is a map of which function calls which. It is what lets a tool say+what breaks when you change something. We check ours against each language's+own compiler: the compiler resolves every call site, and we count how many of+our edges it confirms (**precision**) and how many of its edges we found+(**recall**). We did not write any of these answer keys and cannot tune them.++Repowise main `4fccb31ec`, freshly indexed from source for every cell,+2026-10-03. No competitor was run in this section; the five-tool comparison is+the [next section](#8-the-same-question-against-an-answer-key-we-do-not-control).++| Language | Repository | Set | Answer key | Compiler edges | Precision | Recall |+|---|---|---|---|---:|---:|---:|+| Go | gitleaks | dev | scip-go (Go compiler) | 1,596 | **0.996** | 0.985 |+| TypeScript | zod | dev | `tsc` | 1,269 | **0.995** | 0.796 |+| TypeScript | hono | dev | `tsc` | 706 | **0.976** | 0.806 |+| Java | jsoup | dev | scip-java (javac) | 4,329 | 0.781 | 0.682 |+| Java | gson | dev | scip-java (javac) | 1,760 | 0.737 | 0.701 |+| Java | javapoet | held out | scip-java (javac) | 698 | 0.725 | 0.695 |+| C# | FluentValidation | dev | scip-dotnet (Roslyn) | 629 | **0.916** | 0.506 |+| C# | Polly (4 projects) | dev | scip-dotnet (Roslyn) | 1,172 | 0.731 | 0.200 |+| C# | AutoMapper | held out | scip-dotnet (Roslyn) | 3,157 | 0.863 | 0.321 |+| C | redis | dev | scip-clang (clang) | 34,349 | **0.950** | 0.709 |+| C | git | held out | scip-clang (clang) | 28,118 | **0.979** | 0.830 |+| C++ | fmt | dev | scip-clang (clang) | 1,596 | 0.750 | 0.368 |+| Rust | ripgrep | dev | rust-analyzer | 4,555 | 0.852 | 0.532 |+| Rust | fd | dev | rust-analyzer | 231 | 0.899 | 0.619 |+| Python | httpx | dev | jedi (not a compiler) | 1,753 | **0.983** | 0.696 |+| Python | click | dev | jedi (not a compiler) | 2,246 | **0.958** | 0.717 |+| Python | typer | held out | jedi (not a compiler) | 2,099 | **0.960** | 0.782 |+| Python | attrs | held out | jedi (not a compiler) | 1,328 | 0.500 | 0.368 |++**Dev and held out.** Dev repositories are the ones accuracy fixes were+developed against. Held-out repositories were chosen in advance and never used+to pick a fix, so they are the fairer read of what you will see on your own+code. On every held-out repository except attrs, main is more precise than the+released 0.54.0 build was: javapoet by 30 points, AutoMapper by 4, typer by 2.+On git it is level within 0.2 points while recall rose 21 points.++**How it is scored.** An edge matches when the caller and the callee land on the+same declaration line the compiler names. The SCIP indexers come from each+language's own compiler toolchain. Python has no compiler that resolves calls,+so the reference is jedi, a static analyser: it is the closest available+reference, and a call that resolves only at run time is outside what it can+see. Every cell carries a 95% confidence interval in the published artifacts.++**Limits, stated plainly:**++- **C# Polly lost precision.** It fell from 0.774 on 0.54.0 to 0.731 on main,+ while recall rose from 0.125 to 0.200. Constructor calls now become edges, and+ repowise keeps one graph node per overload set, so a call to the right+ constructor can land on the wrong overload and the compiler counts it as+ wrong. Polly is also measured on 4 of its projects only.+- **Python attrs is low at 0.50 precision.** attrs exports its API through+ chained module aliases (`s = attributes = attrs`), and calls through those+ aliases bind to the alias line. That gap is open.+- **Overloads explain much of the Java and C# gap.** Counting a call that lands+ on any overload of the right method as correct, Java reads 0.76 to 0.80 and C#+ 0.83 to 0.93.+- **Recall is below 0.80 almost everywhere** outside Go and TypeScript. Dynamic+ dispatch is the main reason, as the next section shows for Go.+- **Not reported:** TypeScript through SCIP (that method did not agree with+ `tsc`, so TypeScript comes from `tsc` directly), Kotlin (not run), C++ and+ Rust held-out repositories (no answer key could be built), Ruby and PHP (no+ answer key exists).+- Do not compare recall across languages. It depends on how many entry points a+ codebase has and how much it dispatches dynamically.++Reference versions: scip-go 0.2.7, scip-java 0.13.1, scip-dotnet 0.2.14,+scip-clang 0.3.1, rust-analyzer 1.96.0, jedi 0.20.0. Per-repository scores,+confidence intervals and the scoring code:+**[graph/](https://github.com/repowise-dev/repowise-bench/tree/master/graph)**.---<a id="8-the-same-question-against-an-answer-key-we-do-not-control"></a><a id="is-the-call-graph-correct"></a>-## Is the call graph correct+## Call graph against a compiler, five tools-A call graph is a map of which function calls which. It is what lets a tool answer-"what breaks if I change this". There are exactly two ways to get one wrong: **miss-calls that are real**, or **invent calls that are not there**. Every tool trades one-against the other, and either number on its own is trivially gamed. Draw an edge-between everything and you find 100% of the real calls. Draw one edge you are sure-of and 100% of your edges are correct.+There are two ways to get a call graph wrong: miss calls that are real, or+invent calls that are not there. Either number alone is easy to game. Draw an+edge between everything and you find every real call. Draw one edge you are+sure of and every edge you drew is correct. So we report the pair.-So the pair is the only honest reading, and this is the pair. **Precision** is the-share of a tool's edges that are real. **Recall** is the share of the real call-graph it found.+The answer key is the **Go team's own call graph** (`golang.org/x/tools`, RTA,+over the fully type-checked program) and, on TypeScript, **`tsc`'s own+resolution** of every call site. Anyone with the toolchain can regenerate both.+Five tools, seven cells, **37,853 compiler edges**. Each cell is+precision / recall.-| cell | repowise | CodeGraph 1.5.0 | codebase-memory-mcp 0.10.8 | Graphify 0.9.31 | code-review-graph 2.3.7 |+| Cell | repowise main (2026-10-03) | CodeGraph 1.5.0 | codebase-memory-mcp 0.10.8 | Graphify 0.9.31 | code-review-graph 2.3.7 ||---|---|---|---|---|---|-| cobra (with tests) | 0.972 / 0.684 | 0.929 / 0.763 | 0.912 / 0.743 | 0.971 / 0.433 | 0.997 / 0.174 |-| gitleaks (no tests) | 0.976 / 0.955 | 0.972 / 0.920 | 0.934 / 0.967 | 0.997 / 0.886 | 0.759 / 0.026 |-| gitleaks (with tests) | 0.974 / 0.914 | 0.971 / 0.895 | 0.922 / 0.945 | 0.995 / 0.832 | 0.800 / 0.032 |-| syft (no tests) | 0.943 / 0.513 | 0.872 / 0.508 | 0.635 / 0.542 | 0.771 / 0.447 | 0.968 / 0.201 |-| syft (with tests) | 0.950 / 0.322 | 0.864 / 0.338 | 0.673 / 0.361 | 0.802 / 0.273 | 0.966 / 0.086 |-| zod (no tests) | 0.992 / 0.703 | 0.729 / 0.373 | 0.987 / 0.694 | 0.825 / 0.248 | 0.932 / 0.652 |-| hono (no tests) | 0.977 / 0.731 | 0.805 / 0.684 | 0.949 / 0.686 | 0.980 / 0.688 | 0.966 / 0.691 |--> ### **In all seven cells, no tool that finds as much of the call graph as we do gets more of it right.**--**Nobody is above us on both numbers, anywhere.** Read across any row: the tools-scoring higher than us on precision are below us on recall, every time, and the-tool above us on recall is below us on precision, every time.--That claim names no threshold, which is what makes it worth something. It cannot be-tuned by picking a cutoff, and adding a competitor can only break it. **Two-competitors were added after it was first written and it held.**--**The answer key is not ours.** Every graph-quality number in this field, ours-included, is normally scored against something the publisher controls, so a tool-that is confidently wrong in a consistent way scores well and a reader has no way-to check. Here the key is the **Go team's own call graph** from-`golang.org/x/tools/go/callgraph/rta` over the fully type-checked program, and on-TypeScript the **`tsc` type checker's own resolution** of every call site. We did-not write either, we cannot tune either, and anyone with the toolchain can-regenerate both. Five tools, seven cells, **37,853 oracle edges**.--### What we lose, at full size--**We do not draw the biggest map.** We lead recall in the two TypeScript cells and-in **none of the five Go cells**. On cross-file coverage across 35 repositories,-codebase-memory-mcp separates from us on 15 and we separate on none. Those are real-losses and they are why the row above says we lose that column.--**We are the most precise tool outright in only one cell of seven**, tied in one-more, and beaten in five. Against the two tools this experiment started with,-CodeGraph and codebase-memory-mcp, we are the most precise in seven of seven, and-that narrower claim should always carry its label.--**What the oracle adds is the price of a bigger map, not an excuse for ours.** On-syft, more than a third of what the coverage leader emits is a call the Go-compiler says does not exist, and the highest precision anywhere on this page,-0.997, comes from a graph holding **17% of the calls in the repository**. A graph-that small is very hard to be wrong in and not much use to walk.--**Graphify's result is the one worth taking seriously** and should not be lumped in-with that: 89% recall against our 95% on gitleaks is a narrow, real trade.+| cobra (with tests) | **0.995** / 0.705 | 0.929 / 0.763 | 0.912 / 0.743 | 0.971 / 0.433 | 0.997 / 0.174 |+| gitleaks (no tests) | **0.991** / **0.971** | 0.972 / 0.920 | 0.934 / 0.967 | 0.997 / 0.886 | 0.759 / 0.026 |+| gitleaks (with tests) | **0.989** / 0.939 | 0.971 / 0.895 | 0.922 / 0.945 | 0.995 / 0.832 | 0.800 / 0.032 |+| syft (no tests) | **0.979** / 0.539 | 0.872 / 0.508 | 0.635 / 0.542 | 0.771 / 0.447 | 0.968 / 0.201 |+| syft (with tests) | **0.982** / 0.341 | 0.864 / 0.338 | 0.673 / 0.361 | 0.802 / 0.273 | 0.966 / 0.086 |+| zod (no tests) | **0.995** / **0.796** | 0.729 / 0.373 | 0.987 / 0.694 | 0.825 / 0.248 | 0.932 / 0.652 |+| hono (no tests) | **0.976** / **0.806** | 0.805 / 0.684 | 0.949 / 0.686 | 0.980 / 0.688 | 0.966 / 0.691 |++The competitor columns are August 2026 measurements at the pinned versions+above; they were not re-run. The repowise column is main, freshly indexed, and+its recall is higher than our August figure in all seven cells.++> **In all seven cells, no tool that finds as much of the call graph as we do+> gets more of it right.**++Read across any row. Every tool more precise than us finds less of the graph,+and every tool that finds more is less precise. The claim names no threshold,+so it cannot be tuned by picking a cutoff, and adding a competitor can only+break it. Two competitors were added after it was first written, and it held.++**Where we lose.** We are not the tool that draws the biggest map. We lead+recall in three cells (both TypeScript cells and gitleaks without tests) and+trail in four. Across 35 repositories of cross-file coverage, measured in August,+codebase-memory-mcp is ahead of us on 15 and we are ahead on none. We are the+most precise tool outright in three cells; Graphify is more precise on both+gitleaks cells and on hono, and code-review-graph on cobra.++**What a bigger map costs.** On syft, more than a third of what the coverage+leader emits is a call the Go compiler says does not exist. The highest+precision anywhere in the table, 0.997, comes from a graph holding 17% of the+calls in the repository. Graphify's trade is a real one: 89% recall against our+97% on gitleaks, at slightly higher precision.<details>-<summary><b>Method, limits, and what this does not show</b></summary>--**The two methods agree.** On Go, the hand-graded audit below read 29/30 = 96.7%-for us and 29/30 = 96.7% for CodeGraph. The Go compiler, over roughly 1,600 edges-on the same repository, says **97.6% and 97.2%**. Two unrelated methods, one person-reading source and one type checker, land within about a point on both arms. That-is the result we care about most, and it is not a competitive one. It also removes-the n=30 ceiling: hand-grading caps a precision estimate at roughly ±13 points near-a 95% rate, while an oracle grades every edge, so n becomes the size of the-repository.--**Where our own miss goes**, decomposed rather than waved at. On syft without tests-we miss 3,846 of the oracle's 7,898 edges: **44% of that miss is dynamic dispatch-alone** and a further **39% is dispatch with a closure at one end**, and the two-buckets overlap so neither is the whole gap. Interface dispatch is the ceiling and-nobody in the comparison has cleared it, at 6.5 distinct possible targets per call-site. Matching that recall means emitting six edges where one is right, which is-the behaviour the precision table charges other tools for.--**Do not compare recall across rows.** It swings from 0.03 to 0.97, driven by how-many entry points the oracle had (4 on gitleaks, 268 on syft-with-tests), not by-tool quality. Only within-row comparisons carry meaning, and a pooled recall over-these cells would be meaningless.--Further limits:--- **Two languages, seven cells, five repositories.** This is not a nine-language- claim and must not be quoted as one. The hand-graded audit below is the- nine-language number.-- **A contradicted edge is very strong evidence, not proof.** RTA is unsound under- reflection and `go:linkname`, so a genuinely dynamic edge can land in that- bucket. It applies to all five arms equally and the gaps are far too large to be- explained by it, which is why the metric is named *precision against the oracle*- rather than precision.-- **Edges the oracle cannot speak about are charged to nobody** and reported at- full size. That bucket is 0.4% to 11.1% of a tool's output on the Go cells and- 16% to 46% on the TypeScript ones, where dependencies are not installed in the- pinned corpus.-- **A library has no `main`, so RTA has no roots**; cobra is analysed through its- test binaries only, and the two variants of a repository answer different- questions, so report both or say which you used.-- **Two oracle languages, and the programme stops at two.** The other corpus- languages either need a per-repository build that does not exist on the- measurement machine, have no sound call-graph tool (Rust), or admit no oracle- even in principle (Python, Ruby, PHP). The hand-graded audit below is the- permanent method there, not a stopgap.-- **The TypeScript with-tests variants are void and quoted nowhere.** The pinned- corpus installs no dependencies, so a third of call sites go unresolvable.-- **The three newer competitors are priced on these two languages only**, and- none was ever entered into the hand-graded audit in either direction. That- audit is a two-tool claim and this one a five-tool claim; neither transfers to- the other's languages.-- **Two of the five arms are read through an adapter we wrote**, and both now- beat us on precision somewhere, so the reading matters. Graphify tags 93% of- its call edges `INFERRED` and we score all of them, the choice least- favourable to us. code-review-graph is scored on resolved rows only, the- choice most favourable to it. Both are argued in the benchmark and either can- be recomputed from the artifacts.--Full method, per-cell artifacts, the twenty hand-confirmed identities the protocol-required, and the graded pre-registration including the two predictions that-missed:+<summary><b>Method and limits</b></summary>++**Two methods agree.** On Go, the hand-graded audit below read 29/30 for us and+29/30 for CodeGraph. The compiler, over about 1,600 edges on the same+repository in August, read 97.6% and 97.2%. A person reading source and a type+checker land within about a point on both tools.++**Where our misses go.** On syft without tests, in August, 44% of our missed+edges were dynamic dispatch and a further 39% dispatch with a closure at one end+(the buckets overlap). Interface dispatch averages 6.5 possible targets per call+site there. Matching the coverage leader's recall means emitting six edges where+one is right, which is the behaviour the precision column charges for.++**Do not compare recall across rows.** It runs from 0.03 to 0.97, driven by how+many entry points the compiler had (4 on gitleaks, 268 on syft with tests), not+by tool quality.++- **Two languages, seven cells, five repositories.** This five-tool comparison+ is not a claim about other languages. The [per-language table](#accuracy-by-language)+ covers six more for repowise alone, and the hand-graded audit below is the+ nine-language comparison against another tool.+- **A contradicted edge is strong evidence, not proof.** RTA is unsound under+ reflection and `go:linkname`. That applies to every tool equally, and the gaps+ are far larger than it could explain.+- **Edges the compiler cannot judge are charged to nobody**: 0.4% to 11.1% of a+ tool's output on Go, 16% to 46% on TypeScript, where the pinned corpus installs+ no dependencies. For the same reason the TypeScript with-tests variants are+ void and quoted nowhere.+- **A library has no `main`**, so cobra is analysed through its test binaries.+- **Two arms are read through an adapter we wrote.** Graphify tags 93% of its+ call edges `INFERRED` and we score all of them, the choice least favourable to+ us. code-review-graph is scored on resolved rows only, the choice most+ favourable to it.++Full method, per-cell artifacts and the pre-registration, including the two+predictions that missed:**[graph/experiments/g4-oracle-anchored](https://github.com/repowise-dev/repowise-bench/tree/master/graph/experiments/g4-oracle-anchored)**.</details>@@ -192,81 +225,61 @@ missed:<a id="7-edge-precision"></a>-## The same question, hand-graded across nine languages+## Call edges hand-graded across nine languages-**Two sections ask whether the edges are true, because there are two ways to-find out and each covers what the other cannot.** Above, a compiler is the judge-across five tools, which is unarguable but reaches only two languages. Here,-people read the source across nine languages and two tools, which reaches far-wider but is graded by us. Where both exist they agree to within about a point.+The compiler sections reach more tools or more languages. This one reaches both,+by having people read the source: 280 call edges per tool (560 in all), across+nine languages, every row read with its imports and enclosing scope open.+Measured in August 2026 against CodeGraph 1.5.0.-30 rows per language per tool, seed 2026, stratified by resolution strategy,-every row read from source with its imports and enclosing scope open.-Java is the one cell read at 40 rows rather than 30, because it was a single-repository and that repository turned out to be an outlier; it was widened to a-second one.--| | correct / n | 95% CI |+| | Correct / n | 95% CI ||---|---|---|| **repowise** | **240/280 = 85.7%** | [81.1, 89.3] |-| **CodeGraph 1.5.0** | **164/280 = 58.6%** | [52.7, 64.2] |+| CodeGraph 1.5.0 | 164/280 = 58.6% | [52.7, 64.2] || Language | repowise | CodeGraph | ||---|---|---|---|-| typescript | 29/30 | 7/30 | separates |-| go | 29/30 | 29/30 | tie |-| csharp | 30/30 | 20/30 | separates |-| python | 28/30 | 19/30 | separates |-| kotlin | 27/30 | 13/30 | separates |-| swift | 23/30 | 19/30 | tie |-| cpp | 22/30 | 16/30 | tie |-| rust | 22/30 | 13/30 | tie |-| java *(n=40)* | 30/40 | 28/40 | tie |--**Read our number the other way round: roughly fourteen percent of our call edges-are wrong.** That is the number we plan against, and rust, cpp and java are where-it concentrates.--**Four of nine cells separate. Five are ties and we report them as ties.** At n=30-the interval runs about ±16 points near 60%, so a point-estimate gap inside two-overlapping intervals is not a win. C++ is a tie despite looking like a 20-point-lead.+| TypeScript | 29/30 | 7/30 | separates |+| Go | 29/30 | 29/30 | tie |+| C# | 30/30 | 20/30 | separates |+| Python | 28/30 | 19/30 | separates |+| Kotlin | 27/30 | 13/30 | separates |+| Swift | 23/30 | 19/30 | tie |+| C++ | 22/30 | 16/30 | tie |+| Rust | 22/30 | 13/30 | tie |+| Java *(n=40)* | 30/40 | 28/40 | tie |++**Four of nine cells separate and five are ties**, and we report the ties as+ties. At n=30 the interval is about ±16 points near 60%, so C++ is a tie despite+looking like a 20-point lead.++**Read our number the other way round: about one call edge in seven is wrong.**+Rust, C++ and Java are where that concentrates. The compiler-graded table above+agrees for those three, and adds C#: the hand audit read C# at 30/30, while the+compiler reads 0.73 to 0.92 across three repositories. For C#, plan against the+compiler figure.<details>-<summary><b>Method, limits, and what this does not show</b></summary>--**One repository goes clearly to them.** On `seastar` CodeGraph reads 6/10 against-our 5/10, the only repository in the audit, on any language, where they beat us on-a clear margin. Our misses there are chained calls on an untyped receiver; they-infer the callee's declared return type and validate against it, so a failed-inference costs them an edge instead of buying them a wrong one. On `aria2` both-sides read 10/10 and they resolve 24,950 distinct call edges to our 9,486.-Precision is not the only reading.--**This audit was graded by us, on both sides.** That is the strongest form of a-weak thing, and it is exactly what the oracle section above exists to check. The-two agree where both exist, to within about a point, but only the oracle has an-answer key we did not produce.--**All nine cells are measured at one commit**, and the resolver has moved since.-Every corpus repository's call population was diffed site by site against the-current tip: three cells are byte-identical, six moved, and **of the 280 graded-rows twelve moved — every one of them a row graded `wrong`, with no row graded-`correct` moving at all.** Re-reading at the tip can therefore only raise the-number, so 85.7% is a floor. The table is not redrawn here because a cell whose-population moved needs a fresh proportional draw rather than a twelve-row patch.-No cell was measured on the 0.44.0 release.--**All 600 graded rows are published**, one file per cell, each row carrying the call-site, the declaration the tool bound it to, the verdict and the reason it was-given. A script in the same directory rebuilds every table above from them and-fails if it disagrees. 560 of those rows enter the pooled figures: C++ is read at-50 rows per side and enters at a seeded 30 so that it carries the same weight as-every other language. Every row on the page now carries its own verdict.--**Precision is one of two halves.** We lead no recall cell against the oracle, and-lose cross-file coverage on 15 of 35 repositories in the same bench. A precision-win is a claim about the edges a tool draws, never about how many.+<summary><b>Method and limits</b></summary>++30 rows per language per tool, seed 2026, stratified by resolution strategy.+Java is read at 40 rows because its first repository was an outlier, so a second+was added. C++ is read at 50 rows per side and enters the pooled figure at a+seeded 30, so every language carries the same weight. All 600 graded rows are+published, each with the call site, the declaration it was bound to, the verdict+and the reason, with a script that rebuilds every table and fails if it+disagrees.++**One repository goes to CodeGraph.** On `seastar` CodeGraph reads 6/10 against+our 5/10. Our misses there are chained calls on an untyped receiver. On `aria2`+both read 10/10 and CodeGraph resolves 24,950 distinct call edges to our 9,486,+so precision is not the only reading.++**We graded both sides.** That is what the compiler sections exist to check.++**Measured at one commit.** Diffed site by site against a later tip, 12 of the+280 graded rows moved, and every one of them had been graded wrong. Re-reading+at the tip could only raise the figure, so 85.7% is a floor.**[graph/experiments/g1-edge-precision/rows](https://github.com/repowise-dev/repowise-bench/tree/master/graph/experiments/g1-edge-precision/rows)**@@ -278,73 +291,73 @@ win is a claim about the edges a tool draws, never about how many.## Finding the right files-Before a tool can save an agent any work, it has to point at the right code.-**Grading is deterministic and no LLM judge is involved anywhere in this number**,-which makes it the most reproducible result on the page: ContextBench ships gold-file spans, and a tool either returns them or it does not.+Before a tool can save an agent work, it has to point at the right code.+ContextBench ships the files a fix actually needed, and a tool either returns+them or it does not. No LLM judge is involved, which makes this the most+reproducible result on the page.++The 112 instances were split 70 / 42 by instance id before any work started, and+the 42 stayed sealed until the final measurement. Every figure below is from the+sealed 42. Measured August 2026 on repowise `081a59fa`.-**Every number below comes from instances this work has never seen.** The 112-instances were split 70 / 42 by instance id, **pinned before any of it started**,-and the 42 were kept sealed until the final measurement.+<div align="center">+<picture>+ <source media="(prefers-color-scheme: dark)" srcset="../.github/assets/bench/file-coverage-dark.svg" />+ <img src="../.github/assets/bench/file-coverage.svg" alt="File coverage on 42 sealed ContextBench instances: repowise get_answer 0.876, repowise search_codebase 0.742, CodeGraph 0.610, Graphify 0.546, code-review-graph 0.445, cocoindex 0.361" width="100%" />+</picture>+</div>| Tool | File coverage | n | Precision | Files served ||---|---:|---:|---:|---:|| **repowise** (`get_answer`) | **0.876** | 42 | 0.087 | 19.2 |-| **repowise** (`search_codebase`) | **0.742** | 42 | **0.168** | **8.2** |+| **repowise** (`search_codebase`) | **0.742** | 42 | 0.168 | 8.2 || CodeGraph | 0.610 | 42 | 0.093 | 14.0 || Graphify | 0.546 | 42 | 0.033 | 34.5 |-| code-review-graph | 0.445 | 42 | 0.240 | 5.4 |+| code-review-graph | 0.445 | 42 | **0.240** | 5.4 || cocoindex | 0.361 | 41 | 0.092 | 7.1 |-Head to head per instance against CodeGraph: `get_answer` **19 wins, 1 loss, 22-ties, sign test p = 0.00004**. `search_codebase` **13 wins, 3 losses, 26 ties,-p = 0.021**.--**These are two different tools with two different profiles, and we would rather-say so than average them into one claim.** `get_answer` finds the most, from a list-of ~19 files. `search_codebase` finds fewer but is the most efficient per file-served, **0.742 from 8.2 files**, better coverage-per-file than anything else in-the table. If you are paying by the token, that is the row to read.+Per instance against CodeGraph: `get_answer` **19 wins, 1 loss, 22 ties, sign+test p = 0.00004**; `search_codebase` 13 wins, 3 losses, 26 ties, p = 0.021.-**Precision is not our column.** code-review-graph's 0.240 is nearly three times-ours. Part of that is mechanical, since precision rises for whoever returns fewest-files, which is exactly why files-served is a column here and not a footnote.+The two repowise tools have different profiles. `get_answer` finds the most,+from a list of about 19 files. `search_codebase` finds 0.742 from 8.2 files, the+best coverage per file served in the table. Precision is not our column:+code-review-graph's 0.240 is the highest, partly because precision rises for+whoever returns the fewest files.<details>-<summary><b>Method, limits, and what this does not show</b></summary>+<summary><b>Method and limits</b></summary>-**Why this is not benchmark tuning.** We first ran this and came **last, at-0.228**. We published that. The cause was a query-time gate discarding most-candidates before ranking ever happened. Fixing that path is what moved the number,-and it is a fix any user gets rather than a change shaped around these questions.-The check on that claim is the split, and it points the right way:+**Why this is not tuned to the questions.** Our first run came last, at 0.228,+and we published it. A query-time gate was discarding most candidates before+ranking; fixing it is what moved the number, for every user. The split is the+check:-| | other half (n=70) | **sealed half (n=42)** |+| | Other half (n=70) | **Sealed half (n=42)** ||---|---:|---:|| repowise (`get_answer`) | 0.810 | **0.876** || repowise (`search_codebase`) | 0.684 | 0.742 |-| CodeGraph | **0.6093** | **0.6095** |--Overfitting makes the unseen half score worse. Ours scores better, on both tools.-CodeGraph, which nobody tuned against either half, scores the same on both to three-decimal places, so the two halves are equally hard and the gap is about the tool-rather than the questions. **We do not quote a pooled 112-instance figure**, though-it is easy to compute and would be 0.835. Averaging the halves loses the only-number that matters.--**The cocoindex row is not from the same sitting**, and it is printed rather than-smoothed away: measured 2026-08-09 against the other five from 2026-08-02 to-2026-08-06, same instances, same gold spans, same deterministic grading. Its n is-41 because one named instance never answered even when queried alone, so it is-excluded rather than counted as a zero, which makes its row *better* than counting-it would (0.361 against 0.353). Its placing was-[pre-registered before its index existed](https://github.com/repowise-dev/repowise-bench/tree/master/configs),-including the prediction that it would come last.--**This is retrieval, not task success.** It says we find the right files, not that-an agent using us writes better code.--Measured on repowise `081a59fa` (between v0.37.0 and v0.38.0). Raw cells:+| CodeGraph | 0.6093 | 0.6095 |++Overfitting makes the unseen half score worse. Ours scores better, and CodeGraph,+which nobody tuned against either half, scores the same on both, so the halves+are equally hard. We do not quote a pooled 112-instance figure.++**The grader never reads `confidence`.** File coverage compares returned paths+with the gold files and nothing else. A `get_answer` reply can cover every gold+file and report `confidence: low`, or the reverse, without moving this number.+The 0.876 says nothing about how well that field is calibrated.++**cocoindex was measured on 2026-08-09**, the others between 2026-08-02 and+2026-08-06, on the same instances and grading. Its n is 41 because one instance+never answered; excluding it is kinder to it than scoring a zero (0.361 against+0.353). Its last place was+[pre-registered before its index existed](https://github.com/repowise-dev/repowise-bench/tree/master/configs).++**This is retrieval, not task success.** It shows we find the right files, not+that an agent using us writes better code.++Raw cells:**[rung8](https://github.com/repowise-dev/repowise-bench/tree/master/results/bakeoff_2026_08/rung8)**.</details>@@ -355,137 +368,77 @@ Measured on repowise `081a59fa` (between v0.37.0 and v0.38.0). Raw cells:## What changes in a real agent loop-The section a skeptic should read first. Every question in `django/django`'s-question set, 48 of them, across five question shapes. Six arms: repowise, four-competing tools, and a bare agent with no tools. Byte-identical prompt, full-advertised tool surface per arm, a freshly built index on the same pinned commit,-and a bare control verified free of local hooks. Two things have to be true for a-tool to be worth mounting: the agent has to actually call it, and the loop has to-get leaner when it does.+Every question in `django/django`'s question set, across five question shapes.+Six arms: repowise, four competing tools, and a bare agent with no tools. Same+prompt, each tool's full advertised surface, a fresh index on the same pinned+commit. Codex (`gpt-5.6-sol`), August 2026, repowise `081a59fa`. Five of the 48+questions hit an API usage cap in every arm, so the figures cover the 43 all six+arms completed.-**48 questions on Codex (`gpt-5.6-sol`).** Every tool called on every question, so-this is like for like.+<div align="center">+<picture>+ <source media="(prefers-color-scheme: dark)" srcset="../.github/assets/bench/agent-output-tokens-dark.svg" />+ <img src="../.github/assets/bench/agent-output-tokens.svg" alt="Output tokens an agent writes to reach an answer across 43 django questions on Codex: repowise 1,250, CodeGraph 1,383, Serena 1,550, Graphify 1,658, code-review-graph 1,710, bare agent 1,828 baseline" width="100%" />+</picture>+</div>| Tool | Agent used it | Output tokens | vs bare agent | Tool calls | Leaner on | p ||---|---:|---:|---:|---:|---:|---:|| **repowise** | **44 / 44** | **1,250** | **-31.6%** | **3.8** | **37 of 44** | **<0.0001** |-| CodeGraph | 44 / 44 | 1,383 | **-24.4%** | 4.0 | 37 of 44 | **<0.0001** |+| CodeGraph | 44 / 44 | 1,383 | -24.4% | 4.0 | 37 of 44 | <0.0001 || Serena | 43 / 43 | 1,550 | -14.8% | 10.1 | 35 of 43 | <0.0001 || Graphify | 43 / 43 | 1,658 | -8.9% | 7.4 | 31 of 43 | 0.003 || code-review-graph | 43 / 43 | 1,710 | -6.0% | 7.2 | 26 of 43 | 0.046 || *bare agent (control)* | 0 / 44 | 1,828 | baseline | 7.2 | n/a | n/a |-A third less output than working with no tool at all, reached in **3.8 tool calls-against the bare agent's 7.2**, opening 3.0 files instead of 7.2. One answered-question replacing roughly six greps, visible in the call counts rather than-inferred.--**CodeGraph is a genuine second at -24.4%**, and the honest reading is that we lead-a field in which more than one tool works. Correcting for testing five tools at-once, three reductions are solid and two are marginal.--**We ran this on three agent harnesses**, because the answer depends on the-harness as much as on the tools: Codex above, Claude Code, and a local `qwen3:8b`-under Ollama where inference is free and the win shows up as time instead. Both-other tables are in the fold with their controls, including the one whose token-column fails its own control and is therefore quoted for nobody.+A third less output than working with no tool, in 3.8 tool calls against 7.2,+opening 3.0 files instead of 7.2. CodeGraph is a genuine second at -24.4%; more+than one tool in this field works. Correcting for testing five tools at once,+three reductions are solid and two are marginal. The saving grows with the work:+on the harder half of questions it was 34.3%, on the easier half 27.2% (a+post-hoc split).<details>-<summary><b>Method, limits, and what this does not show</b></summary>--5 of the 48 questions are missing from every arm equally because the run hit an API-usage cap. Paired comparisons are unaffected; the figures are over the 43 all six-arms completed.--**The saving grows with the work.** Splitting at the median by how much the bare-agent needed: easier half 27.2%, harder half **34.3%**, correlation +0.379.-Pre-computed structure replaces exploration, and harder questions contain more-exploration to replace. That split is post-hoc and the pre-registered comparisons-are the stronger evidence.--**Second harness: 15 questions on Claude Code (`claude-sonnet-5`).**--| Tool | Tools advertised | Schema cost (chars) | Agent used it | Output tokens | vs bare | Leaner on | p |-|---|---:|---:|---:|---:|---:|---:|---:|-| **repowise** | 10 | 17,561 | **15 / 15** | **2,420** | **-15.9%** | **12 of 15** | **0.035** |-| CodeGraph | 1 | 1,567 | 13 / 15 | 2,540 | -11.7% | 10 of 15 | 0.302 |-| Serena | 29 | 29,050 | 4 / 15 | 2,551 | -11.3% | 8 of 15 | 1.000 |-| code-review-graph | 30 | 28,118 | **0 / 15** | 2,768 | -3.8% | 10 of 15 | 0.302 |-| Graphify | 10 | 5,482 | 3 / 15 | 2,878 | 0.0% | 7 of 15 | 1.000 |-| *bare agent (control)* | 0 | 0 | n/a | 2,877 | baseline | n/a | n/a |--The "agent used it" column is the interesting one: under Claude Code most of these-tools were barely called at all, code-review-graph never once and Graphify three-times in fifteen. Nothing differed about the servers, questions or indexes between-the two harnesses. Claude Code loads MCP tool schemas on demand, so the agent has-to go looking before it can call anything, and frequently never does. **Treat that-column as unstable, including our own 15 of 15.** Reruns on later days returned 4-of 15 and then 3 of 15 for us, and 2 of 14 for CodeGraph. Whether the collapse is-the model or the harness is **inconclusive**: the same 15 questions on Opus,-against bands fixed before the run (12+ means the model, 6 or fewer means schema-deferral, 7 to 11 inconclusive), came back at 7. That run's token column fails its-own control (-9.3% when the tool was called against -10.7% when it was not), so no-token figure is quoted from it for any tool including ours.--**Third harness: `qwen3:8b` under Ollama via opencode**, same 15 questions, same-seed. Every cell called its tool, 15 of 15 on both rows.--| Row | Agent used it | Output tokens | vs bare | Leaner on | p | Wall clock | vs bare |-|---|---:|---:|---:|---:|---:|---:|---:|-| **repowise, full surface** | **15 / 15** | **1,319** | **-40.8%** | **15 of 15** | **0.00006** | **117s** | **-27.5%** |-| **repowise, local-only tools** | **15 / 15** | **1,172** | **-47.9%** | **15 of 15** | **0.00006** | **96s** | **-41.5%** |-| *bare agent (control)* | 0 / 15 | 2,336 | baseline | n/a | n/a | 171s | baseline |--**Two repowise rows, never combined.** `get_answer` writes its answer using a-hosted model, so a row using it is not a local-only result. The second row switches-it off and leaves only tools that run against the local index. We verified the-restriction holds rather than assuming it: instructed directly and repeatedly to-call `get_answer`, that agent could not reach it in any of its 15 cells. repowise-roughly doubles the tokens fed in on a single step while cutting steps from 3.3 to-2.1 and halving what the model writes, which is why the win shows up as time:-reading a large payload once is cheap on a GPU, generating text token-by-token-across several rounds is not.--**What this section will not claim:**--- **This is work saved, not quality.** A blind judge scored every tool in the- field, ours included, a fraction **below** the bare agent on the 48-question run,- in a range of 0.04 to 0.25 points on a 10-point scale, all smaller than the 0.69- points by which this benchmark moves when rerun unchanged. No tool here- measurably changed answer quality in either direction. Ours is at the low end of- that band and we will say so if it becomes a real effect. "No significant- difference" is not parity, and an equivalence claim needs a test we have not run.- On the local run the full surface scored +1.32 against the bare agent, above the- judge's measured 0.69 noise, but the win-loss count is 10 to 5 at p = 0.30 and- removing the single best question drops it to +0.99; the local-only row is- **+0.20, a null**. The defensible statement is narrower: the local-only- configuration answers about as well as a bare agent in roughly half the wall- clock, on hardware you already own.-- **The quality columns cannot be compared across tables**, because grading a model- with a judge from its own family is a known bias, so the judges differ.-- **Not a universal saving.** One repository, one commit, one prompt. Three- harnesses, which is more than most published numbers in this category, and still- not a constant.-- **Adoption is not a stable property of a tool.** Whether an agent calls a- codebase server at all depends more on the harness than on the server. Any- adoption figure, ours included, is only meaningful with its harness and its date- attached. Ordering across four instruments was Sonnet 3 to 4, Opus 7, Codex 15,- local `qwen3:8b` 15.-- **No dollar figure**, and the control that retired it is the most useful thing we- ran. code-review-graph, having never called its server once in 15 questions and- carrying 28,118 extra characters of schema that should cost *more*, measured- **43% cheaper than the bare agent**. The cause is prompt caching: whichever arm- runs first warms it. An arm's position in the cycle correlated **-0.487** with its- dollar cost. Output tokens are never cached and correlate **+0.010**, so those are- what we publish. If you see a token claim anywhere that does not say whether it- controls for cache state and arm ordering, ask about that first.+<summary><b>Other harnesses, and what this does not show</b></summary>++**Claude Code (`claude-sonnet-5`), 15 questions.** repowise was called on 15 of+15 and cut output by 15.9% (p = 0.035); no other tool reached significance.+Whether an agent calls a tool at all was unstable on this harness: reruns on+later days returned 4 of 15 and then 3 of 15 for us, and 2 of 14 for CodeGraph.+Claude Code loads tool schemas on demand, so the agent has to go looking first.+A run on Opus failed its own control, so no token figure is quoted from it.++**Local `qwen3:8b` under Ollama, 15 questions.** Every cell called its tool.++| Row | Output tokens | vs bare | Leaner on | Wall clock | vs bare |+|---|---:|---:|---:|---:|---:|+| repowise, full surface | 1,319 | -40.8% | 15 of 15 | 117 s | -27.5% |+| repowise, local-only tools | 1,172 | -47.9% | 15 of 15 | 96 s | -41.5% |+| *bare agent* | 2,336 | baseline | n/a | 171 s | baseline |++`get_answer` writes its answer with a hosted model, so the local-only row+switches it off; we confirmed the agent could not reach it. On a GPU, reading+one large payload is cheap and generating text over several rounds is not, which+is why the win shows up as time.++**What this section does not claim:**++- **Work saved, not quality.** A blind judge scored every tool, ours included,+ within 0.04 to 0.25 points of the bare agent on a 10-point scale, inside the+ 0.69 points this benchmark moves when rerun unchanged. No tool measurably+ changed answer quality. On the local run, the local-only configuration+ answered about as well as a bare agent in about half the wall clock.+- **One repository, one commit, one prompt**, and `django` is in every model's+ training data. Nothing here measures a multi-hour engineering task.+- **No dollar figure.** Prompt caching made whichever arm ran first look cheap:+ arm position correlated -0.487 with dollar cost. Output tokens are never cached+ and correlated +0.010, so those are what we publish.Raw data, every cell including failures:**[rung9](https://github.com/repowise-dev/repowise-bench/tree/master/results/bakeoff_2026_08/rung9)**-(48-question Codex) and+(Codex) and**[rung6](https://github.com/repowise-dev/repowise-bench/tree/master/results/bakeoff_2026_08/rung6)**-(the 15-question runs). What producing them cost in runs, machine hours and API-spend, plus the per-tool setup traps that were most of the real effort:+(the 15-question runs). Setup traps per tool and what the runs cost:[head-to-head](https://github.com/repowise-dev/repowise-bench/tree/master/head-to-head).</details>@@ -493,71 +446,57 @@ spend, plus the per-tool setup traps that were most of the real effort:---<a id="5-code-health-predicts-defects"></a>+<a id="code-health-predicts-defects"></a>## Code health predicts defectsA health score is worth something only if the files it flags are the files that-break. Scores are taken at a historical commit, bug fixes are counted over the+break. Each file is scored at a historical commit, bug fixes are counted over thefollowing six months, and nothing after the scoring commit feeds the score.+Measured May 2026.-Across **21 repositories, 9 languages, 2,826 files**: **ROC AUC 0.737** (95% CI-0.683 to 0.787), ranging 0.55 to 0.86 across individual repos. It beats recent-churn by +0.100 AUC and prior-defect history by +0.117 (DeLong p < 1e-9). On-PROMISE/jEdit, a dataset it never saw which carries no git history at all, it holds-at **0.76 to 0.78**.+Across **21 repositories, 9 languages and 2,826 files**: **ROC AUC 0.737** (95%+CI 0.683 to 0.787), 0.55 to 0.86 per repository. It beats recent churn by +0.100+AUC and prior-defect history by +0.117 (DeLong p < 1e-9). On PROMISE/jEdit, a+dataset it never saw, it holds at 0.76 to 0.78.-**Against CodeScene**, the closest commercial product and the only other vendor in-this category with a published empirical defect study. Both tools scored the same-2,770 files at the same leakage-free commit against the same labels:+**Against CodeScene**, the closest commercial product. Both tools scored the same+**2,770 files** (the files CodeScene also scores) at the same commit against the+same defect labels:| Paired test | repowise | CodeScene | ||---|---:|---:|---|| Recall at a 20%-of-lines review budget | **0.173** | 0.074 | p = 0.003 || Effort-aware ranking (Popt) | **0.607** | 0.462 | p = 0.003 |-| Defect density, size-normalized (below 4.0 : 8.0 and above) | **2.18x** | 0.56x | p = 0.003 |-| Discrimination (ROC AUC) | 0.731 | 0.705 | p = 0.054, marginal |+| Defect density, size-normalized | **2.18x** | 0.56x | p = 0.003 |+| Discrimination (ROC AUC) | 0.731 | 0.705 | p = 0.054, not significant || Precision at a 20%-of-lines review budget | 0.580 | **0.636** | p = 0.64, a tie |-Ranking by repowise health surfaces **2.3x the defects under a fixed review-budget**.+Ranking by repowise health surfaces **2.3x the later-defective files** within the+same review budget.-**Where CodeScene is ahead, and why it is a real choice.** Its precision lead is-not significant, but the behaviour behind it is worth having: it flags about **27-files** where we flag **132**, trading recall for a short list a team will actually-work through. That is a deliberate operating point, not a weaker model.--**The axis we lost outright.** CodeScene's "Code Red" study reports a **-0.58**-correlation between Code Health and mean issue-resolution time, on proprietary Jira-data. We could not replicate it on open data: across 17 repositories and 271 files,-GitHub PR merge time correlated at **-0.09** (95% CI -0.19 to +0.14), and six-queue-independent effort signals all came back flat with every interval spanning-zero. Merge time on GitHub largely measures maintainer review-queue availability.-**The business-impact axis remains CodeScene's, unreplicated on open data.**+**Where CodeScene is ahead.** It flags about 27 files where we flag 132, so its+list is shorter and slightly more precise: a deliberate operating point that a+team can work through quickly. CodeScene also reports a -0.58 correlation between+its score and issue-resolution time on proprietary data. We could not replicate+that on open data (GitHub merge time, 17 repositories, 271 files: -0.09, interval+spanning zero), so the business-impact axis stays CodeScene's.<details>-<summary><b>Method, limits, and what this does not show</b></summary>--**It is not better than raw file size at discrimination.** LOC-only scores 0.742-against our 0.737 (p = 0.92, a tie). Where it wins is effort-aware ranking, Popt-+0.134 (95% CI +0.080 to +0.198): the same discrimination as counting lines, much-better at ordering a fixed review budget, and unlike a line count it says why.--**The signal is weak among files of similar size.** Holding size fixed, within-band-AUC runs 0.525 / 0.572 / 0.593 / 0.718 across NLOC quartiles, so it survives-cleanly only in the largest quartile, and a positive control confirms that collapse-is a real absence rather than too few positives to detect one. A prior-defects-baseline also still beats us on Popt by 0.085 even while losing on AUC.--Our AUC edge over CodeScene is **marginal**, and so is our raw defect-density lead-(16.9x against 14.2x, p = 0.65 on a heavy tail); the size-normalized version is the-one that reaches significance and the one to cite.--Marker weights are **calibrated against a real defect corpus, not hand-tuned**:-every file scored at a commit preceding the bug window so nothing leaks backward,-and an L2-logistic fit with file size as an explicit control, so a marker only-earns weight for defect lift *beyond* being big. Only **26** of the 53 detectors-are permitted to move the defect number, because that is the number carrying these-accuracy claims.+<summary><b>Method and limits</b></summary>++**Not better than file size at discrimination.** Lines of code alone score 0.742+against our 0.737 (p = 0.92, a tie). The gain is in effort-aware ranking, Popt++0.134 (95% CI +0.080 to +0.198), and unlike a line count the score says why.++**Weak among files of similar size.** Holding size fixed, AUC runs 0.525 / 0.572 /+0.593 / 0.718 across size quartiles. A prior-defects baseline still beats us on+Popt by 0.085 while losing on AUC.++**Weights are fitted, not hand-set:** a size-controlled logistic fit on files+scored before their bug window, so a detector earns weight only for defect lift+beyond being large. Of the 53 registered detectors, 25 move the defect score;+the others feed the maintainability or performance score, or are advisory.Reports:**[BENCHMARK\_REPORT.md](https://github.com/repowise-dev/repowise-bench/blob/master/health-defect/BENCHMARK_REPORT.md)** ·@@ -567,104 +506,140 @@ Reports:----<a id="6-indexing-time-the-row-we-lose"></a>+<a id="scale"></a>++## Scale++Three large public repositories, indexed end to end on one laptop. Keyless+index only (`repowise init --no-prose`, no model calls, a stand-in embedder),+one run each, on a Windows 11 laptop with 31 GB of RAM and 32 hardware threads,+under a guard that limited the run to 8 CPUs and 2 parse workers.++| Repository | Tracked files | First index | Peak memory | Index on disk | Build |+|---|---:|---:|---:|---:|---|+| dotnet/aspnetcore v8.0.31 | 15,510 | 20 min | 4.0 GiB | 1.4 GiB | 0.54.0 release |+| elastic/elasticsearch v8.19.22 | 34,315 | 54 min | 9.8 GiB | 4.2 GiB | 0.54.0 release |+| elastic/elasticsearch v8.19.22 | 34,315 | 52 min | 7.8 GiB | 4.2 GiB | pre-release, with two memory fixes |+| dotnet/runtime v8.0.31 | **58,924** | **120 min** | **11.7 GiB** | 7.7 GiB | pre-release, with four memory and speed fixes |++Measured 2026-10-01 and 2026-10-02. The pre-release builds were main at that+time plus memory fixes that have since merged; as of 2026-10-03, main contains+all of them, and three further memory fixes merged after these runs that no+number here includes. Peak memory is the whole process tree. The pre-release+build produced byte-identical outputs to main on a determinism check.++**Updates.** On aspnetcore (0.54.0), an update with no changes took 4.1 s. An+update for a one-file commit took 469 s at 3.8 GiB, because the health pass+re-scores the whole repository, so its cost follows repository size, not change+size. Updates were not measured on the larger two.++**What we have not measured.** Nothing larger than 58,924 files; the Linux kernel+(92k files) was not run. Only Windows. One run per configuration, and walls on+this machine move by up to 2x between runs.+---++<a id="test-intelligence"></a>++## Test intelligence++Repowise links tests to the code they exercise from the call graph, with no+coverage run needed. To check those inferred links, we compared them against a+real `coverage run --contexts=test` on this repository, over a slice with+complete per-test attribution (37 test files, 159 production files). Measured+August 2026.++| Question | Precision | Sample | Recall |+|---|---:|---|---|+| Which tests reach this file | **95.7%** | 44 of 46 claims | 27.7%; 77% among files where most executed lines are inside function bodies |+| Which tests to run for a change | **97.5%** | 47 targets, all answered | n/a |++Recall reads low because coverage counts a file as run when it was merely+imported. Code that a framework calls by convention, with no static caller, is+not reached. Inferred answers are file-level; measured coverage, when you ingest+it, is line-level. Repowise labels which kind each answer is, and an empty answer+means unknown, never "no tests". Details:+[Test intelligence](layers/TEST_INTELLIGENCE.md#accuracy-and-limits).++---++<a id="6-indexing-time-the-row-we-lose"></a><a id="what-it-costs-to-run"></a>## What it costs to run-**Two different questions, and mixing them up is the easiest mistake to make about-this page.** One is what it costs to build a call graph, measured against tools that-build only a call graph. The other is what it costs to build everything repowise-builds. We lead the first and lose the second, and both are below.+Two different questions. One is the cost of building a call graph, against tools+that build only a call graph. The other is the cost of building everything+repowise builds. We lead the first and lose the second. Both measured August+2026.### Building the call graph-Graph construction only, on the same 35 repositories, all five tools: 175 cells,-0 failed, three timed runs each after a discarded warmup, nothing restored from-cache.+Graph construction only, 35 repositories, five tools: 175 cells, none failed,+three timed runs each after a discarded warmup, nothing restored from cache.| Tool | Median build | Median peak memory | Fastest on ||---|---:|---:|---:|-| **repowise** | **2.77s** | **75 MB** | 14 of 35 |-| CodeGraph | 3.65s | 757 MB | **16 of 35** |-| codebase-memory-mcp | 6.21s | 1,113 MB | 5 of 35 |-| code-review-graph | 9.97s | 361 MB | 0 |-| Graphify | 12.23s | 860 MB | 0 |--**Memory is a clean sweep and it is the number that matters most in practice.** We-are the lowest-memory tool on **35 of 35 repositories**, no exceptions, and the gap-widens with repository size: 64 MB against CodeGraph's 749 MB under 1,000 files,-152 MB against 1,164 MB above it. Our worst repository in the whole corpus is 468 MB.-Three of the four other tools exceed that on repositories a tenth the size, and one-reaches **5,523 MB** on the same repository where we use 468 MB. That is the-difference between running in a normal CI container and not.--**Speed is level, not a win, and this page used to claim otherwise.** CodeGraph is-fastest on 16 repositories to our 14: we lead under 1,000 files and trail above-it, because we do more resolution work per file. The other half of that trade is-the precision table above.+| **repowise** | **2.77 s** | **75 MB** | 14 of 35 |+| CodeGraph | 3.65 s | 757 MB | **16 of 35** |+| codebase-memory-mcp | 6.21 s | 1,113 MB | 5 of 35 |+| code-review-graph | 9.97 s | 361 MB | 0 |+| Graphify | 12.23 s | 860 MB | 0 |++**Memory: lowest on 35 of 35 repositories, about 10x below the next tool.** The+gap widens with size: 64 MB against CodeGraph's 749 MB under 1,000 files, 152 MB+against 1,164 MB above. Our worst repository is 468 MB; one other tool reaches+5,523 MB on that same repository. That is the difference between fitting in a+normal CI container and not.++**Speed is roughly tied.** CodeGraph is fastest on 16 repositories to our 14: we+lead under 1,000 files and trail above, because we do more resolution work per+file. The precision tables are the other half of that trade.### Building the whole index-This is what `repowise init` actually costs, and here we are the slowest in the-field on every repository measured. On `django/django`:+This is what `repowise init` costs, and here we are the slowest tool measured.+On `django/django`:| Tool | Index time | What it builds ||---|---:|---|-| CodeGraph | 16.4s | call graph |-| code-review-graph | 44.8s | call graph |-| Graphify | 141.5s | call graph, communities |-| **repowise** (`--no-prose`) | **366.8s** | five layers, below |-| **repowise** (default, prose on) | **1,058s** | five layers plus generated documentation |--That is **22x** CodeGraph like for like and **135x** with prose on. Both numbers-ship and the 22x is not the user-facing one.--**The 22x is not a graph number.** The call graph itself is done in about 3-seconds. The rest of that clock builds four more layers on `django/django`: a-90,477-edge graph with communities and flow tracing, git history mined across-2,630 files, **3,392 wiki pages** embedded for natural-language search,-architectural decisions, and code health at **5,317 findings**.--So "22x slower" and "the index contains categorically more" are both true and-neither cancels the other. **If a call graph is all you want, CodeGraph builds one-in 16 seconds and you should use it.** The dishonest move would be quoting the ratio-without the column beside it, which is why the column is there. It is also a-one-time cost: updates after the first index are incremental.+| CodeGraph | 16.4 s | call graph |+| code-review-graph | 44.8 s | call graph |+| Graphify | 141.5 s | call graph, communities |+| **repowise** (`--no-prose`) | **366.8 s** | graph, git history, search index, decisions, health |+| **repowise** (default, prose on) | **1,058 s** | the same, plus generated documentation |++That is **22x** CodeGraph like for like, and 135x with prose on. The call graph+itself takes about 3 seconds. The rest builds a 90,477-edge graph with+communities and flow tracing, git history across 2,630 files, 3,392 wiki pages+embedded for search, architectural decisions, and 5,317 health findings. If a+call graph is all you need, CodeGraph builds one in 16 seconds. Updates after the+first index are incremental.<details>-<summary><b>Method, limits, and what this does not show</b></summary>+<summary><b>Method and limits</b></summary>-**The two tables have different denominators and must never be quoted against each-other.** The graph table is construction only: no documentation, no embeddings, no-health pass. The index table is a full `repowise init`.+**Never quote the two tables against each other.** The graph table is+construction only; the index table is a full `repowise init`.-**One caveat runs in our favour on the graph table and is stated rather than-buried.** Our graph is built in memory and discarded, so serialisation cost is-excluded from our column, while CodeGraph writes a real SQLite index (115 MB on one-repository). Our memory figures come from a subprocess arm built precisely so peak-RSS could be read the way every competitor's already was.+**One caveat favours us on the graph table.** Our graph is built in memory and+discarded, so serialisation is excluded from our column, while CodeGraph writes a+real SQLite index. Our memory figures come from a subprocess arm built so peak+RSS is read the same way as every competitor's.-**CodeGraph's own headline speed claim is incremental re-sync**, roughly 0.3s to-fold one file into a 4,400-file project. That is unmeasured here and we expect to+**CodeGraph's headline speed claim is incremental re-sync**, about 0.3 s to fold+one file into a 4,400-file project. That is not measured here, and we expect tolose it.-**Seconds alone is not a quality claim.** CodeGraph produces more distinct call-edges than we do on several repositories.--**The corpus caps near 2,000 files**, so genuinely large-repository behaviour is-measured on the top 10 of 35 rather than at scale. Single machine, single OS: the-memory ratios are unlikely to invert but absolute numbers will move.+Single machine, single OS: the memory ratios are unlikely to invert, but+absolute numbers will move. The [scale section](#scale) covers repositories far+larger than this corpus, for repowise alone. A fitted build-cost curve for all+five tools across a 12x size range is in+[head-to-head](https://github.com/repowise-dev/repowise-bench/tree/master/head-to-head#the-result-nobody-in-the-field-had-published-a-build-cost-curve);+our exponent is 0.906, which we read as sublinear work, not efficiency, because+symbol density halves across the range.-A fitted build-cost curve for five tools across a 12x repository-size range, which-nobody in this field had published, is in-[head-to-head](https://github.com/repowise-dev/repowise-bench/tree/master/head-to-head#the-result-nobody-in-the-field-had-published-a-build-cost-curve).-Our exponent is sublinear at 0.906, and we publish that as sublinear **work** rather-than efficiency, because symbol density halves across the range. That is a product-finding against us.--Full per-repository tables:+Per-repository tables:**[graph/experiments/g6-build-cost](https://github.com/repowise-dev/repowise-bench/tree/master/graph/experiments/g6-build-cost)**.</details>@@ -676,165 +651,112 @@ Full per-repository tables:## Command-output compression-`repowise distill <cmd>` compresses command output *before* the agent reads it:+`repowise distill <cmd>` compresses command output before the agent reads it:errors first, exit code preserved, every omission recoverable through an inline-`[repowise#<ref>]` marker.+`[repowise#<ref>]` marker. One run per command on one repository, July 2026.| Command | Raw tokens | Distilled | Saved ||---|---:|---:|---|| `pytest -q` (11 failures) | 3,374 | 1,317 | **61%**, all 11 `FAILED` lines kept || `git log -50` | 3,064 | 331 | **89%** |-| `git diff` (30 commits) | 62,833 | 8,635 | 86%, **unsupported, see below** |+| `git diff` (30 commits) | 62,833 | 8,635 | 86%, **needs re-measurement, see below** || `git log --oneline -30` | 321 | 321 | 0%, already compact || `git status` (clean) | 83 | 83 | 0%, too small to distill |-**One row above is inflated, by the mechanism that broke RTK's number.** A saving-may only count output the agent would actually have received, and the host-truncates a command result at 30,000 characters. `git diff`'s 62,833 raw tokens is-eight times that cap and predates it. **Treat that 86% as unsupported until-re-measured.** `pytest` and `git log` sit under the cap and are unaffected.--<details>-<summary><b>Method, limits, and what this does not show</b></summary>+**The `git diff` row overstates the saving.** A saving may only count output the+agent would have received, and the host truncates a command result at 30,000+characters; 62,833 raw tokens is far past that cap. Treat that 86% as unsupported+until re-measured. The `pytest` and `git log` rows sit under the cap.-**There is a comparable tool and we have not run it.**+The 0% rows show the guard working: distill never inflates small output.[RTK](https://blog.jetbrains.com/ai/2026/07/rtk-claude-code-token-savings/) does-this and essentially only this. It is also the tool whose advertised 60 to 90% came-back **7.6% more expensive** under JetBrains' rerun. So this table is a-before-and-after of our own output, which is the weaker design. Only a paired run-would settle it.--The two 0% rows are the net-positive guard working: distill never inflates small-output. One run per command on one repository, so these are point measurements.-Reduction is also not comprehension: the bytes removed are measured, the evidence-they were safe to remove is narrower.--Full guide: **[docs/agent/DISTILL.md](agent/DISTILL.md)**--</details>+the same job and we have not run it head to head, so this table is a+before-and-after of our own output. Guide: [docs/agent/DISTILL.md](agent/DISTILL.md).---<a id="3-loading-one-commits-context-the-easy-number"></a><a id="loading-one-commits-context-the-easy-number"></a>-## Loading one commit's context, the easy number+## Loading a commit's context-This is the measurement almost every tool in this category publishes, and we are-labelling it as such. Over the 30 most recent non-merge commits of `pallets/flask`,-counted with deterministic `tiktoken` (`cl100k_base`):+How many tokens an agent reads to take in one commit. Over the 30 most recent+non-merge commits of `pallets/flask`, counted with `tiktoken` (`cl100k_base`), on+repowise main `4fccb31ec` with a keyless index, 2026-10-03:| Strategy | Tokens per commit ||---|---:|-| naive, full contents of every changed file | 13,984 |-| `git diff` only | 1,408 |-| **`get_context`** | **393** |--**35.6x fewer than naive, pooled**, 29.3x as a mean of per-commit ratios, and 3.6x-pooled against `git diff`.--**What this does not tell you.** It is one payload, not a session. It says our-representation of a commit is smaller than the commit, not that an agent finishes-faster. The honest version of that question is the agent loop above, where the-answer on the same harness is a much more modest 15.9%.--<details>-<summary><b>Method, limits, and what this does not show</b></summary>+| Full contents of every changed file | 13,984 |+| **`get_context` on the changed files** | **2,586** |+| `git diff` only | 1,412 |-**Lead with the pooled figure.** Pooled is sum over sum, so it weights each commit-by the tokens actually at stake. A mean of per-commit ratios does not: a one-line-commit where `get_context` returns 40 tokens contributes a huge ratio that counts-equally against one saving a hundred thousand. That is how a 35x becomes a 209x in-a press release.+`get_context` is 5.4x smaller than reading the changed files, and larger than the+raw diff, because it returns every symbol in each file with its signature and+location. One payload is not a session; the agent-loop section is the measure+of work saved.-Raw CSV:-**[token\_efficiency\_flask30](https://github.com/repowise-dev/repowise-bench/blob/master/results/bakeoff_2026_08/rung1/token_efficiency_flask30_2026-08-01.csv)**--</details>+An earlier version of this page reported 393 tokens per commit. That figure came+from a harness run that was receiving error responses in place of context, and it+has been withdrawn.---## Limits that apply to the whole page-<details>-<summary><b>Four limits that are not specific to any one section</b></summary>--- **Every retrieval number here is Python or Go.** The graph sections are wider, at- nine languages hand-graded and two compiler-graded, and they are the only- sections that are. A JavaScript/TypeScript corpus (`mui/material-ui`, six tools,- a 12x size range) is built and half graded, but **its sealed half is unrun and- nothing from it is quoted here.** Publishing the development half is what that- split exists to prevent, so the row arrives when the sealed half is evaluated,- once.-- **We index code files only, which depresses our own coverage on repositories- with documentation.** Deliberate: on doc-heavy repositories the docs outweigh the- code, and indexing them polluted the index and degraded retrieval for the code- questions the tool exists to answer. On the JavaScript corpus above, **21% of- gold files are `.md` or `.json`**, unreachable to us by construction rather than- by ranking. cocoindex makes the opposite trade, is the only arm of six to- retrieve any documentation gold, and pays for it on code gold.-- **The agent-loop section is one repository**, `django/django` at one commit, which- is in every model's training data, and it measures single-session questions of- roughly four to nine turns each. **Nothing on this page measures a long- multi-hour engineering task**, and we will not imply a number for one.-- **Command-output compression has no head-to-head and one row that needs- re-measuring.**--</details>-----+- **Retrieval and agent-loop numbers are Python and Go**, and the agent loop is+ one repository at one commit. The graph sections are wider: nine languages+ hand-graded, Go and TypeScript against a compiler across five tools, and six+ more languages against a compiler for repowise alone. A JavaScript and+ TypeScript retrieval corpus is built, but its sealed half is unrun, so nothing+ from it is quoted.+- **We index code files only.** On repositories where documentation or JSON+ files are part of the answer, that lowers our retrieval score by construction.+- **Dates differ by section.** Graph accuracy is current main; retrieval, agent+ loop and cost are August 2026; health is May 2026. Each section says which.+- **Command-output compression has no head-to-head**, and one row needs+ re-measuring.## How to read a number on this page-In July 2026 JetBrains took two popular token-saving tools and reran their headline-claims on real agent work.--| Tool | Advertised saving | What JetBrains measured |-|---|---|---|-| [Caveman](https://blog.jetbrains.com/ai/2026/07/speak-to-ai-agents-like-cavemen-tosave-tokens/) | 65% | **8.5%** |-| [RTK](https://blog.jetbrains.com/ai/2026/07/rtk-claude-code-token-savings/) | 60 to 90% | **7.6% more expensive** at low reasoning effort (p = 0.004) |--Greptile's 82% became 45% under Augment's rerun. The pattern is consistent enough-to be a rule: in this category, the advertised number and the reran number are-different numbers. **Measuring one context load is easy. Measuring an agent loop is-hard.** Agents re-read, backtrack, re-plan and re-explore, so a compression that-looks like 90% on one payload routinely nets out near zero across a session.--This page publishes both and labels which is which. We reran our own headline on a-second and then a third harness and published what came back: it moved a number we-had been quoting, and the agent-loop section carries the correction.--Four rules apply to every number here, each because breaking it produced a wrong-published figure at least once. **Pre-register before spending**, as its own commit,-so a favourable result cannot become a different question afterwards. **Seal a-half**, split by instance id before any work begins, evaluated once. **Precision and-files-served beside coverage**, never averaged into one figure. **Prove an arm was-alive and its extractor works before recording a zero**, because a dead server, a-wrong tool name and a broken output parser all score exactly like a bad tool. The-method end to end, with every gate and the failure that put it there:+In this category, advertised savings often shrink when someone else reruns them.+In July 2026 JetBrains reran two token-saving tools on real agent work:+[Caveman](https://blog.jetbrains.com/ai/2026/07/speak-to-ai-agents-like-cavemen-tosave-tokens/)+advertised 65% and measured 8.5%;+[RTK](https://blog.jetbrains.com/ai/2026/07/rtk-claude-code-token-savings/)+advertised 60 to 90% and measured 7.6% more expensive. Shrinking one payload is+easy to measure. A whole agent session is harder, because agents re-read and+re-explore.++Four rules apply to every number here:++- **Pre-register before running**, as its own commit, so a good result cannot+ become a different question afterwards.+- **Seal a held-out set**, split before any work begins and evaluated once.+- **Show precision and recall together**, and files served beside coverage.+ Never average them into one figure.+- **Prove a tool was running before recording a zero.** A dead server and a bad+ tool score the same.++The method end to end:**[THE\_LOOP.md](https://github.com/repowise-dev/repowise-bench/blob/master/head-to-head/THE_LOOP.md)**.-----## Where the depth is-This page is the summary. Everything below it is the evidence, in-**[repowise-bench](https://github.com/repowise-dev/repowise-bench)**: what each run-cost in machine hours and API spend, the build-cost curve, one page per competitor-with every setup trap, all nine method gates, every launch command and exclusion-with its reason, one pre-registration per scored run committed before that run spent-anything, and every graded cell including the runs we invalidated with their-invalidation notes attached.+This page is the summary. The evidence is in+**[repowise-bench](https://github.com/repowise-dev/repowise-bench)**: every graded+cell, every pre-registration, the runs we invalidated with their notes, and what+each run cost.| Start here | For ||---|---|-| **[graph/](https://github.com/repowise-dev/repowise-bench/tree/master/graph)** | The graph-quality bench: hand-graded precision across nine languages, the compiler oracle on Go and TypeScript, cross-file coverage and build cost across five tools |-| **[head-to-head](https://github.com/repowise-dev/repowise-bench/tree/master/head-to-head)** | The retrieval and agent-loop bench: who wins what, what each index can rank at all, and what it all cost to produce |+| **[graph/](https://github.com/repowise-dev/repowise-bench/tree/master/graph)** | Graph accuracy: hand-graded precision across nine languages, the compiler comparison on Go and TypeScript, per-language scores for repowise, cross-file coverage and build cost |+| **[head-to-head](https://github.com/repowise-dev/repowise-bench/tree/master/head-to-head)** | Retrieval and agent-loop results, per-tool setup notes, and what producing them cost |+| **[health-defect](https://github.com/repowise-dev/repowise-bench/tree/master/health-defect)** | The defect-prediction study and the CodeScene comparison || **[results/bakeoff\_2026\_08](https://github.com/repowise-dev/repowise-bench/tree/master/results/bakeoff_2026_08)** | Every graded cell and every verbatim response |-| **[repro/README.md](https://github.com/repowise-dev/repowise-bench/blob/master/repro/README.md)** | Per claim: what it costs to reproduce, how long it takes, and which ones need credentials we cannot hand you |+| **[repro/README.md](https://github.com/repowise-dev/repowise-bench/blob/master/repro/README.md)** | Per claim: what it costs to reproduce and which ones need credentials |-**Found a problem with one of these numbers, or want your tool in the field?**-Adding a competitor is a YAML block, no Python and no runner change. See+Found a problem with a number, or want your tool in the field? Adding a+competitor is a YAML block. See[CONTRIBUTING.md](https://github.com/repowise-dev/repowise-bench/blob/master/CONTRIBUTING.md).<sub>Tool versions as measured: CodeGraph 1.5.0, Graphify 0.9.31, Serena 1.6.2.dev0,@@ -842,7 +764,8 @@ code-review-graph 2.3.7, cocoindex as of 2026-08-09, codebase-memory-mcp 0.10.8.## See also-- [The five intelligence layers](layers/INTELLIGENCE_LAYERS.md)-- [Code health methodology](layers/CODE_HEALTH.md)+- [The intelligence layers](layers/INTELLIGENCE_LAYERS.md)+- [Code health](layers/CODE_HEALTH.md)+- [Test intelligence](layers/TEST_INTELLIGENCE.md)- [MCP tool reference](agent/MCP_TOOLS.md)-- [Roadmap](../ROADMAP.md), including the interface-dispatch recall ceiling+- [Roadmap](../ROADMAP.md)diff --git a/docs/LINEAGE.md b/docs/LINEAGE.mdnew file mode 100644index 000000000..47f571e0e--- /dev/null+++ b/docs/LINEAGE.md@@ -0,0 +1,201 @@+# What Repowise is built on++Repowise did not invent the ideas it runs on. Its call graph, its git-history+signals, its health score and its refactoring plans each come from published+software-engineering research, some of it fifty years old and some of it from last+year. This page lists that research: what each work contributed, where it lives in+the code, and how we checked it on real repositories.++It exists for the same reason [BENCHMARKS.md](BENCHMARKS.md) does. A score that says+a file is risky is only worth acting on if you can see where the method came from and+how well it held up.++**The rule for a row.** A work appears here only if its contribution can be stated in+one sentence and pointed at code that implements it. Works we measured and rejected+are listed separately, and so are well-known methods we do *not* implement, so nobody+reads this page as a longer list than it is.++Paths are relative to `packages/core/src/repowise/core/` unless they say otherwise.++**At a glance:** about 45 published works, from McCabe (1976) to 2025, across graph+theory, mining software repositories, object-oriented metrics, defect prediction,+program analysis, refactoring and information retrieval.++---++## The call graph++| Year | Work | What we take from it | Where it lives |+|---|---|---|---|+| 1972 | Tarjan, *Depth-first search and linear graph algorithms*, SIAM J. Computing | Strongly connected components: import and call cycles | `ingestion/graph/_metrics.py` |+| 1993 | Eades, Lin and Smyth, *A fast and effective heuristic for the feedback arc set problem*, Inf. Processing Letters | Which edge to cut to break a cycle | `analysis/health/refactoring/break_cycle.py` |+| 1998 | Brin and Page, *The anatomy of a large-scale hypertextual web search engine*, WWW7 | PageRank centrality: which files and symbols the rest of the code leans on | `ingestion/graph/_metrics.py` |+| 2001 | Brandes, *A faster algorithm for betweenness centrality*, J. Mathematical Sociology | Bridge files that sit between parts of the system | `ingestion/graph/_betweenness.py` |+| 2008 | Blondel et al., *Fast unfolding of communities in large networks*, J. Stat. Mech. | Louvain communities (the fallback) | `analysis/communities.py` |+| 2019 | Traag, Waltman and van Eck, *From Louvain to Leiden: guaranteeing well-connected communities*, Scientific Reports | Leiden communities: module boundaries, and how Split File partitions a file | `analysis/communities.py`, `analysis/health/refactoring/split_file.py` |++**How we checked it.** The call graph is graded against answer keys we did not write:+the Go team's own call graph, the TypeScript checker, and the SCIP indexers built on+each language's compiler. Precision and recall per language are in+[BENCHMARKS.md](BENCHMARKS.md#accuracy-by-language).++---++## Git history++| Year | Work | What we take from it | Where it lives |+|---|---|---|---|+| 2000 | Mockus and Votta, *Identifying reasons for software changes using historic databases*, ICSM | Classifying a commit as a bug fix from its message | `ingestion/git_indexer/_constants.py` |+| 2005 | Nagappan and Ball, *Use of relative code churn measures to predict system defect density*, ICSE | Churn relative to size as a defect signal | `analysis/health/biomarkers/churn_risk.py` |+| 2005 | Ostrand, Weyuker and Bell, *Predicting the location and number of faults in large software systems*, TSE | Files with past faults have more future faults | `analysis/health/biomarkers/prior_defect.py` |+| 2005 | Śliwerski, Zimmermann and Zeller, *When do changes induce fixes?*, MSR | SZZ: linking a fix back to the change that introduced the bug (calibration labels) | `analysis/change_risk/model.py` |+| 2006 | Kim et al., *Automatic identification of bug-introducing changes*, ASE | The annotation-graph refinement of SZZ (calibration labels) | `analysis/change_risk/model.py` |+| 2007 | Kim et al., *Predicting faults from cached history*, ICSE | Recently fixed code is likely to be fixed again | `analysis/health/biomarkers/prior_defect.py` |+| 2009 | Hassan, *Predicting faults using the complexity of code changes*, ICSE | Change entropy: how scattered a file's changes are | `ingestion/git_indexer/co_change.py` |+| 2009 | D'Ambros, Lanza and Robbes, *On the relationship between change coupling and software defects*, WCRE | Files that change together without an import link | `analysis/health/biomarkers/co_change_scatter.py` |+| 2011 | Bird et al., *Don't touch my code! Examining the effects of ownership on software quality*, ESEC/FSE | Ownership and minor contributors as a defect signal | `analysis/health/biomarkers/ownership_risk.py` |+| 2013 | Kamei et al., *A large-scale empirical study of just-in-time quality assurance*, TSE | The change-level features behind change risk: lines added and deleted, files, directories, subsystems, entropy, author experience | `analysis/change_risk/features.py` |++**How we checked it.** Change risk is a logistic model over the Kamei features,+evaluated leave-one-repository-out on 4,102 commits from 7 repositories: AUC 0.772,+against 0.766 for churn alone. That margin is small and the docs say so; the model's+value is in the reasons it gives, not in beating churn.+[Method](architecture/change-risk.md)++---++## Code health++| Year | Work | What we take from it | Where it lives |+|---|---|---|---|+| 1976 | McCabe, *A complexity measure*, TSE | Cyclomatic complexity | `analysis/health/complexity/cyclomatic.py` |+| 1987 | Karp and Rabin, *Efficient randomized pattern-matching algorithms*, IBM J. R&D | Rolling-hash clone detection: duplicated blocks and Extract Helper | `analysis/health/duplication/rabin_karp.py` |+| 1994 | Chidamber and Kemerer, *A metrics suite for object oriented design*, TSE | Weighted methods per class as god-class evidence | `analysis/health/refactoring/extract_class.py` |+| 1995 | Hitz and Montazeri, *Measuring coupling and cohesion in object-oriented systems* | LCOM4: a class that is really several classes | `analysis/health/complexity/class_analysis.py` |+| 1995 | Bieman and Kang, *Cohesion and reuse in an object-oriented system*, SSR | Tight class cohesion, as Extract Class evidence | `analysis/health/complexity/class_analysis.py` |+| 2001 | El Emam et al., *The confounding effect of class size on the validity of object-oriented metrics*, TSE | Size is a confound: weights are fitted with file size as a control | `analysis/health/scoring.py` |+| 2006 | Lanza and Marinescu, *Object-Oriented Metrics in Practice*, Springer | God Class and Brain Method detection strategies | `analysis/health/biomarkers/god_class.py`, `brain_method.py` |+| 2007 | Meszaros, *xUnit Test Patterns*, Addison-Wesley | Test smells: tests that assert nothing | `analysis/health/biomarkers/assertion_free_test.py` |+| 2018 | Campbell, *Cognitive Complexity*, SonarSource | Nesting-weighted complexity per function | `analysis/health/complexity/cyclomatic.py` |+| 2025 | Zhu et al., FSE ([doi:10.1145/3715741](https://doi.org/10.1145/3715741)) | Recognising mocks, so a test that only checks its own mocks is flagged | `analysis/health/mocks/lexicon.py` |++**How it is evaluated.** The evaluation design follows the defect-prediction+literature, not our convenience:++| Year | Work | What we take from it |+|---|---|---|+| 1988 | DeLong, DeLong and Clarke-Pearson, *Comparing the areas under two or more correlated ROC curves*, Biometrics | Paired AUC comparison against another tool |+| 2009 | Mende and Koschke, *Revisiting the evaluation of defect prediction models*, PROMISE | Popt: ranking quality per line reviewed, not per file |+| 2010 | Mende and Koschke, *Effort-aware defect prediction models*, CSMR | Recall at a fixed 20% review budget |+| 2010 | Jureczko and Madeyski, *Towards identifying software project clusters with regard to defect prediction*, PROMISE | jEdit releases as an external, held-out dataset |++**How we checked it.** Every file in 21 open-source repositories across 9 languages+(2,826 files) was scored at a fixed commit and checked against the next six months of+bug fixes: ROC AUC 0.737 [0.683, 0.787], and 0.76 to 0.78 on the held-out jEdit+releases. Against CodeScene on the 2,770 files both tools scored, ranking by Repowise+surfaces 2.3x the defective files inside a 20% review budget (p = 0.003).+[BENCHMARKS.md](BENCHMARKS.md#5-code-health-predicts-defects)++---++## Program analysis and refactoring++| Year | Work | What we take from it | Where it lives |+|---|---|---|---|+| 1973 | Kildall, *A unified approach to global program optimization*, POPL | The worklist fixpoint for dataflow over a control-flow graph | `analysis/health/dataflow/reaching.py` |+| 1986 | Aho, Sethi and Ullman, *Compilers: Principles, Techniques, and Tools* | Reaching definitions with GEN and KILL sets: which values flow into and out of a span | `analysis/health/dataflow/reaching.py` |+| 1992 | Opdyke, *Refactoring Object-Oriented Frameworks*, PhD thesis, UIUC | Preconditions that make a refactoring behaviour-preserving | `analysis/health/refactoring/preconditions.py` |+| 1999 | Fowler, *Refactoring: Improving the Design of Existing Code* | Extract Method, Extract Class, Move Method and the smells that call for them | `analysis/health/refactoring/` |++Extract Method plans are built on that dataflow pass: it finds the exact span to lift+and infers the parameters and return values a behaviour-preserving signature needs.++---++## Architecture across repositories++| Year | Work | What we take from it | Where it lives |+|---|---|---|---|+| 2006 | MacCormack, Rusnak and Baldwin, *Exploring the structure of complex software designs*, Management Science | Propagation cost: how far a change can travel | `workspace/architecture_metrics.py` |+| 2013 | Sturtevant, *System design and the cost of architectural complexity*, PhD thesis, MIT | Core-periphery roles for services | `workspace/architecture_metrics.py` |+| 2014 | Baldwin, MacCormack and Rusnak, *Hidden structure: using network methods to map system architecture*, Research Policy | The cyclic core and the dependency-structure view | `workspace/architecture_metrics.py` |++---++## Search and retrieval++| Year | Work | What we take from it | Where it lives |+|---|---|---|---|+| 2009 | Robertson and Zaragoza, *The probabilistic relevance framework: BM25 and beyond*, Foundations and Trends in IR | BM25 ranking over full-text search, with per-column weights | `persistence/search.py` |+| 2009 | Cormack, Clarke and Büttcher, *Reciprocal rank fusion outperforms Condorcet and individual rank learning methods*, SIGIR | Merging full-text and semantic results | `packages/server/src/repowise/server/mcp_server/tool_search.py` |++**How we checked it.** On 42 sealed retrieval tasks, held out from every round of+tuning and graded without a model, `get_answer` found 0.876 of the files a fix needed,+against 0.610 for the next tool (sign test p = 0.00004).+[BENCHMARKS.md](BENCHMARKS.md#1-finding-the-right-files)++---++## Decisions++| Year | Work | What we take from it | Where it lives |+|---|---|---|---|+| 2011 | Nygard, *Documenting Architecture Decisions* | The architecture decision record format Repowise reads and writes | `analysis/decisions/adr.py` |+| 2018 | Kopp et al., MADR (Markdown Architectural Decision Records) | The MADR variant of the same format | `analysis/decisions/adr.py` |++---++## Where we changed the original++Research methods are written for a study. A tool has to run on any repository, in any+language, every day. Where we departed from the source, we say so:++- **Weights are fitted.** The original detectors use fixed thresholds. Repowise fits+ the weight of each detector on a defect corpus, with file size as a control, and+ ships only the fitted constants. A detector that cannot show defect lift beyond size+ does not move the defect score; 25 of the 53 detectors do.+- **God Class uses simpler inputs.** Lanza and Marinescu's strategy reads attribute+ access across classes (ATFD). Repowise uses size, method count and the presence of a+ brain method, and gates Brain Method on how central the function is in the graph.+- **Change entropy is decayed and capped.** Older changes weigh less, and very large+ commits are left out so one mass rename does not dominate a file's history.+- **Missing evidence reads as no signal.** A class with no detected member references+ reports LCOM4 of 1, not a high value; an unsupported language produces no health+ findings, not guessed ones.+- **Weak signals stay weak.** Developer congestion and knowledge loss calibrated+ weakly, so their weights are floored at 0.5 and 0.4 instead of being removed or+ inflated.++## What we measured and did not ship++| Idea | Source | What happened |+|---|---|---|+| Refactoring-aware SZZ for bug-introducing commits | Śliwerski et al. 2005 and its successors | 74.5% precision against an 80% gate. Not shipped; SZZ only labels offline calibration data |+| PageRank centrality as a defect predictor | Brin and Page 1998 | Rejected as a predictor on the calibration corpus. Kept for ranking pages and plans, not for the defect score |+| Code naturalness as a defect predictor | Ray et al., *On the naturalness of buggy code*, ICSE 2016 | Tested on the calibration corpus, rejected |+| Change bursts as a defect predictor | Nagappan et al., *Change bursts as defect predictors*, ISSRE 2010 | Tested on the calibration corpus, rejected |++The calibration also reports one result against us: file size alone matches the health+score on ROC AUC (0.742, p = 0.92). The score's measured advantage is in effort-aware+ranking, finding more defects per line reviewed, and that is the claim we make.++## What we do not implement++So nobody infers it from the list above: Repowise does **not** compute Halstead+metrics, the Maintainability Index, the full Chidamber-Kemerer suite (CBO, RFC, DIT),+program slicing in Weiser's sense, or rapid type analysis. Its call graph resolves+receivers by declared type instead.++## What is actually new++The pieces are everybody's. What Repowise adds is putting them in one index and+making them answer questions together: the graph decides which code a git signal+belongs to, health scores the code the graph parsed, tests reach code through graph+edges, and decisions attach to the files they govern. Every answer carries how it was+derived and how far to trust it, it is served to people and to coding agents the same+way, and the whole thing is measured in public against answer keys we did not write.++---++Spotted a wrong attribution, or a technique we use that is missing here? Open an+issue. This page is kept to the same standard as the code it describes.diff --git a/docs/README.md b/docs/README.mdindex 78bd4e382..4bc931fa6 100644--- a/docs/README.md+++ b/docs/README.md@@ -1,119 +1,135 @@# repowise documentation-Index your codebase once. Your agent stops greping, your team stops guessing-which PR is dangerous, and both get their answers from the same place.+Index your codebase once. Your agent stops grepping, your team stops guessing+which change is dangerous, and both get their answers from the same place.++To set up with an agent, paste this into it:++```text+Read https://docs.repowise.dev/setup.md and set up Repowise in this repository.+```++By hand: the [Quickstart](start/QUICKSTART.md) gets you indexed and connected to+your agent in under five minutes, with no API key.<div align="center"><img src="../.github/assets/one-index.svg" alt="One index producing code health, a dependency graph, git history, generated docs, architectural decisions, and ten MCP tools" width="100%" /></div>-New here? **[Quickstart](start/QUICKSTART.md)** gets you indexed and connected to-your agent in under five minutes, with no API key.+For a map of every layer and how they feed each other, read+[layers/INTELLIGENCE_LAYERS.md](layers/INTELLIGENCE_LAYERS.md).----## Pick your path+## I want to...-| You are | Start here | Then |-|---|---|---|-| **A developer wiring up an AI agent** | [Quickstart](start/QUICKSTART.md) | [Supported agents](agent/INTEGRATIONS.md) · [MCP tools](agent/MCP_TOOLS.md) · [Hooks](agent/HOOKS.md) · [Distill](agent/DISTILL.md) |-| **Living in an editor** | [VS Code extension](agent/VSCODE.md) | [Codex](agent/CODEX.md) · [opencode](agent/OPENCODE.md) |-| **A team lead watching what ships** | [Change risk](layers/CHANGE_RISK.md) | [Code health](layers/CODE_HEALTH.md) · [Bug history](layers/BUG_HISTORY.md) |-| **Running many repos** | [Workspaces](scale/WORKSPACES.md) | [Worktrees](scale/WORKTREES.md) · [Auto-sync](scale/AUTO_SYNC.md) |-| **Evaluating repowise to buy it** | [Commercial](business/COMMERCIAL.md) | [Security & compliance](business/SECURITY_COMPLIANCE.md) · [Benchmarks](BENCHMARKS.md) · [Roadmap](../ROADMAP.md) |-| **Contributing** | [Architecture](architecture/README.md) | [CONTRIBUTING](../.github/CONTRIBUTING.md) |+### Give my agent context----+| Doc | What it covers |+|-----|----------------|+| [start/QUICKSTART.md](start/QUICKSTART.md) | Install, index, and connect your agent |+| [agent/INTEGRATIONS.md](agent/INTEGRATIONS.md) | Which agents are supported at what depth, and how to add one |+| [agent/MCP_TOOLS.md](agent/MCP_TOOLS.md) | The ten task-shaped tools, what each answers, and the opt-in extras |+| [agent/HOOKS.md](agent/HOOKS.md) | Context and warnings that reach the agent without it asking |+| [agent/DISTILL.md](agent/DISTILL.md) | Compress noisy command output before your agent reads it |+| [layers/GRAPH.md](layers/GRAPH.md) | The dependency graph, and how much to trust each edge |+| [layers/LANGUAGE_SUPPORT.md](layers/LANGUAGE_SUPPORT.md) | What works per language: 26 parsed to a full AST, 40 on the support ladder |++### Find what to fix first-## Get started+| Doc | What it covers |+|-----|----------------|+| [layers/CODE_HEALTH.md](layers/CODE_HEALTH.md) | Defect risk, maintainability and performance risk per file, and the Fix first queue |+| [layers/REFACTORING.md](layers/REFACTORING.md) | Graph-aware refactoring plans, from Extract Method to Performance Fix |+| [layers/OWNERSHIP.md](layers/OWNERSHIP.md) | Owners, bus factor, and where knowledge is at risk |+| [start/DASHBOARD.md](start/DASHBOARD.md) | Every view in the local dashboard |++### Delete dead code safely| Doc | What it covers ||-----|----------------|-| [start/QUICKSTART.md](start/QUICKSTART.md) | Install, index your repo, and connect your agent in under 5 minutes |-| [start/USER_GUIDE.md](start/USER_GUIDE.md) | The everyday guide: how the pieces fit and the workflows they support |-| [start/DASHBOARD.md](start/DASHBOARD.md) | Every view in the local web dashboard, and what each one answers |-| [start/CI.md](start/CI.md) | Gate pull requests on patch coverage, documentation drift and new security findings, with the GitHub Action and GitLab template |+| [layers/DEAD_CODE.md](layers/DEAD_CODE.md) | Unreachable files, unused exports and zombie packages, with confidence tiers and deletion readiness |-## Connect your AI agent+### Know if a change is risky| Doc | What it covers ||-----|----------------|-| [agent/INTEGRATIONS.md](agent/INTEGRATIONS.md) | Which agents are supported at what depth, generated from the code, plus the recipe for adding one |-| [agent/MCP_TOOLS.md](agent/MCP_TOOLS.md) | The ten task-shaped tools, what each answers, and worked multi-tool examples |-| [agent/HOOKS.md](agent/HOOKS.md) | Proactive delivery: context and warnings that arrive without the agent asking |-| [agent/DISTILL.md](agent/DISTILL.md) | `repowise distill`: compress noisy command output before your agent reads it |-| [agent/CHANGE_REVIEW.md](agent/CHANGE_REVIEW.md) | Reviewing a change from Python, with or without a checkout, and what each evidence state means |-| [agent/VSCODE.md](agent/VSCODE.md) | The VS Code extension: health in the gutter, risk before you push, dashboards in the editor |-| [agent/CODEX.md](agent/CODEX.md) | Wiring repowise into the Codex CLI |-| [agent/OPENCODE.md](agent/OPENCODE.md) | Wiring repowise into opencode |+| [layers/CHANGE_RISK.md](layers/CHANGE_RISK.md) | Score a commit, a `base..head` range or uncommitted work against your repo's own history |+| [layers/BUG_HISTORY.md](layers/BUG_HISTORY.md) | Which files and functions keep getting fixed, and how recently |+| [layers/SECURITY.md](layers/SECURITY.md) | The local scan: 22 pattern kinds plus a symbol-name scan, history scanning for secrets, and its limits |+| [start/CI.md](start/CI.md) | Gate pull requests with the GitHub Action or the GitLab template |-## The intelligence layers+### Know which tests matter| Doc | What it covers ||-----|----------------|-| [layers/INTELLIGENCE_LAYERS.md](layers/INTELLIGENCE_LAYERS.md) | Overview of the five layers: graph, git, docs, decisions, code health |-| [layers/CODE_HEALTH.md](layers/CODE_HEALTH.md) | Defect risk, maintainability, and performance from 51 deterministic detectors |-| [layers/REFACTORING.md](layers/REFACTORING.md) | Concrete, graph-aware refactoring plans (Extract Class, Move Method, Break Cycle) |-| [layers/CHANGE_RISK.md](layers/CHANGE_RISK.md) | Score any commit or `base..HEAD` range 0-10 for defect risk |-| [layers/BUG_HISTORY.md](layers/BUG_HISTORY.md) | Which files and symbols actually get bug-fixed, and how recently |-| [layers/TEST_INTELLIGENCE.md](layers/TEST_INTELLIGENCE.md) | Coverage ingestion, untested hotspots, and running only the tests a diff touches |-| [layers/GRAPH.md](layers/GRAPH.md) | The dependency graph: what is in it, how every edge is resolved, and how much to trust each one |-| [layers/DECISIONS.md](layers/DECISIONS.md) | Architectural decisions mined from your repo and from your own agent sessions |-| [layers/DEAD_CODE.md](layers/DEAD_CODE.md) | Unreachable files, unused exports, and zombie packages by confidence tier |-| [layers/DOC_DRIFT.md](layers/DOC_DRIFT.md) | Documentation claims checked against the tree: what is checked, what is not, and how to read a finding |-| [layers/SECURITY.md](layers/SECURITY.md) | The local pattern scan: what the sixteen patterns catch, what they do not, and how far to trust the result |-| [layers/LANGUAGE_SUPPORT.md](layers/LANGUAGE_SUPPORT.md) | What works per language: 26 parsed to a full AST, 40 on the five-rung ladder |-| [layers/WIKI.md](layers/WIKI.md) | The generated wiki: page types, what `update` re-renders, styles, output language |--## Scale it+| [layers/TEST_INTELLIGENCE.md](layers/TEST_INTELLIGENCE.md) | Coverage ingestion, untested hotspots, and the tests a diff touches, with or without a report |++### Generate and keep docs current| Doc | What it covers ||-----|----------------|-| [scale/WORKSPACES.md](scale/WORKSPACES.md) | Multi-repo intelligence: cross-repo contracts, co-changes, and federated MCP |-| [scale/WORKTREES.md](scale/WORKTREES.md) | Linked git worktrees seed their index from the base checkout, with no flags |-| [scale/AUTO_SYNC.md](scale/AUTO_SYNC.md) | Keep the index fresh automatically on every commit |+| [layers/WIKI.md](layers/WIKI.md) | The generated wiki: page types, styles, and what `update` re-renders |+| [layers/DOC_DRIFT.md](layers/DOC_DRIFT.md) | Claims in your markdown checked against the tree |+| [scale/AUTO_SYNC.md](scale/AUTO_SYNC.md) | Keep the index fresh on every commit |++### Understand why code is this way++| Doc | What it covers |+|-----|----------------|+| [layers/DECISIONS.md](layers/DECISIONS.md) | Architectural decisions mined from your repo and your agent sessions |++### Run across many repos++| Doc | What it covers |+|-----|----------------|+| [scale/WORKSPACES.md](scale/WORKSPACES.md) | Multi-repo intelligence: cross-repo contracts, co-changes, one MCP server |+| [scale/WORKTREES.md](scale/WORKTREES.md) | Linked git worktrees seed their index from the base checkout || [../docker/README.md](../docker/README.md) | Running repowise in Docker |-| [../examples/](../examples/) | Copy-paste walkthroughs (Codex setup and more) |-## Reference+### Evaluate repowise for my company| Doc | What it covers ||-----|----------------|-| [reference/CLI_REFERENCE.md](reference/CLI_REFERENCE.md) | Every command and flag |-| [reference/CONFIG.md](reference/CONFIG.md) | `.repowise/config.yaml`, `health-rules.json`, and environment variables |-| [reference/COMPUTED_GLOSSARY.md](reference/COMPUTED_GLOSSARY.md) | Definitions for every computed metric and term repowise reports |-| [reference/TELEMETRY.md](reference/TELEMETRY.md) | What anonymous telemetry collects, and how to turn it off |-| [reference/UPGRADING.md](reference/UPGRADING.md) | Notes for upgrading between versions |-| [CHANGELOG.md](CHANGELOG.md) | Release history |+| [business/SECURITY_COMPLIANCE.md](business/SECURITY_COMPLIANCE.md) | What leaves your machine, what is stored, and the answers your security team wants |+| [business/COMMERCIAL.md](business/COMMERCIAL.md) | Hosted tier, enterprise options and commercial licensing |+| [BENCHMARKS.md](BENCHMARKS.md) | Every published number with its sample size and method, including the rows we lose |+| [LINEAGE.md](LINEAGE.md) | The published research each layer is built on, where it lives in the code, and how we checked it |+| [../ROADMAP.md](../ROADMAP.md) | What we are building next, and what we are not |-## Evidence+### Fix a problem| Doc | What it covers ||-----|----------------|-| [BENCHMARKS.md](BENCHMARKS.md) | Every published number with its sample size and its test: call-graph precision judged by a compiler we do not control, retrieval against the agent-context field, defect prediction against CodeScene, distillation, and the rows we lose |+| [start/TROUBLESHOOTING.md](start/TROUBLESHOOTING.md) | `repowise doctor`, then install, indexing and answer problems |+| [reference/UPGRADING.md](reference/UPGRADING.md) | Notes for upgrading between versions |-## Teams & business+### Contribute| Doc | What it covers ||-----|----------------|-| [business/COMMERCIAL.md](business/COMMERCIAL.md) | Hosted tier, enterprise (on-prem, SSO/SCIM), and commercial licensing |-| [business/SECURITY_COMPLIANCE.md](business/SECURITY_COMPLIANCE.md) | What leaves your machine, what gets stored, and the answers your security team wants |-| [ROADMAP.md](../ROADMAP.md) | What we are building next, what we are not building, and why there are no dates on it |+| [../.github/CONTRIBUTING.md](../.github/CONTRIBUTING.md) | How to set up, test and submit a change |+| [architecture/README.md](architecture/README.md) | How repowise is built: pipelines, graph algorithms, health internals |-## Architecture & internals+----How repowise is built, for contributors and the curious.+## Reference| Doc | What it covers ||-----|----------------|-| [architecture/ARCHITECTURE.md](architecture/ARCHITECTURE.md) | The system: package layout, pipelines, MCP server |-| [architecture/code-health.md](architecture/code-health.md) | Health internals: marker computation and calibrated weights |-| [architecture/graph-algorithms.md](architecture/graph-algorithms.md) | Every graph algorithm, the intuition plus the math |-| [architecture/language-support.md](architecture/language-support.md) | The language pipeline and how the tiers work |-| [architecture/chat.md](architecture/chat.md) | Codebase chat: agent loop, streaming, artifact panel |-| [architecture/structurizr-export.md](architecture/structurizr-export.md) | Export the architecture as Structurizr DSL and render it anywhere |-| [architecture/editor-files.md](architecture/editor-files.md) | How `CLAUDE.md` and `AGENTS.md` get generated |-| [architecture/savings-accounting.md](architecture/savings-accounting.md) | The agent savings ledger: one canonical event per interaction, priced from its own evidence |-| [architecture/deep-dives.md](architecture/deep-dives.md) | Systems not covered elsewhere |-| [architecture/pluggable-storage.md](architecture/pluggable-storage.md) | The capability seams: storage, graph, vector, CLI, MCP |-| [design/theme-tokens.md](design/theme-tokens.md) | Resolved design tokens and the WCAG contrast matrix |+| [reference/CLI_REFERENCE.md](reference/CLI_REFERENCE.md) | Every command and flag |+| [reference/CONFIG.md](reference/CONFIG.md) | `.repowise/config.yaml`, `health-rules.json`, and environment variables |+| [reference/MCP_RESPONSE_FIELDS.md](reference/MCP_RESPONSE_FIELDS.md) | The field dictionary for every MCP response, including the `_meta` envelope |+| [reference/HOOKS_REFERENCE.md](reference/HOOKS_REFERENCE.md) | The exact settings entries each hook install writes |+| [reference/CHANGE_REVIEW_API.md](reference/CHANGE_REVIEW_API.md) | Reviewing a change from Python, with or without a checkout |+| [reference/WORKSPACE_CONTRACTS.md](reference/WORKSPACE_CONTRACTS.md) | How a workspace names, reads and matches cross-repo contracts |+| [reference/COMPUTED_GLOSSARY.md](reference/COMPUTED_GLOSSARY.md) | Definitions for every computed metric |+| [reference/TELEMETRY.md](reference/TELEMETRY.md) | What anonymous telemetry collects, and how to turn it off |+| [start/USER_GUIDE.md](start/USER_GUIDE.md) | The everyday guide to how the pieces fit |+| [CHANGELOG.md](CHANGELOG.md) | Release history |++Agent and editor pages: [Codex](agent/CODEX.md) ·+[opencode](agent/OPENCODE.md) · [Hermes](agent/HERMES.md) ·+[VS Code](agent/VSCODE.md) ·+[Claude Code as a provider](agent/CLAUDE_CODE_PROVIDER.md) ·+[all integrations](agent/INTEGRATIONS.md).diff --git a/docs/agent/DISTILL.md b/docs/agent/DISTILL.mdindex 0c2b45b7f..99885ad23 100644--- a/docs/agent/DISTILL.md+++ b/docs/agent/DISTILL.md@@ -1,85 +1,84 @@-# Distill — index-aware output distillation+# Distill: index-aware output distillation-AI coding agents burn most of their context window on *output they never-needed*: 300 lines of passing-test dots to find 4 failures, a full `git log`-to learn "what changed recently", a 60k-token diff to review one hunk. Distill-compresses that output **before the agent reads it** — errors first, structure-preserved, everything reversible.+Coding agents spend much of their context window on output they never needed:+300 lines of passing tests around 4 failures, a full `git log` to learn what+changed recently, a 60k-token diff to review one hunk. Distill compresses that+output before the agent reads it. Errors come first, structure is kept, and+everything dropped can be restored.-Distill is a **capability**, not a sixth intelligence layer: it reuses the-index the five layers already build (symbol bounds, graph centrality,-hotspots) to decide *what to keep*, instead of compressing blind.+Distill is a capability, not an intelligence layer of its own. It reuses what+the index already knows (symbol bounds, graph centrality, hotspots) to decide+what to keep. It needs no LLM key, and its filters work on repos with no index.```bashrepowise distill pytest -x # run pytest, print a compact errors-first renderingrepowise expand a1b2c3d4e5f6 # restore anything that was omitted-repowise saved # tokens & dollars saved so far+repowise saved # tokens and dollars saved so far```-**Guarantees** (enforced by the engine, asserted by tests):+**Guarantees** (enforced by the engine and asserted by tests):-- **Errors always survive.** Every error/failure-classified line in the raw+- **Errors survive.** Every line classified as an error or failure in the rawoutput appears in the distilled rendering.-- **Fully reversible.** Raw output is stored *before* any marker is emitted;- `repowise expand <ref>` round-trips it byte-for-byte.-- **Fallback to raw.** Any filter error, storage failure, or non-improvement- prints the original output unchanged. Distillation can never lose output.-- **Net-positive only.** Output is only distilled when it actually gets- smaller (marker included) — small outputs pass through untouched.-- **Exit codes preserved.** `repowise distill <cmd>` is a drop-in wrapper in- scripts and agent tool calls alike.+- **Reversible.** Raw output is stored before any marker is printed;+ `repowise expand <ref>` returns it byte-for-byte.+- **Fallback to raw.** A filter error, a storage failure or a rendering that is+ not smaller prints the original output unchanged.+- **Net-positive only.** Output is distilled only when it gets smaller, marker+ included. Small outputs pass through untouched.+- **Exit codes preserved.** `repowise distill <cmd>` is a drop-in wrapper for+ scripts and agent tool calls.---## The surfaces-### 1. `repowise distill <cmd>` — the executor+### 1. `repowise distill <cmd>`: the executor-Runs the command (shell semantics preserved), captures stdout+stderr, picks a-filter by command shape (then by content sniff), and prints the compact-rendering. Eleven filters ship:+Runs the command (shell semantics preserved), captures stdout and stderr, picks+a filter by command shape and then by content, and prints the compact+rendering. These filters ship:| Filter | Commands | What it keeps ||---|---|---|-| `test_output` | pytest, jest, vitest, cargo test, go test | failures + assertion details + summary; collapses pass parades |-| `build_output` | npm/tsc/cargo/go builds | errors and warnings grouped; strips progress/boilerplate |-| `lint_output` | eslint/biome, ruff/flake8/mypy, clippy, golangci-lint | errors verbatim; warnings grouped by rule id with counts + file:line anchors; fixable totals |-| `install_output` | pip, uv, poetry, npm/pnpm/yarn install, cargo install, brew, bundle, composer | what actually changed + the final summary; drops resolver and download progress. An all-satisfied run collapses to `install: ok (no changes)` |-| `infra_plan` | `terraform`/`tofu plan`, `helm diff`/`upgrade` | the `Plan: N to add…` summary and the resources that change; drops unchanged-resource dumps. A no-op collapses to `plan: no changes` |+| `test_output` | pytest, jest, vitest, cargo test, go test | failures, assertion details and the summary; collapses runs of passes |+| `build_output` | npm/tsc/cargo/go builds | errors and warnings grouped; strips progress and boilerplate |+| `lint_output` | eslint/biome, ruff/flake8/mypy, clippy, golangci-lint | errors verbatim; warnings grouped by rule id with counts and file:line anchors; fixable totals |+| `install_output` | pip, uv, poetry, npm/pnpm/yarn install, cargo install, brew, bundle, composer | what changed and the final summary; drops resolver and download progress. An all-satisfied run collapses to `install: ok (no changes)` |+| `infra_plan` | `terraform`/`tofu plan`, `helm diff`/`upgrade` | the plan summary and the resources that change. A no-op collapses to `plan: no changes` || `git_status` | `git status` | porcelain-style compact status |-| `git_log` | `git log` | recent subjects + counts |-| `git_diff` | `git diff`/`show` | stat + the most relevant hunks |-| `search_results` | grep / rg floods | grouped-by-file digest with per-file counts and anchors |+| `git_log` | `git log` | recent subjects and counts |+| `git_diff` | `git diff`/`show`, `gh pr diff` | stat plus the most relevant hunks (`--stat` output has its own `git_diff_stat` filter) |+| `search_results` | grep / rg floods | digest grouped by file, with per-file counts and anchors || `file_listing` | ls / tree / find | grouped tree rendering |-| `logs` | anything log-shaped | template-collapse with counts (timestamps/ids normalized) |+| `logs` | anything log-shaped | template collapse with counts (timestamps and ids normalized) |-Dropped content is stored in the omission store and referenced inline:+Dropped content goes to the omission store and is referenced inline:```[repowise#a1b2c3d4e5f6: 230 lines omitted (~6.1k tokens); restore: repowise expand a1b2c3d4e5f6]```-### 2. `repowise expand <ref>` — the reversal+### 2. `repowise expand <ref>`: the reversal```bashrepowise expand a1b2c3d4e5f6 # full original outputrepowise expand a1b2c3d4e5f6 -q "FAILED" # only the matching lines-repowise expand "[repowise#a1b2…]" # a pasted whole marker works too+repowise expand "[repowise#a1b2...]" # a pasted whole marker works too```Looks in the current repo's store first, then the user-level fallback store.-MCP clients without a shell can resolve the same refs through-`get_symbol("repowise#<ref>")` — see [MCP](#5-mcp-response-budget--_metaomitted).+MCP clients without a shell resolve the same refs with+`get_symbol("repowise#<ref>")` (see [MCP](#5-mcp-response-budget-_metaomitted)).### 3. The command-rewrite hook (Claude Code + Codex)-A PreToolUse hook rewrites noisy agent commands —-`pytest -x` → `repowise distill pytest -x`. The default posture is `allow`, so-the rewrite runs without a prompt; it is a bounded substitution, never a new-command (see [Safety model](#safety-model-in-one-place)). Set-`permission: ask` in `.repowise/config.yaml` to review each rewrite instead.-`repowise init` offers to install the hook (default: yes); install manually-with:+A PreToolUse hook rewrites noisy agent commands: `pytest -x` becomes+`repowise distill pytest -x`. Rewrites run without a prompt by default+(`permission: allow`). That is a bounded substitution, never a new command (see+[Safety model](#safety-model-in-one-place)). Set `permission: ask` in+`.repowise/config.yaml` to approve each rewrite. `repowise init` offers to+install the hook (default: yes), or install it yourself:```bashrepowise hook rewrite install # or opt in during `repowise init`@@ -87,63 +86,51 @@ repowise hook rewrite statusrepowise hook rewrite uninstall # removes only the repowise entries```-The hook covers both Claude Code shell tools — Bash and, on Windows,-PowerShell. Existing installs are widened automatically on the next-`install`/`init`.--**Codex.** When `~/.codex` exists, `install` also covers the Codex CLI, with-two honest caveats its hook protocol imposes:--- Codex applies a PreToolUse command rewrite only from **version 0.137**;- on older builds the hook entry is skipped (a rewrite response would error- on every shell call). `repowise hook rewrite status` reports what your- build can do.-- Codex has **no ask-with-mutation** — a rewrite can only be auto-allowed,- never shown for approval. So under Codex, rewrites fire **only for families- resolving to `permission: allow`** in `.repowise/config.yaml`, which the- default posture already is; any family you set to `ask` passes through- unchanged there rather than being silently escalated.--The hook entry lands in `~/.codex/hooks.json` (one install covers every-repo); Codex requires new hooks to be reviewed — run `/hooks` inside Codex to-trust it. Independently of any hook, `install` maintains a marker-managed-**"Output Distillation" section in the repo's `AGENTS.md`** teaching the-agent to run `repowise distill <cmd>` voluntarily and to `repowise expand`-markers instead of re-running commands — this works on every Codex version,-including ones with no usable rewrite hook. `uninstall` removes the section-and restores your AGENTS.md byte-for-byte.--The hook is deliberately conservative. It never rewrites:--- redirections, compound commands (`>`, `&&`, `;`, backticks, `$()`) and- almost all pipes. Two safe shapes are carved out: a trailing `2>&1`- (distill merges stderr into its capture anyway), and, on macOS/Linux only,- a single pipe into a bare stdin filter — `head`, `tail`, `grep`, `egrep`,- `fgrep`, `rg` — which runs unchanged inside distill's own shell- (`pytest -q | head -50` → `repowise distill "pytest -q | head -50"`).- The `grep`/`rg` pattern-file forms (`-f`, `--file`) are excluded: they read- a file as configuration rather than filtering stdin-- watch/follow modes (`--watch`, `tail -f`, …)-- anything on the trivial-command ignore-list, or already-prefixed commands-- PowerShell-native constructs: `Verb-Noun` cmdlets, `& "path"` invocations,- backtick continuations — and, from PowerShell, alias tokens (`ls`, `cat`,- `find`, …) that don't mean what their unix namesakes mean+The hook covers both Claude Code shell tools: Bash and, on Windows, PowerShell.++**Codex.** When `~/.codex` exists, `install` also covers the Codex CLI, with two+limits its hook protocol imposes:++- Codex applies a PreToolUse rewrite only from **version 0.137**. On older+ builds the hook entry is skipped. `repowise hook rewrite status` reports what+ your build can do.+- Codex cannot show a rewritten command for approval. Under Codex, rewrites fire+ only for families resolving to `permission: allow` (the default); a family set+ to `ask` passes through unchanged.++The Codex entry lands in `~/.codex/hooks.json`, so one install covers every+repo. Codex asks you to review new hooks: run `/hooks` inside Codex to trust it.+Separately, `install` maintains a marker-managed "Output Distillation" section+in the repo's `AGENTS.md` that teaches the agent to run `repowise distill <cmd>`+itself and to `repowise expand` markers instead of re-running commands. This+works on every Codex version. `uninstall` removes the section and restores your+`AGENTS.md` byte-for-byte.++The hook is conservative. It never rewrites:++- redirections and compound commands (`>`, `&&`, `;`, backticks, `$()`), and+ almost all pipes. Two shapes are allowed: a trailing `2>&1` (distill merges+ stderr anyway), and on macOS/Linux a single pipe into `head`, `tail`, `grep`,+ `egrep`, `fgrep` or `rg`, which runs unchanged inside distill's own shell+ (`pytest -q | head -50` becomes `repowise distill "pytest -q | head -50"`). The+ `grep`/`rg` pattern-file forms (`-f`, `--file`) are excluded.+- watch and follow modes (`--watch`, `tail -f`)+- trivial commands on the ignore list, and commands already prefixed+- PowerShell-native constructs (`Verb-Noun` cmdlets, `& "path"` invocations,+ backtick continuations) and, from PowerShell, alias tokens such as `ls`,+ `cat` and `find` that do not mean what their unix namesakes mean- commands in repos that have not opted into repowise (no `.repowise/` upward)-These decisions are made by tokenizing the command, not by scanning it for-punctuation, so on macOS/Linux a shell character sitting inside quotes is-treated as the text it is: `pytest -k "a|b"` is a normal test run, not a-pipeline, and is rewritten as one. On Windows the blunter rule stands — any-`|`, `&`, `;`, `<`, `>` or backtick anywhere in the command, quoted or not,-passes it through untouched — because `cmd.exe` re-parses metacharacters that-the argv quoting Windows uses does not cover. The conservative answer always-wins the ambiguity; an unrecognized shape is left alone.--**The allowlist trap.** A rewrite changes the command string, so a Claude Code-permission rule you already had — say `Bash(git diff:*)` — no longer matches-the rewritten `repowise distill git diff …`, and the permission prompt comes-back for commands you had already allowed. The fix is one extra allow rule-covering the distill prefix:+On macOS/Linux these decisions come from tokenizing the command, so a shell+character inside quotes is text: `pytest -k "a|b"` is a test run and is+rewritten. On Windows any `|`, `&`, `;`, `<`, `>` or backtick anywhere in the+command passes it through untouched, because `cmd.exe` re-parses metacharacters+that argv quoting does not cover. An unrecognized shape is always left alone.++**Allow rules under `ask`.** A rewrite changes the command string. If you set+`permission: ask`, an allow rule you already had, such as `Bash(git diff:*)`, no+longer matches `repowise distill git diff ...`, and the prompt comes back. One+extra rule covers the distill prefix:```jsonc// ~/.claude/settings.json@@ -155,60 +142,51 @@ covering the distill prefix:}```-`repowise hook rewrite install` offers to add these for you (or pass-`--allow-rule` / `--no-allow-rule` to decide non-interactively). The default-posture stays `ask` — the rule only stops *double*-asking for commands you've-already vetted; `repowise distill` runs the wrapped command unchanged and-never widens what it can do.+`repowise hook rewrite install --allow-rule` adds these for you. Under the+default `allow` posture they are not needed. `repowise distill` runs the+wrapped command unchanged and never widens what it can do.-Per-repo behavior is configured under `distill.commands` in-`.repowise/config.yaml` — see [Configuration](#configuration). Declining the-`repowise init` prompt writes `distill.commands.enabled: false`, so a hook-installed globally from another repo stays inert in this one. A multi-repo-workspace `init` asks once and records the verdict in **every selected-repo**; `repowise hook rewrite install -w` re-enables them all later. The-hook answers in well under 100 ms (stdlib-only hot path, no database).+Per-repo behavior lives under `distill.commands` in `.repowise/config.yaml` (see+[Configuration](#configuration)). Declining the `repowise init` prompt writes+`distill.commands.enabled: false`, so a hook installed globally from another+repo stays inert in this one. A multi-repo workspace `init` asks once and+records the answer in every selected repo; `repowise hook rewrite install -w`+re-enables them all later.`repowise init` also adds a short "Output Distillation" section to the managed-`CLAUDE.md`, teaching the agent to prefer `repowise distill <cmd>` voluntarily-and to expand markers instead of re-running commands — this works in **any**-agent that runs shell commands, hook or no hook.+`CLAUDE.md`, so any agent that runs shell commands can use distill voluntarily,+hook or no hook.-### 4. Read intelligence — skeletons and stale-read notices+### 4. Read intelligence: skeletons and stale-read noticesThe index knows every symbol's line bounds, so repowise can render a **file-skeleton** — every signature, imports, and the bodies of only the most central-symbols — without parsing anything at query time:+skeleton**: every signature, the imports, and the bodies of only the most+central symbols, without parsing anything at query time:```get_context(["src/big_module.py"], include=["skeleton"])```-A typical 600-line file skeletonizes to ~15% of its full tokens with every-signature present. Body selection is importance-ranked (symbol PageRank,-hotspot bit, query match) — this is where the index makes distillation-smarter than blind truncation.+Body selection is ranked by importance (symbol PageRank, hotspot bit, query+match), which is where the index beats blind truncation.-The existing PostToolUse hook complements this passively:+The PostToolUse hook adds passive read surfaces (details in+[HOOKS.md](HOOKS.md#posttooluse-enrichment-on-tool-calls)):-- **Skeleton replacement** — an unbounded `Read` of a large indexed file is- served *as* its skeleton, elision ranges intact, once per file per session- (opt-in: `hooks.read_skeleton`). The advisory version of this — a one-line- nudge naming the skeleton's token cost — was retired after 516 firings- showed no effect above the base rate.-- **Re-read collapse** — a `Read` of a file this session already read, at the- same range, with no intervening edit and an identical content hash, is served- as a short notice naming the earlier read (opt-in: `hooks.read_reread`).- Never applied twice in a row, so one more Read always returns the content.-- **Stale-read notice** — after an `Edit`/`Write`, a later `Read` of the same- file warns that earlier excerpts predate the edit.-- **Search digest** — grep floods (≥50 lines) get a compact grouped-by-file- digest ordered by graph centrality.+- **Skeleton replacement** (opt-in, `hooks.read_skeleton`): an unbounded `Read`+ of a large indexed file is served as its skeleton, once per file per session.+- **Re-read collapse** (opt-in, `hooks.read_reread`): a `Read` of a range this+ session already read, with no edit between and identical bytes, is served as+ a short notice. Never twice in a row, so one more Read always returns content.+- **Stale-read notice**: after an `Edit`/`Write`, a later `Read` of the same file+ warns that earlier excerpts predate the edit.+- **Search digest**: grep floods (50 or more lines) get a digest grouped by file+ and ordered by graph centrality.-### 5. MCP response budget — `_meta.omitted`+### 5. MCP response budget: `_meta.omitted`-All MCP tool responses were always token-budgeted; before Distill, truncation-was silent. Now every drop goes through the same omission store:+MCP tool responses are token-budgeted, and every drop goes through the same+omission store:```jsonc"_meta": {@@ -220,40 +198,37 @@ was silent. Now every drop goes through the same omission store:}```-`get_symbol` resolves omission refs as well as symbol ids — the-`repowise#<12-hex>` shape is unambiguous next to `path/to/file.py::Name`, and-the optional `query` parameter searches within the stored content. Resolving-refs adds no new tool to the surface. See [MCP_TOOLS.md](MCP_TOOLS.md).+`get_symbol` resolves omission refs as well as symbol ids. The+`repowise#<12-hex>` shape cannot be confused with `path/to/file.py::Name`, and+the optional `query` parameter searches within the stored content. See+[MCP_TOOLS.md](MCP_TOOLS.md).-> **Note:** MCP tool calls now record a *counterfactual* saving — what raw file-> exploration the curated answer replaced — as `mcp:<tool>` rows in the same-> ledger (see [`repowise saved`](#repowise-saved--the-savings-report)). Response-> truncation is folded into that delta (the delivered size is measured after the-> budget cap), so it is never double-counted.+MCP calls also record a counterfactual saving (the raw file exploration the+answer replaced) as `mcp:<tool>` rows in the same ledger. Truncation is folded+into that figure, so it is never counted twice.---## The omission store-`.repowise/omissions/omissions.db` — a SQLite sidecar (WAL), deliberately-separate from `wiki.db` so hook-time writes never contend with indexing.-Falls back to `~/.repowise/omissions/` when the current directory is not-inside a repowise repo.+`.repowise/omissions/omissions.db` is a SQLite sidecar, kept separate from+`wiki.db` so hook-time writes never contend with indexing. Outside a repowise+repo it falls back to `~/.repowise/omissions/`.-- Content is keyed by a 12-hex truncated SHA-256 — the same ref that appears- in markers, so one store serves the CLI, the hook, and MCP.-- **Durable across sessions** by design: an agent resuming work tomorrow can- still expand yesterday's markers.-- Pruning is TTL + size-cap based (`7 days` / `50 MB` by default, configurable),- applied opportunistically on write. The most recent row is never evicted, so- a just-rendered marker cannot dangle.+- Content is keyed by a 12-hex truncated SHA-256: the same ref that appears in+ markers, so one store serves the CLI, the hook and MCP.+- It is durable across sessions: an agent resuming tomorrow can still expand+ yesterday's markers.+- Pruning is by age and size (7 days and 50 MB by default, configurable),+ applied on write. The most recent row is never evicted, so a marker just+ printed cannot dangle.----## `repowise saved` — the savings report+## `repowise saved`: the savings report```bash-repowise saved # per-operation rollup + totals + est. dollars+repowise saved # per-operation rollup, totals, estimated dollarsrepowise saved --by surface # distill vs hooks vs MCPrepowise saved --by agent # which agent the savings went torepowise saved --by day # daily rollup@@ -262,75 +237,51 @@ repowise saved --missed # savings raw commands left on the tablrepowise saved --missed --missed-days 30```-Savings are priced at each event's own rate, captured when the event was-recorded, so a later price change never rewrites what a past saving was worth.-Events recorded without a rate are reported as unpriced rather than valued at-today's model, and the command prints both figures. Saved tokens are input the-agent never had to read, so the input rate applies; the pricing table is the-one `repowise costs` uses. Token counts are chars/4 estimates.+One ledger covers the distill command and hook path, the hooks that replace a+tool result, and MCP calls (each answer counted against the raw exploration it+replaced). `--by surface` separates them.-The report covers every capture surface in one ledger: the **distill-command/hook path**, the **hooks that replace a tool result**, and **MCP-calls**, where each answer is counted against the raw exploration it replaced.-`repowise saved --by surface` separates them.+Savings are priced at each event's own rate, captured when it was recorded, so+a later price change never rewrites a past saving. Events recorded without a+rate are reported as unpriced. Saved tokens are input the agent never read, so+the input rate applies, from the same pricing table `repowise costs` uses. Token+counts are chars/4 estimates.The total keeps two kinds of evidence apart. **Measured** savings compare a-known before and after — bytes that really moved. **Inferred** savings estimate-the exploration an answer replaced. The command prints the split, and calls the-total "estimated" whenever any of it is inferred.--### Missed savings — `repowise saved --missed`--The adoption feedback loop: how many tokens did raw commands waste that a-filter would have caught? `--missed` scans your local Claude Code transcripts-for Bash/PowerShell tool calls in this repo that were **not** routed through-`repowise distill`, classifies each with the same router the engine uses, and-estimates the foregone savings using each filter's *conservative* fixture-floor (the per-filter minimums asserted in CI, not the medians — the estimate-undersells on purpose). The scan covers the last 7 days by default-(`--missed-days N` for more); plain `repowise saved` appends a one-line-summary when there is anything to report.--The scan is read-only and best-effort: malformed or absent transcripts mean-an empty report, never an error.--> **Privacy:** the scan stays entirely local. Commands and outputs are read-> from your own transcript directory (`~/.claude/projects/…`) on this-> machine; nothing is uploaded, recorded, or sent anywhere. Codex transcripts-> are not yet scanned.--The local dashboard goes further. The Costs page leads with a **savings hero-card** that combines two surfaces into one honest number:--- **Distill** — the `repowise distill` command/hook ledger above.-- **MCP tool savings** — what each tool answer replaced: the raw file- exploration the agent would have done without it (`source='mcp:*'`).- `get_symbol` stands in for reading the whole file, `get_context` for the files- its skeletons summarise, `search_codebase` for opening each cited file; the- estimate undersells. Tools without a counterfactual estimator still contribute- their response-budget truncation drops.--The dollar figure is **priced at the coding agent's actual model**, not a flat-guess: saved tokens are *input* the agent never read, so they are worth that-agent's input rate. The dashboard detects the model from your local agent-transcripts — the most-recent model that touched the repo across Claude Code-(`~/.claude/projects/…`) and Codex (`~/.codex/sessions/…`) — and falls back to-a sensible default when nothing is detectable. Detection is read-only and stays-on this machine. The missed-savings scan rides along as an "unlock more"-prompt.+known before and after. **Inferred** savings estimate the exploration an answer+replaced. The command prints the split and calls the total "estimated" whenever+any of it is inferred.++### Missed savings: `repowise saved --missed`++How many tokens did raw commands waste that a filter would have caught?+`--missed` scans your local Claude Code transcripts for Bash/PowerShell calls in+this repo that did not go through `repowise distill`, classifies each with the+engine's own router, and estimates the foregone savings from each filter's+conservative fixture floor (the per-filter minimums asserted in CI, not the+medians). The default window is 7 days (`--missed-days N` for more). Plain+`repowise saved` appends a one-line summary when there is something to report.++The scan is read-only and local. Commands and outputs are read from your own+transcript directory (`~/.claude/projects/...`); nothing is uploaded. Codex+transcripts are not scanned.++The local dashboard's Costs page leads with a savings card that combines the+distill ledger with MCP tool savings. The dollar figure uses the input rate of+the model your coding agent actually ran, detected from local Claude Code and+Codex transcripts, with a default when nothing is detectable.----## `repowise corrections` — recurring command fumbles+## `repowise corrections`: recurring command fumbles-The same transcript reader, pointed at a different waste: commands the agent-got *wrong* and then fixed. The scan finds consecutive runs of the same base-command where the first failed (the transcript records a real exit code) and-a later variant succeeded, classifies the fumble, and aggregates the-recurring rules:+The same transcript reader, pointed at commands the agent got wrong and then+fixed. It finds consecutive runs of one base command where the first failed+(the transcript records a real exit code) and a later variant succeeded,+classifies the fumble, and aggregates recurring rules:```bash-repowise corrections # report-only (default window: 30 days)+repowise corrections # report only (default window: 30 days)repowise corrections --days 60repowise corrections --write # maintain the managed guidance block```@@ -342,49 +293,44 @@ repowise corrections --write # maintain the managed guidance block| unknown flag | `pytest` does not support `--looponfail` || missing arg | `tool` needs `--required-thing` |-Classification is deliberately precision-first: apart from the structural-wrong-tool case, a rule only forms when the error text corroborates it (the-dropped flag or path is actually named by the error) — a red-green dev loop-re-running tests with different selections never becomes a "correction".-Wrong-path rules consult the symbol index when available and note where the-corrected target actually lives.--`--write` (strictly opt-in) maintains a short **"Known command corrections"**-managed block — most-frequent rules first, at least 2 occurrences each,-capped at 10 lines — between `REPOWISE_CORRECTIONS` markers in the repo's-`.claude/CLAUDE.md` (and `AGENTS.md` when one exists), so the next agent-session is told up front. Re-running `--write` refreshes the block in place;-when no rule clears the threshold anymore, the block is removed. Content-outside the markers is never touched.+Classification favors precision. Apart from the structural wrong-tool case, a+rule forms only when the error text names the dropped flag or path, so a+red-green loop re-running tests with different selections never becomes a+"correction". Wrong-path rules consult the symbol index when available.-The same privacy contract as the missed scan applies: read-only, best-effort,-entirely local.+`--write` (opt-in) maintains a **"Known command corrections"** block between+`REPOWISE_CORRECTIONS` markers in the repo's `.claude/CLAUDE.md` (and+`AGENTS.md` when one exists): the most frequent rules first, at least 2+occurrences each, at most 10. Re-running refreshes it; when no rule clears the+threshold the block is removed. Content outside the markers is never touched.+The same privacy contract as the missed scan applies.---## Measured savings-On a public OSS repository (microdot), one run per command, tokens estimated-chars/4 — the same estimator the ledger uses:+On a public OSS repository (microdot), one run per command, tokens estimated as+chars/4 (the ledger's estimator):| Command | Raw tokens | Distilled | Saved ||---|---:|---:|---:|-| `pytest -q` (11 failures) | 3,374 | 1,317 | **61%** — all 11 `FAILED` lines preserved |+| `pytest -q` (11 failures) | 3,374 | 1,317 | **61%**, all 11 `FAILED` lines kept || `git log -50` | 3,064 | 331 | **89%** || `git diff` (30 commits of history) | 62,833 | 8,635 | **86%** |-| `git log --oneline -30` | 321 | 321 | 0% — already compact, passed through |-| `git status` (clean tree) | 83 | 83 | 0% — too small to distill, passed through |+| `git log --oneline -30` | 321 | 321 | 0%, already compact, passed through |+| `git status` (clean tree) | 83 | 83 | 0%, too small to distill, passed through |-The 0% rows are the net-positive guard working: distill never bloats small-output. In an end-to-end agent spot-check on the same repo (a seeded-11-failure bug), the agent diagnosed the exact root-cause line and fix from-the distilled test output — identical conclusion to the raw-output run.+The 0% rows are the net-positive guard: distill never bloats small output. In an+agent spot-check on the same repo (a seeded 11-failure bug), the agent found the+root-cause line and fix from the distilled test output, the same conclusion as+the raw-output run.-Fixture-suite medians across the core filters: ≥60% reduction on-test/build/lint output with zero error-line loss (asserted in CI).+Fixture-suite medians across the core filters: at least 60% reduction on+test, build and lint output with zero error-line loss (asserted in CI).-Install and infra-plan logs compress harder, because they are mostly repeated-boilerplate. Measured on real captures from this repository, one run each:+Install and infra-plan logs compress harder because they are mostly repeated+boilerplate. Measured on real captures from the repowise repository, one run+each:| Command | Saved ||---|---:|@@ -392,8 +338,8 @@ boilerplate. Measured on real captures from this repository, one run each:| `pip install` (everything already satisfied) | collapses to `install: ok (no changes)` || `terraform plan` | **80.2%** |-Every one of those still round-trips through `repowise expand`, and the error-line in a failing install survives verbatim.+Each still round-trips through `repowise expand`, and the error line in a+failing install survives verbatim.---@@ -406,8 +352,8 @@ distill:enabled: true # master switch for this repocommands:enabled: true # the command path (CLI + hook rewrites)- permission: allow # ask | allow | off — hook posture (default allow)- families: # per-filter overrides+ permission: allow # ask | allow | off: hook posture (default allow)+ families: # per-filter overrides: ask | allow | off | denytest_output: allow # auto-allow rewrites for test runsgit_diff: deny # never rewrite git diff heredisabled_filters: [] # filters to skip entirely, e.g. [logs]@@ -416,10 +362,10 @@ distill:max_mb: 50 # size cap, oldest pruned first```-Everything defaults sensibly with no block present. `repowise doctor`-validates the block (unknown keys, bad permission values, unknown filter-names, non-positive store sizing) and reports the store size against its cap-and whether the rewrite hook is installed.+With no block present, these defaults apply. `repowise doctor` validates the+block (unknown keys, bad permission values, unknown filter names, non-positive+store sizes) and reports the store size against its cap and whether the rewrite+hook is installed. Full key reference: [CONFIG.md](../reference/CONFIG.md#the-distill-block).---@@ -427,18 +373,18 @@ and whether the rewrite hook is installed.| Risk | Mitigation ||---|---|-| A filter eats a critical line | errors-first invariant + fixture tests + `expand` recovery + fallback-to-raw |-| Silent permission escalation | rewrites default to `allow`, which is not an escalation: a rewrite is always `repowise distill <one recognized command>` drawn from a closed family set, never an arbitrary command smuggled behind the wrapper. Set `permission: ask` to approve each one. Codex has no ask primitive, so only families explicitly set to `allow` rewrite there |-| Marker with nothing behind it | content stored *before* the marker renders; store failure ⇒ raw output |-| Compound-command semantics | compound commands, substitution and redirects are never rewritten. The one pipe shape that is (a single stage into a bare stdin filter, macOS/Linux only) is passed through as one quoted token and runs verbatim in distill's shell; anything that could break out of that quoting bails |-| Unindexed or stale repo | filters work index-free; index only improves ranking |-| Store growth | TTL + size cap, pruned on write; `repowise doctor` reports size |+| A filter eats a critical line | errors-first invariant, fixture tests, `expand` recovery, fallback to raw |+| Silent permission escalation | a rewrite is always `repowise distill <one recognized command>` from a closed family set, never an arbitrary command behind the wrapper. Set `permission: ask` to approve each one. Codex has no ask primitive, so only families resolving to `allow` rewrite there |+| Marker with nothing behind it | content is stored before the marker prints; a store failure prints raw output |+| Compound-command semantics | compound commands, substitution and redirects are never rewritten. The one pipe shape that is (macOS/Linux only) is passed as one quoted token and runs verbatim in distill's shell; anything that could break out of that quoting bails |+| Unindexed or stale repo | filters work without an index; the index only improves ranking |+| Store growth | age and size caps, pruned on write; `repowise doctor` reports size |---## See also-- [CLI_REFERENCE.md](../reference/CLI_REFERENCE.md) — `distill`, `expand`, `saved`, `hook rewrite`-- [MCP_TOOLS.md](MCP_TOOLS.md) — `_meta.omitted`, skeleton include, `get_symbol` ref overload-- [CONFIG.md](../reference/CONFIG.md) — the `distill:` block-- [INTELLIGENCE_LAYERS.md](../layers/INTELLIGENCE_LAYERS.md) — the five layers whose index Distill reuses+- [CLI_REFERENCE.md](../reference/CLI_REFERENCE.md#repowise-distill-command): `distill`, `expand`, `saved`, `hook rewrite`+- [MCP_TOOLS.md](MCP_TOOLS.md): `_meta.omitted`, the skeleton include, `get_symbol` refs+- [CONFIG.md](../reference/CONFIG.md#the-distill-block): the `distill:` block+- [HOOKS.md](HOOKS.md): every agent hook, including the read surfacesdiff --git a/docs/agent/HOOKS.md b/docs/agent/HOOKS.mdindex ac381fdf9..6041d39de 100644--- a/docs/agent/HOOKS.md+++ b/docs/agent/HOOKS.md@@ -1,17 +1,18 @@# Hooks-repowise installs a set of lightweight hooks so context reaches your agent, and-your index stays fresh, with zero effort on your part. They fall into two-families:+repowise installs small hooks so context reaches your agent and your index+stays fresh without you asking. There are two families:-- **Git hooks** keep the wiki and graph in sync with your code.-- **Agent hooks** feed graph, git, health, and decision context into Claude Code- and Codex at the exact moments the agent needs it.+- **Git hooks** keep the wiki and graph in step with your code.+- **Agent hooks** feed graph, git, health and decision context into Claude Code+ and Codex at the moment the agent needs it.-Every agent hook shares the same guarantees: **no LLM calls, no network**, only-local SQLite (`wiki.db`) and `git` reads. They are import-isolated (cold start-under ~500ms), and any failure exits `0` silently, so a broken environment never-crashes or blocks your agent.+Every agent hook makes **no LLM calls and no network calls**. It reads the local+index (`wiki.db`) and `git`. Any failure exits `0` silently, so a broken+environment never blocks your agent.++For the exact settings entries each install writes, see the+[hooks reference](../reference/HOOKS_REFERENCE.md).---@@ -21,248 +22,175 @@ crashes or blocks your agent.|------|--------|--------------|----------|--------------|| **Post-commit auto-sync** | git | `repowise init` by default (`--no-hook` skips it), or `repowise hook install` | every `git commit` | Runs `repowise update` in the background so the wiki tracks your code || **Pre-commit security check** | git | `repowise hook install --security` (opt-in) | every `git commit` | Runs `repowise security check --staged` and blocks a commit that adds a finding at or above `high` ([details](../layers/SECURITY.md#before-a-commit---staged)) |-| **SessionStart context** | Claude Code | `repowise init` | session `startup` / `resume` / `clear` | Live index-freshness line, core-tool trust rule, and the standing decisions relevant to this session |-| **PostToolUse enrichment** | Claude Code | `repowise init` | `Grep` / `Glob` / `Read` / `Edit` / `Write` / repowise MCP calls | Graph context on searches, read-intelligence notices, and edit-time "governed by" decision notices |-| **Wrong-path rescue** | Claude Code | `repowise init` | a `Read` / `Edit` / `Write` / `Grep` / `Glob` / `NotebookEdit` that failed on a path this tree does not have | Names the file when exactly one indexed file carries that basename; silent otherwise |-| **Coverage re-ingest** (opt-in) | Claude Code | `hooks.coverage_reingest: true`, then the next `repowise coverage add` / `init` / `update` | `Bash` / `PowerShell`, success or failure | Re-ingests a fresh full-suite coverage report in the background; a process start per shell command in that repo |-| **Command-rewrite (distill)** | Claude Code | `repowise hook rewrite install` (opt-in) | `Bash` / `PowerShell` | Rewrites noisy commands to `repowise distill <cmd>`; auto-allowed by default, set `permission: ask` to approve each one |-| **Codex context + staleness** | Codex | `repowise init --codex` | SessionStart / edit / shell | Reminds Codex to use the MCP tools and flags stale context after edits |+| **SessionStart context** | Claude Code | `repowise init` | session `startup` / `resume` / `clear` | Index-freshness line, core-tool pointer, and the standing decisions relevant to this session |+| **PostToolUse enrichment** | Claude Code | `repowise init` | `Grep` / `Glob` / `Read` / `Edit` / `Write` / repowise MCP calls | Graph context on searches, read notices, edit-time decision and bug-history notices |+| **Wrong-path rescue** | Claude Code | `repowise init` | a path tool call that failed on a path this tree does not have | Names the file when exactly one indexed file carries that basename; silent otherwise |+| **Coverage re-ingest** (opt-in) | Claude Code | `hooks.coverage_reingest: true`, then the next `repowise coverage add` / `init` / `update` | `Bash` / `PowerShell` | Re-ingests a fresh full-suite coverage report in the background |+| **Command rewrite (distill)** | Claude Code, Codex | `repowise hook rewrite install`, or Yes at the `repowise init` prompt | `Bash` / `PowerShell` | Rewrites noisy commands to `repowise distill <cmd>` |+| **Codex context and staleness** | Codex | `repowise init --codex` | SessionStart, edits, shell | Reminds Codex to use the MCP tools and flags stale context after git operations |-Every agent hook records what it said and whether the agent acted on it — see-[`repowise hook stats`](#hook-efficacy-repowise-hook-stats).+Each agent hook records what it said and whether the agent acted on it. See+[Hook efficacy](#hook-efficacy-repowise-hook-stats).---## Git hook: post-commit auto-sync-The wiki, graph, and health scores are only as current as your last index. The-post-commit hook closes that gap: after every commit it runs `repowise update`-in the background, so documentation, dependency edges, and code-health follow-your code without you thinking about it. Your terminal is never blocked.+The wiki, graph and health scores are only as current as your last index. After+every commit this hook runs `repowise update` in the background, so docs,+dependency edges and health follow your code. Your terminal is never blocked.```bash-repowise hook install # install for the current repo-repowise hook install --workspace # install for all repos in the workspace-repowise hook status # check whether the hook is installed-repowise hook uninstall # remove it+repowise hook install # current repo+repowise hook install --workspace # every repo in the workspace+repowise hook status+repowise hook uninstall```-The hook is **marker-delimited**, so it coexists safely with other tools' hooks-(linters, formatters, commit-msg checks) in the same `post-commit` file: repowise-only ever touches the block between its own markers. See-[AUTO_SYNC.md](../scale/AUTO_SYNC.md) for the full sync model, including how git worktrees-seed from the base checkout.+The hook is marker-delimited, so it coexists with other tools' hooks in the same+`post-commit` file: repowise only touches the block between its own markers. See+[Keeping the index fresh](../scale/AUTO_SYNC.md) for every sync method, and+[WORKTREES.md](../scale/WORKTREES.md) for how worktrees seed from the base checkout.-> Prefer to keep updates manual? Skip this hook and run `repowise update`-> yourself. The agent hooks below will remind you when the index falls behind.+Prefer manual updates? Skip this hook and run `repowise update` yourself. The+SessionStart hook tells the agent when the index falls behind.---## Claude Code agent hooks-Installed automatically during `repowise init` into your global-`~/.claude/settings.json`. Existing user hooks are always preserved, and legacy-repowise entries are migrated in place on the next `init` / `update`. All of them-route through the `repowise-augment` console script (a standalone entry point-that does not load the full CLI).--`repowise init --no-editor-setup` skips this whole group, along with the MCP-server registration that shares the same file. Reach for it when the repo is-temporary (a scratch clone, a worktree, a benchmark loop) and you do not want-that machine-wide config to move. `REPOWISE_SKIP_EDITOR_SETUP=1` is the same-switch for CI. The index itself is identical either way. To register the repo-afterwards, re-run `repowise init` in it without the flag.--### SessionStart, live freshness + relevant decisions--The generated `CLAUDE.md` is static between reindexes, so it can't say whether-the index is current *right now*. This hook adds a short per-session block so the-agent starts with calibrated trust instead of discovering staleness mid-task:--- **Index current** → one line saying so, plus the core-tool pointer.-- **Update running** → a positive "catching up" notice (never a stale scare).-- **Index behind** → indexed vs `HEAD` with a changed-file count, and the- target-scoped trust rule (a `stale_warning` fires only when a file a response- actually served has changed).--It also carries the **relevance-ranked standing decisions** for this session.-repowise scores the repo's active decisions against the session's likely working-set (dirty and staged files, files changed on the branch vs `main`, the previous-session's edited files, and branch-name tokens), expanded one hop through import-edges and co-change partners. The top few land under a hard ~400-token cap.-Relevance or silence: nothing clears the floor, nothing is injected, and-decisions are never shown just for being high-confidence. Repo-wide rules mined-from your own corrections are the one exception: a rule like "use the shared-logger, not print" applies everywhere rather than to specific files, so it-competes at a flat base relevance.--### PostToolUse, enrichment on every tool call--One hook covers several jobs, matched on-`Grep`, `Glob`, `Read`, `Edit`, `Write`, and repowise MCP calls:--**Grep/Glob enrichment.** When Claude Code runs a broad or zero-result search,-repowise appends focused context pulled straight from `wiki.db`:+`repowise init` installs these into your global `~/.claude/settings.json`.+Existing user hooks are preserved, and older repowise entries are updated in+place on the next `init` or `update`. They all run the `repowise-augment` entry+point, which does not load the full CLI.++`repowise init --no-editor-setup` skips this group, along with the MCP server+registration in the same file. Use it for a scratch clone, a worktree or a+benchmark loop where you do not want machine-wide config to change.+`REPOWISE_SKIP_EDITOR_SETUP=1` does the same in CI. The index is identical+either way; re-run `repowise init` without the flag to register the repo later.++### SessionStart: freshness and relevant decisions++The generated `CLAUDE.md` is static between reindexes, so it cannot say whether+the index is current right now. This hook adds a short block at session start:++- **Index current**: one line saying so, plus the core-tool pointer.+- **Update running**: a "catching up" notice.+- **Index behind**: indexed commit vs `HEAD` with a changed-file count, and the+ trust rule (a `stale_warning` fires only when a file a response served has+ changed).++It also carries the **standing decisions** most relevant to this session.+repowise scores active decisions against the likely working set (dirty and+staged files, files changed on the branch vs `main`, the previous session's+edited files, branch-name tokens), expanded one hop through imports and+co-change partners. The top few land under a hard 400-token cap. If nothing+clears the relevance floor, nothing is injected. Working agreements mined from+your own corrections (for example "use the shared logger, not print") name no+file, so they compete at a flat base relevance.++### PostToolUse: enrichment on tool calls++One hook, matched on `Grep`, `Glob`, `Read`, `Edit`, `Write` and repowise MCP+calls, does several jobs.++**Search enrichment.** On a broad or zero-result search, repowise appends+context from the index:| Field | What it tells the agent ||-------|------------------------|-| **Symbols** | Functions, classes, and methods defined in the file |-| **Imported by** | Which files depend on this file (reverse dependency) |-| **Depends on** | What this file imports (forward dependency) |-| **Git signals** | Hotspot status, bus factor, and owner |--So an agent that greps for `PageGenerator` immediately knows what depends on it,-what it depends on, and that it is a hotspot, without a separate MCP call:+| **Symbols** | Functions, classes and methods defined in the file |+| **Imported by** | Files that depend on this file |+| **Depends on** | What this file imports |+| **Git** | Hotspot status, bus factor and owner |```[repowise] 2 related file(s) found:- packages/core/.../page_generator.py- Symbols: function:_now_iso, class:PageGenerator, method:__init__- Imported by: init_cmd.py, update_cmd.py, generation/__init__.py- Depends on: context_assembler.py, base.py, models.py- Git: HOTSPOT, bus-factor=1, owner=RaghavChamadiya+ src/billing/invoice_service.py+ Symbols: class:InvoiceService, method:__init__, function:_now_iso+ Imported by: api/routes.py, jobs/reconcile.py+ Depends on: models.py, tax.py+ Git: HOTSPOT, bus-factor=1, owner=alice```-**Search-flood digests.** A grep that returns 50+ matches also gets a compact-per-file digest: every matched file with its match count and two anchor line-numbers, ranked by graph centrality when the index can rank them, and an explicit-`(N more files, M matches)` tail for anything past the top ten.--With `hooks.search_digest: true` in `.repowise/config.yaml`, written by the same-yes/no as the rewrite hook, and toggled afterwards with `repowise hook-search-digest install | uninstall | status`, that digest *replaces* the raw-match list rather than riding alongside it. Re-run the search scoped to a file it-names, or read those lines directly, to see any match in full. Savings appear in-`repowise saved` under the `search_digest` filter, and a repo with it off still-gets the counterfactual number.--Two cases are deliberately left alone. A **single-file context grep** (`-C`,-`-A`, `-B`) is never digested: that context is exactly what the agent asked for,-and Claude Code renders those results without a path prefix, so they are not-parsed as a multi-file flood in the first place. And `files_with_matches` results-carry no match text to replace: the file list is already a digest.--**Read-intelligence.** On `Read` of an indexed file, repowise emits a per-file-stale-read notice when the file changed after the session's previous read of it,-and points at the cheaper `get_context(..., include=["skeleton"])` for-structure-level questions.--With `hooks.read_skeleton: true` in `.repowise/config.yaml` — which `repowise-init` writes from the same yes/no as the rewrite hook, and which `repowise hook-read-skeleton install | uninstall | status` toggles afterwards — that pointer-becomes an action: an-unbounded `Read` of a large indexed file returns the file's *skeleton* instead of-the file, once per file per session. Signatures stay, keeping their real line-numbers; bodies collapse to `... N lines (a-b)` markers carrying 1-indexed ranges,-so the agent can range-read any elided span back — the same reversibility contract-`repowise distill` makes for shell output. Reading the file again with no range-returns it whole. Savings appear in `repowise saved` under the `read_skeleton`-filter. In a repo that has it off, the same Reads are still *measured*, and-`repowise saved` reports what they would have saved — a number about size only,-never about whether the agent could work from a skeleton.--One consequence is worth knowing: a Read the agent saw only as a skeleton still-satisfies Claude Code's read-before-edit precondition, so an `Edit` (especially-with `replace_all`) or a `Write` could touch bodies it never saw. Editing such a-file raises a one-line warning, once per file, until the file is read in full.--**Re-reads of unchanged files.** With `hooks.read_reread: true` — same consent,-toggled afterwards with `repowise hook read-reread install | uninstall |-status` — a `Read` of a file the session already read comes back as a short-notice naming that earlier read, instead of the content. The content is already-in context a few tool calls up, so sending it twice buys nothing.--The gate is arithmetic rather than a judgement: the same range must have been-served, no `Edit` or `Write` may have come between, and the bytes must hash the-same. When they do not, the agent gets the file *and* a line saying it changed-on disk without an edit in this session — a `git checkout`, a formatter, another-agent. That is worth more than the bytes were, because nothing else in the-session can discover it.--Two rules bound how wrong this can be. It is **never applied twice in a row for-the same file**, so if a context compaction dropped the earlier copy, one more-Read always returns the content; the notice says exactly that. And a Read that-any surface replaced records no content observation at all, so the agent is-never told it already has bytes that a skeleton stood in for.--That second rule is also what makes this a **narrow surface rather than a broad-one**, and it is worth knowing before you turn it on. The two Read surfaces are-substitutes, and the skeleton takes the valuable half first: on a large indexed-file the first Read is served as a skeleton and records no content observation,-so the second Read has nothing to compare against and is served whole. What-reaches the collapse is small files, unindexed files, and third-and-later reads-of the same range. Expect a modest, steady trim rather than a headline. Savings-appear under the `read_reread` filter, and a repo with it off still gets the-counterfactual, so you can see what it would have done before enabling it.--**Searches that time out.** On Windows a `Glob` can exhaust ripgrep's 20-second-budget and return nothing at all — not "no matches", nothing, after twenty-seconds of waiting. A glob is a path query and the index already holds every-path, so repowise answers it offline at the moment the failure happens, naming-the matching paths and counting any it did not list. The win here is wall clock-rather than tokens. Brace expansion (`{a,b}`) is declined rather than-half-matched, and zero indexed matches stays silent: the index having nothing to-say is not the tree having no such file.--**Edit-time "governed by" decisions.** When the agent edits a file governed by an-architectural decision (via `decision_node_links`), it gets a one-line notice-with the rationale, at most once per session per decision and only a few times-per session total. This is how a decision reaches the agent at the moment it is-about to violate (or honor) it.--Every injected decision id is recorded locally in-`.repowise/sessions/sessions.db`. On the next `repowise update`, the session miner-checks whether the guidance was followed or contradicted by your corrections in-that session, and relaxes or bumps the decision's staleness accordingly, so-guidance that stops being true stops being injected. This is the feedback loop-behind "learns from your sessions" (see the [README](../../README.md) and-[decisions layer](../layers/INTELLIGENCE_LAYERS.md)).-------## PostToolUseFailure, the wrong-path rescue--An agent that knows a file exists but guesses the wrong directory for it gets-back "Path does not exist" and burns a turn hunting. The index already knows-where that filename lives, so the failure is answerable at the moment it-happens:+**Search-flood digests.** A grep that returns 50 or more lines also gets a+per-file digest: each matched file with its match count and anchor line+numbers, ranked by graph centrality when the index can rank them, with a+`(N more files, M matches)` tail past the top ten. With `hooks.search_digest:+true`, the digest replaces the raw match list. Single-file context greps (`-C`,+`-A`, `-B`) and `files_with_matches` results are never digested.++**Read notices.** On a `Read` of an indexed file, repowise warns when the file+changed after the session's previous read of it, and points at+`get_context(..., include=["skeleton"])` for structure-level questions.++**Skeleton reads** (opt-in, `hooks.read_skeleton: true`). An unbounded `Read`+of a large indexed file returns the file's skeleton, once per file per session.+Signatures keep their real line numbers; bodies collapse to `... N lines (a-b)`+markers, so the agent can range-read any elided span. Reading the file again+with no range returns it whole. A skeleton read still satisfies Claude Code's+read-before-edit check, so an `Edit` or `Write` on such a file raises a one-line+warning, once per file, until the file is read in full.++**Re-read collapse** (opt-in, `hooks.read_reread: true`). A `Read` of a file+the session already read comes back as a short notice naming the earlier read.+It applies only when the same range was served, no `Edit` or `Write` came+between, and the bytes hash the same. When the bytes differ, the agent gets the+file plus a line saying it changed on disk outside this session. It is never+applied twice in a row for the same file, so one more Read always returns the+content after a context compaction. Because a skeleton read records no content,+this mostly trims small files, unindexed files and third reads of a range.++**Glob timeouts.** On Windows a `Glob` can exhaust its 20-second budget and+return nothing. A glob is a path query and the index holds every path, so+repowise answers it from the index at that moment. Brace expansion (`{a,b}`) is+declined, and zero indexed matches stays silent.++**Edit-time notices.** When the agent edits a file governed by an architectural+decision, it gets a one-line notice with the rationale, once per session per+decision and under a per-session cap. A file with a repeated bug-fix history+gets a one-line heads-up too (see [bug history](../layers/BUG_HISTORY.md)).++Every injected decision id is recorded in `.repowise/sessions/sessions.db`. On+the next `repowise update`, the session miner checks whether your corrections+followed or contradicted that guidance and adjusts the decision's staleness, so+guidance that stops being true stops being injected. See+[decisions](../layers/DECISIONS.md).++The three opt-in read and search keys are written by the rewrite-hook question+in `repowise init`, and toggled per repo with `repowise hook read-skeleton`,+`hook read-reread` and `hook search-digest` (`install | uninstall | status`).+Each one's savings appear in `repowise saved` under its own filter, and a repo+with the key off still gets the counterfactual number. Key details:+[CONFIG.md](../reference/CONFIG.md#the-hooks-block).++### PostToolUseFailure: wrong-path rescue++An agent that guesses the wrong directory for a file gets "Path does not exist"+and burns a turn hunting. The index knows where that filename lives:```-[repowise] core/git_indexer/fix_events.py is not in this tree.-The only indexed fix_events.py is core/ingestion/git_indexer/fix_events.py+[repowise] billing/invoice.py is not in this tree.+The only indexed invoice.py is src/billing/invoice.py```-It speaks only when the basename resolves to **exactly one** indexed file that-is still on disk. Everything else is silence, and each case is a distinct way-to be confidently wrong:--- **An ambiguous basename.** Naming one of a dozen `registry.py` is worse than- saying nothing, because the agent has no cheap way to tell a rescue from a- fact.-- **A directory target.** "Which file did you mean" is not the question a- missing directory asks.-- **A path in another checkout.** A sibling worktree has its own index; this- one has no standing to answer for it.-- **A failure Claude Code already answered.** It prints its own "Did you mean"- for some of these, and repeating it is worse than silence.-- **The path that just failed.** The index can hold a row for a file that is- not on disk right now, and pointing back at the failed path is the worst- thing this surface could say.--Most path-not-found failures therefore get silence, and that gap is the design-rather than a shortfall in it. `repowise hook stats` reports what this surface-actually did on your machine.+It speaks only when the basename resolves to exactly one indexed file that is+still on disk. It stays silent for an ambiguous basename, a directory target, a+path in another checkout, a failure Claude Code already answered with its own+"Did you mean", and the path that just failed. Most path-not-found failures+therefore get silence by design.----## Command-rewrite hook (distill), opt-in+## Command-rewrite hook (distill)-Most of what an agent reads from a shell command is noise: 300 lines of passing-tests around 4 failures, full commit bodies for "what changed recently". The-rewrite hook intercepts noisy `Bash` / `PowerShell` commands and rewrites them to-[`repowise distill <cmd>`](DISTILL.md), which compresses the output errors-first-before the agent reads it, exit code preserved and every omission reversible.+Most of what an agent reads from a shell command is noise. The rewrite hook+rewrites noisy `Bash` / `PowerShell` commands to+[`repowise distill <cmd>`](DISTILL.md), which compresses output errors-first+before the agent reads it, keeps the exit code, and makes every omission+reversible.```bashrepowise hook rewrite install # or answer Yes at the `repowise init` prompt@@ -270,152 +198,82 @@ repowise hook rewrite statusrepowise hook rewrite uninstall```-- Defaults to **`allow`**, so a rewrite runs without a prompt. That is not a+- Rewrites run without a prompt by default (`permission: allow`). This is not apermission escalation: a rewrite is always `repowise distill <one recognized- command>` from a closed family set, never an arbitrary command smuggled- behind the wrapper. Set `permission: ask` under `distill.commands` in- `.repowise/config.yaml` to approve each one instead.-- Never rewrites compound commands, redirections, or watch modes. The one pipe- shape it handles (macOS/Linux) is a single stage into `head`, `tail`, `grep`- or `rg`, quoted whole so it runs unchanged inside distill's own shell.-- Installing also adds `Bash(repowise distill:*)` / `PowerShell(repowise distill:*)`- to `permissions.allow`, so an already-approved command family doesn't start- re-prompting just because its string changed.--Per-repo behavior lives under `distill.commands` in `.repowise/config.yaml`-([CONFIG.md](../reference/CONFIG.md)). Track what it saved with `repowise saved`.+ command>`. Set `permission: ask` under `distill.commands` in+ `.repowise/config.yaml` to approve each one.+- Compound commands, redirections and watch modes are never rewritten. On+ macOS/Linux, a single pipe into `head`, `tail`, `grep` or `rg` is rewritten as+ one quoted command.+- With `permission: ask`, a rewritten string no longer matches allow rules such+ as `Bash(git diff:*)`. `repowise hook rewrite install --allow-rule` adds+ `Bash(repowise distill:*)` / `PowerShell(repowise distill:*)` to cover that.++Codex support, the safety model and the per-repo config are in+[DISTILL.md](DISTILL.md#3-the-command-rewrite-hook-claude-code--codex).---## Codex hooks-Written to project-local `.codex/hooks.json` by `repowise init --codex` (they do-not touch your global `~/.codex/config.toml`):+`repowise init --codex` writes project-local `.codex/hooks.json` (your global+`~/.codex/config.toml` is untouched):-- **SessionStart** → a short developer note reminding Codex to use the repowise- MCP tools for architecture, search, risk, decisions, and dead-code analysis,+- **SessionStart**: a short note reminding Codex to use the repowise MCP tools,plus the same relevance-ranked standing decisions Claude Code receives.-- **PostToolUse** (the shell tool, and `apply_patch` / `Edit` / `Write`) → after+- **PostToolUse** (the shell tool, and `apply_patch` / `Edit` / `Write`): aftera successful `git commit`, `merge`, `rebase`, `cherry-pick` or `pull`, compares- `HEAD` against the last indexed commit and flags that indexed context may be- stale, pointing at `repowise update`.--`UserPromptSubmit` is no longer registered here. It carried no matcher, so it-fired on every prompt, and what it returned was the-static MCP-usage note `SessionStart` had already delivered in the same session:-one process start per prompt to repeat a block the agent was holding.--An install that predates the retirement is repaired in place the next time this-repo's `.codex/hooks.json` is written, because the merge is additive and could-not otherwise retire anything. That means `repowise agents refresh`, or a-`repowise init` that selects Codex (the interactive checklist pre-ticks it when-the repo is already wired, and `--codex` selects it outright). A plain-`repowise init --yes` does not write Codex config at all, so it does not-migrate either. A hook *you* wrote on that event is left alone, and a timeout-you raised yourself is never moved.+ `HEAD` with the last indexed commit and flags that context may be stale.Codex names its shell tool `shell_command` on current releases and `Bash` on-older ones, so the matcher covers both. The shared Claude Code hook deliberately-does *not* watch shell commands: it has `Read` / `Grep` / `Glob` tools and a-SessionStart freshness line, so the shell adds cost without adding reach. Codex-has neither, which is why it keeps the surface. The opt-in exceptions are the-decision capture prompt and the coverage re-ingest (see-[What gets written where](#what-gets-written-where)).--Full Codex setup: [CODEX.md](CODEX.md).+older ones, so the matcher covers both. Codex has no `Read` / `Grep` / `Glob`+tools for repowise to enrich, which is why it watches the shell and Claude Code+does not. Full setup: [CODEX.md](CODEX.md).---## Hook efficacy: `repowise hook stats`The agent hooks keep a local ledger in `.repowise/sessions/sessions.db`: what-each hook said, and whether the agent went on to do what it pointed at.+each hook said, and whether the agent then did what it pointed at. The verdict+comes from your own Claude Code transcripts, so the numbers are yours. Nothing+leaves the machine.-```sh+```bashrepowise hook stats # per-surface firing counts and action rates-repowise hook backfill --all-projects # seed it from your existing transcripts+repowise hook backfill --all-projects # seed the ledger from existing transcripts```-The verdict comes from your own Claude Code transcripts — a firing is paired-with the tool calls that followed it — so the numbers are yours, not a-benchmark. `repowise update` classifies recent sessions; `hook backfill` covers-history. Nothing leaves the machine.--Notices that ask for nothing (the stale-read warning, the silent-read-after-served measurement) report `n/a` rather than a rate. `hook stats`-also reports hook invocation counts and wall time, including the calls that-returned silence.--> Upgrading from a release before firings were keyed by their text: run-> `repowise hook backfill --reset` once, or older rows are counted separately-> from the replayed ones. It never touches decisions.+Flags and the `--reset` upgrade step: [CLI reference](../reference/CLI_REFERENCE.md#repowise-hook-stats).---## What gets written where-`repowise init` writes these entries into `~/.claude/settings.json` (Claude Code)-and `.codex/hooks.json` (Codex when `--codex` is passed). The opt-in coverage-re-ingest entries go to the repository's own `.claude/settings.local.json`-instead: with `hooks.coverage_reingest: true`, the next `repowise coverage add`,-`init` or `update` that finds stored coverage writes them when the Claude Code-hooks above are installed, and removes them once the key is false or gone-(`repowise uninstall` removes them too). The plugin does not carry them.--| Client | Hook type | Matcher | Command |-|--------|-----------|---------|---------|-| Claude Code | `SessionStart` | `startup\|resume\|clear` | `repowise-augment` [^guard] |-| Claude Code | `PostToolUse` | `Grep\|Glob\|Read\|Edit\|Write\|mcp__.*[Rr]epowise.*__.*` | `repowise-augment` [^guard] |-| Claude Code | `PreToolUse` (opt-in) | `Bash\|PowerShell` | `repowise-rewrite` |-| Claude Code, this repo only | `PostToolUse` and `PostToolUseFailure` | `Bash\|PowerShell` | `repowise-augment --coverage-only` [^guard] |-| Codex | `SessionStart` | `startup\|resume\|clear` | context reminder |-| Codex | `PostToolUse` | `Bash\|shell_command`, `apply_patch\|Edit\|Write` | staleness check |--[^guard]: The command is written wrapped in a presence check rather than as the- bare name:-- ```sh- if command -v repowise-augment >/dev/null 2>&1; then exec repowise-augment; fi- ```-- The Claude Code plugin ships these hooks independently of the CLI, so- "plugin installed, `repowise` not installed" is a supported state — and a- partially written install reaches it too (on Windows an MCP server holds- `repowise.exe` open, so an installer can abort after writing only some- console scripts). Unguarded, either state prints `command not found` on- every matched tool call: non-blocking, unactionable, and endless. The guard- is POSIX (`command -v` + `exec`, verified under `sh`, `bash` and `dash`) and- forwards stdin unchanged, so the hook behaves identically when the script is- present. An older install carrying the bare name is rewritten on the next- `repowise init`. Codex hooks keep the bare name: their execution model is- not documented as shell-based, and a directly `exec`'d guard would try to- run a binary named `if`.--`SessionStart` deliberately excludes `compact`: the block usually survives-compaction in the summary, and re-emitting it there would double it up. `init`-also sets `env.ENABLE_TOOL_SEARCH=true` so the MCP tool schemas load on demand-rather than sitting in every session's standing context (an existing value you-set, including a deliberate `false`, is left untouched).--For manual debugging, the underlying entry points can be run directly:--```bash-repowise-augment # invoked by the agent hooks; prints what it would inject-repowise augment # equivalent Click subcommand-```+| Client | Hook type | Matcher | Written to |+|--------|-----------|---------|------------|+| Claude Code | `SessionStart` | `startup\|resume\|clear` | `~/.claude/settings.json` |+| Claude Code | `PostToolUse` | `Grep\|Glob\|Read\|Edit\|Write\|mcp__.*[Rr]epowise.*__.*` | `~/.claude/settings.json` |+| Claude Code | `PostToolUseFailure` | `Read\|Edit\|Write\|Grep\|Glob\|NotebookEdit` | `~/.claude/settings.json` |+| Claude Code | `PreToolUse` (opt-in rewrite) | `Bash\|PowerShell` | `~/.claude/settings.json` |+| Claude Code | `PostToolUse` (opt-in decision capture prompt) | `Bash\|PowerShell` | `~/.claude/settings.json` |+| Claude Code | `PostToolUse` and `PostToolUseFailure` (opt-in coverage re-ingest) | `Bash\|PowerShell` | the repo's `.claude/settings.local.json` |+| Codex | `SessionStart` | `startup\|resume\|clear` | the repo's `.codex/hooks.json` |+| Codex | `PostToolUse` | `Bash\|shell_command`, `apply_patch\|Edit\|Write` | the repo's `.codex/hooks.json` |++The coverage re-ingest entries are written by the next `repowise coverage add`,+`init` or `update` once `hooks.coverage_reingest: true` is set, and removed when+the key is false or gone. Command strings, the presence guard and the other+settings `init` writes: [hooks reference](../reference/HOOKS_REFERENCE.md).---## Hooks vs MCP tools-The two are complementary:--- **Hooks** are passive, automatic, and cost the agent nothing. They fire on- every search, edit, or session start whether or not the agent is thinking about- graph context.-- **[MCP tools](MCP_TOOLS.md)** are active and on-demand, with richer output.- Reach for them when the agent needs full documentation, a risk assessment,- decision history, or dependency tracing.+- **Hooks** are passive. They fire on every search, edit or session start,+ whether or not the agent is thinking about graph context.+- **[MCP tools](MCP_TOOLS.md)** are on-demand and return richer output: full+ documentation, risk, decision history, dependency paths.-For most day-to-day coding, the hooks supply enough context on their own; the MCP-tools are there for deeper investigation.+For day-to-day coding the hooks supply most of the context; the MCP tools are+there for deeper questions.diff --git a/docs/agent/MCP_TOOLS.md b/docs/agent/MCP_TOOLS.mdindex ed189d307..b37cd53ef 100644--- a/docs/agent/MCP_TOOLS.md+++ b/docs/agent/MCP_TOOLS.md@@ -1,1531 +1,521 @@# MCP Tools Reference-repowise exposes a curated set of tools via the [Model Context Protocol](https://modelcontextprotocol.io) (MCP). These tools give AI coding assistants (Claude Code, Codex, Cursor, Cline, Windsurf) structured access to your codebase intelligence: dependency graph, git history, documentation, and architectural decisions.+repowise serves its codebase intelligence to AI coding assistants (Claude Code, Codex, Cursor, Cline, Windsurf and any other [Model Context Protocol](https://modelcontextprotocol.io) client) as a set of MCP tools. The tools answer questions from the index: the dependency graph, git history, generated docs, decision records, health and dead-code analysis. None of them edits your code. Two opt-in tools go further: `set_finding_status` records a triage verdict in the index, and `generate_refactoring_code` calls your configured LLM to draft a diff.-18 tools are registered in total. A single-repo server advertises 10 by default: exactly the canonical tools. Workspace mode adds the `list_repos` discovery utility, for 11. 7 specialist tools are opt-in where eligible. The surface is configurable; see [Configuring the tool surface](#configuring-the-tool-surface).+18 tools are registered in total. A single-repo server advertises 10 by default: exactly the canonical tools. Workspace mode adds the `list_repos` discovery utility, for 11. 7 specialist tools are opt-in where eligible.-**Start the MCP server:**--```bash-repowise mcp --transport stdio # for Claude Code, Codex, Cursor, etc.-repowise mcp --transport streamable-http # for HTTP clients on port 7338-repowise mcp --transport sse --port 7338 # legacy SSE transport-```--**Auto-setup:** `repowise init` automatically registers the MCP server and installs proactive hooks for Claude Code. `repowise init --codex` writes project-local Codex MCP config and hooks.--**Opting out:** each Claude config holds a single `repowise` MCP key, so indexing a second repo repoints it rather than adding a second entry. Pass `repowise init --no-editor-setup` (or set `REPOWISE_SKIP_EDITOR_SETUP=1`) for a repo you do not want registered: a scratch clone, a worktree, a CI or benchmark run. Nothing about the index changes, and re-running `repowise init` without the flag registers it later. `init` also prints a notice when it is about to repoint an existing entry.-----+This page tells you which tool to call and with what arguments. Every response field, the full `_meta` envelope and the truncation rules are in [MCP_RESPONSE_FIELDS.md](../reference/MCP_RESPONSE_FIELDS.md).## Contents+- [Starting the server](#starting-the-server)+- [Configuring the tool surface](#configuring-the-tool-surface)+- [Reading a response](#reading-a-response)+- [Which tool for which question](#which-tool-for-which-question)+- [Workspace mode](#workspace-mode)+**Canonical tools (default in both modes, 10)**-[get_overview](#get_overview) ·-[get_answer](#get_answer) ·-[get_context](#get_context) ·-[get_symbol](#get_symbol) ·-[search_codebase](#search_codebase) ·-[get_risk](#get_risk) ·-[get_change_risk](#get_change_risk) ·-[get_why](#get_why) ·-[get_dead_code](#get_dead_code) ·-[get_health](#get_health)+[get_answer](#get_answer) · [get_context](#get_context) · [get_symbol](#get_symbol) · [search_codebase](#search_codebase) · [get_risk](#get_risk) · [get_change_risk](#get_change_risk) · [get_why](#get_why) · [get_overview](#get_overview) · [get_health](#get_health) · [get_dead_code](#get_dead_code)**Workspace discovery utility (default in workspace mode, 1)**[list_repos](#list_repos)**Opt-in specialists (7; workspace eligibility still applies)**-[get_architecture](#get_architecture) ·-[get_blast_radius](#get_blast_radius) ·-[get_dependency_path](#get_dependency_path) ·-[get_execution_flows](#get_execution_flows) ·-[generate_refactoring_code](#generate_refactoring_code) ·-[get_conformance](#get_conformance) ·-[set_finding_status](#set_finding_status)+[get_dependency_path](#get_dependency_path) · [get_execution_flows](#get_execution_flows) · [generate_refactoring_code](#generate_refactoring_code) · [set_finding_status](#set_finding_status) · [get_blast_radius](#get_blast_radius) · [get_architecture](#get_architecture) · [get_conformance](#get_conformance)-Also see [Configuring the tool surface](#configuring-the-tool-surface), [Reversible truncation](#reversible-truncation-_metaomitted) and [Unrecognised arguments](#unrecognised-arguments-ignored_arguments).----## The ten flagship tools+## Starting the server-| Tool | Purpose | Typical use |-|------|---------|-------------|-| `get_overview` | Architecture summary | First call on any unfamiliar codebase |-| `get_answer` | One-call RAG Q&A | First call on any code question |-| `get_context` | Rich context for targets | Before reading or modifying code |-| `get_symbol` | Raw source bytes for one symbol | When you need one function/class body |-| `search_codebase` | Hybrid symbol / path / concept search | Finding a symbol or file, or discovering code by topic |-| `get_risk` | Modification risk | Before changing hotspot files |-| `get_change_risk` | What a commit or range newly made worse | Before merging a commit or PR range |-| `get_why` | Architectural decisions | Before structural changes |-| `get_dead_code` | Unreachable code | Cleanup tasks |-| `get_health` | Code-health marker scores | Before refactoring, find the worst files |+```bash+repowise mcp # stdio, for Claude Code, Codex, Cursor+repowise mcp --transport streamable-http # HTTP on port 7338+repowise mcp --transport sse --port 7338 # legacy SSE transport+```-In workspace mode, `list_repos` is also on by default so repository aliases are discoverable. It is unavailable in single-repo mode because the server is already bound to the only repository. See [Supplementary tools](#supplementary-tools).+`repowise init` registers the server with Claude Code and installs its hooks; `repowise init --codex` writes project-local Codex MCP config and hooks. Pass `--no-editor-setup` (or set `REPOWISE_SKIP_EDITOR_SETUP=1`) for a scratch clone, worktree or CI run you do not want registered. Each Claude config holds a single `repowise` entry, so indexing a second repo repoints it; `init` prints a notice before it does.---## Configuring the tool surface-The default surface is deliberately small: fewer, richer tools mean fewer round-trips and less schema overhead per task. What a server advertises is resolved from three things: each tool's `default`/`requires_workspace` metadata, whether the server is in workspace mode, and an optional override.+| Tool | Default | Mode |+|------|---------|------|+| `get_answer`, `get_context`, `get_symbol`, `search_codebase`, `get_risk`, `get_change_risk`, `get_why`, `get_overview`, `get_health`, `get_dead_code` | on | single-repo and workspace |+| `list_repos` | on | workspace only |+| `get_dependency_path`, `get_execution_flows`, `generate_refactoring_code`, `set_finding_status` | opt-in | single-repo and workspace |+| `get_blast_radius`, `get_architecture`, `get_conformance` | opt-in | workspace only |-- **Default (single-repo):** 10 tools, exactly the canonical intelligence set.-- **Default (workspace):** those 10 plus `list_repos`, the workspace discovery utility.-- **Opt-in tools:** `get_dependency_path`, `get_execution_flows`, `generate_refactoring_code`, and `set_finding_status` are eligible in either mode. `get_architecture`, `get_blast_radius`, and `get_conformance` are workspace-only. All seven are off by default.--**Configure it in `.repowise/config.yaml`** under an `mcp.tools` key. Four shapes are supported:+Change the surface in `.repowise/config.yaml` under `mcp.tools`:```yaml-# Adjust the default set with + / - deltas (the common case):-mcp:- tools: ["+get_execution_flows", "-get_dead_code"]--# Or give an explicit allowlist (only these tools):-mcp:- tools: ["get_answer", "get_context", "get_symbol", "search_codebase"]--# Or enable everything available in the current mode:-mcp:- tools: all--# Or select the agent-lean profile (see below):mcp:- tools: lean+ tools: ["+get_execution_flows", "-get_dead_code"] # adjust the default set+# tools: ["get_answer", "get_context"] # explicit allowlist+# tools: all # every tool usable in this mode+# tools: lean # the agent-lean profile```-**Or per launch on the CLI**, which overrides the config block:+Or per launch, which overrides the config block:```bash-repowise mcp --tools "+get_execution_flows" # default set plus one-repowise mcp --tools "get_answer,get_context" # explicit allowlist-repowise mcp --tools lean # agent-lean profile-repowise mcp --all # every available tool+repowise mcp --tools "+get_execution_flows" # default set plus one+repowise mcp --tools "get_answer,get_context" # explicit allowlist+repowise mcp --tools lean # agent-lean profile+repowise mcp --all # every tool usable in this mode```-Workspace-only tools named explicitly in single-repo mode are ignored (they cannot do useful work there). Unknown tool names are ignored with a warning.+The dashboard Settings page has a per-repo toggle for each tool and writes the same `mcp.tools` block.-A tool registered with an `available_when` predicate over repository facts is hidden while that predicate is false, even when an override names it. Raw MCP clients get the list resolved when the server starts, while the in-product chat re-evaluates it on every turn.+- Tokens prefixed `+` or `-` adjust the default set. A list without prefixes is an allowlist.+- Unknown names are ignored with a warning. A workspace-only tool named outside workspace mode is ignored too: it has no workspace graph to read.+- Raw MCP clients see the surface resolved when the server started. Restart the server after changing it.-**The `lean` profile** is the agent-lean surface: `get_answer`, `get_context`, `get_symbol`, `search_codebase`, `get_risk`, and `get_why`, plus `list_repos` in workspace mode (where repo aliases must be discoverable). `get_why` is part of the lean set because why/history questions are the category no code-search surface can answer from the tree alone; a lean profile without it measurably underperforms on exactly those questions. The profile advertises ~2.1k tokens of schema versus ~4.1k for the default surface. That is small enough to keep always loaded, so when a repo has `mcp.tools: lean` configured, `repowise init` skips the tool-search recommendation (the `ENABLE_TOOL_SEARCH` setting that defers MCP schemas behind a lookup round trip) for Claude Code; the six schemas the agent actually reaches for stay in context on every turn. init never turns an existing `ENABLE_TOOL_SEARCH` setting off, since it applies to every MCP server, not just repowise.--**Or from the dashboard:** the Settings page lists every tool with its description and a per-repo toggle, and writes the same `mcp.tools` config for you.-----+- **Default (single-repo):** 10 tools, exactly the canonical intelligence set.+- **Default (workspace):** those 10 plus `list_repos`, the workspace discovery utility.-## Reversible truncation: `_meta.omitted`--Tool responses are token-budgeted. When a response is truncated, the dropped-content is no longer silently lost: it is stored in the repo's-[omission store](DISTILL.md#the-omission-store) and the response's `_meta`-envelope lists how to get it back:--```jsonc-"_meta": {- "omitted": {- "refs": ["a1b2c3d4e5f6"],- "tokens": 5840,- "restore": "repowise expand <ref> (CLI) or get_symbol(\"repowise#<ref>\", query?) (MCP)"- }-}-```+**The `lean` profile** is `get_answer`, `get_context`, `get_symbol`, `search_codebase`, `get_risk` and `get_why`, plus `list_repos` in workspace mode. It is small enough to keep every schema loaded, so when a repo sets `mcp.tools: lean`, `repowise init` skips the Claude Code tool-search recommendation.-Truncated skeleton blocks are replaced in place by a `[repowise#<ref>: ...]`-marker; everything else is captured into one combined document per response.-A response that would still oversize sheds whole blocks, in an order each tool-declares cheapest-loss-first, and reports `truncated: true` alongside the refs.--Every tool is budgeted. Most declare their own shed order; the rest meet a-final size guard that trims the largest blocks and records what it took. Either-way no response is returned unbounded and unflagged, and-`_meta.response_budget` reports the ceiling that applied and the size delivered-under it.-Resolve refs with `repowise expand <ref>` from a shell, or-`get_symbol("repowise#<ref>")` from any MCP client. See-[DISTILL.md](DISTILL.md) for the full reversibility model.--**The `_meta` envelope** (all fields optional, present only when meaningful):--| Field | When present |-|-------|--------------|-| `timing_ms` | Tool wall-time |-| `hint` | A short, conservative follow-up suggestion. On a `get_answer` reply that graded `low` while the index is behind live HEAD, it says to run `repowise update` and ask again before trusting the answer |-| `cached` | Only when `true` |-| `index_age_days` | Days since the last `repowise update` |-| `indexed_commit` | Short (12-char) SHA the index was built against |-| `live_head` | Short (12-char) SHA of the current checkout, whenever `.git/HEAD` is readable. Equal to `indexed_commit` when the index is current |-| `stale_warning` | Only on a real signal: HEAD mismatch **that actually changed files**, or age over ~90 days when git is unreachable. Two commits with identical trees (an empty commit, a no-op merge) report `index_behind` with no warning |-| `index_behind` | Whenever the live-vs-indexed comparison ran: `true` if HEAD has moved (alongside `stale_warning` when served content actually changed), `false` if the commits match. Absent means the comparison could not run (no git, or a repo-level tool that serves no file content) |-| `embedder_degraded` | Whenever an embedder is resolved, `true` or `false`. Absent means none was initialised |-| `embedder`, `embedder_warning` | Only when the embedder fell back to a mock/degraded mode |-| `newer_release` | Once per server process, in the first response after the background check sees a newer repowise on PyPI: the available version, the running one, and that the MCP server needs a restart after upgrading. Repeats only for a later, newer version |-| `response_budget` | Always: `limit_chars` (the ceiling that applied), `tier` (`default` or `expanded`, chosen by whether the call passed an expansion argument), `serialized_chars` (the size delivered) |-| `scope_hint` | `get_context` and `get_answer`, when knowledge-graph layers exist that contain none of the served paths: one sentence naming up to three of them with file counts, so an agent knows which areas the answer did not touch |-| `complete` | When the response served whole units: how many symbol bodies (bounds verified against the live file) or whole files, and that they need not be re-opened. Sliced bodies and partial ranges are never counted |-| `completeness` | Always: `capped`, plus `shown` and `total` summed over every collection a reducing pass counted, and `reason` naming the pass that dropped the most. The sum spans unlike collections and sits inside whatever `limit` the call already applied, so it measures what survived reduction, not what share of the repository you hold. Absent `shown`/`total` means no pass counted rows |-| `floor` | Only when the response carries a count derived by walking the indexed graph: names those fields, whose values are lower bounds because an edge the index failed to resolve is uncounted rather than proven absent |-| `state` | Only when something fired: `degraded` plus `degraded_reasons` mapping each contributing key to its reason (a synthesis reason string, the retrieval legs that broke), `partial`, `truncated`. A coarse roll-up of the response's own flags |--Silence on `stale_warning` means the index is current; don't infer staleness from its absence. `list_repos`, `get_architecture`, `get_blast_radius`, `get_conformance`, and `get_dependency_path` don't carry a freshness envelope at all. Neither does `search_codebase` when a workspace call merges results from several repos, since there is no single indexed commit to compare.+See [CONFIG.md](../reference/CONFIG.md#the-mcp-block) for the config block itself.----## Unrecognised arguments: `ignored_arguments`+## Reading a response-A tool never answers a bad argument with a filter that matches nothing. A value-outside a closed vocabulary is **dropped, not applied** — so the response is the-one you would have got without it — and the tool names what it dropped, at the-top level:+Every tool returns a JSON object with a `_meta` envelope. Most fields appear only when they carry a signal. An agent should check these:-```jsonc-"ignored_arguments": [- { "argument": "kind",- "values": ["unused_exports"],- "valid": ["unreachable_file", "unused_export", "unused_internal", "zombie_package"] }-]-```+| Field | What to do with it |+|-------|--------------------|+| `stale_warning` | Present only when the index is behind in a way that changed served files. Run `repowise update` before trusting file-level detail. Its absence means current. |+| `indexed_commit`, `live_head`, `index_behind` | Which commit the answer describes and whether HEAD has moved since. |+| `complete` | Symbol bodies or whole files served verified against the live file. Do not re-read them. |+| `state` | `degraded`, `partial` or `truncated` when something fired, with reasons. A degraded empty result is a failed read, not an empty repository. |+| `omitted` | Refs to content cut for size. Restore with `repowise expand <ref>` or `get_symbol("repowise#<ref>")`. |+| `response_budget` | The character ceiling that applied and the size delivered. |-The key is absent when every argument was understood, so its presence is the-whole signal. One entry per argument, however many of its values missed.+`get_answer` adds a top-level `confidence` (rates the prose) and `retrieval_quality` (rates the evidence). Full envelope: [MCP_RESPONSE_FIELDS.md](../reference/MCP_RESPONSE_FIELDS.md#the-_meta-envelope).-This exists because the alternative is a lie: `get_dead_code(kind="unused_exports")`-used to filter on the plural, match nothing, and recommend *"No dead code found-matching your filters."* beside a summary counting hundreds of unused exports-([#1496](https://github.com/repowise-dev/repowise/issues/1496)). It covers-`get_dead_code` (`kind`, `tier`, `min_confidence`), `get_context` (`include`)-and `search_codebase` (`kind`, `mode`; an unrecognised `mode` is routed as-`auto`).+### Reversible truncation: `_meta.omitted`-`get_dead_code`'s `min_confidence` additionally accepts the tier names the-response is organised by (`"high"` 0.7, `"medium"` 0.4, `"low"` 0.0) as-well as a float.+Responses fit 24,000 serialized characters by default and 32,000 when the call passes an expansion argument (a nonempty `include`, for most tools). Content cut to fit is stored in the repo's [omission store](DISTILL.md#the-omission-store), and `_meta.omitted` lists the refs that restore it. Capped lists carry `*_total` siblings, so a count is never lost with its rows. Details: [MCP_RESPONSE_FIELDS.md](../reference/MCP_RESPONSE_FIELDS.md#truncation-and-recovery).-`get_health` is the exception to the shape. It reports a misspelled `only` key-under its own older name, `unknown_only_keys`, and everything else —-`refactoring_*` and `performance_*` filters, `scope`, `counts` — in-`ignored_arguments` as a flat map of argument to the value that was dropped:+### Unrecognised arguments: `ignored_arguments`-```jsonc-"ignored_arguments": { "counts": "code-shape", "refactoring_effort": "tiny" }-```--A detail lookup (`finding_id`, `plan_id`, `opportunity_id`) reports `scope` and-`counts` there too: it answers about one stored row, so a population control has-nothing to act on.+A value outside a closed vocabulary (a misspelled `kind`, `include` key or `mode`) is dropped, never applied as a filter that matches nothing. The response then names it in `ignored_arguments`, with the valid values. The key is absent when every argument was understood. Per-tool shapes: [MCP_RESPONSE_FIELDS.md](../reference/MCP_RESPONSE_FIELDS.md#ignored-arguments).----## `get_overview`--Architecture summary, module map, and entry points.--**Parameters:**--| Parameter | Type | Required | Description |-|-----------|------|----------|-------------|-| `repo` | string | No | *(workspace only)* Target repo alias, or `"all"` |-| `include` | list[string] | No | Opt-in blocks, any combination of `"content"`, `"outline"`, `"tour"`, `"decisions"`, `"graph"`, `"ownership"` (see below) |--**Returns (default):** `title`, `content_md` (the overview essay's summary section), `key_modules` (name, path, outline section), `entry_points`, `architecture` (layer names and file counts in stack order, the test layer left out; `layer_order` only for an index without knowledge-graph layers), `code_health`, `git_health`, `next_actions` (per horizon, week and quarter: the first three stored actions as `id`, `tier`, `title`, `impact`, `done_when`, `target`, with `total` and `by_tier`; `unavailable` names stores the index predates; replaced by `next_actions_reason` when they cannot be read), `_meta`, and in workspace mode a `workspace` footer. The response's `more` field names the opt-in blocks.--**Opt-in blocks** — omitted unless named in `include`, and not computed at all when they are not:--| `include` key | Adds |-|---|---|-| `"content"` | the full overview essay in `content_md` |-| `"outline"` | `outline` — the stored wiki page tree, two rungs deep. `key_modules[].section` indexes into it |-| `"tour"` | `guided_tour` and `reading_order` — onboarding walks |-| `"decisions"` | `key_decisions`. `get_why` is the richer route |-| `"graph"` | `community_summary` — code-community clusters |-| `"ownership"` | `knowledge_map` — top 3 owners by files owned |--**When to use:** First call on any unfamiliar codebase. Gives the agent a mental map before diving into specifics. Skip on later calls in the same session; it doesn't change mid-session.--**Example calls:**--```-get_overview()-get_overview(include=["outline", "content"])-```--> **Output-schema change.** `guided_tour`, `reading_order`, `key_decisions`,-> `community_summary` and `knowledge_map` moved behind `include`; `outline` did-> too (`include=["outline"]` previously only deepened an always-present tree).-> `key_modules` no longer carries `description`, `page_id` or `parent_page_id`,-> and `architecture.layers[].description` is gone.+## Which tool for which question++| Question | Tool |+|----------|------|+| How does X work? Where is Y? Why is Z like this? | `get_answer` |+| I am new to this repo. What is the shape of it? | `get_overview` |+| What is in this file, who calls this function, who owns it? | `get_context` |+| Show me this file's structure without reading all of it | `get_context(include=["skeleton"])` |+| Give me the body of this one symbol | `get_symbol` |+| Find the symbol, file or page matching this name or topic | `search_codebase` |+| Is this file risky to touch? What has broken here before? | `get_risk` |+| What could my set of changed files break, and which tests should run? | `get_risk(changed_files=[...])` |+| What did this commit, range or uncommitted diff make worse? | `get_change_risk` |+| Why was it built this way? Is there a decision governing it? | `get_why` |+| What should we fix first? Which files are least healthy? | `get_health` |+| Is the file I just edited healthier or worse? | `get_health(targets=[...])` |+| What can we delete? | `get_dead_code` |+| Which repos does this server serve? | `list_repos` (workspace) |+| How are these two files connected? | `get_dependency_path` (opt-in) |+| What runs when this entry point is called? | `get_execution_flows` (opt-in) |+| Turn this refactoring plan into a diff | `generate_refactoring_code` (opt-in) |+| Record that a refactoring plan is a false positive | `set_finding_status` (opt-in) |+| Which services in other repos depend on this one? | `get_blast_radius` (workspace, opt-in) |+| How coupled is the whole system? | `get_architecture` (workspace, opt-in) |+| Does the system obey our declared dependency rules? | `get_conformance` (workspace, opt-in) |----## `get_answer`--One-call RAG: retrieves over the wiki, gates synthesis on confidence, and returns a cited 2-5 sentence answer.--**Parameters:**--| Parameter | Type | Required | Description |-|-----------|------|----------|-------------|-| `question` | string | Yes | Natural language question about the codebase |-| `scope` | string | No | Repository-relative path prefix to restrict retrieval to |-| `include` | list[string] | No | `["evidence"]` returns the expanded projection: no confidence-keyed trimming, and a larger response budget |-| `repo` | string | No | *(workspace only)* Target repo alias |--**Returns:** A synthesized answer with file/symbol citations, a `confidence`-label (`high`, `medium`, `low`) rating the prose, and a `retrieval_quality`-label (`high`, `partial`, `weak`) rating the evidence under it. Every grade-returns an answer; what changes is how much evidence rides along with it.-A `high` answer can be cited directly and sheds `retrieval`, `best_guesses`,-`candidates` and `fallback_targets`, keeping one quote and, when the answer is-not already grounded, one symbol body. A `medium` answer keeps one body, two-quotes and the top two evidence rows. A `low` answer keeps two bodies and the-top three evidence rows, and `fallback_targets` appears only when nothing else-was served. Read the rows the reply names rather than calling `search_codebase`-again. `include=["evidence"]` skips this trimming entirely.--When synthesis cannot run at all, no provider resolvable or the call failed, the-response carries a top-level `degraded` naming the reason, is built from-retrieval and mined rationale with no LLM involved, and keeps the fullest-evidence shape whatever it graded. It also raises `_meta.state.degraded`.-`confidence` there is graded from the retrieval actually served, not from the-missing prose: on `no-llm-provider` it is `medium` unless `retrieval_quality` is-`weak`, in which case `low`. Any other reason, a configured provider whose call-failed, stays `low`, because a retry can still produce a real answer. `high` is-unreachable on this path, since the `answer` string is assembled boilerplate.--`_meta.complete` names the symbol bodies served whole from live source, with-bounds checked against the file; do not re-open those. `_meta.scope_hint` names-up to three knowledge-graph layers holding none of the served paths, so an agent-knows which areas the answer did not touch. When an answer grades `low` and the-index is behind live HEAD, `_meta.hint` says to run `repowise update` and ask-again before trusting it.+## Default tools-Two path-bearing blocks, with different jobs:+### `get_answer`-| Field | Job | Confidence-gated? |-|-------|-----|-------------------|-| `retrieval` | **Evidence.** Enriched hits (summary, snippet, key symbols) to re-read when the prose needs checking. Shrinks as confidence rises, because a trustworthy answer needs less of it. | Yes |-| `candidates` | **Navigation.** The ranked shortlist of files retrieval resolved, one `{path, lines?}` entry each, up to 5. | Shape-gated: the default projection drops it at every confidence |+Answers a how, where or why question in one call: it runs hybrid retrieval over the wiki and the symbol index, then synthesises a cited answer. Use it first for any question about the code. Skip it when you already know the exact file or symbol; go to `get_context` or `get_symbol`.-`candidates` is built whenever retrieval resolved anything, including on high-confidence answers where `retrieval` is deliberately empty, but the default projection drops it; ask for it with `include=["evidence"]`. It is where to look next; it is not evidence that the answer is right.+| Parameter | Type | Default | Meaning |+|-----------|------|---------|---------|+| `question` | string | required | The question, in plain language |+| `scope` | string | none | Path prefix to restrict retrieval to, e.g. `"src/auth/"` |+| `repo` | string | default repo | Workspace repo alias. `"all"` is not supported |+| `include` | list[string] | none | `["evidence"]` returns the full evidence projection with a larger budget |-**Retrieval legs:** three, fused by Reciprocal Rank Fusion: full-text and vector search over wiki pages, plus the structural symbol index. The symbol leg is keyed on the content words of the question rather than on whether it happens to carry an identifier-shaped token, so "how does an incremental update persist symbols" reaches the same rows as `_persist_symbols`. It exists because a generated file page renders only the *public* symbol table: a private helper or a local name is not in the text the other two legs index.+**Key return fields:** `answer`, `confidence` (`high` / `medium` / `low`, rates the prose), `retrieval_quality` (`high` / `partial` / `weak`, rates the evidence), `citations`, `symbol_bodies` (live bodies of symbols the answer names), `retrieval`, `best_guesses` and `fallback_targets` (on low confidence), `episodes` (dated facts bearing on the question), `degraded` (synthesis could not run), `_meta.scope_hint` (areas the answer did not touch).-**When to use:** First call on any code question. Collapses search, read, and reason into one round-trip. On a low grade, start from the evidence rows the reply already carries; reach for `search_codebase` only when `retrieval_quality` is `weak`.--**Example call:**+A `high` answer can be cited directly. On `low`, read the rows the reply names before searching again. Without an LLM provider the tool still answers from retrieval, marked `degraded`.```get_answer(question="How does the authentication flow work?")```------## `get_context`--The workhorse tool. Returns docs, symbols, ownership, freshness, and community membership for any combination of files, modules, or symbols.--**Parameters:**+### `get_context`-| Parameter | Type | Required | Description |-|-----------|------|----------|-------------|-| `targets` | list[string] | Yes | File paths, module names, or symbol IDs. Batch multiple targets in one call. Symbol ids take the same `"path/to/file.py::Name"` form `get_symbol` accepts, with the same `::` / `.` / `/` separator normalisation, so an id from either tool works in the other. |-| `include` | list[string] | No | Additional data to include: `"full_doc"` (full wiki markdown), `"callers"` (who calls this, symbol targets), `"callees"` (what this calls, symbol targets), `"ownership"` (primary owner, bus factor, contributor count), `"last_change"` (last commit date + author), `"metrics"` (PageRank, betweenness, percentiles), `"community"` (cluster membership + neighbors), `"decisions"` (full decision records; default returns titles only), `"skeleton"` (file targets only; the file with bodies elided: every signature, imports, and the bodies of the most central symbols, token-budgeted; typically ~15% of the full file's tokens), `"health"` (code-health score and biomarkers), `"doc_drift"` (the documents that name this file, and whether those documents carry drift of their own). An empty `callers`, `callees` or `used_by` list sits beside a `*_basis` object: the language, how many call edges the index resolved for it, the share of those that are guesses, and a note that unbound call sites are not counted, so an empty list means no resolved edge, not proof of none |-| `compact` | boolean | No | Default `true`. Set `false` for full structure block and importer list. |-| `repo` | string | No | *(workspace only)* Target repo alias; `"all"` is not supported |+A triage card for files, modules or symbols: summary, symbols with signatures and line numbers, hotspot and fix-history cues, ownership and decisions. It returns relationships, not source. Use it before reading or editing code, and batch every target in one call. Do not call `get_symbol` once per signature; use `include=["skeleton"]` or read the file.-**Returns per target:** Documentation summary, symbols defined, ownership percentages, freshness score, co-change partners, architectural decisions governing the file. With `include` options: call graph, graph metrics, community membership, a body-elided skeleton (never source bodies). `include=["decisions"]` returns three lanes: `decisions` (accepted and governing, full records), `candidates` (proposed, nobody has accepted them, at most 3) and `history` (accepted then withdrawn, at most 2); a dismissed record is in none of them.+| Parameter | Type | Default | Meaning |+|-----------|------|---------|---------|+| `targets` | list[string] | required | File paths, module paths, or `"path::Symbol"` ids |+| `include` | list[string] | none | Any of `full_doc`, `ownership`, `last_change`, `callers`, `callees`, `metrics`, `community`, `decisions`, `skeleton`, `health`, `doc_drift` |+| `compact` | bool | `true` | `false` adds the structure block, imports and docstrings |+| `repo` | string | default repo | Workspace repo alias. `"all"` is not supported (returns an error) |-A file with no indexed symbols (README, config, plain data) gets a-`docs.file_preview` instead of an empty symbol list: line and character counts,-plus the heading spine for markdown or the first non-blank lines otherwise.-Counts and verbatim excerpts only, nothing inferred.+**Key return fields:** `targets` keyed by target, each with title, summary, symbols, `hotspot`, `fix_history` (files with counted bug fixes), `episodes`, decision titles, `file_preview` (for files with no symbols), `resolved_to` (a symbol miss falling back to its file); `dropped_targets` and `recovery` when the budget forced targets out; `_meta.complete` for whole files served.-When the symbol half of a `path::Name` target does not resolve but the file-half does, the reply is that file's card with `resolved_to` naming the file and-a `note` saying which symbol was not found. The file's symbol list is where the-correct id is, so this is a partial answer rather than a dead end.--**Several targets under one budget:** each target gets an equal share of the-response budget, and a small target's unused share passes to the others. A-target over its share degrades in place: symbols first, then its optional-blocks, largest first (named per target in `dropped_blocks`), then symbols down-to one. A miss's `suggestions` and a symbol's ambiguity `candidates` are never-shed. When even the identity cards cannot all fit, whole targets are dropped-first so the rest keep their detail; the response then leads with-`dropped_targets` and, when it fits, a `recovery` call: the same call narrowed-to the dropped targets.--**When to use:** Before reading or modifying code. Pass all relevant targets in one call to minimize round-trips. In workspace mode, enriched with cross-repo co-change and contract data.--**Example calls:**+`skeleton` renders a file with bodies elided: every signature, the imports, and the bodies of its most central symbols, with line ranges on every elision. `doc_drift` lists the documents that name the file and whether they carry drift. An empty `callers` or `callees` list comes with a `*_basis` saying how much of that language's calls the graph resolved; read it before concluding nothing calls a symbol.```-get_context(targets=["src/auth/middleware.ts"])-get_context(targets=["middleware", "api/routes", "payments"], include=["callers", "metrics"])-get_context(targets=["src/auth"], compact=false, include=["community"])+get_context(targets=["src/auth/middleware.ts", "src/api/routes.ts"], include=["callers"])get_context(targets=["src/big_module.py"], include=["skeleton"])-get_context(targets=["src/auth/service.py"], include=["doc_drift"])```-**Skeletons:** with `include=["skeleton"]`, file targets gain a structure-level-rendering sliced from the index's persisted symbol bounds (no parsing at query-time): every signature, the import preamble, and the bodies of the top symbols-ranked by graph centrality / hotspot / query match. Elision markers carry-1-indexed line ranges so you can range-`Read` anything back. For-structure-level questions ("what's in this file", "which function handles X")-this replaces a full file read at a fraction of the cost.--**Documentation drift, in reverse:** with `include=["doc_drift"]`, a file target-gains the documents that name it. The drift detector files a finding against the-*document*, so this is the only direction that answers "what documentation would-my change invalidate". Two separate claims ride in the block and must not be-merged: `references` lists documents that name this file and still resolve to-it, while `documents_with_drift` says a listed document carries some assertion-that no longer holds --- anywhere in it, not necessarily about this file.-`references_basis` states both limits, and is emitted on an empty answer too,-since a reference the detector cannot resolve is not listed. Documents dropped-by the repo's exclude rules are counted in `references_excluded` rather than-silently missing.--An answer the store cannot support is a refusal, not an empty list, because "no-document mentions this file" is a far stronger claim than "no drift findings".-`{"unavailable": "not_computed"}` means the table exists but the pass has never-run, which happens between an upgrade and the first update that does any work;-`"index_predates_doc_drift"` means the index is older than the table; and-`"drift_read_failed"` means the read failed for some other reason, where-reindexing is not the fix.+### `get_symbol`----+Returns verified, line-numbered source for one symbol, a live line range, or an omission ref. It is a follow-up read, not an entry point: use it on an id another response gave you, for a body that was elided, or to restore truncated output. `get_answer` already ships `symbol_bodies`.-## `get_symbol`--Raw source bytes for one indexed symbol with exact line bounds, cheaper and-safer than `Read` + offset math. The only tool that returns actual source code.-Also resolves **omission refs** (`repowise#<12-hex>`) from truncated responses.--**Parameters:**--| Parameter | Type | Required | Description |-|-----------|------|----------|-------------|-| `symbol_id` | string | Yes | One of three forms: `"path/to/file.py::SymbolName"` (canonical, from `get_context`'s symbol list; normalises `::` / `.` / `/` separators across languages), `"path/to/file.py:140-180"` (a live range read, 200 lines max), or an omission ref `"repowise#<12-hex>"` / a pasted whole `[repowise#...]` marker. |-| `query` | string | No | Omission refs only: return just the stored lines matching this regex (or substring). Ignored for symbol ids and range reads. |-| `context_lines` | int | No | Extra source lines before/after the symbol (0-50, default 0) |-| `depth` | int | No | Follow the call graph outward from this symbol and include what it calls, with bodies (1-3, default 1 = this symbol only). Out-of-range values clamp. |-| `repo` | string | No | *(workspace only)* Usually omitted; `"all"` is not supported |-| `reference` | object | No | A structured `continuation_reference` or `fetch_reference` emitted by this tool; pass it unchanged to retain both id and repository scope. |--**Returns:** For a symbol id or range: the source (bounded at ~600 lines,-each line prefixed with its file line number in the same format as a `Read`-result), its exact start/end line numbers, kind, and a `truncated` flag; on a-miss, an `error` with the closest matches (`fallback_lines` from a live grep).-When several indexed symbols match the id (overloads, re-exports, conditional-definitions) the response has `ambiguous: true` and a `candidates` list with-every matching body — none is silently chosen; candidates past the response-budget appear in `not_rendered` with a `fetch_with` range read. For an-omission ref: the stored content plus provenance (`source`, `created_at`,-`original_tokens`).--With `depth` above 1 the response also carries `callee_bodies`: the symbols-this one calls, transitively, each with its `depth` (hops from the root), its-source, and a `verified` flag. Every symbol appears once, at the shallowest-depth it was reached from. Callees past the response budget are listed in-`not_rendered` with the `fetch_with` range that retrieves them, so a bounded-walk never looks like a complete one.--**When to use:** When you need the body of one function or class: pipe the-`symbol_id` straight from `get_context`'s symbol list. Use the line-range form-for anything that falls between symbols. Or when a response's `_meta.omitted`-lists refs you want back and you have no shell for `repowise expand` (e.g.-Claude Desktop).--Reach for `depth=2` when you are following a call chain: reading a body,-finding the next name in it, then fetching that one. The graph already holds-those edges before the first call, so one `depth=2` call replaces the whole-sequence of round trips.--**Example calls:**+| Parameter | Type | Default | Meaning |+|-----------|------|---------|---------|+| `symbol_id` | string | none | `"path/file.py::Name"`, `"path/file.py:140-180"` (live range, 200 lines max), or `"repowise#<ref>"` |+| `id` | string | none | Alias for `symbol_id` |+| `reference` | object | none | A `continuation_reference` or `fetch_reference` this tool emitted; pass it unchanged |+| `context_lines` | int | `0` | Extra lines before and after, 0 to 50 |+| `depth` | int | `1` | 2 or 3 also returns the bodies it calls, transitively |+| `query` | string | none | Omission refs only: keep stored lines matching this regex or substring |+| `repo` | string | default repo | Workspace repo alias. `"all"` is not supported |++**Key return fields:** `source` (up to about 600 lines, each prefixed with its line number), start and end lines, `kind`, `truncated` with a continuation to pass back, `ambiguous` and `candidates` when several symbols match, `callee_bodies` (with `depth` above 1), `not_rendered` (bodies past the budget, each with a range read to fetch it), fallback lines from a live grep on a miss.```-get_symbol(symbol_id="src/auth/service.py::AuthService")-get_symbol(symbol_id="src/auth/service.py::login", context_lines=10)get_symbol(symbol_id="src/auth/service.py::login", depth=2)-get_symbol(symbol_id="src/auth/service.py:140-180")-get_symbol(symbol_id="repowise#a1b2c3d4e5f6")get_symbol(symbol_id="repowise#a1b2c3d4e5f6", query="FAILED")```------## `search_codebase`--Hybrid code search over repowise's indexes. A single tool that, depending on-the shape of the query, searches the indexed **symbols**, **file paths**, or-the **wiki**, instead of forcing a fallback to Grep for identifiers.--**Parameters:**--| Parameter | Type | Required | Description |-|-----------|------|----------|-------------|-| `query` | string | Yes | Identifier, path, or natural-language query |-| `limit` | int | No | Max results (default 5) |-| `mode` | string | No | `auto` (default) \| `concept` \| `symbol` \| `path` \| `hybrid` |-| `kind` | string | No | `implementation` \| `test` \| `config` \| `doc` |-| `symbol_kind` | string | No | Restrict symbol hits by kind (`function`, `class`, `method`, ...) |-| `page_type` | string | No | Restrict to one page type. The two you will reach for are `file_page` (the always-on per-file docs) and `module_page` (the subsystem/concept pages). Other stored types (`repo_overview`, `layer_page`, `scc_page`, `api_contract`, `infra_page`, `symbol_spotlight`) also filter. |-| `repo` | string | No | *(workspace only)* Target repo alias, or `"all"` to search across workspace |--**Modes:**--- **`auto`** (default) routes by query shape:- - an **identifier** (`GitIndexer`, `index_repo`) -> searches indexed symbols;- - a **path** (`core/ingestion/indexer.py`) -> searches file pages;- - **prose** ("how do we handle retries?") -> wiki-semantic search;- - mixed prose + identifier -> **hybrid** (symbol hits first, then concept pages).-- **`concept`** forces the original wiki-semantic behavior.-- **`symbol`** / **`path`** force the structural search.--**Returns:**+### `search_codebase`-- *Symbol hits*: `{type: "symbol", symbol_id, name, kind, file, start_line, end_line, signature, next: "get_symbol"}`. Ranked by exact-name/qualified-name match, query-token coverage, then graph centrality (PageRank / betweenness / entry-point); non-test before test unless `kind="test"`.-- *File hits*: `{type: "file", page_id, file, title, next: "get_context"}`.-- *Concept hits*: ranked wiki pages with `relevance_score`, `snippet`, `target_path`, and `sources`, the retrievers that found the page (`fts`, `vector`, or both). A `symbol_spotlight` page's `target_path` is a page identifier of the form `file.py::Symbol`; those hits also carry `file` with the openable path. **Read `file` when present.** `target_path` is for piping into `get_symbol`, not for opening.+Hybrid search that routes by the shape of the query: identifiers search the symbol index, paths resolve files, prose runs wiki-semantic search, and mixed queries run both. Use it when you want ranked hits themselves: enumerating matches, resolving an identifier to a `symbol_id`, scoping a later `get_context`. For a question, call `get_answer`; it runs this retrieval internally.-Alongside `results`, the response carries **`candidates`**: up to `limit`-distinct files worth opening next, one `{path}` entry each, best first.+| Parameter | Type | Default | Meaning |+|-----------|------|---------|---------|+| `query` | string | required | Identifier, path or natural-language text |+| `limit` | int | `5` | Max results |+| `mode` | string | `"auto"` | `auto`, `concept`, `symbol`, `path` or `hybrid`. An unknown mode runs as `auto` |+| `kind` | string | none | `implementation`, `test`, `config` or `doc` |+| `symbol_kind` | string | none | Filter symbol hits, e.g. `function`, `class`, `method` |+| `page_type` | string | none | One page type, usually `file_page` or `module_page` |+| `repo` | string | default repo | Workspace repo alias, or `"all"` to search every repo |-Every entry is a real file path, and that is the difference between the two-blocks. `results` ranks *pages*, and a page is not always a file: a-`module_page` is named by a structural group key that reads exactly like a-directory, an `scc_page` by `scc-<hash>`, an `onboarding` page by a slot name.-Ranking those is correct; opening them is not. `candidates` resolves symbol-pages to their file, collapses several symbols of one file to a single entry,-skips every page that names no file, and backfills from below the result-window so a slot spent on a module page does not also cost you a file.--**If your next move is a Read, read `candidates`.** If you are enumerating-matches or resolving a `symbol_id`, read `results`.--Tombstoned and `exclude_patterns`-excluded results are filtered. In workspace-mode, structural and concept searches both federate across repos and merge-(with `get_overview`, `get_dead_code` and `get_why` with a query, one of the-four tools that accept `repo="all"`).--**When to use:** Locating a function/class/method by name, resolving a-path-shaped query, or discovering pages by topic: the symbol/file shapes pipe-directly into `get_symbol` / `get_context`.--**Example calls:**+**Key return fields:** `results` (symbol hits carry `symbol_id`, `file`, line bounds and `signature`; file hits carry `file`; concept hits carry `relevance_score`, `snippet` and `sources`), `candidates` (up to `limit` distinct openable file paths, best first). If your next move is a Read, read `candidates`: some `results` are pages that are not files.```-search_codebase(query="GitIndexer index_repo") # -> symbol hits-search_codebase(query="core/ingestion/indexer.py") # -> file hits-search_codebase(query="rate limit OR throttle OR retry") # -> wiki pages+search_codebase(query="GitIndexer index_repo")search_codebase(query="login", mode="symbol", symbol_kind="method")```------## `get_risk`--Modification risk assessment for files or a set of changed files.--**Parameters:**--| Parameter | Type | Required | Description |-|-----------|------|----------|-------------|-| `targets` | list[string] | No | File paths to assess |-| `changed_files` | list[string] | No | Files in a PR/changeset for blast radius analysis; passing this switches the response into PR-directive mode |-| `include` | list[string] | No | Opt-in blocks: `graph` (typed `dependents`, `consumers`, `cross_repo_links`, structural `impact_surface`, `direct_risks`), `churn` (`change_magnitude`, `risk_type`, `change_pattern`) |-| `repo` | string | No | *(workspace only)* Target repo alias |+### `get_risk`-**Returns:** Per-file `hotspot_score` (0-1 churn percentile), `health_score` (0-10), hotspot status, direct directed `dependents_count`, historical `co_change_partners` (each with a recency-decayed `weight`, not an integer count, and a `direction` of `a_to_b`, `b_to_a` or `undirected`, where `a` is the assessed file and `b` the partner; `conf_ab` and `conf_ba` are the share of each file's own commits that touched the other, and both are omitted on an index written before those commit totals were recorded, which also reads as `undirected`), blast radius, recommended reviewers, test gap analysis, and security signals. With `include=["graph"]`, `dependents` preserves direct versus transitive structural reach, `consumers` contains typed contract consumers only, and `cross_repo_links` retains both repository identities, direction, relationship type, evidence kind, and file- or repository-level granularity. Package-manifest links are repository-level and never invent a target file. Every typed relationship collection carries matching total/emitted/truncated fields. `relationship_analysis` distinguishes available-empty analysis from unavailable, degraded, partial, and source-truncated artifacts and retains artifact generation/provenance fields. Structural reach is not proof of runtime breakage.+What history says about touching a file: hotspot score, bug-fix record, owners, co-change partners, dependents and security signals. Pass `changed_files` for PR mode, where the response leads with a `directive` naming what may break and which tests to run. To review a commit or diff itself, use `get_change_risk`.-> **Unresolved targets.** A target naming no indexed file comes back as-> `{"resolved": false, "unresolved_reason": ...}` with no counts — omitted-> rather than zeroed, since a structural zero reads as a measured one. Reasons:-> `unsupported_target_kind` (a `module:` id; risk is scored per file),-> `directory`, `not_indexed` (run `repowise update`), `no_such_path`.+| Parameter | Type | Default | Meaning |+|-----------|------|---------|---------|+| `targets` | list[string] | `changed_files` | File paths to assess |+| `changed_files` | list[string] | none | Files in a change; switches on PR mode |+| `include` | list[string] | none | `graph` (typed dependents, consumers, cross-repo links), `churn`, `scales` (units and calibration, identical per call) |+| `repo` | string | default repo | Workspace repo alias. `"all"` is not supported |-> **Opt-in blocks.** `impact_surface` and `direct_risks` are pagerank floats an agent cannot rank; `change_magnitude`, `risk_type` and `change_pattern` restate numbers printed beside them. All five are computed regardless and feed `risk_summary`; `include` only decides whether they ship. `global_hotspots` accompanies a multi-target call only, being ambient orientation that a single named file does not need; it ranks by fix history the same way `defect_profile` does.+**Key return fields:** per file: `hotspot_score` (0 to 1), `health_score` (0 to 10), `dependents_count`, `co_change_partners`, owners, test gaps, `security_signals`. In PR mode, `directive` with `may_break`, `may_break_tests`, `missing_cochanges`, `test_recommendations` (each `measured` or `inferred`), `tests_to_run`, `tests_to_run_basis`, `next_calls`, and the 0 to 10 `structural_impact_score`. A target naming no indexed file returns `resolved: false` with a reason, never zeroed counts.-> **Scales.** Every response carries the facts that stop a misreading: unit,-> range, calibration status, and whether the value is authoritative. The-> per-field dictionary behind them never varies between calls, so it ships only-> with `include=["scales"]` - ask for it once per session, not per call. That-> dictionary describes indexed file values: `hotspot_score`, `owner_pct`, and-> `recent_owner_pct` are 0-1 ratios, while `risk_type` is an-> uncalibrated category. In PR mode, `structural_impact_score` is a deterministic,-> uncalibrated 0-10 structural-exposure heuristic; `localized` is below 4,-> `moderate` is 4 to below 7, and `broad` is 7 or above. It is not a runtime--> breakage probability and is not authoritative for live change review.-> Deprecated `overall_risk_score` remains an exact alias, with migration metadata.-> Direct rows expose raw, unbounded `structural_score` values in-> pagerank-weighted-hotspot units. These are not comparable to-> `get_change_risk.score`. Coverage and gap fields are percentages from 0-100.-> Every emitted float is rounded to 4 significant digits.--When `changed_files` is passed, the exact serialized response starts with a `directive` block. Its core lists are the local blast radius: `may_break` (production files in structural reverse-import reach of the diff, candidates for review rather than proven breakage), `may_break_tests` (test files reached the same way, kept separate so a burst of tests doesn't crowd production impact out of the capped list), `missing_cochanges` (historical co-changers absent from the diff), compatibility `missing_tests`, and additive typed `test_recommendations`. Every recommendation retains `basis`, all retained `bases`, repository identity, source files, and evidence: `measured` means the per-test coverage map found the test; `inferred` means structural reachability found a candidate and is not coverage proof. `tests_to_run` preserves the older measured-first/fallback id projection and `tests_to_run_basis` remains `measured`, `inferred`, or `none`; `files_without_measured_tests` carries the narrower typed coverage claim. Matching total/emitted/truncated/omitted fields describe each exact pre-cap population. `coverage_analysis`, `test_inference_analysis`, and `test_analysis` distinguish available-empty evidence from unavailable, stale, partial, or degraded analysis; unavailable coverage with `tests_to_run: []` never means that no tests are needed. The full typed population is shared with the REST blast-radius response, while any budget-omitted directive rows are recoverable through the response omission marker. In workspace mode the directive also carries the cross-repo fallout of the changed repo:--- `will_break_consumers`: deprecated compatibility name for services in *other* repos that structurally depend on this one. Rows carry `claim: structural_reach`, `runtime_breakage_claim: false`, both repository roles, direction, distance, and aggregated edge kinds; the sibling `will_break_consumers_semantics` is `structural_reach_only`. Matching total/emitted/truncated fields describe the exact pre-cap structural population, while `cross_repo_relationship_analysis` labels unavailable or partial edge provenance.-- `missing_cross_repo_cochanges`: services in other repos that historically co-change with this one but aren't in the diff.-- `breaking_changes`: compatibility-named list of provider incompatibilities and comparison warnings since the last index. Entries carry `contract_id`, `kind`/`severity`, optional request/response `side`, `comparison_source`/`comparison_key`, and endpoint-exposed `impacted_consumers` (repo, service, file). OpenAPI findings require complete sides with matching fidelity; unsupported or changed extraction evidence becomes uncertainty, not field removals. Consumer links do not prove field use or runtime failure. See [Breaking-Change Guard](../scale/WORKSPACES.md#breaking-change-guard).-- `conformance_violations`: declared dependency-rule breaches the diff's repo participates in, each with the offending `source`/`target` services, the `rule` (e.g. `frontend !-> db`), and `edge_kind`. See [Architecture Conformance](../scale/WORKSPACES.md#architecture-conformance).-- `dependency_cycles`: circular service dependencies involving this repo, each with the participating `nodes` and `length`.--`directive.next_calls` names what to call next, built from this response alone: `get_change_risk()` for the diff itself, then `get_context(targets=<first five may_break>, include=["callers"])` when `may_break` is non-empty. `tests_to_run` is already the answer to which tests, so it gets no call. Each entry has the shape every next call has (see `get_health`'s `fix_first`).--> **Output-schema change.** `directive.will_break` is now `directive.may_break`,-> and `directive.will_break_tests` is now `directive.may_break_tests`. Both are-> a reverse-import reachability walk: `get_risk` is given a file list, never a-> diff, so it cannot know whether the symbol an importer actually uses changed.-> The old name promised a precision the analyzer does not have.-> `will_break_consumers` temporarily keeps its old key for compatibility, but-> it is structural reach only and is explicitly marked deprecated. Only-> `breaking_changes` comes from incompatible contract diffing.--**When to use:** Before modifying files, especially hotspots. Understand what could break, who to involve in review, and whether tests cover the affected area.--**Example calls:**+Dependent counts are a floor over the indexed graph, and structural reach is not proof of runtime breakage. `structural_impact_score` is an uncalibrated heuristic, not a probability.```get_risk(targets=["src/auth/middleware.ts"])get_risk(changed_files=["src/api/routes.ts", "src/middleware/cors.ts"])-get_risk(targets=["src/auth/middleware.ts"], include=["graph", "churn"])```----+### `get_change_risk`++Reviews one commit, a `base..head` range, or uncommitted work by comparing the two revisions directly; it needs no index refresh. It reports what the change newly made worse across defect, maintainability and performance, which tests cover the changed lines, and the changed files' bug-fix record. Use it before merging. Use `get_risk` for an indexed file's history.++| Parameter | Type | Default | Meaning |+|-----------|------|---------|---------|+| `revspec` | string | uncommitted work | A commit, `base..head`, or `base...head` (from the merge-base). Pass `HEAD` when the tree is clean |+| `extensions` | list[string] | all | File suffixes to count, e.g. `[".py", ".ts"]` |+| `exclude_patterns` | list[string] | none | Gitignore-style paths to omit; combined with a root `.riskignore` |+| `include_paths` | list[string] or string | all | Gitignore-style paths to keep, as a list or one comma-separated string |+| `baseline` | int | `200` | Recent commits sampled for percentile ranking; `0` disables percentiles |+| `include` | list[string] | none | `findings` (every finding), `diagnostics` (raw score mechanics), `scales` (units) |+| `finding_id` | string | none | Expand one `health_delta` finding |+| `repo` | string | default repo | Workspace repo alias. `"all"` is not supported |++**Key return fields:** `directive` (`status` of `review_required`, `review_recommended`, `clear_in_analyzed_scope` or `unknown`, with reasons and next actions), `health_delta` (introduced, worsened and resolved findings with `status` and `top_findings`), `impacted_tests`, `patch_coverage` (when coverage is stored), `fix_history`, `branch_overlap`, `independent_changes`, `diff_shape`, `risk_percentile`, `cross_repo` (workspace mode).-## `get_change_risk`--Review one commit, a `base..head` range, or uncommitted work. Unlike-`get_risk`, which evaluates indexed files and can report blast radius, this-compares the two revisions directly and needs no index refresh.--**Parameters:**--| Parameter | Type | Required | Description |-|-----------|------|----------|-------------|-| `revspec` | string | No | Commit, `base..head` range, or `base...head` (diffed from the merge-base) to score. Omit it to score uncommitted work, or pass `HEAD` when the tree is clean |-| `repo` | string | No | *(workspace only)* Target repo alias |-| `extensions` | list[string] | No | File suffixes to count, such as `[".py", ".ts"]` |-| `exclude_patterns` | list[string] | No | Gitignore-style paths to omit; combined with root `.riskignore` rules |-| `baseline` | int | No | Recent commits to sample for percentile ranking (default `200`; `0` disables every percentile, `risk_percentile` and `fix_history.percentile` alike) |-| `include` | list[string] | No | `"findings"` for every change finding, `"diagnostics"` for the raw score mechanics, `"scales"` for units and calibration |-| `finding_id` | string | No | Expand one `health_delta` finding by its id |--**Returns:** `directive` leads with a `status`-(`review_required`, `review_recommended`, `clear_in_analyzed_scope`, `unknown`),-a headline, bounded reasons, and concrete next actions.--`health_delta` is what the change newly made worse, across defect,-maintainability and performance. Both revisions are analysed from their own-content, so a finding present at head is reported only when the diff explains-it. `scope` counts changed, eligible, analysed, skipped and failed files, and-`status` distinguishes `available` from `partial` and `unavailable` — a-`partial` comparison is never a clean bill, and `skipped` says why each file-was left out. `introduced`, `worsened` and `resolved` are totals;-`top_findings` carries the three most actionable, with `findings_total` and a-recovery call for the rest.--Each finding names its `dimension`, `biomarker`, `severity`, `path`, `symbol`-and head-side `lines`, a `reason`, and an `attribution` — `basis`-(`added_lines`, `changed_symbol`, `changed_call_edge`, `new_file`,-`file_change`, `context_change`, `unknown`) with a `confidence`. Identity-ignores line numbers, so moving code introduces nothing and a rename carries-its findings across. Performance findings carry `opportunity_id` and-`opportunity_rank` and are ordered by opportunity rank and actionability, never-by defect impact, which is zero for them by construction. `inspect` gives the-exact `finding_id` call that expands one; ids are bound to the two revisions-that produced them. A finding that exactly matches a stored one also carries a-`health_reference` for `get_health(finding_id=...)`.--`fix_history` reports the recency-weighted bug-fix record of the files the-change touches, with `files` naming where the pressure sits and `percentile`-ranking it against the same measure over the repo's own recent commits. It is-the part that separates a small edit to a fragile file from a large edit to a-safe one. `available` is false when the history walk could not run.--`diff_shape` is one sentence ranking the diff's size against the repo's recent-commits, alongside the top-level `risk_percentile`, `review_priority`,-`classification` and `is_fix`. It is size and spread, never a danger verdict.-`working_tree` says whether uncommitted work was the subject.--The raw 0-10 `score` is not on the wire by default. It ranks 0.99 against lines-added on every repository measured, so the percentile beside it already carried-the ranking while the number invited being read as a probability. It remains-available, with `fallback_band`, behind `include=["diagnostics"]`.--`include=["diagnostics"]` adds the raw mechanics: `score`, `fallback_band`,-`risk_authority`, `score_measures`, `score_unit`, `baseline_sample_size`,-`features` and `drivers`. `include=["scales"]` adds each field's kind, unit, range,-calibration and thresholds. Both are identical on every call, so ask once.--It also returns `impacted_tests`, whose `tests_to_run` names the tests the-per-test coverage map proves execute the change's changed *lines* (line-precise,-so a narrower set than `get_risk`'s file-level `tests_to_run` — same field name,-because it is the same concern). It is capped at ten, with `total` and-`truncated` reporting the overflow and the tail written to the omission store:-on `truncated: true` the response carries an `omission_marker`, and-`repowise expand <ref>` returns the rest. Its `line_coverage` buckets flag-`untested_changes` (covered file, uncovered change), `stale_test_candidates`-(covered lines whose guarding test file is absent from the diff), `covered`, and-`no_coverage_data` (files absent from the map). When no map is ingested the-change is never reported as untested: `status` becomes `inferred` when the-import graph can name test files reaching the change (candidates, file-level, no-line attribution, and `line_coverage` stays empty because reaching cannot speak-to lines), and `no_map` ("run the full suite") when it cannot. `basis` carries-the same distinction in one word, always present: `measured`, `inferred`, or-`none`. `tests_to_run_kind` says what each entry is: `test_id` (a coverage-map-test id, measured) or `test_file` (inferred), null when `basis` is `none`;-`get_risk`'s directive carries the same field beside `tests_to_run_basis`. Build the-measured map with `coverage run --contexts=test` followed by-`repowise coverage add`.--When the index stores coverage, the response also carries `patch_coverage`:-the share of the change's executable lines the stored coverage ran, the same-computation and JSON shape `repowise coverage check --format json` gates on.-`patch_coverage_pct` is null when no changed line is executable, files the-coverage never names read `not_in_report` rather than 0%, and-`scope.freshness` is `stale` when the coverage was measured at another commit-than the change's head (for uncommitted work: ingested before the newest-edit), so its line numbers may describe other code. `path_gates` lists the-path-scoped gates in-`coverage.gates`, each judged on the changed files its globs match (`gate`-reads `fail` when one that is not informational fails). They are judged only-on coverage measured at the change's head and valid config; otherwise they-read `no_data`, and `scope.config_errors` names each invalid entry. The block-is absent when no coverage is stored.-Each file row carries `risk`: `fix_pressure` (recency-weighted bug-fix weight-from the checkout's git history), `dependents`, `hotspot` and `bug_magnet`-(from the index), `risky`, `reasons` and `basis` (`git_and_index`, `git`,-`index` or `unavailable`). Rows are listed riskiest first, and `risky`-(`file_count`, `covered_line_count`, `coverable_line_count`,-`patch_coverage_pct`, `threshold`, `gate`) summarizes coverage over the risky-files, null when no row's risk was assessed.-Each measured row also carries `hints`, one per uncovered range (the first-eight): `range`, `symbol` (the innermost indexed symbol, null outside any),-`tests` (up to three test files to extend, best first), `basis` (`per_test`,-`call_graph`, `import_graph` or `none`) and `total` (how many qualified before-the cap). `per_test` is measured (per-test coverage ran nearby lines);-`call_graph` and `import_graph` are inferred from the graph. `hints` is null-when the index could not be read. When changed lines are uncovered,-`directive.next_actions` gains one line naming the scope and the tests to-extend, or, when the coverage is stale, saying to re-run the tests first.--Without a `revspec`, `patch_coverage` covers everything a push would bring,-diffed from the merge-base with the CI or default base branch: `scope.label`-`origin/main...working tree` on a dirty tree (untracked files included),-`origin/main...HEAD` on a clean one, plain `working tree` when no base-resolves. For uncommitted work freshness is by time: `current` when the last-coverage ingest came after the newest edit to the changed files, else-`stale`. After a full test run the augment hook re-ingests a fresh report in-the background, and `get_change_risk` reads it once the ingest finishes.--In workspace mode the response also carries `cross_repo`, and every-`cross_repo.consumers[]` row gains a `tests` block: a `state` (`measured`,-`inferred`, `none` or `unresolved`), up to five `tests_to_run` rows carrying-`test_file`, `test_id`, `basis`, `via` and `confidence`, `total` and-`truncated` for the overflow (the tail goes to the omission store), and-`unresolved_reason` / `unresolved_detail` when the join could not be followed.-Only `tests_to_run` is capped at five; `total` is the true number of tests-found for that consumer, so `truncated: true` means `total` minus five went to-the omission store. `unresolved_reason: "lookup_failed"` is the state every row-lands in when the join itself failed, as opposed to one link that could not be-followed, and `unresolved_detail` names what failed.--> **Output-schema change.** `impacted_tests.tests` is now-> `impacted_tests.tests_to_run`, matching `get_risk`'s directive.--`is_fix` is the defect benchmark's keyword rule read over the commit subject,-not the conventional-commit type, so a `feat:` commit whose subject says it-fixes something reads true; the rule is frozen for comparability rather than-tuned. `fix_history.overlap` is the tuned view: it applies a diff-shape filter-on top of that rule, counting only commits that actually edited production-code. `fix_history` itself runs the same unfiltered rule, and `overlap` is the-one part of the three that needs an index.--When the changed files carry counted bug fixes, `fix_history.overlap` reports-per file how many past bug-fix commits touched it-(`fix_count`), how many of the change's lines fall inside the ranges one of-those fixes replaced (`overlapping_lines`), and how long ago the most recent-was (`last_fix_days_ago`). `total_fixes` counts distinct commits, not rows,-and `files` is capped at ten with `truncated` reporting overflow.--Each file also carries how much of *this* change sits in it — `changed_lines`-and `share_of_change` — with `changed_lines_in_fixed_files` as the total across-them. That join is what lets the response say where the risk sits rather than-only that some touched file has a past: the score is whole-change, so when one-returned file holds at least half the changed lines, `concentration` names it.--`overlapping_lines` is labelled `approximate` in the payload, and that label is-load-bearing: a past fix's ranges are numbered against its own parent commit,-so anything that moved lines in between shifts them. Read it as "this-neighbourhood has been patched before", not "this exact line". The per-file-`fix_count` beside it carries no such caveat. The whole block is aggregate and-never names the commit that introduced a bug: file-level SZZ measured 74.5%-precision on this repo's frozen judgments, which is enough to count fixes and-not enough to accuse one commit of causing them. The block is absent entirely-on an index with no fix history.--In workspace mode the response also holds `cross_repo`, built from the artifacts-of the last `repowise update --workspace` rather than from a graph traversal. It-appears when the commit touches a file that provides a contract some other repo-consumes, or when the compatibility report attributes a finding to one of the-changed files. `consumers[]` names each link with its `provider_file`, the-consumer's `repo`, `file` and `contract_id`, the `contract_type` and the-`match_type` that joined them, plus `provider_symbol_id` and `symbol_id` when the-link is symbol-level; `consumer_repos` lists the other repos in one place.-`breaking_changes[]` carries the `contract_id`, `type`, `kind`, `severity`,-`detail`, `provider_file` and `impacted_repos` of each incompatibility or-comparison warning, together with side/source/key evidence when available.-Impacted repositories are directly linked endpoint exposure, not proof of a-runtime failure. `breaking_changes_available` says whether a detection pass ran at-all, so an empty list reads as silence rather than as an all-clear, and-`breaking_changes_as_of` stamps that half only. `consumers` is capped at ten and-`breaking_changes` at five, with `consumers_truncated` and-`breaking_changes_truncated` counting what the caps left out; the block answers-whether this commit crosses a repo boundary, and `get_blast_radius` is the tool-for the full traversal. It is absent outside workspace mode, without contract-artifacts, and when the commit touches no published file.--`branch_overlap` names the other open branches editing the files this change-edits. It is git-only, so it appears whether or not the repo is indexed; an index-only orders the shared files and adds the history rows. `base` and `current` name-the two ends of the comparison, and each `branches[]` entry carries the branch-name, `ahead` and `behind` commit counts, `last_commit` (the date), and `files[]`.-Every file row states its `basis` in words, either `same file` or-`co-change pair, N of M commits`, the second carrying the `partner` file of this-change it pairs with and appearing only under a branch that already shares a file-directly, at most three per branch. `scanned`, `total` and `truncated` report the-branch scan itself, which is bounded to the newest 50 branches by committer date.-There is no score and no percentage in the block. `branches` is capped at five and-each entry's `files` at ten, both through the shared response budget: a capped-list gains `<key>_total`, `<key>_emitted`, `<key>_truncated`, `<key>_omitted` and-`<key>_reduced_reason` beside it, and the omitted rows go to the omission store,-so on `branches_truncated` or `files_truncated` the response carries an-`omission_marker` and `repowise expand <ref>` returns the rest. `truncated` is a-different fact: it reports the branch scan bound, not a cap. The block is absent when the change has-no counted files, when no other branch edits a shared file, and when the scan-exceeds its 20-second ceiling or git cannot answer.--`independent_changes` says when the diff is several changes rather-than one. It groups the changed files by connectivity, over index edges (imports,-calls, type references, framework and dynamic edges), stored co-change pairs, and,-when `revspec` is a `base..head` range, the files each commit of that range-touched, which links them to each other. A single commit and uncommitted work-carry no commit evidence, so only a range reads it. Only a changed file that is in-the index, is not a test, and is written in a language whose resolver can emit an-import edge is eligible to be grouped: docs, config and data files are never in a-group, not even through a co-change pair, and tests never join or connect one.-`count` is the number of groups; each `groups[]` entry lists its `files` and its-`bridging_files`, the files that alone hold the group together, where moving one-out would split it (named only for groups of three or more files, and most often-the file two commits of the range share). `ungrouped_files` carries every changed-file left out of the grouping, and `summary` names the reasons in those terms:-docs, config, tests, files not in the index, or files it has never linked. `basis` states in words-what was actually checked, and its sentence changes with the subject: it names a-shared commit alongside the import, call, type reference and co-change pair when-the commits of a range were read, and omits it when there were none to read. Under-either wording it is a claim about this index and not about the code. There is no-separate key saying which sentence you got; the sentence itself says it.-`ungrouped_files` is capped at ten through the shared response budget, so a capped-list gains `ungrouped_files_total`, `_emitted`, `_truncated`, `_omitted` and-`_reduced_reason` beside it and the omitted names go to the omission store,-recoverable with `repowise expand <ref>`; nothing else in the block is capped. The-block needs an index and is absent without one, and it is absent whenever the diff-is one change: fewer than two changed files, fewer than two of them eligible to be-grouped, or fewer than two groups surviving. Under a response over budget it is-the first thing shed, ahead of `diff_shape`; `branch_overlap` sheds after-`fix_history` and before `cross_repo`.--The freshness envelope is scoped to the files this change edits, whether or not-the repo is indexed: `branch_overlap` reads files on other branches, and that-never widens what the response is about.--**When to use:** Before merging a commit or PR range, especially when you need-to assess the change itself rather than the risk of an already-indexed file.--**Example calls:**+Trust `health_delta.status`: `partial` means files were skipped and the change is not cleared. `diff_shape` describes size and spread, never danger. An empty diff returns `status: "nothing_to_score"`.```get_change_risk()-get_change_risk(revspec="main..HEAD", extensions=[".py"], exclude_patterns=["tests/"])-get_change_risk(revspec="main..HEAD", include=["findings"])+get_change_risk(revspec="main..HEAD", exclude_patterns=["tests/"])get_change_risk(revspec="main..HEAD", finding_id="chf_27a13be11e7ee33f")```------## `get_why`--Architectural decision intelligence. Falls back to git archaeology when no decision records exist for a path, and further to a rationale comment mined live from the source when neither decisions nor git history explain the "why".--**Parameters:**--| Parameter | Type | Required | Description |-|-----------|------|----------|-------------|-| `query` | string | No | Natural language question about decisions, OR a file/module path |-| `targets` | list[string] | No | File paths to anchor an NL `query` search to |-| `repo` | string | No | *(workspace only)* Target repo alias, or `"all"` (only when `query` is given) |-| `id` | string | No | A decision id or `ev_...` evidence id previously emitted by `get_why`; resolves it directly without relevance search |-| `reference` | object | No | An emitted evidence reference object; pass it unchanged to retain both id and repository scope. |--**Modes:**--1. **NL search**: pass a question, optionally anchored to `targets`: `get_why(query="why JWT over sessions?")` -> searches decision records.-2. **Path-based**: pass a file path as `query`: `get_why(query="src/auth/service.ts")` -> returns three lanes, `decisions` (accepted, governing), `candidates` (nobody accepted them) and `history` (accepted and since replaced), plus the file's origin story.-3. **Health dashboard**: no `query`: `get_why()` -> stale decisions, conflicts, ungoverned hotspots, retired records and accepted records that name no file.-4. **Reference lookup**: pass `id`: `get_why(id="ev_...")` -> the exact evidence and supporting decision in one call.--**Returns:** Matching decision records with title, rationale, alternatives considered, affected files, staleness score. Health mode returns stale decisions, conflicts, ungoverned hotspots, `retired_decisions` and `unscoped_decisions`.--Two health-mode lanes `counts` reported as a bare number now name their records: `retired_decisions` (superseded, deprecated, dismissed — each row carries its `lane`) and `unscoped_decisions` (accepted records naming no file). Five rows each, ranked, remainder in `_meta.omitted`; full sizes stay in `counts`, split across the three status keys for the retired lane. `active` stays count-only.+### `get_why`-`answer_basis` names the strongest lane the response rests on: `decision`, `episode`, `rationale`, `archaeology`, `documentation`, or `candidate`. Only `decision` is a ruling; the rest are evidence to weigh. Absent when no lane was served, and on the health dashboard.+Why code is shaped the way it is: decision records, with git archaeology and mined rationale comments as fallbacks when no decision governs a path. Call it before a refactor or before diverging from an existing pattern.-**The lane a record is in decides whether it binds you, and path mode puts it in one.** `decisions` holds accepted records: somebody accepted each in a recorded event naming the reason, the scope, the evidence and the accepter, so treat them as constraints. `candidates` holds records something inferred and nobody has agreed to; read them as hints and never as rules, and note the `candidates_note` beside them says so too. Nothing produces an acceptance except an explicit `repowise decision confirm` or a committed ADR that says it is accepted, so a candidate that has recurred across fifty sessions is still a candidate.+| Parameter | Type | Default | Meaning |+|-----------|------|---------|---------|+| `query` | string | none | A question, or a file or module path. Omit for the decision health dashboard |+| `targets` | list[string] | none | Paths to anchor a question to, or to ask about on their own with no `query` |+| `id` | string | none | A decision id or `ev_...` evidence id from an earlier response |+| `reference` | object | none | An evidence reference this tool emitted; pass it unchanged |+| `repo` | string | default repo | Workspace repo alias, or `"all"` (only with a `query`) |-Do not read the lane off `status`. That column is a projection kept in step for readers that predate the split, and a record can carry `status: "active"` with no acceptance behind it at all. An accepted record instead carries a `currency`: `active` (still describes its code), `needs_review` (its files have moved, and it still binds), `uncheckable` (it names nothing, so nothing can check it), `superseded` or `dismissed`. A candidate carries `review_state: "open"` and no `currency`.+Modes: a question searches decision records; a path returns that file's `decisions`, `candidates` and `history` plus its origin story and `alignment`; `targets` alone asks about those files; no arguments returns the health dashboard (stale decisions, conflicts, ungoverned hotspots); `id` resolves one record directly.-Path mode's `alignment` counts the lanes separately and they sum to `governing_count`, which is every record naming the file: `active_count` is what governs it, `deprecated_count` what was accepted and withdrawn, `uncheckable_count` what was accepted but names nothing, `candidate_count` what is merely awaiting review. `score` is derived from `active_count` alone, so a file with `active_count: 0` is ungoverned however many candidates name it.--The `candidates` and `history` lanes are capped at three rows each and shed first under response-budget pressure, so an absent lane means the budget was tight, not that it was empty. `get_overview`, `get_risk` directives and `get_answer` serve accepted records only; a candidate reaches none of them as an instruction.--**When to use:** Before architectural changes, understand existing intent and constraints. After changes, record new decisions.--**Example calls:**+**Key return fields:** `decisions` (accepted, binding), `candidates` (inferred, nobody accepted them: hints, not rules), `history` (accepted, since withdrawn), `answer_basis` (the strongest lane served; only `decision` is a ruling), `alignment`, `git_archaeology`, `evidence_refs`.```-get_why(query="rate limiting")+get_why(query="why JWT over sessions?")get_why(query="src/payments/processor.ts")-get_why(query="why is caching split from the eviction path?", targets=["src/cache"])get_why()```------## `get_dead_code`+### `get_overview`-Unreachable code, unused exports, unused internals, and zombie packages, sorted by confidence tier with cleanup impact estimates. Flag-based, not include-list-based.+The architecture map: summary, key modules, entry points, layers, code and git health, and the next actions for the week and quarter. Call it once, first, in an unfamiliar repo. Skip it afterwards; it does not change within a session.-**Parameters:**+| Parameter | Type | Default | Meaning |+|-----------|------|---------|---------|+| `include` | list[string] | none | Any of `content` (full essay), `outline` (wiki page tree), `tour` (guided tour and reading order), `decisions`, `graph` (community clusters), `ownership` (top owners) |+| `repo` | string | default repo | Workspace repo alias, or `"all"` for the cross-repo topology |-| Parameter | Type | Required | Description |-|-----------|------|----------|-------------|-| `repo` | string | No | *(workspace only)* Target repo alias |-| `kind` | string | No | Restrict to one finding kind: `unreachable_file` \| `unused_export` \| `unused_internal` \| `zombie_package` |-| `min_confidence` | float | No | Minimum confidence floor (default `0.4`; `0.7`+ is cleanup-ready only) |-| `safe_only` | boolean | No | Deletion-ready findings only, excluding anything with runtime-load risk (default `false`) |-| `limit` | int | No | Max findings per tier, clamped to 25 (default 20) |-| `tier` | string | No | Restrict to one tier: `high` (>= 0.7) \| `medium` (0.4 to 0.7) \| `low` (< 0.4) |-| `directory` | string | No | Path-prefix filter |-| `owner` | string | No | Primary-owner filter |-| `group_by` | string | No | Roll findings up by `directory` or `owner` instead of listing them flat |-| `include_internals` | boolean | No | Include private/underscore symbols (default `false`) |-| `include_zombie_packages` | boolean | No | Include zombie-package findings (default `true`) |-| `no_unreachable` | boolean | No | Exclude `unreachable_file` findings (default `false`) |-| `no_unused_exports` | boolean | No | Exclude `unused_export` findings (default `false`) |-| `finding_id` | string | No | Resolve an emitted stable finding `id` directly in one call |--**Returns:** Dead code findings grouped by confidence tier (high >= 0.7, medium 0.4 to 0.7, low < 0.4). Each finding includes: file path, kind, confidence score, line count, and cleanup impact estimate. In workspace mode, confidence is lowered on findings other repos still import. `summary.call_resolution_basis` lists, per language, how many call edges the index resolved and what share are guesses, which is the graph the findings rest on. `summary.scope` says the totals cover every open finding while `filtered_findings` and the tiers cover what the filters kept; `summary.filters` echoes each non-default filter.--**When to use:** Cleanup tasks, not a targeted fix. Conservative by design: `safe_only` excludes dynamically-loaded patterns and framework-decorated functions.--**Example calls:**+**Key return fields:** `title`, `content_md` (summary section by default), `key_modules`, `entry_points`, `architecture`, `code_health`, `git_health`, `next_actions`, `more` (names the opt-in blocks), and a `workspace` footer in workspace mode.```-get_dead_code()-get_dead_code(min_confidence=0.7, tier="high", safe_only=true)-get_dead_code(kind="unused_export", group_by="owner")+get_overview()+get_overview(include=["outline", "content"])```----+### `get_health`++Code-health scores and findings from the stored analysis, across defect risk, maintainability and performance. With no `targets` it returns a dashboard led by `fix_first`, one ranked queue of what to fix first. With `targets` it scores those files. It never recomputes: commit, then run `repowise update`, to see new numbers. No LLM calls.++| Parameter | Type | Default | Meaning |+|-----------|------|---------|---------|+| `targets` | list[string] | none (dashboard) | File paths or `module:<name>`. Misses are named in `unresolved` |+| `include` | list[string] | none | Blocks: `biomarkers`, `refactoring`, `trend`, `coverage`, `accuracy`, `signals`, `churn_complexity`, `doc_drift`, `semantics`, `unverified`. Dimension filters: `performance`, `defect`, `maintainability`, `advisory` |+| `only` | list[string] | none | Keep just these top-level keys; identity, totals and recovery fields always survive |+| `limit` | int | `20` | Max rows in every ranked list, capped at 50; `0` for none |+| `cursor` | int | `0` | Offset into a ranked list; the `recovery` block names the next call |+| `fix_id` | string | none | Open one `fix_first` item in full |+| `finding_id`, `plan_id` | string | none | Open one finding or refactoring plan by id |+| `opportunity_id` | string | none | Open one opportunity: `perf...` for performance, `refop...` for a refactoring |+| `refactoring_view` | string | `"diversified"` | `diversified`, `canonical` or `file_spread` |+| `refactoring_scope` | string | `fix_first` without targets, `all` with | Which open refactoring opportunities to list |+| `refactoring_type`, `refactoring_confidence`, `refactoring_effort` | string | none | Refactoring queue filters |+| `performance_view` | string | `detail` | `detail` or `summary` |+| `performance_context` | string | `production` | `production`, `tooling`, `test`, `unknown` or `all` |+| `performance_boundary`, `performance_confidence`, `performance_actionability`, `performance_sort` | string | none | Performance queue filters |+| `scope` | string | `"all"` | `production` drops test files from every figure |+| `counts` | string | `"everything"` | `code_shape` drops the git-derived half of the score |+| `repo` | string | default repo | Workspace repo alias. `"all"` is not supported |-## `get_health`+Only one of `fix_id`, `finding_id`, `plan_id`, `opportunity_id` per call; passing two returns `mode: "conflict"`.-Code-health marker scores: the same deterministic markers the-`repowise health` CLI computes, across three signals (defect risk,-maintainability, performance), exposed for agentic workflows. Zero LLM calls.-Use it to inspect stored health analysis before a change and after committing-health-relevant changes and running `repowise update`: neither re-calling-`get_health` nor updating an uncommitted working tree recomputes those metrics.+**Key return fields:** `mode`, `fix_first` (`lead`, up to five `items` each with a `next_call`, `totals`), `kpis`, `gap_analysis`, `worst_files`, `high_leverage_files` (ranked by `weighted_deficit`), `top_findings`, `unresolved`, and the opt-in blocks you named. `_meta.health_analysis` says whether stored analysis exists and which commit it describes.-**Safe recipes:**+The response is bounded. Pair `include` with `only` to keep one block, e.g. `get_health(include=["refactoring"], only=["refactoring_opportunities"])`.-```text+```get_health(only=["fix_first"])get_health(fix_id="fix1_...")-get_health(targets=["path"], include=["refactoring"])-get_health(targets=["module:path"], only=["modules","metrics"])-get_health(include=["trend"], only=["trend"])-get_health(include=["accuracy"], only=["accuracy"])+get_health(targets=["src/api/server.py"], include=["signals"])+get_health(include=["performance"], only=["performance_summary"])get_health(include=["coverage"], only=["coverage"])-get_health(include=["performance","refactoring"], only=["performance_opportunities","refactoring_plans"])-get_health(include=["refactoring"], only=["refactoring_opportunities"], limit=6)-get_health(opportunity_id="refop2_...")```-**Parameters:**--| Parameter | Type | Required | Description |-|-----------|------|----------|-------------|-| `targets` | list[string] | No | File paths, or `module:foo` to expand a module's file set. Empty means dashboard mode. |-| `include` | list[string] | No | Opt-in blocks (default response stays lean): `"biomarkers"` (findings in dashboard mode), `"refactoring"` (structured, graph-aware refactoring plans; see below), `"trend"` (snapshot diff + declining / predicted-decline alerts), `"coverage"`, `"accuracy"` (the "does the score find the bugs?" stat, dashboard mode), `"signals"` (per-file process / people / topology signals, targeted mode), `"churn_complexity"` (churn x complexity quadrant points, dashboard mode), `"doc_drift"` (documentation whose claims about the tree no longer hold), `"semantics"` (`_meta.health_semantics`, the legend for weighted-deficit points and percentiles; the same on every call, so off by default), and a dimension name (`"performance"` / `"defect"` / `"maintainability"` / `"advisory"`) to filter findings to that pillar. `"advisory"` is the only way to reach the advisory markers: `assertion_free_test` and `mock_saturated_test`. They carry a health impact of exactly zero, so they are out of every impact-ranked list by default: asking for the dimension is what returns them, and their absence from an unfiltered response is not a clean bill. |-| `only` | list[string] | No | Keep just these top-level keys. `include` adds blocks, `only` subtracts them. `mode`, `_meta`, `unresolved`, `known_modules` and each kept list's `*_total` sibling always survive. The three `include` **block** names work as aliases: `biomarkers`→`findings`, `accuracy`→`defect_accuracy`, `refactoring`→`refactoring_plans`. Note that `refactoring_plans` is the raw-per-detector list and is now **opt-in**: `include=["refactoring"]` leads with-`refactoring_opportunities`, the composed unit, and emitting both would ship two-representations of the same work in one response. The `include` **dimension** names (`performance`, `defect`, `maintainability`, `advisory`) do not — they filter rows inside several blocks and have no single key to resolve to, so they land in `unknown_only_keys`. Nor does `signals`, which merges into `metrics[].signals` — in targeted mode, where `signals` applies, name `metrics` instead. |-| `repo` | string | No | *(workspace only)* Target repo alias |-| `limit` | int | No | Max rows in **every** ranked list (default 20, capped at 50). `0` means no rows; the `*_total` siblings still report the true counts. |-| `fix_id` | string | No | Resolve one `fix_first` item `id` (`fix1_...`) to the full item: action steps, gain, effort, risk, verify, context and the next call. Answers `mode: "fix_item"`, with `found: false` and `item: null` when the id is no longer in the queue. |-| `finding_id` | string | No | Resolve an emitted stable health-finding `id` directly in one call. |-| `plan_id` | string | No | Resolve an emitted stable refactoring-plan `id` directly in one call. |-| `opportunity_id` | string | No | Resolve one opportunity `id` directly. The prefix picks the pillar: `perf...` is a performance cause, `refop...` a composed refactoring. Mutually exclusive with `fix_id` and the two above; passing more than one returns `mode: "conflict"` naming them, never answering about whichever was checked first. |-| `refactoring_type` / `refactoring_confidence` / `refactoring_effort` | string | No | Queue filters over the same read model and vocabulary the REST route uses. An unrecognized value is reported back in `ignored_arguments` rather than silently narrowing to nothing. |-| `refactoring_scope` | string | No | Which open opportunities `refactoring_opportunities` lists. `fix_first` (the default with no `targets`) keeps only what Fix first would take, by the same eligibility and exclusion rules; `refactoring_opportunities_hidden` then counts the rest of the filtered set, `{"total": n, "by_reason": {...}}`, by Fix first's exclusion reasons. `all` lists the full inventory, and is the default when `targets` name files. `refactoring_opportunities_scope` says which scope the total counts. |-| `refactoring_view` | string | No | Named ordering for `refactoring_opportunities`. `diversified` (default) round-robins the rank order over cause, refactoring type and area, because the ranked head is a genuine run of ties; `canonical` is the published rank order verbatim, ties and all; `file_spread` asked for one row per file, which a composed opportunity satisfies by construction, so it resolves onto the diversified order. Both older values keep working. It also selects the legacy `refactoring_plans` list's view, where `diversified` resolves to that list's historical `canonical` default. |-| `cursor` | int | No | Zero-based offset into a ranked collection; the `recovery` block names the exact next call. |-| `performance_view` | string | No | `detail` (default) or `summary`. `summary` keeps identity, counts and plan state and drops the explanatory fields. |-| `performance_context` | string | No | `production` (default) / `tooling` / `test` / `unknown` / `all`. The summary block is scoped to the same context as the queue; `repository_total` stays the count over every context. |-| `performance_boundary` | string | No | `db` / `network` / `filesystem` / `subprocess` / `lock` / `none`. |-| `performance_confidence` | string | No | Evidence confidence: `high` / `medium` / `low`. Fix safety and actionability are separate facets. |-| `performance_actionability` | string | No | `plan_ready` / `advisory` / `investigate` / `expected`. Unset means all but `expected` (real repetition with nothing to change), which is still counted in the facet and `repository_total`. |-| `performance_sort` | string | No | `rank` (default) / `leverage` / `observations`. |-| `scope` | string | No | Which files every figure describes: `all` (default) or `production`. Narrowing drops test files from the headline, the distribution and every ranked list. Tests score higher than production code, so `production` lowers the number without a defect having been found. |-| `counts` | string | No | What the score counts: `everything` (default, the calibrated number) or `code_shape`, which removes the git-derived half. Change history rises as a file is worked on, so it answers what a repository has been through rather than what its code is like — `code_shape` is the reading that answers "is this code getting better". Files with no stored split are reported in `unscored_files` rather than counted. Findings from history are dropped, not re-scored. |--An unrecognized `scope` or `counts` falls back to the default and is named in-`ignored_arguments`, so a misspelling never answers a different question under-the name you asked for. Both are echoed on the response.--**Returns:** Dashboard mode (no `targets`) returns `fix_first`, repo-level KPIs-(hotspot health, average health, worst performer, maintainability / performance-pillar averages), the lowest-scoring files, and a per-module NLOC-weighted-rollup. Targeted mode returns per-file marker findings with severity,-per-dimension scores, and the score breakdown. Each finding carries a `dimension`-(`defect` / `maintainability` / `performance` / `advisory`). Only the first three-score; `advisory` describes and never deducts, which is why a file can carry an-advisory finding and a deduction of zero.--**Lead with `fix_first`.** One ranked queue across refactoring, performance and-code-shape findings, built once in core from stored rows; REST-(`/api/repos/{id}/health/fix-first`), `repowise health` and the generated-CLAUDE.md render the same object. The block carries `lead` (equal to-`items[0]`), up to five `items` in a compact projection (`id`, `tier`, `kind`,-`title`, `target`, `why`, `gain`, `effort`, `confidence`, `next_call`),-`totals` (`candidates`, `eligible`, `shown`, and `excluded` counted by reason:-`test`, `tooling`, `unknown`, `generated`, `expected`, `no_strategy`, `no_plan`,-`below_min_worth`, `history_only`, `vendored`, `docs_example`, `deprecated`,-`inherent_dispatch`, `small_function`, `no_concrete_step`, `low_value_kind`; a-performance cause enters only from the performance default queue, whose reasons-`unknown`, `expected` and `no_strategy` are), `by_improves`, `model_version`, `basis`, and `detail_call`,-the `get_health(fix_id=...)` call that returns the lead in full. `tier` is-`now` (worth doing, and the plan is safe to start), `next` (worth doing, needs-judgment) or `later`. Items are ordered by value first, the larger of the recoverable health and the size of the problem (CCN, lines and nesting past their bars, a critical or brain-method finding, a hot file), then by tier, so a function far past every bar leads as "Break up ..." ahead of tidy work that is safer to start. A performance fix's value is its cost: a production, entry-reachable database or network call in a loop that grows with the data shares the top band with the largest functions, and a loop of unknown size that no entry point reaches drops one step. A cause whose marker has not cleared the bar for leading is listed, never first. `kind` is `refactor` (one file's composed refactoring),-`perf_fix` (one intervention, its sinks listed under it) or `finding` (a file's-strongest code-shape finding with no plan). Always the production population,-whatever `scope` says. `limit` caps the items, never which one leads.--**History is context, not a task.** History markers (churn, ownership,-co-change, prior fixes, governance) never create an item or lead one: they ride-on the item for the same file under `context`, and a file whose only findings-are history is counted in `totals.excluded.history_only`.--**Nothing is dropped silently.** Any `targets` entry that matched nothing is-named in `unresolved` with a reason (`not_indexed` → run `repowise update`,-`no_such_path`, `excluded`, `not_measured` → indexed, but carrying no stored-split for the reading `counts` asked for, `no_such_module`; a missed module name also returns-`known_modules`). Missing stored analysis is explicitly unavailable rather than-fabricated as a healthy score. A-target set that resolves to nothing still answers in targeted mode rather than-falling back to the repo dashboard. Every capped list carries a `*_total`-sibling — including under `only`, which retains it automatically. `unresolved`-and `known_modules` survive any `only` projection too, for the same reason-`mode` does: a caller who has to ask for the error report in order to see it-does not have an error report.--`_meta.health_analysis` explicitly labels the result as stored analysis,-states that the call did not recompute it, distinguishes index/live-Git facts-from source-byte verification, and gives the exact commit-then-update refresh-precondition. Its `status` is `available`, `provenance_unknown` (metrics exist-but no row recorded the commit they were computed against, with `reason:-"analysis_commit_not_recorded"`), or `unavailable` (no stored analysis at all).-`provenance_unknown` is a gap in attribution, not a failed analysis. `_meta.health_analyzed_at` dates the health pass, which is separate from-indexing and can lag it, and `_meta.health_analyzed_commit` says which commit-those scores were computed against. The incremental update path rescores only-the files that changed, so the metrics table can hold rows from several passes-at once; when it does, `_meta.health_analyzed_commits_distinct` says how many-and the reported commit is the newest pass's. Both fields are omitted rather-than guessed when no row records a commit.--**The response is bounded.** `include` only *adds* blocks, and the dashboard's-five ranked lists compose: `include=['refactoring']` on a mid-size repo lands-near the host's tool-result cap, past which the host rejects the whole result-and you get nothing. Pair `include` with `only` —-`get_health(include=['refactoring'], only=['refactoring_plans'])` keeps one-block. Anything that would still overflow is shed in the order this tool declares,-never silently: the response carries `truncated: true`, the `*_total` /-`*_emitted` / `*_reduced_reason` siblings describe what was there, and-`_meta.omitted` names refs that restore the dropped rows. Re-requesting one-block with `only` also recovers it.--**Finding types must earn their place.** `core/analysis/finding_registry.py`-records each finding type as `validated`, `provisional` or `hidden`. Hidden-types never appear in `get_health`, `get_dead_code`, overview, priorities or-wiki prompts (the analyzers still run). Provisional types appear only when asked-for — `include=["unverified"]` here, `kind=` on `get_dead_code` — and each row-carries `verification: "unverified"`. `get_dead_code` reports what it held back-in `summary.withheld_types`.--**Test material is bucketed, not hidden.** Every metric row carries `is_test`-(distinct from `has_test_file`: "is this file a test" vs "is this file tested").-In dashboard mode the ranked finding lists are split — `top_findings` /-`findings` carry production findings, `test_findings` carries the test half, and-`top_findings_total + test_findings_total` is the whole open set. Defect risk in-a test asks a different question from defect risk in the code it covers, and at-the default limit a quarter of the headline list was describing the test suite.-Targeted mode is never split: you named the files, so you get their findings.-The ranked file lists split the same way: `worst_files` and-`high_leverage_files` are production files,-and `test_worst_files` ranks the test files. `kpis.worst_performer_path` is the-worst production file and `kpis.worst_test_path` the worst test file. The KPI-averages still include test files — excluding them would move the repo's-headline score, which is a scoring change, not a display one.--**Leverage, not just lowness.** `average_health` is NLOC-weighted (the number the-badge and dashboard surface), so a few large low-scoring files hold it down. To-make that actionable rather than a mystery:--- `kpis.average_health_unweighted` is the plain file mean and- `kpis.average_health_weighting` is `"nloc"`. When the weighted and unweighted- numbers diverge, the gap is telling you to chase *big* files, not the long tail.-- `gap_analysis` (dashboard mode) reports the net weighted points the average must- recover to reach the target score (8.0), how many files sit below it, and how- few of them carry the whole gap (`files_to_reach_target`) or half of it- (`files_for_half_gap`). This reframes a repo-wide number as a short worklist.-- Every metric row carries `weighted_deficit = (8 - score) x nloc`: how much the- repo headline recovers if that file reaches 8.0. `high_leverage_files`- (dashboard mode) is the top-N ranked by it, distinct from `worst_files`, which- sorts by raw score and ranks a 30-line file at 1.0 equal to a 1,200-line file at- 1.0 that moves the average ~40x more.-- `weighted_deficit` and `gap_analysis.weighted_gap_points` share one unit,- *score-points x NLOC*, which compares against itself and nothing else. Every- `high_leverage_files` row also carries `share_of_repo_gap_pct`, the same quantity- over the gross deficit of all below-target files- (`gap_analysis.weighted_gross_gap_points`), so the shares are bounded by 100%- and sum to 100% by construction; that plus- `gap_analysis.files_to_reach_target` is what answers "is this worth doing".-- `_meta.health_semantics` (with `include=["semantics"]`) gives the numerator, gross-deficit denominator,- nonnegative unbounded scale and direction. These are deterministic heuristic- triage points, not probabilities, normalized score points, percentages or- guaranteed improvement.-- `kpis.non_code_files` and `kpis.average_health_code_only` say how much of the- headline is markdown/JSON/YAML. No biomarker walks those files, so they score- a mechanical 10.0 meaning "nothing looked at this" — on this repo, 233 of- 3,314 rows, lifting `average_health` from 7.31 to 7.47. `average_health`- itself deliberately still counts them, so the tool, the badge, the snapshots- and the web UI all report the same number.--**One score, not two.** A metric row carries `score` (the defect dimension and-the headline), `maintainability_score` and `performance_score`. There is no-`defect_score` in the response: it was set from the same value as `score` on-every row, and two names for one number cost a reader a source dive to pick-between them. The field to rank on is neither — it is `weighted_deficit`.--**`primary_biomarker` names a discrete cause.** It prefers the strongest-*discrete* finding over a continuous one. `coverage_gradient` fires on every-file that has coverage data at all, so on a well-covered repo it used to win the-max-impact tiebreak nearly everywhere and headline the list with "N% of lines-uncovered" — true, and equally true of every other file. The gradient still-counts in full toward `total_deduction` and the score, and still leads a file-that has no discrete finding.--The opt-in enrichments:--- **`accuracy`** returns a `defect_accuracy` block: of the K least-healthy files, how- many were recently bug-fixed vs the repo-wide base rate (precision@K + `lift`),- with a per-K table and the flagged files. Silent (`null`) on repos with too- little history to be honest (< 25 scored files or < 5 recently-fixed files).-- **`signals`** adds a `signals` object on each targeted metric: prior-defect count,- change scatter, 90-day churn, primary / recent owner, and graph in / out- degree. Honest `null` per field when the underlying row is absent (never an- imputed zero).-- **`doc_drift`** returns a `doc_drift` block: `findings` (each naming the- **document** to edit, its line, the `target` it wrongly claims exists, a- `reason` sentence, `kind`, `origin` and `confidence`, plus `suggestion` and- `suggestion_basis` only when a likely replacement was found), plus `findings_total`,- `documents` and the high/medium/low `confidence` split. Its `basis` field is- load-bearing: the detector checks only references it can resolve, most- references in a typical repository are uncheckable by design, and a finding is- evidence to check rather than a proven defect. Every document is re-checked- against the live tree on every `init` and every `update`, so a finding does- not outlive the sentence that caused it; the exception is a document that- could not be read on a given run, whose existing rows are left alone rather- than deleted. An index written before findings were stored reports- `{"unavailable": "index_predates_doc_drift"}` rather than an empty list. `repowise doc-drift` serves the same rows in a- terminal, with the evidence lines this block leaves out.-- **`churn_complexity`** returns `churn_complexity` points (one per recently-changed- file: 90-day commit count, max CCN, NLOC, score, churn percentile): the- refactor zone where volatility and tangle collide.-- **`refactoring`** returns `refactoring_opportunities`: one composed unit per- file, carrying the diagnosis it leads with (`lead_biomarker`), whether it- actually addresses that diagnosis (`addresses_primary_problem`, tri-state -- `null` means no dominant finding was recorded, which is not `false`), its- ordered `steps` with a `mechanical` / `judgment` `applicability` each, and- counts for the evidence behind it. Ordered by `refactoring_view`. With no- `targets` it lists only what Fix first would take (`refactoring_scope`), and- `refactoring_opportunities_hidden` counts the rest by reason. A step- carrying `relocated_by` names an earlier step that moves its symbol to another- file: locate the symbol again before applying it, because the step's own- `file_path` and span describe where the symbol was.- `get_health(opportunity_id="refop...")` returns the full ordered steps, the- member plan payloads, the validation profile and structured `next_actions`;- `only=["refactoring_evidence"]` plus `cursor` pages the evidence. `found`- says whether the id named a stored opportunity; `status` is its triage- lifecycle (`open`, `acknowledged`, `resolved`, `false_positive`).-- A refactoring reaches a bare `get_health()` through `fix_first`, whose- `next_call` is the exact `opportunity_id` call that opens it. **`refactoring_summary`**- (`only=["refactoring_summary"]`) is the rollup by type, effort, confidence,- lifecycle and mechanical-vs-judgment, with facets, and a `next_call` that- lists the opportunities.-- Every next call (Fix first's `next_call`, the refactoring and performance- summaries' `next_call`, actions' `commands`, `get_risk`'s- `directive.next_calls`) has one shape: `purpose`, `mcp` (the call as text,- e.g. `get_health(opportunity_id="refop3_...")`), `cli` (the terminal line- that gives the same answer, or null), and `tool` + `arguments`, the same call- structured.-- **`refactoring_plans`** is the raw per-detector list, unchanged and still- addressable by `plan_id`, but **opt-in**: name it in `only` to get it. It- returns ranked, structured refactoring plans (not template- strings): `extract_class` (the cohesion `groups` to split into), `extract_helper`- (clone `occurrences` + `suggested_site`), `move_method` (`{method, from_class,- to_class}`), and `break_cycle` (the import `cut_edges`). Each plan carries its- `evidence`, `impact_delta`, `effort_bucket`, `blast_radius`, and an `id` you can- hand to `generate_refactoring_code`. The list is capped to `limit` and ranked- file-leverage-first (by the file's `weighted_deficit`, then per-plan impact), so- plans on the files that move the headline surface first; `refactoring_plans_total`- reports the full count behind the cap. Each plan echoes its- `file_weighted_deficit`. Full shapes in [`docs/layers/REFACTORING.md`](../layers/REFACTORING.md).-- A requested empty plan list includes `refactoring_plans_status.reason`:- `no_applicable_findings`, `plan_analysis_indeterminate`,- `no_eligible_targets`, or `analysis_unavailable`. The structured-analysis- fallback includes a concrete `get_symbol` or `get_context` source call and- explicitly names the two facts the stored data cannot distinguish: no- supported transformation vs a disabled or failed detector.- Projection exclusion emits no plan warning; a zero-row cursor window is- reported separately as `request_window_empty`.-- **dimension filter** narrows the returned findings to one pillar, e.g.- `include=["biomarkers", "performance"]`.--**Performance findings rank on `perf_rank`, not on `health_impact`.** Every-performance finding carries `health_impact: 0` — the pillar is deliberately-never blended into the score — so ranking them by impact ordered them by nothing-and the cap kept whichever the tie broke to. Each performance row now carries an-integer `perf_rank` (absent on defect and maintainability rows, which rank on-`weighted_deficit`), and the returned list is ordered by it *within* each impact-tier, so the defect ordering is untouched. It is an ordering key, not a score:-nothing is blended into `score` / `performance_score`. It adds three signals the-row already carries, so you can re-rank on your own weights from the same-payload:--| signal | reads | why |-|---|---|---|-| the marker | `biomarker_type` | superlinear (`nested_loop_quadratic`, `nested_loop_with_io`, `sql_cartesian_join` 6) > lock-serialized (5) > one crossing per iteration, or one proven on a hot path (4) > in-loop CPU/allocation (3) > repeated acquisition with no boundary (2); an unweighted marker takes the floor (1) |-| the boundary | `details.boundary_kind` | `subprocess` 5 > `network`/`db` 4 > `lock` 3 > `filesystem` 2. A process spawn in a loop is not a stat in a loop |-| the call shape | `details.cross_function` | +1. An intra-function loop is usually visibly bounded at the call site; a cross-function N+1 is the one nobody sees by reading the loop |--These are the same weights the causal opportunity ranking reads. Two tables-used to answer "which marker costs more" and had drifted apart on markers both-named, so a finding and the opportunity built from it could disagree about the-same evidence.--Centrality is read off the marker rather than a column:-`hot_path_sync_io` and `nested_loop_quadratic` are only ever emitted for a-function the perf ranker called hot (top-quintile call-graph in-degree), so-their presence is already the proof. It is not proof that a request reaches-the function, and neither marker claims that. `hot_path_sync_io` is never-emitted in tests, tooling, examples, or generated or vendored code. Deliberately not-`severity` — that column grades `hot_path_sync_io` below `io_in_loop` and takes-only two values across a whole repo's perf findings.-- **`refactoring`** also emits `suggestion_legend`: `biomarker_type` → the prose- suggestion for that type, once per response rather than per finding. Join on- `biomarker_type`. It is keyed off the ranked finding head and does not vary- with `only`, so it can carry an entry for a block a projection dropped —- extra rows in a lookup table, never a missing one. Note it explains the- **findings**, not the- plans it ships beside: the two sets differ (no plan kind is sourced from- `coverage_gradient`).--**When to use:** Before opening a PR, to self-check the files you changed-(`targets=[...], include=["signals"]`) and confirm you are not regressing the-worst files. Before refactoring, find the worst-scoring files and what to fix-first (`include=["accuracy", "churn_complexity"]`). Pair with `get_risk` on-hotspots.--**Example calls:**+### `get_dead_code`-```-get_health(only=["fix_first"]) # cheapest useful call: what to fix first-get_health(fix_id="fix1_...") # one item in full: steps, verify, risk-get_health() # fix_first, kpis, gap_analysis, high_leverage files-get_health(include=["accuracy", "churn_complexity"])-get_health(include=["biomarkers", "performance"]) # only performance findings-get_health(targets=["src/api/server.py"], include=["signals"])-get_health(targets=["module:src.api"], include=["trend", "refactoring"])-get_health(include=["accuracy"], only=["accuracy"]) # the block, without the dashboard again-get_health(only=["top_findings"]) # + top_findings_total, automatically-get_health(only=["kpis"], limit=0) # headline numbers, no rows at all-```+Unreachable files, unused exports, unused internals and zombie packages, tiered by confidence. Use it for a cleanup pass, not a targeted fix. `safe_only` leaves out anything with runtime-load risk such as dynamic imports and framework-registered functions.++| Parameter | Type | Default | Meaning |+|-----------|------|---------|---------|+| `kind` | string | all | `unreachable_file`, `unused_export`, `unused_internal` or `zombie_package` |+| `min_confidence` | float or string | `0.4` | Confidence floor; also accepts `"high"` (0.7), `"medium"` (0.4), `"low"` (0.0) |+| `tier` | string | all | `high` (0.7 and up), `medium` (0.4 to 0.7) or `low` |+| `safe_only` | bool | `false` | Deletion-ready findings only |+| `limit` | int | `20` | Max findings per tier, clamped to 25 |+| `directory` | string | none | Path-prefix filter |+| `owner` | string | none | Primary-owner filter |+| `group_by` | string | none | `directory` or `owner` rollup |+| `include_internals` | bool | `false` | Also scan private symbols (more false positives) |+| `include_zombie_packages` | bool | `true` | Monorepo package findings |+| `no_unreachable` | bool | `false` | Skip unreachable-file findings |+| `no_unused_exports` | bool | `false` | Skip unused-export findings |+| `finding_id` | string | none | Open one finding by its id |+| `repo` | string | default repo | Workspace repo alias, or `"all"` to aggregate across repos |-### Coverage: the stored report and how far to trust it--`include=["coverage"]` returns the stored per-file rows and a repo-wide-`summary`. Every row carries `covered_line_count` beside-`total_coverable_lines`; targeted mode adds the `covered_lines` array. The-summary's `freshness.status` is `current` when the report was measured at the-indexed commit, `stale` when at another one (its line numbers may describe-code that has moved), and `unknown` when either commit is missing.-`report_paths` says how the report's own file entries mapped at ingest-(`total`, `matched`, `unmatched`, `ambiguous`, and a short `unmatched_sample`);-it is null for coverage stored before that record existed. `source_formats`-lists every report format merged.--In dashboard mode the block also carries `history`, one point per complete-(not partial) ingest (`ingested_at`, `ingested_commit_sha`, `line_coverage_pct`,-`branch_coverage_pct`), oldest first, newest 10, with `history_total`,-`history_emitted` and `history_reduced_reason: "limit"` when cut.--In targeted mode each row also carries a `decay` block: how many of the-report's covered lines are unchanged since it ran (`confirmed_lines`) and how-many have moved since (`invalidated_lines`, now unknown rather than uncovered).-`decay.drifted` is true once a fifth of a file's measurement has moved. It is a-per-file statement about lines, separate from the summary's commit-level-`freshness`.--### Performance: one lead, then drill down--A bare `get_health()` ranks performance work inside `fix_first`: one `perf_fix`-item per intervention (the loop or function you change), with the sinks it-covers under it, and a `next_call` that opens the opportunity. An opportunity-by id adds `plan_steps`, `validation` (tests and commands) and `siblings`: other-causes observed on the same lines. A rejected filter value is echoed with its-accepted values in `ignored_arguments`. Performance findings carry-`health_impact: 0` by construction, so they never rank in the impact-ordered-finding lists. Rows with no safe plan, or whose repetition is expected, stay-out of `fix_first` and are counted in its `totals.excluded`.+**Key return fields:** `tiers` (findings per tier, each with path, kind, confidence, line count and cleanup impact), `summary` (totals over every open finding, `filtered_findings`, `filters`, `withheld_types`, `call_resolution_basis`), `by_directory` or by-owner rollups with `group_by`. In workspace mode, confidence drops on findings another repo still imports.```-get_health(only=["fix_first"]) # the lead-get_health(include=["performance"], only=["performance_summary"])-get_health(include=["performance"], only=["performance_opportunities"], performance_context="all")-get_health(opportunity_id="perf3_...") # the cause, its plan, its evidence-get_health(opportunity_id="perf3_...", only=["performance_evidence"], cursor=3)+get_dead_code()+get_dead_code(tier="high", safe_only=true)+get_dead_code(kind="unused_export", group_by="owner")```-Ids are stable within a performance model version and are never translated-across one, because grouping decides membership and two models disagree about-it. An id from an older model resolves to `model_state.state: "stale_model"`-with `refresh_required`, rather than failing to match and reading as "no plan".-`found` says whether the id named a stored opportunity; `lifecycle_status` is-its lifecycle (`open` or `resolved`). Evidence rows carry the finding's public `finding_id`, which round-trips through-the `finding_id` selector; storage row ids are republished on every analysis and-are never emitted.-----## Supplementary tools--These are registered and on by default (in the modes noted) but are not part-of the ten-tool headline set.+## Workspace default### `list_repos`-Lists the repos this server is serving. No parameters.+Lists the repos this server serves. On by default in workspace mode only; a single-repo server is bound to its one repository. No parameters.-**Returns:** In workspace mode, `workspace: true`, the workspace root, the default repo alias, and every configured repo's `alias`, config-relative `path`, and `absolute_path`. Any of those emitted identities can be passed unchanged as `repo` to workspace-aware tools. In single-repo mode, `workspace: false` and a single `"default"` alias.--**When to use:** Discovering the `repo` aliases to pass to other tools, especially in workspace mode.+**Key return fields:** `workspace` (bool), `workspace_root`, `default_repo`, and per repo its `alias`, config-relative `path` and `absolute_path`. Any of those can be passed unchanged as `repo` to other tools.```list_repos()```-### Workspace-only tools--*(Available only when the server is started inside a workspace; see [Workspace Mode](#workspace-mode).)*--#### `get_blast_radius`+----Cross-repo structural and historical reach from a changed service.+## Opt-in tools-| Parameter | Type | Required | Description |-|-----------|------|----------|-------------|-| `targets` | list[string] | Yes | Node ids (`repo` or `repo::service/path`) or repo aliases |-| `max_depth` | int | No | Reachability depth (1-8, default 3) |-| `include_behavioral` | bool | No | Include co-change (behavioral) edges (default `true`) |+Enable with `mcp.tools: ["+name"]` or `repowise mcp --tools "+name"` (see [Configuring the tool surface](#configuring-the-tool-surface)). The last three are workspace-only.-**Returns:** The impacted services ranked by uncalibrated `score` (0-1 relative-path weight, not a breakage probability), each with `distance` (hops),-`structural` (a real dependency vs co-change only), and the edge kinds that-carried the impact; plus `impact_score_semantics`, `impacted_repos`,-`structural_count` / `behavioral_count`, `total_impacted`, and unresolved targets.+### `get_dependency_path`-Each `symbol_targets[].consumers[]` row also carries a `tests` block of the same-shape as `get_change_risk`'s: which tests in that consumer repo guard the-symbol's contracts, capped at five with the tail in the omission store, and a-named reason when a link could not be followed.+The shortest dependency path between two files or modules. When none exists it returns context to debug the gap: nearest common ancestors, shared neighbours, community analysis and bridge suggestions.-**When to use:** Before changing a high-fan-out provider, see who structurally-consumes it across repo boundaries. Structural reach outweighs historical-co-change in ranking, but neither is a runtime-breakage claim. Reads the same-system graph the [Live System Map](../scale/WORKSPACES.md#live-system-map) renders.+| Parameter | Type | Default | Meaning |+|-----------|------|---------|---------|+| `source` | string | required | Source file or module path |+| `target` | string | required | Target file or module path |+| `repo` | string | default repo | Workspace repo alias. `"all"` is not supported |```-get_blast_radius(targets=["backend"])-get_blast_radius(targets=["mono::services/auth"], max_depth=2, include_behavioral=false)+get_dependency_path(source="src/api/routes.py", target="src/db/models.py")```-#### `get_conformance`--Architecture governance: does the live system graph obey the declared dependency rules, and are there circular service dependencies?+### `get_execution_flows`-**Opt-in.** Off by default even in workspace mode; enable with `mcp.tools: ["+get_conformance"]`. Named in single-repo mode it is ignored, since it needs the workspace graph. The same findings still surface in the `get_risk` PR-mode directive (`conformance_violations` / `dependency_cycles`) without opting the tool in.+The top-scored entry points and a breadth-first call trace from each, marking where a flow crosses community boundaries. Use it to see what runs end to end.-| Parameter | Type | Required | Description |-|-----------|------|----------|-------------|-| `repo` | string | No | Limit findings to those involving this repo alias |--**Returns:** `violations` (each with the offending `source`/`target` services, the `rule_source`/`rule_target` matchers that fired, and the `edge_kind`), `cycles` (each with the participating `nodes` and `length`), and the `violation_count` / `cycle_count` / `rules_evaluated` rollups.--**When to use:** Before a refactor that changes service boundaries, or to audit whether the live architecture still matches the intended one. Rules are declared under `conformance:` in `.repowise-workspace.yaml`. See [Architecture Conformance](../scale/WORKSPACES.md#architecture-conformance).+| Parameter | Type | Default | Meaning |+|-----------|------|---------|---------|+| `top_n` | int | `10` | Entry points to trace |+| `max_depth` | int | `8` | Max trace depth per flow |+| `entry_point` | string | none | Trace from this symbol, overriding `top_n` scoring |+| `repo` | string | default repo | Workspace repo alias. `"all"` is not supported |```-get_conformance()-get_conformance(repo="frontend")+get_execution_flows(entry_point="src/cli/main.py::main", max_depth=4)```-#### `get_architecture`+### `generate_refactoring_code`-The one evaluative read of the whole system: how coupled is it, where is the architectural core, and a single 1-10 architecture score. Deterministic, structural edges only (co-change excluded). No parameters.+Turns one refactoring plan from `get_health(include=["refactoring"])` into generated code and a unified diff, grounded on the plan and the source spans it names. Extract Class results include an LCOM4 before-and-after check. It calls the repo's configured LLM provider, and caches by content hash so an unchanged plan never regenerates.-**Returns:** `score` (1-10), `architecture_type` (`core-periphery` or `hierarchical`), `propagation_cost_pct` (share of other services the average service reaches), `core_size` / `core_ratio` / `core_members` (the largest cyclic group), `cycle_count`, `conformance_violations`, a `role_breakdown` (count of Core / Shared / Control / Peripheral services), and a one-line `summary`.+Adding the tool to the surface is the opt-in step. Generation then runs unless `.repowise/config.yaml` sets `refactoring.llm.enabled: false`, in which case the plan is returned with `generation.available: false` and `reason: "disabled"`. With no provider configured it returns `error: "no_provider"`.-**When to use:** Before a cross-service refactor, or to gauge and compare overall system structure over time. See [Architecture Metrics](../scale/WORKSPACES.md#architecture-metrics).+| Parameter | Type | Default | Meaning |+|-----------|------|---------|---------|+| `suggestion_id` | string | required | A plan `id` from `get_health` |+| `repo` | string | default repo | Workspace repo alias. `"all"` is not supported |```-get_architecture()+generate_refactoring_code(suggestion_id="a1b2c3d4")```-### Opt-in tools--*(Registered but off by default in every mode; enable with `mcp.tools: ["+name"]` or `repowise mcp --tools "+name"`. See [Configuring the tool surface](#configuring-the-tool-surface).)*--#### `get_dependency_path`--Shortest dependency path between two files or modules.+### `set_finding_status`-| Parameter | Type | Required | Description |-|-----------|------|----------|-------------|-| `source` | string | Yes | Source file or module path |-| `target` | string | Yes | Target file or module path |-| `repo` | string | No | *(workspace only)* Target repo alias; `"all"` is not supported |+Records a durable verdict on one refactoring plan in the index, so the triage survives later analysis runs.-**Returns:** The dependency path when one exists. When no direct path exists, visual context instead: nearest common ancestors, shared neighbors, community analysis, and bridge suggestions, to help debug architectural silos.+| Parameter | Type | Default | Meaning |+|-----------|------|---------|---------|+| `suggestion_id` | string | required | A plan `id` or `public_id` from `get_health` |+| `status` | string | required | `open`, `acknowledged`, `resolved` or `false_positive` |+| `reason` | string | `"agent"` | Free-text audit note stored on the row |+| `repo` | string | default repo | Workspace repo alias. `"all"` is not supported |-**When to use:** Understanding how two parts of the codebase are (or aren't) connected, or why an expected dependency doesn't show up.+`false_positive` plans are never re-emitted. `acknowledged` stays visible but stops counting as unheard. `resolved` stays resolved even if the detector still fires. `open` resets. Returns the new status and its timestamp.```-get_dependency_path(source="src/api/routes.py", target="src/db/models.py")+set_finding_status(suggestion_id="a1b2c3d4", status="false_positive", reason="the class is a DTO")```-#### `get_execution_flows`--Top entry points and their call traces: how the codebase actually executes.+### `get_blast_radius`-| Parameter | Type | Required | Description |-|-----------|------|----------|-------------|-| `top_n` | int | No | Number of top entry points to trace (default 10) |-| `max_depth` | int | No | Max trace depth per flow (default 8) |-| `entry_point` | string | No | Trace from a specific symbol, overriding `top_n` scoring |-| `repo` | string | No | *(workspace only)* Target repo alias; `"all"` is not supported |+*Workspace only.* Which services in other repos are structurally reachable from a changed service, ranked by an uncalibrated 0 to 1 path weight. Structural edges (HTTP, gRPC, events, packages) outrank co-change. It is not a runtime-breakage probability.-**Returns:** Scored entry points with BFS call-path traces showing which functions are called in sequence, and whether the flow crosses community boundaries.+| Parameter | Type | Default | Meaning |+|-----------|------|---------|---------|+| `targets` | list[string] | required | Node ids (`repo` or `repo::service/path`), repo aliases, or the symbol id of a published symbol |+| `max_depth` | int | `3` | Reachability depth, 1 to 8 |+| `include_behavioral` | bool | `true` | Include co-change edges |-**When to use:** Understanding runtime call flow through an unfamiliar system, or tracing what a specific entry point actually does end to end.+**Key return fields:** impacted services with `score`, `distance` and edge kinds, `impacted_repos`, `total_impacted`, `impact_score_semantics`, and `symbol_targets` naming the consuming symbols across each contract link, with the tests that guard them.```-get_execution_flows()-get_execution_flows(entry_point="src/cli/main.py::main", max_depth=4)+get_blast_radius(targets=["backend"])```-#### `generate_refactoring_code`+### `get_architecture`-Turns one structured refactoring plan from `get_health(include=["refactoring"])` into actual generated code and a unified diff, grounded on the plan plus the real source spans it references. For Extract Class, the result includes an LCOM4 before/after self-check.+*Workspace only.* Whole-system structure: propagation cost (the share of other services the average service reaches), the largest cyclic core, service roles, and a deterministic 1 to 10 score. Structural edges only. No parameters.-**Opt-in:** it must be added to the tool surface (`mcp.tools: ["+generate_refactoring_code"]`). Generation itself is on unless the repo sets `refactoring.llm.enabled: false` in `.repowise/config.yaml`; then a valid plan id still resolves, returning the canonical plan plus `generation: {available: false, reason: "disabled"}`. Without a configured provider it returns `error: "no_provider"`. When enabled, it uses the repo's configured LLM provider/model (bring your own key) and caches results by a content hash, so an unchanged plan never regenerates.--**Parameters:**--| Parameter | Type | Required | Description |-|-----------|------|----------|-------------|-| `suggestion_id` | string | Yes | The `id` of a plan returned by `get_health(include=["refactoring"])` |-| `repo` | string | No | *(workspace only)* Target repo alias |--**When to use:** After `get_health(include=["refactoring"])` surfaces a plan you want turned into an applyable diff, with the tool opted in and a provider configured.+**Key return fields:** `score`, `architecture_type`, `propagation_cost_pct`, `core_members`, `cycle_count`, `role_breakdown`, `summary`.```-generate_refactoring_code(suggestion_id="a1b2c3d4")+get_architecture()```-#### `set_finding_status`+### `get_conformance`-Records a durable disposition on one refactoring plan — the write half of the findings triage loop. `get_health(include=["refactoring"])` and the generated task prompts ask an agent to flag false positives, but until this tool there was nowhere to record that verdict from inside the agent loop. The status is stored on the plan row, and the analyzer's finalizer never re-emits a `false_positive` plan, so the triage survives every later run.+*Workspace only.* Cross-repo dependencies that break the rules declared under `conformance:` in `.repowise-workspace.yaml`, plus circular service dependencies. Without opting in, the same findings still reach `get_risk`'s PR-mode directive as `conformance_violations` and `dependency_cycles`.-| Parameter | Type | Required | Description |-|-----------|------|----------|-------------|-| `suggestion_id` | string | Yes | The `id` or `public_id` of a plan from `get_health(include=["refactoring"])` |-| `status` | string | Yes | One of `open`, `acknowledged`, `resolved`, `false_positive` |-| `repo` | string | No | *(workspace only)* Target repo alias |-| `reason` | string | No | Free-text audit note stored on the row (defaults to `"agent"`) |+| Parameter | Type | Default | Meaning |+|-----------|------|---------|---------|+| `repo` | string | whole workspace | Limit findings to those involving this repo |-- `false_positive` — the plan is wrong for this repository; it is never re-emitted on future runs.-- `acknowledged` — real, but the team is consciously not acting now; stays visible, stops counting as unheard.-- `resolved` — the change landed (or the code moved on); a person-resolved plan stays resolved even if the detector still fires.-- `open` — reset a prior decision.--**When to use:** After `get_health(include=["refactoring"])` surfaces a plan you have judged, so the verdict becomes durable state instead of a one-off remark.+**Key return fields:** `violations` (source and target services, the rule that fired, `edge_kind`), `cycles`, `violation_count`, `cycle_count`, `rules_evaluated`.```-set_finding_status(suggestion_id="a1b2c3d4", status="false_positive", reason="false alarm: the class is a DTO")+get_conformance(repo="frontend")```----## Workspace Mode+## Workspace mode-In workspace mode (initialized with `repowise init .`), all tools accept an optional `repo` parameter:+A server started inside a [workspace](../scale/WORKSPACES.md) serves every repo in it from one process. `repowise mcp --no-workspace` forces single-repo mode for a nested repo. Every tool with a `repo` parameter takes an alias, path or absolute path from `list_repos`; omitting it targets the default repo.-- **Omit `repo`**: queries the default (primary) repo-- **`repo="backend"`**: targets a specific repo by any `alias`, `path`, or `absolute_path` emitted by `list_repos`-- **`repo="all"`**: queries across all workspace repos (fully supported by `search_codebase`; `get_context` and `get_overview` also accept it; not supported by `get_symbol`, `get_dependency_path`, or `get_execution_flows`)+Only four tools answer across every repo with `repo="all"`. The rest return an error naming the available aliases.-The MCP server automatically enriches responses with cross-repo intelligence:-- **Co-change partners** from other repos surfaced in `get_context` and `get_risk`-- **API contract links** (HTTP, gRPC, topics) between repos-- **Package dependencies** between repos-- **Cross-repo blast radius** via the workspace-only `get_blast_radius` tool, and a cross-repo `directive` in `get_risk` PR-mode-- **Breaking-change guard**: provider incompatibilities and explicit comparison uncertainty, plus endpoint-exposed consumers, in the `get_risk` PR-mode `breaking_changes` directive-- **Architecture conformance**: declared dependency-rule violations and dependency cycles via the workspace-only, opt-in `get_conformance` tool, and `conformance_violations` / `dependency_cycles` in the `get_risk` PR-mode directive-- **Architecture metrics**: whole-system coupling (propagation cost), the cyclic core, per-service roles, and a deterministic 1-10 architecture score via the workspace-only `get_architecture` tool+| Tool | `repo="all"` |+|------|--------------|+| `search_codebase` | Yes: federates the search and merges results |+| `get_overview` | Yes: the cross-repo topology (co-changes, package dependencies, API contracts), with no single-repo detail |+| `get_dead_code` | Yes: aggregates findings across repos |+| `get_why` | Only with a `query`; not for the dashboard or an `id` lookup |+| `get_answer`, `get_context`, `get_symbol`, `get_risk`, `get_change_risk`, `get_health` | No |+| `get_dependency_path`, `get_execution_flows`, `generate_refactoring_code`, `set_finding_status` | No |+| `get_conformance` | No `"all"` value; it covers the whole workspace when `repo` is omitted |+| `get_blast_radius`, `get_architecture`, `list_repos` | No `repo` parameter; they always read the whole workspace |----+Single-repo tools also gain cross-repo evidence in workspace mode:-## Proactive Hooks (Complementary)+- `get_context` adds a per-target `cross_repo` block: co-change partners in other repos and contract consumers or providers.+- `get_risk` in PR mode adds `will_break_consumers` (structural reach only), `missing_cross_repo_cochanges`, `breaking_changes`, `conformance_violations` and `dependency_cycles` to its directive.+- `get_change_risk` adds `cross_repo`: consumers of the contracts the change touches, breaking changes, and the consumer-side tests to run.+- `get_dead_code` lowers confidence on findings another repo still imports.-In addition to the MCP tools above, `repowise init` installs AI-agent hooks (Claude Code and Codex) that provide **passive, automatic** context enrichment:+See [WORKSPACES.md](../scale/WORKSPACES.md) for setup, contracts and the system graph.-- **Claude Code PostToolUse**: broad or zero-result `Grep`/`Glob` calls can be enriched with graph context, git operations can trigger stale-wiki notices, and with `hooks.coverage_reingest: true` (opt-in) a full test run, passing or failing, re-ingests its fresh coverage report in the background ([details](HOOKS.md#what-gets-written-where)).-- **Codex SessionStart**: Codex receives concise repowise MCP workflow guidance when a session starts.-- **Codex PostToolUse**: after edits or git operations, Codex receives a freshness reminder when indexed context may be stale.+----Hooks are lightweight reminders. MCP tools are for deeper, on-demand investigation. See [Auto-Sync](../scale/AUTO_SYNC.md) and [Codex Integration](CODEX.md) for details.+Hooks complement these tools with passive context: see [HOOKS.md](HOOKS.md) and [CODEX.md](CODEX.md).diff --git a/docs/architecture/ARCHITECTURE.md b/docs/architecture/ARCHITECTURE.mdindex 1d269d4fb..0295ce42e 100644--- a/docs/architecture/ARCHITECTURE.md+++ b/docs/architecture/ARCHITECTURE.md@@ -10,14 +10,16 @@ they fit together. Read this before contributing.### Package READMEs-For per-package detail (installation, full API reference, all CLI flags, file maps):+Each Python and web package has a short contributor README that says what the+package is and where its code lives. The detail is in this document and in+[`CLI_REFERENCE.md`](../reference/CLI_REFERENCE.md).| Package | README | What it covers ||---------|--------|----------------|-| `packages/core` | [`packages/core/README.md`](../../packages/core/README.md) | Ingestion, generation, persistence, providers; all key classes with code examples |-| `packages/cli` | [`packages/cli/README.md`](../../packages/cli/README.md) | CLI entrypoints and flags; full surface in [`CLI_REFERENCE.md`](../reference/CLI_REFERENCE.md) |-| `packages/server` | [`packages/server/README.md`](../../packages/server/README.md) | All REST API endpoints, 11 MCP tools, webhook setup, scheduler jobs |-| `packages/web` | [`packages/web/README.md`](../../packages/web/README.md) | Every frontend file with purpose, API client, hooks, components, pages |+| `packages/core` | [`packages/core/README.md`](../../packages/core/README.md) | Ingestion, analysis, generation, persistence and providers |+| `packages/cli` | [`packages/cli/README.md`](../../packages/cli/README.md) | The `repowise` command; full surface in [`CLI_REFERENCE.md`](../reference/CLI_REFERENCE.md) |+| `packages/server` | [`packages/server/README.md`](../../packages/server/README.md) | REST API, the MCP server (18 registered tools, 10 on by default), webhooks, scheduler |+| `packages/web` | [`packages/web/README.md`](../../packages/web/README.md) | The local dashboard and how to run it in development |---@@ -78,7 +80,7 @@ For per-package detail (installation, full API reference, all CLI flags, file ma│ Three Stores │ │ Consumers ││ │ │ ││ SQL (wiki pages, │ │ Web UI MCP Server GitHub Action │-│ jobs, symbols, │ │ (Next.js) (11 tools) (CI/CD) │+│ jobs, symbols, │ │ (Next.js) (18 tools) (CI/CD) ││ versions) │ │ ││ │ │ repowise CLI ││ Vector (LanceDB / │ │ (init, update, watch, │@@ -532,15 +534,15 @@ architecture.**Named binding resolution** (`NamedBinding` dataclass in `ingestion/models.py`) ensuresthat aliased imports, barrel re-exports, and namespace imports resolve to the correct-definition site. The parser's `_extract_import_bindings()` produces bindings for each-import statement, and `GraphBuilder.build()` populates `Import.resolved_file` from them.+definition site. `extract_import_bindings()` in `ingestion/extractors/bindings/` (one extractor+per language) produces bindings for each import statement, and `GraphBuilder.build()` populates `Import.resolved_file` from them.Barrel files (`__init__.py`, `index.ts`) are followed one hop to resolve re-exports.**Two-tier graph isolation:**Symbol nodes and their `DEFINES`/`HAS_METHOD`/`CALLS` edges are stored in the same`DiGraph` as file nodes, but `file_subgraph()` returns a view containing only `file`-and `package` nodes. All file-level metrics (PageRank, betweenness, SCCs, Louvain)+and `package` nodes. All file-level metrics (PageRank, betweenness, SCCs, community detection)run on this subgraph so that the large number of symbol nodes does not distort centralityscores.@@ -552,7 +554,7 @@ After graph construction, the builder computes:Logged as warnings. Require special generation handling (see below).- **Betweenness centrality**: identifies "bridge" symbols whose removal woulddisconnect the graph. These are the most critical to document well.-- **Community detection (Louvain)**: discovers logical modules even when the+- **Community detection (Leiden when the `graph-extra` extra is installed, Louvain otherwise)**: discovers logical modules even when thedirectory structure doesn't reflect them. These communities become module pages.**Circular dependency handling:**@@ -1089,9 +1091,10 @@ and supports two transports:Canonical reference: [`docs/agent/MCP_TOOLS.md`](../agent/MCP_TOOLS.md).A single-repo server advertises **10** tools by default (the canonical set).-Workspace mode adds `list_repos`. Six specialists are registered but opt-in:+Workspace mode adds `list_repos`. Seven specialists are registered but opt-in:`get_architecture`, `get_blast_radius`, `get_dependency_path`,-`get_execution_flows`, `generate_refactoring_code`, and `get_conformance`.+`get_execution_flows`, `generate_refactoring_code`, `get_conformance`, and+`set_finding_status`.| Tool | What it answers | When to call ||------|----------------|-------------|diff --git a/docs/architecture/README.md b/docs/architecture/README.mdindex 931803364..9aaed73d7 100644--- a/docs/architecture/README.md+++ b/docs/architecture/README.md@@ -8,9 +8,15 @@ curious; you don't need them to *use* repowise (start with|-----|--------|| [ARCHITECTURE.md](ARCHITECTURE.md) | The single end-to-end system reference: stores, provider abstraction, the init and maintenance pipelines, git intelligence, dead code, decisions, MCP/REST/UI layers, and key design decisions || [deep-dives.md](deep-dives.md) | Algorithm-level treatment of the hard parts: dead-code detection, the decision/ADR system, search & vector-store internals, incremental updates & webhooks, the change-cascade algorithm |-| [code-health.md](code-health.md) | The code-health layer as a full vertical slice: pipeline, the 25+ markers, scoring, trends, persistence schema, and extension points |+| [code-health.md](code-health.md) | The code-health layer as a full vertical slice: pipeline, the 53 registered detectors, scoring, trends, persistence schema, and extension points |+| [change-risk.md](change-risk.md) | The change-risk model: features, offline calibration and why the score is reported as a size statistic, PR structural impact, and the public risk-scale inventory |+| [bug-history.md](bug-history.md) | Bug-fix history internals: fix-shape classification, the `fix_events` table, the bug-magnet rollup, and the inducing-commit (SZZ) work that was measured and not shipped |+| [decisions.md](decisions.md) | Decision capture internals: the source registry, evidence verification, confidence, currency, session mining and hook delivery |+| [test-intelligence.md](test-intelligence.md) | The inferred test tier: call and import walks, their measured precision and recall, the resolution-origin filter, and why nothing is stored || [language-support.md](language-support.md) | The language pipeline internals and the step-by-step recipe for adding a new language |-| [graph-algorithms.md](graph-algorithms.md) | The graph algorithms (PageRank, betweenness, Tarjan SCC, Louvain, shortest path) with the math and complexity |+| [graph-algorithms.md](graph-algorithms.md) | The graph algorithms (PageRank, betweenness, Tarjan SCC, Leiden/Louvain community detection, shortest path) with the math and complexity |+| [refactoring.md](refactoring.md) | Refactoring intelligence internals: detector registration, plan shape, ranking, REST routes, and code generation |+| [savings-accounting.md](savings-accounting.md) | The savings accounting contract: what counts as a measured reduction versus an inferred avoidance, and what a total may claim || [chat.md](chat.md) | The Codebase Chat feature: schema, `ChatProvider` protocol, SSE streaming, the agentic loop, and per-provider notes || [structurizr-export.md](structurizr-export.md) | Exporting the architecture as Structurizr DSL: the two-file model, health and layer tags, and how to render it || [editor-files.md](editor-files.md) | How `CLAUDE.md` / `AGENTS.md` generation works (no LLM: pure DB + filesystem derivation) and how to add a new editor file |diff --git a/docs/architecture/bug-history.md b/docs/architecture/bug-history.mdnew file mode 100644index 000000000..f7a7f382b--- /dev/null+++ b/docs/architecture/bug-history.md@@ -0,0 +1,89 @@+# Bug-fix history internals++Contributor reference for how fix history is stored and rolled up, and the+record of the inducing-commit work that was measured and not shipped.+User-facing behaviour is in [docs/layers/BUG_HISTORY.md](../layers/BUG_HISTORY.md).++## Pipeline++1. **Candidate selection.** `is_fix_commit` in+ `core/ingestion/git_indexer/_constants.py` applies the keyword include and+ exclude rules. The same rule feeds the health `prior_defect` marker and+ change-risk `fix_history`, so every fix count in the product counts the same+ commits.+2. **Shape classification.** `classify_fix_shape` in+ `git_indexer/fix_shape.py` reads the candidate's `-U0` diff and assigns one+ of `SHAPE_KINDS` (`code_fix`, `test_only`, `doc_only`, `config_other`,+ `comment_only`, `empty`). Only `code_fix` is counted.+3. **Events.** `git_indexer/fix_events.py` writes one `fix_events` row per+ counted fix per file it touched.+4. **Attribution and rollup.** `core/analysis/health/fix_attribution.py`+ matches events to symbol spans and computes the per-file decayed mass and+ `bug_magnet`.++## The `fix_events` table++```+fix_events(repository_id, fix_sha, file_path, shape_kind,+ old_ranges_json, fixed_at, ...)+```++`old_ranges_json` holds the line ranges the fix replaced, numbered on the+fix's own parent commit. Storing events, not only counts, is what makes+per-symbol attribution and changed-line overlap possible without re-walking+git. A full index populates the table; `repowise update` extends it.++`GitMetadata` carries the rollups: `prior_defect_count` (filtered),+`prior_defect_raw_count` (keyword only), `bug_magnet`, `last_fix_at`, and+`fix_symbol_counts_json` (`path::Name -> count`, written in descending order).++## Rollup++`FIX_HALF_LIFE_DAYS = 90.0` and `BUG_MAGNET_MASS = 3.0` in+`fix_attribution.py`. The half-life came from a 60 / 90 / 180-day sweep on the+21-repo health calibration corpus: decay was the only lever that moved the+`prior_defect` coefficient (0.15 undecayed to 0.23 decayed), and 90 and 180+tied inside noise. The rollup recomputes the whole repo on every run because+decay changes it with no file change; on this repo that is two indexed queries+and about a quarter second.++Change-risk `fix_history` uses its own 365-day half-life+(`analysis/change_risk/fix_history.py`), tuned separately on its ranking gate.++`last_fix_at` must be serialized with a timezone. SQLite's DATETIME bind drops+`tzinfo`, and a naive ISO string parses as local time in JavaScript, which+pushes a fresh fix into the future west of UTC and makes its age vanish. The+fix lives once, upstream of every consumer.++## Inducing-commit attribution (SZZ): measured, not shipped++A refactor-aware SZZ pass was built: blame each fix's replaced lines back+through history, walk through behaviour-preserving moves, rank candidates by+overlap. On 53 hand-judged rows it reached **74.5%** top-candidate precision+against an 80% gate. A UI naming a commit and its author would be wrong about+one time in four with no way for the reader to tell which, so the pass was+removed and no surface names an inducing commit. Removing it also saved 9.1 s+of index time on this repo and 5.9 s on zod.++Findings worth keeping for anyone who tries again:++- **Refactor-aware blame helps.** Walking through moves lifted strict precision+ from 70.6% to 74.5%; all 14 initial false calls were refactors inheriting+ moved lines.+- **Overlap ranking beats earliest-commit ranking by 12 points** on judged+ rows. Earliest-commit first looked like 80%, but 7 of its 16 unjudged answers+ were the repo's initial commit, and all 7 were judged not plausible.+- **Two failure modes belong to SZZ itself**: an initial import has no earlier+ commit to name, and package splits that move lines wholesale while+ re-sorting imports defeat the carried-through test.++### AI vs human introduced bugs: not built++A per-KLOC comparison of defects introduced by agent-authored and+human-authored commits was gated on a concentration check, and failed it: the+top 5 inducing commits held 30.3% of traced mass against a 20% ceiling, with+the initial import alone at 13.9%. Any aggregate over inducing commits inherits+that concentration and mostly describes who authored the five largest commits.++What ships answers "this code has broken before, recently, and roughly here".+Nothing answers "and this commit caused it".diff --git a/docs/architecture/change-risk.md b/docs/architecture/change-risk.mdnew file mode 100644index 000000000..37503fd08--- /dev/null+++ b/docs/architecture/change-risk.md@@ -0,0 +1,129 @@+# Change risk internals++Contributor reference for the change-risk model, its calibration, and every+public risk scale. User-facing behaviour is in+[docs/layers/CHANGE_RISK.md](../layers/CHANGE_RISK.md).++Code lives in `packages/core/src/repowise/core/analysis/change_risk/`+(`features.py`, `model.py`, `baseline.py`, `fix_history.py`, `service.py`),+`analysis/risk_semantics.py` (the scale contracts) and `analysis/pr_blast.py`+(PR structural impact). The runtime is deterministic, makes no LLM or network+calls, and runs no blame: SZZ labelling exists only in the offline+calibration.++## The model++A plain L2-logistic over standardized, `log1p`-compressed Kamei change metrics+(`la`, `ld`, `nf`, `nd`, `ns`, `entropy`, `exp`):+`logit = intercept + sum(coef_i * z_i)`. Every feature's contribution is exact+and reported as a driver. `exp` is optional; when the author cannot be+resolved it is `null` and contributes zero, since `0` is a real value (first+commit) the model reads as risk-raising. `nf`, `nd` and `ns` have small+negative coefficients from collinearity with `la`; they stay in the logit as+fit and are hidden as drivers because the explanation would contradict itself.++Constants are hard-coded in `_CONSTANTS` in `model.py`. Only constants ship.++## Calibration++- **Corpus**: 4,102 commits (662 inducing) from a 7-repo, 5-language slice:+ clap, pydantic, fd, gin, fastify, bat, chi.+- **Labels**: AG-SZZ bug-inducing commits, with a 120-day right-censoring gap.+ `age_days` is excluded as a label-availability artifact.+- **Evaluation**: leave-one-repo-out pooled out-of-fold AUC against a+ churn-only baseline.+- **Result**: 0.772 for the model vs 0.766 for churn-only (delta +0.0068, 95%+ CI [-0.0003, +0.0131]).++That margin is small, and it is the number the constants were selected on. It+is not evidence that the score ranks danger.++### Why the score is reported as a size statistic++- `la` carries a coefficient 7.6x the next largest; scoring by `la` alone+ reproduces the full score within 0.12-0.16 points on every repo tried.+- On 47 constructed within-repo pairs (repowise, flask, django, zod) of a small+ dangerous change and a large boring one, the score ranks the dangerous change+ higher in 0 of 47; fix density does in 46 of 47. The pairs are built so size+ alone scores 0, so this is a falsification test, not an accuracy estimate.+- A PR-granularity refit (`--first-parent` merge spans) with two+ size-orthogonal features scored worse: pooled LOO AUC 0.769, against 0.776+ for the current feature set and 0.780 for churn-only. Lines added alone+ matched or beat the fitted model in five of six repos.++The cause is the labels. A commit is marked inducing when a later fix's blame+lands on a line it wrote, and a bigger commit writes more lines, so the label+is size-biased. Any size-orthogonal feature scores near chance against it (fix+density lands at 0.46-0.57 AUC). The constants were kept and the score is+labelled as what it is: `risk_authority` and `score_measures` say so on every+payload.++`fix_history` carries no AUC for the same reason. Its evidence is the 46/47+pair gate and that its top-ranked files in this repo are the ones with the+longest bug-fix records. It uses a 365-day half-life (`FIX_HALF_LIFE_DAYS` in+`fix_history.py`) and `DEEP_WALK_LIMIT = 20_000`.++## PR structural impact (`get_risk`)++`structural_impact_score` in `pr_blast.py`:++```+combined = 0.5 * mean + 0.5 * max of pagerank * (1 + temporal_hotspot) over changed files+file_term = 8 * (1 - exp(-10 * combined)) # 0-8+reach_term = 2 * min(transitive_dependents / 20, 1) # 0-2+score = min(file_term + reach_term, 10)+```++Bands: `localized` below 4, `moderate` 4 to below 7, `broad` 7 and up.+Deterministic and uncalibrated. Co-change evidence is reported beside it and+never enters it. `overall_risk_score` is an exact deprecated alias, described+by `overall_risk_score_compatibility`.++The fixture corpus `tests/fixtures/risk_scale_corpus.json` pins the+distribution:++| Control | Score | Band |+| --- | ---: | --- |+| documentation / low signal | 0.01 | localized |+| small ordinary source | 0.34 | localized |+| historical fixes, limited reach | 0.34 | localized |+| co-change only | 0.08 | localized |+| moderate multi-file | 4.27 | moderate |+| structurally broad, little history | 7.34 | broad |+| genuinely broad high control | 9.91 | broad |++No fixture saturates at 10, ordinary controls stay below the moderate+threshold, and all three bands are occupied.++## Public scale inventory++| Public value | Producer | Unit / range | Authority |+| --- | --- | --- | --- |+| `get_change_risk.score` / REST `score` / stored `change_risk_score` | Logistic model over diff size, spread, entropy, author experience | normalized points 0-10, single-commit scale | Supporting; not a probability |+| `risk_percentile` | Mid-rank of the score among filtered recent commits | percentile 0-100 | Authoritative with `classification` |+| `review_priority` / `classification` | Percentile terciles at 33.33 and 66.67 | category | Authoritative |+| `fallback_band` | Score thresholds at 4 and 7, only without a baseline | category | Absolute fallback |+| `fix_history.density` | Churn-weighted, recency-decayed prior fixes | decayed fixes, unbounded | Uncalibrated, separate evidence |+| `get_risk` `hotspot_score` | Repo-relative churn percentile | ratio 0-1 | Uncalibrated component |+| `get_risk` `health_score` | Code-health model | 0-10, higher is healthier | Benchmarked file signal |+| `structural_impact_score` | Formula above | normalized points 0-10 | Uncalibrated; not authoritative |+| `direct_risks[].structural_score` | `pagerank * (1 + temporal_hotspot)` | unbounded | Uncalibrated; `risk_score` is an alias |+| `direct_risks[].temporal_hotspot` | Decayed per-commit churn (half-life 180 d, each commit up to 3.0) | unbounded | Input to `structural_score` |+| `direct_risks[].churn_percentile` | Rank of `temporal_hotspot` in the repo | 0-1 | Rank only |+| `direct_risks[].is_hotspot` | Top-quartile churn plus absolute activity floors | boolean | Use instead of re-deriving from `churn_percentile` |+| `cochange_warnings[].score` | Commits in which the pair co-changed | count | Historical evidence only |+| workspace `impacted[].score` | Strongest path product of edge confidence, kind weight, `0.6` per hop | 0-1 | Uncalibrated ranking |+| dashboard hotspot triage index | `40% churn pct + 35% bus-factor tier + 25% temporal activity` | 0-100 | Client-side orientation |++`risk_authority`, `structural_impact_scale`, `overall_risk_score_compatibility`+and `impact_score_semantics` carry the guard tier (unit, range, calibration,+authority) on every payload. The reference tier (corpus, formula, components,+the full `risk_scales` dictionary) never varies, so MCP returns it only for+`include=["scales"]`; the CLI `--format json` output carries it always.++## Recalibrating++The calibration script lives in the separate benchmark repository+(`health-defect/jit_calibration.py`). Paste the regenerated constants into+`_CONSTANTS` in `model.py`, and update the corpus figures in `model.py` and+`risk_semantics.py` together.diff --git a/docs/architecture/code-health.md b/docs/architecture/code-health.mdindex 9a7304728..4f1fb9d6e 100644--- a/docs/architecture/code-health.md+++ b/docs/architecture/code-health.md@@ -39,8 +39,8 @@ them. Its only writes are to its own four tables.┌────────────────────────────────┼────────────────────────────────┐▼ ▼ ▼CLI (rich) MCP tools (FastMCP) Web dashboard- health, status, get_health, get_risk, /repos/[id]/health{,- health --trend get_context, get_overview /coverage,/refactoring-targets}+ health, status, get_health, get_risk, /repos/[id]/code-health+ health --trend get_context, get_overview (tabs, ?tab=...)```Three architectural rules govern the whole layer:@@ -251,7 +251,7 @@ The returned report rides on `PipelineResult.health_report`. Thentrend tracking (rolling 50-row window per repo), and a second`{path: total_deduction}` map covering only the files whose score is held atthe floor. Both maps come from `trends.snapshot_file_maps`, which the other two-snapshot writers (`repowise health` and `repowise update --full`) also call — a repo+snapshot writers (`repowise health` and `repowise update --full`) also call; a repowhose writers disagreed would get a history whose depth changed depending onwhich command last wrote it.@@ -307,8 +307,8 @@ are **56 marker ids**. They divide by what each is permitted to affect:| Defect-scoring | 25 | `defect` (11 of them also `maintainability`) || Performance | 22 | `performance` only || SQL | 3 | `maintainability` only |-| Governance | 3 | nothing — the finding surfaces, the score is untouched |-| Advisory | 3 | nothing — measured by construction, kept out of impact-ranked lists unless requested |+| Governance | 3 | nothing: the finding surfaces, the score is untouched |+| Advisory | 3 | nothing: measured by construction, kept out of impact-ranked lists unless requested |The authority is `scoring._BIOMARKER_DIMENSIONS`. Any biomarker **not** listedthere defaults into `defect`, which is why every `sql_*` and every performance@@ -330,7 +330,7 @@ golden guarantee (§6).† The three governance markers carry a category and a weight, but the pass thatwrites them runs *after* scoring completes and never touches-`HealthFileMetric.score` — so in practice they never deduct. They are counted+`HealthFileMetric.score`, so in practice they never deduct. They are countedin the table above because `scoring.py` maps them, not because they move anumber.@@ -472,6 +472,45 @@ a language that maps neither field simply emits no blocks.intersects the clone report with assertion spans. Both gate on`coverage.is_test_file(path)` so production code is never touched.+### 5.4 Advisory markers and their precision record++`advisory` is not in `DIMENSIONS`, so no weight, category or cap table can key+on it; "never deducts" is structural. A marker must appear in both+`_BIOMARKER_DIMENSIONS` and `_ADVISORY_HOME`, or it still deducts from defect.+Advisory findings carry `health_impact == 0.0`, never count toward a change's+introduced/worsened totals, never make a review verdict blocking, and stay out of+impact-ranked lists unless the dimension or marker is named.++A marker earns weight by clearing roughly 70% hand-labelled precision on a real+corpus. Precision is measured per language because it does not transfer:++- `mock_saturated_test`: 67% on Python (32 findings) and 40% on TypeScript (30),+ each the full population at the shipped gate. Both predate the assertion count+ gaining more oracle shapes, which can only suppress findings, so they are now a+ ceiling. It stays advisory until two false-positive families separate:+ `Fake*` value-object builders passed into real logic, and boundary isolation+ where the assertion reads a real artifact. Both turn on whether an assertion+ observes a double or production output, a dataflow question the pass does not+ ask.+- `assertion_free_test`: at least 71% on TypeScript (31 findings, the full+ population of two corpora; a floor, since three false-positive families closed+ after labelling) and 86% on Python (29 labelled, a systematic sample of 172). It+ does not report on Go or Java. A test that delegates its oracle to a helper is+ resolved by name in the same file and through the call graph across files. A+ mock verification counts as an oracle here, the opposite of+ `mock_saturated_test`; the tiers are in `complexity/assertions.py`.+- `hidden_coupling`: see §5 (near-chance AUC; non-inferior without it).++### 5.5 Worked example: the dispatch rule++`nested_complexity` increments depth on each `if` / `for` / `while` / `try` /+`switch` node and fires at depth 4 or more, severity scaling with depth. A function+that is mostly one dispatch on one value (a `switch`, `match` or same-subject `if`+chain holding at least 60% of its decision points) is judged by the code around the+dispatch and its heaviest arm: `nested_complexity` skips the two levels the+`switch` and its `case` open, and `complex_method`, `brain_method` and `bumpy_road`+read the CCN outside the dispatch plus the CCN of its most complex arm.+`biomarkers/registry.py` is an **explicit list**, not auto-discovery:keeps the registration order deterministic and lets tests inject extrasvia `registered_biomarkers(extra=...)`.@@ -507,8 +546,8 @@ scored at the pre-window commit (T0, no leakage) and an L2-logistic regression,with NLOC as an explicit control, fits each marker's defect lift beyond filesize. The runtime stays deterministic; only the learned constants ship. Thefull calibration, with confidence intervals, is published in the-[benchmark report](https://github.com/repowise-dev/repowise-bench/blob/master/health-defect/BENCHMARK_REPORT.md)-and reproduced by `local-stash/calibrate_health_weights.py`.+[benchmark report](https://github.com/repowise-dev/repowise-bench/blob/master/health-defect/BENCHMARK_REPORT.md);+§6.2 describes the protocol.| Weight | Markers | Rationale ||---|---|---|@@ -521,14 +560,188 @@ and reproduced by `local-stash/calibrate_health_weights.py`.| 1.0 | `prior_defect` | Neutral by design: largely redundant with the other process signals, kept for its explanatory value. || 0.5 (floored) | `developer_congestion`, `dry_violation`, `low_cohesion`, `brain_method`, `primitive_obsession`, `bumpy_road` | Fire widely but proved weak under leakage-free scoring; kept as maintainability and parity signals, not disabled. || 0.4 (de-rated) | `knowledge_loss` | Weakest of the floored set. |+| 0.5 (not fitted) | `error_handling` | Excluded from calibration (AUC-neutral); floored and capped at 0.5 per file. |The same marker stream feeds the **maintainability** signal under anindependent, expert-set weight table (the floored smells deduct at full weightthere), and the **performance** signal under its own bounded `performance`category. The three signals share one scoring kernel against separateweight/category/cap tables and never feed back into each other; see the-[user guide](../layers/CODE_HEALTH.md#three-health-signals-defect-risk-maintainability-and-performance)-for what each signal surfaces and why the overall score stays the defect score.+[user guide](../layers/CODE_HEALTH.md#the-three-signals)+for what each signal surfaces. The headline stays the defect score: the band+cutoffs and every accuracy figure are claims about that pillar, and+`test_scoring_dimensions` locks `score_file(...)["defect"]` byte-for-byte to the+pre-split single score. Blending the pillars would need a written rationale and a+recalibration corpus.++### 6.2 Validation and calibration++**Protocol.** Every file is scored at a commit T0 that precedes the bug window, in+a detached worktree whose history stops at T0. Labels are the bug-fix commits of+the following six months. An L2-regularized logistic regression with NLOC as an+explicit control column fits each marker's lift beyond size, with+leave-one-repo-out cross-validation. Only the learned constants ship.++**Leakage runs both ways.** Scoring at HEAD lets the bug-fix commits manufacture+the churn, co-change and congestion the process markers then "detect". Scoring+naively at T0 silences every recency-windowed marker, because a worktree six+months back sees an empty "last 90 days". Recency windows are therefore anchored to+the worktree's own HEAD (`as_of`), not wall-clock time. The evidence that this+matters: `developer_congestion` shipped at 1.5 under HEAD scoring and came back at+coefficient -0.08 under T0 scoring, hence its 0.5 floor.++**What calibration rejected.** Graph centrality (PageRank), code naturalness,+change bursts and error-handling density were declared equivalent to null by+bootstrap TOST, not merely non-significant. Size-relative scoring (lifts small+files, costs 0.07 to 0.12 AUC overall), function-level prediction (Popt at or+below random effort ordering) and recalibrating on SZZ labels (lower out-of-fold+AUC than the shipped weights) were all run and refuted.++**Two corpora.** The cross-project result uses 21 repositories, 9 languages,+2,826 files (379 defect-bearing). The CodeScene comparison uses the 2,770 of those+files both tools could score. The band-separation figures (16.9x raw, 2.18x+size-normalized, At risk vs 8.0 and above) come from the paired 2,770-file set.++| Cross-project result (2,826 files) | Value |+|---|---|+| Mean ROC AUC | 0.737 (95% CI 0.683 to 0.787) |+| Held out: PROMISE jEdit 4.0 / 4.1 (no git history, structural markers only) | 0.761 / 0.776 |+| vs. recent churn | +0.100 AUC (DeLong p = 5e-10) |+| vs. prior-defect history | +0.117 AUC (DeLong p = 3e-15) |+| Per-repo range | 0.55 (axios) to 0.86 (zod) |++Intervals come from a two-stage repo-cluster bootstrap (2,000 seeded replicates,+repositories then files), since the repository is the unit of generalization.++Limits measured on the same corpus:++- LOC alone scores AUC 0.742 (delta -0.001, p = 0.92). The win is effort-aware:+ Popt +0.134 (95% CI +0.080 to +0.198).+- Within NLOC quartiles AUC is 0.525 / 0.572 / 0.593 / 0.718. A positive control+ injecting a size-orthogonal signal of the headline magnitude was recovered with+ power 0.998 to 1.000 in those same quartiles, so the collapse is a property of+ the score, not of sample size.+- A prior-defects baseline beats the score on Popt by 0.085 (95% CI 0.035 to+ 0.141).++| Paired test vs CodeScene (2,770 files) | Repowise | CodeScene | p |+|---|---:|---:|---|+| Recall at 20%-of-lines budget | 0.173 | 0.074 | 0.003 |+| Popt | 0.607 | 0.462 | 0.003 |+| Defect density, size-normalized | 2.18x | 0.56x | 0.003 |+| ROC AUC | 0.731 | 0.705 | 0.054 (not significant) |+| Defect density, raw | 16.9x | 14.2x | 0.65 (not significant) |+| Precision at 20%-of-lines budget | 0.580 | 0.636 | 0.64 (tie) |+| Partial rho beyond size | -0.148 | -0.137 | both beat size (tie) |++CodeScene flags 27 files at its most severe level where Repowise flags 132 At+risk: a more conservative operating point, not a calibration flaw. Its "Code Red"+correlation of -0.58 with issue-resolution time (proprietary Jira data) did not+replicate on open GitHub data: -0.09 (95% CI -0.19 to +0.14), and six+queue-independent effort signals were flat.++**In-product check.** `defect_accuracy.py` ranks files by score, takes the 20+lowest and counts those touched by a fix commit in the trailing 180 days against the+repo base rate. It stays silent below 25 scored files or 5 fixed files. Because+`prior_defect` is an input, this is an association on indexed history, not a+forward test.++Full method and data:+[repowise-bench/health-defect](https://github.com/repowise-dev/repowise-bench/tree/master/health-defect)+([COMPARISON_REPORT.md](https://github.com/repowise-dev/repowise-bench/blob/master/health-defect/COMPARISON_REPORT.md)).++### 6.3 Fix first ranking (`fix_first/`, `worth.py`)++Fix first is built once in `analysis/health/fix_first/build.py` from stored rows+and rendered unchanged by `get_health`, `repowise health`, the dashboard and the+generated CLAUDE.md. Every excluded unit is counted in `totals.excluded` under one+reason:++| Reason | What it leaves out |+|---|---|+| `test`, `tooling`, `generated`, `vendored`, `docs_example` | Path-role classification. Tooling includes build files by type wherever they sit (Gradle, CMake, Makefiles, Bazel, MSBuild props/targets, root `setup.py`, `noxfile.py`, crate-root `build.rs`, bundler configs) |+| `unknown`, `expected`, `no_strategy`, `no_plan` | Performance causes with no classified context, expected repetition, no supported strategy, or no stored safe plan |+| `below_min_worth` | A refactoring recovering under `MIN_WORTH` (0.5) or with no steps; a loop-built string that is bounded or not in production |+| `history_only` | Only git-history findings |+| `deprecated` | Deprecated functions |+| `inherent_dispatch` | One dispatch holds 60% or more of the decision points (labels: 8 of 9 such rows rejected), unless a duplicate also sits in it |+| `small_function` | Under 30 code lines and CCN 15 (on 67 labelled rows this cut drops 13 rejected and 4 accepted) |+| `no_concrete_step` | No first edit with a file and line or a named group |+| `low_value_kind` | Extract Class (0 of 14 accepted), Move Method (0 of 34), low cohesion (0 of 46), long method (0 of 10), long parameter lists (0 of 2) |++Order: value first (the larger of health recovered and size: CCN 20/40/80/150,+lines 100/200/400/800, nesting 5/6/8, a critical or brain-method finding, one step+more on a hot file); a performance fix's value is its cost, with production,+entry-reachable, data-growing DB or network calls in the top band. Then tier+`now` / `next` / `later`, with at most `HEAD_PER_KIND` (3) of the first `HEAD` (5)+places per kind. `worth.py` decides `later` for every default list: a function+under CCN 40, 200 lines and nesting 6 (nesting under 8 only counts at 100 lines or+more) and not both CCN 25 and nesting 5; a dominant dispatch; nesting that is one+else-if or ternary chain; long but CCN under 20 and nesting under 5; a single+complex condition or error site. From CCN 80, 400 lines or nesting 8 a function is+worth doing regardless. The hot-file bonus orders items but never lifts one out of+`later`. Each item carries up to `MAX_TESTS` (5) tests from its validation profile.++The findings list uses the same rule: worth-doing findings lead and the rest carry+`lower_priority`. Class-design findings and history-only findings are lower+priority there; a performance opportunity is unless production code runs it over+growing data.++### 6.4 Performance opportunities (`perf/`)++**Sinks.** `io_in_loop` resolves a call inside a loop through the graph's+resolver and classifies the target against the I/O-boundary lexicon+(`io_boundaries.py`: db, network, filesystem, subprocess, lock). It fires only on a+real round-trip, never on a query-builder chain. A non-local sink is found by+walking the resolved call graph backwards up to three hops (`crossfn.py`), and the+`caller -> ... -> sink` path is attached.++**Gates.** Two markers use call-graph centrality as a precision gate: they fire+only in functions at or above the 80th percentile of distinct direct callers, and+never under two. That measures how widely a function is called, not request+reachability. `blocking_sync_in_async` skips tests, tooling, examples, generated+and vendored code.++**Opportunities.** One opportunity per intervention site within one cost family+and boundary: the function holding the loop, a shared helper every caller passes+through, or `<file>::__module__` for top-level code. Sinks, call sites and+same-family co-signals (`io_in_loop` with `nested_loop_with_io`) are members. Two+loops in one function are one site, because findings carry the sink's line. Each+carries one `actionability` state (`plan_ready`, `advisory`, `investigate`,+`expected`). The default queue holds production `plan_ready` and `advisory`, ranked+by `rank_score`; actionability only breaks ties. Excluded reasons are counted in+`default_queue`.++**Plans.** Proven means the transformation, not the runtime: parallelizing awaits+against a DB or network client is advisory with a `bounded_concurrency`+prerequisite. A loop already walking chunks is `loop_already_chunked`. When+`io_in_loop` and `serial_await_in_loop` fire on one call they name each other in+`siblings`. Plans list their edits (`mechanical` only when proven) and validation+tests (coverage, call graph, import graph, then `via: "name-match"`). An+opportunity links to a `performance_fix` refactoring plan only by exact+`opportunity_id`.++**Dataflow promotion.** `serial_await_in_loop` and `nested_loop_quadratic` are+promoted to `dataflow_verified: true` when a def/use pass proves iterations+independent. Promotion changes the wording, not the score.++**ORM lazy loads.** `lazy_load_in_loop` (sync SQLAlchemy, Django) needs the loop's+rows to resolve to one model through the cross-file model index and the attribute+to be a lazy relation the query does not eager-load. Weight is read per finding+from `details.orm` (`_PERFORMANCE_ORM_WEIGHT`): Django measured 29/32 (90.6%) on a+held-out sample and carries 0.7; SQLAlchemy, with under 30 held-out labels, stays at+0.4. `unbounded_read_reduced_in_memory` (Python) flags an unlimited query whose rows+are deduplicated per key in code.++**Dialects.** `perf/dialects/` registers 12 dialects over 18 language tags (TS/JS+also serve Vue and Svelte script blocks; C# serves Razor). C shares the C++ grammar+but has no perf dialect. A language without one emits nothing.++**Soundness.** Dynamic dispatch, monkeypatching and callbacks-as-values produce no+call edge. ORM lazy loads are seen only where the rows can be typed. Chains beyond+three hops are not followed; an unmodelled library has no classified sinks. The+published linter comparison and ranking quality figures are in+[repowise-bench/perf-detection](https://github.com/repowise-dev/repowise-bench/tree/master/perf-detection).---@@ -550,8 +763,8 @@ KPI cards.## 8. Trends (`trends.py`)State-free: callers pass an oldest-first list of snapshot rows. Both alerts-run over every metric in `_ALERT_METRICS` — hotspot health, the composite-headline and maintainability — skipping any a snapshot never recorded:+run over every metric in `_ALERT_METRICS` (hotspot health, the composite+headline and maintainability), skipping any a snapshot never recorded:- **Declining Health**: current is ≥ `DECLINE_THRESHOLD` (default 0.5)below the snapshot `DECLINE_LOOKBACK` (5) positions back. Fires on the@@ -566,7 +779,7 @@ improved. It carries the same numbers and the opposite reading, because adecline the code shape did not contribute to has nothing to act on andreporting it in error red tells a reader their refactoring made things worse.Maintainability is code shape already, so it has no halves to split and never-softens — it is the fall that always deserves the alarm.+softens; it is the fall that always deserves the alarm.`recent_kpis(history, limit=10)` returns a newest-first serialised viewfor the CLI table and MCP `get_health(include=["trend"])` response.@@ -596,7 +809,7 @@ health-breakdown response, and the standalone trend route (§13).The stored score is clamped to `[SCORE_FLOOR, SCORE_MAX]`, so files 12.9 and9.1 points deep both persist as `1.0` and their series is flat however much ofthe work gets done. The second snapshot map (`per_file_deductions_json`) keeps-the pre-clamp deduction for exactly those files — for every other file the+the pre-clamp deduction for exactly those files; for every other file thededuction is `SCORE_MAX - score`, so there is nothing to store.Each point therefore carries `unclamped_score` alongside `score`:@@ -605,7 +818,7 @@ It is the series `_file_declining` runs on, so `declining` describes the linethat can actually move, and a file getting worse *below* the floor now trips it.Snapshots written before this existed have no deduction map. Their floored-files stay flat, which is correct — the depth was never measured, and inventing+files stay flat, which is correct; the depth was never measured, and inventingone would be worse than a flat line.### Per-file signals (`signals.py`)@@ -675,6 +888,15 @@ refreshed when its `public_id` is detected again, and `resolved` reopens.| `status` | lifecycle || `created_at`, `updated_at` | datetime |+**Identity and triage** (`finding_identity.py`). The public id+(`finding_<digest>`) is anchored on the enclosing function or class name plus+the finding's line offset into it, so an edit above the symbol keeps the id; a+file-level finding keeps absolute lines. Metric values are not part of the id;+only the detail keys a marker needs to tell two findings apart (a coupling+partner, an error kind) are. Each index replaces open findings, and a re-detected+finding matching a triaged row updates that row's evidence in place:+`acknowledged` and `false_positive` stay, `resolved` reopens.+### `health_file_metrics`One row per file (unique on `(repository_id, file_path)`). Read directly@@ -758,6 +980,13 @@ Defined in `tool_health/tool.py`, which dispatches to one module per mode and bl| `"coverage"` | per-file coverage rows + summary || `"refactoring"` | deterministic `suggestion` text on every finding || `"trend"` | snapshot diff + alerts + last 10 KPI rows |+| `"accuracy"` | the in-product defect-accuracy block (§6.2) |+| `"performance"` | performance opportunities (§6.4) |+| `"signals"`, `"churn_complexity"`, `"doc_drift"` | per-file signals, churn x complexity points, drift |+| `"defect"`, `"maintainability"`, `"advisory"` | dimensions for the ranked findings list (default: defect + maintainability) |+| `"unverified"`, `"semantics"` | provisional finding types (labelled), and the shared points/percentile legend |++`request.py::_KNOWN_INCLUDES` is the authority.### Enrichments on existing tools@@ -776,7 +1005,7 @@ Every response carries the standard `_meta` envelope via `build_meta()`.## 13. REST surface-`packages/server/src/repowise/server/routers/code_health/` — a package, one+`packages/server/src/repowise/server/routers/code_health/`: a package, onemodule per surface, sharing `scope.py`, `counts.py`, `file_filters.py` and`statuses.py`. All routes under `/api/repos/{repo_id}/health/`:@@ -815,20 +1044,17 @@ Auth is the standard `verify_api_key` dependency from## 14. Web dashboard-Three routes under `/repos/[id]/health/`:--| Route | What it shows |-|---|---|-| `/health` | KPI cards, lowest-scoring file table, top findings, **per-module rollup** (added in Phase 4) |-| `/health/coverage` | Coverage summary, untested-hotspot warnings, module-level bars, per-file drill-down |-| `/health/refactoring-targets` | Cards sorted by impact-per-effort, each with severity, marker, score, NLOC, effort bucket, **deterministic suggestion** |+One page, `packages/web/src/app/repos/[id]/code-health/page.tsx`, with a map whose+lens switches between health, maintainability, performance and churn, and tabs+selected by `?tab=`: Overview (`triage`, leads with Fix first), Performance,+Findings, Tests (`coverage`), Dead code, Doc drift, Security and Blast radius+(`impact`). `scope` and `counts` controls sit above the tabs. The older+`/repos/[id]/health/*` routes are redirects into this page;+`code-health/refactoring-targets` and `code-health/trend` remain as sub-pages.-Plus a sidecar `HealthRisksPanel` on the Hotspots, Ownership, and Graph-pages: surfaces the lowest-scoring files inline without touching the-shared table/graph components. The **Hotspots & churn** tab carries the-`ChurnComplexityQuadrant` (fed by `GET /churn-complexity`), toggleable in-place with the existing churn × bus-factor scatter; the file Health tab-carries the per-function "Functions by churn" blame table.+The churn lens is paired with `ChurnComplexityQuadrant` (fed by+`GET /churn-complexity`); the file Health tab carries the per-function+"Functions by churn" blame table.All visual primitives live in `packages/ui/src/health/` so the hosted`frontend/` repo (separate git checkout) can reuse them: port is mostly@@ -858,8 +1084,9 @@ agent in noise. The Jinja stanza lives in`.repowise/health-rules.json` is user-authored (the **only** JSON file in thelayer) and is loaded by `HealthConfig.load(repo_path)`. It carries repo-wide and-per-path `disabled_biomarkers` and `severity_overrides`, keyed by an fnmatch-glob over the repo-relative POSIX path (`path`, with `path_glob` and `glob` as+per-path `disabled_biomarkers` and `severity_overrides`, keyed by a+gitignore-semantics glob over the repo-relative POSIX path (every matching rule+applies; disabled sets union) (`path`, with `path_glob` and `glob` asaccepted aliases). `to_analyzer_config(file_paths)` resolves the globs toper-file disabled sets, which the engine honors in `_evaluate_file()`. Theschema and examples are in the@@ -869,7 +1096,7 @@ schema and examples are in the## 17. Performance-Plan §4 P4.6 targets **< 30 s on a 3,000-file synthetic repo**. The+The budget is **< 30 s on a 3,000-file synthetic repo**. Theparallel path in `HealthAnalyzer.analyze_async()` parallelises tree-sitterparsing across worker threads (`asyncio.gather` + `asyncio.to_thread`).tree-sitter releases the GIL on parse, so this scales on single-process@@ -886,7 +1113,7 @@ Other perf notes:repos with low duplication.- **Walker re-parses files** because `ParsedFile` doesn't retain atree-sitter `Tree` across the ingestion boundary. Acceptable (~1 ms- per file); switching to a shared parse cache is a Phase 5 stretch.+ per file); a shared parse cache would remove it.- **No N² loops in scoring.** Category aggregation is O(findings).---@@ -951,12 +1178,12 @@ a new override key (beyond `disabled_biomarkers`), extend## 20. Where the layer deliberately stops-A short list of things the v1 layer **does not** do, by design. Future-phases may revisit; the constraints kept v1 shippable.+Things the layer **does not** do, by design.-- **No LLM-generated suggestions.** `suggestions.py` is static- templates. An optional LLM mode is Phase 5, gated behind an explicit- flag.+- **No LLM in scoring or suggestions.** `suggestions.py` is static+ templates. LLM code generation exists only behind the explicit+ `--generate-code` flag and the opt-in `generate_refactoring_code` MCP tool,+ and never feeds a score.- **No symbol-level scoring.** Score lives at the file granularity tomatch how engineers think about refactor units. Symbol-level CCNstill feeds the file score via `all_functions`.@@ -972,7 +1199,7 @@ phases may revisit; the constraints kept v1 shippable.categorical layer is the five absolute bands (Excellent / Good / Fair /Needs work / At risk, `grading.py`); a letter on top would be a thirdoverlapping scale with arbitrary cliffs. Every surface that shows a band- word or a band colour reads that one vocabulary — there is no second+ word or a band colour reads that one vocabulary; there is no secondramp.---diff --git a/docs/architecture/dead-code.md b/docs/architecture/dead-code.mdnew file mode 100644index 000000000..80c635330--- /dev/null+++ b/docs/architecture/dead-code.md@@ -0,0 +1,182 @@+# Dead-code analysis internals++Contributor reference for the dead-code layer. For what users see and how to+read findings, start with [layers/DEAD_CODE.md](../layers/DEAD_CODE.md). The+algorithm walkthrough in [deep-dives.md](deep-dives.md#1-dead-code-detection)+covers the graph basics; this page records the rules that shape confidence.++All code lives in `packages/core/src/repowise/core/analysis/dead_code/` unless+a path says otherwise.++## Module map++| Module | Role |+|--------|------|+| `analyzer.py` | `DeadCodeAnalyzer`: the four detectors, confidence scoring, caps, rescues. |+| `constants.py` | Never-flag globs, framework decorator lists, dynamic-dispatch name patterns, non-package directories, language sets. |+| `risk_factors.py` | Path risk tokens, `SAFE_CONFIDENCE_THRESHOLD` (0.7), `RISK_CAP_CONFIDENCE` (0.4), `effective_safe_to_delete`. |+| `serving.py` | Read side shared by MCP, REST and workspace merges: `TIER_FLOORS`, filtering, tiering, cross-repo adjustment. |+| `file_reachability.py` | The single "can anything reach this file?" answer. |+| `entry_shape.py` | Files shaped like programs (shebang, Python main guard). |+| `dynamic_markers.py` | Source-text runtime-loader markers per file extension, plus `dynamic_*` graph edges. |+| `name_occurrences.py` | The "is the name written anywhere else?" cap for unused exports. |+| `published_api.py` | Packaged .NET projects: public symbols held at `0.30`. |+| `c_name_uses.py`, `cpp_reachability.py` | C/C++ name uses outside declarations; directory-granular reachability. |+| `contract_methods.py` | COM and runtime contract method names (`QueryInterface`, `AddRef`, `Release`, `IFACEMETHODIMP` / `STDMETHODIMP` implementations). |+| `csharp_reachability.py` | Name-granular reachability for C# (namespaces, not files). |+| `go_reachability.py`, `go_name_uses.py` | Package-granular reachability and package-qualified name uses for Go. |+| `jvm_reachability.py`, `jvm_name_scope.py` | Package-granular reachability and bare-name scope for Java, Kotlin and Scala. |+| `kotlin_multiplatform.py` | `expect` / `actual` pairs. |+| `module_strings.py` | Python modules named by dotted-path strings (entry-point tables, Celery, Django settings). |+| `jsx_prop_guards.py` | JSX components rendered only behind a prop that is never supplied. |++Related: `analysis/finding_registry.py` decides which finding kinds a default+surface shows (`unused_internal` is `hidden`). Storage and triage live in+`persistence/crud/analysis/dead_code.py`. The analyzer runs from+`pipeline/phases/analysis.py` (init), `pipeline/incremental.py` (update), and+the CLI's live path in `cli/commands/dead_code_cmd.py`.++## Flow++1. `analyze(config)` runs the enabled detectors in order: unreachable files,+ unused exports, unused internals, zombie packages. Config keys:+ `detect_*` toggles, `min_confidence`, `dynamic_patterns`, `whitelist`+ (path-exact; no user config feeds it today).+2. Post-passes cap or drop findings (C# scope names, C/C++ name uses, the+ name-occurrence search, unindexed-file names, published API).+3. Findings below `min_confidence` are dropped and counted in+ `hidden_below_threshold`. Init and update store at the `0.4` floor.+4. Persistence replaces `open` rows only. Rows a user marked `acknowledged`,+ `resolved` or `false_positive` survive, and a new finding matching one by+ (file, kind, symbol) is not re-inserted.+5. Read time: `effective_safe_to_delete` re-derives safety. It only ever+ downgrades a stored flag, so rows written under older rules stay honest.++## Unreachable files++Candidate: file-node in-degree 0 after entry points, never-flag paths and the+rescues in `file_reachability.py`. Base confidence comes from git+(`_git_age_confidence`, `_GIT_AGE_RUNGS`):++| Condition | Confidence |+|-----------|-----------|+| No git row at all | `0.50` (`NO_GIT_SIGNAL_CONFIDENCE`) |+| Commits in the last 90 days | `0.40` |+| No commits in 90 days, last commit 365+ days ago | `1.00` |+| ... 180+ days ago | `0.90` |+| ... 90+ days ago | `0.80` |+| File under 30 days old | `0.55` |+| Otherwise | `0.70` |++Caps to `0.40`: a dynamic-import marker in the same directory, any path risk+factor, a namespace-import language (C#). `unreachable_file` is in+`REVIEW_ONLY_KINDS`, so `safe_to_delete` is always false. `lines` is the+physical line count, or unknown when the source was not read.++### Path risk tokens (`risk_factors.py`)++Filename split on `. _ -`; directory segments matched whole.++| Factor | Filename tokens | Directory segments |+|--------|-----------------|--------------------|+| `config` | `config`, `configs`, `configuration`, `conf`, `settings`, `setting`, `setup` | `config`, `configs`, `settings` |+| `environment` | `env`, `environment`, `environ`, `dotenv` | `env`, `environments` |+| `bootstrap` | `bootstrap`, `startup`, `entrypoint` | `bootstrap` |+| `database` | `database`, `db`, `schema`, `seed`, `seeds`, `migration`, `migrations`, `datastore`, `sqlite` | `database`, `db`, `migrations` |+| `script` | none | `scripts`, `bin`, `tasks` |+| `asset` | `sw`, and the pair `service` + `worker` | `public`, `static`, `www` |++Excluded on purpose: broad names (`app`, `main`, `index`, `core`, `base`,+`util`) would cap ordinary modules, and `assets` is the bundled-source+convention in Vite, Vue and Angular.++## Unused exports++Candidate: a public, top-level symbol (methods, fields, properties and enum+members are skipped) with no `imports` edge naming it (or `*`, or a TypeScript+`export { local as alias }`), and no inbound edge in+`REACHABILITY_USE_EDGE_TYPES` (`ingestion/models.py`).++Treated as used without an edge (`_used_without_an_edge`): framework+decorators and their suffixes (`_FRAMEWORK_DECORATORS`,+`_FRAMEWORK_DECORATOR_SUFFIXES`), `@SuppressWarnings("unused")`, names matching+`_DEFAULT_DYNAMIC_PATTERNS` (`*Plugin`, `*Handler`, `*Adapter`, `*Middleware`,+`*Mixin`, `*Command`, `register_*`, `on_*`, `*_view`, `*_endpoint`,+`*_route`, `*_callback`, `*_signal`, `*_task`), framework inner classes+(Python `Meta`), same-file TS/JS type uses, and Python local references.++Base confidence (`_unused_export_confidence`):++- `0.30` if deprecated (name suffix `_DEPRECATED` / `_LEGACY` / `_COMPAT`, or a+ deprecation annotation in any supported form).+- `1.00` when an importer of the file names specific symbols, so absence is+ evidence.+- `0.60` (`UNPROVEN_EXPORT_CONFIDENCE`) otherwise: no importer names symbols (a+ C `#include`, a C# `using`, no importer), C, C++ and Objective-C symbols, C#+ extension-method classes.++Caps to `0.40`: an interface in a file with no `implements` edges, a contract+method, a path risk factor, a dynamic marker in the directory, the name+written elsewhere in the repository (`name_occurrences.py`, including+non-code files; also applied when the search could not run), and the name+appearing in a source file that was skipped during indexing. Deletion-ready+only at `>= 0.7` with no risk factor.++### Language rescues++- **C, C++, Objective-C** (`c_name_uses.py`): a finding is dropped when any+ name its declaration introduces is written outside a declaration, in code or+ in `.def`, `.asm` or `.s` files. Typedefs contribute tag and aliases, enums+ their enumerators. Comments and prose strings do not count. COM methods+ declared with `IFACEMETHODIMP` / `STDMETHODIMP` are never reported.+- **Java, Kotlin, Scala** (`jvm_name_scope.py`): a name written in another+ file in scope, or further down its own file, is a use. Classes named by a+ `build.gradle(.kts)` (`classname`, `implementationClass`, `mainClass`) are+ live.+- **C#**: a file a .NET reference assembly (`ref/*.cs`) lists is never+ reported; packaged projects are held at `0.30` by `published_api.py`.+- **Go**: package-qualified and same-package uses (`go_name_uses.py`).+- **Rust**: `unused_internal` is disabled; rustc's `dead_code` lint covers it.+ Proc-macro entry points are live.+- **Python**: a literal `__all__` raises listed names to `public`; it never+ rescues a finding on its own.++## Unused internals and zombie packages++`unused_internal`: private symbols with no `calls` edge and no cross-file+importer, base `0.65`, never deletion-ready. Hidden on default surfaces by+`finding_registry.py` (measured 0.5% precision, 1,511 findings).++`zombie_package`: a top-level directory with no inter-package import edges,+base `0.50`, never deletion-ready. `_NEVER_PACKAGE_DIRS` skips non-package+directories (`.github`, `.vscode`, `.devcontainer`, `docs`, `examples`,+`scripts`, `assets`, `static`, `public`, `tests`, `benches`, `fuzz` and+similar), and directories holding only Dockerfiles, Makefiles and shell+scripts are run, not imported.++## Never flagged++`never_flag_path` and `_NEVER_FLAG_PATTERNS` in `constants.py` cover: entry+points (`__init__.py`, `__main__.py`, `conftest.py`, `manage.py`, `wsgi.py`,+`asgi.py`, `setup.py`, `main.go`), build files (Gradle, Maven, CMake,+Makefiles, Meson, Bazel, MSBuild, `build.rs`, bundler configs), shell scripts,+files a CI workflow, task file, manifest or script names by path, framework+routes (Next.js, SvelteKit, Nuxt, Remix, ASP.NET minimal APIs, Blazor and+Razor code-behind), test layouts, generated code (protoc, Qt, Bison/Flex, SWIG,+Cython, Roslyn `*.g.cs`, Dart `*.g.dart`, `**/generated/**`), reflective+loading (Alembic and Django migrations, EF configurations, COM class+factories), vendored trees, build artifacts, and non-code languages from the+language registry. Read the constants for the exact globs; they change more+often than this page.++## Serving++`TIER_FLOORS` in `serving.py` is `{"high": 0.7, "medium": 0.4, "low": 0.0}`,+built from the `risk_factors.py` constants. The TypeScript mirror is+`DEAD_CODE_CONFIDENCE` in `packages/types/src/dead-code.ts`, so CLI, web and+MCP tier identically.++`adjust_cross_repo` runs after tiering in workspace mode: a file with+cross-repo co-change partners gets confidence `* 0.5` and a+`cross_repo_note`; otherwise an `unused_export` in a repo that other repos+depend on as a package gets `* 0.3`.diff --git a/docs/architecture/decisions.md b/docs/architecture/decisions.mdnew file mode 100644index 000000000..bbe5e8dec--- /dev/null+++ b/docs/architecture/decisions.md@@ -0,0 +1,173 @@+# Decisions: capture internals++How the decision layer captures, verifies and ranks records. For what a user+does with them, see [layers/DECISIONS.md](../layers/DECISIONS.md).++Code lives under `packages/core/src/repowise/core/analysis/decisions/` (sources,+policy, lifecycle, provenance), `packages/core/src/repowise/core/sessions/`+(transcript mining) and `packages/cli/src/repowise/cli/commands/augment_cmd/`+(hook delivery).++## Source registry++`policy.py` defines one `SourceSpec` per source: whether it has a deterministic+stage, whether it has a model stage, its authority (`machine` or `human`) and+whether it ships on. One resolved policy backs the CLI, the REST API and the+index pipeline, so `repowise decision config show` prints what the next `init`+or `update` will run.++| Key | Deterministic stage | Model stage | Default |+|-----|---------------------|-------------|---------|+| `inline_marker` | yes | yes | on |+| `git_archaeology` | no | yes | on |+| `adr` | yes | yes | on |+| `pr` | no | yes | on |+| `comment` | no | yes | on |+| `session` | yes | yes | off |+| `session_discovery` | no | yes | off |+| `conventions` | yes | no | off |+| `cli` | yes | no | always (not switchable) |++A source with no deterministic stage reports `skipped_no_provider` when no+model is configured; a hybrid source falls back to its deterministic stage.+`session_discovery` and `conventions` were added after the presets existed, so+a stored preset that enumerates its sources does not silently gain them.++Source details:++- **ADR** (`adr.py`): files in a fixed set of ADR directories, up to 60. Nygard+ and MADR headings plus YAML frontmatter map to record fields. A status of+ `accepted` or `approved` maps to `active`. Self-acceptance happens in+ `persistence/crud/decision_ingest.py`: only `adr` is a tracked-artifact+ source, and only when git tracks the evidence file.+- **Inline markers** (`markers.py`): `WHY|DECISION|TRADEOFF|ADR|RATIONALE|REJECTED`+ followed by a colon, case-sensitive, after any comment leader. Up to 5+ continuation lines and 20 lines of context on each side. Fenced Markdown code is skipped.+- **Commits** (`commit_mining.py`, `commit_signals.py`): gated on decision+ verbs before any model call.+- **PR bodies**: only bodies with PR-description structure (`## Why`,+ `## Motivation`, `## Context`, `Closes #`, `Before:` / `After:`), up to 25.+- **Comments** (`rationale_comments.py`): block comments and docstrings on+ high-centrality files, bounded to 30 nodes, and only prose with a rationale+ cue ("because", "instead of", "trade-off", "we chose", "deliberately").+- **Conventions** (`conventions.py`): a wrapper of an I/O library that most+ files reach the library through while few import it directly. Go and Java+ are counted by package. Up to 10 per index.++## Evidence verification++Every produced field (`decision`, `rationale`, `source_quote`) is checked+against the verbatim source span the extractor recorded (`provenance.verify_quote`).++| Verdict | Fires when | Confidence multiplier |+|---------|-----------|-----------------------|+| `exact` | The normalized quote is a substring of the span | 1.0 |+| `fuzzy` | Token overlap with the span is at least 0.6 | 0.85 |+| `unverified` | Neither, or no span to check | 0.6 |++An ungrounded field is cleared. A candidate whose every produced field is+ungrounded is rejected. A candidate with no source text at all is kept but+stamped `unverified`.++## Confidence++`provenance.py` computes:++```+earned = 0.5 * (top_rank / 9) * (filled_fields / BODY_FIELDS) # completeness scaling, when known+conf = (0.4 + earned + min(0.12, 0.04 * (corroborating_sources - 1)))+ * verification multiplier+ * 0.85 if top_rank <= 2+clamped to [0, 0.99]+```++The rank ladder (`SOURCE_RANK`): `cli` 9, `session` 8, `adr` 7, `pr` 7,+`commit` / `git_archaeology` / `conventions` 6, `inline_marker` 4, `comment` 3.+Retired sources keep their rungs (`changelog` 5, `readme_mining` 3,+`code_comment` 2) so old rows still rank. `session` sits above `adr` because a+transcript carries what a person said while deciding; an ADR is a later+write-up. When two sources describe one decision, headline fields come from the+higher rank and the other is kept as a corroborating evidence row.++## Currency++`lifecycle.effective_currency` derives the shown currency from the stored one:+`superseded`, `dismissed` and an explicit `needs_review` are returned as set.+An `active` record becomes `uncheckable` with no scope, `stale` when every+named path is gone at HEAD (`head_artifacts.py`, renames followed, shallow+clones treated as present), `needs_review` at staleness 0.5 or above, and+`active` otherwise. Working agreements (`kind: agreement`) are repository-wide+and always read `active`. Only `active` and `needs_review` govern.++## Supersession and conflict detection++`evolution.detect_supersessions_and_conflicts` used embedding similarity to+scope candidate pairs, then opposing verbs or reversal phrases, with a model+tiebreaker. It is disabled (`SEMANTIC_SUPERSESSION_ENABLED = False`): among+records from one repository, high cosine similarity was the baseline, not+evidence of a shared topic. It was the only writer of `supersedes` and+`conflicts_with` edges. The diff-driven pass on `update` (`reverts.py`) still+marks a decision a commit reversed, and `deprecate --superseded-by` records the+successor on the record.++## Session mining++`sessions/miners/decisions.py` reads coding-agent transcripts through adapters+in `sessions/adapters/` (`claude_code`, `codex`). Only `claude_code` is read+unless `decisions.harnesses` names more. Codex stores sessions by date, so the+repository a line belongs to is decided by its recorded working directory;+sessions imported into Codex from another harness are skipped. Reading is+incremental from a cursor.++Three stages:++1. **Gates.** A user correction needs a pushback lead ("no,", "don't",+ "actually,", "instead"). An explicit choice needs a decision verb paired+ with a causal marker. A dead end needs three consecutive failures of the+ same command (`DEAD_END_FAILURES = 3`).+2. **Structuring.** One batched model call per update, capped at 60+ candidates (`MAX_STRUCTURED_PER_UPDATE`). Every field must quote the+ transcript; an ungrounded `source_quote` rejects the candidate.+3. **Promotion.** Seen in two distinct sessions, or one direct correction, the+ record is written as a `proposed` candidate with `source: session`. Repeat+ observations add evidence rows.++Staging lives in `.repowise/sessions/sessions.db`. Transcripts never leave the+machine; only distilled decision text is stored. The same switch gates+injected-decision feedback and update-time hook-efficacy replay.++### Broad discovery++`session_discovery` (`analysis/decisions/discovery/`) makes one model call per+update over newly read user and assistant prose, bounded by+`discovery.max_sessions` (1-24, default 12) and `discovery.max_input_tokens`+(2,000-60,000, default 30,000). It reuses the `session` source's read, so it+depends on that source being on. Each prose span carries an id; a candidate+must cite sent ids, quote near-verbatim, overlap the claim's content words,+and name only paths the cited turns' tools touched. Bundled claims are flagged+for splitting. Unsent prose stays queued oldest-first; a provider failure+requeues and retries on the next two updates.++## Review state and lanes++Candidates carry a review row (`open`, `accepted`, `merged`, `needs_split`,+`dismissed`) with the raising lane, extractor version, bundling flag, whether+any scope is named, and a priority equal to the acceptance check's verdict at+the last index. Re-extraction refreshes signals but never review state.+The five dashboard lanes are a join onto the acceptance log, not a filter on+status, so `GET /api/repos/{id}/decisions?lane=...` and `/decisions/lane-counts`+differ from `/decisions/counts`.++`decision status` keeps no run log: it derives its figures from records,+review rows, staging queues and `decision_extraction` cost rows, so spend is+all-time and a queue a store predates is reported absent.++## Hook delivery++`augment_cmd/decision_inject.py`: session start scores records against the+working set expanded one hop, multiplies relevance by confidence and freshness,+and injects under `_TOKEN_CAP = 400`, with candidates in a separate section+under `_CANDIDATE_TOKEN_CAP = 120`. Edit-time notices go through+`decision_node_links`, once per session per decision. Injected ids are recorded+in the sessions sidecar so the next `update` can tell whether guidance was+followed; that verdict does not change staleness.diff --git a/docs/architecture/deep-dives.md b/docs/architecture/deep-dives.mdindex d2b2deac0..aa523e30a 100644--- a/docs/architecture/deep-dives.md+++ b/docs/architecture/deep-dives.md@@ -1,6 +1,6 @@-# Repowise Deep Dives — Complete Guide+# Repowise Deep Dives-This document covers systems that are referenced but not fully explained in `architecture-guide.md` and `graph-algorithms-guide.md`. Each section is self-contained with full intuition, implementation details, and the math behind the algorithms.+This document covers systems that are referenced but not fully explained in [ARCHITECTURE.md](ARCHITECTURE.md) and [graph-algorithms.md](graph-algorithms.md). Each section is self-contained with full intuition, implementation details, and the math behind the algorithms.---diff --git a/docs/architecture/editor-files.md b/docs/architecture/editor-files.mdindex 1c6602def..683578b57 100644--- a/docs/architecture/editor-files.md+++ b/docs/architecture/editor-files.md@@ -32,9 +32,8 @@ By embedding codebase intelligence and MCP workflow guidance directly into`CLAUDE.md`, Claude Code treats it as project context and naturally reaches forRepowise tools without being prompted.-The same principle applies to Cursor's `.cursor/rules` / `cursor.md`, GitHub-Copilot's `.github/copilot-instructions.md`, and any other file an AI editor-auto-loads at session start.+The same principle applies to Codex's `AGENTS.md`, Cursor's rules, and any+other file an AI editor auto-loads at session start.---@@ -544,8 +543,10 @@ from .cursor_md import CursorMdGenerator # add this line### Step 5 — Hook into init and update-In `_maybe_generate_claude_md()` (or extract a more generic-`_maybe_generate_editor_files()` helper):+Add a sibling of `maybe_generate_claude_md()` in+`packages/cli/src/repowise/cli/editor_integrations/claude.py` (Codex's+`maybe_generate_agents_md()` in `codex.py` follows the same contract), or extract a+generic helper both can share:```python# After existing CLAUDE.md generationdiff --git a/docs/architecture/graph-algorithms.md b/docs/architecture/graph-algorithms.mdindex 83089e97a..fe7520a53 100644--- a/docs/architecture/graph-algorithms.md+++ b/docs/architecture/graph-algorithms.md@@ -35,7 +35,7 @@ Framework edges are detected automatically when the tech stack includes Django,**`file_subgraph()`: keeping metrics accurate across both tiers**-All centrality algorithms (PageRank, betweenness, SCCs, Louvain) operate on the file-level subgraph, not the full graph. `file_subgraph()` returns a view of the `DiGraph` containing only `file` and `package` nodes. Without this isolation, the large number of symbol nodes would inflate degree counts and distort every metric. Call `file_subgraph()` before running any file-level algorithm; call the full graph when you need symbol-level traversal.+All centrality algorithms (PageRank, betweenness, SCCs, community detection) operate on the file-level subgraph, not the full graph. `file_subgraph()` returns a view of the `DiGraph` containing only `file` and `package` nodes. Without this isolation, the large number of symbol nodes would inflate degree counts and distort every metric. Call `file_subgraph()` before running any file-level algorithm; call the full graph when you need symbol-level traversal.This is the raw material. Every algorithm below operates on the file subgraph unless otherwise noted.@@ -349,7 +349,7 @@ Result: Three SCCs: {A,B,C}, {D}, {E}. Only {A,B,C} has size > 1, meaning it's a----## 4. Louvain Community Detection+## 4. Community Detection (Leiden, Louvain fallback)### What question does it answer?@@ -363,19 +363,25 @@ Community detection finds these clusters automatically. It looks at the densityThink of it as automated "package detection." Even if your code doesn't use clean package boundaries, community detection reveals the actual subsystem structure from the import graph.-### Why undirected?+### Which algorithm runs-Repowise converts the directed graph to undirected before running Louvain:+`detect_file_communities()` in `core/analysis/communities.py` tries three partitioners in order:-```python-communities = nx.community.louvain_communities(g.to_undirected(), seed=42)-```+1. **Leiden** (`graspologic.partition.leiden`, `random_seed=42`). graspologic ships in the optional `graph-extra` extra (`pip install "repowise[graph-extra]"`), so a default install skips this step.+2. **Louvain** (`networkx.community.louvain_communities`, `seed=42`, `threshold=1e-4`, `max_level=10`). This is what a default install runs.+3. **Directory grouping**: if Louvain raises, files are grouped by their first path segment.++The algorithm that ran is returned alongside the partition. Leiden and Louvain optimize the same objective (modularity, below); Leiden adds a refinement step that guarantees each community is internally connected, which Louvain does not.++Only production files are partitioned. Test and example files are assigned afterwards to the community of a production file they link to, and isolated files each get their own community. Any community holding more than 30% of the files (minimum 20) gets a second partition pass to split it.++### Why undirected?-Why? Import direction doesn't matter for clustering. If `auth.py` imports `crypto.py`, they're related, regardless of which one depends on which. The question isn't "who depends on whom" (that's PageRank's job) but "who belongs with whom."+Repowise builds an undirected graph from the file dependency edges (`imports`, `framework`, `type_use`, `reads`, and the dynamic import and route edges) before partitioning. Why? Import direction doesn't matter for clustering. If `auth.py` imports `crypto.py`, they're related, regardless of which one depends on which. The question isn't "who depends on whom" (that's PageRank's job) but "who belongs with whom."### The math: Modularity-Louvain optimizes a metric called **modularity** (Q). Modularity measures the difference between actual intra-community edges and what you'd expect by random chance.+Both algorithms optimize a metric called **modularity** (Q). Modularity measures the difference between actual intra-community edges and what you'd expect by random chance.```Q = (1/2m) × Σ [ A_ij - (k_i × k_j)/(2m) ] × δ(c_i, c_j)@@ -434,7 +440,7 @@ This repeats until Q stops improving. Each round of aggregation discovers larger**Time complexity:** Nearly O(N) in practice (each node is moved a small constant number of times). This makes it suitable for large codebases.-**Why `seed=42`?** Louvain is non-deterministic; the order you process nodes affects the result. Fixing the random seed makes results reproducible across runs. 42 is a conventional choice (from The Hitchhiker's Guide).+**Why `seed=42`?** Both Leiden and Louvain are non-deterministic; the order you process nodes affects the result. Fixing the random seed makes results reproducible across runs. 42 is a conventional choice (from The Hitchhiker's Guide).### How Repowise uses communities@@ -612,7 +618,7 @@ These algorithms aren't isolated; they feed into each other to create a complete| PageRank | Simple import count | Captures transitive importance, not just direct || Betweenness | Closeness centrality | Betweenness directly identifies bottlenecks; closeness measures average distance which is less actionable || Tarjan's SCC | Brute-force cycle detection | O(N+E) vs O(N³). For large codebases, brute-force is impractical |-| Louvain | Spectral clustering, Girvan-Newman | Louvain is nearly O(N) and handles large graphs. Girvan-Newman is O(N²E), too slow. Spectral requires matrix decomposition, overkill for this use case |+| Leiden, Louvain fallback | Spectral clustering, Girvan-Newman | Both are nearly O(N) and handle large graphs; Leiden also guarantees connected communities. Girvan-Newman is O(N²E), too slow. Spectral requires matrix decomposition, overkill for this use case || BFS shortest path | Dijkstra | Edges are unweighted (an import is an import). BFS is optimal for unweighted graphs and simpler than Dijkstra |---@@ -624,7 +630,7 @@ These algorithms aren't isolated; they feed into each other to create a complete| PageRank | O(E × iterations) ≈ O(E × 50) | O(N) | Converge fallback to uniform || Betweenness | O(N × E) | O(N + E) | Sample k=500 for repos > 30k nodes || SCCs (Tarjan) | O(N + E) | O(N) | None needed, already linear |-| Louvain | ~O(N) empirically | O(N + E) | Seed=42 for determinism |+| Leiden / Louvain | ~O(N) empirically | O(N + E) | Seed=42 for determinism; second pass splits communities over 30% of files || BFS shortest path | O(N + E) | O(N) | Cutoff=3 for entry-point views || In-degree | O(1) per node | O(1) | None needed |diff --git a/docs/architecture/language-support.md b/docs/architecture/language-support.mdindex 2946ec26e..140a0019c 100644--- a/docs/architecture/language-support.md+++ b/docs/architecture/language-support.md@@ -34,7 +34,7 @@ new language needs one only when it hits the same kind of wall.- [Multi-language files (the SFC pattern)](#multi-language-files-the-sfc-pattern)- [Per-language mechanics](#per-language-mechanics) · [Elixir](#elixir) · [F#](#f) · [Objective-C](#objective-c) · [GDScript / Godot](#gdscript--godot) · [VB.NET](#vbnet) · [QML](#qml) · [Flutter widget trees](#flutter-widget-trees)- [Optional language-specific passes](#optional-language-specific-passes)-- [The three code-health dialect registries](#the-three-code-health-dialect-registries)+- [The three code-health dialect registries](#the-three-code-health-dialect-registries) · [test-quality markers per language](#test-quality-markers-per-language)- [Workspace contract extraction](#workspace-contract-extraction)---@@ -871,6 +871,12 @@ when nothing on that chain carries a script, when the script would come fromanother scene's `instance=ExtResource(...)`, or when the resolved script declaresno such function. It is never matched on the method name alone.+A `uid://` reference resolves through the `.uid` sidecar Godot writes for+scripts; scenes get no sidecar, so a scene uid resolves through the `path=` on+its `[ext_resource]` header instead. An `addons/` tree is exempt from dead-code+reporting only when a `project.godot` sits above it, so a plugin's own repo+reports normally.+Two constructs the upstream grammar rejects: `$%UniqueName` (`$` and `%` areseparate node-path forms, so the pair fails), and a bare call to a function named`export` or `onready` (`export()`), which the grammar still reserves at statement@@ -961,9 +967,9 @@ cognitive complexity and per-function markers. Optional additions widen it:metrics (LCOM4, god-class); `assert_kinds` / `assert_call_kinds` addassertion-block smells. See `complexity/README.md`.-Note that C has **no map at all**, despite sharing the C++ grammar, so-`get_language_map("c")` is `None` and the health pass never reaches a dialect for-it. Registering C in the downstream registries would be dead configuration.+C, F# and Objective-C have maps of their own (C separate from C++ despite the+shared grammar), so they get complexity markers, but none of the three has a perf+or dataflow dialect yet.### 2. Performance: `analysis/health/perf/dialects/` (`PERF_DIALECTS`)@@ -992,6 +998,10 @@ dialect reuses them instead of re-deriving them:this statement bind the name" question is per-grammar. (The Python, Ruby andDart dialects predate the hook and keep their own tuned versions.)+Object Pascal's DB and network sinks are gated on file-wide `uses` evidence: a+file importing `FireDAC` gates every `.Open` / `.ExecSQL` / `.Post` in it,+because a Pascal variable's declared type has no textual link back to its unit.+### 3. Dataflow: `analysis/health/dataflow/dialects/` (`DEFUSE_DIALECTS`)Intra-procedural CFG + def/use + reaching definitions, powering **Extract@@ -1015,6 +1025,51 @@ out of that is a wrong suggestion.All three registries are purely additive and degrade to silence: an unmappedlanguage produces no findings rather than wrong ones.+### Test-quality markers per language++The test-quality markers need per-language data beyond the node map, and a+language with no row produces no signal.++- **Assertion count** (`analysis/health/asserts/lexicon.py`) has two tiers. The+ narrow tier is the fixed `assert` / `expect` prefix match the scored+ assertion-block markers are calibrated on. The broad tier adds a per-language+ row plus a repo's `health.assertions` config, feeds only the advisory count,+ and matches names exactly, never as prefixes (prefix families such as `check`+ and `validate` match production functions, not assertions). Oracles that are+ not plain assertion statements are counted at the broad tier too:+ `with pytest.raises(...)`, Java's split `object` + `name` call shape, chai+ property chains (`expect(x).to.be.null`), expression-bodied test lambdas and+ private `_assert_*` helpers.+- **Mock saturation** (`analysis/health/mocks/lexicon.py`, `MOCK_DIALECTS`) has+ rows for Python and JS/TS only. Go has no mock vocabulary. Java is held out+ because Mockito states its checks as `verify(...)`, which counts toward+ `verification_count` and not `assertion_count`, so an over-mocked Java test+ arrives with a near-empty denominator.+- **Assertion-free test** needs a per-language "is this a test case" rule+ (`analysis/health/complexity/test_case.py`: a name prefix for Python, `TestXxx`+ for Go, the `it(...)` / `test(...)` callback for JS/TS, `@Test` for Java) and+ reports only for the detector's `SHIPPING_LANGUAGES` (JS, TS, TSX, Python).+ Mock verification does count as an oracle here. Go and Java are classified but+ not reported: Go hands `*testing.T` to package-level helpers and Java inherits+ base-class helpers, which the oracle resolution below does not reliably reach.+- **Oracle resolution.** A test that delegates its checks is not assertion-free.+ Within a file, the names a function calls resolve against same-file functions+ that assert. Across files, `asserts/oracle_reach.py` uses two lanes: a test+ that is a graph symbol walks its own call edges (depth 2); an anonymous JS/TS+ callback, which has no node, needs both a resolved edge from its file to the+ asserting symbol and an unqualified call by name in its body. Only edges that+ bind one definition are read (`import_merged` is excluded), and an unresolved+ call suppresses nothing, so a gap leaves a false positive and never hides a+ real finding.+- **`.tsx` grammar.** A `.tsx` file is tagged `typescript`, but the complexity+ walk, clone tokenizer and dataflow CFG pick the grammar from the path, so JSX+ parses without error recovery.+- **Every walked function is visible.** Markers read `all_functions`, the full+ walked list, not a map keyed by function name, which would collapse anonymous+ `it` callbacks and same-named methods on different classes to one row.+ `duplicated_assertion_block` has no floor of its own; the `test_quality`+ category cap (0.5) bounds its effect on a file's score.+---## Workspace contract extractiondiff --git a/docs/architecture/refactoring.md b/docs/architecture/refactoring.mdnew file mode 100644index 000000000..1b3ca33cc--- /dev/null+++ b/docs/architecture/refactoring.md@@ -0,0 +1,89 @@+# Refactoring intelligence: internals++User guide: [docs/layers/REFACTORING.md](../layers/REFACTORING.md). This page is for+contributors. Code lives in `packages/core/src/repowise/core/analysis/health/refactoring/`.++## Detectors++Each detector is one module registered with `@register` in `registry.py`; adding a+type is a new file plus that decorator. `detect_refactorings` runs every enabled+detector per file, isolates failures (a crashing detector yields no suggestion for+that file) and applies the `min_confidence` floor. Detectors run inside the health+pass (`HealthAnalyzer` in `analysis/health/engine.py`) on data already computed.++| Module | Method |+|--------|--------|+| `extract_class.py` | LCOM4 union-find components as candidate classes; god-class shape confirmed with WMC and TCC (Lanza-Marinescu). |+| `extract_method.py` | Intra-procedural dataflow (CFG, def/use, reaching definitions, liveness) over functions already flagged `large_method`, `brain_method` or `complex_method`. Spans must be single-exit and statement-bounded; the behavior-preservation gate suppresses what it cannot prove. |+| `extract_helper.py` | Rabin-Karp clone pairs, clustered transitively; the helper site is the community centroid of the involved files. |+| `move_method.py` | Jaccard distance between a method's entity set and each class; fires only when a foreign class is clearly nearer and draws more of its calls. |+| `break_cycle.py` | Import-graph SCCs (one repo-wide index per pass); greedy minimum feedback arc set picks the cut edges. |+| `split_file.py` | Community detection (Leiden via the shared graph helpers) over a weighted intra-file symbol graph of calls, shared helpers, shared imports and co-change; emitted only above a modularity gate. |+| `performance_fix.py` | Wraps causal performance opportunities that have a supported shared intervention. |++The methods follow published work: Fokaefs-Tsantalis class splitting, Bavota+feature-envy distance, MFAS for cycles, Newman modularity for decomposition.++## Plan shape++`RefactoringSuggestion` in `models.py` is the source of truth for every field+(`plan`, `evidence`, `impact_delta`, `effort_bucket`, `blast_radius`, `confidence`,+`source_biomarker`, `validation`). Text is rendered only at the surfaces.+`impact_delta` credits the share of the source finding the plan removes; for+Extract Method that is `ccn_removed / ccn`, `slice_nloc / nloc`, or the larger of+the two for `brain_method`, capped at 1. `naming.py` returns `null` when no fact+anchors a name.++Plan ids come from `identity.py`: a `refac<version>_` prefix over a per-type stable+kernel. Bumping the version re-mints ids; a held id from an older version reports+`stale_model`.++## Ranking++`recommendations.py` owns four components and one score:++```+priority = benefit * (1 + leverage) / (1 + cost + risk)+```++Benefit multiplies, so a plan with no evidence of gain scores zero. Leverage is+weighted health deficit, dependents and entry reach. Blast radius feeds cost and+risk, never benefit. Graph-native and performance plans take benefit from their own+evidence.++## Opportunities++`opportunity.py` composes a file's plans into one opportunity at index time; the+finalizer writes `refactoring_opportunities` and a `refactoring_summaries` row in the+same transaction that reconciles plans. `performance_fix` is excluded (the+performance layer owns those). Plans whose source finding the finding registry+withholds are left out of steps and evidence but stay addressable by id.++Serving (filters, views, scope, paging) lives in `serving.py`; the server reads it+through `services/refactoring_health.py`, which MCP `get_health` and the REST routes+share, so the two surfaces cannot disagree. Scope `fix_first` takes its eligibility+from the Fix-first builder itself.++## REST routes (local server)++```text+GET /api/repos/{repo_id}/refactoring/opportunities ?view= &scope= &file_path= &limit=+GET /api/repos/{repo_id}/refactoring/opportunities/{opportunity_id}+GET /api/repos/{repo_id}/refactoring/summary+GET /api/repos/{repo_id}/refactoring/targets ?refactoring_type= &min_confidence=+GET /api/repos/{repo_id}/refactoring/targets/page ?limit= &offset= &refactoring_type=+GET /api/repos/{repo_id}/refactoring/{suggestion_id}+POST /api/repos/{repo_id}/refactoring/{suggestion_id}/generate-code+GET /api/repos/{repo_id}/refactoring/settings+PUT /api/repos/{repo_id}/refactoring/settings+```++## Code generation++`llm/enrich.py` gathers source spans, builds the prompt, calls the provider, runs the+per-type self-check and caches results under the repo's `.repowise` directory by a+hash of plan, source and model. `llm_enrichment_enabled` treats an unset+`refactoring.llm.enabled` as on. `build_enrichment_provider` resolves the top-level+`provider`/`model` from config, or the first provider with a key in the environment+or `.repowise/.env`. The MCP tool (`tool_refactoring.py`) is registered+`default=False`.diff --git a/docs/architecture/test-intelligence.md b/docs/architecture/test-intelligence.mdnew file mode 100644index 000000000..8f9ffbd8f--- /dev/null+++ b/docs/architecture/test-intelligence.md@@ -0,0 +1,144 @@+# Test intelligence: the inferred tier++How repowise answers "which tests reach this file" and "which tests should I+run" with no coverage report. For the user-facing layer, see+[layers/TEST_INTELLIGENCE.md](../layers/TEST_INTELLIGENCE.md).++Code: `packages/core/src/repowise/core/analysis/test_reachability.py` (the+walks), `analysis/test_selection.py` (runner arguments for `impacted-tests`),+`analysis/health/coverage/` (report parsers, discovery, path matching) and+`persistence/crud/analysis/coverage_map.py` (the measured per-test map).++## Why a graph signal++The measured map (`test_coverage`) is filled only by a coverage report. Most+repositories never ingest one, so without a second signal `tests_to_run` was+empty, `impacted_tests` said "run the full suite", and `untested_hotspot` fell+back to matching filenames.++Filename matching fails both ways. On this repository, five of the six worst+bug-magnet files have no test named for them, while the graph names the tests+that reach them:++| File | Filename convention | Graph |+|---|---|---|+| `ingestion/call_resolver.py` | nothing | 7 test files |+| `analysis/dead_code/analyzer.py` | nothing | 9 |+| `pipeline/persist.py` | nothing | 23 |+| `mcp_server/tool_answer/answer.py` | nothing | 18 |+| `analysis/pr_blast.py` | nothing | 3 |+| `analysis/health/engine.py` | `tests/unit/distill/test_engine.py` | 6, all under `tests/unit/health/` |++The last row is the worse failure: a basename match paired the health engine+with the distill engine's tests.++## Two graphs, two directions++The **call graph** leads: seed the symbols a test file declares, walk+`EXECUTION_EDGE_TYPES` forward, and every symbol reached is one the test can+run. `EXECUTION_EDGE_TYPES` (in `ingestion/models.py`) is+`SYMBOL_USE_EDGE_TYPES` minus `references`, because a name sitting in a+dispatch table is a mention, not a transfer of control. Dead code asks "is this+used", which a mention answers; this asks "is this run".++The **import graph** is the weaker tier: importing a module pulls in every+symbol it defines while the test may touch one function.++Measured against a real `coverage run --contexts=test` on this repository, over+the slice where per-test attribution is complete, both sides seeing the same+37 test files and 159 provably executed production files.++**Forward, "what reaches this file"** (suppresses `untested_hotspot`):++| Walk | Claims | Correct | Precision | Recall |+|---|---:|---:|---:|---:|+| Import graph, 1 hop | 43 | 31 | 72.1% | 19.5% |+| Call graph, 3 hops | 48 | 44 | 91.7% | 27.7% |+| **Call graph, 3 hops, filtered** (ships) | 46 | 44 | **95.7%** | **27.7%** |+| Both unioned | 57 | 45 | 78.9% | 28.3% |++**Reverse, "which tests do I run"** (`tests_to_run`, `impacted_tests`):++| Walk | Targets | Hit rate | Precision |+|---|---:|---:|---:|+| Import graph, 1 hop | 32 | 96.9% | 94.8% |+| Call graph, 3 hops, filtered | 46 | 100.0% | 97.5% |+| Both unioned | 47 | 100.0% | 95.8% |+| **Call graph, else import graph** (ships) | 47 | **100.0%** | **97.5%** |++The tiers combine differently per direction. Unioning costs the forward walk+16.8 points of precision for 0.6 of recall, and a false "something reaches+this" hides a real gap, so forward is call edges only. Reverse falls back: the+import tier is spent only on targets the call graph left silent, which answers+one more target at identical precision.++Both tables compare signals uncapped over the same test set. End to end, with+the walk seeing the whole test set and `MAX_TESTS_PER_TARGET` applied, the+reverse walk scores 100.0% hit and 95.5% precision on the same slice; the gap+is the cap truncating a long list.++## Depth++`DEFAULT_CALL_DEPTH = 3`: 3, 4 and 5 hops return the same 48 claims and 44+confirmations, so the recall ceiling is the call graph's capture rate, not the+depth. `DEFAULT_MAX_DEPTH = 1` for imports: a blanket second hop recovered 20+of 128 misses while adding 50 wrong claims; a facade-only second hop through+`__init__.py` (3 recovered, 5 added) and `conftest.py` transitivity (none+recovered) were also tested and rejected.++## The resolution-origin filter++"Filtered" drops call edges whose `resolution_origin` is `global_unique`: a+name bound to the only symbol carrying it anywhere in the repository, which the+origin vocabulary itself scores at 0.50. Dropping it costs no recall and buys+4.0 points of forward precision (91.7% to 95.7%) and 1.1 of reverse (96.4% to+97.5%). Filtering `receiver_global` too changes no number. A confidence floor+of 0.90 was tried instead and cut recall from 27.7% to 23.3%.++## Why recall reads low++Coverage records a line as run whether a test called into it or Python merely+evaluated the module body on import, so the truth set cannot tell an imported+file from an exercised one. Splitting the 159 truth files by how much of what+ran was inside a function body:++| What ran inside function bodies | Files | Recall |+|---|---:|---:|+| Nothing | 39 | 0% |+| Under a quarter | 19 | 0% |+| A quarter to three quarters | 45 | 7% |+| Over three quarters | 56 | 77% |++Of the 13 missed in the last group, 11 are alembic migrations the framework+invokes by naming convention with no static caller.++## Nothing is stored++The inferred map is read from `graph_edges` when asked and never written to+`test_coverage`:++1. **A consumer cannot mistake it for measured data.** Sharing the table behind+ a marker column would make every existing reader start returning inferred+ rows the day someone forgot to check the marker.+2. **It would be a transitive closure.** Tests reach most production files;+ materialising that is tests x sources rows that go stale whenever the graph+ moves, to answer a bounded breadth-first search over rows already in the+ database.++The health pass computes the whole-repo answer in one multi-source walk, once+per run, and caches it. The per-change walk reads the same edge table+`pr_blast` reads.++## Measured map internals++- `.coverage` is opened read-only as SQLite. Repowise decodes coverage.py's+ `numbits` line bitmaps itself and falls back to the `arc` table when line+ bits are absent, so there is no runtime dependency on coverage.py.+- Rows are indexed both ways: repo plus source file (reverse lookup) and repo+ plus test id (forward). `MAX_TEST_COVERAGE_ROWS = 250_000`.+- Ingest records (formats merged, matched, unmatched and tied paths, a sample of+ misses, the commit measured) are kept as history,+ `COVERAGE_HISTORY_RETENTION = 50`. A local index created before the history+ table changed keeps one record per repository until rebuilt.+- Path matching needs more than the basename when the report names a+ directory: `other/pkg/utils.py` never maps to `src/utils.py`.diff --git a/docs/business/COMMERCIAL.md b/docs/business/COMMERCIAL.mdindex d98cd36c9..302f37637 100644--- a/docs/business/COMMERCIAL.md+++ b/docs/business/COMMERCIAL.md@@ -48,7 +48,7 @@ All of the following ship in `pip install repowise` today, free for internal usecommits, contributor profiles, module health), Documentation (a wiki page permodule and file, freshness scoring, hybrid search), Decision (architecturaldecision records linked to graph nodes, staleness tracking), and Code Health- (51 deterministic detectors, 1–10 score per file, coverage ingestion, trend+ (53 deterministic detectors, 1–10 score per file, coverage ingestion, trendalerts). Full detail on each, and what every layer costs to build:**[INTELLIGENCE_LAYERS.md](../layers/INTELLIGENCE_LAYERS.md)**.- **Zero LLM calls in every analysis layer.** Graph, git, code health, change@@ -160,7 +160,7 @@ the items that matter most to you can be prioritized.| Capability | Open Source (AGPL) | Commercial License ||------------|:------------------:|:------------------:|-| Five intelligence layers | ✅ | ✅ |+| Five foundational intelligence layers plus derived ones (change risk, tests, bug history, security, dead code) | ✅ | ✅ || Ten task-shaped MCP tools (plus `list_repos`) | ✅ | ✅ || Multi-repo workspaces | ✅ | ✅ || Full-tier language support (incl. C# / .NET) | ✅ | ✅ |diff --git a/docs/design/theme-tokens.md b/docs/design/theme-tokens.mdindex f5e70bdbc..de72e46e4 100644--- a/docs/design/theme-tokens.md+++ b/docs/design/theme-tokens.md@@ -1,40 +1,36 @@-# the theme system — Design Tokens & Contrast Matrix (Phase 0)+# Design Tokens and Contrast Matrix-> Source of truth for the redesign. These are the **resolved** semantic token-> values — adjusted from the plan's design-intent table to pass the WCAG-> contrast floors. Run `python3 docs/design/contrast_check.py` to regenerate-> the matrix; it exits non-zero if any required pair regresses (CI gate).+> The **resolved** semantic token values for the product UI, tuned to pass the+> WCAG contrast floors below. Run `python3 docs/design/contrast_check.py` (a+> wrapper for `packages/ui/scripts/contrast-check.py`) to regenerate the+> matrix; it exits non-zero if any required pair regresses.## Floors (both light + dark)-- Body text on primary surfaces — **≥ 7.0** (AAA)-- Secondary / large / status text, accent-on-surface — **≥ 4.5** (AA)-- Interactive / non-text UI (active borders, hints) — **≥ 3.0**+- Body text on primary surfaces: **≥ 7.0** (AAA)+- Secondary / large / status text, accent-on-surface: **≥ 4.5** (AA)+- Interactive / non-text UI (active borders, hints): **≥ 3.0**-## §6 decisions (resolved)+## Theme decisions-1. **Default theme — Dark; explicit two-state toggle.** Preserves current- product behavior; Light is opt-in via the toggle. No "System" option —- product decision (2026-06-04): with the OS in dark mode System and Dark- looked identical and read as redundant; the choice stays explicit.- (`next-themes` `defaultTheme="dark"`, `enableSystem={false}`; the shared- ThemeToggle migrates stale persisted `"system"` values to dark.)+1. **Default theme: Light, with an explicit two-state toggle.** Dark is opt-in+ via the toggle and persists across reloads. There is no "System" option, so+ the choice stays explicit. (`next-themes` `defaultTheme="light"`,+ `enableSystem={false}`; the shared ThemeToggle migrates stale persisted+ `"system"` values to light.)2. **Light-mode accent strategy.** The brand accent stays repowise orange- `#f59520` — it's `--color-accent-fill` (CTAs, active fills, focus ring) in+ `#f59520`: it is `--color-accent-fill` (CTAs, active fills, focus ring) in**both** modes. Only accent *text/icons/links* darken to `#A16215`(`--color-accent-primary`) in light so they clear AA on white **and** warmpaper (bright orange can't be AA as small text). Plum(`--color-accent-secondary`) is used for links-alt **and** data-viz.3. **Gradient intensity.** Restrained in-product (active states, small accentwashes); bold in marketing (full hero meshes).-4. **Dark-mode base — plum-tinted charcoal.** Dark surfaces carry a restrained- violet cast (`#17131d`→`#322a3e`, text `#eeeaf4`/`#a79db3`, lavender-alpha- borders) so the product is visually distinct from the crowd of- neutral-charcoal + orange UIs (user feedback: pure charcoal read "exactly- like Claude"). The tint is the theme system's own plum family at low- saturation; the orange accent still pops. **Light mode is unchanged**- (warm paper). *(Supersedes the earlier "neutral charcoal, Dark Orange- UI-kit" decision.)*+4. **Dark-mode base: plum-tinted charcoal.** Dark surfaces carry a restrained+ violet cast (`#17131d`->`#322a3e`, text `#eeeaf4`/`#a79db3`, lavender-alpha+ borders) so the product is visually distinct from neutral-charcoal + orange+ UIs. The tint is the theme's own plum family at low saturation; the orange+ accent still pops. **Light mode** uses warm paper.### The fill-vs-text split (why two accent tokens)@@ -47,7 +43,7 @@ A single accent can't be both AA-as-text on white *and* vivid-as-a-CTA. So:| `--color-text-inverse` | `#FFFFFF` | `#17131D` | text on an `accent-primary` fill || `--color-text-on-accent` | `#241B2C` | `#17131D` | text on the brand `accent-fill` |-White on the bright `#f59520` fill is only ~2.0:1 — so CTAs use **dark** text+White on the bright `#f59520` fill is only ~2.0:1, so CTAs use **dark** text(`text-on-accent`, 7.26:1). Existing `bg-accent-primary text-text-inverse`fills keep working because `text-inverse` flips per theme (white in light onthe deep `#A16215` = 4.91:1; near-black plum in dark on bright `#F59520` = 8.02:1).@@ -107,7 +103,7 @@ the deep `#A16215` = 4.91:1; near-black plum in dark on bright `#F59520` = 8.02:```Gradients are for hero washes, primary CTAs, brand marks, empty-state art, and-selected/active accents — never behind body text. Text over gradients uses+selected/active accents, never behind body text. Text over gradients uses`--color-text-on-accent` / `--color-text-inverse` with a verified ≥4.5:1 floor.## Community families (graph clustering palette)@@ -122,7 +118,7 @@ a softer `-soft` satellite tint (leaf files). The graph cycles `community_id %These are tokens (`--color-community-1..12` + `-soft`) in `globals.css`, resolvedat runtime by `getCommunityFamily` / `useCommunityFamilies` in`shared/use-theme-tokens.ts` (the canvas can't resolve `var()`, so it reads the-computed token per theme and repaints on theme flip — same mechanism as Mermaid+computed token per theme and repaints on theme flip, the same mechanism as Mermaid/ C4 / `THEME_COLORS`).**Usage rule:** module/centroid nodes use `hub`; file/leaf nodes use the@@ -145,10 +141,10 @@ crossCommunity = plum, internal = sage/green), mirroring `lib/confidence.ts`.| 11 | Deep teal | `#2F6B66` / `#558F89` | `#6FB3AB` / `#4A8780` || 12 | Charcoal mauve | `#5E5360` / `#84778A` | `#A79DB3` / `#786F84` |-Hub-on-canvas contrast is gated at **≥ 3.0:1** (non-text UI) in both modes —+Hub-on-canvas contrast is gated at **≥ 3.0:1** (non-text UI) in both modes;families #1 (orange) and #10 (peach) were deepened in **light** to clear the-floor on the warm paper canvas (`#F4EAE1`): orange `#F59520 → #C0641A`-(1.92 → 3.49), peach `#D9825F → #B85A38` (2.42 → 3.88). The `-soft` satellites+floor on the warm paper canvas (`#F4EAE1`): orange `#F59520 -> #C0641A`+(1.92 -> 3.49), peach `#D9825F -> #B85A38` (2.42 -> 3.88). The `-soft` satellitesare not gated (they sit behind hubs, never alone as the only signal).## Contrast matrix (generated)diff --git a/docs/layers/BUG_HISTORY.md b/docs/layers/BUG_HISTORY.mdindex da98e2eb2..18f14cf17 100644--- a/docs/layers/BUG_HISTORY.md+++ b/docs/layers/BUG_HISTORY.md@@ -1,333 +1,180 @@# Bug-fix history-Repowise records which files and which functions have actually been fixed, how-recently, and how often. The output is a small set of honest numbers: a-per-file fix count over a trailing 180-day window, a per-symbol breakdown, a-decayed "bug magnet" flag, and the age of the most recent fix. They surface in-the pre-edit hook, in `get_risk` and `get_change_risk` over MCP, in the health-drawer, and on the symbols views.+Repowise records which files and which functions have been fixed, how+recently, and how often. You get a per-file fix count over a trailing 180-day+window, a per-symbol breakdown, a decayed **bug magnet** flag, and the age of+the most recent fix.-**No LLM calls.** This is git history, a `-U0` diff pass, and arithmetic. It-runs inside the existing git phase of `repowise init` and `repowise update`.+Churn says a file changes a lot, which is often just where the work is. Fix+history says it breaks a lot, which is more actionable. It does not name the+commit that introduced a bug (see [Accuracy and limits](#accuracy-and-limits)).-The reason to care: a file's churn tells you it changes a lot, which is often-just where the work is. A file's fix history tells you it *breaks* a lot, which-is a different and more actionable thing. On this repo the top bug magnets-(`pipeline/persist.py`, `pipeline/incremental.py`, `dead_code/analyzer.py`,-`update_cmd/command.py`) match the churn hotspot list, arrived at independently-from bug fixes rather than commit counts.+No LLM key. It is git history, a `-U0` diff pass and arithmetic, run inside the+git phase of `repowise init` and `repowise update`.-## Fix-shape filtering: what counts as a bug fix+## Quick start-A commit whose subject matches the fix keywords is a candidate, not a fact.-Plenty of `fix:` commits fix a typo in the README, tighten a test, or bump a-version. Counting those inflates every downstream number.--So each candidate's `-U0` diff is classified into one shape, and only one shape-counts:--| Shape | Counted | Example |-|---|---|---|-| `code_fix` | yes | production source lines changed |-| `test_only` | no | only files under `tests/` or matching a test naming rule |-| `doc_only` | no | only `.md` / `.mdx` / `.rst` / `.txt` / `.adoc` or `docs/` |-| `config_other` | no | only lockfiles, CI YAML, `*.config.*`, dotfile rc |-| `comment_only` | no | code files touched, but the changed lines are comments or docstrings |-| `empty` | no | merge commits and no-op diffs |--The comment-only rule is line-level, not path-level: it strips trailing comments-and compares the code part of each removed and added line, so a commit that-rewords a docstring in a `.py` file does not count as a bug fix.--The classifier was validated against 240 hand-labelled fix commits across four-repos (repowise, flask, django, zod) and agrees with the labels **98.3%** of the-time (repowise, flask and django all 60/60; zod 56/60). The four residual-disagreements are all the same shape, "a file with a code extension changed but-no product code did", and were left as honest disagreements rather than closed-with repo-specific ignore lists.--The effect is real and varies enormously by repo. On flask, **58.7%** of-keyword-matched fix commits are noise and filtering removes half the raw-attributions. On repowise it is 7.0%. That is the point: a number that means the-same thing in both repos is worth more than a bigger number in one.--`GitMetadata` stores both `prior_defect_count` (filtered) and-`prior_defect_raw_count` (unfiltered), so the delta stays inspectable.--One thing filtering deliberately did not do: change the health score.-Re-running the defect-weight calibration on the 21-repo corpus showed filtering-moves pooled OOF AUC by +0.0002, inside noise. The `prior_defect` marker stays-at weight 1.0. Phase 1 shipped for count honesty, not for a score change, and-the doc says so rather than implying a win that was not measured.--## The `fix_events` table--Each counted fix produces one row per file it touched:+Nothing to turn on: index the repo and the counts are there.+```bash+repowise init+repowise risk --target src/pipeline/persist.py # history for one file```-fix_events(repository_id, fix_sha, file_path, shape_kind,- old_ranges_json, fixed_at, ...)-```--`old_ranges_json` holds the line ranges the fix replaced, numbered on the fix's-own parent commit. Persisting the events, rather than only a count, is what-makes per-symbol attribution and changed-line overlap possible later without-re-walking git.--The table is populated during a full index and extended incrementally by-`repowise update`. A pre-existing database picks the new columns up through the-additive reconciler with no migration code; rollups sit at their defaults until-the next update.--## Per-symbol attribution--After both `fix_events` and `wiki_symbols` are persisted, each event's replaced-line ranges are matched against the symbol spans in that file, producing a-`fix_symbol_counts` map per file: which function or class absorbed each fix.--Measured attribution rate, walking the shipped tracer live and joining against-each repo's existing index:--| Repo | Events | In-envelope rate |-|---|---|---|-| django | 2,179 | 97.4% |-| flask | 125 | 96.3% |-| zod | 337 | 81.8% |-| repowise | 1,145 | 96.8% |-"In-envelope" means the event's lines land somewhere between the file's first-and last indexed symbol. Gaps *between* symbols still count as misses, so this-remains a test of the join rather than a restatement of it. The two causes of-the loose-denominator misses were measured, not argued: django and flask are-line drift (symbol spans are at HEAD, ranges are at each fix's parent, so a 2011-fix is matched against lines that moved a decade ago, and 99.8% of django's-events are 3+ years old), and zod is symbol density in test files the parser-indexes with one or two named symbols.+From an agent:-**The counts are approximate and every surface says so.** Symbol spans are-current-tree, fix ranges are historical. Read them as "mostly here", not as an-exact ledger.--## The bug magnet rollup+```python+get_risk(targets=["src/pipeline/persist.py"])+# defect_profile: {fix_count, window, last_fix_days_ago, bug_magnet, top_symbols}+```-Each file's counted fixes are collapsed into one decayed mass:+When an agent is about to edit a file with a recent run of fixes, the pre-edit+hook adds one line to its context:```-mass = Σ 0.5 ^ (age_days / 90)-bug_magnet = mass >= 3.0+[repowise] pipeline/persist.py has been bug-fixed 5x in the last 6 months,+last 2 weeks ago (bug magnet); mostly in run_pipeline.```-The 90-day half-life was picked from a sweep of 60 / 90 / 180 days against the-21-repo calibration corpus. Decay was the only lever that moved the-`prior_defect` coefficient at all (0.15 undecayed to 0.23 decayed); 90d and 180d-tied inside noise and 90d is the tighter of the two.+In the dashboard, look at the health drawer's **Bug history** section and the+**Bug-fixed** facet on the Symbols page.-Read the threshold with the decay in mind. Only a same-day fix is worth a full-1.0, so three real fixes spread over a couple of weeks land near 2.9 and do-**not** flag. In practice the flag needs four recent fixes, or three very recent-ones. That is deliberately the conservative end: the flag exists to interrupt-someone mid-edit, so it should be rare enough to be worth reading.+## Reading the results-The rollup recomputes the whole repo on every index and update rather than-patching the files an update touched, because decay ages a rollup with nothing-in the file changing. It costs about 0.25s on this repo (1,172 events, 739-files): two indexed queries and arithmetic.+| Field | Meaning |+|---|---|+| `fix_count` | Counted bug-fix commits that touched the file in the last 180 days |+| `last_fix_days_ago` / `last_fix_at` | Age of the most recent counted fix |+| `bug_magnet` | The decayed fix mass is 3.0 or more. Present only when true |+| `top_symbols` | Up to three functions or classes that absorbed the most fixes. Approximate |-## The recency contract+The whole `defect_profile` block is omitted for a file with no counted fixes,+and on a repo with no fix data. Absence means no counted fix, not a clean bill+of health (see the limits below).-This is the rule that keeps the feature honest, and it is enforced in the type-definitions rather than left to each surface:+### What counts as a bug fix-> `bug_magnet` is the decayed fix mass past its trigger, so it is a recency-> claim. Any copy that shows it must show `last_fix_at` too.+A commit whose subject matches the fix keywords (`fix`, `bug`, `patch`,+`resolves`, `closes #N`, `fixes #N`, and none of `docs`, `typo`, `bump`,+`deps`, `chore`, `lint`, `format`, `style`) is a candidate. Its `-U0` diff is+then classified, and only `code_fix` counts:-Two consequences:--1. **A magnet claim never appears without its age.** If `last_fix_at` is null,- the flag is dropped rather than shown unanchored. Pre-merge review caught- three surfaces violating this (the health panel badge, the VS Code hover, and- the MCP `defect_profile`); all three now drop the flag, and there is a test- on the null-timestamp path.-2. **The surface goes silent outside the window.** The pre-edit hook does not- fire at all when the last fix is older than 180 days, no matter how large the- historical count is. "This file was fixed nine times, in 2019" is not a- warning, it is trivia.--A related bug worth knowing about, since it is the kind that hides: `last_fix_at`-was originally serialized without a timezone. SQLite's DATETIME bind processor-drops `tzinfo`, so `.isoformat()` produced a string JavaScript parses as local-time. West of UTC, a fix committed two hours ago landed in the future, tripped-the relative-time future-guard, and the age silently vanished from all four-surfaces required to carry it, worst for the freshest fixes. Fixed once-upstream, not four times in the consumers.+| Shape | Counted | Example |+|---|---|---|+| `code_fix` | yes | production source lines changed |+| `test_only` | no | only test files |+| `doc_only` | no | only `.md` / `.mdx` / `.rst` / `.txt` / `.adoc` or `docs/` |+| `config_other` | no | only lockfiles, CI YAML, `*.config.*`, dotfile rc |+| `comment_only` | no | code files touched, but only comments or docstrings changed |+| `empty` | no | merge commits and no-op diffs |-## Where it surfaces+The comment-only rule is line level: a commit that rewords a docstring in a+`.py` file is not a bug fix. Both the filtered count (`prior_defect_count`)+and the raw keyword count (`prior_defect_raw_count`) are stored, so the+difference stays inspectable.-### Pre-edit hook+### The bug magnet flag-When an agent is about to edit a file with a real recent run of fixes, one line-goes into its context:+Each file's counted fixes collapse into one decayed mass:```-[repowise] pipeline/persist.py has been bug-fixed 5x in the last 6 months,-last 2 weeks ago (bug magnet); mostly in run_pipeline.-```--Three gates, all of which must hold: at least 3 counted fixes, a last fix inside-180 days, and one claim per file per session (an atomic ledger, so two racing-hook processes cannot double-fire). A file that also has a governing-architectural decision emits two lines, never more. Wired into both the Claude-Code and Codex hook paths. See [docs/agent/HOOKS.md](../agent/HOOKS.md).--### `get_risk`--Files with counted fixes gain a `defect_profile` block:--```python-get_risk(targets=["packages/core/src/repowise/core/pipeline/persist.py"])-# defect_profile: {fix_count, window, last_fix_days_ago,-# bug_magnet, top_symbols}+mass = sum(0.5 ^ (age_days / 90))+bug_magnet = mass >= 3.0```-`bug_magnet` appears only when true, `top_symbols` is the top three with the-path prefix stripped, and the whole block is omitted on repos with no fix data.-Measured at 46 tokens mean, 80 worst case, against a 150-token ceiling across-803 files. The approximation caveat lives in the tool docstring, stated once,-rather than as a constant string repeated on every row.--Shipping `defect_profile` also let the old keyword `risk_type="bug-prone"`-derivation retire. It used to scan commit subjects for `/fix|bug|patch/`, which-counted doc and test commits, had no recency, and disagreed with the count every-other surface showed. It now reads `bug_magnet` or `prior_defect_count >= 3`.-That does silently reclassify some files: a high-traffic file with 3 fixes in-200 commits now reads `bug-prone` where it read `churn-heavy`.--### `get_change_risk`--A diff or commit range gains a `fix_history.overlap` block: per changed file, how many-past fixes it carries and how many of the change's lines fall inside a past-fix's replaced ranges. The per-file counts are exact (counted by-`COUNT(DISTINCT fix_sha)`, so one commit fixing three files reports one fix, not-three). The line overlap is labelled `approximate` in the payload, because past-ranges are numbered on their own parent commit. Files are ranked by overlap then-count, and the list carries a `truncated` flag rather than silently dropping the-tail.--See [CHANGE_RISK.md](CHANGE_RISK.md) for the rest of that tool.--### Dashboard--- **Health drawer**: a collapsed **Bug history** section with per-symbol counts,- the last-fix age on the toggle, and an approximation footnote.-- **File signals panel**: the bug-fix tile gains a **Bug magnet** badge and- folds the last-fix age into its caption.-- **Symbol detail** (modal and routed page): a **Bug fixes** stat tile beside- **Modifications**. How often it changes next to how often it breaks is the- contrast that earns the cell.-- **Symbols list**: a **Bug-fixed** facet and a per-row fix chip. The filter is- symbol-level, unlike the file-level filters beside it, because a bug-fixed- file says very little about the one function you are reading.-- **VS Code**: one line on the file hover, off the same signals block.--See [docs/start/DASHBOARD.md](../start/DASHBOARD.md) for where each of these-lives.--## What was built, measured, and deliberately not shipped--The most useful thing this layer can tell you is what it refuses to tell you.--### SZZ inducing-commit attribution: built, measured, deleted--The obvious next step from "this line was fixed" is "and here is the commit that-introduced the bug". That is the SZZ algorithm, and it was fully built: a blame-pass walking each fix's replaced lines back through history, refactor-aware, with-overlap-based ranking of the candidates.--Then it was measured against a hand-labelled set. Fifty-three judged rows, five-reviewers. **File-level SZZ tops out at 74.5% top-candidate precision on this-corpus. The gate was 80%.**--74.5% is not a bad algorithm. It is a *good* SZZ implementation. It is also not-good enough to put a commit SHA and a person's name in a UI, because one time in-four the UI would be accusing the wrong commit, and there is no way for the-reader to tell which time. So the blame pass was deleted, `szz.py` is gone, and-**no repowise surface names an inducing commit.** The planned-"(last: `<sha>` `<subject>`)" tail on `get_change_risk` was cut. The planned-inducing-commit link on commit rows was cut.--Deleting it was verified rather than assumed: `inducing_shas_json` appeared in-exactly four places in the tree (the model, the migration, the writer, and the-writer's tests), and neither the hosted backend nor the PR bot referenced-`fix_events` in any spelling. There was no consumer to break. Removing the pass-also gave back 9.1s of index time on this repo and 5.9s on zod.--The findings are kept here so they do not vanish with the code, because they are-reusable by anyone attempting this again:--- **Refactor-aware blame is worth building.** Walking through- behaviour-preserving moves to the parent lifted strict precision from 70.6% to- 74.5%. All 14 of the initial false calls were refactors inheriting moved lines.-- **Overlap ranking beats earliest-commit ranking by 12 points** on judged rows.- Earliest-commit's apparent 80% was an artefact of its wrong answers going- unjudged: 7 of its 16 unknowns were this repo's initial commit, and all 7 came- back "not plausible" once judged. This is the finding most likely to be- re-litigated, so, plainly: it was tested twice, and it settled the other way- from the first read.-- **The two residual failure modes are properties of SZZ, not bugs.** An initial- import has no earlier commit to name, and package splits that move lines- wholesale while re-sorting imports defeat the carried-through test on exactly- the lines that moved.--Restoring the tracer is a `git revert` of one commit, not a re-derivation. If-someone finds a ranking that clears 80% on the frozen label set, the data layer-is ready for it.--### "AI vs human introduced bugs": cut--A repo-level card comparing defects introduced by AI-authored commits against-human-authored ones, per KLOC. It is an obvious feature to want given that-repowise already detects agent provenance on commits, and it was planned.--It was gated on a concentration check before build, and the check failed. **The-top 5 inducing commits held 30.3% of all traced mass against a 20% ceiling, with-the initial import alone at 13.9%.** SZZ error does not average out: when blame-concentrates on a handful of enormous commits, any aggregate built on inducing-commits inherits that concentration, and the aggregate is then mostly a-statement about who happened to author the repo's five biggest commits. A stat-like "AI introduces N% more bugs" is exactly the kind of number that gets-screenshotted and quoted, which raises the bar rather than lowering it.--So it was killed before it was built, and the provenance join it depended on was-cut with it.--### What survived the cuts--Counts and recency survived. Accusations did not. Everything shipped here-answers "this code has been broken before, recently, and roughly here", and-nothing answers "and it was your fault". The first question is answerable from-git with measurable accuracy. The second is not, yet.--## Limitations--- **Symbol counts are approximate.** Current-tree spans against historical line- ranges. Stated on every surface that shows them.-- **Keyword-matched fix detection.** A bug fixed in a commit whose subject does- not say so is invisible. The shape filter improves precision, not recall.-- **The 180-day window is fixed.** A file fixed heavily two years ago and quiet- since reads as clean, which is intentional but is a choice, not a truth.-- **Workspace lookup in the hook does not filter by `repository_id`,** matching- the co-change query beside it. In a workspace whose members lack their own- `.repowise` this degrades to silence rather than a wrong answer.-- **The hook's usefulness is unproven.** It shipped without the planned week of- live-session dogfooding, so whether the notice changes behaviour or becomes- ignorable noise is an open question. If it is noise, the plan is to cut or- reword the hook surface and keep every data layer beneath it.+Only a same-day fix is worth a full 1.0, so three fixes spread over a couple of+weeks land near 2.9 and do not flag. In practice the flag needs four recent+fixes, or three very recent ones. It is meant to interrupt someone mid-edit, so+it is kept rare. The rollup is recomputed for the whole repo on every index and+update, because decay ages it even when nothing in the file changes.++### Recency is part of the claim++`bug_magnet` is a recency claim, so every surface that shows it also shows the+last-fix age. When `last_fix_at` is missing, the flag is dropped instead of+being shown without an age. The pre-edit hook stays silent when the last fix is+older than 180 days, however large the old count.++### Per-symbol counts++Each fix's replaced line ranges are matched against the symbol spans in the+file, which gives the per-function breakdown. Spans come from the current tree+and fix ranges from each fix's parent commit, so read the result as "mostly+here", not as an exact ledger. Every surface that shows symbol counts says so.++### On a change++`get_change_risk` adds a `fix_history.overlap` block: for each changed file,+how many past fixes it carries and how many of the change's lines fall inside a+past fix's replaced ranges. Fix counts are exact (one commit fixing three files+is one fix). Line overlap is marked `approximate`. Files are ranked by overlap+then count, and a `truncated` flag marks a cut list. See+[CHANGE_RISK.md](CHANGE_RISK.md).++## Tuning and suppressing++There are no settings for this layer. The 180-day window, the 90-day half-life+and the 3.0 threshold are fixed.++The pre-edit hook fires only when all three hold: at least 3 counted fixes, a+last fix within 180 days, and no earlier notice for that file in the same+session. A file with a governing architectural decision as well gets two lines,+never more. To turn hooks off, see [HOOKS.md](../agent/HOOKS.md).++## Accuracy and limits++- **Fix-shape classifier**: agrees with hand labels on 98.3% of 240 fix commits+ across repowise, flask, django and zod (60/60 on the first three, 56/60 on+ zod). The four misses are all "a code-extension file changed but no product+ code did".+- **Noise removed varies by repo**: 58.7% of keyword-matched fix commits on+ flask are not code fixes; 7.0% on repowise.+- **Per-symbol attribution** lands inside the indexed symbol range for 97.4% of+ django's 2,179 events, 96.3% of flask's 125, 96.8% of repowise's 1,145 and+ 81.8% of zod's 337. Old fixes drift (symbol spans are current); zod's misses+ are test files with few named symbols.+- **No inducing commit is named.** File-level SZZ reached 74.5% top-candidate+ precision on 53 hand-judged rows, below the 80% bar for putting a commit and+ an author on screen. Counts and recency ship; blame does not.+- **Recall is keyword bound.** A fix whose subject does not say so is+ invisible. The shape filter improves precision only.+- **The window is fixed.** A file fixed heavily two years ago and quiet since+ reads as clean.+- **The health score does not change with filtering**: re-running the defect+ calibration with filtered counts moved pooled AUC by +0.0002.+- **In a workspace**, the hook's lookup does not filter by repository; where+ member repos lack their own `.repowise`, it falls silent.+- **Whether the hook notice changes agent behaviour has not been measured.**++Benchmark method for the health model that reads these counts:+[docs/BENCHMARKS.md](../BENCHMARKS.md#code-health-predicts-defects). The+measured SZZ attempt is written up in+[architecture/bug-history.md](../architecture/bug-history.md).++## Where it shows up++- **Agent hook**: the pre-edit notice, in both the Claude Code and Codex hook+ paths.+- **MCP**: `defect_profile` on `get_risk`; `fix_history.overlap` on+ `get_change_risk`.+- **CLI**: `repowise risk --target FILE`.+- **Dashboard**: the health drawer's collapsed **Bug history** section+ (per-symbol counts, last-fix age); the file signals panel's **Bug magnet**+ badge; a **Bug fixes** tile beside **Modifications** on symbol detail; a+ **Bug-fixed** facet and per-row chip on the Symbols list. See+ [DASHBOARD.md](../start/DASHBOARD.md).+- **VS Code**: one line on the file hover.++## Reference++MCP schemas: [`get_risk`](../agent/MCP_TOOLS.md#get_risk) and+[`get_change_risk`](../agent/MCP_TOOLS.md#get_change_risk). Hook behaviour:+[HOOKS.md](../agent/HOOKS.md).## See also-- [CODE_HEALTH.md](CODE_HEALTH.md): the `prior_defect` marker and the- "does the score find the bugs?" self-check that reads the same counts.-- [CHANGE_RISK.md](CHANGE_RISK.md): `get_change_risk` and its scoring.-- [docs/agent/MCP_TOOLS.md](../agent/MCP_TOOLS.md): `get_risk` and- `get_change_risk` schemas.-- [docs/agent/HOOKS.md](../agent/HOOKS.md): the pre-edit hook.+- [architecture/bug-history.md](../architecture/bug-history.md): the+ `fix_events` table, the rollup, and the SZZ findings.+- [CODE_HEALTH.md](CODE_HEALTH.md): the `prior_defect` marker that reads the+ same counts.+- [CHANGE_RISK.md](CHANGE_RISK.md): `fix_history` on a change.diff --git a/docs/layers/CHANGE_RISK.md b/docs/layers/CHANGE_RISK.mdindex bed24ae4c..5af43223f 100644--- a/docs/layers/CHANGE_RISK.md+++ b/docs/layers/CHANGE_RISK.md@@ -1,538 +1,339 @@# Change risk (`repowise risk`)-`repowise risk` reports on a **change** (a commit or a `base..head` range): the-bug-fix history of the files it touches, and the shape of its diff. It is a-just-in-time / pre-merge signal, complementary to `repowise health`, which-scores files rather than changes.--**Lead with `risk_percentile` and `classification`.** They are the benchmarked,-population-relative authority for live change review. `fix_history` is-complementary evidence about where the change lands. The supporting 0–10 score-measures how big and spread out a change is; it is not a probability. See-[What the score does and does not buy](#what-the-score-does-and-does-not-buy).--Everything below describes diff shape, which is what `repowise risk` reports.-The `get_change_risk` MCP tool leads instead with what the change newly made-worse — it compares the health of both revisions and keeps diff shape as-supporting context. See [MCP tools](../agent/MCP_TOOLS.md#get_change_risk).+`repowise risk` reports on a **change**: a commit, a `base..head` range, or+your uncommitted work. It tells you two things. Where the change sits in the+repository's own distribution of recent commits, by size and spread. And+whether the files it touches have broken before. It scores changes; for scores+on files, use `repowise health`.-```bash-repowise risk # score uncommitted work, else HEAD-repowise risk HEAD # score the last commit-repowise risk abc123 # score a single commit-repowise risk main..HEAD # score a branch / PR range as one change-repowise risk main..HEAD --ext .py # count only .py files-repowise risk main..HEAD -x 'tests/' -x '*.spec.ts' # omit matching paths-repowise risk --format json # machine-readable-```+It needs no LLM key, no network and no index: it runs on `git` plus constants+learned offline. A typical run is one `git log` walk and a sample of recent+commits. Branch overlap and the independent-changes report use the index when+there is one.-It runs in-process: pure `git` + learned constants. **No LLM, no network, and no-blame at runtime**: SZZ labelling lives entirely in the offline calibration.+Lead with `risk_percentile` and `classification`. They are the+population-relative signal for review. `fix_history` is separate evidence about+where the change lands. The 0-10 score is a supporting measure of diff size and+spread, not a probability (see [Accuracy and limits](#accuracy-and-limits)).-## What gets scored+## Quick start-With **no revspec** the subject is the change in front of you: your uncommitted-work (staged, unstaged and untracked) if the tree is dirty, otherwise `HEAD`.-The payload sets `working_tree: true` when it took that path, and the CLI says-so. Naming a revspec — `HEAD` included — always means committed refs.+```bash+repowise risk # uncommitted work, else HEAD+repowise risk HEAD # the last commit+repowise risk main..HEAD # a branch or PR range, as one change+repowise risk main..HEAD --ext .py # count only .py files+repowise risk main..HEAD -x 'tests/' -x '*.spec.ts'+repowise risk --format json+repowise overlap # other branches editing the same files+```-A **merge commit** is scored for the diff it brought onto its first parent, so a-merged PR reads as its own content rather than as an empty change.+From an agent:-## Excluding paths+```python+get_change_risk() # the working tree or HEAD+get_change_risk(revspec="main..HEAD")+get_risk(changed_files=["src/api/routes.ts", "src/middleware/cors.ts"]) # PR mode+```-Use repeatable `--exclude` / `-x` flags with gitignore-style patterns to omit-files from a score. The same filters apply to the requested change and the-recent commits sampled for its percentile, so the comparison remains like for-like. Put project-wide, risk-only rules in a repository-root `.riskignore`;-those patterns apply automatically and are combined with any command-line-flags. For example, `tests/` excludes that directory recursively, while-`test_*.py` excludes matching test filenames anywhere in the repository.+`get_change_risk` leads with what the change newly made worse: it compares the+health of both revisions and keeps diff shape as supporting context. In the+dashboard, the Commits page leads its table with review priority.-## In CI+### What gets scored-`--fail-above-percentile P` turns the command into a gate on-`risk_percentile`: it ranks the diff-shape score against recent commits, and-the absolute 0-10 value never decides.+With no revspec the subject is your uncommitted work (staged, unstaged and+untracked) when the tree is dirty, otherwise `HEAD`. The payload sets+`working_tree: true` when it took that path. Naming a revspec, `HEAD`+included, always means committed refs. A merge commit is scored for the diff it+brought onto its first parent, so a merged PR reads as its own content.-```bash-repowise risk --fail-above-percentile 95 --format github-repowise risk origin/main...HEAD --fail-above-percentile 95 --format markdown-```+## Reading the results-With the flag, or with `--format markdown` or `github`, and no revspec, the-subject is the CI change: the pull request's target branch (from the CI's-variables, else the remote's default branch) `...HEAD`. A CI checkout of a-pull request is a merge commit, so `HEAD` alone would be the wrong change.-Otherwise the default stays "uncommitted work, else `HEAD`".--Exit codes are the shared CI ones: `0` at or below the percentile, `1` above-it, `2` when the gate cannot evaluate. A change with no percentile cannot be-gated: `--baseline 0` turns ranking off, and a shallow clone with fewer than 8-recent commits has nothing to rank against, so fetch full history. At a-percentile P roughly (100 - P)% of changes fail by construction; a failure asks-for a split or a second reviewer, not a code fix.--The table prints the verdict under the review priority. `json` adds-`"gate": {"fail_above_percentile": P, "percentile": <unrounded>, "status":-"pass" | "fail"}`; `percentile` is what the gate compared, since-`risk_percentile` is rounded. `markdown` leads with the verdict ("**Above the-95th percentile gate**: larger and more spread out than 97% of recent-commits"), then the range, the sample, the lines, files and directories, the-fix-history rank, and the files with bug-fix history. `github` writes an-`::error::` when the gate fails (a `::notice::` otherwise) plus the markdown in-the job summary. `markdown` and `github` are not for `--target`, which stays-`table` or `json`.--A revspec git cannot read (an unknown revision, not a repository, a missing-merge-base) exits `2` on every revspec run, gated or not. It used to exit `1`,-which a pipeline would have read as a failed gate.--The GitHub Action runs it as the `risk` check (input-`risk-fail-above-percentile`; empty reports the rank without gating), and the-GitLab template as the `repowise-risk` job when-`REPOWISE_RISK_FAIL_ABOVE_PERCENTILE` is set. See [CI](../start/CI.md).+Read the result in this order:++| Field | What it is | Act on it |+|---|---|---|+| `review_priority` / `classification` | Tercile of the percentile: `low` / `moderate` / `high`, labelled Below typical / Typical / Elevated | Yes: the review signal |+| `risk_percentile` | Where this change's diff shape ranks among the repo's recent commits (0-100) | Yes |+| `fix_history` | Prior bug fixes on the touched files, ranked against recent commits | Yes: where to look |+| `score` | 0-10 diff size and spread, on a single-commit scale | Supporting only |+| `fallback_band` | `low` / `moderate` / `high` from the score | Only present when there was nothing to rank against |+| `independent_changes` | Groups of changed files that nothing links | Consider splitting |-## Fix history: where the change lands+### Fix history: where the change lands-The first block in the result is `fix_history`, and it is the one to act on. It-answers a question the diff shape cannot: **have these files broken before?**+`fix_history` answers a question diff shape cannot: have these files broken+before?```These files have broken before · 82nd percentile of this repo's recent commits-┌──────────────────────────────────────────┬───────┬─────────────┐-│ File │ Lines │ Prior fixes │-├──────────────────────────────────────────┼───────┼─────────────┤-│ core/pipeline/persist.py │ 40 │ 21.6 │-│ cli/commands/update_cmd/command.py │ 6 │ 19.3 │-└──────────────────────────────────────────┴───────┴─────────────┘+File Lines Prior fixes+core/pipeline/persist.py 40 21.6+cli/commands/update_cmd/command.py 6 19.3```-- **Prior fixes** is a count of bug-fix commits that previously touched that- file, **recency-weighted against the change's own date**: a fix from a year- earlier counts a half, from two years a quarter. So the number is- "recent-equivalent fixes", not a raw tally, and a file that broke constantly- and then settled decays away. Anchoring to the change rather than to today- means the same commit scores the same on every re-run.-- **`density`** is the churn-weighted mean of those per-file numbers. Weighting- by churn means the file a change mostly edits dominates the answer rather than- a one-line drive-by next door. It is a *ratio*, so unlike the score it does- not grow with the size of the diff: one line in a file fixed twenty times- outranks a thousand lines in files never fixed at all.-- **`percentile`** ranks that density against the same measure over the- repository's own recent commits, since a bare "3.4 decayed fixes" means- nothing on its own. Ranking against whole commits rather than against- individual per-file numbers is what keeps it readable: a change spread over- several files averages below any single hot file, so a per-file population- pinned every multi-file change to the bottom. Commits that touch no- fix-bearing file stay in the population; they are a legitimate answer to how- much fix pressure a change here usually stands on. It is `null` when the- change touches no fix history, or when fewer than eight sampled commits are- available to rank against.--This comes from one `git log` walk (up to 20 000 commits, memoized per-repository state) using the same bug-fix classifier the indexer uses. It needs-no index, no database and no coverage data, so it is available on a repository-`repowise` has never indexed.--Fix history is read from **before** the change being scored: a commit is ranked-against the fixes that had already landed when it was written, never against-fixes it caused. For a range, the record is read at the fork point the diff-starts from, not at the base branch's current tip.--> **Caveat, stated rather than buried.** The bug-fix classifier is keyword-based-> and shared with the indexer, so `fix_history` under-reports in two ways.->-> It matches `fix`, `bug`, `patch`, `resolves`, `closes #N`, `fixes #N` — and-> misses conventions outside that set. Django's `Fixed #12345` is the notable-> one: on a 4 000-commit sample it classifies 5 commits as fixes where roughly-> 1 800 use that prefix. (Django also uses `Fixed #N` for features, so the-> subject line alone cannot separate the two — which is why the classifier has-> not simply been widened.)->-> It also **excludes** any subject containing `docs`, `typo`, `bump`, `deps`,-> `chore`, `lint`, `format` or `style`. That keeps cosmetic commits out, but-> drops genuine fixes like "fix: docs build crash" with them.->-> Where the classifier fires the ranking is good; where a project's convention-> falls outside it, `fix_history` reads lower than the truth.--## What the diff-shape score measures--The model uses Kamei-style *change* metrics (Kamei et al., "A large-scale-empirical study of just-in-time quality assurance"):--| Feature | Meaning |-|---------|---------|-| `la`, `ld` | lines added / deleted |-| `nf` | files touched |-| `nd`, `ns` | distinct directories / top-level subsystems touched |-| `entropy` | Shannon entropy of the per-file churn distribution (diffusion) |-| `exp` | author's prior commit count (experience) |--`exp` is genuinely optional: when the author cannot be resolved — a diff-only-caller with no local history, or a name whose regex breaks the `git rev-list`-lookup — it is reported as `null` and contributes exactly zero to the logit.-Nothing is imputed, because `0` is a real value meaning "first ever commit" and-the model reads it as a risk-raising signal.--These are properties of the *diff*, so the score is a change-level signal rather-than a file-size proxy. The risk is a plain L2-logistic over standardized,-log-compressed features (`logit = intercept + Σ coefᵢ·zᵢ`), so every feature's-push on the risk is exact and reported as an attributable driver (the same-linear / per-finding-attributable contract the file health score holds).--## What the score does and does not buy--The score is a **diff-size statistic**. That is a measured claim, not a hedge:--- `la` (lines added) carries a coefficient 7.6× the next largest, and scoring by- `la` alone reproduces the full seven-feature score to within 0.12–0.16 points- on every repository tried.-- On a hand-picked set of small-but-dangerous changes versus large-but-boring- ones (47 within-repo pairs across repowise, flask, django and zod), the score- ranks the dangerous change above the boring one in **0 of 47** pairs. Ranking- by fix density alone gets **46 of 47**.-- That set is constructed, not held out: the pairs were chosen so that the- dangerous change is always the smaller one, which means ranking by lines added- scores 0 by construction and any signal genuinely independent of size scores- near-perfectly. It is a falsification test — "can the score ever do this?" —- and not an accuracy estimate. Its value is that the score failed it- completely, on cases a reviewer would call obvious.--A refit was measured and rejected rather than shipped. Regrouping the corpus to-PR granularity (`--first-parent` merge spans) and adding two size-orthogonal-features made accuracy *worse*: pooled leave-one-repo-out AUC 0.769 for the-refit against 0.776 for the current feature set and 0.780 for a churn-only-baseline. Per repository, **lines added alone matches or beats the fitted model-in five of six repos**. The reason is the labels: a commit is marked-defect-inducing when a later bug-fix's blame points back at a line it wrote, and-a larger commit writes more lines, so the label is itself size-biased. Any-deliberately size-orthogonal feature scores near chance against it — fix density-lands at 0.46–0.57 AUC — which is a fact about the labels, not about the-feature. So the model constants are unchanged and the score is reported as what-it demonstrably is.--`risk_authority` and `score_measures` state this in every payload;-`include=["scales"]` adds the per-field dictionary.-Read the result in this order:--- **Review priority** / **classification** / **percentile**: where this change's- *diff shape* sits in the repo's own distribution. This is the authoritative- population-relative review signal.-- **`fix_history`**: uncalibrated historical evidence about where the change- lands, reported separately rather than folded into a probability.-- **`score`** (0–10 normalized points): supporting diff size and spread,- offline-calibrated and corpus-anchored to a single commit. It ranks 0.99- against lines added on every repository measured, so `get_change_risk` keeps- it behind `include=["diagnostics"]` rather than on the wire; the CLI and the- REST range endpoint still report it.-- **`fallback_band`**: the heuristic-thresholded absolute `low` / `moderate` /- `high` model-score band. Present- *only* when there was no baseline to rank against (a shallow repo, or- `--baseline 0`), which is why it is not a peer of the review priority.--The score's absolute band is also **unit-blind**. Its corpus is individual-commits (baseline: 10.5 lines added, 1.7 files), so a squash-merged PR or a-`base..head` range is several commits' worth of diff read against a one-commit-scale and skews high: two-thirds of commits can read "high" while ranking-normally for *that* repo. The payload states the assumption in `score_unit`.--Each **driver** is reported relative to *the model's baseline commit* (the-calibration-corpus mean), not this repo, so a `+19 / −1` change can legitimately-read "more lines added than baseline" while still ranking `Below typical` for a-repo of large commits. The signed contribution and colour (red raised the raw-score, green lowered it) carry the direction; the label only states the-feature's standing, never an absolute verdict.--`nf`, `nd` and `ns` enter the logit exactly as fit but are **not reported as-drivers**. Their coefficients are small and negative — collinearity with `la`-(size), not a finding that touching more files is safer — so as an explanation-they contradict themselves: the label reads "more directories than baseline"-while the contribution is protective. Hiding an explanation we cannot stand-behind is the honest interim; a refit is the real fix.--The `repowise risk` CLI samples the repo's recent commits live (`--baseline`,-default 200) to compute this percentile; in the web UI it is precomputed from the-indexed commit history.+- **Prior fixes** counts bug-fix commits that touched the file before the+ change, weighted by age against the change's own date: a fix one year earlier+ counts a half, two years a quarter. The same commit scores the same on every+ re-run.+- **`density`** is the churn-weighted mean of those per-file numbers. It is a+ ratio, so it does not grow with diff size: one line in a file fixed twenty+ times outranks a thousand lines in files never fixed.+- **`percentile`** ranks that density against the same measure over the repo's+ recent commits. It is `null` when the change touches no fix history, or when+ fewer than eight sampled commits are available.++History is read from before the change: a commit is never ranked against fixes+it caused. For a range, the record is read at the fork point. Up to 20,000+commits are walked, once per repository state.++A bug fix is a commit whose subject matches `fix`, `bug`, `patch`, `resolves`,+`closes #N` or `fixes #N`, and does not contain `docs`, `typo`, `bump`, `deps`,+`chore`, `lint`, `format` or `style`. The same rule feeds the health layer and+[bug-fix history](BUG_HISTORY.md).++### The diff-shape score++The score uses Kamei-style change metrics: lines added and deleted (`la`,+`ld`), files touched (`nf`), directories and top-level subsystems touched+(`nd`, `ns`), the entropy of churn across files, and the author's prior commit+count (`exp`). `exp` is `null` when the author cannot be resolved and then+contributes nothing. The model is a plain logistic over standardized features,+so each driver's push is exact.++Drivers are stated relative to the model's baseline commit (10.5 lines added,+1.7 files), not to this repo. A small change can read "more lines added than+baseline" and still rank below typical in a repo of large commits. `nf`, `nd`+and `ns` enter the model but are not reported as drivers: their small negative+weights are collinearity with size, not evidence that touching more files is+safer.++The absolute score is calibrated on single commits. A squash-merged PR or a+`base..head` range is several commits' worth of diff and skews high on it; the+percentile does not have that problem. `score_unit` states this in every+payload.### What the sample is anchored to-The sample is the recent history the change is measured against, and where it-starts depends on what is being scored:+`repowise risk` samples recent commits live (`--baseline`, default 200). The+web UI uses the indexed commit history.++| Subject | Sample runs back from |+|---|---|+| A commit in `HEAD`'s history | `HEAD` |+| A commit on another branch | that commit |+| A `base..head` range | the merge-base of the two sides |+| Uncommitted work | `HEAD` |++A change never ranks against itself. One walk is reused for every change scored+against the same anchor and filters in a process, so a long-running MCP server+pays for it once.++### PR mode (`get_risk` with `changed_files`)++`get_risk(changed_files=[...])`, or `repowise risk --target FILE+--changed-file FILE`, answers a different question: what does this set of+files reach, and what is missing from it. The response opens with a+`directive` block:++| Field | Meaning |+|---|---|+| `may_break` | Production files in structural reverse-import reach of the diff. Candidates for review, not proven breakage |+| `may_break_tests` | Test files reached the same way, listed separately |+| `missing_cochanges` | Files that historically change with these but are not in the diff |+| `missing_tests`, `test_recommendations`, `tests_to_run` | Test gaps and which tests to run; each recommendation says whether it is `measured` (coverage) or `inferred` (graph reach) |+| `next_calls` | What to call next |+| `summary` | One sentence over all of the above |++In a workspace the directive also carries cross-repo fields:+`will_break_consumers` (repos that structurally depend on this one; structural+reach, not a runtime claim), `missing_cross_repo_cochanges`,+`breaking_changes` (incompatible provider contract changes since the last+index), `conformance_violations` and `dependency_cycles`. See+[Cross-Repo Blast Radius](../scale/WORKSPACES.md#cross-repo-blast-radius) and+[Breaking-Change Guard](../scale/WORKSPACES.md#breaking-change-guard).++PR mode also reports `structural_impact_score`: a deterministic, uncalibrated+0-10 heuristic over PageRank, churn and transitive dependents, banded+`localized` (below 4), `moderate` (4 to below 7) and `broad` (7 and up). It is+not a probability and does not decide review. `overall_risk_score` is an exact+deprecated alias.-| Subject | Sample runs back from | Why |-|---|---|---|-| A commit already in `HEAD`'s history | `HEAD` | Ranked against how the repo commits *now*. |-| A commit that is not (another branch) | that commit | `HEAD`'s history is not its cohort. |-| A `base..head` range | the merge-base of the two sides | The range's own commits stay out of the distribution it is measured against. |-| Uncommitted work | `HEAD` | Same subject, same cohort as scoring `HEAD`. |+## Independent changes-A change never ranks against itself: the target's own score is removed from the-sample before the percentile is taken.+This answers whether the files in front of you are one change or several. It+is a structural property of the diff, with no score.-Because the sample depends only on that anchor and the active filters, and not-on the individual change, one walk is reused for every change scored against the-same history in the same process. A long-running MCP server pays for the walk-once and answers subsequent `get_change_risk` calls without repeating it. A new-commit, a moved branch, or a different set of filters produces a different-anchor or a different filter set, and so a fresh walk.+A changed file is grouped only when it has a file node in the index, is not a+test file, and is in a language whose resolver can emit an import edge. Docs,+config, data and test files are never grouped and never bridge two groups.+Three kinds of link join eligible files:-## Independent changes+- **Index edges** between two different changed files: imports, calls, type+ references, framework and dynamic edges, reads.+- **Stored co-change pairs**: files history moves together.+- **Shared commits**, for a `base..head` range only: files one commit touched+ belong together.++Files left out are listed as `ungrouped_files` and never counted as an+independent change. `bridging_files` names the files that alone hold a group of+three or more together. The `basis` field states in words exactly what was+checked; a missing edge is a claim about this index, not about the code.-This is a structural property of the diff, not a score: it answers whether the-files in front of you are one change or several. The changed files are grouped by-connectivity, and one group is separate from another only because nothing-connected them.--Not every changed file is eligible to be grouped. A file is grouped only when it-has a file node in the index, is not a test file, and is written in a language-whose resolver can emit an import edge at all. That rules out documentation,-configuration and data files: a lockfile or a markdown page is a node with no-edges, so "nothing links it" is a fact about the language rather than about this-change. It rules out test files too, because no index records the tie from a test-to the code it exercises, and an integration test otherwise attaches to whichever-shared harness the diff happens to contain. Files that fail either test are never-placed in a group, not even through a co-change pair, and they never bridge two-groups: only eligible files are in the graph at all, and any link with one end-outside that set is dropped before grouping.--Three kinds of link connect two eligible files:--- **Index edges.** Every non-containment edge whose two ends belong to two- different changed files: imports, calls projected to the files that own them,- type references, framework and dynamic edges, and reads. Edges whose ends belong- to the same file are dropped, which is how the containment types (`defines`,- `has_method`) stay out without being enumerated, and an edge to a dependency- outside the repository links nothing.-- **Stored co-change pairs.** Two files history moves together are one change even- when nothing imports anything.-- **Commit co-membership, for a `base..head` range only.** The files one commit- touched are linked to each other, so a file two commits share joins them. That- is the author's own statement that the files belong together, which is stronger- evidence than any edge, and it is what `bridging_files` most often names. A- single commit and the working tree carry no such evidence: one commit says- nothing about what moves with what, so the range's commit list is read only when- the subject is a range.--A group also has to carry at least one file the index has in fact linked to-something, whether through an edge, a stored co-change partner, or a shared-commit. Elsewhere, an absent edge is not evidence, so the files fall out of the-grouping instead.--Everything not grouped is reported as `ungrouped_files` and never counted as an-independent change, and the summary says so in those terms: *N changed file(s)-is/are left out of the grouping: docs, config, tests, files not in the index, or files it has never-linked.* `bridging_files` names the articulation points of a group's own subgraph,-the files that alone hold the group together: move one out and the group splits.-It is computed only for groups of three or more files, since removing either end-of a pair leaves one file, which is not a split.--There is no score, no percentage, and no adjective on the result. The `basis`-field states in words what was actually checked, and it is not the same sentence-in both cases. When the commits of a range were read, it is *no import, call, type-reference, co-change pair or shared commit in the index and this range links one-group to another*; when there were no commits to read, the shared commit drops out-of the sentence and it is *no import, call, type reference or co-change pair in-the index links one group to another*. That wording is deliberate in either form:-a missing edge is a claim about this index, not about the code, and the sentence-never claims a check that did not happen.--The report is silent unless it has something to say. Nothing is produced when-fewer than two changed files exist, when fewer than two of them are eligible to be-grouped, or when fewer than two groups survive. A change that is one change gets-no report.--It surfaces in two places. `repowise risk` prints the groups under the driver-table, naming each group's files, its bridging files, and the first ten ungrouped-paths; `--format json` carries the same object under `independent_changes`. In-`get_change_risk` it is `independent_changes`, which needs an index-and is absent without one.--## Calibration & accuracy--Constants are learned offline against the defect corpus (AG-SZZ bug-inducing-commits as labels, time-ordered evaluation with a right-censoring gap, and a-leave-one-repo-out comparison to the churn-only baseline). On a 7-repo,-5-language slice the pooled leave-one-repo-out AUC is **0.772 vs 0.766 for-churn-only** (Δ +0.0068, 95% CI [-0.0003, +0.0131]).--Read that number for what it is. A churn-only baseline scores 0.766 on the same-labels, and lines-added alone scores higher still, so the margin measures very-little. It is reported because it is the number the constants were selected on,-not as evidence the score ranks danger — for that claim, see-[What the score does and does not buy](#what-the-score-does-and-does-not-buy),-where it fails.--**`fix_history` carries no AUC of its own, deliberately.** Its evidence is the-47-pair ranking gate (46/47) and the fact that the files it ranks highest in-this repository are the ones with the longest bug-fix records. It scores near-chance against the SZZ labels, which — as above — is a property of those labels.-Quoting a number from a benchmark that structurally cannot see the signal would-be worse than quoting none.--Only learned constants ship; the runtime stays deterministic, zero-LLM, and-free of new dependencies. Recalibrate via-`repowise-bench/health-defect/jit_calibration.py`; the constants live in-`packages/core/src/repowise/core/analysis/change_risk/model.py`.--## PR structural-impact scale (`get_risk`)--PR-mode `get_risk` answers a different question. Its-`structural_impact_score` is a deterministic, uncalibrated structural-exposure-heuristic in normalized points from 0 to 10. It combines the mean and maximum-of `pagerank * (1 + temporal_hotspot)` across changed files, maps that component-to at most 8 points with `8 * (1 - exp(-10 * combined))`, then adds at most 2-points for `min(transitive_dependents / 20, 1)`. Bands are `localized` below 4,-`moderate` from 4 to below 7, and `broad` from 7. It is not a probability and is-not authoritative for live change review. Historical co-change evidence is-reported separately and never enters this structural score.--`overall_risk_score` remains an exact deprecated alias so older clients keep-the same value and unit. `overall_risk_score_compatibility` names the migration-to `structural_impact_score`; the two fields cannot contradict.--The deterministic fixture corpus records the retained scale's distribution:--| Control | Score | Band |-| --- | ---: | --- |-| documentation / low signal | 0.01 | localized |-| small ordinary source | 0.34 | localized |-| historical fixes, limited reach | 0.34 | localized |-| co-change only | 0.08 | localized |-| moderate multi-file | 4.27 | moderate |-| structurally broad, little history | 7.34 | broad |-| genuinely broad high control | 9.91 | broad |--The numeric formula was retained, so the before-and-after numeric distributions-are identical; the correction changes names, units, authority, and labels. No-fixture saturates at 10, ordinary controls remain below the moderate threshold,-and all three bands are occupied. The corpus lives in-`tests/fixtures/risk_scale_corpus.json`.--## Public scale inventory--| Public value | Producer and evidence | Unit / range | Calibration and authority |-| --- | --- | --- | --- |-| `get_change_risk.score` / REST `score` / stored `change_risk_score` | Offline-fitted logistic model over live diff size, spread, entropy, and author experience; deterministic at runtime | normalized points, 0-10; calibrated at single-commit granularity | Benchmarked on 4,102 commits across 7 repositories; supporting signal, not a probability or review authority |-| `risk_percentile` | Mid-rank of the same score among filtered recent commits | percentile rank, 0-100 | Population-relative benchmark; authoritative with `classification` for live change review |-| `review_priority` / `classification` | Shared percentile terciles at 33.33 and 66.67 | category | Authoritative population-relative label |-| `fallback_band` | Shared score thresholds at 4 and 7, emitted only without a usable baseline | category | Heuristic thresholds on the benchmarked model score; absolute fallback, not population-relative |-| `fix_history.density` | Churn-weighted, recency-decayed prior bug fixes on touched files | recency-weighted prior fixes, unbounded | Uncalibrated historical heuristic; separate evidence, never folded into a probability |-| `get_risk` `hotspot_score` | Repository-relative churn percentile from the index | ratio, 0-1 | Uncalibrated normalized component |-| `get_risk` `health_score` | Indexed code-health model | health points, 0-10; higher is healthier | Benchmarked file-health signal, not interchangeable with change-risk points |-| `structural_impact_score` | PR structural formula above | normalized points, 0-10 | Deterministic and uncalibrated; not authoritative. `overall_risk_score` is an exact deprecated alias |-| `direct_risks[].structural_score` | `pagerank * (1 + temporal_hotspot)` | raw pagerank-weighted-hotspot value, unbounded | Uncalibrated within-change structural weight. `risk_score` is an exact deprecated alias |-| `direct_risks[].temporal_hotspot` | Exponentially-decayed sum of per-commit churn (halflife 180 d); each commit contributes up to 3.0 | raw decayed churn, unbounded (observed max 42.6 on this repository); use `churn_percentile` for a normalised 0–1 rank | Intermediate input to `structural_score`; not a ratio|-| `direct_risks[].churn_percentile` | Rank of `temporal_hotspot` among the repository's files | percentile rank, 0–1 | Comparable across repositories as a rank, not as activity; reported beside `structural_score`, never folded into it |-| `direct_risks[].is_hotspot` | The index's hotspot verdict: top-quartile churn AND its absolute activity floors | boolean | Read this rather than re-deriving a hotspot from `churn_percentile`, which omits the floors |-| `cochange_warnings[].score` | Number of historical commits in which the pair co-changed | raw commit count, 0+ | Historical evidence only; cannot become structural or runtime-breakage evidence |-| workspace `impacted[].score` | Strongest path product of edge confidence, edge-kind weight, and `0.6` per hop | relative path weight, 0-1 | Deterministic, uncalibrated ranking heuristic; not a probability or change-review authority |-| dashboard hotspot triage index | `40% * churn percentile + 35% * bus-factor tier + 25% * bounded temporal activity` | heuristic points, 0-100 | Client-side, uncalibrated orientation only; labelled adjacent to the chart |--Machine-readable `risk_authority`, `structural_impact_scale`,-`overall_risk_score_compatibility`, and `impact_score_semantics` carry these-definitions on every payload. They ship the guard tier - unit, range,-calibration status, authority - by default. The reference tier (fitting corpus,-formula, component breakdown, and the full `risk_scales` dictionary) is-identical on every call, so MCP returns it only for `include=["scales"]`; the-CLI `--json` output and this table carry it unconditionally.--## Cross-repo change risk (workspace mode)--> **Note:** This section describes `get_risk` in PR mode (`changed_files`).-> `get_change_risk` carries its own `cross_repo` block, built from workspace-> contracts rather than from the graph traversal described here.--In a workspace, a change rarely stops at the repo boundary. When `get_risk` is-called in PR mode (`changed_files`), its `directive` block gains two cross-repo-fields derived from the [system graph](../scale/WORKSPACES.md#system-graph):--- `will_break_consumers`: deprecated compatibility name for services in *other*- repos that structurally depend on the changed repo (a contract or package- import). This is structural reach for review, not a runtime-breakage claim.-- `missing_cross_repo_cochanges`: services in other repos that historically- co-change with the changed repo but aren't in the diff. Correlation, not a- call, so they read as "may drift," not "will break."--The same reachability powers the `get_blast_radius` MCP tool, the-`GET /api/workspace/blast-radius` endpoint, and the Live System Map's-blast-radius ripple. Structural edges outweigh behavioral co-change in the-ranking (one named constant, `BEHAVIORAL_EDGE_WEIGHT`, in-`packages/core/src/repowise/core/workspace/blast_radius.py`). See-[Cross-Repo Blast Radius](../scale/WORKSPACES.md#cross-repo-blast-radius) for the full-model.--The directive carries a third cross-repo field, `breaking_changes`, when a-provider contract in the changed repo has an incompatibility or an explicit-comparison warning. The compatibility-named key includes removed operations,-legacy proto/signature rules, and direction-aware OpenAPI request/response rules-for complete sides with matching fidelity. Where-`will_break_consumers` is topology ("who depends on this repo"),-`breaking_changes` is schema-level truth ("this specific contract changed in a-way that is incompatible"), each entry listing its side, schema source/fidelity,-and endpoint-exposed consumer files across repos. Those links do not prove exact-field use, runtime failure, or deployment safety. It is computed by diffing the-current contracts against the previously-indexed set during-`repowise update --workspace`; unsupported/unresolved sides and source, fidelity,-or selected-response transitions surface as uncertainty rather than field-removals. Compatible directional changes and new endpoints do not appear. See-[Breaking-Change Guard](../scale/WORKSPACES.md#breaking-change-guard) for the full model.+The report is silent unless there are at least two groups. `repowise risk`+prints the groups under the driver table; `--format json` and+`get_change_risk` carry them as `independent_changes` (absent without an+index).## Branch overlapBranch overlap answers one question: which other open branches edit the filesthis change edits. It is git first, so it works in a fresh clone with no index.-The scan starts from one `git for-each-ref` over `refs/heads` and-`refs/remotes`, sorted by committer date, newest first. Symbolic remote heads-are skipped, refs sharing a tip commit collapse to one entry (the local name-where there is one, otherwise the shortest remote name), and the base ref, the-change's own ref, and any ref whose tip is the base's or the change's commit are-dropped. Branches stacked on the current one, and the ones it is stacked on, are-dropped too: either shares every file by construction, which is one change split-over two refs rather than two changes racing. The scan is bounded to the newest-50 branches by committer date, raised or lowered with `--limit`, and what was-left out is reported through `scanned` and `total` rather than dropped silently.--For each scanned branch the file list is `git diff --name-only base...branch`,-which is what that branch did since it forked, not what the base did in the-meantime. That list is intersected with the change's own changed files. Noise-paths (workflows, lockfiles, generated and vendored files, localization blobs)-are removed from both sides by the shared noise filter, and dependency manifests-(`package.json`, `pyproject.toml`, `go.mod`, `Cargo.toml`, `Gemfile`,-`composer.json`, `requirements.txt`, `setup.py`, `pom.xml`, `build.gradle`,-`build.gradle.kts`) are removed with them, because every dependency bump edits-the manifest and sharing one says nothing about the work. A branch with no shared-file after that filtering produces no entry, ever.--Every row states its basis in words, and there are exactly two:--- **`same file`.** A direct hit: both branches change that path.-- **`co-change pair, N of M commits`.** A secondary row, shown only beneath a- branch that already has a direct hit. It names a file the other branch edits- that is not itself a direct hit and that the index's stored co-change record- pairs with one of this change's files, with that file carried as `partner`. It- is made only when both counts are recorded and the pair moved together at least- half the time (`N * 2 >= M`), and at most three appear per branch, so one file- that pairs with everything cannot drown out the files truly shared.--An entry also carries `ahead` and `behind` (commits only on the branch, then only-on the base, from one `git rev-list --left-right --count`) and the date of its-last commit. Branches are ordered by how many files they share directly, then by-most recent commit, then by name; the co-change rows are appended afterwards and-never reorder the list. There is no score and no percentage anywhere in the-output. With an index, the shared files inside an entry are ordered by the same-hub metric the PR blast radius uses (temporal hotspot and pagerank of the file);-that metric orders rows and is never rendered. Without an index the git answer-stands and the shared files are alphabetical, with no co-change rows.--What it does not claim: it compares paths, not hunks, so a shared file is a-reason to talk to the other branch's author and not a predicted merge conflict.-The co-change rows are historical correlation, never a structural link.--Silence is the default. `repowise overlap` prints one line when no other branch-edits a shared file, naming how many branches were scanned of how many exist, and-`get_change_risk` omits the `branch_overlap` block entirely in that case. The-block is also omitted when the branch scan times out.+```bash+repowise overlap # HEAD against the trunk+repowise overlap --base main --branch feature/x --limit 100+```++It scans local and remote branches, newest committer date first, up to+`--limit` (default 50). It drops the base, the change's own ref, refs pointing+at either tip, and branches stacked on or under this one. For each branch it+takes `git diff --name-only base...branch` (what that branch did since it+forked) and intersects it with this change's files. Noise paths (workflows,+lockfiles, generated, vendored and localization files) and dependency+manifests (`package.json`, `pyproject.toml`, `go.mod`, `Cargo.toml` and+similar) are removed from both sides first. A branch with no shared file after+filtering produces no entry.++Every row states its basis, one of two:++- **`same file`**: both branches change that path.+- **`co-change pair, N of M commits`**: shown only under a branch that already+ has a direct hit. A file the other branch edits that history pairs with one+ of yours at least half the time. At most three per branch. Needs an index.++Each entry carries `ahead`, `behind` and the date of its last commit. Branches+sort by shared files, then recency, then name. There is no score.++It compares paths, not hunks: a shared file is a reason to talk to the other+author, not a predicted merge conflict. `repowise overlap` prints one line when+nothing overlaps, naming how many branches were scanned of how many exist.+`get_change_risk` omits the `branch_overlap` block when nothing overlaps or the+branch scan times out.++## Tuning and suppressing++- **`--exclude` / `-x PATTERN`** (repeatable, gitignore-style) omits files. The+ same filter applies to the sampled commits, so the comparison stays like for+ like.+- **`.riskignore`** at the repository root holds project-wide, risk-only+ patterns. They apply automatically and combine with `-x`.+- **`--ext`** counts only the listed suffixes.+- **`--baseline N`** sets the sample size; `0` turns ranking off and leaves only+ `fallback_band`.++## In CI++`--fail-above-percentile P` turns the command into a gate on `risk_percentile`.+The absolute 0-10 value never decides.++```bash+repowise risk --fail-above-percentile 95 --format github+repowise risk origin/main...HEAD --fail-above-percentile 95 --format markdown+```++With the flag, or with `--format markdown` or `github`, and no revspec, the+subject is the CI change: the pull request's target branch (from CI variables,+else the remote's default branch) `...HEAD`. A CI checkout of a PR is a merge+commit, so `HEAD` alone would be the wrong change.++Exit codes are the shared CI ones: `0` at or below the percentile, `1` above,+`2` when the gate cannot evaluate. A change with no percentile cannot be gated:+`--baseline 0` turns ranking off, and a shallow clone with fewer than 8 recent+commits has nothing to rank against, so fetch full history. A revspec git+cannot read also exits `2`. At percentile P roughly (100 - P)% of changes fail+by construction; a failure asks for a split or a second reviewer, not a code+fix.++`json` adds a `gate` object with the unrounded percentile it compared.+`markdown` leads with the verdict, then the range, sample, size and fix-history+rank. `github` writes an `::error::` (or `::notice::`) and puts the markdown in+the job summary.++The GitHub Action runs it as the `risk` gate (input+`risk-fail-above-percentile`; empty reports the rank without gating), and the+GitLab template as the `repowise-risk` job when+`REPOWISE_RISK_FAIL_ABOVE_PERCENTILE` is set. See [CI](../start/CI.md).++## Accuracy and limits++- **The score is a diff-size statistic.** Lines added carries a weight 7.6x the+ next feature, and `la` alone reproduces the full score within 0.12-0.16+ points on every repo tried.+- **Against churn-only it barely moves.** Leave-one-repo-out AUC on 4,102+ commits across 7 repositories (AG-SZZ labels): 0.772 for the model, 0.766 for+ churn alone (95% CI of the gap includes zero).+- **The score cannot rank danger.** On 47 hand-picked within-repo pairs of a+ small dangerous change and a large boring one, the score picks the dangerous+ one in 0 of 47. Fix density picks it in 46 of 47. The set is a falsification+ test, built so size alone scores 0; it is not an accuracy estimate.+- **`fix_history` has no AUC of its own.** It scores near chance against SZZ+ labels, which are size-biased; quoting that number would mislead.+- **The fix classifier is keyword based.** It misses conventions outside its+ list (Django's `Fixed #N` is the notable one) and drops genuine fixes whose+ subject says `docs` or `style`. Where a project's convention falls outside+ it, `fix_history` reads low.+- **Absolute scores skew high on ranges and squash merges**: the scale is one+ commit. Use the percentile.+- **Overlap compares paths, not hunks.**++Method, the refit that was measured and rejected, and every public scale are in+[architecture/change-risk.md](../architecture/change-risk.md).++## Where it shows up++- **CLI**: `repowise risk`, `repowise overlap`.+- **MCP**: `get_change_risk` for a diff, `get_risk` for files and PR mode.+- **Dashboard**: the Commits page, led by review priority.+- **REST**: `GET /api/repos/{repo_id}/risk/range`.+- **CI**: the GitHub Action `risk` gate and the GitLab `repowise-risk` job.++## Reference++### `repowise risk [REVSPEC]`++| Flag | Default | Meaning |+|---|---|---|+| `--path DIR` | `.` | Repository path |+| `--ext SUFFIXES` | all | Comma-separated suffixes to count (`.py` or `.ts,.tsx`) |+| `--baseline N` | 200 | Recent commits to rank against; `0` disables ranking |+| `--exclude`, `-x PATTERN` | none | Gitignore-style exclusion, repeatable |+| `--target`, `-t PATH` | none | Report history for these files instead of a change (needs an index) |+| `--changed-file PATH` | none | With `--target`: PR mode, leads with the directive |+| `--fail-above-percentile P` | none | CI gate on `risk_percentile` |+| `--format` | `table` | `table`, `json`, `markdown`, `github` |+| `--full` | off | Emit the complete payload as JSON |++### `repowise overlap`++| Flag | Default | Meaning |+|---|---|---|+| `--base REF` | trunk | Base ref |+| `--branch REF` | `HEAD` | The change to compare |+| `--path DIR` | cwd | Repository path |+| `--limit N` | 50 | Branches to diff, newest first |+| `--format` | `table` | `table` or `json` |++Full CLI entries: [`repowise risk`](../reference/CLI_REFERENCE.md#repowise-risk-revspec),+[`repowise overlap`](../reference/CLI_REFERENCE.md#repowise-overlap). MCP+schemas: [`get_change_risk`](../agent/MCP_TOOLS.md#get_change_risk),+[`get_risk`](../agent/MCP_TOOLS.md#get_risk).++## See also++- [architecture/change-risk.md](../architecture/change-risk.md): calibration,+ the scale inventory, and PR structural-impact internals.+- [BUG_HISTORY.md](BUG_HISTORY.md): per-file and per-symbol fix counts.+- [CODE_HEALTH.md](CODE_HEALTH.md): file-level scores.+- [Workspaces](../scale/WORKSPACES.md): cross-repo blast radius and contracts.diff --git a/docs/layers/CODE_HEALTH.md b/docs/layers/CODE_HEALTH.mdindex 32609321e..42abdc730 100644--- a/docs/layers/CODE_HEALTH.md+++ b/docs/layers/CODE_HEALTH.md@@ -1,763 +1,274 @@# Code Health-Repowise scores every file in your repo from 1 to 10 on three signals: **defect-risk**, **maintainability**, and **performance risk**. The score is computed by-static analysis over tree-sitter ASTs and git history, with **no LLM calls and-no cloud requirement**, and the defect signal's weights are calibrated against a-real bug corpus rather than hand-tuned.--<div align="center">-<img src="../../.github/assets/health-loop.svg" alt="repowise code-health loop: deterministic markers fan into three signals (defect risk, maintainability, performance), the graph and git history locate where risk concentrates, and refactoring intelligence emits concrete plans your agent executes" width="100%" />-</div>--The loop is **measure → locate → fix**: score every file, use the dependency-graph and git history to find where risk concentrates, then emit a concrete-refactoring plan an agent can execute.--**Linters check patterns. This predicts risk.** A linter flags a line matching a-known-bad shape. The health score estimates which files are likely to harbor the-next bug, ranks them, and has been validated forward in time against real defect-history. It uses signals a linter has no access to — churn, ownership,-co-change, blast radius, hotspots — and even its one linter-adjacent pillar-(performance) follows the call graph *across* files, which a file-local linter-cannot do.+Repowise scores every source file from 1 to 10 and reports three co-equal+signals: **code health** (the defect-risk score, calibrated against real bug+history), **maintainability** and **performance risk**. It reads the parse tree,+the dependency graph and git history. It needs no LLM key and no network, and it+runs as part of `repowise init` and `repowise update`. The analysis is budgeted at+under 30 seconds for a 3,000-file repository, a budget a test enforces.++It finds code shapes and change patterns that go with bugs, slow code and code+that is hard to change. It does not run your code, measure runtime, or prove a+bug exists. A low score means "look here first", not "this is broken".## Quick start```bash-repowise init # full index, populates health tables-repowise health # KPIs + 20 worst-scoring files + top findings-repowise update # re-score only changed files on each subsequent run+repowise init # full index; scores every file+repowise health # Fix first, KPIs, lowest-scoring files, top findings+repowise next # the short "Do next" list across every layer+repowise update # re-scores only the files that changed```-Open `http://localhost:7777/repos/<id>/health` for the dashboard once-`repowise serve` is running.--## How a marker is actually computed--Every marker is a deterministic function of the parse tree, the dependency-graph, and git metadata. Two worked examples, one simple and one that shows why-this is not a linter.--**`nested_complexity` — pure AST.** tree-sitter parses the file into a syntax-tree. A walker descends it tracking control-flow depth, incrementing on each-`if` / `for` / `while` / `try` / `switch` node. If any function reaches depth 4-or more, the marker fires, and severity scales with the depth it reached. A-function that is mostly one dispatch on one value (a `switch`, `match` or-same-subject `if` chain holding at least 60% of its decision points) is judged-by the code around that dispatch and its heaviest arm: `nested_complexity`-does not count the two levels the `switch` and its `case` open, and-`complex_method`, `brain_method` and `bumpy_road` read the CCN outside the-dispatch plus the CCN of its most complex arm. No heuristics about intent, no model, no-sampling. The same commit produces the same finding forever.--**`io_in_loop` — AST plus call graph.** The walker finds a call inside a loop-body and resolves the callee through the same resolver the dependency graph-uses. The resolved target is classified against an I/O-boundary lexicon into-`db` / `network` / `filesystem` / `subprocess` / `lock`, and the marker fires-only on an actual round-trip (a `.execute`, an awaited HTTP call, a-`subprocess.run`) — never on a query-builder chain or a same-named pure helper.--The part a linter cannot reach: **the loop and the I/O call do not have to be in-the same function.** If the sink is not local, Repowise walks the resolved call-graph backwards up to three hops and reports the finding with its full-`caller → … → sink` path attached. clippy, ruff's `PERF` rules, ESLint and-golangci-lint all analyze one function at a time, so a loop in one file whose-database call lives in another is structurally invisible to them.--**Why deterministic is a feature, not a limitation.** The score is reproducible,-diffable between commits, and safe to gate CI on. It is also the precondition-for everything in the validation sections below: a score that returns a-different number each time it runs cannot be calibrated against a defect corpus-at all. An LLM asked to rate the same file twice gives two answers, and neither-can be checked against what actually broke.--The full analysis is budgeted to finish in **under 30 seconds on a 3,000-file-repo** — enforced by a CI test-(`tests/integration/test_health_perf_benchmark.py`), including a variant with a-per-line blame index built for every file.--## The markers, and what each is allowed to do--Repowise ships **53 registered detectors (56 marker ids)**, but only **25 are-permitted to move the headline number**. That restriction is deliberate: the-defect score carries published accuracy claims, so only markers that earned-their weight against a bug corpus may affect it.--| Tier | Markers | What it may do |-|---|---:|---|-| **Defect-scoring** | **25** | Calibrated weights; moves the 1-10 score |-| **Performance** | **22** | Own pillar, own cap; never touches the defect score |-| **Maintainability-only (SQL)** | **3** | Maintainability only |-| **Governance** | **3** | Surfaces as a finding; never deducts |-| **Advisory** | **3** | Measured and reported; never deducts, and stays out of impact-ranked lists unless asked for |--Nothing is inert, but "doesn't move the number" means three different things:--- **Floored (8 markers).** `developer_congestion`, `low_cohesion`,- `brain_method`, `bumpy_road`, `primitive_obsession`, `dry_violation` and- `error_handling` at ×0.5; `knowledge_loss` at ×0.4. These fire widely and- matter for readability, but proved weak defect predictors under leakage-free- calibration, so they were demoted rather than deleted — and most of them earn- **full weight in maintainability**, where they belong.-- **Advisory (21 of the 22 performance markers).** Sub-1.0 weights inside a- category capped at 2.0. Only `io_in_loop` carries full weight.-- **Non-scoring (3 governance markers).** `ungoverned_hotspot`,- `stale_governance` and `contradictory_decision` are written by an additive- pass that runs *after* scoring completes, so they surface in `get_risk`,- `get_context` and `CLAUDE.md` without ever deducting.--### Markers that count toward two signals--Maintainability is not a separate detector set. It is a **second weighting of-the same evidence**, with its own category table and its own caps.--| Markers | In defect | In maintainability |-|---|---|---|-| `low_cohesion`, `brain_method`, `primitive_obsession`, `dry_violation`, `error_handling` | ×0.5 (floored — weak bug predictors) | **×1.0 (full weight)** |-| `god_class`, `large_method`, `nested_complexity` | ×1.13 / ×1.25 / ×1.34 (calibrated) | **×1.0** |-| `sql_high_complexity`, `sql_select_star`, `sql_update_delete_without_where` | — | ×0.7 |--So **11 markers feed maintainability: the 8 shared above plus the 3 SQL-only-ones.** The first group is the design in miniature — a smell that doesn't-predict bugs gets demoted in the pillar that claims to predict bugs, and paid in-full in the pillar that claims to measure how hard code is to work with.+```python+get_health() # repo KPIs, worst files, top findings+get_health(targets=["src/api/server.py"]) # one file in detail+get_health(include=["refactoring", "accuracy"])+```-One display nuance: scoring membership and a finding's displayed "home" pillar-are different. `god_class`, `large_method` and `nested_complexity` score in both-pillars but display as defect findings, so a pillar-filtered view shows 8-maintainability findings while 11 markers actually feed the number.+In the dashboard (`repowise serve`), open **Code Health** for a repository:+`/repos/<id>/code-health`. Its tabs are Overview, Performance, Findings, Tests,+Dead code, Doc drift, Security and Blast radius.-The complete marker roster, per-marker definitions, and the full weight tables-live in-[`docs/architecture/code-health.md`](../architecture/code-health.md#5-the-markers-and-their-categories).+Illustrative `repowise health` output:-## The score+```+Fix first (3 of 41 eligible; tests, tooling and history-only files left out)+ 1. Extract the retry branch from parse_config now · effort M · +1.4 health+ src/config/loader.py:212+ CCN 47 and nesting 6 in a file changed 19 times this quarter.+ verify: pytest tests/unit/test_loader.py-Each file starts at 10.0 and marker findings deduct from it. Deductions are-capped per category, so no single category can dominate:+Code health: 7.4/10 [Good] · Hotspot: 6.1/10 · Worst: 2.3/10 (src/config/loader.py)+```-| Category | Cap |-|---|---|-| Organizational | −3.5, less on cleaner code (below) |-| Structural complexity | −2.5 |-| Test coverage | −2.0 |-| Test coverage gradient | −2.0 |-| Size & complexity | −1.5 |-| Duplication | −1.0 |-| Test quality | −0.5 |-| Error handling | −0.5 |--When a category exceeds its cap, every finding in it is **scaled proportionally**-rather than truncated, so each finding's reported impact stays linear and-attributable — you can always explain exactly why a file scored what it scored.-The final score is clamped to `[1.0, 10.0]`.--**Organizational is the largest cap, and it is the git-derived one.** Churn,-ownership, co-change, congestion and prior fixes describe what a file has been-*through*, not what its code is like — and they rise as a file is worked on. So-a week of genuine refactoring can lower a file's structural deduction while its-history deduction climbs, and the headline barely moves. Each file therefore-stores its deduction as two numbers, `structure_deduction` and-`history_deduction`, which sum to the total. Nothing is hidden: the split is on-every metric row, and the Counts control below reads it back.--Git history on its own measures activity, so its cap follows the code: history-can cost a file at most 1.0, plus one point for each point the other categories-deduct, up to 3.5. A file with no code-shape finding never reads below 9.0 from-history alone.--Three repo-level KPIs: **Hotspot Health** (NLOC-weighted average over files the-git layer classifies as hotspots), **Average Health** (NLOC-weighted over all-files), and **Worst Performer**.+## The three signals-### Bands+All three are computed from one stream of findings by one scoring routine, each+with its own weights and caps. They are never blended into one number.-| Band | Score | Colour | Meaning |-|---|---|---|---|-| **Excellent** | `≥ 8.5` | green | Low-risk, maintainable |-| **Good** | `7.0 – 8.5` | green | Sound; nothing demanding attention |-| **Fair** | `5.5 – 7.0` | gold | Rising complexity or process risk |-| **Needs work** | `4.0 – 5.5` | amber | Worth scheduling |-| **At risk** | `< 4.0` | red | High-risk; concentrates defects |--The bands are **absolute, not percentile**, so a score means the same thing-behind a firewall as it does against a public corpus. Excellent and Good share-one green and are told apart by the word; green starts at 7.0 because a 7 is-not a warning.--The `4.0` cutoff is the empirical one and it has never moved. On the 2,770-file-measured corpus, files below it carry **16.9× the per-file defect rate of files-at 8.0 and above** (95% CI 8.6–29.0), and **2.18×** the defects-per-KLOC once-file size is normalized out (95% CI 1.00–3.58). Both numbers belong together:-the raw ratio is the headline, and the size-normalized one is the proof it is-not simply an artifact of large files being large. The boundaries above 4.0 are-reading points on that same scale rather than separate defect-rate claims.--`8.0` survives as `TARGET_SCORE`, the score a refactoring's leverage is-measured against. It is deliberately not a band edge — moving it would reorder-every recommendation.--The bands are defined once in core (`analysis/health/grading.py`) and mirrored-in `@repowise-dev/types`, with a parity test on each side.--### What the score counts--Two controls decide which files a figure describes and which half of the-deduction it counts. Both reach the CLI, MCP and every page, and both are-echoed on the response, so a number always states what it is.--| Control | Values | What it does |+| Signal | What it answers | How its weights were set ||---|---|---|-| **Scope** | `all` (default) · `production` | Which files are in the population. Non-code files are never scored. |-| **Counts** | `everything` (default) · `code_shape` | Whether the git-derived half of the deduction counts. |+| **Code health** (defect risk) | Which files are most likely to carry the next bug | Fitted against a defect corpus; the headline number, the band and the badge |+| **Maintainability** | How hard the code is to read and change | Set by hand against its own caps; never fitted to bugs |+| **Performance risk** | Where the code's structure wastes work (I/O in loops, quadratic building, blocking calls) | Per-marker precision weights in one capped category |-```bash-repowise health --scope production # exclude tests-repowise health --counts code_shape # the code-shape reading-```+Some smells fire often but predict bugs weakly (low cohesion, brain methods,+duplicated code, broad error handling). They are down-weighted in code health+and counted at full weight in maintainability, where they belong.-```python-get_health(scope="production", counts="code_shape")-```+## Reading the score-**`code_shape` answers "is my code getting better".** It subtracts-`history_deduction` and rescores from what is left, so a week of refactoring-shows up as the improvement it was rather than being masked by the churn that-the work itself created. It is arithmetic over the two stored columns, not a-second scoring pass: history findings are dropped rather than re-weighted, and-a file with no stored split is reported in `unscored_files` rather than counted-as a ten.--**`production` lowers the headline**, which surprises people: test files-generally score higher than the code they cover, so removing them removes the-easy end of the distribution. Nothing was found; the population changed.--`everything` remains the calibrated number every accuracy claim on this page is-about, and the health badge always reports it — so a projected reading can-never be mistaken for the published one.--## How the weights were calibrated--This is the part that separates "we wrote some heuristics" from "we measured-which heuristics predict bugs."--**The protocol.** Every file is scored at a commit that *precedes* the bug-window (T0), in a detached git worktree whose history stops there. Defect labels-are then collected from the following six months. Nothing after the scoring-commit can reach the score. An L2-regularized logistic regression with **NLOC as-an explicit control** fits each marker's defect lift *beyond file size*, with-leave-one-repo-out cross-validation for generalization. Only the learned-constants ship; the runtime stays fully deterministic.--**The trap that makes this hard.** Leakage here is two-sided, and both sides are-silent failures:--- Score at HEAD and the bug-fix commits themselves *manufacture* the churn,- co-change and congestion that the process markers then "detect." The score- looks excellent and means nothing.-- Score naively at T0 and every recency-windowed marker silently stops firing,- because a worktree six months in the past sees an empty "last 90 days."--Getting both right requires anchoring the recency windows to the worktree's own-HEAD rather than wall-clock time. The evidence that this is load-bearing rather-than ceremonial: `developer_congestion` shipped at weight **1.5** under HEAD-scoring, and came back at coefficient **−0.08** under correct T0 scoring. It is-now floored to 0.5. A shipped constant was wrong by a factor of three, and only-the leakage-free protocol revealed it.--**What calibration rejected.** The markers listed as floored above are there-because the corpus said so. Candidate signals were also tested and formally-declared equivalent-to-null using bootstrap TOST equivalence testing rather than-a mere failure to reach significance: graph centrality (PageRank), code-naturalness, change bursts, and error-handling density all came back flat. Whole-experiments were run and refuted — size-relative scoring (lifts small files but-costs 0.07–0.12 AUC overall), function-level prediction (Popt at or below random-effort-ordering), and recalibrating on SZZ labels (*lower* out-of-fold AUC than-the shipped weights, so the shipped weights were kept).--Full methodology, per-marker coefficients and the rejected-candidate tables:-[repowise-bench/health-defect](https://github.com/repowise-dev/repowise-bench/tree/master/health-defect).--## Does the score find the bugs?--### On your repo, at index time--After an index, Repowise checks the score against your repo's own history and-prints a one-line callout:+Each file starts at 10. Every finding deducts by its severity times its marker+weight, deductions are capped per category, and the result is clamped to 1-10.+When a category passes its cap, its findings are scaled down together, so each+finding's reported impact still adds up to the file's score.-```-Does the score find the bugs? 16/20 lowest-health files had a bug fix in the-last 6 months, 3.3x the 24% baseline (80% vs 24%).-```+### Bands-It ranks every file by score, takes the 20 lowest, and counts how many a `fix:`-commit touched in the trailing ~180 days, contrasted against the repo-wide base-rate. Agents can read the same stat over MCP with-`get_health(include=["accuracy"])`, so a coding agent can confirm the score is-trustworthy on *this* repo before acting on it.+| Band | Score | Meaning |+|---|---|---|+| **Excellent** | 8.5 and above | Low risk, easy to change |+| **Good** | 7.0 to 8.5 | Sound; nothing demands attention |+| **Fair** | 5.5 to 7.0 | Rising complexity or process risk |+| **Needs work** | 4.0 to 5.5 | Worth scheduling |+| **At risk** | under 4.0 | Where defects concentrate |++Bands are absolute, not percentiles, so a 6.2 means the same thing in every+repository. The 4.0 cutoff is the measured one: on the 2,770-file paired+benchmark corpus, files under 4.0 carried 16.9 times the per-file defect rate of+files at 8.0 and above, and 2.18 times the defects per thousand lines once size is+taken out. The other edges are reading points on the same scale.++Repository KPIs: **average health** and **hotspot health** (both weighted by code+lines; hotspot health covers only the files the git layer marks as hotspots), plus+the **worst performer**. A file in a language health has no support for is counted+as not analysed, never scored as a 10.++### What moves the score++The **53 registered detectors** fall into these families. Category caps keep any+one family from dominating a file.++| Family | Detectors | Cap on code health | What it catches |+|---|---:|---|---|+| Structural complexity | 6 | 2.5 | Deep nesting, brain methods, god classes, low cohesion, bumpy roads, complex conditions |+| Size and complexity | 3 | 1.5 | Long or high-CCN functions, primitive obsession |+| Duplication | 1 | 1.0 | Cloned blocks |+| Error handling | 1 | 0.5 | Swallowed or catch-all handlers, `unwrap`/panic, discarded Go errors |+| Change history | 10 | 1.0 to 3.5 | Churn, change entropy, co-change scatter, ownership, knowledge loss, prior bug fixes |+| Test coverage | 3 | 2.0 + 2.0 | Untested hotspots, coverage gaps, a continuous deduction for uncovered lines (needs a coverage report) |+| Test quality | 4 | 0.5 | Oversized or duplicated assertion blocks; two advisory markers (below) |+| SQL | 4 | maintainability or performance only | Overly complex SQL, `SELECT *`, `UPDATE`/`DELETE` with no `WHERE`, cartesian joins |+| Performance | 21 | performance only | See [Performance findings](#performance-findings) |++Twenty-five detectors move code health. The SQL and performance detectors move+only their own signals. Three markers are **advisory**: they are measured and+listed, carry an impact of exactly zero, and stay out of impact-ranked lists unless+you ask for the advisory dimension. Three more **governance** findings (an+ungoverned hotspot, stale governance, a contradictory decision) come from the+[decisions layer](DECISIONS.md) and never deduct.++**History amplifies code risk; it does not stand alone.** Git history can cost a+file at most 1.0 point, plus one point for each point its code-shape findings+deduct, up to 3.5. A simple file that changes often reads as busy, not broken, and+never drops below 9.0 from history alone. Each file stores its deduction in two+halves (structure and history) so you can see which one holds it down.++<details>+<summary>All 53 detectors by family</summary>++| Family | Detector ids |+|---|---|+| Structural complexity | `nested_complexity`, `brain_method`, `god_class`, `low_cohesion`, `bumpy_road`, `complex_conditional` |+| Size and complexity | `complex_method`, `large_method`, `primitive_obsession` |+| Duplication | `dry_violation` |+| Error handling | `error_handling` |+| Change history | `change_entropy`, `churn_risk`, `co_change_scatter`, `code_age_volatility`, `developer_congestion`, `function_hotspot`, `knowledge_loss`, `ownership_risk`, `prior_defect`, `hidden_coupling` (advisory) |+| Test coverage | `untested_hotspot`, `coverage_gap`, `coverage_gradient` |+| Test quality | `large_assertion_block`, `duplicated_assertion_block`, `assertion_free_test` (advisory), `mock_saturated_test` (advisory) |+| SQL | `sql_high_complexity`, `sql_select_star`, `sql_update_delete_without_where` (maintainability); `sql_cartesian_join` (performance) |+| Performance | `io_in_loop`, `nested_loop_with_io`, `serial_await_in_loop`, `blocking_sync_in_async`, `hot_path_sync_io`, `blocking_io_under_lock`, `lock_in_loop`, `resource_construction_in_loop`, `lazy_load_in_loop`, `unbounded_read_reduced_in_memory`, `string_concat_in_loop`, `nested_loop_quadratic`, `membership_test_against_list_in_loop`, `list_insert_zero_in_loop`, `regex_compile_in_loop`, `json_parse_in_loop`, `defer_in_loop`, `goroutine_in_unbounded_loop`, `array_spread_in_reduce`, `pd_concat_in_loop`, `pandas_iterrows_in_loop` |-It stays silent on repos with too little history to be honest (fewer than 25-scored files, or fewer than 5 recently-fixed files). One caveat it discloses:-`prior_defect` is itself one down-weighted input to the score, so this is an-association on indexed history, **not** a leakage-free forward prediction. That-is what the next section is for.+</details>-### Across projects, leakage-free+### Ownership and knowledge-loss signals-Every file graded at a commit preceding the bug window, then checked against-what actually broke. **21 repositories, 9 languages, 2,826 files, 379-defect-bearing.**+Three history detectors read who has worked on a file, from git authorship alone:-| Result | Value |+| Detector | Fires when ||---|---|-| Cross-project mean ROC AUC | **0.737** (95% CI 0.683–0.787) |-| Held-out (PROMISE / jEdit 4.0, 4.1) | **0.761** and **0.776** |-| vs. recent churn | **+0.100 AUC** (DeLong p = 5e-10) |-| vs. prior-defect history | **+0.117 AUC** (DeLong p = 3e-15) |-| Per-repo range | 0.55 (axios) to **0.86** (zod) |--ROC AUC measures how often the score ranks a known-buggy file worse than a clean-one: 0.5 is a coin flip, 1.0 is perfect. The **jEdit result is the important-one** — it is a public academic defect dataset that played no part in-calibration, scored from a single commit snapshot with no git history available,-so only the structural markers could fire. That is the main evidence the signal-is not overfit to our corpus.--Confidence intervals come from a two-stage repo-cluster bootstrap (2,000 seeded-replicates, resampling repositories then files within them), because the unit of-generalization is the repository, not the file.--### The limits, stated plainly--These are published because a validation section that only contains wins is not-a validation section.--- **The score does not beat raw file size on discrimination.** LOC-only scores- AUC 0.742 against our 0.737 (ΔAUC −0.001, p = 0.92). Where it wins decisively- is effort-aware ranking: **ΔPopt +0.134 (95% CI +0.080 to +0.198)**. Same- discrimination as counting lines, far better at ordering a fixed review- budget — and unlike a line count, it tells you *why* a file is risky.-- **Within a size band, most of the signal disappears.** Holding file size- fixed, within-band AUC is 0.525 / 0.572 / 0.593 / 0.718 across NLOC quartiles.- The two smallest quartiles straddle 0.5. Discrimination survives cleanly only- in the largest quartile.-- **That last finding is a real absence, not a measurement artifact.** A- purpose-built positive control injected a synthetic size-orthogonal signal of- the headline magnitude into the same corpus, band membership and clustering,- and recovered it with power 0.998–1.000 in exactly the quartiles where the- real signal vanishes. So the collapse is a property of the score, not of the- sample size.-- **A prior-defects baseline still beats us on effort-aware ranking.**- ΔPopt −0.085 (95% CI −0.141 to −0.035). "Re-inspect whatever broke before" is- a hard heuristic to beat per line reviewed. Our edge is discrimination plus an- attributable explanation, not raw bugs-found-per-LOC.--## Head to head against CodeScene--CodeScene is the closest commercial product and the only other vendor in this-category with a published empirical defect study. Both tools scored the **same-2,770 files at the same leakage-free commit against the same labels**, through-the same estimator, with paired tests resampling the same repositories for both.--| Paired test | Repowise | CodeScene | p |-|---|---:|---:|---|-| Recall @ 20%-of-lines review budget | **0.173** | 0.074 | **0.003** |-| Effort-aware ranking (Popt) | **0.607** | 0.462 | **0.003** |-| Defect density, size-normalized | **2.18×** | 0.56× | **0.003** |-| Discrimination (ROC AUC) | 0.731 | 0.705 | 0.054 — *not significant* |-| Defect density, raw | 16.9× | 14.2× | 0.65 — *not significant* |-| Precision @ 20%-of-lines budget | 0.580 | **0.636** | 0.64 — *a tie* |-| Beyond file size (partial ρ) | −0.148 | −0.137 | *both beat size — a tie* |--Ranking by Repowise health surfaces **2.3× the defects under a fixed review-budget**. Three of seven axes are significant wins; the rest are honest ties or-losses:--- **AUC is not a win.** p = 0.054 is above 0.05. The correct statement is "at- least as good, consistent small edge," not "significantly better."-- **CodeScene's precision lead is a real design choice.** It flags **27** files- at its most severe level where we flag **132** At risk. That is a deliberately more conservative- operating point: a short list a team will actually work through, traded- against recall. If you want a handful of files to fix this quarter rather than- the ranking that catches the most defects, that operating point is better, and- it is not a calibration flaw.-- **We could not replicate their business-impact result.** CodeScene's "Code- Red" study reports a Pearson correlation of −0.58 between Code Health and- issue-resolution time on proprietary Jira data. On open GitHub data we- measured **−0.09 (95% CI −0.19 to +0.14)**, and six queue-independent effort- signals (commit span, review rounds, changes-requested, and others) all came- back flat with every interval spanning zero. The likely reason is structural —- GitHub merge time measures maintainer review-queue availability, not change- difficulty — but the honest summary is that **the business-impact axis remains- CodeScene's, unreplicated on open data.**--Reports:-[BENCHMARK_REPORT.md](https://github.com/repowise-dev/repowise-bench/blob/master/health-defect/BENCHMARK_REPORT.md) ·-[COMPARISON_REPORT.md](https://github.com/repowise-dev/repowise-bench/blob/master/health-defect/COMPARISON_REPORT.md)--## Three health signals: defect risk, maintainability, and performance--The three scored signals are computed from the same marker stream by one shared-scoring kernel against independent weight, category and cap tables. They are-co-equal views, **never blended into a single number**. A fourth dimension,-`advisory`, carries markers that never score at all. See below.--**Defect risk** is the calibrated headline: the number on the dashboard ring,-the band, the badge, and every accuracy claim above.--**Maintainability** exists because not every smell predicts bugs. The markers-the defect calibration floored are real problems about how hard code is to read-and change; floored inside a defect-framed score they do two unhelpful things at-once — nudge a calibrated number with noise, and get no credit for the problem-they actually describe. Maintainability weights are expert-set, tuned only-against their own category budget, never against the defect corpus.--**Performance risk** flags code whose *structure* wastes work — redundant I/O,-quadratic accumulation, contention — rather than measured runtime. It is-deliberately high-precision and low-recall.--The headline stays exactly the defect score, byte-for-byte, locked by a golden-test against the pre-split implementation. Blending would require a written-rationale and a recalibration corpus, because the band cutoffs, the 16.9×-separation and every AUC figure above are claims about the defect pillar-specifically.--Every finding carries a `dimension` naming its pillar, and all three surface-identically: `summary.*_average` on the REST overview, `kpis.*` on MCP-`get_health`, per-file scores on every metric row, and a line each in `CLAUDE.md`-and `repowise status`.--### The fourth dimension: `advisory`, which never scores--Some things are worth measuring and impossible to calibrate. No defect corpus-labels how much of a test is mock setup, so a mock-density marker has no-calibration story and never will. Weighting it anyway would move a number that-claims to predict bugs using evidence that says nothing about bugs.--`advisory` is where those markers live. It is deliberately **not** one of the-three scored dimensions: there is no weight table, category or cap keyed on it,-so "never deducts" is structural rather than a promise. Its findings carry a-`health_impact` of exactly `0.0`, are never counted in a change's-introduced/worsened totals, and can never make a review verdict blocking.--They are also out of every impact-ranked list, which is a statement about-ranking rather than about existence: a zero-impact row appended to a list-ordered by impact reads as a deduction that rounded away. Ask for the dimension-and you get it, and a surface that ranks nothing -- one file's findings in an-editor, or a list already filtered to one marker -- gets it without asking.--| Marker | Languages | What it measures |-|---|---|---|-| `assertion_free_test` | Python · TypeScript / JavaScript | A test case that runs the code under test and checks nothing |-| `mock_saturated_test` | Python · TypeScript / JavaScript | Mock-setup statements per assertion in a test function |-| `hidden_coupling` | every language with git history | Files that change together with no import between them |--`hidden_coupling` is here for the other reason: it was measured and carried no-signal. Alone it ranks defect-prone files near chance (AUC about 0.55), and a-pre-registered test on 12 repositories no earlier health study used found the-score without it non-inferior at predicting defects, so it stopped deducting.--A marker earns weight by clearing the house precision bar (roughly 70%-hand-labelled) on a real corpus. `mock_saturated_test` has not, and precision is-measured per language because it does not transfer: **67%** on Python (32-findings) and **40%** on TypeScript (30 findings), each the complete population-at the shipped gate. Both figures predate the assertion count gaining the-oracle shapes listed in-[LANGUAGE_SUPPORT.md](LANGUAGE_SUPPORT.md#code-health-coverage), which grew this-marker's denominator on Python and TypeScript. A larger denominator lowers the-ratio, so the change can only suppress findings, never add one — but the two-numbers above were measured against the smaller denominator and are now a-ceiling rather than a current reading.--It stays advisory until two false-positive families are separated: value-object builders named `Fake*` passed as input to real logic, and-boundary isolation where the assertion reads a real artifact. Both turn on-whether an assertion observes a double or production output, which is a dataflow-question the pass does not ask — and on the TypeScript sample that one question-accounted for every false positive.--`assertion_free_test` measures at least **71%** on TypeScript — a floor-rather than a current reading, three false-positive families having been closed-since the labelling (31 findings, the complete-population of two corpora; an earlier pass published 51% for the larger-pre-change population, which re-labels to 43.1% against this rubric) and **86%** on Python (29 hand-labelled, a systematic-sample of 172), and does not report on Go or Java at all; the per-language-reasoning is in-[LANGUAGE_SUPPORT.md](LANGUAGE_SUPPORT.md#code-health-coverage). A test that-delegates its oracle to a helper is resolved in the same file by name, and in-another file when the call graph binds the call, which is what closed most of-the TypeScript gap. The marker stays advisory on both: neither population is large-enough to settle the bar.--It asks a question with a yes-or-no answer rather than a-threshold, which is why it can be stated plainly: a test case with no oracle of-any kind — no assertion, no mock verification, no `raise`/`throw` the author-wrote by hand, and no call to an assertion helper from a position the statement-scan cannot classify. Four questions, every one of them yes or no and none of-them a threshold. A **mock verification counts as an-assertion here**, the opposite of the marker above it, because a test whose only-oracle is `verify(...)` does check something. The two markers read the same call-for two different questions; the tiers that keep them apart are in-`complexity/assertions.py`.--Per-language coverage, and why Go and Java are blocked for each marker:-[LANGUAGE_SUPPORT.md](LANGUAGE_SUPPORT.md#code-health-coverage).--## Performance risk--The pillar's core is **`io_in_loop`** — a database call, network request,-filesystem read or subprocess spawn executed once per loop iteration, the-classic N+1 — resolved through a shared I/O-boundary classifier and traced-across function boundaries as described above. Around it sit detectors for-quadratic string building, blocking calls inside `async` functions, connection-churn, lock contention, serial awaits that could fan out, membership tests-against lists, and language-specific shapes.--Two markers use call-graph centrality as a *precision gate* rather than a sort-key, firing only in a repo's most-called functions (at least the 80th-percentile-number of distinct direct callers, and never fewer than two), which keeps a-noisy shape reviewable. That shows how widely a function is called, not that a-request reaches it. The blocking-call marker is also skipped in tests, tooling,-examples, and generated or vendored code.--**Standard linters do not find this class of problem.** On a 12,600-file-benchmark, clippy, ruff, ESLint and golangci-lint together found **0** of the-cross-function I/O-in-loop cases; Repowise surfaced 557 findings of which ~87-span function boundaries, and 95%+ fell in categories ruff has no rule for.-Findings are ordered by impact rather than raw count (**NDCG 0.755** against-0.292 for severity-only ranking).--**Languages.** Performance analysis ships as a self-contained dialect plugin per-language — 11 dialects covering 16 language tags, including Python,-TypeScript/JavaScript, Java, Go, C#, Rust, Kotlin, Scala, Ruby, C++ and Dart. A-language without a dialect emits no findings, never a wrong one.--**Verified findings.** Two advisory markers (`serial_await_in_loop`,-`nested_loop_quadratic`) can be *promoted* when a def/use dataflow pass proves-the loop iterations are independent. A promoted finding asserts its fix instead-of hedging and carries `dataflow_verified: true`. The check is conservative —-anything it cannot prove stays advisory — and it changes the wording, not the-score.--**Soundness limits, by design.** Performance is a static signal, so it-under-reports rather than over-reports. Dynamic dispatch, monkeypatching and-callbacks-as-values produce no call edge and are invisible; an ORM lazy load-fires on attribute access with no visible call, so it is seen only where-`lazy_load_in_loop` can type the loop's rows (below) and missed everywhere else;-chains beyond three hops are not followed; an unmodelled library has no-classified sinks. We call this performance **risk**, never measured performance,-and it never folds into the defect score.--One honest caveat about actionability: `io_in_loop` reports a per-iteration-round-trip, it does not claim the work is avoidable. Database and network-findings are usually batchable; filesystem ones often are not, since deleting N-files genuinely needs N unlinks. The finding still tells you where the time-goes.--Each opportunity carries one actionability state: `plan_ready` (a proven-strategy on a reliable call path), `advisory` (a strategy whose prerequisites-are not all proven), `investigate` (no supported strategy yet) or `expected`-(the repetition is real and there is nothing to change: a filesystem or-subprocess boundary, or a loop already walking chunks). The default queue-leaves `expected` out and still counts it.--A few more things keep the plans honest:--- **Proven means the transformation, not the runtime.** Parallelizing awaits- against a database or network client is advisory with a- `bounded_concurrency` prerequisite, however clean the dataflow. A loop that- already walks its input in chunks (`range(0, n, CHUNK)`, `batched(...)`) is- reported as `loop_already_chunked` rather than told to batch.-- **One line, one problem.** When `io_in_loop` and `serial_await_in_loop` fire- on the same call, each names the other in `siblings`, and batching is queued- right before the parallelize variant.-- **Plans carry steps and validation.** Each performance plan lists its edits- (`mechanical` only when the strategy is proven) and the tests that validate- it: coverage, then call graph, then import graph, then a test named for the- file (`via: "name-match"`).--**Over-fetch.** `unbounded_read_reduced_in_memory` (Python, advisory) flags a-query with no limit or aggregate whose rows are then deduplicated per key in-code, in the same function or one same-file helper. It is the one shape here-that is not a loop around I/O: the query runs once and returns too much, and-the fix is to select one row per key in the database.--**ORM lazy loads.** `lazy_load_in_loop` (Python: sync SQLAlchemy and Django)-flags a lazy relationship read on each iteration of a loop over model rows, the-N+1 an attribute access hides. The loop's rows must resolve to one SQLAlchemy or-Django model through the cross-file model index, and the attribute must be a-lazy relation the producing query does not already eager-load. Async functions-are skipped, so SQLAlchemy coverage is sync only. Its plan is-`eager_load_relationship`, advisory, and names the eager load that replaces it-(`selectinload(...)`, `select_related(...)`). Each finding records its ORM in-`details.orm`, and the weight and the right to lead are read per finding:-Django measured 29/32 = 90.6% on a fresh held-out sample and carries weight 0.7-and may lead; SQLAlchemy has fewer than 30 held-out labels and stays at the-advisory 0.4 and never leads. One marker, one opportunity id: the ORM is a fact-on the finding, not a second marker.--Methodology and raw data:-[perf-detection](https://github.com/repowise-dev/repowise-bench/tree/master/perf-detection).--### Product flow--The web Code Health page has a dedicated **Performance** tab. It leads with a-bounded list of causal opportunities rather than a flat wall of observations:-the boundary and execution context, shared intervention, affected call-site and-file totals, confidence, and resolution provenance. Expanding an opportunity-shows caller-to-sink paths; raw findings remain canonical and load as a-separately paged evidence drill-down.--**One opportunity per intervention.** An opportunity is the place you edit,-within one cost family and boundary: the function holding the loop, or a helper-every caller reaches the sink through. The sinks it reaches, its call sites and-co-signals of one family (`io_in_loop` with `nested_loop_with_io`) are members,-so one loop reaching three sinks is one opportunity with three-`terminal_sinks`, not three opportunities. Every opportunity names its-`intervention_symbol` and `intervention_kind` (`function`, `shared_helper`, or-`module` for top-level script code, named `<file>::__module__`). Two loops in one-function are one edit site today, because findings carry the sink's line and not-the loop's.--**The default queue** (MCP `get_health` and the REST list) holds production-opportunities with a strategy: `plan_ready` and `advisory`. Test, tooling and-unclassified code, `expected` repetition and `investigate` causes (no-supported strategy, so no safe plan) are one filter away, and the summary's-`default_queue` block counts each reason it leaves out, so nothing is dropped-silently. Fix first (the dashboard lead, `repowise health` and the Overview's-Do next) takes its performance items through the same predicate and counts the-same reasons in `totals.excluded`.--**Order.** The queue is ranked by value (`rank_score`) first; actionability only-breaks a tie, so at equal value a plan-ready cause comes before an advisory one,-and a cheap plan never outranks a costlier cause.--When the deterministic service can describe a safe intervention, the-opportunity links by its exact stable `opportunity_id` to the matching-`performance_fix` plan on the existing **Refactoring** page. It never guesses a-nearby plan. When no safe plan exists, Code Health says so and keeps the raw-evidence available.+| `knowledge_loss` | A file that is still changing has one real author (bus factor 1), and that author is gone from recent commits or now makes under 20% of them |+| `ownership_risk` | A file with 5 or more commits has 3 or more minor contributors (each under 5% of commits) or no author above 40% |+| `developer_congestion` | 5 or more contributors, the file is in the top quarter of churn for the repository, and nobody owns half of it |-## Fix first+A file nobody touches never fires `knowledge_loss`: stable code with a departed+author is low risk. On a small team (few active contributors) the first two are+capped at low severity unless the file is a hotspot, since one owner per file is+the normal way such a team works. The `small-team` profile (see+[Configuration](#configuration)) lowers them further. Ownership and bus factor per+file also appear in `get_context(include=["ownership"])` and on the dashboard.-**Fix first** is one ranked list of what to fix in a repository, built once in-core (`analysis/health/fix_first/`) from stored rows and rendered the same way by-`get_health()`, `repowise health`, the Code Health page, the Overview's Do next and-the generated CLAUDE.md. An item is one unit of work: a file's composed-refactoring, one performance intervention (the loop you change), or the strongest-code-shape finding of a file that has no plan. Each item says what to change and-where (file and line), why it matters there, the first concrete edit, the gain, the-effort and the tests to run after.+### Test-quality smells-**What may be an item.** A unit must be shipped code with a concrete first edit-that is worth doing. Everything else is left out and counted in-`totals.excluded`, one reason per unit:+`large_assertion_block` and `duplicated_assertion_block` fire only in test files+and share a 0.5 cap. Two advisory markers ask harder questions with no defect+corpus to calibrate against, so they never deduct:-| Reason | What it leaves out |-|---|---|-| `test` | Test files. `scope=all` keeps them, labelled. |-| `tooling` | Scripts, tools, benchmarks, CI files, migrations, code under a directory whose role is unknown (docs, demos), and build files by type wherever they sit (Gradle scripts, `CMakeLists.txt`, `*.cmake`, Makefiles, Bazel files, MSBuild `.props` / `.targets`, root `setup.py`, `noxfile.py`, crate-root `build.rs`, bundler configs). |-| `generated` | Generated files. |-| `vendored` | Third-party code: `vendor/`, `third_party/`, `node_modules/` and similar directories, minified files, and a library with its own license header among served assets. |-| `docs_example` | Documentation code and examples: `docs/`, `docs_src/`, `examples/`, tutorials, samples. |-| `unknown` | A performance cause whose code context could not be classified. |-| `expected` | A performance cause whose repetition is expected. |-| `no_strategy` | A performance cause with no supported fix strategy. |-| `no_plan` | A performance cause with no stored safe plan to quote. |-| `below_min_worth` | A refactoring that recovers under 0.5 health on its file, or has no steps; a string built in a loop that is bounded or not in production code. |-| `history_only` | A file whose only findings come from git history (churn, ownership, co-change). History is context on an item, never the item. |-| `deprecated` | A function marked deprecated. |-| `inherent_dispatch` | A function where one dispatch on one value holds 60% or more of its decision points, unless a duplicate also sits in it (then it is listed as `later`). On the dev labels that share held 9 complexity rows, 8 of them rejected. |-| `small_function` | A complexity unit under 30 code lines and under CCN 15. On the 67 labelled dev rows that cut drops 13 rejected and 4 accepted. |-| `no_concrete_step` | No first edit with a file and a line or a named group: a cycle with no import line to cut, a move with no destination, a split with no named groups, a class finding with no member groups. |-| `low_value_kind` | A kind raters found not worth doing: Extract Class (0 of 14), Move Method (0 of 34), low cohesion (0 of 46), long method (0 of 10), and long parameter lists (0 of 2, thinly measured). |--**Order.** Items rank by value first: the larger of the health they recover and-the size of the problem (CCN 20, 40, 80 and 150; 100, 200, 400 and 800 lines;-nesting 5, 6 and 8; a critical or brain-method finding; one more on a hot file),-so a function far past every bar leads ahead of tidy work. A performance fix's-value is its cost: a production, entry-reachable database or network call in a-loop that grows with the data shares the top band. Then tier: `now` (worth doing,-and the plan is safe to start), `next` (worth doing, the fix needs judgment),-`later`. No kind takes more than 3 of the first 5 places while another has an item-worth doing.+- `assertion_free_test` (Python, TypeScript/JavaScript): a test that runs code and+ checks nothing. No assertion, no mock verification, no hand-written raise, no+ call to an assertion helper.+- `mock_saturated_test` (Python, TypeScript/JavaScript): mock setup statements per+ assertion in a test function.-**Lower priority.** A `later` item is real and stays listed, after every `now` and-`next` item, with the reason it can wait as its tier reason. Code shape decides-(`analysis/health/worth.py`, one rule for every default list); the hot-file bonus-orders items but never lifts one out of `later`, and a later item's rank facts-read "value within later" and its shape-only size. A function-size problem is-`later` when the function is under CCN 40, 200 lines and nesting 6 (nesting under-8 counts only in a function of 100 lines or more) and not both CCN 25 and nesting-5; when one dispatch on one value holds 60% of its decisions; when its nesting is-one else-if or ternary chain; or when it is long with CCN under 20 and nesting-under 5, however long. Otherwise, from CCN 80, 400 lines or nesting 8 a-function is worth doing whatever its branching. A complex condition and a single error site (a swallowed or broad-catch, an unwrap or panic) are `later` too: each is a local fix. The same rule orders the-default findings list (`get_health`, the REST findings list): findings worth doing-first lead, and each other one carries `lower_priority`, the reason it can wait.-Class-design findings (god class, low cohesion, long parameter lists) and findings-that rest on git history alone are lower priority there. A performance-opportunity carries `lower_priority` unless production code runs it over data-that grows.--**Verify.** Each item carries up to 5 tests from its stored validation profile,-each with how it reaches the changed code (call graph, import graph, a matching-name or a coverage report), and the command that runs them. A finding with no-plan says so and names none.--## Refactoring targets+Add house assertion helpers with the `assertions:` block in+[CONFIG.md](../reference/CONFIG.md#the-assertions-block). Per-language coverage is+in [LANGUAGE_SUPPORT.md](LANGUAGE_SUPPORT.md#code-health-coverage).-```bash-repowise health --refactoring-targets-```+### Scope and counts-A score tells you a file is in trouble; a refactoring target names the fix.-Repowise emits structured suggestions computed deterministically during the-health pass from data it already has — the call graph, the cohesion model, the-clone pairs, git co-change. No re-parse, no LLM, inside the same budget. Seven-detectors ship:+Two controls decide what a figure describes. Both work on the CLI, MCP and every+dashboard page, and every response echoes them.-| Type | What it names |-|---|---|-| **Extract Class** | The cohesion groups an incohesive class should split into — exact methods and fields per group |-| **Extract Method** | A behavior-preserving extraction, with the IN/OUT signature inferred from reaching-definitions and liveness analysis |-| **Extract Helper** | A clone's exact occurrences and where the shared helper should live |-| **Move Method** | A feature-envy method and the class it actually belongs to |-| **Break Cycle** | The minimal set of import edges to invert to break a dependency cycle |-| **Split File** | The cohesive files an oversized module should decompose into, plus the import edits in every dependent |-| **Performance Fix** | A proven shared intervention for a causal performance opportunity, with affected call sites and caller-to-sink paths |--Each suggestion is structured data, not a string: a `plan`, the `evidence` that-justifies it, the `impact_delta` it recovers, an `effort_bucket`, and a-`blast_radius` of callers and co-changing files that must move with it. Ranking-is graph-aware. The canonical recommendation service separates benefit and-leverage from cost and risk; a larger blast radius raises cost and risk and is-never presented as benefit. Performance plans retain detector-native benefit-even though they intentionally recover zero defect-health points.+| Control | Values | Effect |+|---|---|---|+| `scope` | `all` (default), `production` | `production` drops test files. It lowers the headline because tests usually score higher than the code they cover. |+| `counts` | `everything` (default), `code_shape` | `code_shape` removes the history half of each deduction. Use it to ask "is my code getting better" after a week of refactoring raised churn. |-```python-get_health(include=["refactoring"]) # ranked structured plans-get_health(targets=["src/api/server.py"]) # one file in detail-```+`everything` is the calibrated number every accuracy claim refers to, and the+badge always reports it.-Full reference: **[REFACTORING.md](REFACTORING.md)**.+## Fix first++**Fix first** is one ranked list of what to fix. `repowise health` prints the top+3, `get_health()` and the dashboard Overview tab show more, and `repowise next`+mixes its top items with work from other layers. Each item is one unit of work: a+file's refactoring plan, one performance intervention, or the strongest code-shape+finding in a file with no plan. It says what to change and where, why, the first+edit, the expected gain, the effort and the tests to run afterwards.++What stays out, and is counted as excluded with a reason: test files, tooling,+generated and vendored code, docs and examples, files whose only findings come+from git history, changes that would recover under half a point, items with no+concrete first edit, and kinds that reviewers found not worth doing.++Items rank by value first (health recovered, or how far the code sits past size+and complexity bars), then by tier: `now` (worth doing, safe to start), `next`+(worth doing, needs judgment), `later` (real but can wait; the reason is shown).+No single kind takes more than 3 of the first 5 places while another kind has work+worth doing.-## Trends and coverage+### Do next: `repowise next`-Every health run writes a snapshot (rolling 50 per repo) storing both repo KPIs-and a per-file score map, which powers two repo-level alerts — **Declining-Health** and **Predicted Decline** — and a per-file score trajectory with a-delta and declining flag. Trends stay silent on thin history rather than showing-a misleading single point.+`repowise next` prints the short cross-layer list the dashboard calls **Do next**+and `get_overview` returns as `next_actions`. It reads stored actions from the+index, so it is instant and makes no analysis pass. Fix first items sit beside+work from other layers: a live secret, fresh regressions, fragile files, files+with knowledge loss, broken doc references, a dead-code batch, missing or stale+coverage, and decisions waiting for review.```bash-repowise health --trend+repowise next # top 5 for this week (or the quarter when the week is empty)+repowise next --horizon quarter --all+repowise next --json # the whole stored view```-Coverage reports light up the test-coverage markers. **LCOV**, **Cobertura**,-**Clover**, **JaCoCo**, **Go coverprofile** and a normalized JSON format are-auto-detected:+Rows are grouped as **Now**, **Worth planning** and **Improve what Repowise can+see** (steps such as adding a coverage report that make other answers sharper).+Run `repowise update` to refresh the list after you change code.++## Performance findings++Performance risk flags structure that wastes work. It does not measure runtime.++| Kind | Examples |+|---|---|+| I/O per loop iteration (N+1) | `io_in_loop`, `nested_loop_with_io`, `lazy_load_in_loop` (Django and sync SQLAlchemy) |+| Concurrency | `serial_await_in_loop`, `blocking_sync_in_async`, `blocking_io_under_lock`, `lock_in_loop`, `goroutine_in_unbounded_loop` |+| Repeated setup | `resource_construction_in_loop`, `regex_compile_in_loop`, `json_parse_in_loop`, `defer_in_loop` |+| Quadratic or wasteful data work | `string_concat_in_loop`, `nested_loop_quadratic`, `membership_test_against_list_in_loop`, `list_insert_zero_in_loop`, `array_spread_in_reduce`, pandas `concat`/`iterrows` in loops |+| Over-fetch | `unbounded_read_reduced_in_memory`, `sql_cartesian_join` |++The loop and the I/O call do not have to sit in the same function: Repowise follows+the resolved call graph up to three hops and attaches the caller-to-sink path.+Per-function linters cannot see that case.++Findings group into **opportunities**, one per place you would edit (the function+holding the loop, or a helper every caller goes through). Each opportunity carries+one state: `plan_ready` (a proven fix strategy), `advisory` (a strategy with+unproven prerequisites), `investigate` (no supported strategy) or `expected` (the+repetition is real and nothing should change, such as deleting N files). The+default queue holds production `plan_ready` and `advisory` opportunities, ranked by+cost first; state only breaks ties. Everything left out is counted by reason.++Performance analysis covers Python, TypeScript/JavaScript (including Vue and Svelte+scripts), Java, Go, C# (including Razor), Rust, Kotlin, Scala, Ruby, C++, Dart and+Pascal. A language without support emits no performance findings, never a guessed+one. Dynamic dispatch, monkeypatching and callbacks passed as values leave no call+edge, so those cases are missed.++## Refactoring++Each low-scoring file can carry deterministic refactoring plans (Extract Method,+Split File, Break Cycle, Performance Fix and others) with evidence, gain, effort and+blast radius. `repowise health --refactoring-targets` prints the stored queue. See+[REFACTORING.md](REFACTORING.md).++## Trends, coverage and badges++Every run stores a snapshot (the last 50 per repository) with repository KPIs and+per-file scores. Alerts read them: **declining** (the score fell 0.5 or more+against five snapshots back), **predicted decline** (three falls in a row), and+**history drag** (either of those, where only the git-history half moved, so there+is nothing in the code to act on).+`repowise health --trend` prints the last 10 snapshots. Trends stay quiet on thin+history.++Coverage reports feed the test-coverage detectors. LCOV, Cobertura, Clover,+JaCoCo, Go coverprofile and a Repowise JSON format are detected automatically.+With no report, only `untested_hotspot` runs, judging a hotspot by whether any test+reaches it; the line-coverage detectors stay silent, so no file is treated as+uncovered.```bashpytest --cov --cov-report=lcov:coverage.lcov@@ -765,9 +276,13 @@ repowise coverage add coverage.lcovrepowise health```+`repowise health --badge` prints README Markdown: a static shields.io badge for the+current score and the live form served by a running Repowise server+(`/api/repos/<id>/health/badge.json`, or `badge.svg`).+## Configuration-Per-file overrides live in `.repowise/health-rules.json`:+Per-file rules live in `.repowise/health-rules.json`:```json{@@ -780,43 +295,96 @@ Per-file overrides live in `.repowise/health-rules.json`:}```-`path` is a gitignore-semantics glob over the repo-relative path. The-`small-team` profile demotes the process and people signals a 1-3 person repo-cannot support; an explicit `severity_overrides` key always wins over it.--**Only the severity label is tunable.** The per-marker weight multipliers and-the category caps are the calibrated constants the published accuracy numbers-rest on, and they are deliberately not overridable — so a team's local policy can-never silently change what those numbers mean.--## Finding identity and triage--Every finding carries a stable public id (`finding_<digest>`). Inside a known-function or class it is anchored on that symbol's name plus the finding's line-offset into it, so an edit above the symbol does not change the id; a-file-level finding keeps its absolute lines. Metric values (a CCN, a churn-figure, a coverage percentage) are not part of the id, so a finding keeps its-id while its numbers move; only the few detail keys a marker needs to tell two-of its findings apart (a coupling partner, an error kind) are.+- `path` uses gitignore semantics over the repo-relative path. Every matching rule+ applies. A single `*` stops at a directory, so write `src/legacy/**` for a whole+ subtree.+- `small-team` lowers the people and process signals a 1-3 person repository+ cannot support. An explicit `severity_overrides` entry wins over the profile.+- Only severity labels are tunable. Weights and category caps are the calibrated+ constants the published accuracy rests on, so local policy cannot change what+ those numbers mean.++Full schema: [CONFIG.md](../reference/CONFIG.md#the-health-rulesjson-file).+Refactoring detectors and their confidence floor are set in the `refactoring:`+block of `.repowise/config.yaml`.++**Triage.** Each finding has a stable id that survives edits above it and changes+in its metric values. Mark a finding `acknowledged`, `false_positive` or `resolved`+in the dashboard, over REST, or with the opt-in `set_finding_status` MCP tool.+Triage survives re-indexing; a `resolved` finding that is detected again reopens.++## Accuracy and limits++Measured leakage-free: each file scored at a commit before the bug window, then+checked against the bug fixes of the next six months.++- **Defect prediction:** ROC AUC **0.737** (95% CI 0.683 to 0.787) across 21+ repositories, 9 languages and 2,826 files; per-repository range 0.55 to 0.86.+- **Held-out data:** 0.761 and 0.776 on PROMISE jEdit 4.0 and 4.1, a public dataset+ with no git history that played no part in calibration.+- **Against CodeScene** (same 2,770 files, same commit, same labels): recall at a+ 20%-of-lines review budget **0.173 vs 0.074** (p = 0.003), surfacing 2.3 times the+ defects. AUC 0.731 vs 0.705 is not significant (p = 0.054), and CodeScene's+ precision at that budget is a tie (0.636 vs 0.580, p = 0.64).+- **Calibration:** marker weights come from a regression on that defect corpus with+ file size as an explicit control, so a marker earns weight only for risk beyond+ being big.+- **Not better than file size at discrimination:** lines of code alone scores AUC+ 0.742 (a tie). The gain is in ordering a fixed review budget and in saying why.+- **Weak among files of similar size:** within size quartiles AUC runs 0.525 to+ 0.718; the signal holds cleanly only in the largest quartile.+- **Business impact not replicated:** the score did not track GitHub merge time on+ open data.+- **On your repository:** after an index, Repowise reports how many of the 20+ lowest-scoring files had a bug fix in the last 180 days against the repository's+ base rate (`get_health(include=["accuracy"])`). It is an association on indexed+ history, not a forward test, and stays silent below 25 scored files or 5 fixed+ ones.+- **Performance risk** is a static, high-precision, low-recall signal. It+ under-reports by design.++Method, sample sizes and every test: [BENCHMARKS.md](../BENCHMARKS.md#code-health-predicts-defects).+How the calibration works: [architecture/code-health.md](../architecture/code-health.md#62-validation-and-calibration).++## Where it shows up++| Surface | What you get |+|---|---|+| CLI | `repowise health`, `repowise next`, a line in `repowise status` |+| MCP | `get_health`; health fields on `get_risk`, `get_context(include=["health"])` and `get_overview` |+| Dashboard | Code Health page: Overview (Fix first), Performance, Findings, Tests |+| CLAUDE.md | Health KPIs and Fix first items in the generated file |+| Editor | Per-file findings in the [VS Code extension](../agent/VSCODE.md) |-Triage survives re-indexing. Each index replaces the open findings, and a-re-detected finding whose id matches a triaged row updates that row's evidence-(lines, severity, reason, details, impact) in place: `acknowledged` and-`false_positive` stay as they are, and `resolved` reopens, because the finding-marked fixed was detected again.+## Reference-## Incremental updates+`repowise health [PATH]`-`repowise update` re-scores only changed files. Findings and metrics for-unchanged files stay put; no nightly full re-index.+| Flag | Effect |+|---|---|+| `--file PATH` | Report one file (repo-relative path) |+| `--module PREFIX` | Report only files under this path prefix |+| `--scope all\|production` | Population to report on (default `all`) |+| `--counts everything\|code_shape` | Include or drop the history half of the deduction (default `everything`) |+| `--format table\|json\|md` | Output format; `json` and `md` do not write to the index |+| `--refactoring-targets` | Print the stored refactoring queue |+| `--recompute` | With `--refactoring-targets`: analyze the working tree in-process |+| `--generate-code SELECTOR` | Opt-in: generate code and a diff for one suggestion with the configured LLM (needs an API key) |+| `--trend` | Print the last 10 snapshots |+| `--badge` | Print README badge Markdown |+| `--repo ALIAS` / `--no-workspace` | Pick a workspace repository, or force single-repo mode |+| `-v`, `--verbose` | Show pipeline debug logs |++`repowise next [PATH]` takes `--horizon week|quarter`, `--all` (20 rows, not 5),+`--repo`, `--no-workspace` and `--format`/`--json`.++MCP: [`get_health`](../agent/MCP_TOOLS.md#get_health).## See also-- [`docs/architecture/code-health.md`](../architecture/code-health.md): internals,- the full marker roster, and the complete weight tables.-- [`docs/BENCHMARKS.md`](../BENCHMARKS.md): every published number with its- sample size and test.-- [DOC_DRIFT.md](DOC_DRIFT.md): the other thing this layer checks, your own- documentation against the tree.+- [architecture/code-health.md](../architecture/code-health.md): internals, weight+ tables, calibration, Fix first and performance ranking rules.+- [BENCHMARKS.md](../BENCHMARKS.md): every published number with its sample and test.- [REFACTORING.md](REFACTORING.md) · [TEST_INTELLIGENCE.md](TEST_INTELLIGENCE.md) ·- [BUG_HISTORY.md](BUG_HISTORY.md)+ [BUG_HISTORY.md](BUG_HISTORY.md) · [DOC_DRIFT.md](DOC_DRIFT.md) ·+ [LANGUAGE_SUPPORT.md](LANGUAGE_SUPPORT.md#code-health-coverage)diff --git a/docs/layers/DEAD_CODE.md b/docs/layers/DEAD_CODE.mdindex d4d5bb45f..08749abd4 100644--- a/docs/layers/DEAD_CODE.md+++ b/docs/layers/DEAD_CODE.md@@ -1,374 +1,247 @@# Dead CodeRepowise reports the files nothing imports, the exported symbols nothing uses,-and the packages nothing depends on, each with a confidence score and the-evidence behind it. Pure graph traversal and SQL: no LLM calls, no network, and-it finishes in under 10 seconds on any repo size.+and the packages nothing depends on. Each finding carries a confidence score,+a deletion-readiness flag and the evidence behind both. The analysis is graph+traversal plus git history: it needs no LLM key, makes no network calls, and+runs as part of `repowise init` and `repowise update`.-The layer surfaces candidates. You decide. Static reachability can prove that-nothing *imports* a file; it can never prove nothing *loads* it. Everything below-is built around that asymmetry.+Treat every finding as a candidate. Static reachability can show that nothing+*imports* a file; it cannot show that nothing *loads* it. The confidence+scores, caps and exemptions below all follow from that gap.## Quick start```bash-repowise init # dead-code findings populate during indexing-repowise dead-code # the report-repowise dead-code --safe-only --min-confidence 0.8 # cleanup-ready only+repowise init # findings are computed during indexing+repowise dead-code # the report+repowise dead-code --safe-only # deletion-ready findings onlyrepowise dead-code --kind unused_export --format json-repowise dead-code --repo backend # workspace, one repo``````-repowise dead-code+ Kind File / Symbol Confidence Ready? Lines Reason+ unused_export formatLegacyDate 100% ✓ 14 Public symbol 'formatLegacyDate' has no importers+ unreachable_file src/legacy/parser.ts 100% ✗ 212 File has no importers (in_degree=0)+ unreachable_file config/feature_flags.py 40% ✗ 31 File has no importers (in_degree=0)- 23 findings · 4 safe to delete-- ✓ utils/legacy_parser.ts file 1.00 safe to delete- ✓ auth/session.ts file 0.92 safe to delete- ✓ helpers/formatDate export 0.71 safe to delete- ✗ analytics/v1/tracker.ts file 0.41 recent activity, review first-```--From an agent:--```python-get_dead_code()-get_dead_code(min_confidence=0.8, tier="high", safe_only=True)-get_dead_code(kind="unused_export", group_by="owner")+Cleanup-candidate lines: 14 (high 2, medium 1 confidence)```-## The four finding kinds--| Kind | What it means | How it is computed | Base confidence |-|------|---------------|--------------------|-----------------|-| `unreachable_file` | No file in the repo imports this one. | File-node in-degree of 0 on the dependency graph, after entry points and the never-flag allowlist are removed. | Scored from git age (below) |-| `unused_export` | A public symbol nothing imports. | No `imports` edge names the symbol (or `*`, or a TypeScript `export { local as alias }` rename), and no `calls` / `extends` / `implements` / `method_implements` / `dispatches_to` / `framework_binds` / `reads` / `references` / `type_use` edge reaches it. Member kinds (`method`, `field`, `property`, `enum_member`) are excluded from this pass across all languages since they are accessed through their container, not imported by name. Top-level `export const` primitives, objects, and arrays in TypeScript and JavaScript are fully evaluated (they are always importable by name). | `1.00` when an importer of the containing file names one of its symbols (so the file is alive and its importers list what they take, but not this symbol); `0.60`, never safe to delete, when no importer names a symbol of the file (no importer at all, a C `#include`, a C# `using`), for every C, C++ and Objective-C symbol (the preprocessor reaches symbols through macros and typedef aliases no edge records), and for a C# class holding extension methods (called as `x.Method()` without naming the class); `0.30` when the name ends in `_DEPRECATED` / `_LEGACY` / `_COMPAT` — then capped as described below |-| `unused_internal` | A private or underscore-prefixed symbol nothing calls. | No `calls` edge, and no cross-file importer pulls the name (which would mean a dispatch-table lookup). Off by default. | `0.65` |-| `zombie_package` | A whole top-level package no other package imports. | No inter-package import edges into it. Never marked safe to delete. | `0.50` |--`unused_internal` is opt-in (`--include-internals`); `zombie_package` is on by-default and can be turned off with `--no-include-zombie-packages`. Passing-`--kind` overrides both toggles, so `--kind unused_internal` enables internals on-its own.--### An unused export must survive a search for its own name--The base confidences above score how strong the evidence for deadness is. They-do not ask whether a use would have been visible at all — and "no import edge"-only means "unused" in a language where using a symbol requires importing it.-That holds in Python and TypeScript. It is false for a same-package Kotlin or-Go reference, a same-translation-unit C++ type, an intra-crate Rust path, and a-C# or Swift member of the same module, where a use needs no import. The same-blind spot swallows a use written through an aliased import, an attribute call-on an imported module, or a handler named by string from infrastructure config.--So before an `unused_export` keeps a confidence above `0.40`, the repository is-searched for the symbol's name. If the name is written anywhere other than the-declaration itself — in any indexed file, including non-code ones such as-Terraform or YAML — the finding is capped at `0.40`, loses `safe_to_delete`,-and its reason names where the name was found. The same cap applies when the-search could not be run at all, because not looking and looking without finding-are different results.--The check only ever suppresses. A name written solely in a comment counts as a-use, so this costs recall and buys precision, which is the trade the top tier-exists for. Nothing is removed from the report: the cap lands exactly on the-default `min_confidence`, so a capped finding still appears as a review-candidate and only stops claiming to be deletion-ready.--Two caveats on the numbers. `unused_internal` is disabled entirely for Rust,-because rustc's own `dead_code` lint already reports unused private items with-macro expansion and type information this analysis cannot match, so a private-Rust helper is never flagged here. And the `lines` count on file and package findings is an-estimate (symbol count times ten), not a real line count, so treat the-"reclaimable lines" roll-up as an order of magnitude rather than a figure.--A .NET reference assembly (`ref/*.cs`) is the compile-time API of a library.-Its files are never reported, and neither is a C# file or type it lists: that-is public API, used outside the repository.--### A C or C++ name written anywhere else is a use--C, C++ and Objective-C symbols are used in ways that carry no edge: a callback-passed to `SetTimer`, a function reached through a `#define` alias, a P/Invoke-export named in a C# `[LibraryImport]`, an icall registered in a table header,-a `.def` EXPORTS line, an assembly label. So an `unused_export` or-`unused_internal` in these languages is dropped when any of the names its-declaration introduces is written outside a declaration of it, in any code-file or in a `.def`, `.asm` or `.s` file. A typedef contributes its tag and-every alias (`typedef struct _X {...} X, *PX;`), and an enum its enumerators.-Declarations do not count (the header prototype of a `.cpp` function), and-neither do comments, prose strings and documentation. A COM method declared-with `IFACEMETHODIMP` or `STDMETHODIMP` fills an interface slot and is never-reported.--### How unreachable-file confidence is scored--An orphaned file that nobody has touched in a year is a much stronger signal than-one added last week. Confidence starts from git activity:--| Condition | Confidence |-|-----------|-----------|-| No commits in 90 days, last touched over a year ago | `1.00` |-| No commits in 90 days, last touched over 180 days ago | `0.90` |-| No commits in 90 days, last touched over 90 days ago | `0.80` |-| No commits in 90 days, no other signal | `0.70` |-| No commits in 90 days, but the file is under 30 days old | `0.55` (may be work in progress) |-| Still being committed to | `0.40` |--Then it only ever goes down. These caps apply:--- **Dynamic imports nearby.** If any file in the same directory uses a runtime- loader, confidence is capped at `0.40`.-- **Imported by namespace.** A C# `using` names a namespace, never a file, and- a same-namespace `new T()` needs no `using` at all, so a C# file is capped at- `0.40` whatever its age.-- **Its type is named elsewhere.** When another file writes the name of a type- the file declares, confidence is capped at `0.40` and the evidence names- that file: Java, C# and Swift use a type from its own package or module- without an import.-- **Runtime-load risk factors.** If the path looks like config, environment,- bootstrap, database, script, or runtime-asset code, confidence is capped at- `0.40` and the finding carries an evidence line explaining why. These are- exactly the files wired up by a config key or a string path rather than an- import, so "nothing imports it" is weak evidence. The full token set, matched- against the filename split on `. _ -` and against directory segments:-- | Factor | Filename tokens | Directory segments |- |---|---|---|- | `config` | `config`, `configs`, `configuration`, `conf`, `settings`, `setting`, `setup` | `config/`, `configs/`, `settings/` |- | `environment` | `env`, `environment`, `environ`, `dotenv` | `env/`, `environments/` |- | `bootstrap` | `bootstrap`, `startup`, `entrypoint` | `bootstrap/` |- | `database` | `database`, `db`, `schema`, `seed`, `seeds`, `migration`, `migrations`, `datastore`, `sqlite` | `database/`, `db/`, `migrations/` |- | `script` | — | `scripts/`, `bin/`, `tasks/` |- | `asset` | `sw`, plus the token pair `service` + `worker` | `public/`, `static/`, `www/` |-- Broad identifiers (`app`, `main`, `index`, `core`, `base`, `util`) are- deliberately excluded: they would cap ordinary modules. `src/assets/` is- likewise absent, because that is the Vite / Vue / Angular convention for- *bundled* source, which is imported normally.--## Confidence tiers and `safe_to_delete`--A finding is presented as **safe to delete** only when confidence is at or above-`0.70` **and** the path carries no runtime-load risk factor **and** the name does-not match a dynamic-dispatch pattern (`*Plugin`, `*Handler`, `*Adapter`,-`*Middleware`, `*Mixin`, `*Command`, `register_*`, `on_*`, `*_view`,-`*_endpoint`, `*_route`, `*_callback`, `*_signal`, `*_task`). Zombie packages are-never safe to delete regardless of confidence.--The safety re-derivation is monotonic: it only ever downgrades a stored flag,-never upgrades it, so findings written by an older version stay honest.--The two surfaces bracket the tiers differently, which is worth knowing before you-compare numbers:--| Surface | High | Medium | Low | Default floor |-|---------|------|--------|-----|---------------|-| CLI (`repowise dead-code`) | `>= 0.7` | `0.4` to `0.7` | `< 0.4` | `--min-confidence 0.4` |-| MCP (`get_dead_code`) | `>= 0.8` | `0.5` to `0.8` | `< 0.5` | `min_confidence=0.4` |--Both surfaces share the same default floor (`0.4`), so both return the same-findings. Only the tier labelling differs: MCP brackets both cutovers higher-(`0.8`/`0.5` against the CLI's `0.7`/`0.4`), because an agent acting on a-finding is riskier than a human reading a table. The two rows are deliberately-different, so do not flatten them to matching numbers.--When findings exist below the floor — for example, deprecated exports-(confidence `0.30`, see the kind table above) — the CLI prints a dim footer-line: `N finding(s) hidden below threshold; pass --min-confidence 0.0 to see-them.` The report and its buckets are unaffected; the footer is the only-visible change.--## What is exempt by construction--Before anything is scored, repowise removes what it knows is framework-loaded,-generated, or convention-wired. Static reachability is simply the wrong tool for-these, so they are never flagged rather than flagged and down-weighted.--| Group | Examples |-|-------|---------|-| Entry points | Anything the graph marked `is_entry_point`, plus `__init__.py`, `__main__.py`, `conftest.py`, `manage.py`, `wsgi.py`, `asgi.py`, `setup.py`, `main.go` |-| Build files | Any file a build tool runs by name, classified by type in `code_origin`: Gradle `*.gradle(.kts)`, `pom.xml`, `CMakeLists.txt`, `*.cmake`, Makefiles, `meson.build`, Bazel `BUILD` / `*.bzl`, MSBuild `.props` / `.targets`, crate-root `build.rs`, `magefile.go`, `noxfile.py`, bundler configs. A directory holding only build files is not a package |-| Shell scripts | `*.sh`, `*.bash`, `*.zsh`. Invoked by name from CI configs and Makefiles; static reachability is meaningless |-| Programs | Any file whose first line is a shebang, and any Python file with a top-level `if __name__ == "__main__":` block. Nothing imports an entry point |-| Files a runner names | A file a CI workflow (`.github/workflows/`, `.gitlab-ci.yml`, `.circleci/`, `.buildkite/`), build file (above), task file (`Dockerfile`, `tox.ini`), manifest (`pyproject.toml`, `package.json`, `setup.cfg`) or shell script names by path. A doc that names a file only caps it at `0.40` |-| Build-named JVM classes | A class whose fully-qualified name a `build.gradle(.kts)` quotes: `esplugin { classname '...' }`, `implementationClass`, `mainClass` |-| Framework routes | Next.js `page.tsx` / `layout.tsx` / `route.ts` / `middleware.ts`, SvelteKit `+page.svelte`, Nuxt `pages/*.vue`, Remix entry files, ASP.NET minimal-API `Apis/` / `Endpoints/`, Blazor and Razor code-behind |-| Test files | `*_test.go`, `*.test.ts`, `*.spec.ts`, `*_test.cc`, `*Test.java`, `**/tests/*.rs`, `src/test/java/`, MSTest and xUnit project layouts, `__tests__/`, `__mocks__/` |-| Generated code | protoc `*.pb.go` / `*.pb.cs` / `*.pb.cc`, Qt MOC/UIC/RCC, Bison/Flex, SWIG, Cython, stringer, MapStruct `*MapperImpl.java`, Dagger, AutoValue, Roslyn `*.g.cs`, Dart `*.g.dart` / `*.freezed.dart`, `**/generated/**` |-| Reflective loading | Alembic `versions/*.py`, Django migrations, EF entity configurations, COM `*ClassFactory.cpp`, Win32 `*NativeMethods.cs`, ETW event classes |-| Vendored trees | `vendor/`, `third_party/`, `deps/`, `external/`, `extern/`, `contrib/`, `submodules/` |-| Build artifacts | `build/`, `cmake-build-*/`, `_deps/`, `*.min.js`, `*.bundle.js` |-| Non-code languages | Config and infra languages from the language registry, plus anything the parser could not identify |--Symbols decorated by a framework are treated as live too: pytest fixtures, Flask-and FastAPI routes, Django `admin.register` and signal receivers, Celery tasks,-Click and Typer commands, and the JVM stereotype and routing annotations-(`@Component`, `@Service`, `@RestController`, `@Entity`, `@KafkaListener`,-`@GetMapping`, `@Test`, JAX-RS `@Path` / `@GET`), and JMH `@State`,-`@BenchmarkMode` and `@Benchmark`. Every annotation of a Java or Kotlin-declaration counts, not only the first. A Java, Kotlin or Scala export whose-name is written in another code file, or further down its own file, is-dropped: those languages use a type from its own package by its bare name,-with no import. Decorator *suffixes* are matched-too, so `@my_local_group.command` and `@api.get` register even when the receiver-has a project-local name.--Two more targeted rescues: an `interface` in a file with no incoming `implements`-edges is capped at `0.40` (implementor detection is heuristic, and missing-evidence is not evidence of absence), and COM contract methods-(`QueryInterface`, `AddRef`, `Release`) are capped the same way because they are-dispatched through native vtables.--**Symbol-kind exemptions in the unused-export pass.** Class methods, struct-fields, properties, and enum members are never evaluated as unused exports across-all languages — they are only reachable through their enclosing container, not-independently importable. Top-level `export const` declarations in TypeScript and-JavaScript (`export const API_URL = "..."`, `export const config = {...}`) are-*not* exempt: they are module-level named exports and are fully evaluated. Non-exported-constants (declared without `export`) are automatically private and never reach-the pass.--Zombie-package detection additionally ignores directories that are not packages-at all: `.github`, `.vscode`, `.devcontainer`, `docs`, `scripts`, `assets`,-`static`, `public`, `tests`, `benches`, `fuzz`, and their siblings, and any-directory whose code is only Dockerfiles, Makefiles and shell scripts, which-are run rather than imported.--## Dynamic-import awareness--When a file uses a runtime loader, repowise assumes its neighbours may be reached-through it and caps their confidence at `0.40`. Detected markers include:--- **Python**: `importlib.import_module`, `__import__(`, `importlib.reload`, `pkgutil.iter_modules`-- **JS/TS**: dynamic `import(`, `require.context(`, `import.meta.glob(`,- `React.lazy(`, `next/dynamic`, `jest.mock(` / `vi.mock(`, and the- `'use server'` / `'use client'` boundary directives--This is a text scan over source, grouped by file extension. Languages without-markers in the table get no dynamic-import protection, which is one of the honest-limits below.--## Workspaces: cross-repo consumers--In a workspace, a file that is dead inside its own repo may still be the surface-another repo depends on. `get_dead_code` checks every finding against the-cross-repo graph before returning it:--- **The file cross-changes with files in other repos.** Confidence is halved and- the finding gains a `cross_repo_note` naming those repos. This is a behavioral- signal drawn from git co-change history, not an import edge, so read it as- "something over there moves when this moves".-- **Another repo depends on this one as a package**, the finding is an- `unused_export`, and there was no co-change signal. Confidence is cut to 30% of- its value with a note saying the export may be consumed and should be verified.--The adjustment runs after tiering, so it lowers the displayed confidence without-moving a finding between tiers. `repo="all"` aggregates every workspace repo,-sorts by confidence then size, and tags each finding with its repo alias.--The CLI has no cross-repo pass yet: `repowise dead-code --repo <alias>` analyzes-that one repo in isolation. See [WORKSPACES.md](../scale/WORKSPACES.md) for how-the cross-repo graph is built.--## Known false-positive sources--The layer is conservative, but it is still a static analysis over a static graph.-These are the cases where a finding is most likely wrong:--- **Reflection and string-keyed dispatch.** A class instantiated from a name in a- config file, a handler looked up in a registry dict, a Java class loaded by- `Class.forName`. The dynamic-pattern name list and the `.register` decorator- suffix catch the common shapes; nothing catches all of them.-- **Dynamic imports in unmodelled languages.** The marker table covers Python,- JS/TS, Java, Kotlin, Ruby, PHP, Go, Swift, Scala, Rust, C#, and C/C++.- Languages without dynamic import markers or framework hints do not detect- runtime loading, so an orphan in those unmodelled languages carries no- dynamic-import cap.-- **Entry points the graph did not mark.** A binary target, a CLI script, or a- serverless handler that neither the allowlist nor the entry-point pass- recognized reads as unreachable every time. If you see a whole directory light- up, that is usually the cause.-- **Barrel re-exports.** `__init__.py` and index barrels are exempt from being- flagged as unreachable *files*, but they are deliberately not exempt in the- unused-export pass: a symbol defined in a barrel that nobody imports should- still be reported. A symbol re-exported through a barrel to external callers- can therefore surface as an unused export.-- **Python `__all__` is read for visibility, never as a rescue.** A literal- module-level `__all__` (list, tuple or set of string constants) raises the- names it lists to `public`, even underscore-prefixed ones; a name it omits- keeps its name-based visibility, because the list is often stale and- demoting on absence would hide a genuinely dead export. Membership is capped- at the visibility label: it sets no export marker, mints no edge, and- suppresses no finding, so declaring a public API there does not by itself- rescue a symbol. Lists built at runtime (comprehensions, `+=`,- concatenation) are treated as absent. The rescues that do apply are the- `__init__.py` exemption, the dunder-name skip, and intra-module reference- tracking.-- **Test-only usage reads as usage.** A test file's import produces a real graph- edge, so a symbol only its tests touch is not flagged. That is deliberate, but- it also means repowise will not tell you a symbol is *exclusively* exercised by- tests. There is no "used only in tests" classification.-- **Recently added code.** A file under 30 days old with no importers is capped- at `0.55` precisely because it is often unfinished, not dead.-- **Interfaces and abstract bases.** Reached only through implementors, which is- heuristic detection. Capped, not suppressed.--The evidence list on every finding tells you which of these applied. Read it-before deleting anything.--## CLI reference--| Flag | Description |-|------|-------------|-| `--min-confidence` | Minimum confidence threshold (default `0.4`) |-| `--safe-only` | Only findings marked safe to delete |-| `--kind` | `unreachable_file`, `unused_export`, `unused_internal`, `zombie_package` |-| `--format` | `table` (default), `json`, `md` |-| `--include-internals` / `--no-include-internals` | Private and underscore symbols (default: off) |-| `--include-zombie-packages` / `--no-include-zombie-packages` | Unused declared packages (default: on) |-| `--no-unreachable` | Skip unreachable-file findings |-| `--no-unused-exports` | Skip unused-export findings |-| `--repo` | Workspace mode: target one repo (defaults to primary) |-| `--workspace` / `--no-workspace` | Force workspace or single-repo mode |--Full command reference: [CLI_REFERENCE.md](../reference/CLI_REFERENCE.md#repowise-dead-code-path).--## The `get_dead_code` MCP tool--Findings come back grouped into high, medium, and low tiers, each with the file-path, kind, confidence, line count, and a cleanup impact estimate.--| Parameter | Default | Notes |-|-----------|---------|-------|-| `kind` | all | One of the four finding kinds |-| `min_confidence` | `0.4` | `0.7` and above is cleanup-ready only |-| `tier` | all | `high` (`>= 0.8`), `medium`, `low` |-| `safe_only` | `false` | Deletion-ready only, excluding runtime-load risk |-| `limit` | `20` | Per tier, clamped to 25 |-| `directory` / `owner` | none | Path-prefix and primary-owner filters |-| `group_by` | none | Roll up by `directory` or `owner` instead of a flat list |-| `include_internals` | `false` | Private and underscore symbols |-| `include_zombie_packages` | `true` | |-| `no_unreachable` / `no_unused_exports` | `false` | |-| `repo` | primary | Workspace mode |--This is a tool for cleanup sweeps, not targeted fixes. Turn it off entirely with-`mcp: {tools: ["-get_dead_code"]}` in `.repowise/config.yaml`. Parameter details:-[MCP_TOOLS.md](../agent/MCP_TOOLS.md#get_dead_code).--## Where else it shows up--- The generated `CLAUDE.md` lists dead-code candidates alongside hotspots and- decisions.-- Contributor profiles carry a dead-code burden per author.-- Module health folds dead-code percentage into its 0-100 composite.-- `repowise update` recomputes findings for changed files only.+From an agent, `get_dead_code()` or `get_dead_code(tier="high", safe_only=True)`.+In the dashboard, open the **Dead code** view of a repository.++## What it finds++| Kind | Meaning | Shown by default |+|------|---------|------------------|+| `unreachable_file` | No file in the repository imports this file, and it is not an entry point. | Yes |+| `unused_export` | A public, top-level symbol that no import, call, inheritance or type reference reaches. | Yes |+| `zombie_package` | A top-level package that no other package in the repository imports. | Yes |+| `unused_internal` | A private or underscore-prefixed symbol that nothing calls. | No, opt in |++Methods, fields, properties and enum members are never reported as unused+exports: they are reached through their container. `unused_internal` is+withheld from the dashboard and default MCP responses because its measured+precision was very low ([Accuracy and limits](#accuracy-and-limits)); ask for+it with `--include-internals` or `include_internals=True`.++## Reading the results++### Confidence++Confidence runs from `0.0` to `1.0` and measures how strong the evidence for+deadness is. It starts from a base value per kind and can only be lowered:++- **Unreachable files** are scored from git activity. A file untouched for a+ year or more starts at `1.00`; one still being committed to starts at `0.40`;+ a file under 30 days old starts at `0.55`, because it is often unfinished.+- **Unused exports** start at `1.00` when the file has importers and none of+ them takes this symbol. When that check is impossible (no importer names+ individual symbols, or the language reaches symbols through a preprocessor)+ the start is `0.60`. Deprecated symbols start at `0.30`.+- **Unused internals** start at `0.65`. **Zombie packages** start at `0.50`.++Several caps then pull a finding down to `0.40`, the review tier: a runtime+loader in the same directory, a path that looks like config or bootstrap code,+an unused export's name written somewhere else in the repository, or a name that+appears in a file repowise could not index. Each cap adds a line to the+finding's evidence saying why.++### Tiers++| Tier | Confidence | Meaning |+|------|-----------|---------|+| High | `>= 0.7` | No references found. Strong cleanup candidate. |+| Medium | `0.4` to `0.7` | Likely unused, but something indirect may reach it. Review first. |+| Low | `< 0.4` | Plausibly used at runtime (dynamic loading, reflection, published API). Investigate first. |++The CLI, the dashboard and `get_dead_code` use the same tier floors, so a+finding lands in the same tier on every surface. The default `min_confidence`+is `0.4` everywhere, which hides the low tier unless you ask for it.++### `safe_to_delete`++A finding is deletion-ready (the CLI's "Ready?" column) only when its+confidence is `0.7` or higher and its path carries no runtime-load risk factor+(config, environment, bootstrap, database, script or served-asset paths).+Unreachable files, zombie packages and unused internals are never marked+deletion-ready, whatever their confidence. In practice `safe_to_delete` means+"an unused export with strong evidence".++### Exempt by construction++Some files and symbols are never reported, because static reachability is the+wrong tool for them. This covers entry points and programs (`__main__.py`,+`main.go`, files with a shebang), build and CI files and anything they name by+path, test files, generated code, vendored trees, framework routes such as+Next.js `page.tsx`, and symbols registered by a framework decorator or+annotation (pytest fixtures, Flask and FastAPI routes, Celery tasks, Spring+stereotypes, and similar). Names that follow dynamic-dispatch conventions+(`*Handler`, `*Plugin`, `register_*`, `on_*` and others) are treated as used.+The full lists live in [the internals doc](../architecture/dead-code.md).++### Empty results++An empty report means no finding cleared the confidence floor, not that the+repository has no dead code. The CLI prints how many findings were hidden below+the floor. An index whose dead-code pass failed also reads as empty; pass+`--min-confidence 0.0` to compute live.++## Tuning and suppressing++- **Confidence floor.** `--min-confidence` (CLI) or `min_confidence` (MCP).+ The index stores findings at `0.4` and above, so the CLI answers from the+ index at that floor or higher and computes a fresh analysis below it.+- **Scope.** `--kind`, `--no-unreachable`, `--no-unused-exports`,+ `--include-internals`, `--no-include-zombie-packages`. Passing `--kind`+ overrides the individual toggles.+- **Triage in the dashboard.** Each finding has a status: `open`,+ `acknowledged`, `resolved` or `false_positive`. Set it per row or resolve in+ bulk. A finding you have acted on stays out of the open list across later+ `repowise update` runs; it is not re-opened.+- **Excluding paths.** `.repowiseIgnore` files and `exclude_patterns` in+ `.repowise/config.yaml` remove paths from indexing (see+ [CONFIG.md](../reference/CONFIG.md)). An excluded file's own imports leave+ the graph too, so a file only it imported can then surface as unreachable.+- **In source.** A Java or Kotlin symbol annotated `@SuppressWarnings("unused")`+ is treated as deliberately unused and is not reported. A symbol marked+ deprecated (by annotation, or by a `_DEPRECATED`, `_LEGACY` or `_COMPAT`+ name suffix) drops to `0.30`, below the default floor.+- **Turning the MCP tool off.** `mcp: {tools: ["-get_dead_code"]}` in+ `.repowise/config.yaml`.++There is no allowlist file for dead-code findings; use the dashboard status.++## False positives and dynamic code++When a file contains a runtime loader, repowise assumes its neighbours in the+same directory may be reached through it and caps their confidence at `0.40`.+Markers are recognised in Python, JavaScript and TypeScript, Java, Kotlin,+Ruby, PHP, Go, Swift, Scala, Rust, C# and C/C++. Examples: `importlib.import_module`,+dynamic `import(` and `require.context(`, `Class.forName(`, `ServiceLoader.load(`,+`const_get(`, `plugin.Open(` and `//go:embed`, `NSClassFromString(`,+`Activator.CreateInstance(`, and `dlopen(` / `LoadLibrary(` in C. Dynamic edges+found by the language extractors count the same way.++The cases where a finding is most likely wrong:++- **Reflection and string-keyed dispatch** that matches no marker, naming+ convention or framework hint: a class named in a YAML file, a handler looked+ up in a registry dict.+- **Entry points the graph did not recognise**, such as a serverless handler or+ a binary target with an unusual layout. A whole directory lighting up is+ usually this.+- **Barrel re-exports.** Barrel files are exempt as unreachable files, but a+ symbol re-exported through one for external callers can surface as an+ unused export.+- **Published libraries.** Code consumed outside the repository has no+ importer inside it. Packaged .NET projects are detected and held at `0.30`;+ other ecosystems are not.+- **Test-only usage counts as usage.** A symbol only its tests import is not+ reported. There is no "used only in tests" classification.++The evidence list on each finding says which signals applied. Read it before+deleting anything.++## Workspaces++In a workspace, a file dead inside its own repo may be the surface another repo+depends on. `get_dead_code` checks each finding against the cross-repo data+before returning it:++- If the file changes together with files in other repos (git co-change),+ confidence is halved and the finding gains a `cross_repo_note` naming those+ repos.+- Otherwise, if the finding is an unused export and another repo depends on+ this one as a package, confidence is cut to 30% of its value with a note to+ verify consumers.++The adjustment runs after tiering, so it lowers the displayed confidence+without moving a finding between tiers. `repo="all"` merges every workspace+repo and tags each finding with its alias. The CLI has no cross-repo pass:+`--repo <alias>` analyzes one repo in isolation (the primary repo by default).+See [WORKSPACES.md](../scale/WORKSPACES.md) for how cross-repo data is built.++## Accuracy and limits++- `unused_internal` was measured at under 1% precision on 1,511 hand-labelled+ findings from one TypeScript/Python monorepo, which is why it is withheld by+ default. Private symbols used within their own file are often invisible to+ the graph.+- Whole unreachable files were measured well below the deletion-ready bar+ (build scripts, manifests and runtime loaders read files by path), so they+ are never marked `safe_to_delete`.+- Unused-export confidence is only as good as the graph's call and import+ resolution for that language. `get_dead_code` reports this per language in+ `summary.call_resolution_basis`.+- Languages without dynamic-import markers get no dynamic-import cap.+- Rust private items are never reported: rustc's `dead_code` lint covers them+ with type and macro information this analysis lacks.+- The name search that caps unused exports counts a name written only in a+ comment as a use. This trades recall for precision.++For benchmark method across layers, see [BENCHMARKS.md](../BENCHMARKS.md).++## Where it shows up++- **CLI** `repowise dead-code`; **MCP** `get_dead_code` (on by default).+- **Dashboard:** the Dead code view (tiers, a safe-to-delete pile, an owner+ leaderboard, status triage) and the dead-code signal on the graph view.+ Module health includes each module's dead-code share; contributor profiles+ carry dead-code files and lines per owner.+- **Editor:** the VS Code extension's findings tree.++## Reference++### `repowise dead-code [PATH]`++| Flag | Default | Description |+|------|---------|-------------|+| `--min-confidence FLOAT` | `0.4` | Minimum confidence. Below `0.4` the analysis runs live. |+| `--safe-only` | off | Only deletion-ready findings. |+| `--kind KIND` | all | `unreachable_file`, `unused_export`, `unused_internal` or `zombie_package`. Overrides the toggles below. |+| `--format` | `table` | `table`, `json` or `md`. |+| `--include-internals` / `--no-include-internals` | off | Report unused private symbols. |+| `--include-zombie-packages` / `--no-include-zombie-packages` | on | Report packages nothing else imports. |+| `--no-unreachable` | off | Skip unreachable-file findings. |+| `--no-unused-exports` | off | Skip unused-export findings. |+| `--repo ALIAS` | primary repo | Workspace mode: analyze this repo. |+| `--no-workspace` | off | Force single-repo mode inside a workspace. |++The CLI reads stored findings when the index is at `HEAD` and the request fits+what was stored; otherwise it analyzes the working tree live, and says which.+JSON rows carry `kind`, `file_path`, `symbol_name`, `confidence`, `reason`,+`safe_to_delete`, `risk_factors`, `lines` and `primary_owner`. Full command+reference: [CLI_REFERENCE.md](../reference/CLI_REFERENCE.md#repowise-dead-code-path).++MCP parameters and response shape: [`get_dead_code` in MCP_TOOLS.md](../agent/MCP_TOOLS.md#get_dead_code).## See also-- [INTELLIGENCE_LAYERS.md](INTELLIGENCE_LAYERS.md): how dead code fits the wider index.-- [CODE_HEALTH.md](CODE_HEALTH.md): the scoring layer that shares the same graph and git data.-- [LANGUAGE_SUPPORT.md](LANGUAGE_SUPPORT.md): which languages parse into the graph the analysis walks.+- [architecture/dead-code.md](../architecture/dead-code.md): exemption lists,+ risk tokens, per-language rescues and confidence scoring internals.+- [CODE_HEALTH.md](CODE_HEALTH.md): the health layer, built on the same graph+ and git data.+- [GRAPH.md](GRAPH.md): the dependency graph the analysis walks.+- [LANGUAGE_SUPPORT.md](LANGUAGE_SUPPORT.md): which languages parse into that graph.diff --git a/docs/layers/DECISIONS.md b/docs/layers/DECISIONS.mdindex ed47a079d..aa80fc1b6 100644--- a/docs/layers/DECISIONS.md+++ b/docs/layers/DECISIONS.md@@ -1,94 +1,250 @@# Architectural Decisions-Repowise mines the *why* out of your repo: the ADRs, commit bodies, PR-descriptions, and `# WHY:` comments where your team already wrote down its-reasoning, plus the choices you make in coding-agent sessions. Every record is-tied to the files it governs, backed by a verbatim quote, tracked for staleness,-and pushed back at your agent at the moment it is about to violate or honor it.+Repowise finds the reasoning your team already wrote down (ADR files, `# WHY:`+comments, commit messages, pull request bodies) and, if you opt in, the choices+you make in coding-agent sessions. Each record is tied to the files it governs,+backed by a verbatim quote, checked for staleness against the code, and handed+to your agent when it is about to edit those files.++It does not decide anything for you. A machine can propose a candidate; only a+person (or a committed ADR) makes it a decision that governs.++Cost: capture runs inside `repowise init` and `repowise update`. ADR files,+inline markers, manual entry and the conventions source work with no LLM key.+Commit, pull request and comment mining need a model, and are skipped with a+stated reason when no provider is configured. Delivery to agents is a local+SQLite lookup with no network.## Quick start```bash-repowise init # extraction runs as part of indexing-repowise decision candidates # what is awaiting review; none of it governs-repowise decision confirm a1b2c3d4 # accept one: this is what makes it govern-repowise decision export # write the accepted ones to a tracked file-repowise decision list # everything, decisions and candidates alike-repowise decision health # stale, conflicting, ungoverned hotspots-repowise decision add # guided interactive capture+repowise init # capture runs as part of indexing+repowise decision candidates # what is awaiting review+repowise decision confirm a1b2c3d4 # accept one; this is what makes it govern+repowise decision list # everything, decisions and candidates+repowise decision health # stale, unscoped, ungoverned hotspots+repowise decision export # write accepted ones to .repowise/decisions.yaml+```++Record one yourself, from a script or an agent:++```bash+repowise decision add --title "JWT over sessions" \+ --decision "Authenticate API calls with signed JWTs" \+ --rationale "Stateless horizontal scaling" \+ --affects src/auth+```++From an agent over MCP:++```python+get_why(query="why JWT over sessions?") # search decisions+get_why(query="src/payments/processor.ts") # what governs this file+get_why() # governance health```-## Three things, not one+In the dashboard, the **Decisions** page of a repository lists records by+review lane and lets you accept or dismiss candidates.++```+repowise decision health++ Decision Health++ Active decisions 14+ Proposed (needs review) 3+ Stale decisions 2+ Deprecated 1++ Stale decisions (2):+ 9f3c1a44 JWT over sessions (staleness: 0.72)+ 2b70de91 EventBus stays in-process (staleness: 0.58)++ Ungoverned hotspots (1):+ payments/processor.ts+```-Repowise distinguishes what it observed, what it inferred, and what you agreed-to, because conflating them is how a machine ends up writing your team's rules.+## What a decision record is++Repowise keeps three things apart:| | What it is | Can a machine create it? | Does it govern? ||---|---|---|---|| **Episode** | An evidenced event: a transcript span, a commit, a structural change | Yes | No |-| **Candidate** | A durable choice inferred from that evidence | Yes | No |-| **Decision** | A constraint you accepted | No, except a committed ADR | Yes |+| **Candidate** | A durable choice inferred from that evidence. Stored with status `proposed` | Yes | No |+| **Decision** | A constraint someone accepted. Stored with status `active` | No, except a committed ADR | Yes |++Only decisions reach your agent as rules, count toward path alignment, or land+in the generated `CLAUDE.md`. Recurrence, confidence and a model's verdict make+a candidate worth reading. None of them accepts it.-Only the third reaches your agent, scores path alignment, or lands in the-generated `CLAUDE.md`. Recurrence, confidence, a model's verdict and a-successful implementation are all evidence that a candidate is worth reading;-none of them is an acceptance.+A record carries a title, the decision, context, rationale, rejected+alternatives, consequences, tags, the files or modules it governs (its scope),+its source, a confidence, a staleness score, and one or more evidence rows+(file and line, or commit, plus the quote).-**Acceptance is a recorded event**, not a status somebody set. It carries a-reason, a scope, an evidence reference, and who accepted it, and repowise-refuses to store one missing any of those rather than filling in a blank:+**Acceptance** is a recorded event, not a status flag. It carries a reason, a+scope, an evidence reference and who signed it. Repowise refuses to store an+acceptance that is missing any of those:```$ repowise decision confirm 4b6ddc58Cannot accept 4b6ddc58: no scope: name the files or modules it governsSupply the missing parts with --reason, --scope or --evidence.-$ repowise decision confirm 4b6ddc58 --scope packages/core/src/repowise/core/ingestion+$ repowise decision confirm 4b6ddc58 --scope src/ingestionDecision 4b6ddc58 accepted (governing)```-The log is append-only. Re-accepting after review, sending a decision back to-review, superseding it and withdrawing it each add a row, so the history of who-granted and withdrew authority survives every later action.+The acceptance log is append-only. Accepting, re-accepting, withdrawing,+superseding and dismissing each add a row, so the history of who granted and+withdrew authority survives.-### Upgrading from before the split+## Where decisions come from-Before this, a record reached `active` by appearing in two coding-agent-sessions, with nobody involved. `repowise decision migrate` classifies those:+| Source | Key | Reads | Needs an LLM key | On by default |+|--------|-----|-------|------------------|---------------|+| ADR files | `adr` | `adr/`, `adrs/`, `docs/adr/`, `docs/adrs/`, `docs/decisions/`, `decisions/`, `architecture/`, `doc/adr/` (up to 60 files); Nygard and MADR headings, YAML frontmatter | No (parsed directly; a model stage is optional) | Yes |+| Inline markers | `inline_marker` | `WHY:`, `DECISION:`, `TRADEOFF:`, `ADR:`, `RATIONALE:`, `REJECTED:` in any comment syntax. Uppercase only, like `TODO:` | No (a model stage is optional) | Yes |+| Commits | `git_archaeology` | Commit messages that carry a decision verb (migrate, switch to, replace, adopt, deprecate, drop, ...) | Yes | Yes |+| Pull request bodies | `pr` | Squash-merge and PR commit bodies that read like a PR description | Yes | Yes |+| Code comments | `comment` | Rationale prose in comments on the most central files | Yes | Yes |+| Agent sessions | `session` | Your local coding-agent transcripts | No for the gates; the structuring step uses a model when one is configured | No |+| Session discovery | `session_discovery` | One broad model pass over new transcript prose per update | Yes | No |+| Conventions | `conventions` | Import edges: a wrapper most files reach a library through | No | No |+| Manual entry | `cli` | `repowise decision add`, the dashboard form | No | Always |++A committed ADR whose status says `accepted` or `approved` and that names what+it governs accepts its own decision, with the file recorded as the accepter. It+is the only non-human acceptance, and it only applies to a file git tracks. A+draft, an uncommitted file, or one with no status lands as a candidate.++The same decision found by two sources becomes one record with two evidence+rows. Headline fields come from the more authoritative source; the other is+kept as corroboration and raises confidence.++Internals of each source, the evidence check, the confidence formula and the+session miner: [architecture/decisions.md](../architecture/decisions.md).++## Lifecycle++Stored status is one of `proposed` (a candidate), `active` (accepted),+`deprecated`, `superseded` or `dismissed`. An accepted decision also has a+**currency**, which says whether it still describes the code:++| Currency | Meaning | Still governs? |+|----------|---------|----------------|+| `active` | Accepted, and still describes the code | Yes |+| `needs_review` | Accepted, but at least half of the files it names have changed since it was recorded | Yes |+| `uncheckable` | Accepted, but names no file or module, so it cannot be checked against the code | Agents editing a file are not given it |+| `stale` | Accepted, but every file it names is gone at HEAD (renames followed) | No |+| `superseded` | Replaced by a later decision the person named | No |+| `dismissed` | Authority withdrawn; kept for history | No |++`active`, `needs_review`, `uncheckable` and `stale` are derived from the code+on every read. `superseded` and `dismissed` are set by a person.++**Staleness score.** Each record carries `staleness_score` between 0 and 1: the+fraction of its governed files that have been committed to since the decision+was recorded. A file the repository does not track counts as changed. At 0.5 or+above a record is stale for `decision list --stale-only`, the health summary,+and the `needs_review` currency. A record that names no file scores 0.0 and is+counted separately as unscoped. `repowise decision show` and `get_why` on a+path also ask git directly for one record and print what changed.++**Sticky review.** A dismissed candidate is kept as a tombstone, so re-indexing+never proposes it again. An accepted decision is never walked back to+`proposed` by a later extraction.++**Supersession.** `repowise decision deprecate ID --superseded-by ID2` records+the successor. `repowise update` also marks a decision that a new commit+reversed. Automatic detection of `supersedes` and `conflicts_with` edges by+text similarity is switched off, because similarity between records from one+repository did not reliably mean a shared topic.++## Reviewing candidates-```-$ repowise decision migrate+| Command | What it records |+|---------|-----------------|+| `decision confirm ID...` | Accept one or more candidates. `--reason`, `--scope`, `--evidence` fill gaps. `--preview` writes nothing. |+| `decision dismiss ID...` | Tombstone. On an accepted decision this also withdraws its authority. |+| `decision deprecate ID --superseded-by ID2` | Retire with an explicit successor. |+| `decision merge ID INTO_ID` | Fold a candidate into an accepted decision. The old id keeps resolving. |+| `decision dedupe` | Fold duplicate candidates in one sweep. Dry run until `--apply`. |+| `decision split ID` | Flag a candidate as bundling two choices. Never splits it for you. |-Decision migration (dry run)+In a batch, each id goes through the same acceptance check and is applied on+its own, so one refusal does not stop the rest.- Total legacy records 504- Kept as decisions 0- Reclassified as candidates 502- Dismissed tombstones 2- Already migrated 0+The Decisions page splits records into five lanes that do not overlap:- Why records became candidates:- 381 never accepted: it was awaiting review- 121 active by recurrence, not by a person: a session row carries no acceptance event+| Lane | What is in it |+|------|---------------|+| Active | Accepted and still describes the code. These are the rules. |+| Candidates | Never accepted. Governs nothing. Ordered so the ones that can be accepted as-is come first. |+| Needs review | Accepted, but the files it names have moved. Still binds. |+| Uncheckable | Accepted, but names no file or module. |+| History | Accepted, then withdrawn, superseded or dismissed. |++Accept from the UI goes through the same check as `decision confirm`: a+candidate missing a reason, scope or evidence has Accept disabled with the gap+named.++## How to add one++```bash+repowise decision add # guided prompts+repowise decision add --title T --decision D [--affects PATH ...] # no prompts```-**Expect your standing decisions to shrink**, and on many repositories to-empty. That is the correction, not a regression: those records were never-accepted, and everything is still there under `repowise decision candidates`,-ready to be accepted deliberately. Nothing is deleted, every id keeps-resolving, retirements you performed are preserved, and re-running changes-nothing. `repowise init` and `repowise update` run the same classification, so-a store cannot sit half-migrated with the status column and the acceptance log-disagreeing.+| Flag | Meaning |+|------|---------|+| `--title`, `--decision` | Both together switch to non-interactive mode and print the new id |+| `--context`, `--rationale` | Why it was needed, why this choice |+| `--alternative` | A rejected alternative. Repeatable |+| `--consequence` | A tradeoff accepted. Repeatable |+| `--affects` | A file or module it governs. Repeatable |+| `--tag` | A tag. Repeatable |+| `--evidence-commit` | A commit the decision was made in. Repeatable |+| `--kind` | `architectural` (default) or `agreement` |+| `--format json` | Machine-readable output with the full id |++What gets stored depends on how you add it:++- **Guided prompts** record an acceptance, signed by you. If an architectural+ decision names no files it is kept as a candidate, because it cannot be+ checked against the code.+- **Flags** (`--title` and `--decision`) store a candidate (`proposed`). A+ person answering the prompts has reviewed the decision; a script or agent+ inferring one has not. Promote it with `repowise decision confirm ID`.+- `--kind agreement` records a rule about how the work is done ("never push+ without review"). It governs the whole repository, needs no files, and+ reaches an agent at session start.++## Governance++**Who may accept.** Each acceptance records who signed it and what kind of+signer: `person`, `agent`, or `import` (the committed manifest or an ADR). The+person defaults to `git config user.name`; `--as` records a different identity.++An agent may withdraw authority (`dismiss`, `deprecate`) by signing as itself+with `--agent SLUG`. Granting authority is refused for an agent unless the+repository allows it:++```bash+repowise decision config agent-acceptance --on # off by default+repowise decision confirm ID --agent claude_code --session "$SESSION_ID"+```-## The tracked manifest+The dashboard badges any authority record a person did not sign, and+`decision show --format json` returns the signature.-Accepted decisions are the one part of this layer that belongs to the-repository rather than to your machine, so they live in a file you commit:+**The tracked manifest.** Accepted decisions belong to the repository, so they+live in a file you commit:```bash-repowise decision export # store -> .repowise/decisions.yaml-repowise decision import # .repowise/decisions.yaml -> store+repowise decision export # store -> .repowise/decisions.yaml+repowise decision import # .repowise/decisions.yaml -> store``````yaml@@ -104,648 +260,132 @@ decisions:currency: activesource: sessionaccepted_at: '2026-09-01T13:04:22+00:00'- accepted_by: Raghav+ accepted_by: Jane Doe```-Ordered by id and rendered from a fixed field order, so two machines holding-the same decisions produce the same bytes and a one-line change is a one-line-diff. Written atomically and fsynced; an unchanged render is not rewritten at-all. Comments are not preserved — it is generated, and says so in its own-header.--`export` un-ignores the file in your `.gitignore` if a `.repowise/` rule was-hiding it, leaving everything else under `.repowise/` ignored, including nested-ones elsewhere in the tree.--On import the file wins: it is what a colleague's commit changed and the store-is the copy. A decision the file has and the store does not is created and-accepted with the file itself as the accepter — the only non-human accepter-there is, and only because it is version controlled and reviewable. An entry-that differs is re-accepted from the file. An entry missing a reason or a scope-is skipped rather than guessed at, and an entry the file no longer holds is-left alone: a missing line is as likely a bad merge as a deliberate removal.--Two guards on the write. A store with no acceptances refuses to overwrite a-non-empty committed file, because deleting the index is a documented recovery-step and it must not delete your decisions with it. And a file written by a-newer repowise is refused rather than silently downgraded.+The file is ordered by id with a fixed field order, so a one-line change is a+one-line diff. `export` un-ignores this one file if a `.repowise/` rule in+`.gitignore` was hiding it. On `import` the file wins: new entries are created+and accepted with the file as accepter, changed entries are re-accepted, entries+missing a reason or scope are skipped, and entries the file no longer holds are+left alone (a missing line may be a bad merge). An empty store never overwrites+a non-empty committed file, and a file written by a newer repowise is refused.+Candidates and episodes stay out of the file.-Episodes and candidates stay out of it. They are evidence and inference, they-turn over constantly, and committing them would put a machine's opinions under-review as though they were the team's.--## Review actions--| Command | What it records |-|---------|-----------------|-| `decision confirm ID...` | Accept, optionally editing the reason and scope on the way. Takes many ids. |-| `decision confirm ID` on a decision | Reaffirm it after review. |-| `decision confirm\|dismiss\|deprecate ID --agent SLUG` | A machine signing as itself. Only `confirm` needs the repository's permission. |-| `decision merge ID INTO_ID` | Fold a candidate into an existing decision. The old id resolves to the target. |-| `decision dedupe` | Fold candidates that duplicate another candidate, in one sweep. Dry run until `--apply`. |-| `decision split ID` | Flag a candidate as bundling two choices. Never splits it for you. |-| `decision dismiss ID...` | Tombstone it. On an accepted decision this also withdraws its authority. Takes many ids. |-| `decision deprecate ID --superseded-by ID2` | Retire it with an explicit lineage edge. |--`merge` folds one pair and needs the target to be an accepted decision, because-that merge changes what governs. `dedupe` is the other case: a store that has-been indexed many times accumulates the same decision written several ways, and-none of those records governs anything yet. It folds each duplicate into the-candidate it duplicates — evidence, governed files and all — and only ever into-a record it is directly measured against, so a chain of near-neighbours is not-collapsed into one decision. Run it without `--apply` first; it prints exactly-what it would fold.--`confirm` and `dismiss` take a list because a repository accumulates candidates-faster than anyone reviews them one command at a time. Every id in a batch goes-through the same acceptance contract as a single one, and each is applied on its-own, so one refusal commits the rest and reports itself instead of ending the-run. `--preview` runs the batch through that contract and rolls it back, so the-report is what the write would have said rather than a second opinion about it.--Similarity never supersedes anything: an edge exists because somebody named the-successor. Merging and superseding retire ids that may already be written down-somewhere, so both leave an alias and the retired id keeps resolving.--The same actions are on the Decisions page, split into five review lanes that-partition the repository, so the counts add up and no record is in two of them:--| Lane | What is in it |-|------|---------------|-| Active | Accepted, and still describes the code it names. These are the rules. |-| Candidates | Never accepted. Governs nothing, reaches no agent. Carries the evidence quote it was drawn from, so it can be accepted or dismissed from the row. Ordered so the ones the acceptance contract would take come first. |-| Needs review | Accepted, but the files it names have moved. Still binds. |-| Uncheckable | Accepted, but names no file or module, so nothing can check it against the code and no agent editing a file will be given it. |-| History | Accepted and then withdrawn, superseded or dismissed. |--The lane is a join onto the acceptance, not a filter on the status column, so-`GET /api/repos/{id}/decisions?lane=candidates` and-`GET /api/repos/{id}/decisions/lane-counts` are different questions from-`/decisions/counts`, which groups the status projection. Every listed record-also carries a `currency` field; a record with no `currency` is a candidate.-The lane is applied before the page is cut, so a page of a lane is a page of-that lane.--A candidate carries a review row, written where it is captured and refreshed-for every open candidate on each index. It records which lane raised it, the-version that extracted it, whether the lane saw the candidate bundling more-than one choice, and whether the record names any scope at all. It also carries-the review priority the Candidates lane and `repowise decision candidates`-order on, which is the acceptance contract's verdict as of the last index and-nothing else: a reviewer meets the candidates they can finish first, and-confidence and recency break the ties. Re-extraction refreshes those signals-but never the review state, so a dismissal or a split request survives the next-index.--A store indexed before the row existed has candidates without one. Until the-next index writes them a row they rank as unjudged, alongside the ones the-contract refused, and `repowise decision status` counts them so the gap is-visible rather than silent.--Accepting from the UI goes through the same acceptance contract as-`repowise decision confirm`, so a candidate missing a reason, a scope or an-evidence reference has its Accept disabled with the gap named beside it rather-than failing after the click. Dismissing stays available: a candidate that-cannot be accepted is the one most worth tombstoning.--```-repowise decision health-- Decision Health-- Active decisions 14- Proposed (needs review) 3- Stale decisions 2- Deprecated 1-- Stale decisions (2):- 9f3c1a44 JWT over sessions (staleness: 0.72)- 2b70de91 EventBus stays in-process (staleness: 0.58)-- Ungoverned hotspots (1):- payments/processor.ts-```--`repowise decision health` scores governance. `repowise decision status` reports-capture: the effective policy and preset, every source with the reason it made-no call, what each has captured and when, the review lanes, how much of the-unreviewed backlog is ready to accept against how much is blocked, the age of-that backlog, the staging queues, and the model spend booked to decision-extraction.--Nothing records a capture run. The extraction report and the broad lane's-discovery report are rendered to the progress line and discarded, and-`.repowise/state.json` carries nothing per source, so `status` derives every-figure from a durable trace instead: the records and their review rows, the-staging queues, and the `decision_extraction` cost rows. Two things follow, and-the command says both rather than papering over them. Spend is all-time with-the last call named, because no stored boundary says which call belonged to-which run. And a queue that a store predates is reported absent, not as zero.--From an agent:--```python-get_why(query="why JWT over sessions?") # NL search-get_why(query="src/payments/processor.ts") # what governs this file-get_why(query="why is caching split?", targets=["src/cache"])-get_why() # health dashboard-```--## Where decisions come from--Six capture sources run at index time, each a pass over the repo, its git-history, or the graph it just built.--| Source | Key | Reads | Notes |-|--------|-----|-------|-------|-| ADR files | `adr` | `adr/`, `adrs/`, `docs/adr/`, `docs/adrs/`, `docs/decisions/`, `decisions/`, `architecture/`, `doc/adr/` | Nygard and MADR headings plus YAML frontmatter, parsed without an LLM. Up to 60 files. A document that explicitly says `accepted`/`approved`, is committed, and names what it governs accepts its own decision, with the file recorded as the accepter; it is the only non-human acceptance there is. A draft, an uncommitted file, or one with no Status section lands as a candidate. |-| Inline markers | `inline_marker` | `# WHY:` / `# DECISION:` / `# TRADEOFF:` / `# ADR:` / `# RATIONALE:` / `# REJECTED:` | Any comment syntax (`#`, `//`, `--`, `/*`, `*`). The keyword is case-sensitive and must be capitalised, like `TODO:` and `FIXME:` — otherwise ordinary prose ("# Rejected: nothing to extract.") becomes an architectural decision. Up to 5 continuation lines, plus 20 lines of surrounding context. Fenced code blocks in Markdown are skipped. |-| Git archaeology | `git_archaeology` | Commit messages | Gated on 19 decision verbs (migrate, switch to, replace, adopt, deprecate, drop, rewrite, split, revert, and the rest). |-| PR bodies | `pr` | Squash-merge and PR commit bodies | A body only qualifies when it looks like a PR description (`## Why`, `## Motivation`, `## Context`, `Closes #`, `Before:` / `After:`). Up to 25 bodies. |-| Code comments | `comment` | Block comments and docstrings on high-centrality files | Bounded to 30 nodes, and to prose carrying a rationale cue ("because", "instead of", "rather than", "trade-off", "we chose", "deliberately"). Centrality-bounded on purpose: comment archaeology across a whole repo is noise. |-| Conventions | `conventions` | Import edges and wrapper bodies | Off by default (`repowise decision source set conventions --on`), no model. A wrapper of an I/O library (HTTP client, database driver, subprocess, filesystem, lock) that most files reach the library through while few import it directly; the record states the counts and names the direct importers. Go and Java are counted by package. Up to 10 per index. |--Three more sources sit outside the index-time set: `session` (deterministic-gates over your coding-agent transcripts, below, off unless you enable it),-`session_discovery` (one broad model pass over the same transcript prose,-below, also off unless you enable it), and `cli` (a decision you typed-yourself, the most authoritative source there is).--### Controlling capture--Every source is switchable, and so is the model. One resolved policy backs the-CLI, the API, and the index pipeline, so what `config show` prints is what the-next `init` or `update` will actually run.+**Capture policy.** Every source can be switched, and so can the model:```bash-repowise decision config show # resolved policy, per source, with reasons-repowise decision config preset local_only-repowise decision source list+repowise decision config show # resolved policy, with reasons+repowise decision config preset local_only # default | off | local_only | balanced | fullrepowise decision source set comment --off-repowise decision source set adr --no-llm # keep the parse, skip the model-repowise decision source set session_discovery --on # the broad lane, opt-in-repowise decision config discovery --max-sessions 6 # and its per-update budget-repowise decision llm --off # no decision extraction reaches a model+repowise decision source set adr --no-llm # keep the parse, skip the model+repowise decision llm --off # no decision extraction calls a model```-```-repowise decision config show-- Decision capture on · LLM extraction off · preset local_only- No LLM provider configured; model stages are skipped.--Source Status LLM Why-inline_marker deterministic_only no Decision LLM extraction is off. The deterministic stage still runs.-git_archaeology disabled no This source is switched off.-adr deterministic_only no Decision LLM extraction is off. The deterministic stage still runs.-pr disabled no This source is switched off.-comment disabled no This source is switched off.-session deterministic_only no Decision LLM extraction is off. The deterministic stage still runs.-session_discovery disabled no This source is switched off.-conventions disabled no This source is switched off.-cli always_on - Manual entry is always available.-- Discovery budget: up to 12 session(s), 30,000 input tokens per update.-```+| Preset | What it runs |+|--------|--------------|+| `default` | What a config with no `decisions:` block gets: ADR, markers, commits, PR bodies, comments, with the model on. Sessions, discovery and conventions off |+| `off` | No capture. Manual entry still works |+| `local_only` | ADR, markers and sessions, deterministic only. No model calls |+| `balanced` | `default` minus comments, plus sessions and session discovery |+| `full` | Every source, model on |-The presets are `default`, `off`, `local_only`, `balanced`, and `full`; editing-any individual switch afterwards reads as `custom`. `default` is what a config-with no `decisions:` block resolves to, and it is deliberately not `full`: a-source added after those switches existed stays off until you ask for it, so an-upgrade never starts a model call you did not enable. A stored preset that-lists its sources is treated the same way: that list is what the preset-covered when it was written. Re-apply the preset to pick up a source added to-it since. Mutations take `--dry-run` and-`--format json`, write atomically, and preserve every unrelated key in-`config.yaml`.--Three things the switches deliberately do **not** do. Turning a source off stops-new capture from it; it deletes nothing already stored. A decision you have-already accepted keeps governing after the source that found it is switched off,-because authority comes from your acceptance and not from the source staying on.-And `llm: false` is a complete mode rather than a broken one: transcripts are-still read, markers and ADRs still parsed, episodes still recorded, manual-entry unaffected. Status says *skipped*, never *failed*.--`skipped_no_provider` is the same idea for a missing API key: an LLM-only source-reports that it had nothing to run with, and a hybrid source falls back to its-deterministic stage.--Full block reference: [CONFIG.md](../reference/CONFIG.md).--## Evidence: verified, fuzzy, unverified--Every produced field (`decision`, `rationale`, `source_quote`) is checked against-the verbatim source span the extractor recorded. This is the anti-hallucination-gate, and it matters most for the generated sources: the page generator *writes*-candidate decisions rather than only reading them, and the gate is what stops a-fluent invention from being stored as institutional memory.--| Verdict | Fires when | Effect on confidence |-|---------|-----------|----------------------|-| `exact` | The normalized quote is a substring of the source span. | No penalty. |-| `fuzzy` | Token overlap with the source span is at least 0.6 (a paraphrase or a reflow). | Multiplied by 0.85. |-| `unverified` | Neither, or there was no source span to check against. | Multiplied by 0.6. |--An ungrounded field is cleared, not kept. A candidate whose every produced field-is ungrounded is rejected outright. A candidate with no source text at all is-kept but stamped `unverified`: repowise never fabricates a rejection it cannot-justify.--Confidence then rises with how authoritative the source is and how many-independent sources corroborate it:+Switching a source off stops new capture; it deletes nothing, and accepted+decisions keep governing. With the model off or no provider configured, model+stages report `skipped`, never `failed`. Full block reference:+[CONFIG.md](../reference/CONFIG.md#the-decisions-block).-```-confidence = 0.4 + 0.5 * (best_source_rank / 9)- + min(0.12, 0.04 * (corroborating_sources - 1))- x verification penalty-```+`repowise decision status` reports what capture did: the policy, each source+and why it did or did not run, review lanes, backlog age, staging queues and+model spend on decision extraction.-The rank ladder is `cli` 9, `session` 8, `adr` and `pr` 7, `commit` and-`git_archaeology` 6, `conventions` 6, `inline_marker` 4, `comment` 3, and the heuristic tiers-below that. `session` sits above `adr` because a transcript carries what a person-actually said while deciding, and a document is someone's later write-up of it;-with the order reversed the write-up overwrote the words. Retired sources keep-their rungs (`changelog` 5, `readme_mining` 3) so rows written before their-removal still rank instead of dropping to the unknown-source floor. The result is clamped to-`[0, 0.99]`: nothing is ever certain.+## How agents get them-**Sources corroborate, they do not overwrite.** The same decision found in an ADR-and in a commit body becomes one record with two evidence rows. Headline fields-come from the highest-ranked row; the lower-ranked one is kept as corroboration-and pushes confidence up. A decision resting only on a plain code comment is-decayed a further 0.85 so it never reads as confident as an ADR.+Through hooks, without asking (details in [HOOKS.md](../agent/HOOKS.md)):-## The decision graph+- **Session start.** Repowise scores decisions against the session's likely+ working set (dirty and staged files, branch changes, the previous session's+ edits, branch-name tokens), expands one hop through imports and co-change+ partners, and injects the most relevant under a hard ~400-token cap.+ Candidates come in a separate, labelled section with its own small budget.+ If nothing is relevant enough, nothing is injected.+- **Edit time.** When the agent edits a file an accepted decision governs, it+ gets a one-line notice with the rationale, once per session per decision.-Decisions are not a flat list. Typed edges connect them:+Through MCP, `get_why` picks its mode from the call shape:-| Edge | Meaning |+| Call | Returns ||------|---------|-| `supersedes` | The newer decision replaces the older one, and the older record flips to `superseded`. |-| `refines` | Narrows or extends a decision without reversing it. |-| `relates_to` | Same topic, no ordering claim. |-| `conflicts_with` | Two *active* decisions contradict each other. A governance smell, surfaced in `decision health` and in the code-health layer. |--**Automatic detection of these two edges is currently off.** It scoped a-conflict by embedding similarity: two decisions had to share a topic (at least-two shared content tokens after stopword removal) and then either straddle an-opposing verb pair (`adopt` / `use` / `introduce` against `drop` / `remove` /-`deprecate` / `revert`) or carry a reversal signal ("replace", "migrate",-"switch to", "no longer", "in favor of"), with an LLM tiebreaker on the pairs-the heuristic could not call. In practice similarity does not scope: among-descriptions drawn from one repository, a cosine of 0.81 is the baseline rather-than evidence of a shared topic, so the check fired between unrelated records-and retired records that were correct. It returns when a conflict is scoped-structurally — by two decisions touching the same code — with similarity used-only to rank the candidates that test finds.--It was the only writer of these edges, so **no edges exist while it is off** and-lineage is empty. Two things still work: the diff-driven evolution pass on-`repowise update` marks a decision that a new commit reversed, and `repowise-decision deprecate --superseded-by ID` records the successor on the record-itself (the `superseded_by` column, not an edge). Records the detector retired-before it was turned off are restored to `proposed` on the next `init` or-`update`, and its edges are deleted in the same pass.--`supersedes` and `refines` chain into a **lineage**, so `get_why` can answer-"why is auth structured this way?" with `sessions -> JWT -> OAuth2` rather than-three disconnected records. `get_why(query="<path>")` returns the lineage-whenever the chain has more than one node.--Edges accrete rather than clobber, and confirmations are sticky in both-directions: `repowise decision dismiss` keeps a `dismissed` tombstone so-reindexing never re-proposes the same thing, and a confirmed `active` decision is-never walked back to `proposed` by a later extraction.--## Staleness--A decision is only useful while it still describes the code, and that is a-question the repository can answer rather than one worth guessing at. Every-record carries a `staleness_score` between 0 and 1: **the fraction of its-affected files that have been committed to since the decision was recorded.**--- None of them has changed: `0.0`.-- All of them have: `1.0`.-- A file the repository does not track counts as changed, because a record- naming something absent cannot be shown to still hold.--`>= 0.5` is stale everywhere: `repowise decision list --stale-only`, the health-summary, and the staleness column in the CLI table.--The number is a supporting fact; what an accepted decision *is* gets a word:--| Currency | Meaning |-|----------|---------|-| `active` | Accepted, and still describes the code. |-| `needs_review` | Accepted, but the files it names have moved. Still governs — a decision whose code moved is one to re-read, not one to ignore. |-| `uncheckable` | Accepted, but names nothing the repository can be asked about. |-| `superseded` | Replaced by a later decision, through an explicit lineage edge. |-| `dismissed` | Authority withdrawn, kept for history. |--The first three are derived from the code on every read, so a decision moves-between them without anybody touching it. The last two are states a person set,-and the repository does not get to argue with them.--There is deliberately no constant in that definition. An earlier version grew-the score with 90-day commit volume and record age, and added a boost when a-later commit *message* contained words like "migrate away"; both were fitted to-one repository's history and to English commit prose, and the result was a-number that moved for reasons unrelated to whether the code had moved.--**A record that names no file scores 0.0 and is not fresh** — the question-cannot be asked of it at all. `repowise decision health` counts those-separately, as *unscoped*, rather than banking them as current.--For one record on demand, `repowise decision show` and `get_why` on a path go-further and ask git directly, printing what it says: *"nothing in the 3 files it-governs has changed since 2026-05-02"*. That costs a subprocess, so it is served-only where one is affordable — never from an editor hook or during an update.--## Driving decisions from an agent--Every subcommand takes `--format json`. The lifecycle commands (`confirm`,-`dismiss`, `deprecate`, `show`) exit non-zero on an unknown id and emit-`{"error": "decision_not_found", "decision_id": "..."}`, so a scripted confirm-cannot be mistaken for a successful one; `dismiss` skips its prompt under-`--format json` or with `--yes`. `decision config show --format json` returns-the whole source registry with a `status` and a `reason` per source, which is-the machine-readable answer to "will this source run, and if not why not".--Reading decisions back, `active` means somebody accepted it and it binds;-anything else is a candidate and is a hint. The stored status is a projection-of the acceptance event, not the authority itself — the acceptance is — but the-two are kept in step, so a reader that only has the status column stays correct-rather than merely stale. See [MCP_TOOLS.md](../agent/MCP_TOOLS.md#get_why).--### An agent may withdraw authority; granting it is a switch--An acceptance records what signed it — `person`, `agent` or `import` — beside-who. The two are different questions, and only the first one was ever stored:-`accepter` resolves to `git config user.name`, so an agent confirming a-decision signed the maintainer's name and no surface could tell them apart.--The asymmetry follows. A machine can withdraw authority without asking: the-worst a wrong withdrawal does is stop a rule binding, and re-accepting undoes-it. Granting authority mints a constraint nobody agreed to, under a signature-that reads as an acceptance, so it is refused unless the repository says-otherwise:--```bash-repowise decision config agent-acceptance --on # off by default-repowise decision confirm ID --agent claude_code --session $SESSION_ID-```--`--agent` takes the agent's slug and is accepted by `confirm`, `dismiss` and-`deprecate` alike — withdrawing needs no switch, so it is the action an agent-is likeliest to take, and the one it would otherwise take under your name. On-`confirm` it is refused without the switch. What it-records is visibly an agent's: `decision show --format json` returns-`signature` with the kind and the session, the API carries `accepter_kind`-on every decision row, and the dashboard badges any authority record a person-did not sign. A row written before these columns existed reads as-`unrecorded`, not as a person's — those are the two things the field exists to-keep apart.--The same log records withdrawals, so what is shown is who *signed*, never that-they accepted. A record the evolution stage retired reads "Signed by an agent"-beside a superseded status, not "accepted by" anything.--## Session-mined decisions--`repowise update` (docs mode) reads your local coding-agent transcripts and mines-the durable decisions out of them: user corrections, explicit choices with a-stated reason, and failed approaches replaced by working ones. Claude Code-transcripts come from `~/.claude/projects/`, read incrementally from a cursor so-each line is processed once.--Claude Code is the only harness read unless you name another:--```yaml-decisions:- harnesses: [claude_code, codex]-```--Adding one is a real widening rather than a preference. Codex files its-sessions by date rather than by project, so its store holds every session on-the machine and the repository a line belongs to is decided by the working-directory recorded in it, not by where the file sits. Sessions bulk-imported-into that store from another harness are skipped, because the harness that-recorded them already reads them and counting them here would count one-conversation twice.--Three stages, in order:--1. **Deterministic gates.** A user correction needs a pushback lead ("no,",- "don't", "not like that", "actually,", "instead"). An explicit choice needs a- decision verb ("use", "went with", "switched to", "chose", "always", "never")- *paired* with a causal marker. A dead end needs three consecutive failures of- the same command anchor.-2. **One batched LLM structuring call per update**, capped at 60 candidates.- Every produced field must quote the transcript verbatim or it is dropped;- an ungrounded `source_quote` rejects the candidate.-3. **Promotion to a candidate.** A decision seen in two or more distinct- sessions, or one direct user correction, is written with `source: session`- as a candidate. Recurrence is evidence that a candidate is worth reading,- not an acceptance event: `repowise decision confirm` is what makes it- govern. Repeated observations accrete evidence rows and raise confidence- without ever creating authority.--**This source ships off.** What it captured on this repository read as working-agreements from a transcript — "pause before pushing", "avoid duplication" —-rather than decisions the codebase had taken: of 139 records, none carried a-context and 115 carried neither a context nor a rationale. Turn it on with-`repowise decision source set session --on`; the `local_only`, `balanced` and-`full` presets carry it already.--Everything stays on your machine. Transcripts are read locally, staging lives in-`.repowise/sessions/sessions.db`, and only the distilled decision text about the-codebase is stored. `--off` turns the pipeline back off, or keep the transcript-reading and drop only the structuring call with `--no-llm`.--This switch is the transcript boundary, so two things ride on it. Injected--decision feedback is inert without it — its only evidence is this lane's mined-corrections — and update-time hook-efficacy replay stops, which leaves-`repowise hook stats` with what the live hook recorded first-hand and without-the columns only a replay can fill. `repowise hook backfill` fills those on-demand.--### Broad session discovery--Those gates are precision-first, and they are the recall bottleneck: a decision-settled in a paragraph that never opens with "no," or pairs a decision verb-with a causal cue is invisible to them. `session_discovery` is the recall lane-beside them, off unless you enable it:+| `get_why()` | Governance health: counts, stale decisions, candidates awaiting review, ungoverned hotspots |+| `get_why(query="src/auth/jwt.py")` | Three lanes that never mix: `decisions` (accepted, governing; with lineage when there is one), `candidates` (nobody accepted them), `history` (superseded or withdrawn). Plus an origin story from git and an alignment read on the first lane |+| `get_why(query="why is caching split?", targets=[...])` | Ranked decision search, optionally anchored to paths |+| `get_why(query=..., repo="all")` | The same search across every workspace repo |++Only the `decisions` lane is a rule. When no decision covers a path, `get_why`+falls back to git history for that file, then to a rationale comment in the+source. Full parameters: [MCP_TOOLS.md](../agent/MCP_TOOLS.md#get_why).++Accepted decisions also appear in the generated `CLAUDE.md`, `get_overview`,+`get_context`, and as the `governance_risk` flag in `get_risk`. A candidate+never raises `governance_risk`, and reaches `get_answer` labelled `candidate`.++## Accuracy and limits++- Every stored field is checked against the verbatim source span. A field that+ is not grounded is cleared; a candidate with no grounded field is rejected.+- Confidence is capped at 0.99 and rises with source authority and independent+ corroboration. It is a ranking aid, not a probability.+- The `session` source ships off: on this repository it produced 139 records,+ none with a context and 115 with neither context nor rationale.+- Automatic supersession and conflict edges are off, so lineage comes only from+ explicit `--superseded-by` and commit reversals.+- Staleness measures whether governed files changed, not whether the decision+ is still true.+- Commit, PR and comment mining need a model; without one, only ADRs, markers,+ conventions, sessions (gates only) and manual entry produce records.++Method for repowise's measured numbers: [BENCHMARKS.md](../BENCHMARKS.md).++## Where it shows up++| Surface | What you get |+|---------|--------------|+| CLI | `repowise decision ...` |+| MCP | `get_why`; decisions inside `get_overview`, `get_context`, `get_risk`, `get_answer` |+| Hooks | Session-start and edit-time injection |+| Dashboard | Decisions page with review lanes |+| Generated files | Accepted decisions in `CLAUDE.md` |+| Code health | `ungoverned_hotspot`, `stale_governance`, `contradictory_decision` findings |-```bash-repowise decision source set session_discovery --on-```+## Reference-Once per update it makes **one** model call over the user and assistant prose-that update newly read, bounded by `discovery.max_sessions` (default 12) and-`discovery.max_input_tokens` (default 30,000). It reuses the transcript read-the `session` source already performs, so enabling it does not read your-transcripts twice, and it costs one call whether the queue holds one session or-forty.--Each span of prose it sends carries a stable id, and every candidate must cite-the ids it rests on. Before anything is stored:--- the cited spans must all be spans that were actually sent;-- the evidence quote must be verbatim, or near enough that only reflow and- articles separate it from the span's own text. The deterministic lane accepts- a looser paraphrase because its excerpts are the few sentences its gates- already matched; this lane is shown thousands of words, where a partial- overlap with some part of them is not evidence of anything;-- the claim itself must overlap that text on its content words, so a fluent- invention riding a real quote is still rejected;-- a rationale that fails the same check is dropped rather than invented;-- a path the model names is kept only if the cited turns' tools touched it.- Scope is resolved from those files, never from the model, and naming none is- an honest repository-wide claim;-- a claim that bundles two independent choices is flagged for splitting at- review, never split by machine.--Everything that survives is a **candidate**. Like the deterministic lane it is-written `proposed`, needs two distinct sessions to be proposed at all, and only-`repowise decision confirm` makes it govern. The model's own-`implemented_and_validated` verdict is evidence for review, not authority.--Zero calls, and a stated reason, on every path that should not call: capture-off, the source off, `llm: false`, no provider configured, and no new prose-since the last update. Prose that does not fit one update's budget stays queued-and is sent oldest-first next time, and a queued span is never aged out unread;-a provider failure requeues the same prose and retries it on the next two-updates before retiring it. Switching the source off leaves what is already-queued in place, to be drained if you switch it back on. New prose is only-captured while the source is on, so there is no backfill of what was read in-between. It also depends on the `session` source, which is what reads the-transcripts; with `session` off, discovery says so rather than reporting an-empty repository.--## Getting decisions back to your agent--Capture is half the loop. The other half is delivery, and it happens at two-moments without the agent asking (see [HOOKS.md](../agent/HOOKS.md)):--**At session start.** Repowise scores active decisions against the session's-likely working set (dirty and staged files, files changed on the branch versus-`main`, the previous session's edited files, branch-name tokens), expands that-one hop through import edges and co-change partners, and injects the top few-under a hard ~400-token cap. Relevance is multiplied by confidence and by-freshness, so a stale or low-confidence record has to be *much* more relevant to-make the cut. Nothing clears the floor means nothing is injected: decisions are-never shown just for being high-confidence.--**At edit time.** When the agent edits a file governed by a decision (through the-`file` and `module` node links), it gets a one-line notice with the rationale, at-most once per session per decision. This is the moment that matters, right before-the code is written.--Every injected decision id is recorded locally. On the next `repowise update` the-session miner checks whether the guidance was followed or contradicted by your-corrections in that session and stores that verdict on the injection, where-`repowise hook stats` reports the split. It deliberately does not touch the-record's staleness: whether you overrode a decision is a judgement about the-record, staleness is a measurement of the code, and one of those is per-machine-while the other travels with the repository.--Decisions also land in the generated `CLAUDE.md` (accepted records, freshest-first) and in `get_overview()`, `get_context()`, and the `governance_risk` flag-in `get_risk()` PR review.--Every one of those reads is the acceptance join, not the status column. A-candidate does not appear in `get_overview()`'s key decisions, cannot raise a-`governance_risk` directive against a pull request, and reaches `get_answer`'s-prompt labelled `candidate` rather than `active`. That labelling matters: the-row is rendered verbatim into a model prompt, and a candidate named `active`-reads to the model as something the repository has settled on.--## CLI reference--Every subcommand takes an optional trailing `PATH`; in workspace mode it targets-the primary repo. Decision ids accept an 8-character prefix.+Every subcommand takes an optional trailing `PATH`, `--format json`, and+accepts an 8-character id prefix. Lifecycle commands exit non-zero on an+unknown id.| Command | What it does ||---------|--------------|-Recording a decision from the UI follows the same rule as `repowise decision-add`: name the files or modules it governs and the write records an acceptance,-name none and it is stored as a candidate, and the form says which before you-press the button. Re-recording the title of a decision that is already accepted-without naming its scope is refused rather than applied, because it would clear-the scope that decision governs.--| `repowise decision add` | Guided interactive capture: kind, title, context, decision, rationale, rejected alternatives, tradeoffs, affected files, tags. Answering the prompts is an acceptance, recorded as one. An architectural record that names no files is kept as a candidate instead, because a decision that names nothing cannot be checked against the code. `--kind agreement` records the other noun — a rule about how the work is conducted — which is accepted without naming files, because it governs the repository rather than part of it, and reaches an agent at session start rather than when a file is edited. |-| `repowise decision list` | Table of id, title, status, source, confidence, staleness, created date. |-| `repowise decision show ID` | Full record including alternatives, consequences, affected files, and the evidence file and line. |-| `repowise decision add --evidence-commit SHA` | Record which commit a decision was made in. Repeatable. The capture prompt passes it, and it is what stops the prompt asking again for a commit already recorded. |-| `repowise decision confirm ID...` | Accept candidates. Refuses, naming the gap, when one has no reason, scope or evidence; `--reason`, `--scope` and `--evidence` supply them. A refused id does not stop the others. `--preview` writes nothing. `--agent SLUG` signs as an agent rather than as you, and needs `decision config agent-acceptance --on`. |-| `repowise decision dismiss ID...` | Tombstone them. Never re-proposed on reindex. `--preview` writes nothing. |-| `repowise decision deprecate ID` | Retire it, optionally `--superseded-by <ID>`, which writes the lineage edge. |-| `repowise decision candidates` | What is awaiting review, and why each was raised. |-| `repowise decision merge ID INTO_ID` | Fold a candidate into an existing decision. |-| `repowise decision split ID` | Flag a candidate as bundling two choices. |-| `repowise decision export` / `import` | Round-trip accepted decisions through `.repowise/decisions.yaml`. |-| `repowise decision migrate` | Classify pre-split records. Dry run unless `--apply`. |-| `repowise decision health` | Counts, stale decisions, ungoverned hotspots, proposals awaiting review. |-| `repowise decision status` | What capture did: policy and preset, per-source state and why, review lanes and backlog age, staging queues, model spend. |-| `repowise decision config show` | The resolved capture policy: every source, its status, and why. |-| `repowise decision config preset NAME` | Apply `off`, `local_only`, `balanced`, or `full`. |-| `repowise decision config agent-acceptance --on/--off` | Whether an agent may grant a decision authority. Off by default; no preset changes it. |-| `repowise decision config capture-prompt --on/--off` | After a commit whose message states a choice, ask the agent once a session to record it. Off by default; no preset turns it on. `--on` also installs the `Bash\|PowerShell` PostToolUse entry it fires from, because the shared matcher deliberately excludes the shell tools; `--off` removes it, for every repository on the machine. Silent while `decisions.enabled` is off. |-| `repowise decision source list` | The source registry with capabilities and current state. |-| `repowise decision source set SRC --on/--off` | Switch one source. `--llm/--no-llm` switches only its model stage. |-| `repowise decision llm --on/--off` | Master switch for decision-extraction model calls. |--`list` filters:--| Flag | Values |-|------|--------|-| `--status` | `proposed`, `active`, `deprecated`, `superseded`, `dismissed`, `all` (default) |-| `--source` | any source in the rank ladder except the retired ones, plus `all` (default) |-| `--proposed` | Shortcut for `--status proposed` |-| `--stale-only` | Only records with staleness at or above 0.5 |--Full flag reference: [CLI_REFERENCE.md](../reference/CLI_REFERENCE.md#repowise-decision).--## The `get_why` MCP tool--`get_why(query=None, targets=None, repo=None)` dispatches into four modes:--| Call shape | Mode | Returns |-|------------|------|---------|-| No `query` | health | `counts`, `stale_decisions`, `proposed_awaiting_review`, `ungoverned_hotspots`, `conflicts` |-| `query` is a path | path | Three lanes: `decisions`, the accepted decisions governing that file (with `lineage` when the chain is longer than one); `candidates`, records naming it that nobody has accepted; and `history`, records accepted and since superseded or withdrawn. Plus an `origin_story` from git and an `alignment` read scored on the first lane only |-| `query` is a question | search | Ranked decision records plus `related_documentation`, optionally anchored with `targets` |-| `repo="all"` | workspace search | The same records across every workspace repo, each tagged with its alias |--The lanes never mix. `decisions` is what governs, and it is the only lane an-agent should read as a rule. `candidates` carries a `candidates_note` saying in-as many words that nobody has accepted them and how to accept one, and-`history` carries a `history_note` saying its records were accepted and have-since been replaced. When only candidates name a path, `alignment.score` is-`none` and its explanation says there is no accepted decision, rather than-reporting the file as governed. `alignment` also splits its counts four ways,-and they sum to `governing_count`: `active_count`, `deprecated_count`,-`uncheckable_count`, `candidate_count`.--Both non-governing lanes are capped at three rows and are the first things shed-under budget pressure: they are context, and a response that cannot afford the-rules must not spend on the queue. A note is dropped with its lane, so a-sentence never counts rows that are no longer in the payload.--It is designed never to come back empty-handed. If no accepted decision covers-a path, it falls back to git archaeology on that file. If git history is silent-too, it mines a rationale comment live from the source and returns it as-`code_rationale`. Semantic decision search falls back to full-text search when-the vector store is unavailable.--See [MCP_TOOLS.md](../agent/MCP_TOOLS.md#get_why) for parameters and worked-examples.+| `decision add` | Record a decision (see [How to add one](#how-to-add-one)) |+| `decision list` | Table of records. `--status`, `--source`, `--proposed`, `--stale-only` |+| `decision show ID` | Full record with evidence and signature |+| `decision candidates` | What awaits review, and why each was raised |+| `decision confirm` / `dismiss` / `deprecate` | Accept, tombstone, retire |+| `decision merge` / `dedupe` / `split` | Fold or flag candidates |+| `decision export` / `import` | Round-trip `.repowise/decisions.yaml` |+| `decision migrate` | Classify records from older stores. Dry run unless `--apply` |+| `decision health` | Counts, stale records, ungoverned hotspots |+| `decision status` | What capture did, per source |+| `decision config show` / `preset` / `discovery` / `agent-acceptance` / `capture-prompt` | Capture policy |+| `decision source list` / `set SRC --on/--off [--llm/--no-llm]` | Per-source switches |+| `decision llm --on/--off` | Master switch for model calls |++Full flags: [CLI_REFERENCE.md](../reference/CLI_REFERENCE.md#repowise-decision).+MCP: [`get_why`](../agent/MCP_TOOLS.md#get_why).## See also-- [INTELLIGENCE_LAYERS.md](INTELLIGENCE_LAYERS.md): where decisions sit among the five layers.-- [CODE_HEALTH.md](CODE_HEALTH.md): the `ungoverned_hotspot`, `stale_governance`, and `contradictory_decision` findings.-- [HOOKS.md](../agent/HOOKS.md): the SessionStart and edit-time injection hooks in detail.+- [architecture/decisions.md](../architecture/decisions.md): source internals, evidence check, confidence, session mining.+- [INTELLIGENCE_LAYERS.md](INTELLIGENCE_LAYERS.md): where decisions sit among the layers.+- [CODE_HEALTH.md](CODE_HEALTH.md): the governance findings.+- [HOOKS.md](../agent/HOOKS.md): the injection hooks.- [CONFIG.md](../reference/CONFIG.md): the `decisions:` block.diff --git a/docs/layers/GRAPH.md b/docs/layers/GRAPH.mdindex fc62bfb9a..c1ee6a5aa 100644--- a/docs/layers/GRAPH.md+++ b/docs/layers/GRAPH.md@@ -1,23 +1,24 @@# Graph IntelligenceMost tools that draw a code graph will tell you `A calls B`. Very few will tell-you *how sure they are*, and none of the interesting questions can be answered+you how sure they are, and the interesting questions cannot be answeredwithout that.repowise builds a two-tier graph of your codebase, files and symbols, with no-model calls and no network. What makes it worth trusting is not its size. It is-that **every edge carries its own evidence**.+model calls and no network. It needs no LLM key and is built during+`repowise init` and kept current by `repowise update`. What makes it worth+trusting is not its size. Every edge carries its own evidence: the strategy+that produced it and the confidence that strategy earns.<p><img src="https://img.shields.io/badge/17-edge_types-3178C6?style=flat-square&labelColor=0A0A0A" alt="17 edge types" /><img src="https://img.shields.io/badge/39-resolution_origins-059669?style=flat-square&labelColor=0A0A0A" alt="39 resolution origins" />- <img src="https://img.shields.io/badge/26-languages-F59520?style=flat-square&labelColor=0A0A0A" alt="26 languages" />- <img src="https://img.shields.io/badge/22-framework_detectors-7F52FF?style=flat-square&labelColor=0A0A0A" alt="22 framework detectors" />+ <img src="https://img.shields.io/badge/26-full--AST_languages-F59520?style=flat-square&labelColor=0A0A0A" alt="26 full-AST languages" /><img src="https://img.shields.io/badge/0-LLM_calls-1E293B?style=flat-square&labelColor=0A0A0A" alt="zero LLM calls" />- <img src="https://img.shields.io/badge/compiler_graded-7_of_7_cells_undominated-DC2626?style=flat-square&labelColor=0A0A0A" alt="no tool is both more precise and more complete, in 7 of 7 compiler-graded cells" /></p>-**Contents:** [The problem with a plain arrow](#the-problem-with-a-plain-arrow) ·+**Contents:** [Quick start](#quick-start) ·+[The problem with a plain arrow](#the-problem-with-a-plain-arrow) ·[Two stages, and they fail differently](#two-stages-and-they-fail-differently) ·[How good is it, and how we know](#how-good-is-it-and-how-we-know) ·[What is in the graph](#what-is-in-the-graph) ·@@ -31,6 +32,29 @@ that **every edge carries its own evidence**.---+## Quick start++```bash+repowise init # builds the graph; no API key needed+repowise context src/app.py --include callers # who depends on this file, with confidence+repowise serve # dashboard: Graph, Architecture, Coupling, Blast radius+```++From an agent, over MCP:++```text+get_context(targets=["src/app.py::App.run"], include=["callers"])+get_symbol(id="src/app.py::App.run", depth=2)+```++`get_context` lists callers and callees with their confidence; `get_symbol` at+depth 2 also returns the bodies the symbol calls. Both expansions drop call+edges below 0.7, which keeps+the repo-wide name guesses out of them (see+[how an agent should read confidence](#how-an-agent-should-read-confidence)).++---+## The problem with a plain arrowConsider one line of Python:@@ -42,34 +66,33 @@ user.save()To draw an edge, a tool has to answer "what is `user`?". There are three ways todo it, and they are not equally good:-1. **Give up.** Emit nothing. The edge is missing, so a dead-code pass now- thinks `save` is unused and offers to delete it.-2. **Guess.** Find every method named `save` in the repo and pick one. If your+1. Give up. Emit nothing. The edge is missing, so a dead-code pass now thinks+ `save` is unused and offers to delete it.+2. Guess. Find every method named `save` in the repo and pick one. If yourcodebase has `User.save`, `Draft.save` and `Session.save`, you have a two inthree chance of drawing an arrow to the wrong file.-3. **Work out what `user` is**, then resolve `save` on that type.+3. Work out what `user` is, then resolve `save` on that type.Most graphs do (2) and present the result identically to an edge they werecertain about. That is the actual problem. A wrong arrow is worse than a missingone, because a missing arrow looks like missing information and a wrong arrowlooks like an answer.-repowise does (3) where it can, falls back to (2) where it must, and **labels-which one happened, every time**.+repowise does (3) where it can, falls back to (2) where it must, and labels+which one happened on every edge.---## Two stages, and they fail differently-An edge is two claims made by two different pieces of machinery, and the reason-to keep them apart is that they break in opposite directions.+An edge is two claims made by two different pieces of machinery. They are kept+apart because they break in opposite directions.-**Stage one: capture.** Before anything can be resolved, the parser has to-notice that a call was written at all. That is a tree-sitter query per language,-listing the source shapes that count as a call site:+**Capture** is noticing that a call was written at all. It is a tree-sitter+query per language, listing the source shapes that count as a call site:```scheme-; queries/go.scm -- Method call: obj.Method(args)+; Go method call: obj.Method(args)(call_expressionfunction: (selector_expressionoperand: (identifier) @call.receiver@@ -79,138 +102,64 @@ listing the source shapes that count as a call site:) @call.site```-There is one of those files per language, and a shape that is not in it is a-call the graph will never contain. Go alone lists a plain call, a method call, a-package-qualified call, a chained call and a function passed as an argument, and-that last one is captured deliberately as a *reference* rather than a call,-because passing a handler is not invoking it.+A shape that no query matches is a call the graph will never contain. Not low+confidence, not unresolved: absent. Nothing downstream can recover it, because+nothing downstream knows the call was there. Go alone lists a plain call, a+method call, a package-qualified call, a chained call, and a function passed as+an argument. That last one is captured as a reference, not a call, because+passing a handler is not invoking it.-A shape no query matches is invisible. Not low confidence, not unresolved:-**absent**. Nothing downstream can recover it, because nothing downstream knows-the call was there.+**Resolution** is working out what a captured name points at.+`repo.save(draft)` hands you the name `save` and a receiver spelled `repo`, and+the job is to turn that into one declaration in one file. This is where the 39+resolution origins live.-**Stage two: resolution.** Given a captured site, work out what the name points-at. `repo.save(draft)` hands you the name `save` and a receiver spelled `repo`,-and the job is to turn that into one declaration in one file. This is where the-39 origins below live, and it is the `user.save()` problem from the section-above.--| | fails when | costs you | how you find out |+| Stage | Fails when | Costs you | How you find out ||---|---|---|---|-| **capture** | nobody wrote a query for that call shape | **recall.** the edge does not exist | nothing internal can tell you, so it takes an outside answer key |-| **resolution** | the receiver cannot be typed, or the name is ambiguous | **precision** if it guesses, **recall** if it declines | the origin on the edge names the strategy that answered, so a wrong class is found and fixed once rather than per call site |+| Capture | no query covers that call shape | recall: the edge does not exist | nothing internal can tell you; it takes an outside answer key |+| Resolution | the receiver cannot be typed, or the name is ambiguous | precision if it guesses, recall if it declines | the origin on the edge names the strategy, so a wrong class of edge is found and fixed once |-That asymmetry sets the whole design. A missed capture is silent, so it is-measured against a compiler rather than against ourselves. A bad resolution is-loud, so every edge is stamped with the strategy that produced it and nothing is-allowed to launder a guess into a fact.+That asymmetry sets the design. A missed capture is silent, so it is measured+against a compiler. A bad resolution is loud, so every edge is stamped with the+strategy that produced it, and nothing is allowed to launder a guess into a+fact.-**And at the bottom of the ladder repowise declines rather than guesses.** That-is a choice with a price, paid in the recall column below, and it is the right-way round: a missing arrow looks like missing information, a wrong arrow looks-like an answer.+At the bottom of the ladder repowise declines to guess. That costs recall, and+it is the right way round.---## How good is it, and how we know-Two readings of the same question, and we graded only one of them.--### The one we did not grade--On Go the answer key is the **Go team's own RTA call graph** from-`golang.org/x/tools`, computed over the fully type-checked program. On TypeScript-it is the **`tsc` checker's own resolution** of every call site. We wrote-neither, we can tune neither, and anyone with the toolchain can regenerate both.--Seven cells, five repositories, **five tools**, 37,853 oracle edges. Of the call-edges we emit, the share the compiler confirms runs **0.943 to 0.992** per cell.--That number on its own is not the claim, because precision on its own has a cheap-way to win: draw one edge you are certain of and you score 1.000. Two of the five-tools do a version of exactly that. One scores 0.997 on cobra, the highest figure-in the whole experiment, from a graph holding **17% of the calls in the-repository**. Another takes both gitleaks cells from a graph that-finds 89% of the calls where we find 95%. Meanwhile the tool with the best recall emits, on the largest-repository measured, more than a third of its edges as calls the compiler says do-not exist.--**Recall alone is gameable in the other direction and precision alone in this-one, so the claim is the pair:**--> In all seven cells, **no tool that recovers as much of the call graph as we do-> gets more of it right.**--It names no threshold, so it cannot be tuned, and a new competitor can only break-it. Two were added after it was written and it held in all seven cells.--The weaker readings, so nobody has to infer them: most precise outright in one-cell, tied in one more, beaten in five by tools drawing much smaller graphs. And-against the two tools the experiment started with, most precise in seven of-seven, which is the narrower claim it should always be labelled as.--**Two languages, and only two.** C#, Java, Kotlin and C++ each need a toolchain-installed and a working build per repository, and nobody has done that here.-Python, Ruby and PHP can never have an oracle at all, because what a call-resolves to can change at runtime. That is a fact about those languages rather-than a gap in the harness, and it is why the hand-graded reading below is-permanent rather than a stopgap.--[The cells, the method and the graded pre-registration](../BENCHMARKS.md#8-the-same-question-against-an-answer-key-we-do-not-control)--### The one we did--Nine languages, 30 call edges per language per tool (widened to 40 for Java-across two repositories), every row opened in its own file with its imports and-enclosing scope, then the target declaration opened too. **240 of 280 correct-for us (85.7%), 164 of 280 for CodeGraph 1.5.0 (58.6%)**, intervals disjoint.-Four of the nine cells separate and five are ties, reported as ties.--Read our own number the other way round: **roughly fourteen percent of our call-edges are wrong**, concentrated in java, rust and cpp. That is the figure to plan-against, and it is a floor rather than a best case, because every resolver change-since the earliest rows were graded only removes wrong edges.--**The two readings agree.** On Go the hand grade says 96.7% for us and the-compiler says 97.6%, over roughly 1,600 edges rather than 30 rows. A person-reading source and a type checker landing within about a point of each other is-the strongest available evidence that the hand-graded half is accurate rather-than self-serving, and it is the result here we care about most.--[The nine cells, and all 560 graded rows with the reason each was given](../BENCHMARKS.md#7-edge-precision)--### The column we lose--Recall is the other half of the same question, and we do not lead it.--Across the five Go cells our recall runs 0.32 to 0.96 and **we lead none of-them**; codebase-memory-mcp leads four and CodeGraph the fifth. On cross-file-coverage over 35 repositories the same tool separates from us on 15 and we-separate on none. We do lead recall in both TypeScript cells, and we lead it over-the two tools that beat us on precision in every cell, which is the same trade-seen from the other side.--The oracle explains the trade rather than excusing it. **That tool recovers more-of the true call graph and emits far more that is not in it**: on the largest Go-repository measured, more than a third of what it emits is a call the compiler-says does not exist. Coverage rewards drawing edges and never asks whether they-are real, which is why no page in that benchmark prints a coverage number without-a precision number beside it.--Where our own miss actually goes, decomposed on one cell rather than waved at.-On syft without tests we miss **3,846 of the oracle's 7,898 edges**, and-**44% of that miss is dynamic dispatch alone, with a further 39% dispatch with a-closure at one end**. The two buckets overlap, so neither is the whole gap. Interface dispatch is the ceiling and-nobody in the comparison has cleared it: of 3,303 dispatch edges we match 12,-CodeGraph 35, codebase-memory-mcp 81, at 6.5 distinct possible targets per call-site. Matching that recall means emitting six edges where one is right, which is-the behaviour the precision table charges the other tool for.--The obvious cheap fix was priced and refused: giving Go `func` literals a symbol-would recover **50** static edges on that cell, not the 1,309 the raw closure-count suggests, because the rest need the dispatch ceiling cleared first.--[Both tables, the recall decomposition and what it would cost to close](https://github.com/repowise-dev/repowise-bench/tree/master/graph/experiments/g4-oracle-anchored)+Measured on main, freshly indexed from source, on 2026-10-03.++- **Against a compiler, Go and TypeScript.** Seven cells over five+ repositories, 37,853 compiler edges (the Go team's RTA call graph and `tsc`'s+ own resolution). Precision runs 0.976 to 0.995, and recall is higher than our+ August figure in all seven cells. In all seven, no tool in the five-tool+ comparison that finds as much of the call graph gets more of it right. Two+ tools post higher precision in some cells by drawing much smaller graphs.+- **Against compiler indexes for more languages.** Java on jsoup reads 0.781+ precision against javac (0.725 to 0.781 across three repositories). C on git, a+ repository held out from all tuning, reads 0.979 precision at 0.83 recall.+ C#, C++, Rust and Python (against jedi, a static analyser) are in the same+ table.+- **Hand-graded across nine languages.** 280 call edges per tool, 560 in all,+ each read with its imports and enclosing scope open: 240 of 280 correct for+ repowise (85.7%) against 164 of 280 for CodeGraph 1.5.0 (58.6%), intervals+ disjoint. Read the other way round, about one call edge in seven (roughly+ fourteen percent) is wrong, concentrated in Rust, C++ and Java.++Recall is the column we do not lead everywhere. Against the compiler we lead+recall in three of the seven cells and trail in four. Where our misses go is+mostly dynamic dispatch, which no tool in the comparison has cleared without+emitting several wrong edges for each right one.++The per-language table, the five-tool table and the method are in+[BENCHMARKS.md: accuracy by language](../BENCHMARKS.md#accuracy-by-language) and+[BENCHMARKS.md: call graph against a compiler](../BENCHMARKS.md#8-the-same-question-against-an-answer-key-we-do-not-control).+The hand-graded rows are in+[BENCHMARKS.md: hand-graded](../BENCHMARKS.md#7-edge-precision).---@@ -218,64 +167,80 @@ count suggests, because the rest need the dispatch ceiling cleared first.**Two tiers of node.** Files and packages on one tier; functions, classes,methods and interfaces on the other. Third-party packages appear as lightweight-external nodes so a dependency is visible without being documented.+external nodes, so a dependency is visible without being documented.**Two families of edge.** Structure the parser can see (imports, calls,inheritance) and structure only history can see (files that keep changing-together without importing each other). They are kept apart on purpose: a-co-change edge is real signal, but treating it as a dependency would put-"these two files were edited in the same commit" into your import graph.+together without importing each other). They are kept apart on purpose. A+co-change edge is real signal, but treating it as a dependency would put "these+two files were edited in the same commit" into your import graph.-Consumers never hand-roll a filter over this. Three named views are derived-from the vocabulary and shared:+Every consumer reads the graph through one of three shared views, so no two+features disagree about what counts as a dependency:| View | Answers ||------|---------|-| `FILE_DEPENDENCY_EDGE_TYPES` | "what does this file depend on?" Used for communities, cycles, coupling |-| `SYMBOL_USE_EDGE_TYPES` | "what reaches this symbol?" Containment excluded, since a class holding a method is not a use of it |-| `REACHABILITY_USE_EDGE_TYPES` | "does anything use this at all?" The dead-code view |--Before those views existed, four different places each wrote their own edge-filter, and two of them silently counted co-change edges as imports.+| File dependency | "what does this file depend on?" Used for communities, cycles and coupling |+| Symbol use | "what reaches this symbol?" Containment is excluded, since a class holding a method is not a use of it |+| Reachability | "does anything use this at all?" The dead-code view |---## Every edge says how it got there-Each `calls` edge is stamped with a **resolution origin**: the named strategy-that produced it. There are 39, drawn from a closed vocabulary, and each one-carries exactly one confidence.+Each `calls` and `references` edge is stamped with a **resolution origin**: the+named strategy that produced it. There are 39, from a closed vocabulary, and+each one carries exactly one confidence. The table lists the 18 base origins;+the other 21 are receiver-typing variants, covered in+[Typing the receiver](#typing-the-receiver).-| Confidence | Origin | What was actually established |+| Confidence | Origin | What was established ||:---:|---|---|-| **0.95** | `same_file` | The callee is defined in the calling file. A certainty |-| **0.95** | `self_scope` | `self` / `this`, a method on the caller's own class |-| **0.93** | `receiver_same_file` | The receiver names a type declared right here |-| **0.90** | `import_scoped` | The name was imported from the file that defines it |-| **0.90** | `same_package` | A sibling file that needs no import (Go, JVM) |-| **0.88** | `receiver_import` | The receiver's type was found in an imported file |-| **0.85** | `import_merged` | It is in *one* of the imported files. Which one is unattributed |-| **0.75** | `receiver_global` | The `(class, method)` pair exists somewhere in the repo |-| **0.50** | `global_unique` | The name is unique repo-wide. **A guess, and stored as one** |+| 0.95 | `same_file` | The callee is defined in the calling file |+| 0.95 | `self_scope` | `self` / `this`: a method on the caller's own class |+| 0.95 | `enclosing_class` | A bare call, bound to the caller's own class |+| 0.93 | `receiver_same_file` | The receiver names a type declared in this file |+| 0.93 | `scoped_name` | C/C++ `Qualifier::name()`: the class is written at the call site |+| 0.90 | `import_scoped` | The name was imported from the file that defines it |+| 0.90 | `same_package` | A sibling file that needs no import (Go, JVM) |+| 0.90 | `receiver_same_package` | The receiver is a class in the same package (JVM) |+| 0.90 | `self_inherited` | `self` / `this` or Python `super()`, found on exactly one ancestor |+| 0.90 | `enclosing_inherited` | A bare call, found on exactly one ancestor |+| 0.88 | `package_alias` | Go `pkg.Func`, resolved across the whole package |+| 0.88 | `module_alias` | The receiver is an imported module |+| 0.88 | `crate_root` | A Rust crate-scoped reference |+| 0.88 | `receiver_import` | The receiver's type was found in an imported file |+| 0.85 | `import_merged` | It is in one of the imported files; which one is unattributed |+| 0.85 | `same_target` | C/C++: a sibling translation unit of the same build target |+| 0.75 | `receiver_global` | The `(class, method)` pair exists somewhere in the repo |+| 0.50 | `global_unique` | The name is unique repo-wide. A guess, and stored as one |Because every origin has exactly one confidence, the origin distribution and the-confidence histogram are two views of the same data, which is what makes the-stamping checkable rather than decorative.+confidence histogram are two views of the same data. That is what makes the+stamping checkable.++### How an agent should read confidence-**Why this matters in practice.** An agent tracing an execution flow can decline-anything below a threshold and know what it declined. A reviewer looking at a-blast radius can tell "this definitely breaks" from "this shares a method name-with something that breaks". And when the graph is wrong, the origin tells you-*which strategy* was wrong, so it can be fixed once rather than patched per-call site.+| Confidence | Read it as |+|:---:|---|+| 0.93 to 0.95 | Established in the calling file or class. Treat as fact |+| 0.85 to 0.90 | Established through an import, a package or one ancestor. Safe to act on |+| 0.75 | A class and method name match somewhere in the repo. Verify before acting on it alone |+| 0.50 | A unique name, nothing more. A lead, not an answer |++An agent tracing a flow can decline anything below a threshold and know what it+declined. A reviewer reading a blast radius can tell "this definitely breaks"+from "this shares a method name with something that breaks". When the graph is+wrong, the origin names which strategy was wrong, so it is fixed once for every+call site that strategy touched.---## Typing the receiver-Twelve of the 39 origins exist to answer the `user.save()` question properly.-Rather than matching a bare method name, repowise reads the receiver's-**declaration** and resolves the method on that type.+21 of the 39 origins exist to answer the `user.save()` question properly. They+read the receiver's declaration and resolve the method on that type, instead of+matching a bare method name.```javavoid handle(UserRepo repo) { // parameter declares the type@@ -285,36 +250,38 @@ void handle(UserRepo repo) { // parameter declares the type}```-Five receiver shapes are covered, each a separate origin family so each can be-measured on its own:+Each receiver shape is its own origin family, so each can be measured on its+own:-| Shape | Example |-|---|---|-| Locals and parameters | `var repo = new UserRepo()`, `fun f(r: UserRepo)` |-| Fields of the enclosing class | `this.cache.evict(...)` |-| Freshly constructed receivers | `new Foo().bar()` |-| The method's own receiver | Go's `func (s *Server) handle()` |-| Framework-retyped symbols | `@shared_task def add` makes `add` a `Task`, so `add.s()` is `Task::s` |--Shipped for **Java, C#, Python, Go, Kotlin and Swift**. The typed origins share-their untyped twin's confidence deliberately: the inferred type had to declare-the method before any edge was emitted, so the evidence is no weaker. What-differs is how the receiver was named, and naming that difference is the entire-point of an origin.--Languages are added here one at a time, each gated on a sampled precision audit-before it ships. Several are deliberately absent because they failed that gate.-An unsupported language falls back to the weaker origins and says so, which is-the correct degradation.+| Family | Example | Origins |+|---|---|:---:|+| Locals and parameters, including Go's method receiver | `var repo = new UserRepo()`, `func (s *Server) handle()` | 4 |+| Fields of the enclosing class | `this.cache.evict(...)` | 4 |+| Framework-retyped symbols | `@shared_task def add` makes `add` a `Task`, so `add.s()` is `Task::s` | 4 |+| C# extension methods, via the type their `this` parameter names | `items.Paged(10)` | 3 |+| Dotted receivers typed hop by hop through declared fields | `this.a.b.m()` | 2 |+| The declared return type of an inner call | `repo.find(id).save()` | 4 |++The families map onto the same scopes as the base origins (same file, same+package, import, global), and a typed origin shares its untyped twin's+confidence. The inferred type had to declare the method before any edge was+emitted, so the evidence is no weaker. What differs is how the receiver was+named, and naming that difference is the point of an origin.++Declaration scanning covers Java, C#, Python, Go, Kotlin, Swift, TypeScript,+Rust and C++. A language outside that list falls back to the base origins and+says so on every edge.---## Seventeen edge types, because `calls` was doing too many jobsAn edge type is a claim. If one type carries several different claims, every-consumer downstream has to guess which one it is looking at.+consumer downstream has to guess which one it is looking at. The graph has 17+edge types. These six carry the distinctions that matter most when you read an+answer:-| Edge | Claims | Explicitly does **not** claim |+| Edge | Claims | Does not claim ||---|---|---|| `calls` | The parser saw a call expression and resolved its callee | || `references` | Something holds a handle to this function: a dispatch-table entry, a callback field, a registration macro | That it is ever invoked. Enough to make deleting it unsafe, not enough to call it a call |@@ -323,25 +290,28 @@ consumer downstream has to guess which one it is looking at.| `type_use` | A type is referenced in a constructor, method, delegate or record parameter | An import. Weighted below one || `co_changes` | These files keep changing in the same commit | Any code dependency at all |+The other eleven are `imports`, `defines`, `has_method`, `extends`,+`implements`, `method_implements`, `framework`, `reads`, `dynamic_uses`,+`dynamic_imports` and `dynamic_url_route`.+The `references` distinction is not academic. A handler sitting in a dispatchtable is never called anywhere a parser can see. Counting that as "no use"-reported entire registration layers as safe to delete.+would report entire registration layers as safe to delete.-`framework_binds` is separated from `calls` for the same reason in reverse. A-fixture nobody calls and a collaborator nobody constructs are both genuinely-used, by the container. But an inferred wiring hop is not source, and letting it-render as a call would put it into an execution flow as though someone had-written it.+`framework_binds` is separated from `calls` for the opposite reason. A fixture+nobody calls and a collaborator nobody constructs are both used, by the+container. But an inferred wiring hop is not source, and letting it render as a+call would put it into an execution flow as though someone had written it.---## Flows that say why they stoppedAn execution flow walks the call graph from an entry point. Every walk ends, and-a trace that just stops reads identically whether execution really ends there or+a trace that just stops reads the same whether execution really ends there orthe walker ran out of things it could follow.-So a flow never simply ends. It terminates with one of six reasons:+So every flow ends with one of six reasons:| Termination | Meaning ||---|---|@@ -352,11 +322,11 @@ So a flow never simply ends. It terminates with one of six reasons:| `excluded_target` | Every successor was a test, demo or fixture node || `callees_truncated` | Rows were cut before the walk saw them |-Two details carry most of the value. `no_callees` is **deliberately not called a-leaf**, because a symbol whose calls we failed to resolve looks exactly like a-function that genuinely calls nothing, and asserting the second is a claim the-graph cannot support. And when a confidence floor is what stopped the walk, the-flow reports *which origins it declined*, which is the part you can act on.+Two details carry most of the value. `no_callees` is deliberately not called a+leaf: a symbol whose calls we failed to resolve looks exactly like a function+that calls nothing, and asserting the second is a claim the graph cannot+support. When a confidence floor stopped the walk, the flow reports which+origins it declined, which is the part you can act on.---@@ -369,19 +339,19 @@ The graph is not the product. These are:visible as one.- **Dead code.** Reachability over the union view, not over `calls` alone. Asymbol reached only by a framework, a dispatch table or a type reference is- not dead, and each of those is a different edge type for exactly this reason.+ not dead, and each of those is a different edge type for this reason.- **Communities.** Leiden clustering (Louvain as fallback) finds the modules- your codebase actually has, which is frequently not the directory layout.+ your codebase actually has, which is often not the directory layout.- **Centrality.** PageRank over the file tier ranks what everything depends on.Betweenness finds the bridges whose removal splits the graph. Neither is fedco-change edges, because "changes alongside many things" is not "many thingsdepend on it".- **Cycles.** Strongly connected components, which need their own documentation- strategy because nothing in them can be explained before the others.+ order because nothing in them can be explained before the others.- **Execution flows.** Entry point to leaf, ranked, with the termination reasonattached.-- **Framework wiring.** 22 detectors connect routes to handlers, DI- registrations to implementations, and ORM entities to their relationships,+- **Framework wiring.** Framework handlers connect routes to handlers, DI+ registrations to implementations and ORM entities to their relationships,across Django, FastAPI, Flask, Spring, ASP.NET, Rails, Laravel, Next.js,Express, Axum, Gin and more.@@ -389,61 +359,68 @@ The graph is not the product. These are:## Seeing it yourself-**For your agent**, via MCP:+**For your agent**, over MCP:| Tool | Gives you ||---|---|-| `get_context(targets)` | Dependencies, dependents and co-change partners for a file or symbol |+| `get_context(targets)` | Dependencies, dependents and co-change partners for a file or symbol; `include=["callers"]` for callers with confidence |+| `get_symbol(id, depth)` | One symbol's verified body; depth 2-3 adds the bodies it calls || `get_risk(targets)` | What history and the graph say about touching these paths |-| `get_execution_flows()` | Traced flows with their termination reason || `get_dead_code()` | Unreachable files, unused exports and zombie packages, by confidence tier |-| `get_blast_radius()` | Cross-repo impact, in workspace mode |+| `get_execution_flows()` | Traced flows with their termination reason (opt-in) |+| `get_blast_radius()` | Cross-repo impact, in workspace mode (opt-in) |++Opt-in tools are enabled through the tool surface settings; see+[MCP_TOOLS.md](../agent/MCP_TOOLS.md#get_execution_flows).++**In the dashboard**, `repowise serve` gives you the Graph, Architecture,+Coupling, Blast radius, Knowledge graph and Dead code views.-**In the dashboard**, `repowise serve` gives you the graph, architecture,-coupling, blast-radius, knowledge-graph and dead-code views.+**Across a multi-repo workspace**, a service-level system graph built from+contracts and package dependencies adds declared dependency rules, cycle+detection, a dependency-structure matrix and a 1-10 architecture score. Those+rules work between services, not between layers inside one repository. See+[Architecture Conformance](../scale/WORKSPACES.md#architecture-conformance) and+[Architecture Metrics](../scale/WORKSPACES.md#architecture-metrics).-**Everything above is computed without a single model call.** An LLM is an-optional upgrade for prose quality in the wiki. It is never part of building the-graph, which is why the graph is reproducible and why indexing needs no API key.+**Everything above is computed without a model call.** An LLM is an optional+upgrade for prose quality in the wiki. It is never part of building the graph,+which is why the graph is reproducible and why indexing needs no API key.---## Honest ceilings-- **Resolution quality varies by language**, and the [language support- page](LANGUAGE_SUPPORT.md) says how per language. Statically typed languages- with explicit declarations resolve best. Dynamically typed and heavily- reflective code resolves worst, and falls back to the low-confidence origins- rather than pretending.+- **Resolution quality varies by language.** The+ [language support page](LANGUAGE_SUPPORT.md) says what each rung covers.+ Statically typed languages with explicit declarations resolve best.+ Dynamically typed and reflective code resolves worst, and falls back to the+ low-confidence origins.- **Receiver typing reads declarations with per-language patterns**, not a fulltype checker. A declaration the patterns cannot see is a receiver that doesnot get typed, and the call falls back to a weaker origin.- **`dispatches_to` compares no signature.** It matches by method name, so it- names a possible dispatch target rather than a proven override.+ names a possible dispatch target, not a proven override.- **A repo-wide unique name is still a guess**, and `global_unique` at 0.50 iswhere that lives. It is kept because a labelled guess beats a missing edge for- reachability, and it is labelled so nothing downstream can mistake it for a- fact.-- **Method-level dead-code detection is not shipped**, for any language. It was- measured, precision failed, and shipping it would have meant confidently- recommending deletions that were wrong.-- **Roughly fifteen percent of our call edges are wrong**, and **we lead no Go- recall cell**. Both are measured, both are above under [how good is it, and how- we know](#how-good-is-it-and-how-we-know), and neither is buried down here.-- **Two competing tools are more precise than us in five of seven oracle cells.**- Each of them draws a much smaller graph, which is the whole reason, and it is- stated above rather than left out. A precision figure quoted without the recall- beside it is a misuse of this data, including by us.-- **The compiler-graded reading covers two languages.** Go and TypeScript are the- only ones with an oracle, so on the other seventeen the precision figure is the- hand-graded one, at 30 rows per language. That is a smaller n and a method we- ran ourselves; the two agree to within a point where both exist, which is the- reason to trust the half where only one does.-- **A competitor's coverage lead is only priced on two languages too.** The tool- that beats us on cross-file coverage across 35 repositories has an- oracle-anchored precision figure on go and typescript alone. That its extra- edges are mostly wrong is measured there and inferred elsewhere, and the- benchmark says so rather than generalising quietly.+ reachability, and it is labelled so nothing downstream mistakes it for a fact.+- **Unused methods are not reported**, for any language. Method-level precision+ did not hold up when measured, and shipping it would mean confidently+ recommending wrong deletions.+- **About fourteen percent of our call edges are wrong** on the hand-graded+ sample (85.7% correct over 280 rows), and we trail on recall in four of seven+ compiler-graded cells.+- **Overloads cost precision on Java and C#.** The graph keeps one node per+ overload set, so a call to the right method can land on the wrong overload.+ On C#, plan against the compiler figure (0.73 to 0.92), not the hand-graded+ 30 of 30.+- **Precision is published for ten languages, not all 26.** Go, TypeScript,+ Java, C#, C, C++, Rust and Python have a compiler or analyser answer key.+ Kotlin and Swift have the hand-graded sample only, at 30 rows each. The other+ full-AST languages have no published precision figure.+- **A precision figure without the recall beside it is a misuse of this data**,+ including by us. Two tools beat us on precision in some compiler cells, each by+ drawing a much smaller graph.---@@ -453,7 +430,7 @@ graph, which is why the graph is reproducible and why indexing needs no API key.- [architecture/language-support.md](../architecture/language-support.md) · call resolution internals and the contributor recipe- [architecture/graph-algorithms.md](../architecture/graph-algorithms.md) · PageRank, Leiden, betweenness and SCC in detail- [DEAD_CODE.md](DEAD_CODE.md) · how reachability becomes a confidence-tiered report-- [CHANGE_RISK.md](CHANGE_RISK.md) · how live diff-shape review and structural PR impact remain distinct+- [CHANGE_RISK.md](CHANGE_RISK.md) · how live diff-shape review and structural PR impact stay distinct- [reference/COMPUTED_GLOSSARY.md](../reference/COMPUTED_GLOSSARY.md) · every derived metric, defined-- [BENCHMARKS.md §7](../BENCHMARKS.md#7-edge-precision) and [§8](../BENCHMARKS.md#8-the-same-question-against-an-answer-key-we-do-not-control) · the precision numbers on this page, with their sample sizes and intervals-- [repowise-bench/graph](https://github.com/repowise-dev/repowise-bench/tree/master/graph) · the harnesses, the five arms, the graded rows and the pre-registrations behind all of it+- [BENCHMARKS.md](../BENCHMARKS.md#accuracy-by-language) · the accuracy numbers on this page, with sample sizes, intervals and method+- [repowise-bench/graph](https://github.com/repowise-dev/repowise-bench/tree/master/graph) · the harnesses, the graded rows and the pre-registrationsdiff --git a/docs/layers/INTELLIGENCE_LAYERS.md b/docs/layers/INTELLIGENCE_LAYERS.mdindex 2e430ee41..9b38998a5 100644--- a/docs/layers/INTELLIGENCE_LAYERS.md+++ b/docs/layers/INTELLIGENCE_LAYERS.md@@ -1,410 +1,170 @@# The Intelligence Layers-repowise indexes your codebase **once**, builds five foundational layers, then-keeps them in sync on every commit. Four further layers are derived from those-five. This document is the deep dive; the README gives the one-paragraph version-of each and links here for the detail.--<div align="center">-<img src="../../.github/assets/intelligence-layers.svg" alt="repowise's intelligence layers: one index (repowise init) fans into Graph, Git, Docs, Decisions, and Code Health, each surfaced through its signature MCP tool and delivered through task-shaped tools, the CLI, the local dashboard, auto-generated CLAUDE.md/AGENTS.md, and the PR bot" width="100%" />-</div>--The layers compound. The graph locates what git flags, code health scores it,-decisions explain why it is shaped that way, and the docs make all of it-searchable in natural language.--**Foundational**--1. [Graph Intelligence](#graph-intelligence)-2. [Git Intelligence](#git-intelligence)-3. [Documentation Intelligence](#documentation-intelligence)-4. [Decision Intelligence](#decision-intelligence)-5. [Code Health Intelligence](#code-health-intelligence)--**Derived**: built on the five, each with its own reference page--6. [Change Risk](#change-risk) · [Test Intelligence](#test-intelligence) ·- [Bug History](#bug-history) · [Security](#security) ·- [Dead Code](#dead-code-detection)--**Cross-cutting**--- [Proactive context enrichment: hooks](#proactive-context-enrichment-hooks)-- [Auto-sync](#auto-sync-five-ways-to-stay-current)-- [Auto-generated CLAUDE.md](#auto-generated-claudemd)--Everything below is computed **without model calls**, with two named exceptions,-both optional and both skipped entirely on a keyless index: prose quality in the-docs layer, and comment archaeology in the decision layer. Neither is a-requirement for the index.-------## Graph Intelligence--tree-sitter parses your source into a **two-tier dependency graph**: file nodes-and symbol nodes (functions, classes, methods). 26 languages parse to a full-AST; see [`LANGUAGE_SUPPORT.md`](LANGUAGE_SUPPORT.md) for per-language tiers.--A **confidence-scored call resolver** handles import aliases, barrel-re-exports and namespace imports. Every `calls` edge is stamped with one of 39-named resolution origins (`ResolutionOrigin` in `ingestion/models.py`), each-carrying a fixed confidence from 0.95 (`same_file`) down to 0.50-(`global_unique`, a repo-wide name match, labelled as the guess it is). That-includes per-language tiers for Go packages, JVM same-package and C++-same-target, and twelve receiver-typing origins that resolve a call on a-variable by reading the variable's declaration (Java, C#, Python, Go, Kotlin,-Swift). Heritage extraction covers `extends`, `implements` and trait impls.--- **A closed edge vocabulary** keeps inference from reading as fact. A- `calls` edge means the parser saw a call; `references` means something only- holds a handle to the function; `dispatches_to` names an implementation that- *could* answer for a base method; `framework_binds` is wiring a framework- performs and no parser could have seen. Consumers read the derived edge sets- rather than re-deriving their own filter.-- **An execution flow says why it stopped**, from a six-value vocabulary: a real- end, a cycle, an exhausted hop budget, an all-excluded successor set, a- confidence-filtered one, or calls that failed to resolve. A trace that simply- ends is the one thing it never reports.-- **Leiden community detection** (Louvain fallback) finds logical modules even- when your directory structure doesn't reflect them.-- **PageRank, betweenness centrality, SCC analysis, and execution-flow tracing**- from entry points identify your most central, most coupled, and most traversed- code.-- **Framework-aware edges** connect routes to handlers across 22 framework- detectors, including Django, FastAPI, Flask, ASP.NET, Spring, Micronaut,- Quarkus, Jakarta, Express/NestJS, Next.js App Router, Remix, tRPC, Hono,- Gin/Echo/Chi, Axum/Actix/Rocket, Rails, Laravel, TYPO3, Flutter and the- pytest/gtest test runners.--Full detail, including the resolution-origin table, receiver typing worked-through, and what each edge type does and does not claim:-[`GRAPH.md`](GRAPH.md). Every derived metric is defined in-[`COMPUTED_GLOSSARY.md`](../reference/COMPUTED_GLOSSARY.md).-------## Git Intelligence--repowise mines your git history to produce signals no static analysis can find.--**Hotspots**: files in the top quartile of *decayed* churn (exponential-half-life, so last month outweighs last year) that also clear minimum-activity-floors, at least 3 commits in 90 days and a meaningful temporal score. Churn-alone would flag every file a bulk refactor touched; the floors keep the list-about sustained activity. Surfaced by `get_risk()` before your agent edits.--**Ownership**: `git blame` aggregated into per-author percentages. Know who to-ping, and where knowledge silos are.--**Co-change pairs**: files that change together in the same commit *without* an-import link. Hidden coupling AST parsing cannot detect. `get_context()` surfaces-co-change partners alongside direct dependencies.--**Bus factor**: how many authors it takes to cover 80% of a file's commits. A-bus factor of 1 is the classic single-owner risk, surfaced in `CLAUDE.md`.--**Significant commits**: up to 50 meaningful commit messages per file, merges,-dependency bumps, lint-only and bot commits filtered out, with the commit body-retained when it carries decision intent. These feed generation prompts, so the-wiki can explain *why* code is structured the way it is.--**Contributor profiles**: every author gets a page, modules they own, top-files, co-authors, commit category mix, silo modules, bus-factor risk files, and-dead-code burden.--**Module health**: a 0–100 composite per top-level module from silo penalty,-hotspot density, dead-code percentage, average churn, doc coverage and median-bus factor.--**Reviewer suggestions**: paste a PR file list into Blast Radius for a ranked-reviewer list, scored by direct authorship (×1.0), co-change partners (×0.5) and-recency (×0.4), capped at the 5 strongest co-change signals per file.-------## Documentation Intelligence--A wiki for every module and file, rebuilt **incrementally** on every commit.-Deterministic templates render it with no model calls; supplying an LLM key-upgrades the prose rather than enabling the layer.--- **Coverage tracking**: what's documented and what isn't.-- **Freshness scoring** per page, relative to the underlying code.-- **Semantic search via RAG**: hybrid retrieval merging full-text and vector- results through Reciprocal Rank Fusion, with PageRank bias and a 1–2 hop graph- expansion that walks the imports and projected-calls graph for flow-shaped- questions.--A typical single-commit update regenerates only the handful of pages your change-actually touched.--Reference: [`WIKI.md`](WIKI.md).-------## Decision Intelligence--Architectural decisions mined at index time from **five sources**: ADR files-(Nygard/MADR), PR and squash-commit bodies, inline markers, git archaeology, and-centrality-bounded code comments. Two further capture paths exist for humans and-agents: `repowise decision add`, and mining your Claude Code or Codex session-transcripts. **Seven sources in total, six of them deterministic**; comment-archaeology is the one that reads prose with a model, and it is skipped on a-keyless index. Three further sources were retired for producing records nobody-ever acted on, which [`DECISIONS.md`](DECISIONS.md) names.--```python-# WHY: JWT chosen over sessions — API must be stateless for k8s horizontal scaling-# DECISION: All external API calls wrapped in CircuitBreaker after payment provider outages-# TRADEOFF: Accepted eventual consistency in preferences for write throughput-```--Every decision is **evidence-backed**: each rationale traces to a verbatim source-span, and an anti-hallucination substring gate stamps it **exact**, **fuzzy** or-**unverified**. Corroborating sources raise confidence rather than overwrite each-other.--**On lineage, precisely.** Decisions have a typed-edge schema-(`supersedes` / `refines` / `relates_to` / `conflicts_with`), but **no edges are-written today**. The only detector that produced them scoped conflicts by-similarity, which does not scope, so it is disabled and the edges it wrote were-removed. Lineage stays empty until a structural detector replaces it. What does-work: the diff-driven pass on `repowise update` marks decisions a new commit-reversed, and `repowise decision deprecate --superseded-by` records the successor-on the record itself. See [`DECISIONS.md`](DECISIONS.md).--These records surface everywhere your agent already looks: `get_why()` for the-archaeology, governing decisions in `get_context()`, a `governance_risk` flag in-`get_risk()` PR review, a Key Decisions section in `get_overview()`, and the-`ungoverned_hotspot` / `stale_governance` / `contradictory_decision` findings in-the code-health layer.--```bash-repowise decision add # guided interactive capture-repowise decision confirm # review auto-proposed decisions-repowise decision health # stale, conflicting, ungoverned hotspots-```--The "why" usually walks out the door: when a teammate leaves, or when you reopen-your own repo six months later. This keeps it in the codebase.-------## Code Health Intelligence--repowise scores **every file 1–10** on three co-equal signals (defect risk,-maintainability, and performance risk) from a roster of **51 deterministic-detectors**, of which only **26 are permitted to move the defect number**. Pure-static analysis over tree-sitter and git data, budgeted (and CI-tested) to-finish in **under 30 seconds on a 3,000-file repo**.--The defect weights are **calibrated offline against a real bug corpus, not-hand-tuned**: every file scored at a commit preceding the bug window (no-leakage), an L2-logistic fit with NLOC as an explicit control so a marker only-earns weight for defect lift *beyond* file size. Only the learned constants-ship; the runtime stays fully deterministic.--Validated leakage-free across **21 repositories, 9 languages, 2,826 files** at-mean ROC AUC **0.737**, and **0.76–0.78** on the public PROMISE/jEdit dataset-that played no part in calibration.--It does not stop at scoring; the layer closes the loop into concrete,-graph-aware refactoring plans an agent can execute: Extract Class, Extract-Method, Extract Helper, Move Method, Break Cycle and Split File, each with its-plan, recovered impact and blast radius.--```bash-repowise health # KPIs + lowest-scoring files-repowise health --refactoring-targets # ranked by impact / effort-repowise health --trend # snapshots + declining-health alerts-```--The same layer also checks **documentation drift**: the claims your markdown-makes about the tree (a path, a link, a heading, a command), resolved against-the graph, with the refuted ones reported. No model, and it runs on every-update. [`DOC_DRIFT.md`](DOC_DRIFT.md).--Full guide, the calibration story and the head-to-head against CodeScene:-[`CODE_HEALTH.md`](CODE_HEALTH.md).-------## Change Risk--A deterministic live-diff assessment for **a whole commit or `base..head`-range**, computed against the live checkout with no index lookup or model call.-Lead with the benchmarked `risk_percentile` and `classification`, which rank the-diff-shape score against sampled recent commits in the same repo. The supporting-0–10 score is calibrated at single-commit granularity and is not a probability.-Distinct from `get_risk()`, whose PR-blast value is an uncalibrated structural-heuristic over indexed files.--```bash-repowise risk HEAD-repowise risk main..feature-branch-```--Reference: [`CHANGE_RISK.md`](CHANGE_RISK.md).-------## Test Intelligence--Which tests actually exercise the code you changed, which changed files have no-guarding test at all, and per-file coverage merged across every test that touches-it. Coverage ingests from LCOV, Cobertura, Clover, JaCoCo, Go coverprofile or-normalized JSON and feeds the code-health coverage markers. The same report gates-a pull request on its patch coverage with `repowise coverage check`.--**It answers with or without that ingest.** Most repositories never produce a-report, so where one is missing the layer walks the call graph instead: a test-whose calls reach a file executes it, three hops by default, filtered to the call-edges the graph is confident in. Measured against a real-`coverage run --contexts=test`, that reaches **95.7% precision** on what covers a-file and **97.5% on the run list**. Rows are stamped `basis: "measured"` or-`basis: "inferred"`, the measured tier wins outright wherever it can answer, the-two are never averaged, and the inferred tier may never emit a percentage.--The practical payoff is `tests_to_run` in `get_risk()` PR mode: a real list rather-than a guess, so an agent runs the tests that can actually catch its change, on-repositories that have never configured coverage at all.--Reference: [`TEST_INTELLIGENCE.md`](TEST_INTELLIGENCE.md).-------## Bug History--Bug-fix commits attributed back to the files and functions they repaired, giving-each file a fix count, a last-fixed age, and a "bug magnet" flag. This is why the-generated `CLAUDE.md` orders its *files that need care* by **bug-fix history-first, then churn**: a file that keeps breaking is a better warning than a file-that merely changes often.--Reference: [`BUG_HISTORY.md`](BUG_HISTORY.md).-------## Security--A static security scan over the working tree **and the full history**, so a-finding carries the commit that introduced it and its author, not just a line-number. Findings are idempotent across re-scans and surface through-`repowise security`, the REST API and the dashboard.-------## Dead code detection--Pure graph traversal and SQL. No model calls.+Repowise builds one local index with five layers. Everything it does, from dead code+to change risk to the answers your agent gets, is one of those layers or a+combination of them. This page is the map: what each layer holds, what is built on+it, and which page has the detail.```-repowise dead-code-- 23 findings · 4 safe to delete-- ✓ utils/legacy_parser.ts file 1.00 safe to delete- ✓ auth/session.ts file 0.92 safe to delete- ✓ helpers/formatDate export 0.71 safe to delete- ✗ analytics/v1/tracker.ts file 0.41 recent activity — review first+ What you use answers · change risk · Fix first · dead code · impacted tests+ refactoring plans · doc drift · contracts across repositories+ ▲+ Five layers 1 Graph · 2 Git history · 3 Docs · 4 Decisions · 5 Health+ ▲+ One index your source, its git history, your markdown, your coverage```-Conservative by design. `safe_to_delete` requires confidence ≥ 0.70 and excludes-14 dynamically-loaded naming patterns (`*Plugin`, `*Handler`, `*Adapter`,-`*Middleware`, `*Mixin`, `*Command`, `register_*`, `on_*`, `*_view`,-`*_endpoint`, `*_route`, `*_callback`, `*_signal`, `*_task`). Dynamic-import-detection and a per-language framework-convention registry further cut false-positives. repowise surfaces candidates; engineers decide.--Reference: [`DEAD_CODE.md`](DEAD_CODE.md).-------## Proactive context enrichment: hooks--Most MCP tools are passive; the agent has to know to call them. repowise hooks-are active: they act on what the agent is already doing. No LLM calls, no-network, pure local SQLite queries. Installed automatically during-`repowise init`.--There is deliberately **no unconditional pre-search injection**. An earlier-version enriched every `Grep`/`Glob` before it ran and was removed: it added-noise on the majority of searches where the agent had already found what it-wanted. What ships now fires on evidence that the agent needs help:--- **Zero-result rescue**: a grep found nothing, so wiki full-text, fuzzy symbol- and decision matches surface the closest real hit.-- **Flood digest and triage**: a search returning far too much is replaced with- a compact per-file digest of match counts and anchor lines, plus the top files- by PageRank.-- **Wrong-path rescue**: a failed `Read`/`Edit`/`Write` path is resolved against- the index to the file you almost certainly meant.-- **Read skeleton**: an unbounded read of a large indexed file is served as its- skeleton with 1-indexed ranges, once per file per session.-- **Stale-read notice**: a file was edited after this session read it earlier.-- **Decision injection**: editing a file with a governing decision surfaces it- in one line.-- **Session start**: index freshness, the trust protocol, and standing- decisions.-- **Post-commit staleness**: after a successful commit, merge, rebase,- cherry-pick or pull, a notice if the wiki has drifted from HEAD.--Claude Code and Codex are both supported.--> **Related capability:** [Distill](../agent/DISTILL.md) reuses the index-> (symbol bounds, centrality, hotspots) to compress noisy command output and-> large reads before the agent sees them, built *on* the layers, not a layer.-------## Auto-sync: five ways to stay current--| Method | Command | Best for |-|--------|---------|----------|-| **Post-commit hook** | `repowise hook install` | Set-and-forget local development |-| **File watcher** | `repowise watch` | Active development without committing |-| **GitHub webhook** | Configure in repo settings | Teams, CI/CD |-| **GitLab webhook** | Configure in project settings | Teams, CI/CD |-| **Polling fallback** | Automatic with `repowise serve` | Safety net for missed webhooks |--```bash-repowise hook install # post-commit hook (current repo)-repowise hook install --workspace # all workspace repos-repowise watch # or use the file watcher-```--Updates are incremental: only the pages your change actually touched are-regenerated. Full guide: [`AUTO_SYNC.md`](../scale/AUTO_SYNC.md).-------## Auto-generated CLAUDE.md--After every `repowise init` and `repowise update`, repowise regenerates your-`CLAUDE.md` from actual codebase intelligence, not a template. No LLM calls. An-`AGENTS.md` generator shares the same pipeline.--```bash-repowise generate-claude-md-```--The generated section includes: index freshness and a `_meta` explainer, how to-work in this repo, a **trust protocol** stating when a served result may be used-without re-reading, the MCP tool table, an architecture summary, key modules,-entry points, **files that need care** (ordered by bug-fix history, then churn),-code health across all three signals, standing architectural decisions, and your-build/test commands. A user-owned section at the top is never touched.--```markdown-<!-- REPOWISE:START — Do not edit below this line. Auto-generated by Repowise. -->-## Architecture-Monorepo with 4 packages. Entry points: api/server.ts, cli/index.ts.--## Files that need care (bug-fix history first, then churn)-- payments/processor.ts — 19 bug fixes, last fix 2 days ago (bug magnet); 23 commits/90d--## Standing decisions (ask get_why before diverging)-- JWT over sessions (auth/service.ts) — stateless required for k8s horizontal scaling-<!-- REPOWISE:END -->-```+`repowise init` parses the code and reads the history once. `repowise update` keeps+it current after each commit, re-analysing only what changed. Almost everything is+computed without a model: a provider key adds written prose to the docs layer and a+few optional extras, and never switches a layer on.++## 1. Graph: what is connected to what++Every file, symbol, import and call in the repository, parsed from an AST in 26+languages. Each call edge is stamped with how it was resolved (one of 39 named+resolution origins) and how far to trust it, across 17 edge types. Communities,+centrality, cycles and execution flows from each entry point are computed on top.++The graph is what lets every other layer point at the right code: git history+attaches to files the graph knows, health scores what the graph parsed, tests reach+code through graph edges.++| Built on it | What you get | Read more |+|---|---|---|+| Call graph and execution flows | Callers, callees, dependency paths and traced flows, each edge with a confidence | [GRAPH.md](GRAPH.md) |+| Dead code | Unreachable files, unused exports, unused packages, by confidence tier | [DEAD_CODE.md](DEAD_CODE.md) |+| Inferred test links | Which tests reach a file, without a coverage report | [TEST_INTELLIGENCE.md](TEST_INTELLIGENCE.md) |+| Cross-repo contracts | HTTP, gRPC, topic and OpenAPI contracts matched between repositories, with breaking-change detection | [WORKSPACES.md](../scale/WORKSPACES.md) |+| Language coverage | What each of the 40 supported languages gets | [LANGUAGE_SUPPORT.md](LANGUAGE_SUPPORT.md) |++## 2. Git history: where change and breakage concentrate++Churn, hotspots, co-change pairs, ownership, bus factor and bug-fix commits, read from+the repository's own history. Static analysis cannot see any of this: two files with+no import between them can still change together every week.++| Built on it | What you get | Read more |+|---|---|---|+| Ownership and knowledge risk | Owners, bus factor, silos, knowledge loss when a main author goes quiet, suggested reviewers | [OWNERSHIP.md](OWNERSHIP.md) |+| Bug history | Fix commits traced to files and symbols, bug magnets | [BUG_HISTORY.md](BUG_HISTORY.md) |+| Hidden coupling | Files that change together with no import link | [CHANGE_RISK.md](CHANGE_RISK.md) |+| Branch overlap | Other branches editing the same files or their co-change partners | [CHANGE_RISK.md](CHANGE_RISK.md#branch-overlap) |+| Secrets in history | Keys and tokens anywhere in the full history, not only the working tree | [SECURITY.md](SECURITY.md) |++## 3. Docs: the codebase, readable and searchable++A wiki page for every module and file, rendered from the code's structure with no+key, or written by a model when you choose. Full-text, semantic, symbol and path+search over code and pages. Pages carry freshness and confidence, and rebuild+incrementally.++| Built on it | What you get | Read more |+|---|---|---|+| Generated wiki | Module, file, layer, API and onboarding pages, with diagrams | [WIKI.md](WIKI.md) |+| Search and cited answers | `get_answer`, `search_codebase`, `repowise ask` | [WIKI.md](WIKI.md) |+| Doc drift | Your own markdown checked against the tree: every reference to code that is gone | [DOC_DRIFT.md](DOC_DRIFT.md) |+| Agent instructions | A managed `CLAUDE.md` and `AGENTS.md` regenerated from the index | [editor-files.md](../architecture/editor-files.md) |++## 4. Decisions: why the code is shaped this way++Architectural decisions found in ADR files, `# WHY:` and `# DECISION:` comments,+commit and PR history, and your own agent sessions. Each decision is tied to the code+it governs and to the evidence it came from, and it is flagged when the code moves+away from it.++| Built on it | What you get | Read more |+|---|---|---|+| Decision records and lifecycle | Candidates, accepted, superseded and stale decisions, with evidence spans | [DECISIONS.md](DECISIONS.md) |+| `get_why` | The governing decision for a file, or git archaeology when none exists | [DECISIONS.md](DECISIONS.md) |+| Ungoverned hotspots | Heavily changed files no decision explains | [DECISIONS.md](DECISIONS.md) |++## 5. Health and change intelligence: what to fix, and what a change will do++53 deterministic detectors score every file 1 to 10 across defect risk,+maintainability and performance. 25 of them may move the defect score, calibrated+against real bug history. The same machinery judges a change before it merges.++| Built on it | What you get | Read more |+|---|---|---|+| Code health score | Per-file score, bands, trends, badges | [CODE_HEALTH.md](CODE_HEALTH.md) |+| Fix first | A ranked queue of what to fix, by impact and effort (`repowise next`) | [CODE_HEALTH.md](CODE_HEALTH.md#fix-first) |+| Performance findings | N+1 queries, I/O in loops, blocking calls in async code, traced across functions | [CODE_HEALTH.md](CODE_HEALTH.md#performance-findings) |+| Refactoring plans | Extract Class, Extract Method, Extract Helper, Move Method, Break Cycle, Split File, Performance Fix | [REFACTORING.md](REFACTORING.md) |+| Change risk | Where a commit or range ranks against the repo's recent history, and what it made worse | [CHANGE_RISK.md](CHANGE_RISK.md) |+| Tests and coverage gates | Impacted tests, patch coverage, untested hotspots | [TEST_INTELLIGENCE.md](TEST_INTELLIGENCE.md), [CI.md](../start/CI.md) |+| Security signals | Secrets and risky calls, with a pre-commit check and a CI gate | [SECURITY.md](SECURITY.md) |++## Where the layers meet++The most useful answers cross layers. A few examples:++**Untested hotspot (graph + git + health).** The `untested_hotspot` detector fires on+a file that churns heavily in git, is central in the graph (four or more dependents,+or top-decile activity), and is under-tested: low line coverage where you ingested a+report, or no test reaching it through the graph where you did not. It can lead the+Fix first queue.++**What breaks if I change this (graph + git + health).**+`get_risk(changed_files=[...])` reads dependents from the graph, the files that+usually change alongside these from git, and the tests that reach the change. It+leads with a directive: what may break, which co-changes look missing, which tests to+run. Each test row says whether its basis is measured or inferred.++**A warning at edit time (git + agent hooks).** When an agent touches a file with a recent+run of bug fixes, a hook adds one line to its context: how many fixes, how recent,+which function they cluster in. The generated `CLAUDE.md` lists files that need care+in the same order.++**Missing rationale (decisions + git + health).** A hotspot no recorded decision+governs shows up in `repowise decision health`, in `get_why`, and as an+`ungoverned_hotspot` finding in code health.++**Across repositories (graph + git, per repo).** In a workspace, contract links from+the graph and co-change from git combine into cross-repo blast radius, consumer test+impact and architecture rules you can gate in CI. See+[WORKSPACES.md](../scale/WORKSPACES.md).++## Reaching the layers++| Question | CLI | MCP tool |+|---|---|---|+| What is this repository? | `repowise serve` (dashboard) | `get_overview` |+| How does X work, where is Y? | `repowise ask`, `repowise search` | `get_answer`, `search_codebase` |+| What is this file or symbol, who calls it? | `repowise context`, `repowise symbol` | `get_context`, `get_symbol` |+| What breaks if I change it? | `repowise risk` | `get_risk`, `get_change_risk` |+| Why is it built this way? | `repowise why` | `get_why` |+| What should we fix first? | `repowise next`, `repowise health` | `get_health` |+| What can we delete? | `repowise dead-code` | `get_dead_code` |+| Which tests matter? | `repowise impacted-tests` | `get_risk(changed_files=[...])` |++18 MCP tools are registered; a single-repo server exposes 10 by default and the rest+are opt-in. See [MCP_TOOLS.md](../agent/MCP_TOOLS.md) and the+[CLI reference](../reference/CLI_REFERENCE.md). Every derived metric is defined in+the [computed glossary](../reference/COMPUTED_GLOSSARY.md).++## Delivery to agents++Most MCP tools wait to be called. Two mechanisms put the layers in front of the agent+without that:++- **Hooks.** `repowise init` installs agent hooks for Claude Code (and Codex with+ `--codex`). They fire on session start, searches, reads, edits and failed paths, and+ add a line only when there is evidence the agent needs it: a governing decision, a+ bug-magnet file, a stale read, a search that found nothing. No model calls, no+ network. See [HOOKS.md](../agent/HOOKS.md).+- **Generated `CLAUDE.md` and `AGENTS.md`.** `init` and `update` regenerate a managed+ section from the index: architecture summary, key modules, entry points, files that+ need care, code health, standing decisions and your build commands. Anything you+ write outside the markers is kept.++[Distill](../agent/DISTILL.md) uses the same index to compress noisy command output+before the agent reads it.++## Freshness++`repowise init` installs a post-commit hook by default, so every commit runs an+incremental `repowise update` in the background. A file watcher, GitHub and GitLab+webhooks, and a polling fallback under `repowise serve` cover other setups. When the+index falls behind, the session-start hook says so and MCP responses carry a+`stale_warning`. See [AUTO_SYNC.md](../scale/AUTO_SYNC.md).diff --git a/docs/layers/LANGUAGE_SUPPORT.md b/docs/layers/LANGUAGE_SUPPORT.mdindex 9d2654887..5e739fd83 100644--- a/docs/layers/LANGUAGE_SUPPORT.md+++ b/docs/layers/LANGUAGE_SUPPORT.md@@ -1,11 +1,11 @@# Language Support-**26 languages parsed to a full AST · 40 on the five-rung ladder ·-framework-aware where an ecosystem handler exists.** "Do you support X" has five useful answers-rather than two, so every language lands on a rung and the rung says what it-buys you. Everything else in your repo still appears in the wiki and is tracked-through git history. This page is the "what works for my language today"-reference.+**26 languages parsed to a full AST, 40 on the five-rung support ladder, and+framework-aware edges where an ecosystem handler exists.** Every language lands+on one rung, and the rung tells you which parts of the pipeline produce real+output for it. Files in any other language still appear in the wiki and are+tracked through git history. No LLM key is needed for any of this: parsing,+resolution and health markers are all static.<p><strong>Full tier </strong>@@ -43,88 +43,73 @@ reference.**Contents:** [Tiers](#tiers) ·[What the pipeline gives each tier](#what-the-pipeline-gives-each-tier) ·-[Why these graphs are different](#why-these-graphs-are-different) ·+[How accurate is the graph per language](#how-accurate-is-the-graph-per-language) ·[Full tier](#full-tier) · [Good tier](#good-tier) ·[Beyond code files](#beyond-code-files) ·[Code-health coverage](#code-health-coverage) ·-[Known ceilings](#known-ceilings) · [Roadmap](#roadmap)+[Known ceilings](#known-ceilings) · [Roadmap](#roadmap) ·+[Requesting or adding a language](#requesting-or-adding-a-language)---## Tiers-Every language lands in one tier, and the tier decides which pipeline stages-produce meaningful output.-| Tier | Languages | What you get ||------|-----------|--------------|-| **Full** (13) | Python · TypeScript · JavaScript · Svelte · Vue · Java · Kotlin · Go · Rust · C++ · C# · Scala · Ruby | The whole pipeline: AST symbols, import resolution, a resolved call graph, heritage, docstrings, framework edges, **and code-health markers** |-| **Good** (11) | C · Swift · PHP · Dart · Object Pascal · COBOL · GDScript · VB.NET · Elixir · F# · Objective-C | Everything above except the full health suite. Dart and Object Pascal *do* get health markers, and C, F# and Objective-C get the complexity-derived ones; Swift, PHP, COBOL, GDScript, VB.NET and Elixir don't yet. GDScript has a dedicated import resolver and Godot-specific framework edges but no named bindings (see [Known gaps](../architecture/language-support.md#gdscript--godot)) |-| **Partial** (2) | Luau / Roblox · Razor / Blazor | Luau: AST symbols and `require()` resolution (Rojo / `.luaurc` aware), no health markers yet. Razor: a component symbol per file, call edges from `@code` blocks and component tags, C# health markers; no import resolution yet |-| | | ⎯⎯ *tree-sitter parsing stops here. The rungs below are derived from git and imports, not from an AST.* ⎯⎯ |+| **Full** (13) | Python · TypeScript · JavaScript · Svelte · Vue · Java · Kotlin · Go · Rust · C++ · C# · Scala · Ruby | The whole pipeline: AST symbols, import resolution, a resolved call graph, heritage, docstrings, framework edges and code-health markers |+| **Good** (11) | C · Swift · PHP · Dart · Object Pascal · COBOL · GDScript · VB.NET · Elixir · F# · Objective-C | Everything above except the full health suite. Dart and Object Pascal get health markers; C, F# and Objective-C get the complexity-derived ones; the rest get none yet |+| **Partial** (2) | Luau / Roblox · Razor / Blazor | Luau: AST symbols and `require()` resolution (Rojo and `.luaurc` aware), no health markers. Razor: a component symbol per file, call edges from `@code` blocks and component tags, C# health markers, no import edges yet || **Lightweight** (6) | Clojure · Haskell · Lean 4 · Erlang · HTML · QML | A real file-to-file import graph, no symbol-level claims || **Structural** (8) | R · Zig · Julia · Elm · OCaml · Crystal · Nim · D | Git history only: blame, hotspots, co-change. No AST parsing |-The first three rungs are the **26 languages parsed to a full AST**; all five are-the **40** on the ladder. Both numbers are worth stating and neither is worth-stating alone, so if you only take one thing from this page, take the rung your-language sits on rather than either count.--[**SQL / dbt**](#sql--dbt) and [**shell**](#shell) sit outside this ladder on-purpose: each has a coverage shape the tiers cannot describe. SQL is parsed by-sqlglot rather than tree-sitter and gets symbols, wiki pages and health markers-but no call graph, with import edges only inside a dbt project. Shell gets-symbols, `source` edges and function-level complexity, but reachability is-meaningless for a script invoked by name. See-[Beyond code files](#beyond-code-files), along with-[config and data formats](#config-and-data).--Cross-repo contracts (HTTP routes and calls, database tables, queues, sockets)-are listed per language and framework in+Tree-sitter parsing covers the first three rungs: those are the **26 languages+parsed to a full AST**. The bottom two rungs come from git history and import+extraction, and all five together make the **40** on the ladder.++[SQL / dbt](#sql--dbt) and [shell](#shell) sit outside the ladder on purpose,+because neither fits a rung: SQL gets symbols and health markers but no call+graph, and shell gets symbols, `source` edges and complexity but no dead-code+claims. Cross-repo contracts (HTTP routes, gRPC, queues, sockets, database+tables) are listed per language and framework in[WORKSPACES.md](../scale/WORKSPACES.md#api-contract-extraction).## What the pipeline gives each tier| Stage | Full | Good | Partial | Lightweight | Structural ||-------|:----:|:----:|:-------:|:-----------:|:----------:|-| File discovery & git history | ✅ | ✅ | ✅ | ✅ | ✅ |-| AST symbol extraction | ✅ | ✅ | ✅ | — | — |-| Import resolution | ✅ | ✅ | Luau | file-level | — |-| Call graph edges | ✅ | ✅ | ✅ | — | — |-| Heritage (extends / implements) | ✅ | ✅ | — | — | — |-| Named bindings | ✅ | ✅ | — | — | — |-| Code-health markers | ✅ | Dart, Pascal | Razor | — | — |+| File discovery and git history | ✅ | ✅ | ✅ | ✅ | ✅ |+| AST symbol extraction | ✅ | ✅ | ✅ | no | no |+| Import resolution | ✅ | ✅ | Luau | file-level | no |+| Call graph edges | ✅ | ✅ | ✅ | no | no |+| Heritage (extends / implements) | ✅ | ✅ | no | no | no |+| Named bindings | ✅ | ✅ | no | no | no |+| Code-health markers | ✅ | partial, see [coverage](#code-health-coverage) | Razor | no | no || Dead code detection | ✅ | ✅* | ✅ | ✅ | ✅ |-| Semantic search & wiki pages | ✅ | ✅ | ✅ | ✅ | ✅ |+| Semantic search and wiki pages | ✅ | ✅ | ✅ | ✅ | ✅ |-Scala's import resolution is partial: it shares the JVM index with Java and-Kotlin and falls back to parsing SBT / Mill build files. Elixir and F# also-have dedicated resolvers with known gaps, described in their rows below.-Object Pascal and Objective-C resolve through the generic unit-name or-header-stem match, and COBOL resolves literal program calls, not imports.-Every other Full and Good language has a dedicated import resolver. On the Partial rung the two languages-split: Luau resolves `require()` and has no health markers, Razor has C#-health markers and no import edges yet.+Scala, Elixir and F# have dedicated import resolvers with known gaps. Object+Pascal and Objective-C resolve through a generic unit-name or header-stem match,+and COBOL resolves literal program calls, not imports. Every other Full and Good+language has a dedicated import resolver.\* COBOL is excluded from dead-code claims: JCL, schedulers and dynamic program-calls are external entry paths that the repository graph cannot observe.+calls are entry paths the repository graph cannot observe.----+Every `calls` edge records how it was resolved and carries a confidence from+that, and a call on a typed variable resolves through the variable's declared+type. [GRAPH.md](GRAPH.md) covers resolution origins, receiver typing and+execution flows.-## Why these graphs are different+## How accurate is the graph per language-Plenty of tools will hand you a call graph. Three things separate a graph you-can act on from a picture of arrows. This is the short version; the full story,-with the complete origin table and worked examples, is in-[GRAPH.md](GRAPH.md).+Call-graph precision and recall differ by language, because resolution depends+on what the grammar and the type system expose. Measured figures per language,+with the corpus and method behind them, are in+[BENCHMARKS.md](../BENCHMARKS.md#accuracy-by-language). This page states only+what each rung covers, not how well.### Every call edge says how it was resolved, and how much to trust it-Most graphs give you an unlabelled arrow. `A calls B`, but was `B` the only-match in the entire repo, or was it right there in the same file? Those are-wildly different claims, and a tool that renders them identically is asking you-to trust its weakest guess as much as its strongest fact.-Every `calls` edge repowise emits carries a **resolution origin** from a closedvocabulary, and every origin has exactly one confidence:@@ -175,754 +160,222 @@ the MCP tools and the web UI surface which origin produced an edge, so an agentreading an execution flow can tell a fact from an inference instead of treatingboth as source.-### A call on a variable resolves by typing the variable--`user.save()` is only a useful edge if something worked out what `user` is. The-naive approach matches `save` against every method named `save` in the repo and-picks one, which is how graphs end up confidently wrong.--repowise instead reads the receiver's **declaration** and resolves the method on-that type. It covers the receiver shapes that actually occur:--- **locals and parameters**: `var repo = new UserRepo()`, `fun f(r: UserRepo)`-- **fields**: a call on `this.repo`, typed from the enclosing class-- **freshly constructed receivers**: `new Foo().bar()`-- **the method's own receiver**: Go's `func (s *Server) handle()`-- **framework-retyped symbols**: `@shared_task def add` makes `add` a `Task`,- so `add.s()` resolves to `Task::s`--Shipped for **Java, C#, Python, Go, Kotlin and Swift**. Each language was gated-on measured precision before it shipped, and the ones that failed the gate are-recorded as unsupported rather than shipped loose.--### It reports silence instead of guessing--Every capability degrades to *no signal* rather than a wrong one, and each-degradation is deliberate:--- **An execution flow says why it stopped.** A trace that just ends reads the- same whether execution really ends there or the walker ran out of things it- could follow. Six terminations are named separately: a real end, a cycle, an- exhausted hop budget, an all-excluded successor set, a confidence-filtered one,- and calls that failed to resolve. Where a confidence floor stopped the walk, it- reports which origins it declined.-- **A `calls` edge means the callee is callable.** Edges pointing at properties,- constants and modules were withdrawn and re-minted as `references`: "something- holds a handle to this", which is enough to make deleting it unsafe and not- enough to claim it is invoked.-- **Framework wiring is its own edge type.** A pytest fixture injection or a- Spring `@Autowired` field is a `framework_binds` edge, never a `calls` edge.- Nothing there is a call the parser could have seen, and letting it read as one- would put an inferred hop into an execution flow as if it were source.-- **An unmapped language produces no findings, not bad ones.** A language reaches- a health dialect or it stays silent. There is no "best effort" middle.+Receiver typing, execution-flow terminations and the edge types are in [GRAPH.md](GRAPH.md).---## Full tier-Complete pipeline coverage: AST parsing, import resolution, call resolution,-named bindings, heritage, docstrings, framework-aware edges, dynamic-hint-extractors and code-health markers.+AST parsing, import resolution, call resolution, named bindings, heritage,+docstrings, framework-aware edges and code-health markers.| Language | Extensions | Import resolution ||----------|-----------|--------------|| **Python** | `.py` `.pyi` | Source-root-aware module index (`src/`, monorepo `packages/*/src`, PEP 420), `__init__.py` re-export barrels |-| **TypeScript** | `.ts` `.tsx` `.mts` `.cts` | ESM / `require()`, tsconfig path aliases, npm/yarn/pnpm workspaces, `export * from` barrels |-| **JavaScript** | `.js` `.jsx` `.mjs` `.cjs` | `import` / `require()` including CommonJS re-export shapes and member picks |-| **Svelte** | `.svelte` | The TS/JS resolver plus SvelteKit's `$lib` and Node `#`-prefixed subpath imports |-| **Vue** | `.vue` | The TS/JS resolver plus `jsconfig`/`tsconfig` aliases, directory-index components, router `import()` specifiers |-| **Java** | `.java` | `import` / `.*` / `import static` with Maven + Gradle reactor discovery, JPMS, package fan-out |-| **Kotlin** | `.kt` `.kts` | Shares the JVM workspace index with Java, so resolution is cross-language |+| **TypeScript** | `.ts` `.tsx` `.mts` `.cts` | ESM and `require()`, tsconfig path aliases, npm/yarn/pnpm workspaces, `export * from` barrels |+| **JavaScript** | `.js` `.jsx` `.mjs` `.cjs` | `import` and `require()`, including CommonJS re-export shapes |+| **Svelte** | `.svelte` | The TS/JS resolver plus SvelteKit's `$lib` and `#` subpath imports |+| **Vue** | `.vue` | The TS/JS resolver plus config aliases, directory-index components, router `import()` specifiers |+| **Java** | `.java` | Maven and Gradle reactor discovery, JPMS, `import static`, package fan-out |+| **Kotlin** | `.kt` `.kts` | Shares the JVM index with Java, so resolution is cross-language || **Go** | `.go` | Multi-module `go.mod` discovery; a package import fans out to every file in the package || **Rust** | `.rs` | `use crate::` / `super::` / `self::` with `Cargo.toml` |-| **C++** | `.cpp` `.cc` `.cxx` `.h` `.hh` `.hpp` `.hxx` `.inl` `.ipp` `.tpp` `.inc` | `#include` via `compile_commands.json` plus CMake / Bazel header maps, header↔implementation pairing |-| **C#** | `.cs` | `using` / `global using` / aliases via `.csproj` / `.sln`, MSBuild project graph, `partial` class linking |-| **Scala** | `.scala` | The shared JVM index (cross-language with Java/Kotlin), SBT / Mill build parsing as fallback |+| **C++** | `.cpp` `.cc` `.cxx` `.h` `.hh` `.hpp` `.hxx` `.inl` `.ipp` `.tpp` `.inc` | `#include` via `compile_commands.json` plus CMake and Bazel header maps, header/implementation pairing |+| **C#** | `.cs` | `using`, `global using` and aliases via `.csproj` / `.sln`, MSBuild project graph, `partial` class linking |+| **Scala** | `.scala` | The shared JVM index, with SBT and Mill build parsing as fallback || **Ruby** | `.rb` | `require` / `require_relative` with `$LOAD_PATH` probing, Gemfile externals, Rails / Zeitwerk autoloading |-All thirteen also get docstring extraction: Python, Ruby comments, JSDoc,-GoDoc, Rustdoc, Javadoc, Scaladoc, Doxygen and XML doc.-### Framework-aware edges-Routes connect to handlers, DI registrations to implementations, ORM entities to-their relationships:+Routes connect to handlers, DI registrations to implementations, and test+fixtures to the tests that use them.| Language | Frameworks ||----------|-----------|| Python | Django, FastAPI, Flask, Celery, pytest fixtures |-| Ruby | Rails (routes → controller actions, Zeitwerk), RSpec mirror edges |-| Java / Kotlin | Spring (stereotypes, `@RequestMapping`, Spring Data, `@Bean`), Jakarta / JPA, Quarkus, Micronaut, Android manifest |-| C# | ASP.NET (attribute + minimal API), EF Core, gRPC-dotnet, host-builder extensions, CommunityToolkit MVVM |+| Ruby | Rails (routes to controller actions, Zeitwerk), RSpec mirror edges |+| Java / Kotlin | Spring, Jakarta / JPA, Quarkus, Micronaut, Android manifest |+| C# | ASP.NET (attribute and minimal API), EF Core, gRPC-dotnet, host-builder extensions, CommunityToolkit MVVM || Go | net/http, gin, echo, chi, gRPC server registration |-| Rust | Axum, Actix route → handler |+| Rust | Axum, Actix || JS / TS / Svelte | Next.js App Router, Hono / Fastify / Koa / Elysia, Remix / SvelteKit / Astro, tRPC, Express / NestJS, Angular || C++ | GoogleTest, Catch2, Boost.Test, doctest, Google Benchmark, libFuzzer |+| PHP | Laravel, TYPO3 |+| Dart | Flutter route tables, `runApp()` and widget trees |+| GDScript | Godot scenes, autoloads, signal connections, `class_name` |-The dead-code analyzer knows each ecosystem's entry points, generated-file-conventions and never-flag globs, so build products and framework-invoked code-are not reported as unreachable.--### Single-file components (Svelte, Vue)+The dead-code analyzer also knows each ecosystem's entry points and generated+files, so framework-invoked code is not reported as unreachable.-A `.svelte` or `.vue` file is projected into TypeScript at **byte-identical-offsets**, so a component reuses the TypeScript queries, config and all three-health dialects verbatim and every line number points at the real source file.-The component itself becomes a symbol named after the file, `<Foo />` in markup-mints a call edge, and a handler referenced only from `on:click={inc}` or-`@click="inc"` carries an edge instead of reading as dead code.+### Single-file components and Razor-### Razor / Blazor markup (`.razor`, `.cshtml`)--Razor has the same shape, but with no usable grammar available the locator-byte-scans `@code { }`, `@functions { }` and `@{ }` interiors into a C# buffer-at byte-identical offsets, so the C# queries and health markers apply verbatim-and PascalCase markup tags mint component call edges. It sits on the Partial-rung because `@code` members land as call edges rather than symbols and the-directives are blanked, so a Razor file carries no import edges yet.+A `.svelte` or `.vue` file is projected into TypeScript at the same byte+offsets, so components get the TypeScript queries and health markers, and every+line number points at the real source. Razor (`.razor`, `.cshtml`) does the same+with its `@code` blocks projected into C#; it sits on the Partial rung because+it has no import edges yet.---## Good tierAST parsing, symbol extraction and static call resolution, plus the-language-appropriate dependency surface. Languages with imports add named-bindings, heritage and a workspace resolver where their syntax supports them.+dependency surface each language's syntax supports.| Language | Extensions | Dependency resolution ||----------|-----------|--------------|| **C** | `.c` | `#include` via `compile_commands.json` (shares the C++ grammar) |-| **Swift** | `.swift` | SPM `Package.swift` target → directory mapping, intra-module type references, `@main` entry points |-| **PHP** | `.php` | `use` declarations (grouped `use A\{B, C}` included) resolved through PSR-4 from the root and nested `composer.json` files, longest prefix first as composer does; same-namespace and `\Fully\Qualified` class references; Laravel edges (route files to controllers and aliased middleware, registered and discovered listeners, policies, providers, commands by signature), TYPO3 edges |-| **Dart** | `.dart` | `import` / `export` / `part` URIs, `package:` via every `pubspec.yaml`, Flutter route tables and `runApp()` edges. **Health markers included** |-| **COBOL** | `.cbl` `.cob` `.cobol` `.cpy` | Program IDs, sections, paragraphs and data levels; literal `CALL` and `PERFORM` targets resolve to program/procedure symbols. Dynamic calls and `COPY` edges are deliberately silent |--COBOL is Good rather than Full because it has no code-health walker. Its static-dependency surface is narrower than a module language's: literal program calls-and local procedure transfers resolve, while copybook resolution, dynamic-`CALL data-item`, dialect-specific syntax and source-format preprocessing are-explicitly deferred. Dead-code findings are also suppressed because JCL and-scheduler entry points usually live outside the indexed source graph.-| **Object Pascal** | `.pas` `.pp` `.dpr` `.dpk` `.lpr` | `uses` clauses via the generic unit-name → file-stem fallback; project files as entry points. **Health markers included** |-| **GDScript** | `.gd` | `preload(...)` / `load(...)` / `extends "res://..."` resolved as absolute paths from the nearest `project.godot`, so a repo holding many Godot projects keeps each project's `res://` namespace separate, plus scene, autoload and `class_name` edges (see [GDScript / Godot](../architecture/language-support.md#gdscript--godot)) |-| **VB.NET** | `.vb` | `Imports` through the same MSBuild project index C# uses: `.vbproj` / `.sln` parsing, `<RootNamespace>`-aware namespace lookup, NuGet package references |-| **Elixir** | `.ex` `.exs` | `alias` / `import` / `require` / `use` against a `defmodule` index, with the Mix `lib/foo/bar.ex` → `Foo.Bar` convention as the fallback; `alias Foo.{Bar, Baz}` names both modules (no heritage: `use` and `@behaviour` are not inheritance) |-| **F#** | `.fs` `.fsx` `.fsi` | `open` / `open type` against a declared-name index (namespace/module headers), plus the fsproj `<Compile Include>` compile order. `.fsi` signature files contribute imports only, no symbols. A layout that shares one namespace across many single-file projects leaves most `open` targets ambiguous by design, so unused-export findings there sit at the review tier (0.4) rather than being asserted |-| **Objective-C** | `.m` `.mm` `.h` | `#import` / `#include` through the generic header-stem index, which covers the one-class-per-file convention; a framework import names no in-repo file and resolves to nothing |+| **Swift** | `.swift` | SPM `Package.swift` targets, intra-module type references, `@main` entry points |+| **PHP** | `.php` | `use` declarations through PSR-4 from every `composer.json`, same-namespace and fully qualified class references |+| **Dart** | `.dart` | `import` / `export` / `part`, `package:` via every `pubspec.yaml` |+| **Object Pascal** | `.pas` `.pp` `.dpr` `.dpk` `.lpr` | `uses` clauses via unit-name to file-stem match; project files as entry points |+| **COBOL** | `.cbl` `.cob` `.cobol` `.cpy` | Literal `CALL` and `PERFORM` targets; dynamic calls and `COPY` stay silent |+| **GDScript** | `.gd` | `preload` / `load` / `extends "res://..."` from the nearest `project.godot` |+| **VB.NET** | `.vb` | `Imports` through the same MSBuild project index C# uses |+| **Elixir** | `.ex` `.exs` | `alias` / `import` / `require` / `use` against a `defmodule` index, Mix path convention as fallback |+| **F#** | `.fs` `.fsx` `.fsi` | `open` against a declared-name index plus the fsproj compile order; `.fsi` files contribute imports only |+| **Objective-C** | `.m` `.mm` `.h` | `#import` / `#include` through the header-stem index |---## Beyond code files-### SQL + dbt--Parsed by a dedicated sqlglot handler (multi-dialect, error-tolerant) rather-than tree-sitter.--- **DDL symbols**: `CREATE TABLE` / `VIEW` become class-kind symbols with their- columns in the signature; `CREATE FUNCTION` / `PROCEDURE` become function-kind- symbols. Both get wiki pages and `get_symbol` lookups. Set `sql_dialect` in- config for dialect-specific syntax. Any parse problem degrades the file to- passthrough, never a crash, never a guess.-- **dbt lineage**: `{{ ref('model') }}` and `{{ source(...) }}` become real- import edges, so model-level lineage, hotspots, co-change, ownership and- communities all fall out free.-- **App-to-database contracts** (workspace mode), table *providers* (DDL,- migrations, ORM models) pair with table *consumers* (SQL literals and query- builders in app code) on the Live System Map. See [WORKSPACES.md](../scale/WORKSPACES.md).-- **Health markers**: stored routines get cyclomatic complexity, plus- `sql_select_star`, `sql_update_delete_without_where` and `sql_cartesian_join`.- All of them are **uncalibrated by construction** (no defect corpus covers- procedural SQL), so they surface as findings and never move the defect- headline score.--SQL is not placed in a tier because it would misreport in both directions:-Partial would understate the health markers, and any higher tier would claim a-call graph and heritage that a DDL file does not have. Outside a dbt project,-`.sql` files get symbols and pages but no import edges at all.+### SQL / dbt++SQL is parsed by sqlglot (multi-dialect, error-tolerant), not tree-sitter.++- `CREATE TABLE` / `VIEW` become class-kind symbols with columns in the+ signature; `CREATE FUNCTION` / `PROCEDURE` become functions. Set `sql_dialect`+ in config for dialect-specific syntax. A file that fails to parse is kept as+ plain text.+- dbt `{{ ref('model') }}` and `{{ source(...) }}` become import edges, so model+ lineage, hotspots, co-change and ownership work on dbt projects. Outside dbt,+ `.sql` files get symbols and pages but no import edges.+- In workspace mode, tables defined by DDL, migrations or ORM models pair with+ the SQL in app code that reads them. See [WORKSPACES.md](../scale/WORKSPACES.md).+- Stored routines get cyclomatic complexity, plus `sql_select_star`,+ `sql_update_delete_without_where` and `sql_cartesian_join`. These are+ uncalibrated, so they appear as findings and never move the defect score.### Shell-`.sh` / `.bash` / `.zsh`. Function definitions become symbols, `source` / `.`-statements become import edges, and calls to functions defined in the same or a-sourced file resolve to call edges; external binaries like `grep` mint nothing.-Resolution covers literal paths plus the common directory-anchor idioms-(`$SCRIPT_DIR/x.sh`, `$(dirname "$0")/x.sh`, `$BATS_ROOT/lib/x.sh`); genuinely-dynamic paths stay external. Shell also gets **function-level complexity**, with-`&&` / `||` command lists counted toward CCN. No class metrics, heritage or-dead-code flagging: shell scripts are invoked by name, so static reachability is-meaningless.+`.sh` / `.bash` / `.zsh`: functions become symbols, `source` / `.` statements+become import edges (including `$SCRIPT_DIR/x.sh` and `$(dirname "$0")/x.sh`+idioms), and calls to functions in the same or a sourced file resolve. Shell+gets function-level complexity. No dead-code flagging: scripts are invoked by+name, so static reachability says nothing.### Lightweight tier-Clojure, Haskell, Lean 4, Erlang and HTML get a real file-level import graph:-imports extracted per language and resolved against a declared module-name-index. The knowledge graph runs in flow/sparse mode on the result: honest-file-to-file dependencies, no symbol-level claims.--**QML** joins the lightweight tier with imports only: module specs resolve-against a `qmldir`-declared module index and quoted references resolve relative-to the importing file. `.qml` files are **never flagged as dead code**, because-a component is instantiated by type name and loaded by the Qt runtime through-qrc, `Loader` and `qmlRegisterType`, none of which is an import.--**HTML** is import-tier on purpose. It has no functions, classes or calls, so-there are no symbols to claim, but its `<script src>` and `<link href>` become-file-level edges, including the one every Vite/webpack SPA depends on:-`index.html` → `src/main.ts`. References resolve as document- or root-relative-asset paths, never as module specifiers, and a tie yields no edge rather than a-guessed one. `.html` files are **never flagged as dead code**: whether a page is-reachable is not statically decidable.+Clojure, Haskell, Lean 4, Erlang, HTML and QML get a file-level import graph+resolved against a declared module-name index, with no symbol-level claims.+HTML contributes `<script src>` and `<link href>` edges (so `index.html` reaches+`src/main.ts` in a Vite or webpack app). HTML and QML files are never flagged+as dead code, since whether a page or a Qt component is reachable is decided at+runtime.### Config and data-OpenAPI, Protobuf, GraphQL, Dockerfile, Makefile, YAML, JSON, TOML, Terraform and-Markdown appear in the file tree and the wiki, with special handlers extracting-endpoints and targets where applicable.--Godot resource files sit here too: `.tscn` / `.tres` / `.escn`,-`project.godot` and an addon's `plugin.cfg` carry no symbols, but the paths-they name are how a Godot project reaches its scripts, so they are indexed and-read for those paths. See-[GDScript / Godot](../architecture/language-support.md#gdscript--godot).+OpenAPI, Protobuf, GraphQL, Dockerfile, Makefile, YAML, JSON, TOML, Terraform+and Markdown appear in the file tree and the wiki. OpenAPI and Protobuf also feed+workspace contracts, and Terraform feeds AWS Lambda handler edges. Godot resource+files (`.tscn`, `.tres`, `.escn`, `project.godot`, `plugin.cfg`) carry no symbols+but are read for the script paths they name.---## Code-health coverage-Health markers run off a per-language walker map that is **independent** of-`.scm` parsing: a language can parse perfectly for the graph and still need this-map before markers fire. This table is why a language is Full rather than Good.+Health markers need a per-language walker map that is separate from parsing, so+a language can parse fully for the graph and still produce no health findings.+This table is why a language is Full and not Good.| Language | Complexity / nesting | Class metrics | Assertion smells | Mock saturation | Assertion-free test | Extract Method | Performance risk ||----------|:---:|:---:|:---:|:---:|:---:|:---:|:---:|| Python | ✅ | ✅ | ✅ | ✅ | ✅ | ✅ | ✅ || TypeScript / JavaScript | ✅ | ✅ | ✅ | ✅ | ✅ | ✅ | ✅ |-| Svelte · Vue | ✅ | ✅ | ✅ | later | later | ✅ | ✅ |-| Java | ✅ | ✅ | ✅ | blocked | blocked | ✅ | ✅ |-| Go | ✅ | n/a | ✅ | blocked | blocked | ✅ | ✅ |-| Rust | ✅ | ✅ | ✅ | later | later | ✅ | ✅ |-| C++ | ✅ | ✅ | ✅ | later | later | ✅ | ✅ |-| C# | ✅ | ✅ | ✅ | later | later | later | ✅ |-| Kotlin | ✅ | ✅ | ✅ | later | later | blocked | ✅ |-| Scala | ✅ | ✅ | ✅ | later | later | later | ✅ |-| Ruby | ✅ | ✅ | ✅ | later | later | later | ✅ |-| Dart | ✅ | n/a | ✅ | later | later | later | ✅ |-| Object Pascal | ✅ | n/a | ✅ | n/a | n/a | later | ✅ |-| Razor | ✅ | n/a | n/a | n/a | n/a | later | ✅ |+| Svelte · Vue | ✅ | ✅ | ✅ | no | no | ✅ | ✅ |+| Java | ✅ | ✅ | ✅ | no | no | ✅ | ✅ |+| Go | ✅ | n/a | ✅ | no | no | ✅ | ✅ |+| Rust · C++ | ✅ | ✅ | ✅ | no | no | ✅ | ✅ |+| C# · Scala · Ruby | ✅ | ✅ | ✅ | no | no | no | ✅ |+| Kotlin | ✅ | ✅ | ✅ | no | no | no | ✅ |+| Dart | ✅ | n/a | ✅ | no | no | no | ✅ |+| Object Pascal | ✅ | n/a | ✅ | n/a | n/a | no | ✅ |+| Razor | ✅ | n/a | n/a | n/a | n/a | no | ✅ |+| C · F# · Objective-C | ✅ | no | no | no | no | no | no || Shell | ✅ | n/a | n/a | n/a | n/a | n/a | n/a |-Every cell is a deliberate call, not an oversight: a language reaches a dialect-or it stays silent, and an `n/a` records a metric the language cannot carry-rather than one nobody got to. Kotlin's Extract Method is **blocked on the-grammar**, not unscheduled.--**The assertion count** is the denominator of every test-quality ratio, counted-in two tiers (`analysis/health/asserts/lexicon.py`). The narrow tier is the-`assert` / `expect` prefix match that `large_assertion_block` and-`duplicated_assertion_block` are calibrated on, and it never changes. The broad-tier adds a per-language row plus anything a repository declares in the-`assertions:` block, and feeds the advisory count alone, so a vocabulary edit-cannot move a score. Broad-tier names are matched exactly, never as prefixes:-over three corpora the `check` / `validate` / `approve` / `fail` prefix families-matched ~600 production functions under test and not one assertion.--Counted assertions, narrow tier then both, when the tiers were introduced:--| Corpus | Test files | Narrow | Both |-|---|---|---|---|-| Go, stdlib `t.Errorf` idiom | 502 | 87 | **3,454** |-| Go, testify idiom | 476 | 1,586 | **3,656** |-| TypeScript, should.js present | 521 | 9,956 | **11,165** |-| TypeScript, no should.js | 574 | 15,595 | 15,595 |--Go was effectively invisible. Assertion *runs* were byte-identical in every-corpus, which is what makes the broad tier safe.--**Mock saturation** measures a test's mock setup against its assertions, so it-needs both a mock vocabulary and a trustworthy assertion count. The vocabulary-is one data row per language in-`analysis/health/mocks/lexicon.py`; a language with no row produces no signal at-all.--Because the marker is advisory it ships below the 70% bar, so the line is not-drawn on the precision number alone: **a language ships when its false positives-are the families the marker already declares, and is blocked when its failure is-a different one the marker cannot name.** TypeScript's every false positive was-the known assertion-observes-production-output family. The two blocked languages-fail differently, and for the same underlying reason — the assertion count is the-denominator of every ratio, so a language whose assertions cannot be counted-honestly cannot carry the marker at all:--- **Go** — the assertion half is fixed (table above). What still blocks the- marker is the other half: no row in `mocks/lexicon.py` and no hand-labelled- precision figure.-- **Java** — Mockito states a test's real checks as `verify(...)`, which is- deliberately not an assertion here (counting it would hide the very tests this- marker looks for). An over-mocked Java test therefore arrives with one vacuous- `assertDoesNotThrow` as its whole denominator, so the marker misses the tests- it is for and the ratio peaks instead on large controller tests whose setup is- irreducible fixture. Measured at 20% precision over 30 hand-labelled findings,- flat across every threshold tried — but the mechanism above, not that number,- is why it is blocked. Java now **does** carry a verification vocabulary, added- for `assertion_free_test`, and it deliberately does not help here: Mockito- verification is 1,087 calls over those 299 test files, and it reaches- `verification_count` rather than `assertion_count` precisely so that this- marker keeps dividing by state assertions alone.--**Assertion-free test** asks whether a test case checks anything at all, so it-needs two things the other markers do not: a per-function "is this a test case"-rule, and a way to tell that a test checks nothing. The second is deliberately not one-count: `assertion_count` is calibrated and shared with a scored marker, so the-shapes it cannot classify — a hand-rolled `throw`, an assertion bound to a name-— are answered beside it rather than folded into it. The rules live in-`analysis/health/complexity/test_case.py`, one data row per language — a name-prefix for Python, `go test`'s own case-sensitive `TestXxx` rule for Go, the-`it(...)` / `test(...)` callback for JS/TS, and `@Test` and its JUnit siblings-for Java. A language with no row classifies nothing. Go and Java are classified-and counted, which is how their precision below was measured, but the marker-does not report on them — the shipping set is a constant in the detector.--Measured precision, hand-labelled: **71%** on TypeScript (31 findings, the-complete population of two corpora) and **86%** on Python (29 findings, a-systematic sample of 172). **Both are now a floor rather than a current-reading**, as `mock_saturated_test`'s are a ceiling: three false-positive-families have been closed since they were labelled, which removed findings-from the numerator's denominator without re-labelling what survives. On the-TypeScript population the same 22 true findings now sit in 26 rather than 31,-which is 84.6% if none of the five removals took a true finding with it — they-were each read, and none did. The marker ships advisory on both: the bar below is a-property of the marker rather than of one language, and neither reading settles-it on a population this size.--Twenty-nine items cannot settle a 70% bar: 86% on that sample carries a 95%-interval of roughly 68% to 96%, whose lower bound sits below the bar it appears-to clear. The reading is consistent with clearing it and does not demonstrate-it. What moved Python is that a test's oracle now reaches it from a function it-calls in the same file, described below. That barely moved TypeScript, because a-JS/TS suite keeps its helpers in a separate module; resolving the same-delegation *across* a file is what took TypeScript from 43% to 71%, and how it-is done is in the section after next.--An earlier pass published 53% for Python. Re-labelling the pre-change findings-of the same corpus, drawn the same way and read against this pass's rubric,-gives 68.8% (22 of 32), so the gain below is that 68.8 to 86.2 and the 53% is-superseded rather than contradicted. The two passes are not reconciled: 15-points of the difference is labelling, on samples of about thirty, and the-rubric behind the earlier number was not recorded. TypeScript read 52% then and 51% on-the pass after it; re-labelled here it reads 43.1%, so the drift is on both-languages and in the same direction rather than on Python alone. That withdraws-the earlier reading of it as sample noise: what the three passes differ on is-the rubric, and the only defence against that is to state the rubric and-re-label both arms of a comparison together, which is what the section below-does.--### Resolving a test's oracle across a file boundary--A test that hands its checks to a helper is not assertion-free, and the-assertion count is per function, so the helper's checks are invisible from the-test. Resolving that within one file took Python from 68.8% to 86.2%. A JS/TS-suite keeps its helpers in another module, so the same question had to be asked-of the call graph.--It is asked in two lanes, because **a JS/TS test case is not a graph node**:-every one of them is an anonymous callback handed to `it(...)`, and on one-corpus 0 of 7,354 resolved to a symbol, against 5,212 of 5,212 for Python's-named `def test_x`. Resolving from the enclosing module node instead would let-one delegating test speak for every other test in its file, so it is not done.--- **call-edge** — the test is a symbol; walk its own outgoing call edges, depth- bounded. Its job on Python is as much to answer *no* authoritatively as to- answer yes: while it answers, the looser lane below is never consulted.-- **file-edge** — the test is a callback; require both that the *file* has a- resolved call edge to the asserting symbol and that the *function body* calls- it by name. The edge alone is file-scoped and the name alone is repo-scoped;- together they are neither.--Only edges that bind **one** definition are read, which is a harder filter than-the execution index applies. `import_merged` carries 0.85 confidence and means-only "a function of this name exists in one of the files this file imports"; it-was measured resolving `Silent().check()` to `Asserting.check` because the-receiver could not be typed. An extra edge costs a performance pass a cheap-false positive and costs this marker a hidden true finding, so the two filter-differently on purpose.--A suppression is available only where an edge is. An unresolved call is-indistinguishable from a call to something that checks nothing, so it suppresses-nothing and the marker fires — a missing edge leaves a false positive, where a-wrong one would hide a real finding.--Three consequences of that, all leaving findings in place rather than removing-them, and all properties of a repository's layout as much as of the mechanism. A-helper outside a test-named directory is not an oracle, so a suite keeping its-helpers in `src/test-utils/` gets nothing from this while one using-`src/testUtils.ts` does. A re-export is followed one hop, so a single barrel-file works and a barrel of barrels does not. And an oracle living in a-`beforeEach` is an anonymous callback with no symbol, so it is never in the sink-set at all.--Measured on the complete population of two corpora, both arms labelled in one-pass against one rubric: **51 findings at 43.1% to 31 at 71.0%, with none of the-22 true findings lost.** All 20 suppressions trace to three helpers, each read in-full, and each does check something. The sharpest evidence that the oracle is-per function rather than per file: `inspectTreeStructure` and-`testParseSourceCodeDefinitions` live in the same helper module, and only the-second asserts — the first's eleven callers all survive.--An earlier pass published 51% for TypeScript. Re-labelling that same pre-change-population against this pass's rubric gives 43.1%, so the gain is 43.1 to 71.0-and the 51% is superseded rather than contradicted; the difference is that this-pass counts a deliberate, commented "should not throw" as a false positive.--Read that gain with its provenance. Both arms were labelled by one person in one-unblinded pass, by the author of the change, and the re-label moved the baseline-**down** -- the one direction that widens the reported gain. The post figure is-the more defensible of the two, being a complete population rather than a-sample; the 28-point delta is the part carrying the labeller. A second reader on-the same population is the cheapest thing that would settle it. Note also that-22 of 31 is a 95% interval of roughly 53% to 84%, whose lower bound sits below-the 70% bar exactly as Python's 29-item reading does, so neither language-demonstrates the bar on these populations. Four families accounted for the remainder, and three have since been closed-by reading the test rather than the call graph — a wait helper that fails by-throwing, a guard test whose body is a bare `throw`, and an `expect(...)` the-statement scan declines because it is bound to a name or sits inside a nested-function. What remains is a documented no-throw, which is a labelling question-rather than a code one: every test that executes code already fails if that-code raises, so a comment saying "should not throw" names an oracle the marker-cannot distinguish from its absence. None of the four was introduced by-resolving a call.--Python moves on one corpus and not the other, and the split is the point. One-corpus suppresses nothing: its tests keep their oracles inline or in the same-file, where the existing rule already reaches them. The other goes from **501-findings to 346**, and all 155 suppressions trace to just five helper methods ---one of which, a `check_html` on a shared `SimpleTestCase` subclass, accounts for-144 by itself. Every one of the five was read in full and every one asserts, so-the verification is complete by oracle rather than sampled by finding.--That is the base-class idiom listed below as Java's blocker, showing up in-Python: the helper is inherited, so it is neither in the test's file nor named-anything the assertion lexicon recognises. It is also why the depth is **2**-rather than 1: depth 1 resolves only the single largest helper, missing eleven-of the other four's callers, and depth 3 adds four more that reach-`SimpleTestCase._assert_raises_or_warns_cm`, the framework's own `assertRaises`-machinery, which is further than a test's oracle should have to be chased.--A **mock verification counts as an assertion for this marker and not for the one-above**, which is the deliberate inversion described in CODE_HEALTH.md. It-matters only for Java: Python's `assert_called_once`, jest's-`expect(m).toHaveBeenCalled()` and Go's `t.Error` already reach the narrow tier,-so Mockito is the one dialect where verification is invisible. On a 299-test-file-Java corpus it is 1,087 verifications, and counting them takes the marker from-146 findings to 28 — 118 tests that each had a real oracle — while introducing-none.--Counting a test's oracle turned out to be where the work was. Four forms reached-no tier at all before this. Each was found by reading findings that named an-oracle the counter had not seen, and each is answered at the broad tier so that-no calibrated run moves:--| Form | Example | Why it was missed |-|---|---|---|-| Context-manager oracle | `with pytest.raises(ValueError):` | The call is in the `with` header, not a statement |-| Split call shape | `Helpers.assertRejected(...)` | Java states a call as `object` + `name` and exposes no callee node |-| Property-terminated chain | `expect(x).to.be.null` | chai ends in a property, so the statement is not a call |-| Expression-bodied lambda | `it("x", () => expect(a).toBe(b))` | The body is the expression, so there is no statement to classify |--A private helper (`_assert_no_leak(msg)`) was a fifth: the narrow prefixes anchor-at the start of a name and a leading underscore breaks the anchor.--Firing rate fell as they landed — on a 794-test-file Python corpus from 13.0% of-its test cases to 3.7%, and on a 521-test-file TypeScript corpus from 1.6% to-0.4%. A rate is not a precision measurement; what these five changes are measured-to have done is stop the counter missing an oracle that was written down.--**Go and Java are blocked, and for the same structural reason rather than a-threshold.** Both languages conventionally delegate the oracle to a helper, and-the assertion count is per function, so the helper's assertions are invisible-from the test that calls it:--- **Go** — the idiom is to hand `*testing.T` to a helper that asserts, as in- `parsertest.RunFileCases(t, ...)`. Measured on the findings themselves: **81 of- 87** on one corpus and **119 of 121** on another pass the test handle to a- call. That is not a false-positive family a marker can declare and live with,- it is nearly the whole population, and suppressing on it would leave the marker- firing a handful of times per repository.-- **Java** — the same shape without the handle, a helper on the test class that- asserts on the test's behalf. Hand-labelled at **33%** over 18 findings, and 10- of its 12 false positives were this family.--Separating them needs the assertions of a called function to reach its caller.-Within a single file that now happens. Neither language is unblocked by it: Go-hands `*testing.T` to a package-level helper and Java inherits a base-class-method, and both live in another file, so re-admitting either is a measurement-of its own.--**A called function's assertions reach its caller, inside one file.** The-assertion count is per function, so a test that hands its checks to a helper-beside it read as checking nothing. That was the largest false-positive family-in every language the marker reports on. The walk now records the names each-function calls, and the marker resolves them against same-file functions that-assert. It counts nothing and no calibrated marker reads it.--What it suppresses, by corpus:--| Corpus | Findings | Test cases | Rate |-|---|---|---|---|-| 2,012-file Python | 763 to 501 | 17,966 | 4.2% to 2.8% |-| 794-file Python | 194 to 172 | 5,212 | 3.7% to 3.3% |-| 374-file TypeScript | 35 to 33 | 7,354 | 0.5% to 0.4% |-| 562-file TypeScript | 18 to 18 | 4,761 | 0.4% to 0.4% |--Six suppressions were read by hand across the two Python corpora and each is-the declared family: a test whose whole body delegates to a helper that asserts,-or one that parametrises a sibling test by calling it. The resolution matches-names against same-file names and drops the receiver, so a call to an imported-function sharing a local helper's name, or a method on a test-local double-sharing a module-level function's name, suppresses wrongly. Both hide a finding-rather than invent one. No instance turned up in the corpora above, and the-shape is pinned by a test.--What it deliberately does not reach is anything needing the call graph, which-this pass does not consult: `super().test_x(...)`, a package-level Go helper-handed the test handle, and the shared JS/TS fixture module that is most of-TypeScript's remaining error.--**A `.tsx` file is read with the JSX grammar.** It arrives tagged `typescript`,-and the grammar that tag selects fails on the first `<Component />`; everything-after it lands in ERROR recovery. The health pass picks the grammar from the-path instead, in all three places it parses: the complexity walk, the clone-tokenizer and the dataflow CFG. A React suite is read the way a `.ts` suite-always was. What it was costing, measured on two React corpora (374 and 562-`.tsx` files):--| | 374-file corpus | 562-file corpus |-|---|---|---|-| Assertions counted on `.tsx` | 1,367 to 2,085 | 175 to 377 |-| Assertion runs on `.tsx` | 264 to 391 | 30 to 69 |-| `.tsx` files whose clone token stream changed | 365 of 374 | 550 of 562 |-| Function names over 60 chars, a proxy for recovery wreckage | 124 to 0 | 20 to 3 |-| `assertion_free_test` on `.tsx` | 15 to 0 | 0 to 1 |-| Non-`.tsx` files whose walk or token stream moved | 0 of 1,165 | 0 of 1,868 |--Every one of those 15 findings was the grammar rather than the test, which is-why the marker declined `.tsx` files until now. The second corpus's one new-finding is a test that genuinely asserts nothing, hidden by that same rule.--Two calibrated markers read the same walk, and at the time both read it-through the name-keyed `function_metrics`, which kept one row per distinct-function name and so dropped all but one of a file's anonymous `it` callbacks.-The section below removes that key; what follows here is the state as this-grammar fix left it. That key is the-thing to understand here. Recovery had been swallowing whole source spans into-each callback's *name*, and those accidentally unique names were defeating it,-so a `.tsx` file kept rows that a `.ts` file has always lost. Counting names-longer than 60 characters as a proxy for that wreckage: 124 of them on the first-corpus, 0 after. Measured as the share of walked functions that survive the key,-`.tsx` goes from 44.4% to 22.9%, against a `.ts`/`.js` control of 16.0%.--`large_assertion_block` does not move on either corpus (0 and 0, then 1 and 1).-Its floor is fifteen assertions in one run and React cases are short, but the-same name key dropped most of the newly visible runs before that floor was-ever consulted, so the floor was not the whole reason. Removing the key turns-that marker's 0 into a 6 on the second corpus below.--`duplicated_assertion_block` falls from 304 to 262 on the first corpus, and this-is a loss rather than a correction. Six of the lost findings were read by hand-and every one is a genuine run of consecutive assertions, correctly bounded; the-rows were real and the clone partner was real. What was not real is the-mechanism that kept them visible, which was a mangled name defeating a lossy-key. So at this point `.tsx` files under-reported duplicated assertion blocks-exactly as `.ts` files always had, which the next section then fixes for both. The corpus carries more real assertion runs after this-change, 264 to 391, and the marker reports fewer of them. Re-keying-`function_metrics` is the fix for that, and the next section is that change.--Svelte and Vue stay `later` rather than following TypeScript: an SFC is walked as-a TypeScript buffer, but a single-file component is essentially never a test-file, so the row would be untestable rather than useful.--The marker is advisory and never deducts. See-[CODE_HEALTH.md](CODE_HEALTH.md).--### A marker reads the whole walked file, not one row per name--Eleven markers used to read a `function_metrics` map keyed by function name. A-map keyed that way keeps one row per distinct name, and a function name is not-unique inside a file. Every `it(...)` callback in a spec file walks under the name-`it callback`; `__init__` walks under one name on each class in a module. Every-row after the first was dropped before a marker ever saw it. They read-`all_functions` now, the whole walked list, which is what the two advisory-test-quality markers already read for this reason.--How much a file loses to that key is a property of how a language names its-functions rather than of any marker. Measured across the walked functions of-four corpora:--| Corpus | Language | Walked functions | Survive the name key |-|---|---|---|---|-| 1,336 files | TypeScript / TSX | 12,257 | 37.8% |-| 2,194 files | TypeScript / TSX | 17,261 | 72.3% |-| 1,949 files | Python | 12,926 | 91.4% |-| 1,714 files | Python | 30,747 | 87.9% |--These count every walked file, so they do not line up with the test-file-only-`.tsx` figures above and are not meant to.--A JS/TS spec suite is the worst case by a wide margin, because its test bodies-are anonymous callbacks that all walk under one name. Python loses about a tenth-of its rows, mostly same-named methods on neighbouring classes.--What each scoring marker sees, before the change and after. Note that the two-assertion-block markers score but were never fitted on a defect corpus, and they-are the two that move furthest:--| Marker | TS/TSX (1,336 files) | TS/TSX (2,194 files) | Python (1,949 files) | Python (1,714 files) |-|---|---|---|---|---|-| `complex_method` | 496 to 506 | 1,005 to 1,015 | 443 to 468 | 483 to 537 |-| `large_method` | 347 to 382 | 608 to 649 | 192 to 203 | 157 to 168 |-| `nested_complexity` | 232 to 233 | 338 to 341 | 236 to 253 | 286 to 317 |-| `bumpy_road` | 80 to 80 | 161 to 162 | 65 to 71 | 99 to 108 |-| `complex_conditional` | 92 to 96 | 163 to 168 | 67 to 69 | 91 to 101 |-| `primitive_obsession` | 54 to 54 | 106 to 107 | 685 to 859 | 449 to 599 |-| `large_assertion_block` | 0 to 0 | 0 to 6 | 0 to 0 | 11 to 11 |-| `duplicated_assertion_block` | 262 to 2,831 | 175 to 1,711 | 1,098 to 1,108 | 2,360 to 2,477 |--Three markers are missing from that table because they need inputs this-measurement does not build: `brain_method` needs graph centrality, and-`code_age_volatility` and `function_hotspot` need git blame. They read the same-list as the other eight, which is visible in the diff, but how far they move was-not observed.--Most of those cells are single-digit percentages. `complex_method`,-`large_method`, `nested_complexity` and `complex_conditional` each clear ten-percent on one corpus and stay below it on the other three. Two markers move-much further, and they are the two worth reading carefully.--`primitive_obsession` gains a quarter to a third of its findings on both Python-corpora, 685 to 859 and 449 to 599, and almost nothing on either JS/TS one.-Python modules define the same method name on several classes each, and the-row that survived the key was the first of them in the file, so a wide signature-on any of the others was invisible.--`duplicated_assertion_block` is the large one: it rises about tenfold on both-JS/TS corpora and barely moves on either Python one. That is the same fact as-the survival table, seen from the other end.--Six of the gained rows were read in the source, and each of those six is a-correctly bounded run of consecutive assertions. Six labels cannot speak for-2,569 rows, so that is evidence the marker is bounding blocks correctly, not a-precision figure. What does speak for the population is its shape. On the first-corpus the two arms' assertion-run-length histograms have closely matching-profiles while the new arm is about eleven times larger, which is what uniform-thinning looks like and not what a filter that selected for anything would look-like. On that same corpus the marker goes from a median of one finding per test-file to five, with one file at sixty. Neither measurement says whether a gained-row is worth showing a reader; only counts and block lengths were measured.--No floor was added to hold that number down. The name key was an accidental-limiter and replacing it with a deliberate one is a calibration change, which-needs its own defect-corpus evidence rather than a number that looks-comfortable. What bounds the damage today is the `test_quality` category cap of-0.5, which a single medium finding already exceeds: a test file with sixty of-these loses the same half point as a test file with one, so the scores move far-less than the counts do. Whether the finding list itself wants a cap is a-question for the marker, not for the container it reads.--The six findings `large_assertion_block` gains on the second JS/TS corpus were-also read by hand. All six are runs of fifteen or more consecutive assertions in-a single test, which is exactly what that marker declares it looks for, and all-six are anonymous callbacks, so each sat in a file where another function of the-same name had been holding the only row.--Per-marker mechanics, every per-language precision ceiling and the reasoning-behind each `n/a`: [CODE_HEALTH.md](CODE_HEALTH.md) and-[architecture/language-support.md](../architecture/language-support.md).+✅ supported, "no" not available yet, n/a the language cannot carry that metric.+An unmapped language produces no findings, never guessed ones. The two+test-quality markers (mock saturation, assertion-free test) are advisory and+never deduct from a score. Why Go and Java are held back from them, and why+Kotlin's Extract Method waits on its grammar, is in+[architecture/language-support.md](../architecture/language-support.md#test-quality-markers-per-language).+Per-marker detail is in [CODE_HEALTH.md](CODE_HEALTH.md).---## Known ceilings-Stated plainly, because a ceiling you know about is worth more than a claim you-cannot check.-- **Template dialects are invisible.** Django/Jinja, Go templates, ERB,Handlebars, Blade and Thymeleaf parse as HTML and yield nothing.- **Svelte and Vue binding forms are skipped.** `{#each x as y}` and- `v-for="x in xs"` *parse* as JS but mean something else, so they are skipped: a- parse that succeeds with the wrong meaning is worse than a skip. Component- props are set by the parent as markup attributes and never imported by name, so- the unused-export pass is suppressed for both.-- **Object Pascal type kinds collapse.** `record` / `interface` / class-helper /- enum / type alias all report as `kind="class"`, and its `extends` vs- `implements` split is inferred from the `I`-prefix naming convention rather- than guaranteed by the language.-- **COBOL support is intentionally COBOL-85 shaped.** Literal `CALL` and- `PERFORM` resolve, and `.cpy` files are parsed when indexed directly, but- `COPY` does not create an edge. Dynamic program names, dialect extensions and- fixed/free-format preprocessing beyond what the grammar accepts stay silent.-- **Razor has no import edges**, and an attribute-bound handler carries none.-- **Object Pascal's `extends`/`implements` split is a naming heuristic**,- inferred from the `I`-prefix convention rather than a language guarantee.-- **Object Pascal's DB/network performance sinks are gated on file-wide- `uses`-clause evidence, not per-receiver evidence.** A file importing- `FireDAC` gates every `.Open` / `.ExecSQL` / `.Post` call in it, not only- calls on an actual `TFDQuery`, because a Pascal variable's declared type has- no textual link back to the unit it came from the way `client = requests.Session()`- does in Python.-- **Object Pascal has no `assertion_free_test`.** `large_assertion_block` /- `duplicated_assertion_block` / `mock_saturated_test` all read the same- assertion counts and work for it; `assertion_free_test` additionally gates- on a `SHIPPING_LANGUAGES` allowlist that needs its own measured-precision- pass (see the module docstring) before Pascal joins it. It has no mock- dialect either, so `mock_saturated_test`'s mock-setup half stays at zero --- advisory and harmless, never a false `mock_saturated_test` positive.-- **A GDScript `uid://` resolves through the `.uid` sidecar Godot writes for- scripts**, so a `preload` naming one reaches its file. A uid naming a scene- does not: Godot writes no sidecar for `.tscn` / `.tres`, and the- `[ext_resource]` header carrying such a uid also carries a `path=`, which is- what resolves. String dispatch is still unresolved, and a script- without `class_name` gets no class symbol.-- **A Godot `addons/` tree is exempt from dead-code reporting** only when a- `project.godot` sits above it, so a plugin's own repo reports normally.+ `v-for="x in xs"` are not resolved, and the unused-export pass is off for+ both, since props are set by the parent as markup attributes.+- **Object Pascal type kinds collapse** to `kind="class"`, and its `extends`+ vs `implements` split follows the `I`-prefix naming convention.+- **COBOL is COBOL-85 shaped.** `COPY` does not create an edge; dynamic program+ names and dialect extensions stay silent.+- **Razor has no import edges.**+- **GDScript string dispatch is unresolved**, and a script without `class_name`+ gets no class symbol.- **VB.NET leaves a few constructs unparsed**, XML and tuple literals among them.-- **C has no health dialect.** It shares the C++ grammar for parsing but reaches- no health walker map, so it gets graph coverage without markers.+- **C has complexity markers only.** It has no performance or Extract Method+ support.- **Scala import resolution is partial**, and ScalaTest's infix DSL(`x shouldBe y`) is not counted as an assertion.-Per-language mechanics behind these:-[architecture/language-support.md](../architecture/language-support.md).+Mechanics behind these:+[architecture/language-support.md](../architecture/language-support.md#per-language-mechanics).---## Roadmap-| Language | Target | Next |-|----------|--------|------|-| Vue | Full | Options-API `pair`-function members, `v-for` head bindings |-| Svelte | Full | `{#each}` head bindings, object-literal attributes, `.svelte.ts` rune modules |-| Kotlin | Full (health) | Dataflow is **blocked on the grammar** and needs a grammar upgrade or a text-based jump seam |-| Scala | Full (health) | Dataflow dialect, combinator (`.map` / `.foreach`) loop tracking |-| Ruby | Full (health) | Dataflow dialect, LCOM4 via `@ivar` grouping |-| C# | Full (health) | Dataflow dialect |-| Dart | Good | riverpod / get_it dynamic hints, dataflow dialect |-| GDScript | Good | The health dialects (complexity, performance, dataflow) that would take it to Full; the grammar supports all three |-| Object Pascal | Good | A dedicated `uses` resolver; `assertion_free_test` needs a measured-precision pass before it joins its language allowlist |-| COBOL | Good | Copybook resolution, source-format normalization, dialect coverage, health markers |-| VB.NET | Good | Health markers, project-level `<Import Include=...>` as implicit imports |-| Elixir | Good | Health markers, and a call-resolution strategy beyond same-file |-| F# | Good | The health markers beyond complexity, and a resolver that reads the AST index instead of the declared-name regex |-| Objective-C | Good | The health markers beyond complexity, a resolver that reads the Xcode project rather than file stems, and pairing a header with its implementation across files |-| SQL / dbt | — | Column-level blast radius |-| Shell | — | Shebang detection for extensionless executables |-| HTML | Lightweight | A regex import tier for template dialects, gated on a framework manifest |+| Language | Next |+|----------|------|+| Vue · Svelte | Template binding forms (`v-for`, `{#each}`) |+| Kotlin | Extract Method, which needs a grammar that labels control flow |+| Scala · Ruby · C# | Extract Method (dataflow dialect) |+| GDScript | Health markers (complexity, performance, dataflow) |+| Object Pascal | A dedicated `uses` resolver |+| COBOL | Copybook resolution, dialect coverage, health markers |+| VB.NET · Elixir | Health markers |+| F# · Objective-C | Health markers beyond complexity |+| SQL / dbt | Column-level blast radius |+| Shell | Shebang detection for extensionless executables |++---++## Requesting or adding a language++To ask for a language, open an issue on+[GitHub](https://github.com/repowise-dev/repowise/issues) with a public repo+that uses it. To add one yourself, the contributor recipe (a `LanguageSpec`, a+tag, a tree-sitter query, a config entry and the grammar dependency) is in+[architecture/language-support.md](../architecture/language-support.md#adding-a-new-language),+and the general contribution flow is in+[CONTRIBUTING.md](../../.github/CONTRIBUTING.md). A new language starts on the+Good or Partial rung; health markers come from separate registries described in+the same architecture page.---## See also-- **[GRAPH.md](GRAPH.md)** · the graph itself: edge vocabulary, resolution origins, execution flows-- **[architecture/language-support.md](../architecture/language-support.md)** · pipeline internals, the call-resolution architecture, and the contributor recipe for adding a language-- [CODE_HEALTH.md](CODE_HEALTH.md) · health markers and per-language precision-- [WORKSPACES.md](../scale/WORKSPACES.md) · cross-repo contracts and co-change+- [GRAPH.md](GRAPH.md) · edge vocabulary, resolution origins, execution flows+- [architecture/language-support.md](../architecture/language-support.md) · pipeline internals and per-language mechanics+- [CODE_HEALTH.md](CODE_HEALTH.md) · health markers+- [BENCHMARKS.md](../BENCHMARKS.md#accuracy-by-language) · measured accuracy per language+- [WORKSPACES.md](../scale/WORKSPACES.md) · cross-repo contractsdiff --git a/docs/layers/OWNERSHIP.md b/docs/layers/OWNERSHIP.mdnew file mode 100644index 000000000..cf8df5487--- /dev/null+++ b/docs/layers/OWNERSHIP.md@@ -0,0 +1,209 @@+# Ownership and Knowledge Risk++Repowise reads git history to tell you who owns each file and module, how many+people it would take to lose the knowledge behind a file (its bus factor), and+where that knowledge is already at risk: a file still being edited whose main+author has gone quiet, a file with many drive-by authors and no clear owner, or+a busy file with too many hands on it. It also suggests reviewers for a set of+changed files.++It needs no LLM key and makes no network calls. Everything is computed from+`git log` (and `git blame` in the standard index) during `repowise init` and+`repowise update`. It does not read code review data, chat, or issue trackers,+so "owner" means "wrote the most commits or lines", not "is responsible for".++This layer needs git history. In a directory that is not a git repository the+ownership fields stay empty and the detectors below stay silent. A shallow clone+only sees the commits it fetched, so owners and bus factors describe that window;+run `git fetch --unshallow` before indexing for the full picture.++## Quick start++```bash+repowise init # ownership is computed during indexing+repowise context src/auth/session.py --include ownership+repowise risk --target src/auth/session.py # owner, recent owner, bus factor+repowise health --file src/auth/session.py # knowledge_loss and friends, if they fire+```++From an agent:++```text+get_context(targets=["src/auth/session.py"], include=["ownership"])+get_risk(targets=["src/auth/session.py"])+get_risk(changed_files=["src/api/routes.ts", "src/middleware/cors.ts"])+get_overview(include=["ownership"])+```++In the dashboard, open **Owners** for the contributor directory and per-person+profiles, and the **Impact** tab of **Code Health** for reviewer suggestions on a+set of files.++Sample `get_context` ownership block:++```json+"ownership": {+ "primary_owner": "Dana Lee",+ "owner_pct": 0.71,+ "owner_line_pct": 0.83,+ "contributor_count": 4,+ "bus_factor": 1,+ "recent_owner": "Sam Ortiz",+ "recent_owner_pct": 0.6+}+```++## Reading the results++### Per-file fields++| Field | Meaning |+|---|---|+| `primary_owner` | The file's owner. In the standard index this is the author of the most surviving lines by `git blame`; in a fast index (`--mode fast`, no blame) it is the author of the most commits |+| `owner_pct` | The primary owner's share of the file's commits (0-1). Can be empty when the blame owner made none of the indexed commits |+| `owner_line_pct` | The primary owner's share of current lines by blame (0-1). Empty without blame |+| `recent_owner`, `recent_owner_pct` | Who made the most commits to the file in the last 90 days, and their share of those commits. `get_context` only shows these when the recent owner differs from the primary owner |+| `contributor_count` | Distinct authors in the indexed history of the file |+| `bus_factor` | How many of the top authors it takes to cover 80% of the file's commits. 1 means one person wrote at least 80% of them. 0 means no history was indexed for the file, not "no risk" |++Commits that only move a file do not count as authoring it. Author names and+emails go through the repository's `.mailmap`, and GitHub `noreply` addresses+for one login fold into one person, so a contributor who committed from several+addresses is usually counted once. Each file keeps its top 50 authors.++### Module ownership and silos++Per module (top-level directory), the owner is the person who is primary owner of+the most files in it. A module is a **silo** when one person owns more than 80%+of its files. Per file, the dashboard and REST API mark a silo when the primary+owner holds more than 80% of the file's commits.++### Detectors++Three code-health detectors turn these fields into findings. They run with the+rest of [code health](CODE_HEALTH.md) and appear in its findings list.++| Detector | Fires when | Severity |+|---|---|---|+| `knowledge_loss` | The file is still changing (a commit in 90 days, or a hotspot; never a file marked stable), its bus factor is 1, and the primary owner is no longer the recent owner or the recent owner makes under 20% of recent commits | High on a hotspot; medium when both conditions hold; otherwise low |+| `ownership_risk` | The file has 5 or more commits and either 3 or more minor contributors (each under 5% of commits) or no author above 40% | Critical at 6+ minor contributors on a hotspot; high at 5+, or 3+ on a hotspot; medium at 3+; otherwise low |+| `developer_congestion` | 5 or more contributors, the file is in the top quarter of churn for this repository, and nobody holds half the commits | High at 10+ contributors and 20+ commits in 90 days; medium at 7+ contributors or 12+ commits; otherwise low |++A stable file whose author left never fires `knowledge_loss`: code nobody edits+is low risk, and the finding only matters while people are editing around the+lost author's intent.++On a small team (3 or fewer active contributors in 90 days), one owner per file+is the normal way of working. `knowledge_loss` and `ownership_risk` still report+the finding there, but cap it at low severity unless the file is a hotspot, and+add "informational: small team" to the reason.++### Repository-level numbers++| Number | Where | Meaning |+|---|---|---|+| Truck factor | Stats page | Fewest primary owners who together own more than half the owned files. Bot authors are left out |+| Single-owner files | Stats page | Files with bus factor 1 |+| Silo count | Stats page | Modules where one person owns more than 80% of files |+| `git_health.avg_bus_factor`, `git_health.files_with_bus_factor_1` | `get_overview` | Average bus factor and count of bus-factor-1 files |+| `knowledge_map.top_owners` | `get_overview(include=["ownership"])` | The three people owning the most files, by display name only |++### Contributor profiles++Each person in **Owners** has a profile: files and hotspots they own, modules+where they hold more than 80% of files, files they own with bus factor 1, dead+code in files they own, commits in the 90 days before the newest commit, and+the people they share files with most.++### Reviewer suggestions++Two surfaces suggest reviewers for a set of changed files, and they score+differently:++- **Dashboard and REST** (`/api/repos/<id>/reviewer-suggestions`): each author of+ a touched file scores their share of its commits, plus a recency term from the+ file's 90-day activity. Authors of each touched file's five strongest+ co-change partners score at half weight. Each suggestion lists the files it+ owns and the co-change files behind it.+- **`get_risk(changed_files=...)`**: `pr_blast_radius.recommended_reviewers` lists+ up to five primary owners of the affected files, ranked by how many of those+ files they own, then by average ownership share.++## Tuning and suppressing++The detectors use the code-health rules file, `.repowise/health-rules.json`:++```json+{+ "profile": "small-team",+ "disabled_biomarkers": ["developer_congestion"],+ "rules": [+ { "path": "vendor/**", "disabled_biomarkers": ["ownership_risk", "knowledge_loss"] }+ ]+}+```++- The `small-team` profile lowers all three detectors (and a few process+ signals) to low severity.+- `severity_overrides` sets a detector's label; an explicit entry wins over the+ profile.+- History depth: `repowise init --commit-limit N` sets how many commits per file+ are read (default 500, maximum 10000). `--follow-renames` follows a file+ across renames, so ownership survives a move. Both are saved to config.+- `--mode fast` skips blame. Owners then come from commit counts alone and+ `owner_line_pct` stays empty.++Full schema: [CONFIG.md](../reference/CONFIG.md#the-health-rulesjson-file).++## Accuracy and limits++- Ownership is authorship. Someone who reviews every change to a file but never+ commits to it does not appear.+- Squash merges credit the whole change to whoever is recorded as the commit+ author. Pair-programming credit via `Co-authored-by` trailers is not counted+ toward ownership.+- Identity merging is conservative: beyond `.mailmap` and GitHub `noreply`+ folding, two addresses are only merged on guarded evidence, so some people may+ still appear twice. A false merge is treated as worse than a split.+- In the health score's defect calibration, `ownership_risk` showed positive+ lift and carries a higher weight than the default. `developer_congestion` and+ `knowledge_loss` showed weak signal and deduct less than the default; they are+ kept as findings a team can act on. All three share the change-history+ category cap. Method: [BENCHMARKS.md](../BENCHMARKS.md).+- A contributor profile's 90-day commit count is measured back from the newest+ commit, and stays empty when the stored commits do not reach back that far.+- Truncated history (shallow clone, low `--commit-limit`) can make an early+ author vanish, which can both hide and invent a `knowledge_loss` finding.++## Where it shows up++| Surface | What you get |+|---|---|+| CLI | `repowise context --include ownership`, `repowise risk --target` (owner, recent owner, bus factor), `repowise health` findings |+| MCP | `get_context` `ownership` block, `get_risk` owner fields and PR-mode reviewers, `get_overview` ownership summary, `get_health` findings |+| Dashboard | **Owners** directory and profiles, **Stats** people numbers, **Code Health** findings and **Impact** reviewer suggestions |+| Editor | VS Code: file hover shows the owner and their commit share; the **Hotspots & Ownership** view lists module owners and flags silos, with bus factor in each hotspot's tooltip |++## Reference++| Command | Flag | Effect |+|---|---|---|+| `repowise context` | `--include ownership` | Add the ownership block to each target's card |+| `repowise risk` | `--target PATH` | Score a file's history, including owners and bus factor |+| `repowise risk` | `--changed-file PATH` | With `--target`: PR mode; `--format json` includes `recommended_reviewers` |+| `repowise health` | `--file PATH` | One file's findings, including the ownership detectors |+| `repowise init` | `--commit-limit N` | Commits read per file (default 500, max 10000) |+| `repowise init` | `--follow-renames` | Track files across renames |+| `repowise init` | `--mode fast` | Skip blame; commit-count ownership only |++MCP: [get_context](../agent/MCP_TOOLS.md#get_context),+[get_risk](../agent/MCP_TOOLS.md#get_risk),+[get_overview](../agent/MCP_TOOLS.md#get_overview).++## See also++- [Code health](CODE_HEALTH.md): how the ownership detectors feed the score.+- [Change risk](CHANGE_RISK.md): co-change pairs and reviewing a diff.+- [Graph](GRAPH.md): hidden coupling between files that change together.+- [Bug history](BUG_HISTORY.md): which files keep needing fixes.+- [Intelligence layers](INTELLIGENCE_LAYERS.md): how this layer fits with the others.diff --git a/docs/layers/REFACTORING.md b/docs/layers/REFACTORING.mdindex f83eebd2c..5f58944a9 100644--- a/docs/layers/REFACTORING.md+++ b/docs/layers/REFACTORING.md@@ -1,305 +1,192 @@-# Refactoring intelligence (`repowise health --refactoring-targets`)--A health score tells you a file is in trouble. **Refactoring intelligence names-the specific fix.** Every other tool stops at the score, or prints the same-static sentence for every god class in every repo. repowise emits one structured-plan per opportunity: *split `GraphBuilder` into these three cohesive groups*,-*move `resolve_call` to the `resolvers` class where its calls actually land*,-*break the `pipeline ↔ update` import cycle by inverting this one edge*,-*decompose this 900-line module into these four files and rewrite the imports in-the six that depend on it*, computed deterministically from the same graph,-class model, and git data the score is built on.+# Refactoring intelligence-```bash-repowise health --refactoring-targets # ranked plans, biggest win for least effort-repowise health --refactoring-targets --format json-```+A health score tells you a file is in trouble. Refactoring intelligence names the+fix: split this class into these groups, move this method to the class its calls+land in, cut this one import edge to break a cycle, lift these lines into a helper+with this signature. Each plan is structured data computed from the same graph,+class model and git history the health score uses.-It runs **inside the health pass** (`init` / `update`), reusing data already-computed with no re-parse, **no LLM, no network**, inside the same <30s budget. The-LLM layer (code generation) is a separate, strictly opt-in step ([below](#optional-code-generation)).+Plans are built during the health pass of `repowise init` and `repowise update`.+They need no LLM key, no network and no re-parse. Turning a plan into actual code+is a separate, on-request step that does use an LLM+([Code generation](#code-generation)).<div align="center"><img src="../../.github/assets/health-loop.svg" alt="repowise code-health loop: markers fan into three signals, the graph and git history locate risk, and refactoring intelligence emits concrete plans an agent executes" width="100%" /></div>-## The seven detectors--Each detector is a self-contained module registered into a registry (adding a-refactoring type is a new file + a registry entry, like the marker registry).-A detector degrades to **"no suggestion" on any missing signal, never a wrong-one**, and produces stable-sorted, deterministic output.--| Type | What it names | Detection (deterministic) |-|------|---------------|---------------------------|-| **Extract Class** | The cohesion groups an incohesive / god class should split into: the exact methods + fields per group. | LCOM4 union-find components (each disconnected component is a candidate class), with the god-class shape confirmed via Lanza-Marinescu (WMC = Σ McCabe, TCC). |-| **Extract Helper** | A clone's exact occurrences and where the shared helper belongs. | Rabin–Karp clone pairs (line ranges, token count, co-change). The extraction site is the community centroid of the involved files; transitive clones (A↔B, B↔C) are clustered into one suggestion, not pairwise nags. |-| **Move Method** | A feature-envy method and the class it actually belongs to. | The method's entity set (fields/methods it touches, class-qualified) is built from the call graph; Jaccard distance to each class. Fires only when a foreign class is clearly nearer than its own and draws more of the method's calls than its home file does. A class the method instantiates (a result it builds) and an ancestor of its own class are never targets. |-| **Break Cycle** | The minimal set of import edges to invert to break a dependency cycle. | A strongly-connected component in the import graph → greedy minimum feedback arc set (MFAS) over the real edges picks the smallest cut. |-| **Split File** | The cohesive files an oversized module should decompose into: which top-level symbols move to each new file, plus the import edits in every dependent. | Community detection (Leiden, Louvain fallback) over a weighted intra-file symbol graph (direct calls, shared local helpers, shared foreign modules); emits only when the partition's **modularity** clears a decomposability gate. The file-level analog of Extract Class. |-| **Extract Method** | The exact line span to lift out of an oversized/complex method, with the helper's inferred signature: the parameters it needs (IN) and the value it must return (OUT). | Intra-procedural dataflow (CFG + def/use + reaching definitions) over methods a `large_method` / `brain_method` / `complex_method` finding already flagged. Only single-exit, statement-boundary spans that remove real complexity qualify, and IN/OUT comes from liveness over the def/use facts. A span is offered only when the gate can prove it behavior-preserving: every returned value must be written on every path through the span, and a span nested in a loop must carry no state between iterations and must not mutate what the loop iterates. What the gate cannot prove is suppressed, not demoted, so the layer under-reports rather than suggest a rewrite that changes behavior. A span must also be worth doing: at least 5 code lines that remove at least 2 decision points (or, for `large_method`, at least 12 code lines), and a helper that would not carry the finding at the function's own severity, since lifting nearly the whole body moves the smell instead of splitting it. When the best span misses that floor the next-best one that clears it is offered. A JSX component whose decision points sit mostly in its markup (conditional spreads, `&&` or ternaries in attributes and children, template ternaries) gets no extraction: that branching is prop plumbing. |-| **Performance Fix** | A safe shared intervention, affected call sites, and caller-to-sink paths for one causal performance opportunity. | The call-graph opportunity service groups raw findings by stable cause, boundary and context, then emits only supported strategies. Findings with no coherent safe intervention remain visible in Code Health without a plan. |--The algorithms are derived from public academic literature (Fokaefs-Tsantalis-HAC for class splitting, Bavota feature-envy distance, MFAS for cycle breaking,-Newman-Girvan modularity for module decomposition), not from any product.--**Split File is the cross-file wedge made concrete.** It is language-agnostic:-it reads only the already-built graph (`defines` / `calls` edges), so it works-the same on every language with call resolution, and it covers the gap LCOM4-leaves for Go (top-level functions, not class methods). Splitting Go files in the-same package is near-zero blast radius (no import edits); Python/TS get a-back-compat re-export shim, surfaced as `shim_required` on the plan.--## Anatomy of a suggestion--Every suggestion is **structured data, not a string**: the structure is the-source of truth; human-readable text is rendered only at the edges (CLI / MCP /-web).--| Field | Meaning |-|-------|---------|-| `refactoring_type` | `extract_class` \| `extract_helper` \| `move_method` \| `break_cycle` \| `split_file` \| `extract_method` \| `performance_fix` |-| `file_path`, `target_symbol`, `line_start`, `line_end` | What the refactoring acts on. |-| `plan` | The concrete, type-specific plan: the split `groups` (methods + fields), the move `{method, from_class, to_class}`, the clone `occurrences` + `suggested_site`, the cycle + `cut_edges`, the file-split `groups` (`{name, symbols, suggested_file}`) + `residual` core + `shim_required`, or the method-extraction `span` + `params` + `returns` + `suggested_name`. Names are omitted rather than invented: `suggested_name` and a group's `name` / `suggested_file` are `null` whenever no fact anchors a name, and `suggested_site.directory` is `null` unless an occurrence actually lives in that directory. |-| `evidence` | The signals that justify it: `lcom4`, `wmc`, clone token/line counts + `co_change_count`, Jaccard distances, cycle size, or the split's `modularity` + `symbol_count` + `group_count` + intra/cut edge counts. |-| `impact_delta` | The defect-health score the refactoring would recover; `0` for graph-native and performance plans. Their canonical `benefit` comes from detector-native evidence instead. Credit is proportional to what the plan removes, never the whole finding for crossing a threshold. Extract Method credits the share of what the finding measures that moves into the helper: `ccn_removed / ccn` for `complex_method`, `slice_nloc / nloc` for `large_method`, the larger of the two for `brain_method`, capped at 1. Both line counts are code lines (blank, comment and docstring lines excluded, the walker's NLOC rule), and `evidence.slice_nloc` reports the same count. Extract Class credits a `low_cohesion` finding whole (the plan is the full LCOM4 partition) and a `god_class` finding by the NLOC or CCN share that leaves the largest group. Extract Helper credits the block's share of the file's duplicated lines. |-| `effort_bucket` | `S` \| `M` \| `L` \| `XL`, from the target's size. |-| `blast_radius` | What else must move: the callers, co-change partners, and importing files. Extract Method carries `{"scope": "local"}` instead — extraction adds a private helper and changes no signature, so nothing outside the file moves and there is no count to make. |-| `confidence` | `low` \| `medium` \| `high` (drives the `min_confidence` surface gate). For Extract Method, `high` means the step is mechanical and worth handing off: it removes at least 2 decision points, takes at most 4 parameters, and moves at least a tenth of what its finding measures. Only `high` Extract Method steps classify as mechanical. |-| `source_biomarker` | The finding this answers (e.g. `low_cohesion`, `god_class`, `dry_violation`). |--The per-type `plan` / `evidence` / `blast_radius` shapes are documented in full in-`packages/core/src/repowise/core/analysis/health/refactoring/models.py`.--## Ranking: graph-aware, not churn-only--Each detector sorts its own output, but the surfaces show one mixed list, so the-**global** order is what matters. The shared recommendation service owns four-named components and one canonical score:+## Plan types-```-score = (1 + benefit) * (1 + leverage) / (1 + cost + risk)-```+| Type | What the plan names |+|------|---------------------|+| **Extract Class** | The cohesive groups (methods and fields) an incohesive or god class splits into. |+| **Extract Method** | A line span to lift out of a long or complex function, with the helper's parameters and return value. |+| **Extract Helper** | Every occurrence of a duplicated block and where the shared helper belongs. |+| **Move Method** | A method that uses another class more than its own, and the class it belongs in. |+| **Break Cycle** | The smallest set of import edges to invert to break a dependency cycle. |+| **Split File** | The files an oversized module splits into, which symbols go where, and the import edits in each dependent. |+| **Performance Fix** | One shared change for a performance problem, with the affected call sites and caller-to-sink paths. |-`benefit` is recoverable health (an opportunity's `recoverable_health` is the sum-of its credited steps' proportional `impact_delta`, each finding counted once) or-detector-native structural/performance gain;-`leverage` is weighted health deficit, dependents, and reliable entry reach;-`cost` includes effort and change surface; `risk` includes blast radius,-confidence, provenance, and validation quality. Larger blast radius therefore-cannot improve rank by masquerading as benefit. Performance plans keep their-detector-native benefit even with zero health impact. Ties break deterministically.--> **The wedge.** The leading commercial code-health tool ranks refactoring-> targets by **churn alone**, generates code **within-function only**, and ignores-> its own coupling signal at generation time. repowise ranks by graph centrality,-> works **across files** (class splits, method moves, cycle breaks), and feeds the-> co-change + graph context straight into the plan.--## Opportunities: the unit a surface serves--A **plan** is one detector output. An **opportunity** is one file's composed-refactoring: the diagnosis it leads with, its plans ordered so that applying-them in sequence is safe, and the clone groups that support the diagnosis-without instructing a change. The plan stays the addressable atom; a step names-one by `plan_id`.--Composition is deterministic and pure-(`analysis/health/refactoring/opportunity.py`), and it runs at **index time**:-the finalizer writes `refactoring_opportunities` and a one-row-`refactoring_summaries` headline in the same transaction that reconciles the-plans. Folding at read time was measured at 91 ms over 2,283 plans and 787 ms-at ten times that, plus 1,118 ms to rebuild the validation profiles, so every-page cost the whole repository.--| Field | Meaning |-|---|---|-| `opportunity_id` | `refop<model>_<digest>` over the member plan ids. Evidence is deliberately outside the kernel: a clone appearing or vanishing must not rename work an agent holds an id for. |-| `lead_biomarker` | The file's dominant finding, from the one owner of that rule (`analysis/health/models.py::primary_finding`). |-| `addresses_primary_problem` | Tri-state. `null` means no dominant finding was recorded, which is not the same claim as `false`. On the dogfood index it reads 201 true / 380 false / 3 unknown, and the false answers are the honest ones. |-| `steps[].applicability` | `mechanical` or `judgment`, with the reasons and the facts behind it, and the facts it could not establish under `unknowns`. Extraction is the only mechanical class. |-| `steps[].relocated_by` | An earlier step that moves this step's symbol to another file. The step's own `file_path` and span describe where the symbol *was*: locate it again before applying. Any surface that renders an ordered step list must say so. |-| `status` | Rolled up from the member plans' triage. Resolved only when every step is; one step marked a false positive does not resolve the rest. An `acknowledged` step is still outstanding work, so it keeps composing. |-| `steps[].finding_ids` | The findings this step's own target answers (its source biomarker's findings on the same symbol and span, never every finding of that biomarker in the file), so a step round-trips to its diagnosis through `get_health(finding_id=...)`. The key is **omitted entirely** when no finding on the file is addressable by id (a store written before findings carried public ids); an empty list means "none for this cause", which is a different answer. `lead_finding_ids` on the opportunity follows the same rule. |-| `steps[].validation_profile_id` | Points into the opportunity's `validation_profiles`. Resolved once at index time; the test-reachability walk behind it is never repeated per request. |--`performance_fix` plans are excluded by construction: the performance layer-composes, ranks and owns the lifecycle of its own opportunities.--Plans whose `source_biomarker` the finding registry-(`analysis/finding_registry.py`) withholds are left out too, as steps and as-evidence, and so are they from the plan lists, the attention list and the-`get_health` directive. A withheld finding never becomes a file's-`lead_biomarker`. The plan rows stay stored and addressable by id.--**Split File ids across `init` and `update`.** A full index, an incremental-update and a re-score from stored git metadata now build the same symbol graph-for a file at one commit, so its `split_file` plan id is the same on every path:-the co-change edge reads each function's commit set (its 50 most recent-distinct commits) from `git_function_blame` when no blame index is at hand.-A re-score blames, once, the candidate files whose rows predate that column.-This moved Split File groups, so refactoring model 4 re-mints every plan id-once on upgrade; a held `refac3_` id reports `stale_model`.--### Ordering, and `refactoring_view`--The ranked head is honestly flat. On the dogfood index eight of the top ten-score identically: single high-confidence extractions recovering the same-quantised `complex_method` deduction, separated only by file path. They really-are equal under the published factors, so the answer is a queue that spends its-first rows on distinct problems rather than a tiebreaker invented to hide the-tie.--| `refactoring_view` | Order |-|---|---|-| `diversified` *(default)* | Rank order round-robined over (lead biomarker, lead refactoring type, containing area). Falls back to plain rank order when a repository has one cause in one area. |-| `canonical` | The published rank order verbatim, ties and all. What the old default produced. |-| `file_spread` | Asked for one row per file. An opportunity *is* one file's work, so the spread is satisfied by construction; the value resolves onto the diversified order, which is what it was reaching for. |--Test files (tests and test support, by `test_paths.is_test_related_path`) rank-after every production file, in both views and in the plan list, and never lead-the `refactoring_directive`. When only test files have open work the directive-is `clear` with reason `only_test_file_opportunities`.--Both older values keep working. The same parameter also selects the legacy-`refactoring_plans` list's view, where `diversified` resolves to that list's-historical `canonical` default.--## Fix first, and the default scope--Fix first ([CODE_HEALTH.md](CODE_HEALTH.md#fix-first)) ranks refactorings beside-performance fixes and code-shape findings. A refactoring opportunity is a Fix-first-candidate when its file ships, it recovers at least 0.5 health, its lead step is a-kind raters found worth doing, and that step names a concrete edit (lines to lift,-a destination, the import to cut, named groups); a complexity step also needs a-function of at least 30 code lines or CCN 15 that is not mostly one dispatch on one-value.--**The refactoring view lists those by default.** On this repository's index that is-94 of 702 open opportunities: 507 fall below the worth floor, 42 are in tests, 41-are small functions, and 18 are tooling, low-value kinds, docs, generated or-without a concrete step. The eligibility is read from the Fix-first builder itself,-so the view and the queue cannot disagree.--| `scope` | Lists |-|---|---|-| `fix_first` *(default for the open, repository-wide queue)* | Only what Fix first would take. `hidden` counts the rest of the filtered set, `{"total": n, "by_reason": {...}}`, by the same exclusion reasons. |-| `all` | The full inventory. The default when the call names a file, or lists a triaged status, which Fix first never reads. |--The REST route takes `scope`, MCP `get_health` takes `refactoring_scope` and reports-`refactoring_opportunities_scope` and `refactoring_opportunities_hidden`, and facet-counts follow the scope. The web board's Scope control switches between **Worth-doing** and **Full inventory**, and its count line reads "Showing N worth doing; M-more in the full inventory", with the reasons beneath it.--## Surfaces+Each detector reports nothing when a signal it needs is missing, so a gap means+"no suggestion", never a guessed one. Output is deterministic: the same commit+gives the same plans and the same ids.++## Quick start```bash-repowise health --refactoring-targets # ranked table+repowise health --refactoring-targets # the stored queue, in serving order+repowise health --refactoring-targets --format json+repowise health --refactoring-targets --recompute # analyze the working tree instead of the index+repowise health --refactoring-targets --generate-code 1 # code + diff for the top plan (LLM)```+From an agent, through MCP:+```python-# MCP. A bare call leads with fix_first, which ranks refactorings beside the rest.-get_health(only=["fix_first"])+get_health(only=["fix_first"]) # refactorings ranked beside other fixesget_health(include=["refactoring"], only=["refactoring_opportunities"], limit=6)-get_health(include=["refactoring"], only=["refactoring_summary"])-get_health(opportunity_id="refop2_...") # steps, plans, validation-get_health(opportunity_id="refop2_...", only=["refactoring_evidence"], cursor=3)-get_health(plan_id="refac2_...") # one plan, and its owner-get_health(include=["refactoring"], only=["refactoring_plans"]) # the raw list, opt-in-get_health(targets=["src/api/server.py"]) # one file+get_health(targets=["src/api/server.py"], include=["refactoring"]) # one file+get_health(opportunity_id="refop...") # one opportunity's steps+get_health(plan_id="refac...") # one plan```-```text-# REST. Both surfaces read services/refactoring_health.py, so they cannot-# answer differently; a parity suite asserts order, filters, totals and detail.-GET /api/repos/{repo_id}/refactoring/opportunities?view=diversified&limit=20-GET /api/repos/{repo_id}/refactoring/opportunities?scope=all # the full inventory-GET /api/repos/{repo_id}/refactoring/opportunities?file_path=src/api/server.py-GET /api/repos/{repo_id}/refactoring/opportunities/{opportunity_id}-GET /api/repos/{repo_id}/refactoring/summary-GET /api/repos/{repo_id}/refactoring/targets?refactoring_type=extract_class&min_confidence=high-GET /api/repos/{repo_id}/refactoring/targets/page?limit=60&offset=0&refactoring_type=performance_fix-GET /api/repos/{repo_id}/refactoring/{suggestion_id} # one plan + blast-radius detail-```+In the dashboard, the **Refactoring** page lists every type on one board, with+filters for type, confidence and effort. A plan's drawer explains its rank, shows+the tests that reach the code and the command to run them, and can export the+structured plan for an agent. The page never edits code or runs tests on its own.++## Reading the results++A **plan** is one detector's output for one target. An **opportunity** is one+file's refactoring work: the finding it leads with, and its plans as ordered steps+that are safe to apply in sequence. Surfaces list opportunities; a step points to+its plan by `plan_id`.++| Field | What it tells you |+|-------|-------------------|+| `refactoring_type` | Which of the seven plan types this is. |+| `plan` | The concrete change: groups, span and signature, destination class, cut edges, or file split. |+| `evidence` | The measurements behind it, such as LCOM4, clone size, co-change count or modularity. |+| `confidence` | `low`, `medium` or `high`. A `high` Extract Method is mechanical enough to hand to an agent. |+| `effort_bucket` | `S`, `M`, `L` or `XL`, from the size of the target. |+| `blast_radius` | What else has to change: callers, importing files, co-change partners. Extract Method is `local`. |+| `impact_delta` | Health the plan would recover, in proportion to what it removes. `0` for Break Cycle, Split File and Performance Fix, which rank on their own evidence. |+| `source_biomarker` | The health finding the plan answers, such as `god_class` or `complex_method`. |++On an opportunity, each step carries `applicability`: `mechanical` (only+extractions qualify) or `judgment`, with the facts behind the call. When an+earlier step moves a symbol to another file, the later step carries+`relocated_by`: its stored location is where the symbol was before, so find it+again before applying.++Names are never invented. A suggested helper or file name is `null` when nothing+in the code anchors one.++An empty list means no detector found work that clears its gates. It does not mean+the code needs no attention: check the findings in+[CODE_HEALTH.md](CODE_HEALTH.md).++### Scope and order++By default the open, repository-wide list shows only what+[Fix first](CODE_HEALTH.md#fix-first) would take: opportunities in shipped code+that recover real health and name a concrete edit. The response counts the rest+and the reasons they were held back. Ask for the full inventory with `scope=all`+(MCP: `refactoring_scope="all"`; dashboard: **Full inventory**). A call that names+a file, or a triaged status, gets the full inventory automatically.++Many top opportunities score the same. The default order (`diversified`)+round-robins over the lead finding, plan type and area so the first rows cover+distinct problems. `canonical` returns the plain rank order. Test files always rank+after production files.++## Code generation++Code generation is **on by default** but only runs when you ask for it, and it+needs an LLM API key. It never runs during indexing. Set+`refactoring.llm.enabled: false` to turn it off for the dashboard and MCP. The CLI+flag is itself an explicit request and does not read that setting.++You can ask from three places:++- CLI: `repowise health --refactoring-targets --generate-code <rank-or-symbol>`.+- Dashboard: the **Generate code** action on a plan, when the server runs on your+ machine with the checkout on disk.+- MCP: `generate_refactoring_code(suggestion_id=...)`. This tool is off the default+ MCP surface; add it with `mcp.tools: ["+generate_refactoring_code"]`.++Generation reads the plan's real source spans from the working tree, sends the+plan, the source and the graph and co-change context to your configured provider,+and returns refactored code with a unified diff. Where a check is cheap it runs+one: Extract Class compares LCOM4 before and after, Split File checks each new file+is under the size floor with no symbol duplicated, and Extract Method checks the+remaining function's complexity dropped. Results are cached by a hash of plan,+source and model, so an unchanged plan is not paid for twice. Nothing is applied+to your files.++It uses the repository's configured `provider` and `model`, or the first provider+whose API key is set. With generation disabled, the MCP tool returns the plan with+`generation.available: false` and the REST endpoint returns `403`. With no key, MCP+returns `error: "no_provider"`.-A file's opportunities are a **separate call** - the queue filtered by-`file_path` - and deliberately not a field on the file-detail aggregate. That-aggregate is already the widest read in the server and every one of its callers-would pay for this whether or not it renders it, while the surface that wants-it knows it wants it. R2's `(repository_id, status, file_path)` index makes it-one indexed lookup rather than a scan.--The web **Refactoring** page uses one shared plan board and drawer for every-type, including a Performance filter. Its initial data path is bounded and-server-owned: search, type, confidence, effort, canonical sorting, true totals,-and deterministic offset cursors. The older unpaged endpoint remains available-during migration. The drawer explains benefit, leverage, cost, risk and rank;-shows validation basis and provenance, true test totals, capped tests and-commands; and makes affected paths and intervention points navigable. Each card-can still export the structured plan for an agent. It never runs tests, creates-edits, or applies a refactoring automatically.--## Optional code generation--The deterministic plan is the product. The LLM code-generation step is **never-in the indexing hot path**: it runs only when you ask for code for a specific-plan, and it is on by default. Configure it in `.repowise/config.yaml`:+## Configuration++In `.repowise/config.yaml`:```yamlrefactoring:- enabled: true+ enabled: true # set false to skip the refactoring layerdetectors:- disabled: [] # e.g. [move_method]- min_confidence: medium # low | medium | high+ disabled: [] # e.g. [move_method, split_file]+ min_confidence: medium # low | medium | high; unset keeps every planllm:- enabled: true # on by default; set false to disable- provider: null # falls back to the repo's configured provider- model: null+ enabled: true # code generation; set false to disable```-When code generation is enabled (the default), the **Generate code** action on a plan card (or the-endpoint below) gathers the plan's real source spans off the working tree, builds-a behavior-preservation prompt carrying the structured plan **plus the-graph/co-change context** a bare codegen tool throws away, and returns the-refactored code and a unified diff. Where a self-check is cheap and meaningful it-runs one: Extract Class re-walks the generated classes for an **LCOM4 before/after-delta**, and Split File re-walks the generated files to assert each is **below the-size floor** and the symbols are **partitioned with no duplication**. Results are-cached on disk by a content hash (plan + source + model), so the same plan never-pays twice.--```text-POST /api/repos/{repo_id}/refactoring/{suggestion_id}/generate-code-GET /api/repos/{repo_id}/refactoring/settings # read llm.enabled + provider/model-PUT /api/repos/{repo_id}/refactoring/settings # toggle it from the dashboard-```--Code generation needs the working tree on disk (it reads the real source spans),-so it is a local-`serve` capability: it returns `403` when disabled and `404`-when the repo has no accessible checkout. **Apply is out of scope**: the wedge is-the plan and the reviewable diff, not auto-applied edits.--## Configuration--Per-path disables reuse the existing `.repowise/health-rules.json` glob-mechanism, so a refactoring type can be silenced for generated or vendored paths-the same way a marker is:--```json-{- "rules": [- { "path": "src/generated/**", "disabled_biomarkers": ["dry_violation"] }- ]-}-```+Detector names are the snake-case plan types: `extract_class`, `extract_method`,+`extract_helper`, `move_method`, `break_cycle`, `split_file`, `performance_fix`.++Most plans answer a health finding, so per-path marker rules in+`.repowise/health-rules.json` also remove the plans for those findings. See+[CODE_HEALTH.md](CODE_HEALTH.md#configuration).++## Accuracy and limits++- Plans are suggestions for review. Nothing is applied automatically, and+ generated code comes back as a diff.+- Extract Method offers a span only when it can show the extraction keeps behavior+ (every returned value written on every path, no state carried across loop+ iterations). Spans it cannot prove are dropped, so it under-reports by design.+- Extract Method skips spans too small to matter and spans that would carry the+ original finding into the helper. A component whose branching is mostly in its+ markup gets no extraction.+- Move Method never targets a class the method only instantiates, or an ancestor of+ its own class.+- Split File works on any language with call resolution and suggests a split only+ when the groups are clearly separable. Python and TypeScript splits usually need+ a re-export shim, flagged as `shim_required`.+- Plan quality follows graph quality. Where calls are not resolved for a language,+ Move Method and Split File have less to work with.+- Generated code depends on the model you configure. The self-checks catch some+ structural mistakes, not behavior changes; review and test the diff.+- There is no published accuracy benchmark for refactoring plans yet.++## Reference++- MCP: [`get_health`](../agent/MCP_TOOLS.md#get_health),+ [`generate_refactoring_code`](../agent/MCP_TOOLS.md#generate_refactoring_code).+- CLI flags on `repowise health`:++| Flag | Effect |+|------|--------|+| `--refactoring-targets` | Print the stored refactoring queue in the order MCP and the dashboard serve it. |+| `--recompute` | With `--refactoring-targets`: analyze the working tree in-process. Slow on large repos. |+| `--generate-code SELECTOR` | Generate code and a diff for one plan. `SELECTOR` is a 1-based rank or a symbol name. Needs an API key. |+| `--format json` | Machine-readable output. |## See also-- [`docs/layers/CODE_HEALTH.md`](CODE_HEALTH.md): the markers and the three health- signals the suggestions are built on.-- [`docs/layers/INTELLIGENCE_LAYERS.md`](INTELLIGENCE_LAYERS.md): how code health fits- the five-layer index.-- [`docs/agent/MCP_TOOLS.md`](../agent/MCP_TOOLS.md): the `get_health(include=["refactoring"])`- response shape.+- [CODE_HEALTH.md](CODE_HEALTH.md): the findings and health signals plans are built on, and Fix first.+- [INTELLIGENCE_LAYERS.md](INTELLIGENCE_LAYERS.md): how code health fits the rest of the index.+- [architecture/refactoring.md](../architecture/refactoring.md): detector algorithms, ranking, plan ids and REST routes.+- [architecture/code-health.md](../architecture/code-health.md): health pipeline internals.diff --git a/docs/layers/SECURITY.md b/docs/layers/SECURITY.mdindex 8d73ecdae..d2f207dcf 100644--- a/docs/layers/SECURITY.md+++ b/docs/layers/SECURITY.md@@ -1,33 +1,29 @@-# Security Signals+# Security signals-Repowise records a small set of security signals while it indexes: pattern-matches over source text, and symbol names that look security-relevant. Pure-regex and SQL, no LLM calls, no network, no dependency resolution.+Repowise records security signals while it indexes: matches from a registry of+22 patterns over source text, plus a scan for security-relevant symbol names.+It also scans git history for leaked credentials, and gates pull requests on+what a change adds.-This is a floor, not a scanner. The registry holds twenty-two patterns. It has no-model of your framework, no notion of which inputs are attacker-controlled, and-no dataflow: it cannot tell a parameterised query from a concatenated one beyond-what the surrounding characters give away, and it cannot tell whether a-dangerous call is reachable. Everything below is written so you can see exactly-where that floor sits, because a security surface that overstates itself is-worse than one that reports nothing.+It is a pattern registry, not a security scanner. There is no taint analysis,+no framework model and no dependency lookup. Run a real SAST tool and a secret+scanner alongside it.-Run a real SAST tool alongside it. Nothing here replaces one.+No LLM key, no network. Working-tree signals cost a regex pass during+indexing; the CI check needs only git.## Quick start```bash-repowise init # working-tree signals populate during indexing-repowise security scan --history # walk every tracked revision for leaked secrets+repowise init # working-tree signals populate during indexing+repowise security scan --history # walk every revision for leaked secretsrepowise security scan --history --since v1.0.0 --to HEAD-repowise security scan --history --all-patterns --format json-repowise security check origin/main...HEAD # CI gate on what a change adds (see below)+repowise security check origin/main...HEAD # CI gate on what a change adds```-`scan --history` stores what it finds and prints counts, not a list:+`scan --history` stores what it finds and prints counts:```-repowise security scan --historyCommits scanned: 412Blobs scanned: 2318Files scanned: 1604@@ -36,206 +32,122 @@ repowise security scan --historyBy kind: aws_access_key=1, hardcoded_password=2, hardcoded_secret=5```-The findings themselves are read where the stored rows surface, below.+From an agent, `get_risk(targets=["src/api/"])` includes a `security_signals`+block. In the dashboard, findings are on the Security tab of the Code Health+page, and over REST at `GET /api/repos/{repo_id}/security`.-From an agent, through the risk surface:+## Reading the results-```python-get_risk(targets=["src/api/"]) # includes a security_signals block for the target-```--Findings also appear on the Security tab of the code-health page, and at-`GET /api/repos/{repo_id}/security`.+### What the registry catches-## What the registry catches--Twenty-two patterns plus a symbol-name scan, giving twenty-three kinds across-three severities. Severity is a fixed property of the pattern, with one-exception: any finding in test material, or under a directory named-`test`, `tests`, `__tests__`, `__test__`, `fixtures`, `__fixtures__`, `spec`,-`specs`, `mock`, `mocks`, `__mocks__`, `example` or `examples`, is recorded at-`low`. Keys there are mostly fake and calls there do not ship, but a real one-is still worth seeing. Nothing is scored, ranked, or aggregated.+22 pattern kinds plus the symbol-name scan: 23 kinds in all, in three+severities. Severity is fixed per kind, with one exception: a finding in test+material, or under a directory named `test`, `tests`, `__tests__`, `__test__`,+`fixtures`, `__fixtures__`, `spec`, `specs`, `mock`, `mocks`, `__mocks__`,+`example` or `examples`, is recorded at `low`. Nothing is scored or ranked.| Kind | Severity | Matches ||------|----------|---------|-| `eval_call` | high | `eval(...)`, including a receiver chain (`vm.eval(`, `foo.bar.eval(`) |-| `exec_call` | high | `exec(...)`, same receiver handling. Outside Python only a call through a name bound to `child_process` (a namespace, a named import, or the `require` result), covering `execFile` / `execSync` too |+| `eval_call` | high | `eval(...)`, including a receiver chain (`vm.eval(`) |+| `exec_call` | high | `exec(...)`. Outside Python, only a call through a name bound to `child_process` (`exec`, `execFile`, `execSync`) || `pickle_loads` | high | `pickle.loads` |-| `subprocess_shell_true` | high | `subprocess.*` with `shell=True`, including across physical lines |+| `subprocess_shell_true` | high | `subprocess.*` with `shell=True`, also across lines || `os_system` | high | `os.system` |-| `hardcoded_password` | high | an assignment of a quoted literal to a name containing `password`, any case, where the name is code (not inside a comment, docstring or string) |-| `hardcoded_secret` | high | the same for a name containing `api_key`, `apikey`, `secret`, `token` or `access_key`. A snake_case value (a constant holding a key's name) or a template placeholder (`{{ ... }}`, `${...}`, `$(...)`) is skipped, and a value one of the vendor kinds below already reports is not reported twice |-| `aws_access_key` | high | an AWS access key ID (`AKIA`/`ASIA` + 16 chars), regardless of variable name |-| `github_token` | high | a GitHub PAT, OAuth, app, or refresh token (`ghp_`/`gho_`/`ghu_`/`ghs_`/`ghr_`/`github_pat_`), regardless of variable name |-| `slack_token` | high | a Slack token (`xoxb-`, `xoxa-`, `xoxp-`, `xoxr-`, `xoxs-`), regardless of variable name |-| `google_api_key` | high | a Google API key (`AIza` + 35 chars), regardless of variable name, but not from the middle of a base64 run such as a lockfile integrity hash |-| `stripe_key` | high | a Stripe secret or restricted key (`sk_` or `rk_`, then `live_`, `test_` or `prod_`), regardless of variable name. Publishable `pk_` keys are public by design and do not fire |-| `private_key_pem` | high | a PEM header (`-----BEGIN ... PRIVATE KEY-----`) followed by a base64-looking body line, so assembling just the header string does not fire. The break may be an escaped `\n`, which catches keys held on one line in JSON or `.env` files |-| `fstring_sql` | med | an f-string containing `SELECT` and an interpolation |+| `hardcoded_password` | high | a quoted literal assigned to a name containing `password`, any case |+| `hardcoded_secret` | high | the same for `api_key`, `apikey`, `secret`, `token`, `access_key` |+| `aws_access_key` | high | `AKIA`/`ASIA` + 16 chars, any variable name |+| `github_token` | high | `ghp_`/`gho_`/`ghu_`/`ghs_`/`ghr_`/`github_pat_` tokens |+| `slack_token` | high | `xoxb-`, `xoxa-`, `xoxp-`, `xoxr-`, `xoxs-` tokens |+| `google_api_key` | high | `AIza` + 35 chars, not from inside a base64 run |+| `stripe_key` | high | `sk_` or `rk_` then `live_`/`test_`/`prod_`. Publishable `pk_` keys do not fire |+| `private_key_pem` | high | a `-----BEGIN ... PRIVATE KEY-----` header followed by a base64 body line (an escaped `\n` counts) |+| `public_env_secret` | high | `API_KEY`/`SECRET`/`TOKEN`/`PASSWORD` behind `NEXT_PUBLIC_` or `VITE_`, excluding `..._ANON_...` |+| `new_function_call` | high | `new Function(...)` |+| `fstring_sql` | med | a Python f-string with `SELECT` and an interpolation || `concat_sql` | med | `.execute("SELECT ... +` |+| `template_literal_sql` | med | a JS/TS template literal with `SELECT`+`FROM` or `UPDATE`+`SET` and an interpolation || `tls_verify_false` | med | `verify = False` |-| `weak_hash` | low | the words `md5` or `sha1` |-| `unsafe_inner_html` | med | `__html:` (React's `dangerouslySetInnerHTML` shape) assigned a non-literal value |-| `template_literal_sql` | med | a JS/TS template literal containing `SELECT`+`FROM` or `UPDATE`+`SET` and an interpolation |-| `public_env_secret` | high | a secret-shaped name (`API_KEY`, `SECRET`, `TOKEN`, `PASSWORD`) behind a `NEXT_PUBLIC_` or `VITE_` prefix, excluding `..._ANON_...`, outside comments |-| `new_function_call` | high | `new Function(...)` || `reject_unauthorized_false` | med | `rejectUnauthorized: false` |-| `security_sensitive_symbol` | low | a symbol whose name contains `auth`, `token`, `password`, `jwt`, `session` or `crypto` |--The last one is informational. It flags nothing wrong; it marks where the-security-relevant code lives.--Two details worth knowing:--**`eval` and `exec` are resolved properly, the rest are not.** In Python files-they are found by walking the AST, so a match is a real call rather than a-substring, with a bounded lexical fallback when the file does not parse. Other-languages get the lexical path, over source with comments and string literals-masked out, so an `eval(` inside a comment does not fire. Outside Python-`exec_call` carries one more condition: the call has to go through a name the-file binds to `child_process`. The imports are read from raw source because the-module arrives as a string literal that masking would blank.--**Text that describes code is not code.** `pickle_loads`, `os_system`,-`subprocess_shell_true` and `new_function_call` match over the same masked source, so a docstring or a-comment naming the call does not fire. The two keyword secret kinds need their-name to be code, so `password = "..."` in a doctest is an example, not a-password, and `public_env_secret` ignores a comment naming the variable. The-vendor key shapes (`aws_access_key` and the rest) still match anywhere,-comments included, because a real key pasted into a comment is still a leak.-Prose files (`.md`, `.mdx`, `.rst`, `.txt`, `.adoc`) are scanned for secret-kinds only.--**Examples of a credential are not one.** A keyword value that is elided or-templated (`...`, a typographic ellipsis, `<your key>`, `your-`, `xxx`,-`example`, `placeholder`, `dummy`, `fake`, `changeme`), shorter than eight-characters, a short single-case word (`api_key="lmstudio"`, the dummy a local-server's client requires), or a CSS custom property (`--chart-1`) is skipped.--**Two patterns see across lines.** `private_key_pem` needs the body line after-its header, and `subprocess.run(` opening on one line with-`shell=True` several lines down is invisible to a per-line scan, so-`subprocess_shell_true` gets a second pass over the whole source. Continuation-is restricted to lines that begin with indentation and capped at roughly 200-characters, so a closed call cannot reach forward into an unrelated one.--## What the registry does not catch--Stated plainly, because the table above reads like more coverage than it is.--**Most languages.** The registry is shaped by Python, with five patterns added-for JavaScript and TypeScript (below). `pickle.loads`, `os.system`,-`subprocess(shell=True)` and `verify=False` are still Python-only idioms;-`fstring_sql` is a Python f-string. On a Go, Rust or Java codebase, `eval` /-`exec` and the two secret patterns are most of what can fire. A repo in one of-those languages reporting few findings is reporting the registry's shape, not-its own health.--**Anything needing framework semantics.** Missing authorization on a route-handler, permissive CORS on a mutating endpoint, an unvalidated redirect, an-object reference with no ownership check. These require knowing what a route-handler is, which ones mutate, and what counts as a gate. A regex cannot do it-and this layer does not try.--**Anything needing dataflow.** Whether attacker-controlled input actually-reaches a dangerous call is not computed. `eval_call` fires the same on a-constant and on a request parameter.--**Real secret detection.** `hardcoded_password` / `hardcoded_secret` match a-literal assignment to a variable named like a credential, in any case, and the-six value-shape kinds above (AWS, GitHub, Slack, Google, Stripe, PEM) catch a-handful of common vendor formats regardless of variable name. That is still far-short of a real scanner: no entropy scoring, and any format or provider not in-that short list is invisible. For secret scanning proper, run gitleaks or-trufflehog; history mode below is complementary to those, not a replacement.--**Dependencies.** No CVE lookup, no advisory feed, no SBOM. Nothing here looks-outside your source.--**False positives are still possible.** `weak_hash` fires on the word `md5`-anywhere, including in a comment explaining why md5 was removed. The high-kinds are tuned the other way: a `high` should be something to act on, so a-match that is only probably real is dropped.--**Two of them are worth knowing before you read a report.**--`exec` is not a global in JavaScript; the name belongs to `RegExp.prototype.exec`,-so the receiver-chain prefix in the pattern would otherwise match-`re.exec(expr)`, `/x/.exec(s)` and `cellPattern.exec(xml)` (ordinary parsing-code) at `high`. Outside Python the kind is resolved through the file's own-`child_process` bindings, which is where the dangerous call comes from: a-file's own function named `exec`, or a regex `.exec` beside a `spawn` import,-does not fire.--The remaining per-line patterns (`weak_hash`, the SQL kinds, the TLS kinds,-`unsafe_inner_html`) run on raw source, comments included,-so `weak_hash` still fires on a comment explaining why md5 was removed.--**Five patterns for JavaScript and TypeScript, measured the same way as the-`exec_call` and secret-case fixes above** — a 17-repository, 1109-file corpus,-every hit adjudicated by hand. Three of the five needed tightening before they-were worth shipping:--| Kind | First cut | Verified real | After tightening | The noise |-|---|---|---|---|---|-| `unsafe_inner_html` | 15 | 0 | 4 | every hit a source-pinned stylesheet constant referenced by name |-| `template_literal_sql` | 10 | 2 | 2 | prose (`` `Order update failed: ${status}` ``) and a Tailwind class (`select-none`) |-| `public_env_secret` | 9 | 3 | 3 | Supabase anon keys — public by design, protected by row-level security |-| `new_function_call` | 0 | 0 | 0 | — |-| `reject_unauthorized_false` | 0 | 0 | 0 | — |--`unsafe_inner_html` ships at `med`, not `high`, because of that first row: a-name reference cannot be told apart from a pinned constant without dataflow,-so the pattern is a places-to-read signal, not a confirmed sink. The same-gap explains why the value test (`__html:` followed by an identifier, `$`, or-`(`) is written as a positive lookahead rather than a negative one excluding-quotes — tried at the position *before* the preceding `\s*`, a negative-lookahead would trivially pass and let a string literal through anyway.--`public_env_secret` excludes any name containing `ANON`, which is the one-carve-out narrow enough to state as a rule rather than a heuristic: an-`..._ANON_...` key is meant to be public. Nothing else in the corpus noise was-worth a similarly specific exclusion.--`template_literal_sql` requires the SQL verb's companion clause —-`SELECT`+`FROM`, `UPDATE`+`SET` — because the bare verb is an ordinary English-word and fires on both prose and a Tailwind utility class otherwise.--## Working tree versus history--Two scan surfaces share the registry and the storage.--**Working tree** runs during `repowise init` and `repowise update`, over the-files as they exist now. Rows are stored with an empty `commit_sha`. There is no-separate command for it: `repowise security scan` without `--history` is a stub-that prints a hint, because re-scanning the working tree outside indexing would-duplicate what indexing just did.--**History** runs only on `repowise security scan --history`, walking every-tracked revision of every source file. Rows carry the SHA and author date of the-commit that introduced the match. This finds what the working tree cannot: a-credential committed in March and removed in April is absent from HEAD and-present in the clone forever.--History mode reports **only the `SECRET_KINDS` (genuine leaked-credential)-kinds** by default — `hardcoded_password`, `hardcoded_secret`, and the six-vendor value-shape kinds above. The reasoning is asymmetry of decay. A commit-that once called `eval()`-is history doing what history does — the code changed, that is the point, and-reporting every such moment across every revision buries the surface in noise. A-committed secret does not decay. It stays valid until someone rotates it, and-whoever cloned the repo still has it. Pass `--all-patterns` to get the code-smell-kinds across history too, and expect volume.--Both paths land in the same `security_findings` table, with a unique constraint-on `(repository_id, file_path, kind, line_number, commit_sha)`. Re-running either-scan is idempotent.+| `unsafe_inner_html` | med | `__html:` (React's `dangerouslySetInnerHTML`) assigned a non-literal |+| `weak_hash` | low | the words `md5` or `sha1` |+| `security_sensitive_symbol` | low | a symbol named with `auth`, `token`, `password`, `jwt`, `session` or `crypto` |++`security_sensitive_symbol` is informational: it marks where security code+lives, not a problem.++The eight **secret kinds** are `hardcoded_password`, `hardcoded_secret` and+the six vendor shapes (AWS, GitHub, Slack, Google, Stripe, PEM). History mode+and the in-change commit scan report only these by default.++### What fires and what does not++- **`eval` and `exec` in Python** are found by walking the AST, so a match is+ a real call. Other languages use source with comments and strings masked out.+- **Text describing code is not code.** `pickle_loads`, `os_system`,+ `subprocess_shell_true` and `new_function_call` match over masked source, so a+ comment naming the call does not fire. The keyword secret kinds need the name+ to be code, so `password = "..."` in a doctest does not fire.+- **Vendor key shapes match anywhere**, comments included: a real key in a+ comment is still a leak.+- **Prose files** (`.md`, `.mdx`, `.rst`, `.txt`, `.adoc`) are scanned for the+ secret kinds only.+- **Placeholders are skipped**: a keyword value that is elided or templated+ (`...`, `<your key>`, `xxx`, `example`, `placeholder`, `dummy`, `fake`,+ `changeme`, `{{ }}`, `${}`), shorter than eight characters, a short+ single-case word, or a snake_case key name. A value a vendor kind already+ reports is not reported twice.+- **`weak_hash`, the SQL kinds, the TLS kinds and `unsafe_inner_html`** run on+ raw source, comments included.++### Working tree versus history++**Working tree** runs during `repowise init` and `repowise update`. Rows have+an empty `commit_sha`. `repowise security scan` without `--history` only+prints a hint, since indexing already did this.++**History** runs on `repowise security scan --history` and walks every tracked+revision. Rows carry the SHA and author date of the commit that introduced the+match. This finds a credential committed in March and removed in April: absent+from HEAD, present in every clone. By default it reports only the secret kinds.+A committed secret stays valid until rotated; an `eval` that was later removed+is just history. `--all-patterns` adds the code kinds, with a lot more volume.++Both write the `security_findings` table, unique on file, kind, line and+commit, so re-running is idempotent.++### Line verification++A stored line number goes stale as the file moves on, so each served finding is+re-checked against the live file using its stored snippet:++| `line_verified` | `line_number` | Meaning |+|---|---|---|+| `true` | a line | The snippet is on that line (possibly after correction) |+| `false` | a line | The snippet occurs more than once; the line is a guess |+| `false` | `null` | The snippet is gone; the finding is stale |++`security_sensitive_symbol` is checked in place and never relocated, since a+name recurs throughout a file. When the file cannot be read, the stored line is+returned unverified.++Snippets are trimmed to 120 characters with every credential value masked to+its first four characters and `****`. The raw value is never stored.++### A suggested reading order++1. Secret kinds from history mode. Rotate first, then clean history.+2. High working-tree kinds, as places to read. The layer found a call; you+ decide whether it is reachable.+3. `security_sensitive_symbol`, as a map of where auth and crypto code lives.+4. Low findings last. `weak_hash` is a word match.++A low finding count means the registry did not match. On most repos, in most+languages, that is all it means.## In CI: `repowise security check`-A third surface, for pull requests. It needs git and nothing else: no index,-no database, no API key, and it stores nothing.+The PR surface. It needs git only: no index, no database, no API key, and it+stores nothing.```bashrepowise security check origin/main...HEAD@@ -243,150 +155,134 @@ repowise security check --fail-on med --format githubrepowise security check --format sarif > security.sarif```-It judges what the change adds, not the repository:--- **At the change's head**, every file the change touched is scanned as the- head commit has it, and a finding counts only when a line it spans is one- the change added or edited. The two multi-line kinds count when any line of- their span changed, so editing only the `shell=True` line of a call, or only- the body of a PEM key, still counts. In a documentation file (`.md`, `.mdx`,- `.rst`, `.txt`, `.adoc`) only the secret kinds count: prose naming- `pickle.loads` is not a call, but a key pasted into a README is a leak.-- **Inside the change**, every commit in the range is scanned the same way- for the secret kinds only. A key committed in one commit and deleted in the- next is absent from the head and present in every clone; the gate reports it- with the commit that added it and says to rotate it. Merge commits are- skipped, so merging the target branch in does not blame its lines on the- change. A code smell added and removed inside the change does not count.--`--fail-on` names the lowest severity that fails (`high` by default, then-`med`, `low`). A finding under a test, fixture, spec, mock or example path is-`low` (see above), so under the default it shows as a warning and does not-fail; pass `--fail-on low` to fail on those too. Exit codes are the ones every-repowise CI gate uses: `0` passed,-`1` failed, `2` could not evaluate (not a git repository, unknown revision, no-merge-base in a shallow clone, a commit of the change cut off by a shallow-clone, an unreadable baseline). A git error is never read as a clean change.--Every format carries the masked snippet only, because CI logs are often-public: `table`, `json`, `markdown`, `github` (annotations plus the job-summary), `sarif` (for code-scanning upload) and `gitlab` (a GitLab Code-Quality report for the merge request widget). The gate is the registry-above and has its limits: a pass means no pattern matched a changed line.--**Baseline.** `--write-baseline FILE` records the change's findings in a-committed JSON file and exits 0; `--baseline FILE` accepts what it lists, so-only new findings fail. A finding is keyed on its file, kind and masked-matched line, so it stays accepted when an edit above it shifts its line-number. The vendor key shapes mask to their fixed prefix (`AKIA****`), so their-key also takes a one-way hash of the matched text: a different key on an-identical line is a new finding. The file never holds a raw value. The keyword-kinds (`hardcoded_password`, `hardcoded_secret`) stay keyed on the masked line-alone, because a short password could be guessed back from a hash of it; a-different value sharing the first four characters on an identical line stays-accepted. Because the gate sees one change at a time, writing to an existing-baseline adds to its entries (and to those of `--baseline`, when given)-instead of replacing them; remove an entry by deleting it from the file.-`--format sarif` marks accepted findings as suppressed; `--format gitlab`-leaves them out, because that format has no suppression field. In the Code-Quality report a finding at or above `--fail-on` is `critical` when high and-`major` otherwise, and one below it `minor`.--Without a REVSPEC the base comes from the CI's pull-request variables, else-the remote's default branch. A shallow checkout cannot be read commit by-commit, so fetch the full history; the gate exits 2 rather than guess. The-GitHub Action, the GitLab template and the SARIF upload are in-[Repowise in CI](../start/CI.md).+It judges what the change adds:++- **At the head**, every touched file is scanned, and a finding counts only+ when a line it spans was added or edited. In documentation files only the+ secret kinds count.+- **Inside the change**, every non-merge commit in the range is scanned for the+ secret kinds. A key added in one commit and deleted in the next is reported+ with the commit that added it.++`--fail-on` sets the lowest failing severity (`high` by default, then `med`,+`low`). Findings in test paths are `low`, so under the default they warn+without failing. Exit codes: `0` passed, `1` failed, `2` could not evaluate+(not a git repo, unknown revision, missing merge-base or commits in a shallow+clone, unreadable baseline). A git error is never read as a clean change.++Every format carries masked snippets only: `table`, `json`, `markdown`,+`github` (annotations plus job summary), `sarif` and `gitlab` (Code Quality).+In the Code Quality report a failing finding is `critical` when high and+`major` otherwise; one below `--fail-on` is `minor`.++Without a revspec, the base comes from the CI's pull-request variables, else+the remote's default branch. The GitHub Action, GitLab template and SARIF+upload are in [Repowise in CI](../start/CI.md).++## Tuning and suppressing### Silencing one finding: `repowise-security-ignore`-For a single false positive, put `repowise-security-ignore` on the finding's-own line, in any comment syntax; `repowise-security-ignore: aws_access_key`-silences only the kinds it lists. There is no next-line or whole-file form: use-the baseline for that. A silenced finding never fails the gate and is still-reported as suppressed in every format. The exact rules are in the+Put `repowise-security-ignore` on the finding's own line in any comment+syntax; `repowise-security-ignore: aws_access_key` silences only the kinds it+lists. There is no next-line or whole-file form. A silenced finding never fails+the gate and is still reported as suppressed. Exact rules are in the[CLI reference](../reference/CLI_REFERENCE.md#repowise-security-check-revspec).+### Baselines++`--write-baseline FILE` records the change's findings in a committed JSON file+and exits 0; `--baseline FILE` accepts what it lists, so only new findings+fail. An entry is keyed on file, kind and masked line, so it survives line+shifts. Vendor shapes also key on a one-way hash of the matched text, so a new+key on an identical line is a new finding. Keyword kinds do not, since a short+password could be guessed back from its hash. Writing to an existing baseline+adds entries; remove one by deleting it. `sarif` marks accepted findings+suppressed; `gitlab` leaves them out.+### Custom patterns: `security.patterns`-A secret shape of your own goes under `security.patterns` in-`.repowise/config.yaml` as a name, a regex and a severity. Each becomes the-secret kind `custom:<name>`, scanned and masked like the built-in ones, and an-invalid pattern stops the check with exit 2. Keys and limits are in+Add your own secret shapes under `security.patterns` in `.repowise/config.yaml`+as a name, a regex and a severity. Each becomes the secret kind+`custom:<name>`, scanned and masked like the built-in ones. An invalid pattern+stops the check with exit 2. See[the `security:` block](../reference/CONFIG.md#the-security-block).### Before a commit: `--staged`-`repowise security check --staged` checks what the next commit would record:-the staged lines, read from the index. `repowise hook install --security` runs-it as a pre-commit hook that blocks a commit only on a finding at or above-`high`, and lets the commit through when the check cannot run. Skip it once-with `git commit --no-verify`. See+`repowise security check --staged` checks the staged lines, read from the+index. `repowise hook install --security` runs it as a pre-commit hook that+blocks a commit on a finding at or above `high`. Skip it once with+`git commit --no-verify`. See[`repowise hook install`](../reference/CLI_REFERENCE.md#repowise-hook-install).-## Line verification--A finding's `line_number` is written at scan time, and the file moves on. A wrong-line on a security finding is worse than none: it sends the reader to innocent-code while looking authoritative. So the line is re-checked against the live file-every time a finding is served, using the stored snippet: the matched line,-stripped and trimmed to 120 characters, with every credential value on it-masked to its first four characters and `****`. An unmasked snippet is a-substring of the line it came from; a masked one is matched on the text before-the first `****`.+## Accuracy and limits++This is a registry of regular expressions with a little parsing. Read every+result with that in mind.++- **No taint analysis or dataflow.** Whether attacker input reaches a call is+ not computed. `eval_call` fires the same on a constant and on a request+ parameter.+- **No framework semantics.** Missing authorization, permissive CORS,+ unvalidated redirects and missing ownership checks are out of reach.+- **Mostly Python and JS/TS.** `pickle_loads`, `os_system`,+ `subprocess_shell_true`, `tls_verify_false` and `fstring_sql` are Python+ idioms. On Go, Rust or Java, `eval`/`exec` and the secret kinds are most of+ what can fire, and a quiet report reflects the registry's shape.+- **Secret detection is shallow.** No entropy scoring; providers outside the+ six vendor shapes are invisible unless you add custom patterns. Use gitleaks+ or trufflehog for secret scanning proper.+- **No dependency or CVE scanning.**+- **False positives happen.** `weak_hash` fires on `md5` in a comment.+ `unsafe_inner_html` ships at `med` because it cannot tell a name pointing at+ a pinned constant from a dynamic value.+- **The five JS/TS kinds were checked by hand** on a 17-repository, 1,109-file+ corpus. After tightening: `unsafe_inner_html` 4 hits, `template_literal_sql`+ 2 of 2 real, `public_env_secret` 3 of 3 real, `new_function_call` and+ `reject_unauthorized_false` 0.+- **A CI pass means no pattern matched a changed line**, nothing more.++## Where it shows up++- **CLI**: `repowise security scan --history`, `repowise security check`,+ `repowise hook install --security`.+- **MCP**: the `security_signals` block on `get_risk`.+- **Dashboard**: the Security tab of the Code Health page.+- **REST**: `GET /api/repos/{repo_id}/security`.+- **CI**: the GitHub Action's `security` gate (annotations, SARIF) and the+ GitLab template (Code Quality).++## Reference++### `repowise security scan`++| Flag | Meaning |+|---|---|+| `--history` | Walk every tracked revision. Without it the command only prints a hint |+| `--since REV` | Lower bound (exclusive) |+| `--to REV` | Upper bound (inclusive) |+| `--all-patterns` | History mode: report the code kinds as well as secrets |+| `--path DIR` | Repository path |+| `--format` | `table` or `json` |-Three outcomes, on every finding the API returns:+### `repowise security check [REVSPEC]`-| `line_verified` | `line_number` | Meaning |+| Flag | Default | Meaning ||---|---|---|-| `true` | a line | The snippet is on that line. Either it never moved, or it moved and the line was corrected. |-| `false` | a line | The snippet occurs more than once in the file. The line is a guess and the surface should mark it as one. |-| `false` | `null` | The snippet is gone from the file. The finding is stale; showing a line would point at unrelated code. |--`security_sensitive_symbol` is the exception. Its snippet is a bare identifier-rather than a line of code, and an identifier recurs throughout a file, so-relocating on it would land somewhere arbitrary. Those findings are checked in-place and never relocated or withdrawn: if the name is not on the stored line-they are returned unverified rather than treated as gone.--When the file cannot be read at all — the repo is not checked out where the-server expects it — the stored line is passed through unverified rather than-claimed as correct.--## Storage--One table, `security_findings`, written by both scan paths:--| Column | Notes |-|---|---|-| `file_path` | Repo-relative |-| `kind` | One of the kinds in the table above |-| `severity` | `high`, `med`, `low`: fixed per pattern, except a finding in test material is `low`. The Overview reads `med` as `medium` |-| `snippet` | Matched line, trimmed to 120 characters, credential values masked (`AKIA****`); a symbol name for `security_sensitive_symbol`. The raw value is never stored |-| `line_number` | As of scan time; verified at serve time |-| `commit_sha` | Empty for working-tree rows, the introducing commit for history rows |-| `commit_at` | Author date, history rows only |--Indexing against a database that has not yet migrated the table skips-persistence silently; every other write failure is a real error.--## Reading the surface honestly--A useful way to read a repo's findings, in order:--1. **Any `SECRET_KINDS` finding from history mode** (`hardcoded_secret`,- `hardcoded_password`, or one of the vendor value-shape kinds). These- are the findings most likely to be both true and actionable. Rotate first,- remove from history second.-2. **The high-severity working-tree kinds**, as a list of places to read rather- than a list of bugs. The layer found a call; you decide whether it is- reachable.-3. **`security_sensitive_symbol` as a map.** Where the auth and crypto code- lives is useful context for review, and for `get_risk` when you are about to- change one of those files.-4. **Low findings last, and sceptically.** `weak_hash` in particular is a word- match.--And once more, because it is the whole point of this page: a low finding count-means the registry did not match. On most repos, in most languages, that is what-it means and nothing more.+| `--fail-on` | `high` | Lowest severity that fails: `high`, `med`, `low` |+| `--baseline FILE` | none | Accept findings listed in this file |+| `--write-baseline FILE` | none | Add this change's findings to the file and exit 0 |+| `--staged` | off | Check staged changes; takes no revspec |+| `--path DIR` | cwd | A path inside the repository |+| `--format` | `table` | `table`, `json`, `markdown`, `github`, `sarif`, `gitlab` |++Full entries: [`repowise security`](../reference/CLI_REFERENCE.md#repowise-security).+MCP schema: [`get_risk`](../agent/MCP_TOOLS.md#get_risk).++## See also++- [CHANGE_RISK.md](CHANGE_RISK.md): the other CI gate on a change.+- [CODE_HEALTH.md](CODE_HEALTH.md): the Code Health page that hosts the+ Security tab.+- [Repowise in CI](../start/CI.md).diff --git a/docs/layers/TEST_INTELLIGENCE.md b/docs/layers/TEST_INTELLIGENCE.mdindex 082585f49..b11053840 100644--- a/docs/layers/TEST_INTELLIGENCE.md+++ b/docs/layers/TEST_INTELLIGENCE.md@@ -1,26 +1,25 @@# Test Intelligence-Ingest a coverage report and repowise can answer two questions your CI cannot:-which files are risky *and* untested, and which tests a given diff actually-exercises. That second one turns a 4,000-test suite into the 40 tests that guard-the change you just made. The same report also gates a pull request on its-patch coverage, with no index at all (see-[Patch coverage in CI](#patch-coverage-in-ci)).+Test intelligence answers two questions your CI cannot: which files are risky+*and* untested, and which tests a given change exercises. The second turns a+4,000-test suite into the few dozen tests that guard the change you just made.-Without a coverage report there is still an answer, a weaker one. The dependency-graph records which test files import which source files, so repowise can say-*something reaches this* even where nothing measured it. That inferred tier needs-no setup and is always labelled as inferred. See-[The inferred tier](#the-inferred-tier-no-coverage-report-needed).+It works at two levels of evidence. Ingest a coverage report and the answers are+**measured**: this test ran these lines. Without one, repowise still answers+from the dependency graph, which records which tests reach which files; those+answers are **inferred**, always labelled as such, and never turned into a+percentage.-Everything here is an index lookup: no LLM, no network.+No LLM key, no network. Ingesting a report is a parse and a database write;+every question after that is an index lookup. The patch-coverage gate for pull+requests needs only git and a report (see [Patch coverage in CI](#patch-coverage-in-ci)).## Quick start```bash-# 1. Produce a report. Any of these work.+# 1. Produce a report. Any supported format works.pytest --cov --cov-report=lcov:coverage.lcov-coverage run --contexts=test -m pytest # also builds the per-test map+coverage run --contexts=test -m pytest # also records which test ran each line# 2. Ingest it.repowise coverage add coverage.lcov@@ -29,11 +28,14 @@ repowise coverage status# 3. Use it.repowise health # untested hotspots now light up-repowise impacted-tests main..HEAD # the tests guarding this branch-repowise impacted-tests main..HEAD --format list | xargs pytest+repowise impacted-tests main...HEAD # the tests guarding this branch+repowise impacted-tests main...HEAD --format list | xargs pytest```-Longer walkthrough: [examples/health-coverage/](../../examples/health-coverage/).+From an agent: `get_change_risk(revspec="main...HEAD")` returns `impacted_tests`+and `patch_coverage`; `get_risk(changed_files=[...])` returns test+recommendations. In the dashboard, coverage is the **Tests** tab of a+repository's health page.```repowise coverage status@@ -49,46 +51,45 @@ repowise coverage statusRecords: 19,551```-## Two dimensions, one command+Longer walkthrough: [examples/health-coverage/](../../examples/health-coverage/).-`repowise coverage add` stores two different things, and the difference matters-for everything below.+## Measured and inferred-| Dimension | What a row says | Powers |-|-----------|-----------------|--------|-| **Per-file aggregate** | This file is 71% covered, merged across every test. | `untested_hotspot`, `coverage_gap`, `coverage_gradient` in [code health](CODE_HEALTH.md), the coverage dashboard |-| **Per-test map** | Test `tests/test_auth.py::test_login` covered lines 40-58 of `src/auth/service.py`. | `repowise impacted-tests`, `get_change_risk`'s `impacted_tests`, `get_risk`'s `tests_to_run` |-| **Inferred map** (no ingest) | `tests/test_round_trips.py` imports `src/auth/service.py`, so it reaches it. | The fallback under every row above, always labelled `inferred` <!-- repowise-drift-ignore --> |+| | Per-file aggregate | Per-test map | Inferred map |+|---|---|---|---|+| Comes from | Any coverage report | A report that records which test ran each line | The call and import graph, already indexed |+| Says | `src/auth/service.py` is 71% covered | `tests/test_auth.py::test_login` ran lines 40-58 of `src/auth/service.py` | `tests/test_auth.py` reaches `src/auth/service.py` |+| Granularity | File | Line | File |+| May produce a percentage | Yes | Yes | **Never** |+| Goes stale | Yes: tied to the commit it was measured at | Yes | No: read from the current graph |+| Labelled | `measured` | `measured` | `inferred` |+| Powers | Code-health coverage findings, the Tests tab | `impacted-tests`, `impacted_tests`, `tests_to_run` | The fallback under every row, when nothing measured answers |-The aggregate always gets stored. The map is only built when the report carries-per-test contexts. A report without contexts still ingests fine, it just skips-the map. The inferred map is not stored at all: it is read off the graph when-asked.+The two kinds are shown side by side and never averaged. Inference errs in one+known direction: a call edge says control *can* reach a file, not that a run+did. So it is safe as a floor ("something reaches this, do not call it+untested") and unsafe as a quantity.-Both are point-in-time: each ingest replaces the previous rows rather than-appending history. Ingest at the same commit you intend to query, so line numbers-line up.+Each ingest replaces the previous per-file and per-test rows. Ingest at the+commit you intend to query so line numbers match. The repo-wide totals of the+newest 50 ingests are kept as history and drawn as a trend on the Tests tab+once there are three; partial ingests are left out of the trend.## Supported formats| Format | Detected by | Per-test map ||--------|-------------|--------------|-| **LCOV** | Leading `TN:` / `SF:`, or any `TN\|SF\|DA\|BRDA\|LF\|LH\|BRF\|BRH:` line | Yes, when each record carries a non-blank `TN:` test name |+| **LCOV** | `TN:` / `SF:` / `DA:` style lines | Yes, when each record carries a non-blank `TN:` test name || **Cobertura** XML | `<coverage` plus `<packages` or `line-rate` | No || **Clover** XML | `<coverage` plus `<project` | No || **JaCoCo** XML | The JaCoCo doctype, or a `<report` root | No || **Go coverprofile** (`go test -coverprofile`) | Leading `mode:` line | No |-| **coverage.py `.coverage`** | SQLite magic bytes | Yes, when written with `--contexts` |+| **coverage.py `.coverage`** | SQLite file | Yes, when written with `--contexts` || **Normalized JSON** (`repowise-coverage-v1`) | Leading `{` plus `repowise-coverage` or `line_coverage_pct` | No |-A UTF-8 byte-order mark at the start of a report is ignored. Go percentages are-line-based, not the statement percentage `go tool cover -func` prints. A Go-profile records blocks rather than lines, so every line a block spans counts as-executable, and a comment inside an unexecuted block reads as uncovered.--Force a parser with-`--format lcov|cobertura|clover|repowise-json|go-coverprofile|jacoco`. The-normalized JSON shape lets you feed any runner once you map it:+Force a parser with `--format lcov|cobertura|clover|repowise-json|go-coverprofile|jacoco`.+Go percentages are line-based, not the statement percentage `go tool cover`+prints. Any other runner can be fed through the normalized JSON:```json{ "format": "repowise-coverage-v1",@@ -99,353 +100,152 @@ normalized JSON shape lets you feed any runner once you map it:````covered_lines` and `coverable_lines` are optional. Patch coverage needs-`coverable_lines` (the executable-line set); without it a file reads as having-no line data, never as 0%.--With no path argument, `add` auto-discovers `coverage/lcov.info`, `lcov.info`,-`coverage.lcov`, `coverage.xml`, `**/cobertura.xml`, `**/clover.xml`,-`target/llvm-cov/**/*.lcov`, `coverage.out`, `cover.out`,-`**/target/site/jacoco*/jacoco.xml`, a root-module-`build/reports/jacoco/**/*.xml` (pass a nested Gradle module's report-explicitly), and a repo-root `.coverage`. Multiple reports merge with hit-wins:-covered lines and coverable lines are each unioned.--Report paths are matched to indexed files by exact key first, then basename, then-the longest trailing-path overlap. A tie refuses to guess and is reported as-ambiguous rather than mapped to the wrong file. If a whole report comes back-unmatched, set `coverage.strip_prefix` in `.repowise/config.yaml`.--Each ingest records how the report mapped: every format merged, how many of-its paths matched, did not match or tied, a short sample of the misses, and-the commit it was measured at. The coverage summary (REST `/health/coverage`,-MCP `get_health`) reports it as `report_paths`, and its `freshness` says-whether the measurement matches the indexed commit. Stored patch coverage-reads the same record, so its report path counts are real numbers rather than-null.--Ingest records are kept as history, not replaced: each one also stores the-repo-wide line and branch coverage it measured, and the newest 50 are-retained. REST `/health/coverage` returns them oldest first as `history`, and-the Coverage tab draws a trend under the headline figure once there are three-reports. Partial ingests are left out of the trend, since their figure covers-a fragment of the repository. A local index created before this change keeps-a single record per repository (its table still carries the old uniqueness-rule) until it is rebuilt.--A few rules keep that matching honest in monorepos and mixed layouts:--- A match needs more than the basename when the report names a directory:- `other/pkg/utils.py` never maps to `src/utils.py`. A root-level file still- matches under an absolute path, because the whole key is the path's tail.-- A relative path is first tried under the report's own directory and its- parents, nearest first, so `packages/web/coverage/lcov.info` naming- `src/index.ts` maps to `packages/web/src/index.ts`. A remaining tie prefers- the file nearest the report.-- Cobertura paths are joined to each `<source>` root before matching, which- separates files that share a name under different packages.--## Building a per-test map--The map needs a report that records *which test* covered each line. Two paths-exist today.--**coverage.py dynamic contexts** (the main one):+`coverable_lines`; without it a file reads as having no line data, never as 0%.++**Discovery.** With no path, `coverage add` looks for the usual locations:+`coverage/lcov.info`, `lcov.info`, `coverage.lcov`, `coverage.xml`,+`**/cobertura.xml`, `**/clover.xml`, `target/llvm-cov/**/*.lcov`,+`coverage.out`, `cover.out`, `**/target/site/jacoco*/jacoco.xml`,+`build/reports/jacoco/**/*.xml`, a repo-root `.coverage`, and a few more.+Multiple reports merge: covered and coverable lines are each unioned.++**Path matching.** Report paths are matched to indexed files by exact path,+then by the longest matching tail. A tie is reported as ambiguous, never+guessed. A relative path is tried under the report's own directory first, so+`packages/web/coverage/lcov.info` naming `src/index.ts` maps to+`packages/web/src/index.ts`. Cobertura paths are joined to each `<source>`+root. If a whole report comes back unmatched, set `coverage.strip_prefix`.+`coverage add --strict` fails when any report file did not map. The coverage+summary (`get_health`, the REST coverage route) reports how each report mapped+and whether it was measured at the indexed commit.++### Building a per-test map++The per-test map needs a report that records which test covered each line.++**coverage.py dynamic contexts** (the main path):```bashcoverage run --contexts=test -m pytestrepowise coverage add .coverage```-The `.coverage` file is read directly as read-only SQLite. Repowise decodes the-`numbits` line bitmaps itself, so it has no runtime dependency on coverage.py,-and falls back to the `arc` table when line bits are absent. Contexts look like-`tests/test_auth.py::TestLogin::test_ok|run`; the leading path becomes the test's-own file when it is resolvable.--**Per-test LCOV:** a report where each `end_of_record` block carries a distinct-`TN:` name. Blocks with a blank `TN:` are skipped. A bare suite label with no-path resolves to a test id with no test file, which is still usable for "run-these" but not for staleness reasoning.+Repowise reads the `.coverage` SQLite file directly and has no runtime+dependency on coverage.py. Contexts like+`tests/test_auth.py::TestLogin::test_ok|run` become test ids.-If you ran without contexts, `coverage add` says so explicitly rather than-silently producing an empty map.+**Per-test LCOV**: each record carries a distinct `TN:` name. Blank `TN:`+records are skipped.-Rows land in a `test_coverage` table indexed both ways (repo plus source file for-the reverse lookup, repo plus test id for the forward one), capped at 250,000-rows. The CLI reports how many were dropped if you hit the cap.+If the report has no contexts, `coverage add` says so; the per-file aggregate+is still stored. The map is capped at 250,000 rows and the CLI reports how many+were dropped.## Impacted tests-`repowise impacted-tests` diffs a change, looks up the changed *lines* in the-map, and returns the tests whose recorded coverage intersects them.+`repowise impacted-tests` diffs a change, looks up the changed lines in the+per-test map, and returns the tests whose recorded coverage touches them.```bash-repowise impacted-tests # staged changes (the default)+repowise impacted-tests # staged changes (the default outside CI)repowise impacted-tests main...HEAD # a branch or PR, from the merge-baserepowise impacted-tests main..HEAD # a plain rangerepowise impacted-tests abc123 # a single commit-repowise impacted-tests main..HEAD --format list | xargs pytestrepowise impacted-tests main...HEAD --format args --runner pytest # for CI```-| Flag | Values |-|------|--------|-| `--path` | Repo path (defaults to cwd, or the workspace primary) |-| `--staged` | Diff `git diff --cached`. Implied with no range outside CI; in CI the default is the pull request's change |-| `--format` | `table` (default), `json` (full report), `list` (test ids, one per line), `args` (runner arguments, or `:all`) |-| `--runner` | For `--format args`: `auto` (default), `pytest`, `go`, `jest`, `files` |--It always says which path fired, and it never lets a guess pass for evidence:+Each result says how it was found, and a guess never passes for evidence:| Situation | Reported as ||-----------|-------------|-| Changed file has per-test coverage on the changed lines | The exact covering tests, `via: coverage` |+| Per-test coverage on the changed lines | The covering tests, `via: coverage` || The changed file is itself a test | Itself, `via: changed-test` |-| Changed file has no coverage rows, but a test reaches it in the graph | Those test files, `via: call-graph` or `via: import-graph` (direct or transitive), in the "NOT coverage-backed" table |+| No coverage rows, but a test reaches the file in the graph | Those test files, `via: call-graph` or `via: import-graph`, in a separate "NOT coverage-backed" table || No coverage and no graph edge, but a name-shaped match | That file, `via: filename-pattern`, in the same table || None of the above | "unknown, run the full suite to be safe" || No map ingested at all | A prompt to run `coverage add` on a report with contexts |-Changed lines are matched at the commit the map was measured at (the change's-base or head); otherwise, and for deleted or removal-only files, matching is by-file and the graph is asked too. With `--format list` the caveats go to stderr-so the pipe into `pytest` stays clean. The command exits `0` in every one of-these cases, including "no tests found": it is a reporting tool, not a gate.--For CI, `--format args` prints the runner arguments, or `:all` whenever any part-of the answer is not known, with the reasons on stderr. The rules, the-`tests.*` config keys and the CI recipes are in-[CI](../start/CI.md#selecting-the-tests-a-change-needs).+With `--format list` the caveats go to stderr so the pipe stays clean. The+command exits `0` in every case: it reports, it does not gate. `--format args`+prints runner arguments, or `:all` whenever any part of the answer is unknown.+Rules, `tests.*` config keys and CI recipes:+[Selecting the tests a change needs](../start/CI.md#selecting-the-tests-a-change-needs).-A `base...head` range diffs from the merge-base of the two, so it is what a pull-request changed: commits that landed on `base` after the branch forked stay out.-`base..head` is the plain range. Every command that reads a range follows the-same rule, change risk (`repowise risk`, `get_change_risk`) included.+`base...head` diffs from the merge-base, so it is what a pull request changed.+`base..head` is the plain range. Change risk follows the same rule.## Patch coverage in CI-`repowise coverage check [REVSPEC]` is the gate: of the lines a change touched-that the report calls executable, what share did the tests run?+`repowise coverage check [REVSPEC]` gates a change on patch coverage: of the+changed lines the report marks executable, what share did the tests run?```bashrepowise coverage check origin/main...HEAD --report coverage/lcov.info --fail-under 80```-It needs git and a report, nothing else: no index, no ingest, no LLM key.-Report paths are resolved against `git ls-files` rather than an index, so a file-the change adds resolves too. Without `REVSPEC` the base comes from the CI's-pull-request variables, else the default branch; pass it explicitly in CI so the-job says what it measures. Reports come from `--report`, else `coverage.paths`,-else discovery, else (locally) the coverage an index already stores from-`coverage add`.--The same figure is on every surface, from the same computation over stored-coverage: `get_change_risk`'s `patch_coverage` block, and-`GET /api/repos/{id}/health/coverage/patch?base=&head=` (diffs `base...head`; `null`-when nothing is ingested), which the editor's branch-risk view reads. Stored-coverage measured at another commit than the change's head is marked `stale`.--The denominator is changed lines the report marks executable, so a changed-comment or blank line is neither covered nor uncovered. What counts is decided by-the report:--| Status | Meaning | In the % |-|--------|---------|----------|+It needs git and a report, nothing else: no index, no ingest, no LLM key. The+same figure appears in `get_change_risk`'s `patch_coverage` block and in the+REST patch-coverage route the editor reads.++| File status | Meaning | Counts in the % |+|-------------|---------|-----------------|| `measured` | The report names the file and some changed lines are executable | Yes || `no_coverable_changes` | The report names the file, but no changed line is executable | No |-| `not_in_report` | The report does not name the file, but it has an extension the report measures, e.g. a new source file no test loaded | No. Listed as "not in report", never 0% |-| `no_line_data` | The report names the file but gives no executable-line set for it | No |--Changed test files, and file types the report never measures (docs, config), are-out of scope and only counted. A change with no measured lines has nothing to-judge and passes.--The threshold comes from `--fail-under`, else `coverage.fail_under` in-`.repowise/config.yaml` (read at the repository root, even with `--path`); with-neither, the check reports without gating. `--min-coverable-lines` (else-`coverage.min_coverable_lines`) is the small-change tolerance: a change with-fewer changed executable lines than that, counting only lines the report-measures as the percentage does, is reported against the threshold, but a miss-reads `too_small` and does not fail. `--report` takes globs and `PATH=PREFIX`,-`coverage.ignore` leaves generated files out, and Go module paths map to their-`go.mod` directory; see-[Monorepos and matrix jobs](../start/CI.md#monorepos-and-matrix-jobs).-`coverage.gates` adds path-scoped gates: each judges the changed files its-globs match with the same rule (the small-change tolerance stays the whole-change's), a file can count in several, and a path-scoped gate that is not-informational fails the change when it fails, whatever the whole-change-figure (`path_gates` in the JSON, REST and agent output; the read-only-surfaces judge them only on coverage measured at the change's head and valid-config, see [Path-scoped gates](../start/CI.md#path-scoped-gates)).-`repowise coverage suggest-gates` proposes gates from CODEOWNERS, the-top-level packages and, when indexed, the graph's communities. The-gate compares the unrounded figure; the displayed one is floored to one-decimal, so 79.99% reads 79.9% and fails an 80% gate. Exit `0` passes (or had-nothing to judge), `1` is below the gate or a path-scoped gate fails, `2`-means the check could not run:-no report found, readable or matching a repository file, an unknown revision,-no merge-base (a shallow clone), a single commit at a shallow clone's-boundary, bad config, or not a git repository.--`--format github` writes up to 10 `::warning` annotations (riskiest file first,-then largest uncovered range), a notice counting the rest, an `::error::` for-each gate that fails and a notice when the small-change tolerance exempts one,-then appends the markdown summary to `$GITHUB_STEP_SUMMARY`. `markdown` and-`json` suit other CI systems.--### Risk of each changed file--Every row carries the risk of its file (`risk` on each file in `json`, the REST-response and `get_change_risk`'s `patch_coverage` block): `fix_pressure` (the-recency-weighted bug-fix count from git, read at the change's merge-base),-`dependents` (import fan-in from the index graph), the index's `hotspot` and-`bug_magnet` flags, `risky`, `reasons`, and `basis`. Risky reuses existing-rules: with index data for the file, a hotspot or a bug magnet; on git alone-(no index, or a file the change adds, which the index cannot know), the top-quartile of files with bug-fix history, by fix pressure. Ranking only among-files that have been fixed keeps "fixed once" from qualifying in a repository-where most files never were. `basis` is `git_and_index`, `git`, `index` (the-git walk failed) or `unavailable`. `coverage check` always reads git and adds-the index when one opens; a missing or unreadable index never fails it.--Rows, annotations and the markdown table list risky files first, then by fix-pressure, dependents and uncovered lines. The table gains a "Risk" column in-words ("hotspot, bug-fix weight 3.2, 14 dependents", "none known", "unknown"),-and a line states the basis when not every row had index data.--`--fail-under-risky PCT` (or `coverage.fail_under_risky`, validated like-`fail_under`) adds a stricter gate over the risky files' changed executable-lines. `risky` in `json` carries its covered and coverable counts, percentage,-file count and `gate`. It uses the same rule as the other gates, including the-whole change's small-change tolerance: a small change is exempt (`too_small`),-a small risky slice of a big one is not. The overall `gate` fails when the flat-gate, a path-scoped gate that is not informational, or the risky gate fails;-the flat gate's own rule is unchanged. No risky file with a changed executable-line leaves the risky gate not applied (`no_data`, exit `0`). With the flag-set, a shallow clone or a measured row whose risk could not be read exits `2`,-unless the flat, a path-scoped or the branch gate already failed, which is-then reported as the failure.--### Where to add the test--With an index, each uncovered range also says where its test belongs, naming-the test file to extend when one is found (`hints` on each measured row in-`json`, the REST response and `get_change_risk`). `hints` is `null` without an-index or when it could not be read; a measured row with no uncovered range has-an empty list; every uncovered range the index was asked about has a hint, with-basis `none` when nothing names a test. A hint names the innermost indexed-symbol containing the range (`null` outside any), up to three test files (best-first), the `basis` that found them and `total`, how many qualified before the-cap. Evidence is tried strongest first:--| `basis` | The test files are | Evidence |-|---|---|---|-| `per_test` | tests the per-test map says ran other lines of that symbol, or lines within 5 of a range outside any symbol; only when the coverage is `current` for the change | measured |-| `call_graph` | tests whose calls reach that symbol | inferred, symbol-precise |-| `import_graph` | tests that import the file | inferred, file-level |-| `none` | nothing names a test | the range needs a new one |--The index stores symbol lines at the commit it was built from. When the change-ends somewhere else (always, for uncommitted work), each file's symbols are-moved through the diff from that commit, so a range in a function added since-the index has no symbol and falls to file-level evidence.--Hints cover the first 8 ranges of each file, the ones every rendering lists.-The markdown table gains an "Extend" column ("extend tests/test_auth.py-(inferred: calls reach `login`)", "(measured: runs other lines of `login`)",-"(inferred: imports this file)", or "no test reaches this; add one"), the-terminal table the same, and each GitHub annotation ends with the same phrase.-Hints are advice: a missing or unreadable index leaves them `null` and never-changes the verdict.--### Branches on changed lines--When the report has per-line branch data, `coverage check` also reports the-branches taken on changed lines, beside patch coverage and never blended with-it, and lists lines that ran with a branch never taken as partly taken-(`branches` and each row's `partial_ranges` in `json`, the REST response and-`get_change_risk`; `null` when not measured, never 0%). Stored coverage keeps-per-line branches, so the stored path reports the same. `--fail-under-branches-PCT` (or `coverage.fail_under_branches`) gates it. See-[Branches on changed lines](../start/CI.md#branches-on-changed-lines) and the-[`coverage check` reference](../reference/CLI_REFERENCE.md#repowise-coverage-check-revspec).--### Project coverage and coverage outside the change--Patch coverage cannot see a change that deletes a test, or removes a code path-that ran other files. `--base-report PATH` takes a report measured at the-change's base commit and adds `project`: coverage at the base and the head,-gated by `--max-drop P` (or `coverage.max_drop`), and `outside_change`, the-files whose coverage changed on lines the change did not touch, each loss with-the changed files that explain it. Without a base report, an index's ingest at-the base commit gives the totals alone, as `get_change_risk` and the REST-endpoint do (never gated). See-[Project coverage and coverage outside the change](../start/CI.md#project-coverage-and-coverage-outside-the-change)-for recipes and the-[`coverage check` reference](../reference/CLI_REFERENCE.md#repowise-coverage-check-revspec)-for the fields.--A coverage.py `.coverage` database is not a text report: export it with-`coverage lcov` or `coverage xml` first.--CI checkouts are often shallow, which leaves no merge-base to diff from. Fetch-full history (`fetch-depth: 0` on GitHub Actions, `GIT_DEPTH: 0` on GitLab).-The GitHub Action, the GitLab template and the report command for each language-are in [Repowise in CI](../start/CI.md). A new file no test loads must still-appear in the report, or it reads "not in report" and is not counted; the-per-language commands there make sure it does.+| `not_in_report` | The report does not name the file (e.g. a new file no test loaded) | No; listed, never 0% |+| `no_line_data` | The report names the file but gives no executable-line set | No |++Exit `0` passes or had nothing to judge, `1` is below a gate, `2` means the+check could not run (no usable report, unknown revision, shallow clone without+a merge-base, bad config). With an index, each uncovered range also names the+test file to extend.++Everything else lives in [Repowise in CI](../start/CI.md): the GitHub Action+and GitLab template, the per-language report commands,+[path-scoped gates](../start/CI.md#path-scoped-gates),+[risk-weighted gates](../start/CI.md#risk-weighted-patch-coverage),+[branches on changed lines](../start/CI.md#branches-on-changed-lines),+[project coverage and coverage outside the change](../start/CI.md#project-coverage-and-coverage-outside-the-change),+and [monorepos and matrix jobs](../start/CI.md#monorepos-and-matrix-jobs). All+flags and JSON fields: the [`coverage check` reference](../reference/CLI_REFERENCE.md#repowise-coverage-check-revspec).## Untested hotspots-Coverage feeds the [code health](CODE_HEALTH.md) layer's test-coverage markers.-The sharpest of them is `untested_hotspot`, the textbook "write tests before you-refactor" case. It fires only when a file is all three of:--1. **A hotspot.** Flagged as one by the git layer, or 8+ commits in 90 days, or a- temporal hotspot score at or above 0.8.-2. **Centrally depended on.** At least 4 dependents. Below that, a churning file- is usually a leaf one author is iterating on, and flagging it is noise.-3. **Under-tested.** Line coverage below 40%. When no coverage has been ingested- at all, it falls back to firing only when *nothing* says a test touches the- file: no paired test file by name, and no test reaching it in the import- graph. Either signal suppresses the finding, because this is the one place- the layer asserts a negative and the bar for asserting it is no evidence at- all. The graph half is what fixed the long-standing false positive on suites- that name their tests for behaviour rather than for the file under test.--Severity is `CRITICAL` at 15% coverage or less with 10+ dependents, `HIGH` at one-of those two, `MEDIUM` otherwise. Its sibling `coverage_gap` handles the-has-coverage-but-thin case, and `coverage_gradient` applies a continuous-deduction proportional to the uncovered fraction, so a file is penalised in-proportion to how much of it is untested rather than only at a cliff.+Coverage feeds the [code health](CODE_HEALTH.md) layer. The sharpest finding is+`untested_hotspot`, the "write tests before you refactor" case. It fires only+when a file is all three of:-## From an agent+1. **A hotspot**: flagged by the git layer, 8+ commits in 90 days, or a+ temporal hotspot score of 0.8 or more.+2. **Depended on**: at least 4 dependents.+3. **Under-tested**: line coverage below 40%. With no coverage ingested, it+ fires only when *nothing* says a test touches the file: no paired test file+ by name and no test reaching it in the graph.-Two MCP tools carry test information, at two different granularities.+Severity is `CRITICAL` at 15% coverage or less with 10+ dependents, `HIGH` at+one of those two, `MEDIUM` otherwise. `coverage_gap` covers thin-but-present+coverage, and `coverage_gradient` deducts in proportion to the uncovered share.-**`get_risk(changed_files=[...])`** leads with a `directive` block whose-`test_recommendations` names tests for the changed *files*. It is file-level,-scoped to the diff, and capped at ten. Each row keeps its machine-readable-`basis`, repository identity, source files, and evidence:+Why the graph matters here: a filename convention fails both ways. On this+repository, five of the six worst bug-magnet files have no test named for them+yet are reached by 3 to 23 test files in the graph, and the sixth was paired+with a different subsystem's `test_engine.py` by basename alone.-| Row `basis` | Recommendation holds | Means |-|---|---|---|-| `measured` | test node id | the available per-test map found the test covering a changed file |-| `inferred` | test file path | the graph shows the test reaching the change; a candidate, not coverage proof |+## From an agent-`tests_to_run` is the compatibility id projection. It preserves the historical-measured-first fallback and its `measured` / `inferred` / `none` scalar domain;-new consumers use the adjacent typed rows for the additive evidence union.-Total/emitted/truncated/omitted fields cover each exact population before its-cap. `coverage_analysis` separately reports-available, unavailable, partial, degraded, and stale states; an empty list under-unavailable analysis is unknown and never means "no tests are needed".+**`get_risk(changed_files=[...])`** leads with a `directive` whose+`test_recommendations` names up to ten tests for the changed files. Each row+keeps its `basis`:++| `basis` | Holds | Means |+|---|---|---|+| `measured` | A test node id | The per-test map shows the test covering a changed file |+| `inferred` | A test file path | The graph shows the test reaching the change; a candidate, not proof |-Both forms are runnable arguments to pytest. The typed list may contain both,-but every row keeps its own basis; the legacy id list remains measured-first.+`tests_to_run` is the older flat id list, measured first. `coverage_analysis`+says whether coverage is available, partial, degraded or stale.**`get_change_risk(revspec=...)`** returns `impacted_tests`, computed from the-changed *lines*, so it is a strictly narrower and more useful set:+changed *lines*, so it is narrower:```json{@@ -464,234 +264,53 @@ changed *lines*, so it is a strictly narrower and more useful set:}```-The `line_coverage` buckets are the honest breakdown: `untested_changes` is the-strong signal (the file *is* in the map, but nothing covers the lines you-touched), `stale_test_candidates` flags covered lines whose guarding test file is-absent from the diff, and `no_coverage_data` means the file is simply not in the-map.--`get_change_risk` deliberately omits the CLI's filename-pattern guess. An agent-cannot tell a guess from real coverage, and `no_coverage_data` already reports-those files honestly.--### Self-checking before a push--`get_change_risk` without a `revspec` measures its `patch_coverage` block over-everything a push would bring, diffed from the merge-base with the branch CI-would compare against (the CI base variables, else the default branch). On a-dirty tree that is the working tree, untracked files included, labelled e.g.-`origin/main...working tree`; on a clean one it is `origin/main...HEAD`, so the-check does not shrink to the last commit once the work is committed. With no-base to find, it falls back to the plain working-tree diff (`working tree`) or-the commit scored. No commit names uncommitted code, so its freshness is by-time: `current` when the last coverage ingest came after the newest-modification of the changed files on disk, else `stale`. Ceiling: an old-report ingested after the last edit also reads current, because the report's-own run time is not stored; the hook below only ingests reports newer than the-last ingest, which keeps that case rare. Committed changes keep the commit-rule above. When changed lines are uncovered, the directive gains one-`next_actions` line: "N changed executable lines since origin/main are-uncovered; extend <the hinted tests>", or, when the coverage is stale, a line-saying to re-run the tests with coverage first.--The loop can close itself in Claude Code and Codex, opt-in per repository-with `hooks.coverage_reingest: true` in `.repowise/config.yaml`-(`REPOWISE_HOOK_COVERAGE_REINGEST=1` for one session). It costs a process-start after every shell command in that repository. After a Bash command that-runs a whole test suite, such as `pytest`, `python -m pytest`,-`coverage run -m pytest`, `go test ./...`, `npm test`, `npx vitest run`,-`jest`, `cargo test`, `cargo llvm-cov`, `mvn test` or `./gradlew test`,-passed or failed (not interrupted), the augment hook checks the aggregate-reports (`coverage.paths`, else the fixed default locations such as-`coverage/lcov.info`, `coverage.xml` and `coverage.out`). When any is newer-than the index's last coverage ingest, it starts `repowise coverage add` on-every watched report in the background and adds one line to the transcript:-"Re-ingesting coverage from coverage/lcov.info in the background (log:-.repowise/.coverage.log); get_change_risk reads it once the ingest finishes."--The ingest replaces the stored coverage, so the hook is conservative. It acts-only in a repository that has ingested coverage before, and only on a run-that does not target some tests: a path, a `::` node id, `-k`, `-run`, `-t`,-`--testNamePattern`, `--filter`, `--package`, `--workspace`, `-pl`,-`--projects`, or a short flag's argument the runner does not document as a-value (`cargo test -p x`) makes a run partial, and a partial report must not-replace full-suite coverage. It never acts on a coverage.py `.coverage`-database alone, when discovery is customised (`auto_discover: false` without-`paths`, or custom `artifacts` globs), or under `REPOWISE_DB_URL`. A-`.repowise/.coverage.queued` marker keeps a second test run from spawning-over an ingest already running for the same reports. It costs a few file-stats and one SQLite read, only after a test command, and reads only a-repo-local index. A report that only a `**` pattern finds is not watched:-name it in `coverage.paths` (single-level globs such as `coverage/*.xml` are-watched).--Claude Code runs it from repo-local entries ([what gets written-where](../agent/HOOKS.md#what-gets-written-where)); Codex only after a-successful command, so a failing run there may need `repowise coverage add`.--## The inferred tier: no coverage report needed--Everything above needs an ingest. Most repositories never do one, and the layer-used to have nothing to say to them: `tests_to_run` came back empty,-`impacted_tests` said "run the full suite", and `untested_hotspot` fell back to-matching filenames.--Matching filenames fails both ways, and this repository is the proof. Of its six-worst bug-magnet files, five have no file named for them anywhere under `tests/`-and so read as untested, while the graph names the tests that reach them:--| File | Filename convention | Graph |-|---|---|---|-| `ingestion/call_resolver.py` | nothing | 7 test files |-| `analysis/dead_code/analyzer.py` | nothing | 9 |-| `pipeline/persist.py` | nothing | 23 |-| `mcp_server/tool_answer/answer.py` | nothing | 18 |-| `analysis/pr_blast.py` | nothing | 3 |-| `analysis/health/engine.py` | `tests/unit/distill/test_engine.py` | 6, all under `tests/unit/health/` |--The last row is the worse failure. The convention matches on basename alone, so-it paired the health engine with the *distill* engine's tests (a different-subsystem) and called the file tested on the strength of a name collision.--Across the whole index the effect is large. Against the **80** standing-`untested_hotspot` findings the filename convention leaves behind, the call graph-clears **32** of them, 22 high and 3 critical, where the one-hop import walk this-tier used to run clears 11.--The graph already holds the relation. A test whose **calls reach** a source file-executes it, and that is a recorded edge rather than a name-shaped guess. It needs-no setup at all.--### What it is allowed to claim--| | Measured map | Inferred map |-|---|---|---|-| Comes from | a coverage report you ingested | the call graph, already indexed |-| Granularity | lines | files |-| Proves | this test executed these lines | this test's calls reach this file |-| Decays | yes: rows carry the commit they were ingested at, and `test_impact.coverage.freshness` reads `stale` once the index moves past it | no |-| May produce a percentage | yes | **never** |-| Labelled | `basis: "measured"` | `basis: "inferred"` |--The two are shown side by side and are never averaged into one number. The-canonical PR test-impact population evaluates both and retains their separate-evidence; compatibility surfaces remain measured-first. Other test-intelligence-surfaces may use inference only to fill measured silence. Inference's error is-one-sided and known, so it is sound as a floor ("something reaches this, do not-call it untested") and unsound as a quantity.--### The call graph leads, the import graph fills silence--The tier first shipped walking **import** edges one hop, which was the best of the-import-based options and still the wrong graph: a test that imports a module is-weak evidence it runs it. Release 0.44.0 made the call graph good enough to lead-with, so it does.--Dogfooded against a real `coverage run --contexts=test` over a slice where-per-test attribution is complete, both sides seeing the same 37 test files and 159-provably-executed production files:--**Forward, "what reaches this file" (suppresses `untested_hotspot`):**--| Walk | claims | correct | precision | recall |-|---|---:|---:|---:|---:|-| import graph, 1 hop (what shipped before) | 43 | 31 | 72.1% | 19.5% |-| call graph, 3 hops | 48 | 44 | 91.7% | 27.7% |-| **call graph, 3 hops, filtered** (ships) | 46 | 44 | **95.7%** | **27.7%** |-| both unioned | 57 | 45 | 78.9% | 28.3% |--**Reverse, "which tests do I run" (`tests_to_run`, `impacted_tests`):**--| Walk | targets | hit rate | precision |-|---|---:|---:|---:|-| import graph, 1 hop (what shipped before) | 32 | 96.9% | 94.8% |-| call graph, 3 hops, filtered | 46 | 100.0% | 97.5% |-| both unioned | 47 | 100.0% | 95.8% |-| **call graph, else import graph** (ships) | 47 | **100.0%** | **97.5%** |--The call graph wins on both axes, so it leads. **The two tiers are combined-differently per direction, because the measurements differ.** Unioning costs the-forward walk 16.8 points of precision for 0.6 of recall, and a false "something-reaches this" hides a real gap, so forward is call edges only. Reverse falls back-instead of unioning, spending the import tier only on targets the call graph left-silent, which answers one more target at identical precision.--**Depth 3 is where the call walk saturates**: 3, 4 and 5 hops return the same 48-claims and 44 confirmations, so the ceiling is the call graph's capture rate rather-than the depth. The import tier keeps its measured default of one hop.--**"Filtered" is `resolution_origin` doing work.** Not every call edge is equally-trustworthy, and this is the first thing to read the graph's own confidence-vocabulary: `global_unique`, which binds a name to the only symbol carrying it-repo-wide and which the vocabulary itself scores 0.50, is dropped. That buys **4.0-points of forward precision and 1.1 of reverse at no cost to recall**. A confidence-floor of 0.90 was tried instead and cut recall from 27.7% to 23.3%.--**Why recall reads low at 27.7%**, which is a property of the truth set rather than-the walk: `coverage` records a line as run whether a test called into it or Python-merely evaluated the module body on import, so the truth set cannot tell an-imported file from an exercised one. Splitting the 159 truth files by how much of-what ran was inside a function body: 39 ran nothing inside one at all (0% recall,-correctly), 19 ran under a quarter inside (0%), 45 a quarter to three quarters-(7%), and 56 over three quarters (**77%**). Of the 13 missed in that last group, 11-are alembic migrations the framework invokes by naming convention with no static-caller anywhere.--<sub>Reachability walks <code>EXECUTION_EDGE_TYPES</code>, which is the reachability-view minus <code>references</code> and <code>reads</code>: those record a mention-rather than a transfer of control. Dead code asks "is this used", which a mention-answers; this asks "is this run", which it does not. Both depths are keyword-arguments on the walks.</sub>--### Nothing is stored--The inferred map is read off `graph_edges` when asked and never written to-`test_coverage`. Two reasons, and the first is the one that matters:--1. **A consumer cannot mistake it for measured data.** Sharing the table behind- a marker column would make every existing reader work immediately and would- make every existing reader silently start returning inferred rows the day- someone forgot to check the marker. That exact ambiguity, "no data" read as- "not loaded", is the shape of [#1739](https://github.com/repowise-dev/repowise/issues/1739).-2. **It would be a transitive closure.** On this repository tests reach 1,630 of- 2,509 production files; materialising that is O(tests x sources) rows that go- stale the moment the graph moves, to answer a query that is a bounded- breadth-first search over rows already in the database.--The cost of deriving it is what makes that affordable. Indexing gains nothing:-the health pass computes the whole-repo answer in one multi-source walk, measured-at **27 ms** on this 3,700-file repository, once per run and cached. The-per-change walk is one `IN` query at depth 1, over the same edge table-`pr_blast` already reads.--## Empty means unknown, not "no tests"--This is the contract that makes the whole layer safe to act on, and it is worth-stating plainly: **an empty test list never means the change is untested.**--Both tools carry an explicit discriminator alongside the list:--- `get_change_risk` sets `status` to `"map_present"` only when a map exists.- With no map ingested it returns `status: "inferred"` when the graph can name- candidate test files, and `status: "no_map"`, `map_present: false` and a- summary that says "run the full suite" when it cannot. Other degraded statuses- are `no_index` (nothing indexed yet), `unknown` (the git read failed), and- `no_source_line_changes`. `basis` carries the same distinction in one word.-- `get_risk` and the REST blast-radius route consume one canonical- `test_impact` population. The directive lifts capped typed recommendations and- keeps `tests_to_run` as a non-contradictory compatibility projection. Coverage- availability, freshness, and measured-map presence remain explicit at- `pr_blast_radius.test_impact.coverage`; the legacy `guarding_tests` block is a- projection of the same rows rather than a separate derivation.--Only `status: "map_present"` with an empty `tests` list means "the map exists and-nothing in it covers this change". That is a real finding. `status: "inferred"`-is not: it is a candidate list from the import graph, and passing everything in-it does not clear a change. Everything else is an-absence of evidence, and repowise says so rather than implying a clean bill of-health. The same rule runs through the CLI ("unknown, run the full suite"), the-`no_coverage_data` bucket, and the coverage lookup helpers, which document-absence as unknown at every layer.+`untested_changes` is the strong signal: the file is in the map but nothing+covers the lines you touched. `stale_test_candidates` flags covered lines whose+guarding test is absent from the diff. `no_coverage_data` means the file is not+in the map. `get_change_risk` leaves out the CLI's filename-pattern guess.++**Self-checking before a push.** `get_change_risk` without a `revspec` measures+`patch_coverage` over everything a push would bring: the working tree+(untracked files included) or `HEAD`, diffed from the merge-base with the+branch CI compares against. Uncommitted code is `current` when the last ingest+came after the newest edit of the changed files, else `stale`. When changed+lines are uncovered, the directive adds one `next_actions` line naming the+tests to extend, or saying to re-run tests with coverage first.++**Re-ingest after a test run** (opt-in). Set `hooks.coverage_reingest: true`+in `.repowise/config.yaml` (or `REPOWISE_HOOK_COVERAGE_REINGEST=1` for one+session). After the agent runs a whole test suite (`pytest`, `go test ./...`,+`npm test`, `cargo test`, `mvn test`, `./gradlew test` and similar), the hook+re-ingests any watched report that is newer than the last ingest, in the+background, and says so in one transcript line. It acts only in a repository+that has ingested coverage before, never on a run that targets some tests (a+path, `::`, `-k`, `-run`, `--filter`, `-p` and the like), never on a+`.coverage` database alone, and never on a report only a `**` glob finds. It+costs a process start after each shell command in that repository. Claude Code+runs it from repo-local hook entries ([what gets written where](../agent/HOOKS.md#what-gets-written-where));+Codex only after a successful command.++## Empty means unknown++**An empty test list never means the change is untested.** Every surface+carries a discriminator beside the list:++- `get_change_risk` `impacted_tests.status`:++ | `status` | Meaning |+ |---|---|+ | `map_present` | A per-test map exists. An empty `tests` list here is a real finding: nothing in the map covers this change |+ | `inferred` | No map; the graph names candidate test files. Passing them does not clear the change |+ | `no_map` | No map and no graph answer. The summary says to run the full suite |+ | `no_index` | Nothing indexed yet |+ | `unknown` | The git read failed |+ | `no_source_line_changes` | No changed source lines to map |++- `get_risk` keeps coverage availability, freshness and map presence explicit+ in `test_impact.coverage`. An empty list under unavailable analysis is+ unknown, never "no tests needed".+- The CLI prints "unknown, run the full suite", and the per-file lookups report+ a file with no data as `no_coverage_data`, not as uncovered.## Configuration@@ -712,31 +331,67 @@ coverage:fail_under: 80 # patch-coverage gate for `coverage check` (0-100)min_coverable_lines: 5 # small-change tolerance for that gatemax_drop: 0.5 # most project coverage may fall from the base, in points- gates: # path-scoped gates, judged on the files they match+ gates: # path-scoped gates- {name: api, paths: ["/services/api/"], fail_under: 85}- - {name: scripts, paths: ["/scripts/"], fail_under: 50, informational: true}```Coverage is also auto-discovered and ingested during `init` and `update`, and`repowise init --coverage-report <path>` takes explicit reports (repeatable).-Note that `--coverage-report` is test coverage, while `--coverage` controls-*documentation* breadth. Two different things, similarly named.--## CLI reference--| Command | What it does |+`--coverage-report` is test coverage; `--coverage` controls documentation+breadth. Full block: [CONFIG.md](../reference/CONFIG.md).++## Accuracy and limits++Inferred tier, dogfooded on this repository against a real+`coverage run --contexts=test` over a slice with complete per-test attribution+(37 test files, 159 production files provably executed):++- **"What reaches this file"** (suppresses `untested_hotspot`): **95.7%+ precision**, 27.7% recall (46 claims, 44 correct).+- **"Which tests do I run"** (`tests_to_run`, `impacted_tests`): **97.5%+ precision**, 100% of targets answered (47 targets).+- Recall reads low because coverage counts a file as run when it was merely+ imported. Among files where over three quarters of what ran was inside+ function bodies, recall is 77%.+- Framework-invoked code with no static caller (migrations run by naming+ convention, for example) is not reached by the graph.+- Inferred answers are file-level; only measured data is line-level.+- Measured data is only as current as the last ingest; rows are tied to their+ commit and read `stale` once the index moves past it.++How these were measured and why the walk is shaped this way:+[architecture/test-intelligence.md](../architecture/test-intelligence.md).+Benchmark method across layers: [BENCHMARKS.md](../BENCHMARKS.md).++## Where it shows up++| Surface | What you get ||---------|--------------|-| `repowise coverage add [PATHS...]` | Ingest reports. Auto-discovers when no path is given, merges multiple, builds the per-test map when contexts are present. Flags: `--path`, `--format`, `--verbose` |-| `repowise coverage status` | Coverage summary plus test-to-code map counts. Flag: `--path` |-| `repowise coverage check [REVSPEC]` | Patch-coverage gate for CI, no index needed. Flags: `--report`, `--report-format`, `--fail-under`, `--min-coverable-lines`, `--fail-under-risky`, `--fail-under-branches`, `--base-report`, `--max-drop`, `--path`, `--format` |-| `repowise coverage suggest-gates` | Propose path-scoped gates for `coverage.gates` as YAML to paste below `coverage:`; writes nothing. Flags: `--path`, `--format` |-| `repowise impacted-tests [REVSPEC]` | The tests a change exercises, or the runner arguments for CI. Flags: `--path`, `--staged`, `--format`, `--runner` |+| CLI | `repowise coverage add/status/check/suggest-gates`, `repowise impacted-tests` |+| MCP | `get_change_risk` (`impacted_tests`, `patch_coverage`), `get_risk` (`test_recommendations`, `tests_to_run`), `get_health` (coverage summary) |+| Dashboard | The Tests tab on the health page; coverage findings in code health |+| CI | The GitHub Action and GitLab template run `coverage check`; `impacted-tests --format args` selects tests |+| Editor | The branch-risk view reads patch coverage |+| Hooks | Optional re-ingest after an agent's full test run |++## Reference++| Command | What it does | Flags |+|---------|--------------|-------|+| `repowise coverage add [PATHS...]` | Ingest reports; auto-discovers with no path; builds the per-test map when contexts are present | `--path`, `--format`, `--strict`, `--verbose` |+| `repowise coverage status` | Coverage summary plus per-test map counts | `--path` |+| `repowise coverage check [REVSPEC]` | Patch-coverage gate, no index needed | `--report`, `--report-format`, `--fail-under`, `--min-coverable-lines`, `--fail-under-risky`, `--fail-under-branches`, `--base-report`, `--max-drop`, `--path`, `--format` |+| `repowise coverage suggest-gates` | Propose `coverage.gates` YAML; writes nothing | `--path`, `--format` |+| `repowise impacted-tests [REVSPEC]` | The tests a change exercises, or runner arguments | `--path`, `--staged`, `--format table\|json\|list\|args`, `--runner auto\|pytest\|go\|jest\|files` |Full reference: [CLI_REFERENCE.md](../reference/CLI_REFERENCE.md#repowise-coverage).+MCP: [`get_change_risk`](../agent/MCP_TOOLS.md#get_change_risk),+[`get_risk`](../agent/MCP_TOOLS.md#get_risk).## See also-- [CODE_HEALTH.md](CODE_HEALTH.md): the coverage markers and how they deduct from the score.-- [CHANGE_RISK.md](CHANGE_RISK.md): the authoritative review percentile and supporting diff-shape score that `impacted_tests` rides alongside.-- [MCP_TOOLS.md](../agent/MCP_TOOLS.md#get_change_risk): full parameter and response reference.+- [architecture/test-intelligence.md](../architecture/test-intelligence.md): the inferred tier's walks, measurements and storage choice.+- [Repowise in CI](../start/CI.md): gates, the GitHub Action, per-language reports.+- [CODE_HEALTH.md](CODE_HEALTH.md): coverage findings and how they deduct.+- [CHANGE_RISK.md](CHANGE_RISK.md): the review signal `impacted_tests` rides alongside.- [CONFIG.md](../reference/CONFIG.md): the `coverage:` block.diff --git a/docs/agent/CHANGE_REVIEW.md b/docs/reference/CHANGE_REVIEW_API.mdsimilarity index 84%rename from docs/agent/CHANGE_REVIEW.mdrename to docs/reference/CHANGE_REVIEW_API.mdindex 9c93537dc..cd1ec49af 100644--- a/docs/agent/CHANGE_REVIEW.md+++ b/docs/reference/CHANGE_REVIEW_API.md@@ -1,8 +1,10 @@-# Change review — the Python API for reviewing one change+# Change review API++The Python API for reviewing one change.Everything the MCP tools and the CLI know about a change is computed by`repowise.core`, and that core is importable on its own. This page is for a-program that wants the whole review in one object rather than a rendered+program that wants the whole review in one object, not a renderedresponse: a bot commenting on a pull request, a CI job gating a merge, aneditor extension, a service that never has a checkout to score from.@@ -31,10 +33,10 @@ for entry in bundle.manifest:## Reviewing a change you have no checkout of-This is the case a hosted consumer actually has: the bytes arrived from an API+This is the case for a service that receives changes over an API: the bytes arrived from an APIor an artifact, and there is no repository on disk. Build a `RevisionPair`describing the change, hand over both sides' content, and the rest is-identical — same analyzers, same verdict.+identical: same analyzers, same verdict.```pythonfrom repowise.core.analysis.change_health import (@@ -70,13 +72,13 @@ source = MappingRevisionSource(head={"app/service.py": head_bytes},)-# No repo_path: there is no checkout, so the shape score is unsupported rather-# than wrong. Every lane that works from content still answers.+# No repo_path: there is no checkout, so the shape score is reported as unsupported,+# not wrong. Every lane that works from content still answers.bundle = ChangeReviewService(source).review(ChangeReviewRequest())````FileChange` carries a `diff_status` for providers that cannot always supply a-usable patch — set it to `truncated` for a size-capped API page and `binary`+usable patch: set it to `truncated` for a size-capped API page and `binary`for a blob. Without it, "no lines changed" and "we never saw the lines" lookidentical downstream, which is the one thing this type exists to prevent.@@ -91,15 +93,15 @@ carries one of five states:| State | Meaning | Reading it as "fine" would be ||---|---|---|| `available` | Computed. An empty population is a real answer. | correct |-| `partial` | Some of the requested scope was analysed. | wrong — the change is not cleared |-| `unavailable` | The data was required and could not be read. | **wrong — this is a failure** |+| `partial` | Some of the requested scope was analysed. | wrong: the change is not cleared |+| `unavailable` | The data was required and could not be read. | **wrong: this is a failure** || `degraded` | Computation failed or fell back to something weaker. | wrong |-| `unsupported` | Nobody asked, or this provider cannot answer at all. | wrong — nothing was checked |+| `unsupported` | Nobody asked, or this provider cannot answer at all. | wrong: nothing was checked |```pythonfor name in bundle.degraded_lanes:state = bundle.lane(name)- print(f"{name}: {state.state} — {state.reason}")+ print(f"{name}: {state.state}: {state.reason}")```The distinction that matters most in practice: **`unavailable` is not@@ -112,7 +114,7 @@ never render the way "these files have never broken" renders.| Lane | Field | Notes ||---|---|---|-| `manifest` | `bundle.manifest` | Every counted path, its status, diff reliability and added line spans. The one counted universe — the request's extension and exclude filters are applied here, once. |+| `manifest` | `bundle.manifest` | Every counted path, its status, diff reliability and added line spans. The one counted universe: the request's extension and exclude filters are applied here, once. || `risk` | `bundle.risk` | Diff shape ranked against the repository's recent commits. Needs a checkout or a pre-scored result. || `health` | `bundle.health` | What the change newly made worse, each finding naming its attribution basis. || `contracts` | `bundle.contracts` | Changed symbols whose callers sit outside the change, split into test and production callers. A signature change carries `signature_effect` (`compatible`, `breaking`, `unknown`) and a `signature_reason` naming the parameter, judged from the indexed parameter list and return type only; a text-only change is not reported as one. |@@ -124,7 +126,7 @@ never render the way "these files have never broken" renders.Populations are **uncapped**. Core returns everything and says how much thereis; deciding how many rows to show is a surface's job, not core's.-`bundle.directive` is the verdict, its reasons and its next actions, as data —+`bundle.directive` is the verdict, its reasons and its next actions, as data:a status, a plain factual headline, and `ReviewAction`s with stablefingerprints so a caller can tell a repeated action from a new one. It carriesno Markdown, no URLs and no tool-call syntax, because two surfaces that render@@ -137,7 +139,7 @@ it differently still have to reach the same conclusion.The service is synchronous and fetches nothing that needs a database session,a network call or a snapshot artifact. A caller collects those and hands themover; the service decides what they mean. A caller that collects nothing still-gets a bundle, with those lanes honestly marked `unsupported`.+gets a bundle, with those lanes marked `unsupported`.```pythonfrom repowise.core.analysis.change_review import ChangeReviewEvidence, ContractInputs@@ -145,21 +147,21 @@ from repowise.core.analysis.change_review import ChangeReviewEvidence, ContractIbundle = service.review(ChangeReviewRequest(revspec="HEAD"),evidence=ChangeReviewEvidence(- # Run the comparison alongside your other work rather than losing that- # concurrency to a synchronous lane.+ # Run the comparison alongside your other work and pass the result in,+ # so a synchronous lane does not cost you that concurrency.health=precomputed_delta,tests=await analyze_test_impact(session, repo_id, changed_paths),contracts=ContractInputs(graph=graph, base_parsed=base_parsed, head_parsed=head_parsed,- # True when the base side came from the last indexed snapshot- # rather than this change's real base.+ # True when the base side came from the last indexed snapshot,+ # not this change's real base.base_is_snapshot=True,),),)```-`skip={"risk"}` drops a lane you do not want the cost of. It is reported as+`ChangeReviewEvidence(skip=frozenset({"risk"}))` drops a lane you do not want the cost of. It is reported as`unsupported`, never as empty.---@@ -183,6 +185,6 @@ removing or renaming one, or changing what an existing field means, does.## Related-- The same evidence rendered for an agent: [MCP_TOOLS.md](MCP_TOOLS.md)+- The same evidence rendered for an agent: [MCP_TOOLS.md](../agent/MCP_TOOLS.md)- `repowise risk` for the terminal, which reports the change-shape score andthe independent-change split.diff --git a/docs/reference/CLI_REFERENCE.md b/docs/reference/CLI_REFERENCE.mdindex 00d8c06bc..a7d18f918 100644--- a/docs/reference/CLI_REFERENCE.md+++ b/docs/reference/CLI_REFERENCE.md@@ -73,7 +73,8 @@ Grouped by what you're trying to do, not alphabetically. `PATH` and flag details[`workspace set-default`](#repowise-workspace-set-default-alias) ·[`workspace diagnostics`](#repowise-workspace-diagnostics) ·[`workspace check`](#repowise-workspace-check) ·-[`workspace metrics`](#repowise-workspace-metrics-path)+[`workspace metrics`](#repowise-workspace-metrics-path) ·+[`workspace impacted-tests`](#repowise-workspace-impacted-tests-repopath)**Maintenance**[`doctor`](#repowise-doctor-path) ·@@ -1645,7 +1646,7 @@ repowise savings sync --dry-runMine local agent transcripts for recurring command fumbles, consecutive runsof the same base command where the first failed and a later variant succeeded(wrong tool, wrong path, unknown flag, missing argument). Report-only by-default; entirely local. See [DISTILL.md](../agent/DISTILL.md#repowise-corrections--recurring-command-fumbles).+default; entirely local. See [DISTILL.md](../agent/DISTILL.md#repowise-corrections-recurring-command-fumbles).| Flag | Description ||------|-------------|@@ -1805,6 +1806,26 @@ repowise workspace metricsrepowise workspace metrics --format json```+### `repowise workspace impacted-tests <repo:path>...`++Given changed provider files (`repo:path`), list the tests in consumer repos worth running. Each consumer call site lands in one state (measured, inferred, none, unresolved). See [Cross-repo test impact](../scale/WORKSPACES.md#cross-repo-test-impact).++| Flag | Description |+|------|-------------|+| `--path` | Workspace root (default: auto-detect from cwd) |+| `--call-depth` | Call graph walk depth, 1-8 (default: 3) |+| `--import-depth` | Import graph fallback depth, 1-3 (default: 1) |+| `--no-measured` | Exclude coverage-backed recommendations |+| `--no-inferred` | Exclude graph-inferred recommendations |+| `--min-confidence` | Minimum contract link confidence, 0.0-1.0 (default: 0.0) |+| `--target-repo` | Limit to these consumer repo aliases (repeatable) |+| `--format` | `table` (default), `json`, or `list` (one `repo:test-file` per line). `--json` is a deprecated alias |++```bash+repowise workspace impacted-tests backend:app/routers/users.py+repowise workspace impacted-tests backend:app/routers/users.py --target-repo frontend --format list+```+See [Workspaces](../scale/WORKSPACES.md) for the full multi-repo guide.---@@ -1884,9 +1905,10 @@ hook surfaces, never decisions.Manage the Distill command-rewrite hooks (Claude Code + Codex PreToolUse).When installed, noisy agent commands (tests, builds, git status/log/diff,-searches, listings) are rewritten to `repowise distill <command>`, pending-your approval by default, so the agent sees a compact, errors-first-rendering.+searches, listings) are rewritten to `repowise distill <command>` without a+prompt by default (`permission: allow`), so the agent sees a compact,+errors-first rendering. `--allow-rule` seeds a `Bash(repowise distill:*)` /+`PowerShell(repowise distill:*)` allow rule, needed only under `permission: ask`.```bashrepowise hook rewrite install # writes ~/.claude/settings.json (idempotent)diff --git a/docs/reference/COMPUTED_GLOSSARY.md b/docs/reference/COMPUTED_GLOSSARY.mdindex 0be479541..4c2e4acf3 100644--- a/docs/reference/COMPUTED_GLOSSARY.md+++ b/docs/reference/COMPUTED_GLOSSARY.md@@ -131,7 +131,7 @@ workspace overlays, MCP responses, and CLI output.| Lines added/deleted 90d | Recent churn by numstat. | `_index_file()` | `{lines_added_90d: 340, lines_deleted_90d: 87}` || Average commit size | `(lines_added_90d + lines_deleted_90d) / commit_count_90d`. | `_index_file()` | `35.6` || Merge commit count 90d | Number of merge commits touching the file recently. | `_index_file()` | `merge_commit_count_90d: 2` |-| Original path | Earliest path found through rename-follow history. | `_detect_original_path()` | `legacy/auth/session.py` |+| Original path | Earliest path found through rename-follow history. | `detect_original_path()` | `legacy/auth/session.py` || Temporal hotspot score | Exponentially decayed churn score with 180-day half-life. | `_index_file()` | `2.43` || Churn percentile | Rank percentile among indexed files by temporal hotspot score, with 90-day commits as tiebreak. | `_compute_percentiles()` | `0.88` || Hotspot flag | Top churn file: percentile >= 0.75 and has recent commits. | `_compute_percentiles()` | `is_hotspot: true` |@@ -174,8 +174,6 @@ workspace overlays, MCP responses, and CLI output.| Architecture diagram context | Top PageRank nodes, selected edges, communities, SCC groups. | `assemble_architecture_diagram()` | Mermaid graph inputs for 50 nodes and 200 edges || API contract context | Raw API contract plus endpoint/schema hints. | `assemble_api_contract()` | `endpoints: ["GET /users"]`, `schemas: ["User"]` || Infra page context | Raw infra file plus target names. | `assemble_infra_page()` | `Dockerfile`, `Makefile`, `terraform` files |-| Diff summary context | Changed files, symbol diffs, affected pages, trigger commit/diff. | `assemble_diff_summary()` | `{added_files: ["src/new.py"], affected_page_ids: [...]}` |-| Cross-package context | Monorepo boundary summary between packages. | `assemble_cross_package()` | `{source_package: "cli", target_package: "core", coupling_strength: 5}` || Dependency summaries | Summaries of already-generated dependency pages. | `assemble_file_page()` with `page_summaries` | `{ "src/db.py": "Database access layer..." }` || RAG context | Snippets from vector search for related generated pages. | `_generate_file_page_from_ctx()` | `["[file_page:src/schema.py]\nDefines API schema..."]` || Token estimate | `len(text) // 4` heuristic. | `ContextAssembler._estimate_tokens()` | `3200` |@@ -206,8 +204,7 @@ workspace overlays, MCP responses, and CLI output.| Decision record | ADR-like row from code comments, git, docs, or CLI/manual entry. | `DecisionExtractor`, CRUD, CLI | `{title: "Use Redis for sessions", status: "active"}` || Inline marker decision | Decision extracted from comments such as `WHY:`, `DECISION:`, `TRADEOFF:`, `ADR:`. | `scan_inline_markers()` | `# DECISION: cache auth sessions in Redis` || Git archaeology decision | LLM-structured decision inferred from significant commit messages with decision keywords. | `mine_git_archaeology()` | `migrate from REST client to generated OpenAPI client` |-| README-mined decision | Decision extracted from docs such as README, CLAUDE, ARCHITECTURE, DESIGN. | `mine_readme_docs()` | `"We use SQLite by default because setup should be local-first."` |-| Decision source | Provenance of a record. | `DecisionRecord.source` | `inline_marker`, `git_archaeology`, `readme_mining`, `cli` |+| Decision source | Provenance of a record. | `DecisionRecord.source` | `inline_marker`, `git_archaeology`, `adr`, `cli` || Decision confidence | Source-specific extraction confidence. | `DecisionExtractor` | `0.95` inline LLM, `0.70` git signal, `0.60` README mining, `1.0` manual || Affected files | Files linked to a decision from graph neighbors, commit files, or manual input. | `DecisionExtractor` | `["src/auth.py", "src/session.py"]` || Affected modules | Directories the affected files live in, deduped and capped; for records that name no file, the deepest directories mentioned in the decision text. | `resolve_module_nodes()` and `_infer_modules_from_text()` | `["packages/core/src/repowise/core/pipeline", "tests/unit/pipeline"]` |diff --git a/docs/reference/CONFIG.md b/docs/reference/CONFIG.mdindex 89ff597eb..f1fd9d1c8 100644--- a/docs/reference/CONFIG.md+++ b/docs/reference/CONFIG.md@@ -367,7 +367,7 @@ hooks:### The `mcp:` blockControls which tools the MCP server advertises. The default surface is curated-(11 tools in single-repo mode, plus 2 workspace-only tools in workspace mode);+(10 tools in single-repo mode, plus `list_repos` in workspace mode; seven more are opt-in);this block lets you opt extra tools in or trim the set down. The `repowise mcp--tools` / `--all` flags override it for a single launch.diff --git a/docs/reference/HOOKS_REFERENCE.md b/docs/reference/HOOKS_REFERENCE.mdnew file mode 100644index 000000000..c223dbf2f--- /dev/null+++ b/docs/reference/HOOKS_REFERENCE.md@@ -0,0 +1,73 @@+# Hooks reference++The exact entries repowise writes for its agent hooks, and how to run them by+hand. For what each hook does and when it fires, see [HOOKS.md](../agent/HOOKS.md).++## Entries and commands++| Client | Hook type | Matcher | Command | Written to |+|--------|-----------|---------|---------|------------|+| Claude Code | `SessionStart` | `startup\|resume\|clear` | `repowise-augment` (guarded) | `~/.claude/settings.json` |+| Claude Code | `PostToolUse` | `Grep\|Glob\|Read\|Edit\|Write\|mcp__.*[Rr]epowise.*__.*` | `repowise-augment` (guarded) | `~/.claude/settings.json` |+| Claude Code | `PostToolUseFailure` | `Read\|Edit\|Write\|Grep\|Glob\|NotebookEdit` | `repowise-augment` (guarded) | `~/.claude/settings.json` |+| Claude Code | `PreToolUse` (opt-in) | `Bash\|PowerShell` | `repowise-rewrite` | `~/.claude/settings.json` |+| Claude Code | `PostToolUse` (opt-in decision capture prompt) | `Bash\|PowerShell` | `repowise-augment` (guarded) | `~/.claude/settings.json` |+| Claude Code | `PostToolUse` and `PostToolUseFailure` (opt-in coverage re-ingest) | `Bash\|PowerShell` | `repowise-augment --coverage-only` (guarded) | the repo's `.claude/settings.local.json` |+| Codex | `PreToolUse` (opt-in) | the shell tool | `repowise-rewrite --agent codex` | `~/.codex/hooks.json` |+| Codex | `SessionStart` | `startup\|resume\|clear` | `repowise-augment --client codex` | the repo's `.codex/hooks.json` |+| Codex | `PostToolUse` | `Bash\|shell_command`, `apply_patch\|Edit\|Write` | `repowise-augment --client codex` | the repo's `.codex/hooks.json` |++Which switch turns each opt-in entry on:++| Entry | Switch |+|-------|--------|+| Command rewrite | `repowise hook rewrite install`, or Yes at the `repowise init` prompt |+| Decision capture prompt | `repowise decision config capture-prompt --on` ([decisions](../layers/DECISIONS.md)). The entry is per install, so one repo opting in adds a process start on shell calls for every repo on the machine; repos with the policy off return at once |+| Coverage re-ingest | `hooks.coverage_reingest: true` in `.repowise/config.yaml`, applied by the next `repowise coverage add`, `init` or `update` ([test intelligence](../layers/TEST_INTELLIGENCE.md)) |++The Claude Code plugin does not carry the coverage re-ingest entries.+`repowise uninstall` removes them.++## The presence guard++Claude Code commands marked "guarded" are written wrapped in a presence check:++```sh+if command -v repowise-augment >/dev/null 2>&1; then exec repowise-augment; fi+```++The Claude Code plugin ships these hooks independently of the CLI, so "plugin+installed, `repowise` not installed" is a supported state. A partial install+reaches it too: on Windows a running MCP server holds `repowise.exe` open, so an+installer can stop after writing only some console scripts. Without the guard,+either state prints `command not found` on every matched tool call. The guard is+POSIX (`command -v` plus `exec`), forwards stdin unchanged, and behaves+identically when the script is present. An install carrying the bare name is+rewritten on the next `repowise init`.++Codex hooks keep the bare command name, because a Codex hook command is not+documented to run through a shell.++## Other settings `init` writes++- `SessionStart` excludes `compact`: the block usually survives compaction in+ the summary, and emitting it again would double it.+- `env.ENABLE_TOOL_SEARCH=true` is set in `~/.claude/settings.json` so MCP tool+ schemas load on demand. A value you already set, including `false`, is left+ alone.+- `repowise init --no-editor-setup` (or `REPOWISE_SKIP_EDITOR_SETUP=1`) writes+ none of the Claude Code entries above and skips MCP registration.++## Entry points++`repowise-augment` is the import-isolated console script the agent hooks call;+`repowise augment` is the same engine as a CLI subcommand. `repowise-rewrite`+backs the command-rewrite hook. None of them is meant for everyday manual use.+See [CLI reference](CLI_REFERENCE.md#repowise-augment).++## Hook ledger++`repowise hook stats` and `repowise hook backfill` read and seed the local+ledger in `.repowise/sessions/sessions.db`. Flags:+[`hook stats`](CLI_REFERENCE.md#repowise-hook-stats),+[`hook backfill`](CLI_REFERENCE.md#repowise-hook-backfill).diff --git a/docs/reference/MCP_RESPONSE_FIELDS.md b/docs/reference/MCP_RESPONSE_FIELDS.mdnew file mode 100644index 000000000..6feab1eaf--- /dev/null+++ b/docs/reference/MCP_RESPONSE_FIELDS.md@@ -0,0 +1,403 @@+# MCP Response Fields++The field dictionary for every repowise MCP tool response. [MCP_TOOLS.md](../agent/MCP_TOOLS.md) says which tool to call and with what arguments; this page says what comes back and how to read it.++## Contents++- [The `_meta` envelope](#the-_meta-envelope)+- [Truncation and recovery](#truncation-and-recovery)+- [Ignored arguments](#ignored-arguments)+- [Next-call shape](#next-call-shape)+- Per tool: [get_answer](#get_answer) · [get_context](#get_context) · [get_symbol](#get_symbol) · [search_codebase](#search_codebase) · [get_risk](#get_risk) · [get_change_risk](#get_change_risk) · [get_why](#get_why) · [get_overview](#get_overview) · [get_health](#get_health) · [get_dead_code](#get_dead_code) · [list_repos](#list_repos) · [workspace tools](#workspace-tools)++---++## The `_meta` envelope++Every response carries `_meta`. Fields are present only when they carry a signal, so read absence as "nothing to report", never as an error.++| Field | When present | Meaning |+|-------|--------------|---------|+| `contract_version` | Always | Version of the response contract |+| `timing_ms` | Usually | Tool wall time |+| `hint` | Sometimes | A short follow-up suggestion. On a `get_answer` reply graded `low` while the index is behind HEAD, it says to run `repowise update` and ask again |+| `cached` | Only when `true` | The response came from cache |+| `index_age_days` | When a repository is resolved | Days since the last `repowise update` |+| `indexed_commit` | When a repository is resolved | Short SHA the index was built from |+| `live_head` | When `.git/HEAD` is readable | Short SHA of the current checkout; equal to `indexed_commit` when current |+| `index_behind` | When the live-versus-indexed comparison ran | `true` if HEAD moved, `false` if it matches. Absent means the comparison could not run |+| `stale_warning` | Only on a real signal | HEAD moved and the move changed files this response serves, or the index is very old and git is unreachable. Two commits with identical trees set `index_behind` with no warning |+| `index_scope` | When the index records it | Compact description of how the index was built: run mode, provenance, git tier, whether it is whole |+| `embedder_degraded` | When an embedder is resolved | `true` or `false` |+| `embedder`, `embedder_warning` | Only when the embedder fell back to a mock or degraded mode | Which embedder, and why |+| `newer_release` | Once per server process | A newer repowise is published; restart the MCP server after upgrading |+| `response_budget` | Always | `limit_chars` (the ceiling), `tier` (`default` or `expanded`), `serialized_chars` (size delivered) |+| `omitted` | When content was cut | `refs`, `tokens`, `restore`. See [Truncation and recovery](#truncation-and-recovery) |+| `recovery_unavailable` | When the omission store could not be written | Names the storage failure; the cut rows cannot be restored |+| `scope_hint` | `get_answer`, `get_context` | Up to three knowledge-graph layers holding none of the served paths, with file counts: areas the answer did not touch |+| `complete` | When whole units were served | Symbol bodies (bounds verified against the live file) or whole files. Do not re-read them. Sliced bodies and partial ranges are never counted |+| `completeness` | Always | `capped`, plus `shown` and `total` summed over the collections a reducing pass counted, and `reason` naming the pass that dropped most. It measures what survived reduction, not what share of the repository you hold |+| `floor` | When a count comes from walking the graph | Names those fields. Their values are lower bounds: an unresolved edge is uncounted, not proven absent |+| `state` | Only when something fired | `degraded` with `degraded_reasons`, `partial`, `truncated`. A roll-up of the response's own flags |++`list_repos`, `get_architecture`, `get_blast_radius`, `get_conformance` and `get_dependency_path` carry no freshness fields. Neither does a workspace-wide `search_codebase(repo="all")`, since there is no single indexed commit to compare.++---++## Truncation and recovery++Responses fit 24,000 serialized characters by default and 32,000 when the call passes an expansion argument (for most tools, a nonempty `include`; for `get_answer`, `include=["evidence"]`). `_meta.response_budget` reports which ceiling applied.++Content cut to fit is stored in the repo's [omission store](../agent/DISTILL.md#the-omission-store), and `_meta.omitted` says how to get it back:++```jsonc+"_meta": {+ "omitted": {+ "refs": ["a1b2c3d4e5f6"],+ "tokens": 5840,+ "restore": "`repowise expand <ref>`, or get_symbol(\"repowise#<ref>\", query=...)"+ }+}+```++- A truncated skeleton block is replaced in place by a `[repowise#<ref>: ...]` marker. Other cut content goes into one combined document per response.+- A response still over budget sheds whole blocks in an order each tool declares, cheapest loss first, and reports `truncated: true`.+- A capped list keeps its count: it gains `<key>_total`, `<key>_emitted`, `<key>_truncated`, `<key>_omitted` and `<key>_reduced_reason` siblings as they apply. Some blocks also carry an `omission_marker` for their own tail.+- Restore with `repowise expand <ref>` in a shell, or `get_symbol(symbol_id="repowise#<ref>")` from any MCP client. `query` filters the restored lines.+- Include-gated blocks are projections, not omissions: a block you did not ask for is not reported as cut.++---++## Ignored arguments++A value outside a closed vocabulary is dropped, not applied, so the response is the one you would have got without it. The tool names what it dropped:++```jsonc+"ignored_arguments": [+ { "argument": "kind",+ "values": ["unused_exports"],+ "valid": ["unreachable_file", "unused_export", "unused_internal", "zombie_package"] }+]+```++The key is absent when every argument was understood. One entry per argument.++| Tool | Arguments checked | Shape |+|------|-------------------|-------|+| `get_dead_code` | `kind`, `tier`, `min_confidence` | Top-level list, as above |+| `get_context` | `include` | Top-level list |+| `search_codebase` | `kind`, `mode` (an unknown `mode` runs as `auto`) | Top-level list |+| `get_risk`, `get_change_risk` | `include` | Top-level list |+| `get_answer` | `include` | `_meta.ignored_arguments`, a map: `{"include": [...]}` |+| `get_health` | `refactoring_*` and `performance_*` filters, `scope`, `counts` | Top-level flat map of argument to dropped value, e.g. `{"counts": "code-shape"}` |++`get_health` reports a misspelled `only` key under `unknown_only_keys`. A detail lookup (`finding_id`, `plan_id`, `opportunity_id`) also lists `scope` and `counts` there, since a population control has nothing to act on for one stored row.++---++## Next-call shape++Every suggested follow-up (`get_health`'s `fix_first[].next_call`, the refactoring and performance summaries' `next_call`, `get_risk`'s `directive.next_calls`) has one shape:++| Field | Meaning |+|-------|---------|+| `purpose` | Why to make the call |+| `mcp` | The call as text, e.g. `get_health(opportunity_id="refop3_...")` |+| `cli` | The terminal command giving the same answer, or null |+| `tool`, `arguments` | The same call, structured |++---++## `get_answer`++| Field | Meaning |+|-------|---------|+| `answer` | The synthesised answer, with file and symbol citations |+| `confidence` | `high`, `medium` or `low`: rates the prose. `high` is content-grounded and can be cited directly. A why-answer whose named mechanism is absent from the retrieved source is downgraded to `medium` |+| `retrieval_quality` | `high`, `partial` or `weak`: rates the evidence under the prose |+| `symbol_bodies` | Live bodies of the symbols the answer names. Read these before calling `get_symbol` |+| `retrieval` | Evidence rows (summary, snippet, key symbols). Shrinks as confidence rises |+| `candidates` | Ranked shortlist of files retrieval resolved, `{path, lines?}`, up to 5. Navigation, not evidence. Only with `include=["evidence"]` |+| `best_guesses`, `fallback_targets` | On low confidence: where to look, each with a one-line reason |+| `episodes` | A dated fact recorded about this checkout that bears on the question; `still_true` says how current it is |+| `degraded` | Synthesis could not run (no provider, or the call failed). The answer is assembled from retrieval and mined rationale with no LLM, and `confidence` is graded from the retrieval: `medium` unless `retrieval_quality` is `weak`, never `high` |++How much evidence rides along depends on the grade. `high` keeps one quote, and one symbol body when the answer is not already grounded. `medium` keeps one body, two quotes and the top two evidence rows. `low` keeps two bodies and the top three rows. `include=["evidence"]` skips this trimming.++Retrieval fuses three legs by reciprocal rank: full-text and vector search over wiki pages, and the structural symbol index keyed on the question's content words. The symbol leg reaches private helpers that a file page's public symbol table does not list.++---++## `get_context`++Per target, under `targets`:++| Field | Meaning |+|-------|---------|+| title, summary, symbols | Docs summary and the symbols defined, with signatures and line numbers |+| `hotspot` | Churn flag |+| `fix_history` | Only on files with counted bug fixes: count, age, `bug_magnet`. A cue to call `get_risk` |+| `episodes` | Count of dated records bound to the target; `get_why` serves their bodies |+| decisions | Titles by default. With `include=["decisions"]`, three lanes: `decisions` (accepted, governing), `candidates` (proposed, at most 3), `history` (accepted then withdrawn, at most 2) |+| `file_preview` | For a file with no indexed symbols: line and character counts, plus the heading spine for markdown or the first lines otherwise |+| `resolved_to`, `note` | A `path::Name` target whose symbol did not resolve answers with the file's card instead |+| `cross_repo` | Workspace mode: `co_changes_with` partners in other repos and contract `consumers` / `providers` |+| `*_basis` | Beside an empty `callers`, `callees` or `used_by`: the language, how many call edges the index resolved for it, and the share that are guesses. An empty list means no resolved edge, not proof of none |++Opt-in blocks: `full_doc`, `ownership` (primary owner, bus factor, contributor count), `last_change`, `callers`, `callees`, `metrics` (PageRank, betweenness, percentiles), `community`, `skeleton`, `health`, `doc_drift`.++**Skeleton.** Sliced from the index's stored symbol bounds, with no parsing at query time: every signature, the import preamble, and the bodies of the top symbols ranked by centrality, hotspot and query match. Elision markers carry 1-indexed line ranges, so you can read any part back.++**doc_drift.** `references` lists documents that name this file and still resolve to it. `documents_with_drift` says a listed document carries some assertion that no longer holds, anywhere in it, not necessarily about this file. Do not merge the two claims. `references_basis` states the limits; documents dropped by exclude rules are counted in `references_excluded`. An answer the store cannot support is a refusal: `{"unavailable": "not_computed"}` (the pass has not run yet), `"index_predates_doc_drift"`, or `"drift_read_failed"`.++**Several targets under one budget.** Each target gets an equal share, and a small target's unused share passes to the others. A target over its share drops symbols first, then its optional blocks largest first (named in `dropped_blocks`). Miss `suggestions` and ambiguity `candidates` are never shed. When even the identity cards cannot all fit, whole targets go first; the response leads with `dropped_targets` and a `recovery` call narrowed to them.++---++## `get_symbol`++| Field | Meaning |+|-------|---------|+| source | Up to about 600 lines (200 for a range read), each prefixed with its line number in `Read` format |+| start and end lines, `kind` | Exact bounds of the symbol |+| `truncated` | The body was cut; the response carries the continuation to pass back as `reference` |+| `ambiguous`, `candidates` | Several indexed symbols match (overloads, re-exports, conditional definitions). Every candidate body is returned; none is chosen silently |+| `callee_bodies` | With `depth` 2 or 3: each callee's `depth`, source and `verified` flag. Each symbol appears once, at its shallowest depth |+| `not_rendered` | Candidates or callees past the budget, each with a `fetch_with` range read |+| fallback lines | On a miss, the closest matches from a live grep |+| `source`, `created_at`, `original_tokens` | For an omission ref: where the stored content came from |++---++## `search_codebase`++| Field | Meaning |+|-------|---------|+| `results` | Ranked hits. Symbol hits: `type: "symbol"`, `symbol_id`, `name`, `kind`, `file`, `start_line`, `end_line`, `signature`, `next: "get_symbol"`. File hits: `type: "file"`, `page_id`, `file`, `title`, `next: "get_context"`. Concept hits: wiki pages with `relevance_score`, `snippet`, `target_path`, `sources` |+| `sources` | The retrievers that found a concept hit: `fts`, `vector`, or both. A hit found by `fts` alone has no semantic agreement |+| `candidates` | Up to `limit` distinct openable file paths, best first |++Symbol hits rank by exact and qualified name match, query-token coverage, then graph centrality; non-test before test unless `kind="test"`. A `symbol_spotlight` page's `target_path` is a page id (`file.py::Symbol`); open its `file`.++`results` ranks pages, and some pages are not files: a `module_page` is named by a group key that looks like a directory, an `scc_page` by a hash. `candidates` resolves symbol pages to their file, collapses several symbols of one file into one entry, skips pages that name no file, and backfills from below the result window. If the next move is a Read, read `candidates`. Decision records rank below file pages unless the query is why-shaped.++---++## `get_risk`++Per file:++| Field | Meaning |+|-------|---------|+| `hotspot_score` | 0 to 1 churn percentile |+| `health_score` | 0 to 10 |+| `dependents_count` | Direct directed structural dependents; a floor over the indexed graph |+| `co_change_partners` | Historical correlation only. Each has a recency-decayed `weight`, a `direction` (`a_to_b`, `b_to_a` or `undirected`, where `a` is the assessed file), and `conf_ab` / `conf_ba`, the share of each file's commits that touched the other |+| owners, reviewers | Primary and recent owners (`owner_pct`, `recent_owner_pct` are 0 to 1) and recommended reviewers |+| test gaps, `security_signals` | Test coverage gaps and security findings for the file |+| `resolved`, `unresolved_reason` | A target naming no indexed file: `unsupported_target_kind` (a `module:` id), `directory`, `not_indexed`, `no_such_path`. Counts are omitted, never zeroed |++Opt-in blocks: `graph` adds typed `dependents` (direct versus transitive), `consumers` (typed contract consumers only), `cross_repo_links`, `impact_surface` and `direct_risks`, with `relationship_analysis` distinguishing an empty analysis from an unavailable or partial one. `churn` adds `change_magnitude`, `risk_type` and `change_pattern`. `scales` adds the unit, range and calibration of every scalar; it is identical on every call, so ask once. A multi-target call also carries `global_hotspots`.++**PR mode** (`changed_files` passed). The response starts with `directive`:++| Field | Meaning |+|-------|---------|+| `may_break` | Production files in reverse-import reach of the change: candidates for review, not proven breakage |+| `may_break_tests` | Test files reached the same way, kept separate so tests do not crowd out production impact |+| `missing_cochanges` | Historical co-change partners absent from the change |+| `test_recommendations` | Typed rows, each with a `basis`: `measured` (the per-test coverage map found the test) or `inferred` (structural reach, not coverage proof) |+| `tests_to_run`, `tests_to_run_basis` | The test ids or files to run, and whether the list is `measured`, `inferred` or `none`. An empty list with unavailable coverage never means no tests are needed |+| `missing_tests`, `files_without_measured_tests` | Changed files with no measured test |+| `coverage_analysis`, `test_inference_analysis` | Whether each evidence source was available, stale, partial or degraded |+| `structural_impact_score` | Uncalibrated 0 to 10 structural heuristic: `localized` below 4, `moderate` 4 to below 7, `broad` 7 and up. Not a breakage probability. `overall_risk_score` is a deprecated alias |+| `next_calls` | `get_change_risk()` for the diff itself, then `get_context` on the first `may_break` files |++In workspace mode the directive also carries `will_break_consumers` (services in other repos that structurally depend on this one; structural reach only despite the name), `missing_cross_repo_cochanges`, `breaking_changes` (provider incompatibilities since the last index, with impacted consumers; a consumer link does not prove field use), `conformance_violations` and `dependency_cycles`.++Every float is rounded to 4 significant digits. Direct rows' `structural_score` values are in pagerank-weighted hotspot units and are not comparable to `get_change_risk` scores.++---++## `get_change_risk`++| Field | Meaning |+|-------|---------|+| `directive` | `status` (`review_required`, `review_recommended`, `clear_in_analyzed_scope`, `unknown`), a headline, reasons, next actions |+| `health_delta` | What the change newly made worse. See below |+| `risk_percentile`, `review_priority`, `classification`, `is_fix` | The change ranked against the repo's recent commits |+| `diff_shape` | One sentence on size and spread; never a danger verdict |+| `working_tree` | Whether uncommitted work was the subject |+| `fix_history` | Recency-weighted bug-fix record of the touched files, with `percentile` and `overlap` |+| `impacted_tests` | Tests that execute the changed lines |+| `patch_coverage` | Share of changed executable lines the stored coverage ran |+| `branch_overlap` | Other branches editing the same files |+| `independent_changes` | When the diff is several unrelated changes |+| `cross_repo` | Workspace mode: consumers and breaking changes across repos |++`include=["diagnostics"]` adds `score` (the raw 0 to 10 number), `fallback_band`, `risk_authority`, `score_measures`, `score_unit`, `baseline_sample_size`, `features` and `drivers`. `include=["scales"]` adds each field's unit, range and calibration.++**health_delta.** Both revisions are analysed from their own content, and a finding at head is reported only when the diff explains it. `scope` counts changed, eligible, analysed, skipped and failed files; `status` is `available`, `partial` or `unavailable`, and `partial` is never a clean bill. `introduced`, `worsened` and `resolved` are totals; `top_findings` holds the three most actionable, with `findings_total`. Each finding names `dimension`, `biomarker`, `severity`, `path`, `symbol`, head-side `lines`, a `reason`, and an `attribution` (`basis` of `added_lines`, `changed_symbol`, `changed_call_edge`, `new_file`, `file_change`, `context_change` or `unknown`, with a `confidence`). Identity ignores line numbers, so moved code introduces nothing. `inspect` gives the `finding_id` call that expands one; ids are bound to the two revisions. A finding that matches a stored one carries `health_reference` for `get_health(finding_id=...)`. Performance findings carry `opportunity_id` and are ordered by opportunity rank.++**impacted_tests.** `tests_to_run` names tests the per-test coverage map proves execute the changed lines, capped at ten with `total` and `truncated`. `line_coverage` buckets: `untested_changes`, `stale_test_candidates` (covered lines whose test file is absent from the diff), `covered`, `no_coverage_data`. With no coverage map, `status` is `inferred` when the import graph names test files reaching the change, or `no_map` (run the full suite). `basis` is always present: `measured`, `inferred` or `none`; `tests_to_run_kind` is `test_id` or `test_file`. Build the measured map with `coverage run --contexts=test` and `repowise coverage add`.++**patch_coverage.** Present when the index stores coverage; the same computation and JSON shape `repowise coverage check --format json` gates on. `patch_coverage_pct` is null when no changed line is executable; files the report never names read `not_in_report`. `scope.freshness` is `stale` when coverage was measured at another commit than the change's head. `path_gates` judges the path-scoped gates in `coverage.gates`. Each file row carries `risk` (`fix_pressure`, `dependents`, `hotspot`, `bug_magnet`, `risky`, `reasons`, `basis`) and up to eight `hints` per uncovered range: `range`, `symbol`, up to three `tests` to extend, and a `basis` of `per_test` (measured), `call_graph` or `import_graph` (inferred). Without a `revspec`, it covers everything a push would bring, diffed from the merge-base with the default base branch.++**fix_history.** `is_fix` applies a fixed keyword rule to the commit subject, not the conventional-commit type. `overlap` adds a diff-shape filter and needs an index: per file, `fix_count`, `overlapping_lines` (labelled `approximate`: past fix ranges are numbered against their own parent commit, so read it as "this area was patched before"), `last_fix_days_ago`, `changed_lines` and `share_of_change`. `concentration` names a file holding at least half the changed lines. The block never names the commit that introduced a bug, and is absent on an index with no fix history.++**branch_overlap.** Git-only, so it works without an index. Each `branches[]` entry has `ahead`, `behind`, `last_commit` and `files[]`; each file row states its `basis` (`same file`, or a co-change pair with its `partner`). The scan covers the newest 50 branches; `scanned`, `total` and `truncated` describe the scan, not a cap. Absent when no other branch edits a shared file, or when the scan times out.++**independent_changes.** Groups changed files by connectivity over index edges, co-change pairs, and (for a range) shared commits. Docs, config, data files and tests are never grouped. Each `groups[]` entry lists `files` and `bridging_files`; `ungrouped_files` and `summary` name what was left out and why; `basis` states what was checked. Absent when the diff is one change.++**cross_repo.** `consumers[]` names each contract link (`provider_file`, consumer `repo`, `file`, `contract_id`, `contract_type`, `match_type`) and a `tests` block (`state` of `measured`, `inferred`, `none` or `unresolved`, up to five `tests_to_run`). `breaking_changes[]` carries `contract_id`, `kind`, `severity`, `detail` and `impacted_repos`; `breaking_changes_available` says whether detection ran, so an empty list is not an all-clear. Absent outside workspace mode or when the change touches no published file.++---++## `get_why`++| Field | Meaning |+|-------|---------|+| `decisions` | Accepted records: someone accepted each in a recorded event. Treat as constraints |+| `candidates` | Inferred records nobody accepted. Hints, never rules; `candidates_note` says so. At most 3 |+| `history` | Accepted, then withdrawn. At most 2 |+| `answer_basis` | Strongest lane served: `decision`, `episode`, `rationale`, `archaeology`, `documentation`, `candidate`. Only `decision` is a ruling |+| `authority` | Per row: `accepted` or `candidate` |+| `currency` | On accepted records: `active`, `needs_review` (files moved, still binds), `uncheckable` (names nothing), `superseded`, `dismissed`. Candidates carry `review_state: "open"` instead |+| `alignment` | Path mode: `active_count`, `deprecated_count`, `uncheckable_count`, `candidate_count`, summing to `governing_count`. `score` comes from `active_count` alone |+| `git_archaeology` | Present when no decision governs the path |+| `provenance`, `evidence_refs` | Self-contained evidence references. Matching ids mean shared evidence, not independent corroboration |++Read the lane, not `status`: a record can carry `status: "active"` with no acceptance behind it. Only `repowise decision confirm` or a committed ADR marked accepted produces an acceptance.++The `candidates` and `history` lanes shed first under budget pressure, so an absent lane may mean the budget was tight. The health dashboard (no arguments) returns stale decisions, conflicts, ungoverned hotspots, `retired_decisions` and `unscoped_decisions` (five rows each, full sizes in `counts`).++---++## `get_overview`++| Field | Meaning |+|-------|---------|+| `content_md` | The overview essay's summary section; the full essay with `include=["content"]` |+| `key_modules` | Name, path, and outline section |+| `entry_points` | Detected entry points |+| `architecture` | Layer names and file counts in stack order, test layer left out |+| `code_health`, `git_health` | Repo-level health and history summaries |+| `next_actions` | Per horizon (week, quarter): the first three stored actions (`id`, `tier`, `title`, `impact`, `done_when`, `target`) with `total` and `by_tier` |+| `more` | Names the opt-in blocks |+| `workspace` | Workspace mode footer |++Opt-in blocks: `outline` (wiki page tree, two levels), `tour` (`guided_tour`, `reading_order`), `decisions` (`key_decisions`), `graph` (`community_summary`), `ownership` (`knowledge_map`). They are not computed unless named.++---++## `get_health`++### Modes and identity++`mode` is `dashboard` (no targets), `targeted`, `fix_item` (`fix_id`), or `conflict` (two selectors passed). `unresolved` names targets that matched nothing, with a reason: `not_indexed`, `no_such_path`, `excluded`, `not_measured`, `no_such_module` (which also returns `known_modules`). `mode`, `_meta`, `unresolved`, `known_modules` and every kept list's `*_total` survive any `only` projection. `scope` and `counts` are echoed on the response.++`_meta.health_analysis` labels the result as stored analysis that this call did not recompute. Its `status` is `available`, `provenance_unknown` (metrics exist but no row recorded their commit) or `unavailable`. `_meta.health_analyzed_at` and `_meta.health_analyzed_commit` date the health pass, which can lag indexing; `_meta.health_analyzed_commits_distinct` appears when rows come from several passes.++### fix_first++One ranked queue across refactoring, performance and code-shape work, the same object `repowise health` renders.++| Field | Meaning |+|-------|---------|+| `lead` | Equal to `items[0]` |+| `items` | Up to five: `id`, `tier`, `kind`, `title`, `target`, `why`, `gain`, `effort`, `confidence`, `next_call` |+| `tier` | `now` (worth doing, safe to start), `next` (worth doing, needs judgment), `later` |+| `kind` | `refactor` (one file's composed refactoring), `perf_fix` (one intervention with its sinks), `finding` (a code-shape finding with no plan) |+| `totals` | `candidates`, `eligible`, `shown`, and `excluded` counted by reason |+| `by_improves`, `basis`, `model_version`, `detail_call` | Rollup, basis, and the call that opens the lead in full |++Items are ordered by value first, then tier. History markers (churn, ownership, co-change, prior fixes) never create or lead an item; they ride on an item under `context`, and history-only files count under `totals.excluded.history_only`. `fix_first` always describes production code, whatever `scope` says; `limit` caps the items, never which leads.++### Dashboard and metric rows++| Field | Meaning |+|-------|---------|+| `kpis` | Hotspot health, `average_health` (NLOC-weighted), `average_health_unweighted`, `average_health_weighting`, pillar averages, `worst_performer_path`, `worst_test_path`, `non_code_files`, `average_health_code_only` |+| `gap_analysis` | Weighted points the average must recover to reach 8.0, files below it, `files_to_reach_target`, `files_for_half_gap`, `weighted_gap_points`, `weighted_gross_gap_points` |+| `worst_files`, `test_worst_files` | Lowest raw scores, production and test |+| `high_leverage_files` | Ranked by `weighted_deficit`, with `share_of_repo_gap_pct` |+| `top_findings` / `findings`, `test_findings` | Production and test findings; their totals sum to the whole open set |+| metric row | `score` (defect dimension and headline), `maintainability_score`, `performance_score`, `weighted_deficit` = (8 - score) x NLOC, `is_test`, `has_test_file`, `primary_biomarker` |++`weighted_deficit` and the gap fields share one unit, score points x NLOC; they compare only with each other. Non-code files score a mechanical 10.0 because no marker reads them; `average_health_code_only` shows the headline without them. `primary_biomarker` prefers a discrete finding over the continuous `coverage_gradient`. Each finding carries a `dimension`; `advisory` findings (`assertion_free_test`, `mock_saturated_test`) never deduct and appear only with `include=["advisory"]`.++Provisional finding types appear only with `include=["unverified"]`, each marked `verification: "unverified"`. Targeted mode is never split into production and test.++### Opt-in blocks++| `include` | Adds |+|-----------|------|+| `biomarkers` | Findings in dashboard mode |+| `refactoring` | `refactoring_opportunities`, `suggestion_legend` |+| `trend` | Snapshot diff and declining or predicted-decline alerts |+| `coverage` | Stored coverage rows and summary |+| `accuracy` | `defect_accuracy`: of the K least-healthy files, how many were recently bug-fixed versus the base rate (precision@K, `lift`). `null` on repos with too little history |+| `signals` | Targeted mode: per-file prior defects, change scatter, 90-day churn, owners, graph degree. `null` per missing field, never an imputed zero |+| `churn_complexity` | One point per recently changed file: commits, max CCN, NLOC, score, churn percentile |+| `doc_drift` | `findings` naming the document to edit, its line, the `target` it wrongly claims, `reason`, `kind`, `confidence`, optional `suggestion`; plus `findings_total`, `documents`, `basis`. A finding is evidence to check, not a proven defect |+| `semantics` | `_meta.health_semantics`: the legend for deficit points and percentiles |+| `performance`, `defect`, `maintainability`, `advisory` | Filter findings to one dimension; `performance` also adds the performance queue |++`only` aliases: `biomarkers` -> `findings`, `accuracy` -> `defect_accuracy`, `refactoring` -> `refactoring_plans`. Dimension names have no single key and land in `unknown_only_keys`; for `signals` in targeted mode, name `metrics`.++### Refactoring++`refactoring_opportunities` holds one composed unit per file: `lead_biomarker`, `addresses_primary_problem` (`true`, `false`, or `null` when no dominant finding was recorded), ordered `steps` each with a `mechanical` or `judgment` applicability, and evidence counts. With no targets it lists what Fix first would take (`refactoring_scope: fix_first`); `refactoring_opportunities_hidden` counts the rest by reason and `refactoring_opportunities_scope` names the scope. A step with `relocated_by` refers to a symbol an earlier step moves: locate it again before applying.++`get_health(opportunity_id="refop...")` returns full steps, member plans, the validation profile and `next_actions`; `only=["refactoring_evidence"]` with `cursor` pages the evidence. `found` says whether the id matched; `status` is `open`, `acknowledged`, `resolved` or `false_positive`. `refactoring_summary` rolls up by type, effort, confidence and lifecycle.++`refactoring_plans` is the raw per-detector list, returned only when named in `only`: `extract_class` (cohesion `groups`), `extract_helper` (clone `occurrences`, `suggested_site`), `move_method`, `break_cycle` (`cut_edges`). Each carries `evidence`, `impact_delta`, `effort_bucket`, `blast_radius`, `file_weighted_deficit` and an `id` for `generate_refactoring_code`. An empty requested list has `refactoring_plans_status.reason`: `no_applicable_findings`, `plan_analysis_indeterminate`, `no_eligible_targets` or `analysis_unavailable`. See [REFACTORING.md](../layers/REFACTORING.md).++### Performance++Performance findings carry `health_impact: 0`; the pillar never blends into the score. They rank on an integer `perf_rank`, which weighs the marker (superlinear above per-iteration crossings above in-loop CPU), the boundary crossed (`details.boundary_kind`: subprocess, then network or database, then lock, then filesystem) and whether the loop crosses a function call (`details.cross_function`). It is an ordering key, not a score.++A bare `get_health()` ranks performance work inside `fix_first` as `perf_fix` items. `get_health(opportunity_id="perf...")` adds `plan_steps`, `validation` and `siblings`; `only=["performance_evidence"]` with `cursor` pages the evidence. Ids are stable within a performance model version; an id from an older model resolves to `model_state.state: "stale_model"` with `refresh_required`. `lifecycle_status` is `open` or `resolved`. Evidence rows carry a public `finding_id`.++### Coverage++`include=["coverage"]` returns per-file rows (`covered_line_count`, `total_coverable_lines`; targeted mode adds `covered_lines`) and a `summary`. `freshness.status` is `current` (measured at the indexed commit), `stale` or `unknown`. `report_paths` says how the report's entries mapped at ingest (`total`, `matched`, `unmatched`, `ambiguous`, `unmatched_sample`). `source_formats` lists merged formats. Dashboard mode adds `history`, one point per complete ingest, newest 10. Targeted rows add `decay`: `confirmed_lines`, `invalidated_lines` (now unknown, not uncovered), and `drifted` once a fifth of a file's measurement has moved.++---++## `get_dead_code`++| Field | Meaning |+|-------|---------|+| `tiers` | Findings by tier: `high` (0.7 and up), `medium` (0.4 to 0.7), `low`. Each has path, kind, confidence, line count, cleanup impact and a stable `id` |+| `summary` | Totals over every open finding; `filtered_findings` counts what the filters kept; `filters` echoes non-default filters; `scope` says which is which |+| `summary.withheld_types` | Provisional finding types held back; ask with `kind=` to see them, marked unverified |+| `summary.call_resolution_basis` | Per language, how many call edges the index resolved and the share that are guesses: the graph the findings rest on |+| `by_directory`, `by_owner` | Rollups with `group_by` |+| `limit_note` | A `limit` over 25 was clamped |++The 0.7 and 0.4 tier floors are shared by the CLI, the web UI and this tool. See [DEAD_CODE.md](../layers/DEAD_CODE.md).++---++## `list_repos`++`workspace` (bool), `workspace_root`, `default_repo`, and per repo `alias`, config-relative `path` and `absolute_path`. In single-repo mode: `workspace: false`, `workspace_root: null`, `default_repo: "default"`.++---++## Workspace tools++**`get_blast_radius`.** Impacted services with `score` (0 to 1 relative path weight, uncalibrated), `distance` (hops), `structural` (a real dependency versus co-change only) and edge kinds; `impact_score_semantics`, `impacted_repos`, `structural_count`, `behavioral_count`, `total_impacted`, unresolved targets. `symbol_targets[].consumers[]` names consuming symbols across each contract link, each with a `tests` block shaped like `get_change_risk`'s.++**`get_architecture`.** `score` (1 to 10, higher means lower coupling and a smaller core), `architecture_type` (`core-periphery` or `hierarchical`), `propagation_cost_pct`, `core_size`, `core_ratio`, `core_members`, `cycle_count`, `conformance_violations`, `role_breakdown` (Core, Shared, Control, Peripheral counts), `summary`.++**`get_conformance`.** `violations` (offending `source` and `target` services, the `rule_source` / `rule_target` matchers that fired, `edge_kind`), `cycles` (`nodes`, `length`), `violation_count`, `cycle_count`, `rules_evaluated`. A report that never ran says so instead of returning zero counts.++**`get_dependency_path`.** The path when one exists; otherwise nearest common ancestors, shared neighbours, community analysis and bridge suggestions.++**`get_execution_flows`.** Scored entry points, each with a breadth-first call trace and whether it crosses community boundaries.++**`generate_refactoring_code`.** `suggestion_id`, `plan`, the generated code and unified diff, or `generation: {available: false, reason: "disabled"}`, or `error: "no_provider"`.++**`set_finding_status`.** The new status and its timestamp.diff --git a/docs/reference/UPGRADING.md b/docs/reference/UPGRADING.mdindex 7524afbd0..51760d434 100644--- a/docs/reference/UPGRADING.md+++ b/docs/reference/UPGRADING.md@@ -33,6 +33,19 @@ It is a recommendation, not a requirement. Nothing is wiped until you choose to## Behaviour changes worth knowing about+### 0.54.0: health re-scores once, and some git numbers move++**Your first `update` re-scores the whole repository.** The health analyzer version moved from 31 to 36, so the first update after upgrading re-scores every file and takes longer than usual. The parser schema moved from 3 to 5 (Python import resolution and overload declarations), so that same update re-parses files instead of reading them from the parse cache. Nothing needs re-indexing: the store format is unchanged.++**Some scores and counts change on that first update, by design:**++- History windows are measured from the indexed commit, not the wall clock, so re-indexing an unchanged tree on a later day gives the same scores. Set `REPOWISE_GIT_WINDOW_ANCHOR=now` to keep the wall-clock behaviour.+- A file with no code finding loses at most one point to git history, and `hidden_coupling` is still listed but no longer lowers the score. Files that were held down only by history can score higher.+- "Commits" counts non-merge commits everywhere, and one person's noreply address, handle and local email merge into one contributor. Commit and contributor counts can go down.+- Docs, config, workflow and data files gain commit history on the first update. They stay out of hotspot rankings.+- Entry points are read from the package manifest first, so the entry-point list and the dead-code findings that depend on it can shift.+- The full-text search index is widened and refilled the next time the index is opened. It happens once and needs no action.+### 0.53.0: health re-scores once, and workspace contracts re-extract**Your first `update` re-scores the whole repository.** The health analyzer version moved from 21 to 31 over this cycle: performance plans now claim only what the loop proves, two new performance markers landed, and Pascal gained loop, performance and assertion coverage. Each of those changes how a stored health number is computed, so the first update after upgrading re-scores every file and takes longer than usual. The next one is back to normal.diff --git a/docs/reference/WORKSPACE_CONTRACTS.md b/docs/reference/WORKSPACE_CONTRACTS.mdnew file mode 100644index 000000000..17999a331--- /dev/null+++ b/docs/reference/WORKSPACE_CONTRACTS.md@@ -0,0 +1,118 @@+# Workspace contract matching reference++How a [workspace](../scale/WORKSPACES.md) names, reads and matches cross-repo+contracts, for when you need to know why a link did or did not form. Start from+[Extraction Diagnostics](../scale/WORKSPACES.md#extraction-diagnostics), which+names the reason for each unmatched consumer. The support matrix of frameworks+and libraries is in [API Contract Extraction](../scale/WORKSPACES.md#api-contract-extraction).++## Sockets++An endpoint is identified by its path (`socket::/hubs/game`); a message by its+event within a scope, `socket::<scope>#<event>`, where the scope is a socket.io+namespace (`/` unless the file names one) or a broadcast channel. Channels keep+the wire prefix Pusher gives them (`private-orders.{param}`), and Echo's event+names are read the way Echo formats them, so `.listen('OrderShipped')` meets a+Laravel event class `App\Events\OrderShipped` and `.listen('.order.shipped')`+meets `broadcastAs()` returning `order.shipped`. The emitting side is the+provider. socket.io, amqplib's `publish`/`consume`, BullMQ, Redis and NestJS+calls are read only in a file importing the library, since `emit`, `on`,+`publish` and `subscribe` are common method names.++## Topics and queues++A topic, queue or exchange name is read the way a URL is: a literal, or a name+the same file assigns exactly once to a literal (Python, Go, Java, PHP class+constants, and JS/TS `const` including object and enum members such as+`QUEUES.ticketSold`), is resolved; a name built at runtime is skipped, not+guessed. An SQS queue URL is named by its last path segment and an SNS+topic ARN by its last field. A subscription by pattern (NATS `orders.*` /+`orders.>`, Redis `psubscribe`, Kafka `topicPattern`) links to every publisher+whose name it matches. A Laravel job dispatched without a queue runs on the+queue its class declares (`public $queue`, `$this->onQueue()`, `viaQueue()`),+found by the class's fully qualified name; one left on the connection's+default queue is not recorded, since every app has one. RabbitMQ+publishers name an exchange and a routing key while consumers name a queue, so+a queue binding found anywhere in the workspace connects the two: each consumer+of the bound queue links to the exchange's publishers whose routing key the+binding pattern accepts (`*` and `#` follow topic-exchange rules; an empty key+or pattern matches everything, while a key the source does not settle matches+nothing). Such a link carries the exchange as its `contract_id` and the queue as+`consumer_contract_id`. When no consumer of the bound queue is found, the+binding site itself is linked. A publish to the default exchange+(`publish('', 'jobs')`) is a publish to the queue `jobs`.++## Data contracts++Data/DB contracts use the id scheme `data::<table>` and render as a `db` edge in the [system graph](../scale/WORKSPACES.md#system-graph). An ORM model with no explicit table name takes its library's default: the Prisma model name, the TypeORM class name in snake case, the Sequelize model name pluralized, the Eloquent class name in snake case and plural. A service that only models a table (an ORM class, no migration) is linked to the service whose migration or DDL defines that table's schema, when exactly one service defines it; two services that each migrate a table of one name are read as separate databases and not linked. The consumer side (SQL string matching) is heuristic and lower-confidence than the ORM-based providers; unlike HTTP and gRPC, there is no field-level breaking-change diffing for data contracts, only table/route-level removal.++## HTTP routes and calls++HTTP routes are matched on their **full** path: a router mount prefix+(`APIRouter(prefix=...)`, `include_router(prefix=...)`, Express `app.use('/x', router)`,+Go route groups, Laravel `Route::prefix(...)->group(...)` and `Route::group(['prefix' => ...])`)+is stitched onto each handler path before matching. Laravel's `routes/api.php` is+served under `/api` unless `bootstrap/app.php` (`apiPrefix`) or a route provider says+otherwise, and `Route::resource` / `apiResource` expand into the routes they register. A NestJS+route is served at the app's `setGlobalPrefix` (unless its `exclude` list names the route), then+its URI version (`enableVersioning`, `@Version`), then the `@Controller` prefix. A call through an+axios, ky, got or ofetch instance is read with the instance's `baseURL` / `prefixUrl`, also when+another file imports the instance. An Angular `HttpClient` call is read on any receiver typed+or injected as `HttpClient`, with `environment.apiUrl` folded from `src/environments/` and class+fields built on it; a base an interceptor prepends is not read. A client call+whose base URL is an unresolved placeholder (`fetch(\`${API_BASE}/users\`)`) matches+on the host-relative path; the link is **exact** when exactly one workspace service+provides that path and a lower-confidence **candidate** when the target is ambiguous.++## Contracts over REST++`GET /api/workspace/contracts` lists contracts and links,+filterable by `contract_type`, `repo` and `role`. Each contract carries its+line, its ingestion symbol id and the extractor's `meta`; each link carries both+symbol ids and both service boundaries. The request/response `schema` is not on+the list, because it runs to full inline type declarations and only one is ever+needed at a time: fetch it with+`GET /api/workspace/contracts/detail?repo=<alias>&file=<path>&id=<contract-id>`,+which returns that one contract with its schema, its links, and its unmatched+reason. All three parameters are required, since a contract id alone is not+unique across repos.++## Package matching++npm and composer dependencies match in two ways: a local path (`file:` specs,+workspace globs, composer `path` repositories) that resolves into a sibling repo,+or a package name that exactly one sibling repo publishes. For npm, a repo+publishes its root `package.json` and its declared workspace members, so a+vendored or fixture `package.json` never claims a name. For composer, every+`composer.json` up to three directories deep counts, outside `vendor/`, hidden+directories, tests, fixtures and examples, which covers split packages such as+`src/Illuminate/Support/composer.json`. A name two repos publish links to+neither.++Maven matching is filesystem-only and coordinate-based. Repowise resolves local+reactor modules, local parents, properties, and dependency-management versions,+then links an active direct compile/runtime dependency only when exactly one+selected workspace project publishes that `groupId:artifactId`. Test, provided,+system, optional, profile-only, ambiguous, and external dependencies do not create+production package edges. Bounded diagnostics retain the reason for Maven+non-matches. When a repository has a root `pom.xml`, only that declared reactor is+eligible; unrelated nested example or fixture POMs are not treated as producers.++This does **not** execute Maven, read user settings, download artifacts, resolve+plugins/transitive dependencies/imported BOMs, or infer generated sources. A Maven+package edge is also not a published symbol-level code API or a runnable Maven+target recommendation; those capabilities are reported separately and remain+unsupported.++## OpenAPI comparison boundary++The [breaking-change guard](../scale/WORKSPACES.md#breaking-change-guard) diffs+OpenAPI schemas within this boundary.++OpenAPI comparison covers the common supported subset of `3.0.x`, `3.1.x`, and `3.2.x`: JSON/YAML documents; path/query/header/cookie parameters using OpenAPI's default `style`, `explode`, and `allowReserved` behavior; one `application/json` request body; exactly one explicit JSON 2xx response; recursive objects and arrays; exact primitive types; requiredness; normalized nullability; finite homogeneous scalar enums; and bounded same-document JSON Pointer references. Request rules describe values the provider accepts; response rules describe values consumers may receive, so requiredness, nullability, enum values, and constrained/unconstrained enum transitions reverse between the two sides. Operation removal remains the transport-neutral contract rule.++Only complete sides with the same schema source and comparison-fidelity key are field-diffed. Unsupported/unresolved nodes, extraction-strategy changes, and response-selection changes become warning-level uncertainty, never shortened schemas or false removals. Remote/cross-file references, composition, additional-properties semantics, arbitrary JSON Schema constraints, non-JSON or ambiguous media, multiple materially different success responses, and OpenAPI 2.0 are outside this boundary. No network dereferencing occurs.++## Test impact over REST++`GET /api/workspace/test-impact?repo=<alias>&file=<path>`. Repeat `file` for several changed files in the same repo. The response carries the same fields the `--format json` output has: `recommendations` (each with its consumer repo, the `consumer_files` and bound `consumer_symbol_ids` that reached the test, `basis`, `via`, `confidence` and the contract ids that produced it), `unresolved` with a reason per link, `files_analyzed` with the state each landed on, and the `summary` counts. Over the API an empty answer's `summary.reason` is `no_contract_data`, `no_matching_links` or `lookup_failed`.diff --git a/docs/scale/AUTO_SYNC.md b/docs/scale/AUTO_SYNC.mdindex 338c26ce9..711ad157e 100644--- a/docs/scale/AUTO_SYNC.md+++ b/docs/scale/AUTO_SYNC.md@@ -1,298 +1,212 @@-# Keeping Your Wiki in Sync+# Keeping the index fresh-repowise can automatically update your wiki whenever code changes. Choose the-method that fits your setup.+repowise answers from its index, so the index should follow your code. Pick the+method that fits your setup; they can be combined.| Method | Best for | Requires server? ||--------|----------|-----------------|| [Post-commit hook](#1-post-commit-git-hook-recommended) | Solo developers, local repos | No |-| [File watcher](#2-file-watcher) | Local development, rapid iteration | No |-| [GitHub webhook](#3-github-webhook) | Teams, CI/CD, hosted repos | Yes |-| [GitLab webhook](#4-gitlab-webhook) | Teams, CI/CD, hosted repos | Yes |+| [File watcher](#2-file-watcher) | Local development, uncommitted work | No |+| [GitHub webhook](#3-github-webhook) | Teams, a shared repowise server | Yes |+| [GitLab webhook](#4-gitlab-webhook) | Teams, a shared repowise server | Yes || [Polling fallback](#5-polling-fallback) | Safety net for missed webhooks | Yes |+Whichever you use, agents are told when the index is behind: the SessionStart+hook reports indexed commit vs `HEAD` ([HOOKS.md](../agent/HOOKS.md)), and MCP+responses carry a `stale_warning` when a file they served has changed.+---## 1. Post-Commit Git Hook (Recommended)Runs `repowise update` in the background after every local commit. Your-terminal is never blocked. This is the recommended auto-sync method for-local development.--### Setup+terminal is never blocked.-`repowise init` installs the hook by default (interactive runs ask first; `--no-hook` skips it) at the end-of the setup wizard. You can also install or manage it manually:+`repowise init` installs the hook by default (interactive runs ask first;+`--no-hook` skips it). To manage it yourself:```bash-repowise hook install # install for current repo-repowise hook install --workspace # install for all workspace repos-repowise hook status # check if installed-repowise hook uninstall # remove the hook+repowise hook install # current repo+repowise hook install --workspace # every workspace repo+repowise hook status # add --workspace to check every repo+repowise hook uninstall # add --workspace to remove from every repo```-The hook is marker-delimited, so it safely coexists with other tools'-hooks (linters, formatters, etc.) in the same `post-commit` file.--### What happens+The hook is marker-delimited, so it coexists with other tools' hooks in the+same `post-commit` file.-1. You run `git commit`-2. The hook fires in the background (non-blocking)-3. `repowise update` diffs the new commit against the last synced commit-4. Only affected pages are regenerated (typically 3-10 for a small commit)-5. Graph, git stats, dead code, and decisions are also refreshed--### Check hook status--```bash-repowise hook status-repowise hook status --workspace # check all repos in workspace-```--### Remove the hook--```bash-repowise hook uninstall-repowise hook uninstall --workspace-```+After each commit, `repowise update` diffs the new commit against the last+synced one, refreshes the graph, git signals, dead code, health and decisions,+and regenerates the wiki pages the change affects (when the repo was indexed+with docs). See [How updates work](#how-updates-work).---## 2. File Watcher-Watches your working directory for file saves and auto-updates. Useful during-active development when you want the wiki to stay current without committing.+Watches the working directory and updates on save. Use it when you want the+index current without committing.```bash-repowise watch # watch current directory+repowise watch # current directoryrepowise watch /path/to/repo-repowise watch --debounce 5000 # wait 5s after last change (default: 2s)+repowise watch --debounce 5000 # wait 5s after the last change (default: 2s)repowise watch --index-only # no model calls per save+repowise watch --workspace # every workspace repo```Press `Ctrl+C` to stop.-Unlike the post-commit hook, the watcher indexes the working tree: staged,-unstaged and untracked files all reach the index without a commit.--Ignored, so a build or an update of repowise's own files never triggers a run:-`.repowise/`, `.git/`, `node_modules/`, build output and other blocklisted-directories, lockfiles and non-source files, and the files repowise manages-itself (`CLAUDE.md`, `AGENTS.md`, `.mcp.json`, `.claude/`).+The watcher indexes the working tree, so staged, unstaged and untracked files+all reach the index without a commit. It ignores `.repowise/`, `.git/`,+`node_modules/`, build output and other blocklisted directories, lockfiles and+non-source files, and the files repowise manages itself (`CLAUDE.md`,+`AGENTS.md`, `.mcp.json`, `.claude/`), so an update never triggers itself.On a repo indexed with docs, each trigger regenerates the changed files' pages-with a model. Use `--index-only` to keep the index current for free and leave-the prose for a later `repowise update`.+with a model. `--index-only` keeps the index current at no model cost and+leaves the prose for a later `repowise update`.---## 3. GitHub Webhook-For teams or hosted deployments. GitHub sends a push event to your repowise-server, which triggers an incremental update.--### Prerequisites--- repowise server running and reachable from the internet (or GitHub's network)-- A webhook secret (**strongly recommended** for any server exposed beyond localhost)+For a shared repowise server. GitHub sends a push event to the server, which+runs an incremental update.-### Start the server+You need a server reachable from GitHub, and a webhook secret for any server+exposed beyond localhost.```bash-repowise serve # http://localhost:7337+repowise serve # http://localhost:7337repowise serve --host 0.0.0.0 --port 8080 # custom bind-```--### Configure the webhook secret--```bashexport REPOWISE_GITHUB_WEBHOOK_SECRET="your-secret-here"```-### Add the webhook in GitHub--1. Go to your repo **Settings > Webhooks > Add webhook**-2. **Payload URL:** `https://your-server.example.com/api/webhooks/github`-3. **Content type:** `application/json`-4. **Secret:** the same value you set in `REPOWISE_GITHUB_WEBHOOK_SECRET`-5. **Events:** select **Just the push event**-6. Click **Add webhook**+In your GitHub repo, open **Settings > Webhooks > Add webhook**:-### Verify it works+1. **Payload URL:** `https://your-server.example.com/api/webhooks/github`+2. **Content type:** `application/json`+3. **Secret:** the value of `REPOWISE_GITHUB_WEBHOOK_SECRET`+4. **Events:** **Just the push event**-Push a commit and check the webhook delivery in GitHub (Settings > Webhooks >-Recent Deliveries). You should see a `200` response with:--```json-{- "event_id": "...",- "status": "accepted"-}-```+Push a commit and check **Recent Deliveries**. A working setup returns `200`+with `{"event_id": "...", "status": "accepted"}`.### Security (GitHub)-When `REPOWISE_GITHUB_WEBHOOK_SECRET` is set, every incoming GitHub request must-carry a valid HMAC-SHA256 signature in the `X-Hub-Signature-256` header.-A missing `sha256=` prefix or a signature that doesn't match is rejected with-`401 Unauthorized`.+With `REPOWISE_GITHUB_WEBHOOK_SECRET` set, every request must carry a valid+HMAC-SHA256 signature in `X-Hub-Signature-256`. A missing `sha256=` prefix or a+mismatched signature gets `401 Unauthorized`.-When the secret is **not** set, the server falls back to a dev-mode check:-requests from a local/loopback caller are accepted, and everything else is-rejected with `403 Forbidden`. This keeps `repowise serve` usable out of the-box for local testing, but it means an unset secret is not a safe posture for-a server reachable from the internet or GitHub's network. Always set-`REPOWISE_GITHUB_WEBHOOK_SECRET` before exposing the server.+Without the secret, the server accepts only local (loopback) callers and+rejects everything else with `403 Forbidden`. That keeps `repowise serve`+usable for local testing, but it is not safe for a server reachable from the+internet. Set the secret before exposing the server.---## 4. GitLab Webhook-Same concept as GitHub, different endpoint and auth mechanism.--### Configure the webhook token+Same idea as GitHub, with a different endpoint and token check.```bashexport REPOWISE_GITLAB_WEBHOOK_TOKEN="your-token-here"```-### Add the webhook in GitLab+In your GitLab project, open **Settings > Webhooks**:-1. Go to your project **Settings > Webhooks**-2. **URL:** `https://your-server.example.com/api/webhooks/gitlab`-3. **Secret token:** the same value you set in `REPOWISE_GITLAB_WEBHOOK_TOKEN`-4. **Trigger:** enable **Push events**-5. Click **Add webhook**+1. **URL:** `https://your-server.example.com/api/webhooks/gitlab`+2. **Secret token:** the value of `REPOWISE_GITLAB_WEBHOOK_TOKEN`+3. **Trigger:** **Push events**### Security (GitLab)-When `REPOWISE_GITLAB_WEBHOOK_TOKEN` is set, the server compares it against-the `X-Gitlab-Token` header using constant-time comparison. Invalid tokens are-rejected with `401 Unauthorized`.--When the token is **not** set, GitLab behaves exactly like GitHub above:-local/loopback callers are accepted as a dev convenience, and every other-caller is rejected with `403 Forbidden`. Set the token before exposing the-server.+With `REPOWISE_GITLAB_WEBHOOK_TOKEN` set, the server compares it with the+`X-Gitlab-Token` header in constant time and rejects a mismatch with+`401 Unauthorized`. Without the token, only local callers are accepted and every+other caller gets `403 Forbidden`, as with GitHub.---## 5. Polling Fallback-When the server is running, a background job polls all registered repositories-every 15 minutes. If new commits are detected that weren't caught by webhooks-(e.g. webhook delivery failed), it triggers an incremental update automatically.--No configuration needed -- this runs automatically when you start the server-with `repowise serve`.+While `repowise serve` runs, a background job polls every registered repository+every 15 minutes. New commits a webhook missed trigger an incremental update.+No configuration is needed.----## How Updates Work--There are two update paths, depending on the trigger:--### CLI incremental update (`repowise update`)--1. **Diff** -- compare the new HEAD against the last synced commit-2. **Full graph rebuild** -- re-traverse and re-parse all files to rebuild- the dependency graph (needed for cascade analysis)-3. **Incremental git re-index** -- re-index git metadata only for changed files,- then recompute percentiles across the whole repo-4. **Dead code analysis** -- run partial dead code detection on changed files- and their graph neighbors-5. **Decision re-scan** -- re-scan changed files for inline decision markers-6. **Detect affected pages** -- find directly changed files, cascade to 1-hop- importers, capped by an adaptive budget (auto-scaled based on change size)-7. **Regenerate** -- call the LLM only for affected pages-8. **Cross-repo hooks** -- if the repo is part of a workspace, re-run cross-repo- co-change and contract analysis-9. **Persist** -- update the database, search index, and record a GenerationJob-10. **Save state** -- record the new HEAD as the last synced commit--A typical single-commit update touches 3-10 pages and completes in under a-minute.--### Server sync (`POST /api/repos/{id}/sync`)--1. **Full re-traverse** -- re-parse all files and rebuild the dependency graph-2. **Full git re-index** -- re-index git metadata for all files-3. **Dead code analysis** -- full dead code scan-4. **Decision extraction** -- full decision marker scan-5. **Persist** -- update the database-6. **Save state** -- record the new HEAD--Server sync does **not** regenerate wiki pages (no LLM cost). Use-**Full Re-index** (`POST /api/repos/{id}/full-resync`) to also regenerate-all wiki pages.+## How updates work-### Workspace updates (`repowise update --workspace`)+### `repowise update`-Updates all stale repos in parallel (up to 4 concurrent), then re-runs-cross-repo analysis (co-changes, package dependencies, API contracts).-Use `--repo <alias>` to target a single repo within the workspace.--### Dry run--To see what would be updated without actually regenerating:--```bash-repowise update --dry-run-```--### Manual update+1. Diff the new `HEAD` against the last synced commit (or `--since <ref>`).+2. Re-parse and rebuild the dependency graph, re-index git metadata for the+ changed files, and refresh dead code, health and decision markers.+3. Pick the affected wiki pages: the changed files' pages plus their direct+ importers, within a budget scaled to the size of the change+ (`--cascade-budget N` to override).+4. Regenerate those pages with the LLM, when the repo was indexed with docs.+ Affected pages beyond the budget are marked stale, not regenerated;+ `repowise generate --stale` refreshes them when you choose to spend on it.+5. In a workspace, re-run the cross-repo analysis.+6. Save the new `HEAD` as the last synced commit.```bashrepowise update # diff since last syncrepowise update --since abc123 # diff from a specific commit-repowise update --cascade-budget 50 # allow more cascade pages (default: auto)-repowise update --workspace # update all stale workspace repos-repowise update --repo backend # update a specific workspace repo-repowise watch --workspace # auto-update all workspace repos on file change+repowise update --dry-run # show affected pages, regenerate nothing+repowise update --cascade-budget 50 # allow more pages (default: auto)+repowise update --index-only # refresh the index, no model calls```+### Workspaces++`repowise update --workspace` updates every stale repo, then re-runs the+cross-repo analysis (co-changes, package dependencies, API contracts). Each repo+uses its own docs setting; `--docs` / `--no-docs` override it for the run.+`--repo <alias>` targets one repo. See [WORKSPACES.md](WORKSPACES.md).++### Server sync++`POST /api/repos/{id}/sync` re-parses the repo, re-indexes git metadata, and+rescans dead code and decisions. It does not regenerate wiki pages, so it has no+LLM cost. `POST /api/repos/{id}/full-resync` also regenerates every page.+----## Environment Variables Reference+## Environment variables| Variable | Used by | Description ||----------|---------|-------------|-| `GEMINI_API_KEY` | CLI | API key for Gemini LLM provider |-| `OPENAI_API_KEY` | CLI | API key for OpenAI LLM provider |-| `ANTHROPIC_API_KEY` | CLI | API key for Anthropic LLM provider || `REPOWISE_GITHUB_WEBHOOK_SECRET` | Server | HMAC secret for GitHub webhook verification || `REPOWISE_GITLAB_WEBHOOK_TOKEN` | Server | Token for GitLab webhook verification |-| `REPOWISE_DB_URL` | Server | Database URL (default: local SQLite) |-| `REPOWISE_API_KEY` | Server | Bearer token for API authentication |-API keys can also be stored in `.repowise/.env` (auto-gitignored) during-`repowise init`. The `update` and `reindex` commands load this file-automatically.+Provider API keys, the database URL and the API token are listed in+[CONFIG.md](../reference/CONFIG.md#environment-variables). Keys saved to+`.repowise/.env` during `repowise init` are loaded by `update` automatically.---## Troubleshooting-**"No previous sync found"** -- Run `repowise init` first to create the-initial wiki before using any auto-sync method.+**"No previous sync found"**: run `repowise init` first.-**"Already up to date"** -- The wiki is already synced to the latest commit.-Nothing to do.+**"Already up to date"**: the index is already at the latest commit.-**Hook doesn't fire** -- Make sure the hook file is executable:-`chmod +x .git/hooks/post-commit`+**Hook doesn't fire**: make sure the hook file is executable:+`chmod +x .git/hooks/post-commit`. `repowise hook status` shows whether it is+installed.-**Webhook returns 401** -- A secret/token is configured on the server but the-incoming signature did not match it (or, for GitHub, the `sha256=` prefix was-missing). Check that the value matches between your git hosting provider and-the environment variable on the server.+**Webhook returns 401**: a secret or token is set on the server but the+request's signature did not match (or the GitHub `sha256=` prefix was missing).+Check the value matches on both sides.-**Webhook returns 403** -- No secret/token is configured on the server and the-request came from a non-local caller. The unset-secret path only accepts-loopback callers, so a remote provider hitting it is refused. Set-`REPOWISE_GITHUB_WEBHOOK_SECRET` (or `REPOWISE_GITLAB_WEBHOOK_TOKEN`), restart-the server, and re-deliver the payload from your provider's webhook settings-page.+**Webhook returns 403**: no secret or token is set, and the request came from a+non-local caller. Set `REPOWISE_GITHUB_WEBHOOK_SECRET` or+`REPOWISE_GITLAB_WEBHOOK_TOKEN`, restart the server, and redeliver the payload.-**Update is slow** -- If you're catching up on many commits, the first update-may take longer. Subsequent single-commit updates are fast (~30-60s).+**Update is slow**: the first update after many commits does more work.+`repowise update --index-only` skips page regeneration when you only need the+index current.diff --git a/docs/scale/WORKSPACES.md b/docs/scale/WORKSPACES.mdindex dd957aa14..af0b97922 100644--- a/docs/scale/WORKSPACES.md+++ b/docs/scale/WORKSPACES.md@@ -1,6 +1,6 @@# Workspaces, Multi-Repo Support-Repowise workspaces let you index and analyze multiple repositories together. You get per-repo documentation, graphs, and search, plus cross-repo intelligence: co-change detection, API contract extraction, and package dependency mapping.+A workspace indexes several repositories together. Each repo keeps its own docs, graph and search, and the workspace adds cross-repo analysis: co-change, API contracts matched provider to consumer, package dependencies, a service-level system graph, breaking-change and test-impact checks, and architecture rules.---@@ -11,11 +11,14 @@ Repowise workspaces let you index and analyze multiple repositories together. Yo3. [How It Works](#how-it-works)4. [Workspace Commands](#workspace-commands)5. [Cross-Repo Intelligence](#cross-repo-intelligence)-6. [Architecture Metrics](#architecture-metrics)-7. [Web UI](#web-ui)-8. [MCP Integration](#mcp-integration)-9. [File Layout](#file-layout)-10. [FAQ](#faq)+6. [External Systems](#external-systems)+7. [System Graph](#system-graph) and [Extraction Diagnostics](#extraction-diagnostics)+8. [Web UI](#web-ui) and [Live System Map](#live-system-map)+9. [Cross-Repo Blast Radius](#cross-repo-blast-radius), [Breaking-Change Guard](#breaking-change-guard), [Cross-Repo Test Impact](#cross-repo-test-impact)+10. [Architecture Conformance](#architecture-conformance) and [Architecture Metrics](#architecture-metrics)+11. [MCP Integration](#mcp-integration)+12. [File Layout](#file-layout)+13. [FAQ](#faq)---@@ -28,15 +31,14 @@ Use a workspace when your project spans multiple git repositories that are relat- **Microservices** that communicate over HTTP, gRPC, or message topics- Any set of repos where you want to understand **cross-repo dependencies and co-change patterns**-If you only have a single repo, `repowise init` works as before, no workspace needed.+With a single repo, `repowise init` needs no workspace.---## Quick Start-### 1. Organize your repos--Put related repos under a common parent directory:+Put related repos under a common parent directory. The parent can itself be a+git repo (a monorepo with sub-repos).```my-workspace/@@ -45,47 +47,22 @@ my-workspace/shared-libs/ # git repo```-Or, if your workspace root is itself a git repo (e.g., a monorepo with sub-repos):--```-my-project/ # git repo (monorepo)- .git/- backend/ # git repo- frontend/ # git repo-```--### 2. Initialize the workspace-```bashcd my-workspacerepowise init .```-Repowise will:--1. **Scan** for git repositories (up to 3 levels deep)-2. **Prompt you to select** which repos to index-3. **Ask you to pick a primary repo** (the default for MCP queries)-4. **Walk you through provider setup** (LLM provider, model, cost estimate)-5. **Index each repo**, parse files, build graphs, index git history-6. **Generate documentation** for each repo (unless `--no-prose`)-7. **Run cross-repo analysis**, co-changes, API contracts, package deps-8. **Register MCP servers** with Claude Desktop and Claude Code--### 3. Explore+`repowise init .` scans for git repositories up to 3 levels deep, asks which to+index and which is the primary repo (the default for MCP queries), walks you+through provider setup, then indexes each repo, generates its docs (unless+`--no-prose`), runs the cross-repo analysis, and registers the MCP server with+your editors.```bash-# Check workspace status-repowise status --workspace--# List workspace repos-repowise workspace list--# Start the web UI-repowise serve--# Search across all repos-repowise search "authentication flow"+repowise status --workspace # workspace status+repowise workspace list # repos and index state+repowise serve # web UI+repowise search "authentication flow" # search across repos```---@@ -102,7 +79,7 @@ Each repo is indexed independently into its own `.repowise/wiki.db`, the same fo|---------|-------------|-----------|| Per-repo docs, graph, search | Yes | Yes (for each repo) || Co-change detection | Within repo | Within + across repos |-| API contract extraction | No | Yes (HTTP, gRPC, topics) |+| API contract extraction | No | Yes (HTTP, gRPC, topics, sockets, data) || Package dependency mapping | No | Yes || Web UI | Repo pages | Repo pages + workspace dashboard || MCP | One server per repo | One server, all repos |@@ -111,117 +88,37 @@ Each repo is indexed independently into its own `.repowise/wiki.db`, the same fo## Workspace Commands-### `repowise init .`--Initialize a workspace in the current directory. Scans for git repos, prompts for selection, and indexes everything.--**Options:**--| Flag | Description |-|------|-------------|-| `--no-prose` | Parse and analyze, wiki rendered from structure, no LLM (free) |-| `-x, --exclude` | Glob patterns to exclude (e.g., `-x "node_modules/"`) |-| `--yes` | Skip confirmation prompts |-| `--concurrency N` | Max concurrent file parses (default: auto) |--**Example:**--```bash-repowise init . -x "node_modules/" -x "*.lock" -x "vendor/"-```--### `repowise workspace list`--Show all repos in the workspace with their index status.--```bash-repowise workspace list-```--### `repowise workspace add <path>`--Add a new repo to an existing workspace and index it.--```bash-repowise workspace add ../new-service --alias api-gateway-```--### `repowise workspace remove <alias>`--Remove a repo from the workspace (does not delete files).--```bash-repowise workspace remove api-gateway-```--### `repowise workspace scan`--Re-scan the workspace directory for new repos that haven't been added yet.--```bash-repowise workspace scan-```--### `repowise workspace set-default <alias>`--Change which repo is the default for MCP queries.--```bash-repowise workspace set-default backend-```--### `repowise workspace diagnostics`--Explain the cross-repo contract link count, per-repo provider/consumer counts, unmatched consumers grouped by reason, and orphan providers. See [Extraction Diagnostics](#extraction-diagnostics).--```bash-repowise workspace diagnostics # human-readable report-repowise workspace diagnostics --json # raw JSON-repowise workspace diagnostics --repo api # limit to one repo-```--### `repowise workspace check`--Architecture lint: check the declared `conformance:` rules against the system graph and detect dependency cycles. Exits non-zero on any finding, so it gates CI. See [Architecture Conformance](#architecture-conformance).--```bash-repowise workspace check # human-readable report; exit 1 on findings-repowise workspace check --json # raw report JSON-```--### `repowise workspace metrics`--Architecture-complexity metrics: propagation cost, the cyclic core, per-service roles, and a deterministic 1-10 architecture score. See [Architecture Metrics](#architecture-metrics).--```bash-repowise workspace metrics # human-readable summary-repowise workspace metrics --json # raw metrics JSON-```--### `repowise workspace impacted-tests <repo:path>...`--Given one or more changed provider files, list the tests in consumer repos worth running. See [Cross-Repo Test Impact](#cross-repo-test-impact).--```bash-repowise workspace impacted-tests backend:app/routers/users.py-repowise workspace impacted-tests backend:app/routers/users.py --target-repo frontend --format list | cut -d: -f2- | xargs npx vitest run-repowise workspace impacted-tests backend:app/routers/users.py --format json-```+Run these from anywhere inside the workspace. Flags and examples for each are+in the [CLI reference](../reference/CLI_REFERENCE.md#workspace-commands).++| Command | What it does |+|---------|--------------|+| `repowise init .` | Initialize the workspace: scan for repos, select, index (`--no-prose` for a free structure-only wiki, `-x` to exclude globs) |+| `repowise workspace list` | Repos in the workspace with their index status |+| `repowise workspace add <path>` | Add and index a repo (`--alias`, `--no-docs`, `--no-index`) |+| `repowise workspace remove <alias>` | Remove a repo from the workspace (files are not deleted) |+| `repowise workspace scan` | Find repos not yet added |+| `repowise workspace set-default <alias>` | Change the default repo for MCP queries |+| `repowise workspace diagnostics` | Explain the contract link count ([Extraction Diagnostics](#extraction-diagnostics)) |+| `repowise workspace check` | Architecture lint; exits non-zero on findings ([Architecture Conformance](#architecture-conformance)) |+| `repowise workspace metrics` | Propagation cost, cyclic core, service roles, 1-10 score ([Architecture Metrics](#architecture-metrics)) |+| `repowise workspace impacted-tests <repo:path>...` | Tests in consumer repos to run for a provider change ([Cross-Repo Test Impact](#cross-repo-test-impact)) |++The report commands take `--format table|json` (`impacted-tests` also takes+`list`); `--json` is a deprecated alias.---## Cross-Repo Intelligence-When you initialize a workspace with 2+ repos, repowise runs three types of cross-repo analysis:+When a workspace has two or more repos, repowise runs three kinds of cross-repo analysis. Each repo also gets its own [external systems](#external-systems) registry.### Co-Change DetectionAnalyzes git history across repos to find files that frequently change together. For example, if `backend/api/routes.py` and `frontend/src/api/client.ts` are always modified in the same time window, they get a high co-change score.-Useful for:-- Understanding implicit dependencies between repos-- Knowing what frontend files to check when a backend API changes-- Identifying tightly coupled components+Use it to find implicit dependencies between repos, such as the frontend files+to check when a backend API changes.### API Contract Extraction@@ -263,66 +160,23 @@ Scans source files for HTTP routes, gRPC services, database tables, message topi| Socket / WebSocket | SignalR `MapHub<T>("/path")`, FastAPI `@app.websocket("/path")`, `ws` `WebSocketServer({ path })`, NestJS `@WebSocketGateway`; events: socket.io `emit` (server and client), Laravel broadcast events (`broadcastOn` / `broadcastAs`), `Broadcast::on`, Pusher `trigger` | ClientWebSocket `ConnectAsync`, SignalR `HubConnectionBuilder.WithUrl`, NativeWebSocket and WebSocketSharp `new WebSocket(...)`, browser/Node `new WebSocket(url)`; events: socket.io `on` / `@SubscribeMessage`, Laravel Echo `listen` and `useEcho`, pusher-js `bind` |Socket detection is toggled by `detect_socket` in the `contracts:` block below.-An endpoint is identified by its path (`socket::/hubs/game`); a message by its-event within a scope, `socket::<scope>#<event>`, where the scope is a socket.io-namespace (`/` unless the file names one) or a broadcast channel. Channels keep-the wire prefix Pusher gives them (`private-orders.{param}`), and Echo's event-names are read the way Echo formats them, so `.listen('OrderShipped')` meets a-Laravel event class `App\Events\OrderShipped` and `.listen('.order.shipped')`-meets `broadcastAs()` returning `order.shipped`. The emitting side is the-provider. socket.io, amqplib's `publish`/`consume`, BullMQ, Redis and NestJS-calls are read only in a file importing the library, since `emit`, `on`,-`publish` and `subscribe` are common method names.--A topic, queue or exchange name is read the way a URL is: a literal, or a name-the same file assigns exactly once to a literal (Python, Go, Java, PHP class-constants, and JS/TS `const` including object and enum members such as-`QUEUES.ticketSold`), is resolved; a name built at runtime is skipped rather-than guessed. An SQS queue URL is named by its last path segment and an SNS-topic ARN by its last field. A subscription by pattern (NATS `orders.*` /-`orders.>`, Redis `psubscribe`, Kafka `topicPattern`) links to every publisher-whose name it matches. A Laravel job dispatched without a queue runs on the-queue its class declares (`public $queue`, `$this->onQueue()`, `viaQueue()`),-found by the class's fully qualified name; one left on the connection's-default queue is not recorded, since every app has one. RabbitMQ-publishers name an exchange and a routing key while consumers name a queue, so-a queue binding found anywhere in the workspace connects the two: each consumer-of the bound queue links to the exchange's publishers whose routing key the-binding pattern accepts (`*` and `#` follow topic-exchange rules; an empty key-or pattern matches everything, while a key the source does not settle matches-nothing). Such a link carries the exchange as its `contract_id` and the queue as-`consumer_contract_id`. When no consumer of the bound queue is found, the-binding site itself is linked. A publish to the default exchange-(`publish('', 'jobs')`) is a publish to the queue `jobs`.--Data/DB contracts use the id scheme `data::<table>` and render as a `db` edge in the [system graph](#system-graph). An ORM model with no explicit table name takes its library's default: the Prisma model name, the TypeORM class name in snake case, the Sequelize model name pluralized, the Eloquent class name in snake case and plural. A service that only models a table (an ORM class, no migration) is linked to the service whose migration or DDL defines that table's schema, when exactly one service defines it; two services that each migrate a table of one name are read as separate databases and not linked. The consumer side (SQL string matching) is heuristic and lower-confidence than the ORM-based providers; unlike HTTP and gRPC, there is no field-level breaking-change diffing for data contracts, only table/route-level removal.--HTTP routes are matched on their **full** path: a router mount prefix-(`APIRouter(prefix=...)`, `include_router(prefix=...)`, Express `app.use('/x', router)`,-Go route groups, Laravel `Route::prefix(...)->group(...)` and `Route::group(['prefix' => ...])`)-is stitched onto each handler path before matching. Laravel's `routes/api.php` is-served under `/api` unless `bootstrap/app.php` (`apiPrefix`) or a route provider says-otherwise, and `Route::resource` / `apiResource` expand into the routes they register. A NestJS-route is served at the app's `setGlobalPrefix` (unless its `exclude` list names the route), then-its URI version (`enableVersioning`, `@Version`), then the `@Controller` prefix. A call through an-axios, ky, got or ofetch instance is read with the instance's `baseURL` / `prefixUrl`, also when-another file imports the instance. An Angular `HttpClient` call is read on any receiver typed-or injected as `HttpClient`, with `environment.apiUrl` folded from `src/environments/` and class-fields built on it; a base an interceptor prepends is not read. A client call-whose base URL is an unresolved placeholder (`fetch(\`${API_BASE}/users\`)`) matches-on the host-relative path; the link is **exact** when exactly one workspace service-provides that path and a lower-confidence **candidate** when the target is ambiguous.--**Over REST.** `GET /api/workspace/contracts` lists contracts and links,-filterable by `contract_type`, `repo` and `role`. Each contract carries its-line, its ingestion symbol id and the extractor's `meta`; each link carries both-symbol ids and both service boundaries. The request/response `schema` is not on-the list, because it runs to full inline type declarations and only one is ever-needed at a time: fetch it with-`GET /api/workspace/contracts/detail?repo=<alias>&file=<path>&id=<contract-id>`,-which returns that one contract with its schema, its links, and its unmatched-reason. All three parameters are required, since a contract id alone is not-unique across repos.+A topic, queue or route name is resolved when it is a literal or a constant the+same file assigns once; a name built at runtime is skipped, not guessed. HTTP+routes match on their full path, with router mount prefixes stitched on first.+An ambiguous target becomes a lower-confidence **candidate** link, and a unique+one an **exact** link. The exact naming and matching rules for each transport+(socket scopes, pattern subscriptions, RabbitMQ bindings, Laravel queues, ORM+table defaults, framework route prefixes, client base URLs) are in the+[contract matching reference](../reference/WORKSPACE_CONTRACTS.md).++Data/DB contracts use the id scheme `data::<table>` and render as a `db` edge in+the [system graph](#system-graph). The consumer side (SQL string matching) is+heuristic and lower-confidence than the ORM-based providers, and data contracts+get table-level removal checks only, no field-level diffing.++Over REST, `GET /api/workspace/contracts` lists contracts and links (filter by+`contract_type`, `repo`, `role`), and `GET /api/workspace/contracts/detail`+returns one contract with its schema.**Tuning extraction** via the `contracts:` block in `.repowise-workspace.yaml`:@@ -358,30 +212,53 @@ Reads package manifests (`package.json`, `composer.json`, `pyproject.toml`,`Cargo.toml`, `go.mod`, `.csproj`, and Maven `pom.xml`) to detect when one repodepends on another as a package or project.-npm and composer dependencies match in two ways: a local path (`file:` specs,-workspace globs, composer `path` repositories) that resolves into a sibling repo,-or a package name that exactly one sibling repo publishes. For npm, a repo-publishes its root `package.json` and its declared workspace members, so a-vendored or fixture `package.json` never claims a name. For composer, every-`composer.json` up to three directories deep counts, outside `vendor/`, hidden-directories, tests, fixtures and examples, which covers split packages such as-`src/Illuminate/Support/composer.json`. A name two repos publish links to-neither.--Maven matching is filesystem-only and coordinate-based. Repowise resolves local-reactor modules, local parents, properties, and dependency-management versions,-then links an active direct compile/runtime dependency only when exactly one-selected workspace project publishes that `groupId:artifactId`. Test, provided,-system, optional, profile-only, ambiguous, and external dependencies do not create-production package edges. Bounded diagnostics retain the reason for Maven-non-matches. When a repository has a root `pom.xml`, only that declared reactor is-eligible; unrelated nested example or fixture POMs are not treated as producers.--This does **not** execute Maven, read user settings, download artifacts, resolve-plugins/transitive dependencies/imported BOMs, or infer generated sources. A Maven-package edge is also not a published symbol-level code API or a runnable Maven-target recommendation; those capabilities are reported separately and remain-unsupported.+npm and composer dependencies match through a local path that resolves into a+sibling repo, or a package name that exactly one sibling repo publishes; a name+two repos publish links to neither. Maven matching is filesystem-only: it links+an active direct compile/runtime dependency when exactly one workspace project+publishes that `groupId:artifactId`, and never runs Maven or downloads+artifacts. Details: [contract matching reference](../reference/WORKSPACE_CONTRACTS.md#package-matching).++---++## External Systems++Every indexed repo, in a workspace or not, keeps a registry of the third-party+packages it declares. repowise reads dependency manifests during indexing; no+package is downloaded and no build tool runs.++| Ecosystem | Manifests read |+|-----------|----------------|+| npm | `package.json` |+| PyPI | `pyproject.toml` (PEP 621, optional dependencies, Poetry groups), `requirements*.txt` |+| Cargo | `Cargo.toml` |+| Go | `go.mod` |+| NuGet | `*.csproj` |+| Maven | `pom.xml` |+| CMake | `find_package(...)` calls in `CMakeLists.txt` |++Each declaration keeps its name, version, the manifest it came from, and whether+it is a dev dependency. A dependency on the repo's own packages is dropped: it+is part of the codebase, not an external system. Two tags are added from+built-in name lists:++- **Category**: `framework` (fastapi, react, spring, gin), `service` (a network+ dependency such as stripe, openai, an AWS or Azure SDK, a database driver),+ `tool` (eslint, pytest, vite) or `library`, the default. The lists are+ deliberately small, so an unknown package stays `library`.+- **I/O kind**: `db`, `network`, `filesystem`, `subprocess` or `lock`, when the+ package is a known boundary of that kind. Unknown names have none.++The registry also links each package to the import edges that reach it, so you+can see which parts of the codebase use it. In the dashboard this is the+**Packages** tab on a repo's Architecture page: declared packages with their+importing communities, and a drill-down to the importing files. The C4 view+uses the same registry to draw external boxes. Over REST:+`GET /api/repos/{repo_id}/external-systems/summary`.++This is a declared-dependency view. It does not resolve transitive+dependencies or lockfiles, and it does not match a package to a sibling+workspace repo; that is [Package Dependency Scanning](#package-dependency-scanning).---@@ -391,7 +268,7 @@ The contracts, package dependencies, and co-changes above are each a flat list.**Nodes are services, not repos.** A monorepo with three detected service boundaries (a `package.json` / `composer.json` / `go.mod` / `Cargo.toml` sub-directory) shows three nodes; the repo is a grouping attribute on each node. A repo with no sub-boundary collapses to a single repo-root node. Each node carries its provider/consumer counts, the contract types it participates in, and flags for orphan/isolated services.-**Edges are typed and honest.** Every edge carries:+**Edges are typed.** Every edge carries:- a `kind`, `http`, `grpc`, `event`, `package`, `co_change`, or `db`;- a `match_type`, `exact`, `candidate`, `manual`, or `inferred`;@@ -407,9 +284,9 @@ Fetch it over REST with `GET /api/workspace/system-graph`, or explore it visuallWhen the cross-repo link count looks low, diagnostics explain why. Computed alongside contract matching, they report, per repo and contract type, how many providers and consumers were found, which consumers went unmatched (and why), and which providers have no consumer at all.```bash-repowise workspace diagnostics # human-readable report-repowise workspace diagnostics --json # raw JSON-repowise workspace diagnostics --repo api # limit to one repo+repowise workspace diagnostics # human-readable report+repowise workspace diagnostics --format json # raw JSON+repowise workspace diagnostics --repo api # limit to one repo```The report covers:@@ -422,7 +299,7 @@ The report covers:- `external_host`, the call targets a literal third-party host (Stripe, Formspree, ...) that is not a workspace service, so it is intentionally excluded from matching.- **Orphan providers**, endpoints declared but never consumed by any repo.- **Weak links**, matched links below the confidence threshold.-- **Extraction coverage**: how many contracts came from the parsed symbol table (`index`) versus a text dialect (`regex`), and how many HTTP client calls were located but could not be resolved to an endpoint. The HTTP coverage percentage is calls resolved over calls located — it is not total recall, because a call no dialect recognises is not in either number. It is the figure that turns a large orphan-provider count from alarming into explained.+- **Extraction coverage**: how many contracts came from the parsed symbol table (`index`) versus a text dialect (`regex`), and how many HTTP client calls were located but could not be resolved to an endpoint. The HTTP coverage percentage is calls resolved over calls located. It is not total recall, because a call no dialect recognises is not in either number. It is the figure that turns a large orphan-provider count from alarming into explained.The same data is available over REST at `GET /api/workspace/diagnostics` and is embedded in the system graph artifact's `diagnostics` block.@@ -451,29 +328,29 @@ The System Map renders the [system graph](#system-graph) as an always-current di- **Service nodes**, coloured by category (service, frontend, worker, library, external), with a health ring rolled up from the owning repo and small flags for orphan or isolated services.- **Typed edges** distinguished by `kind` (colour + glyph) and by `match_type` (solid for exact/manual, dashed for candidate, dotted for inferred co-change). Behavioral co-change edges read differently from structural contract/dependency edges.-- **Filters** to toggle each edge kind on or off, and a **service ↔ repo** switch that collapses a monorepo's services into one node per repository.+- **Filters** to toggle each edge kind on or off, and a **service / repo** switch that collapses a monorepo's services into one node per repository.- **Drill-down**: click a service to inspect its providers/consumers and connected services; click an edge to see its match type, confidence, weight, and the underlying contract evidence, with a jump to the Contracts view.- A **legend** explaining the edge colours, dash patterns, and the health scale.-The map appears once the workspace has at least two indexed repositories with detected relationships; it shows honest empty states otherwise.+The map appears once the workspace has at least two indexed repositories with detected relationships; it shows an empty state otherwise.---## Cross-Repo Blast Radius-Blast radius answers a single question: **if I change this service, which downstream services and repos are structurally exposed?** It walks the [system graph](#system-graph) *against* its edge direction, a `consumer → provider` edge means changing the provider may impact the consumer, and returns every reachable service ranked by an impact score.+Blast radius answers a single question: **if I change this service, which downstream services and repos are structurally exposed?** It walks the [system graph](#system-graph) *against* its edge direction (a consumer-to-provider edge means changing the provider may impact the consumer) and returns every reachable service ranked by an impact score.Two edge classes are weighted and labelled distinctly:-- **Structural** edges (http / grpc / event / package / db) assert a real dependency, a contract or an import. They propagate impact at full weight and surface under the compatibility-named **will break** field, but mean structural reach rather than certain runtime failure.-- **Behavioral** co-change edges only assert that two files historically *changed together*. They are correlation, not a call, so they propagate at half weight (one named constant, `BEHAVIORAL_EDGE_WEIGHT`) and surface as **may drift**.+- **Structural** edges (http / grpc / event / package / db) assert a real dependency, a contract or an import. They propagate impact at full weight and surface under the compatibility-named **will break** field, but mean structural reach, not certain runtime failure.+- **Behavioral** co-change edges only assert that two files historically *changed together*. They are correlation, not a call, so they propagate at half weight and surface as **may drift**.Each impacted service carries its `distance` (hops from the change) and `score` (0-1, with distance decay and the behavioral weighting baked in). Nearer, structural impact ranks highest.Use it three ways:- **REST**, `GET /api/workspace/blast-radius?target=<node-id-or-repo>&max_depth=3&include_behavioral=true`. `target` is a node id (`repo` or `repo::service/path`) or a repo alias (expands to all its services).-- **MCP**, the `get_blast_radius` tool (workspace mode) gives an agent the impacted set before it touches a high-fan-out provider. The `get_risk` PR-mode directive also gains `will_break_consumers` and `missing_cross_repo_cochanges` so a diff in one repo flags its cross-repo fallout.+- **MCP**, the opt-in `get_blast_radius` tool (workspace mode, `mcp.tools: ["+get_blast_radius"]`) gives an agent the impacted set before it touches a high-fan-out provider. The `get_risk` PR-mode directive also gains `will_break_consumers` and `missing_cross_repo_cochanges` so a diff in one repo flags its cross-repo fallout.- **System Map**, pick a service in the **Blast radius** control above the map; the reachable set ripples (highlighted, the rest dimmed, badges grading intensity), and a side panel lists the impacted services. Click any impacted service to walk the impact outward from there.---@@ -482,13 +359,13 @@ Use it three ways:Where blast radius answers *what could be affected*, the breaking-change guard answers a sharper question: **did a provider contract change incompatibly?** On every `repowise update --workspace`, freshly extracted contracts are diffed against the previously indexed set. Each finding carries the consumer files linked to the endpoint, but that link proves endpoint exposure only. It does not prove use of the changed field, a runtime failure, or deployment safety.-Detected change kinds (a registry, adding a kind is one new rule, never an `if/elif`):+Detected change kinds:| Kind | Severity | Fires when ||------|----------|-----------|| `removed_endpoint` | breaking | A provider route / gRPC method / topic that existed before is gone || `removed_field` | breaking (response) / warning (request) | A request or response field disappeared |-| `field_type_changed` | breaking | A field's type changed (e.g. `string → int64`) |+| `field_type_changed` | breaking | A field's type changed (e.g. `string -> int64`) || `field_number_changed` | breaking | A proto field's wire number changed || `field_required` | breaking | A request field became required, or a new required request field was added; legacy proto/signature behavior is preserved || `field_required_relaxed` | breaking | A required OpenAPI response field became optional |@@ -496,13 +373,16 @@ Detected change kinds (a registry, adding a kind is one new rule, never an `if/e| `field_enum_changed` | breaking | An OpenAPI request enum lost values, or a response enum gained values || `schema_comparison_uncertain` | warning | Schema source/fidelity, completeness, or selected response changed, so field compatibility was not inferred |-OpenAPI comparison covers the common supported subset of `3.0.x`, `3.1.x`, and `3.2.x`: JSON/YAML documents; path/query/header/cookie parameters using OpenAPI's default `style`, `explode`, and `allowReserved` behavior; one `application/json` request body; exactly one explicit JSON 2xx response; recursive objects and arrays; exact primitive types; requiredness; normalized nullability; finite homogeneous scalar enums; and bounded same-document JSON Pointer references. Request rules describe values the provider accepts; response rules describe values consumers may receive, so requiredness, nullability, enum values, and constrained/unconstrained enum transitions reverse between the two sides. Operation removal remains the transport-neutral contract rule.--Only complete sides with the same schema source and comparison-fidelity key are field-diffed. Unsupported/unresolved nodes, extraction-strategy changes, and response-selection changes become warning-level uncertainty rather than shortened schemas or false removals. Remote/cross-file references, composition, additional-properties semantics, arbitrary JSON Schema constraints, non-JSON or ambiguous media, multiple materially different success responses, and OpenAPI 2.0 are outside this boundary. No network dereferencing occurs.+Field-level comparison covers proto, signatures and a supported subset of+OpenAPI `3.0.x` to `3.2.x` (JSON request bodies, one JSON 2xx response, objects,+arrays, primitive types, requiredness, nullability, scalar enums, same-document+references). Anything outside that subset becomes a warning-level+`schema_comparison_uncertain`, never a false removal. The exact boundary:+[contract matching reference](../reference/WORKSPACE_CONTRACTS.md#openapi-comparison-boundary).**Compatible changes stay quiet**: examples include an optional request addition, a request enum widening, a request becoming nullable, a response enum narrowing, a response becoming non-nullable, an additional response field in the supported open-object subset, and a brand-new endpoint. A rename remains a removal plus an addition; no rename inference is attempted.-Endpoint-exposed consumers are resolved from the matched contract links, the same provider↔consumer evidence the [system graph](#system-graph)'s edges are built from. The evidence is direct and endpoint-level: it identifies a consumer file linked to the changed contract, but not the exact field it uses. Transitive structural reach stays the job of blast radius.+Endpoint-exposed consumers are resolved from the matched contract links, the same provider-consumer evidence the [system graph](#system-graph)'s edges are built from. The evidence is direct and endpoint-level: it identifies a consumer file linked to the changed contract, but not the exact field it uses. Transitive structural reach stays the job of blast radius.Use it three ways:@@ -536,7 +416,7 @@ Results are capped per consumer and provider pair so one widely-called helper caUse it three ways:- **CLI**, `repowise workspace impacted-tests <repo:path>...`, the command above.-- **REST**, `GET /api/workspace/test-impact?repo=<alias>&file=<path>`. Repeat `file` for several changed files in the same repo. The response carries the same fields the `--format json` output has: `recommendations` (each with its consumer repo, the `consumer_files` and bound `consumer_symbol_ids` that reached the test, `basis`, `via`, `confidence` and the contract ids that produced it), `unresolved` with a reason per link, `files_analyzed` with the state each landed on, and the `summary` counts. Over the API an empty answer's `summary.reason` is `no_contract_data`, `no_matching_links` or `lookup_failed`.+- **REST**, `GET /api/workspace/test-impact?repo=<alias>&file=<path>` (repeat `file` for several changed files), with the same fields as `--format json`. Field list: [contract matching reference](../reference/WORKSPACE_CONTRACTS.md#test-impact-over-rest).- **Web UI**, a provider contract's page (`/workspace/contracts/detail`) ends in a **Tests to run** section: the tests grouped by consumer repo, each marked measured or inferred and saying whether the coverage map, the call graph or the import graph found it, and a **Could not determine** list naming the consumer file and the reason. A consumer contract has no such section, since tests are found on the consumer side.---@@ -587,20 +467,20 @@ A rule with `allow: false` (the default) is a **deny** rule: a structural depend### Dependency cycles-Independently of any rules, conformance detects **circular dependencies** among services over structural edges (`A → B → … → A`). A cycle means the services cannot be built, deployed, or reasoned about independently. Cycle detection runs even with zero rules declared, so every workspace gets it for free.+Independently of any rules, conformance detects **circular dependencies** among services over structural edges (`A -> B -> ... -> A`). A cycle means the services cannot be built, deployed, or reasoned about independently. Cycle detection runs even with zero rules declared, so every workspace gets it.### Using it- **CLI**, `repowise workspace check` prints violations and cycles and exits non-zero when any are found, so it gates CI (the architecture lint):```bash- repowise workspace check # human-readable report; exit 1 on findings- repowise workspace check --json # raw report JSON (still exits 1 on findings)+ repowise workspace check # human-readable report; exit 1 on findings+ repowise workspace check --format json # raw report JSON (still exits 1 on findings)```It recomputes from the persisted system graph, so editing rules and re-running picks them up without a full re-index.- **REST**, `GET /api/workspace/conformance` returns the report from the most recent update (filterable by `repo`).-- **MCP**, `get_conformance` exposes violations and cycles to an agent; the `get_risk` PR-mode directive gains `conformance_violations` and `dependency_cycles` blocks for the findings the diff's repo participates in.+- **MCP**, `get_conformance` exposes violations and cycles to an agent. It is opt-in (`mcp.tools: ["+get_conformance"]`, see [MCP_TOOLS.md](../agent/MCP_TOOLS.md#get_conformance)). Without it, the `get_risk` PR-mode directive still gains `conformance_violations` and `dependency_cycles` blocks for the findings the diff's repo participates in.- **Conformance view**, the web UI's Conformance page renders a **dependency-structure matrix (DSM)**: services on both axes, each filled cell a dependency tinted by transport, with rule violations ringed red and cycle cells amber. Governance panels list the violations and cycles. Violations also badge the offending edges on the [Live System Map](#live-system-map) (toggle **Conformance**), reusing the same additive overlay as the breaking-change guard.---@@ -623,26 +503,27 @@ Conformance and the cycle finder answer *per-relationship* questions (is this ed### Using it-- **CLI**, `repowise workspace metrics` prints the score, propagation cost, cyclic core, dependency-cycle count, and the per-role service breakdown. CI-friendly plain output; `--json` emits the raw metrics.+- **CLI**, `repowise workspace metrics` prints the score, propagation cost, cyclic core, dependency-cycle count, and the per-role service breakdown. CI-friendly plain output; `--format json` emits the raw metrics.```bash- repowise workspace metrics # human-readable summary- repowise workspace metrics --json # raw metrics JSON+ repowise workspace metrics # human-readable summary+ repowise workspace metrics --format json # raw metrics JSON```- **REST**, `GET /api/workspace/architecture` returns the workspace metrics plus the per-service roles. Computed at request time from the system graph (no separate artifact); the conformance violation count, if a report exists, is folded into the score.-- **MCP**, `get_architecture` gives an agent the score, propagation cost, core members, and role breakdown in one call, the system-structure read to consult before a cross-service refactor.+- **MCP**, `get_architecture` gives an agent the score, propagation cost, core members, and role breakdown in one call, the system-structure read to consult before a cross-service refactor. It is opt-in, like `get_conformance` (`mcp.tools: ["+get_architecture"]`).- **Web**, the **architecture score** appears as a stat on both the Conformance and System Map pages. The DSM header shows score / propagation cost / core size and tints each service's diagonal cell by its role, so the on-diagonal core block stands out. On the Live System Map, toggle **Core** to highlight the cyclic core, and the inspector shows any selected service's role and visibility profile.---## MCP Integration-Workspace init automatically registers MCP servers with Claude Desktop and Claude Code. The MCP server is workspace-aware:+Workspace init registers the MCP server with your editors (skip with `--no-editor-setup`). The MCP server is workspace-aware:- **Default repo context**, queries go to the primary repo unless you specify otherwise-- **Cross-repo tools**, MCP tools can query across repos and return enriched context with co-change and contract data; `get_blast_radius` answers cross-repo downstream impact (see [Cross-Repo Blast Radius](#cross-repo-blast-radius)); `get_conformance` answers architecture rule violations and dependency cycles (see [Architecture Conformance](#architecture-conformance)); `get_architecture` answers whole-system coupling, the cyclic core, and the architecture score (see [Architecture Metrics](#architecture-metrics))-- **Repo parameter**, most tools accept an optional `repo` parameter to target a specific repo. Four also accept `"all"` to query across the workspace: `get_overview` (workspace topology and a summary per repo), `search_codebase` (results from every repo), `get_dead_code` (findings from every repo) and `get_why` with a query (decisions from every repo). The other tools answer about one repo at a time+- **Cross-repo context**, single-repo tools add co-change and contract evidence from the other repos+- **Repo parameter**, most tools accept an optional `repo` parameter to target a specific repo. Four also accept `"all"` to query across the workspace: `get_overview` (the cross-repo topology), `search_codebase` (results from every repo), `get_dead_code` (findings from every repo) and `get_why` with a query (decisions from every repo). The other tools answer about one repo at a time; see [MCP_TOOLS.md](../agent/MCP_TOOLS.md#workspace-mode)+- **Opt-in workspace tools**, off by default and enabled with `mcp.tools` in `.repowise/config.yaml`: `get_blast_radius` ([Cross-Repo Blast Radius](#cross-repo-blast-radius)), `get_conformance` ([Architecture Conformance](#architecture-conformance)) and `get_architecture` ([Architecture Metrics](#architecture-metrics))---@@ -662,18 +543,14 @@ my-workspace/.claude/CLAUDE.md # Workspace-level CLAUDE.md for AI editorsbackend/- .repowise/ # Per-repo index data- wiki.db # SQLite database (pages, graph, symbols, git)- lancedb/ # Vector embeddings- config.yaml # Repo-level config- mcp.json # MCP server config+ .repowise/ # Per-repo index, same as single-repo modefrontend/.repowise/- wiki.db- lancedb/- ...```+What lives in each repo's `.repowise/` is described in+[CONFIG.md](../reference/CONFIG.md#the-repowise-directory).+### What goes in `.gitignore`Add these to your `.gitignore`:@@ -684,7 +561,7 @@ Add these to your `.gitignore`:.repowise-workspace.yaml```-The workspace config and data are local, they reference absolute paths and contain generated analysis that should be rebuilt per-machine.+The workspace config and data are local: they can reference absolute paths and contain generated analysis that should be rebuilt per-machine.---@@ -692,7 +569,7 @@ The workspace config and data are local, they reference absolute paths and conta### Can I add repos that live outside the workspace directory?-Yes. Use `repowise workspace add /path/to/external-repo`. The path will be stored relative to the workspace root if possible, or as an absolute path otherwise.+Yes. Use `repowise workspace add /path/to/external-repo`. The path is stored relative to the workspace root if possible, or as an absolute path otherwise.### What happens if I run `repowise init` (without `.`) in a workspace?@@ -719,7 +596,7 @@ repowise watch --workspace### How do I re-run just the cross-repo analysis?-Currently, cross-repo analysis runs automatically during `repowise init .` and `repowise update --workspace`. To force a re-run, use `repowise init .` again, it will detect existing indexes and only re-run what's needed.+Cross-repo analysis runs during `repowise init .` and `repowise update --workspace`. To force a re-run, run `repowise init .` again; it detects existing indexes and only re-runs what is needed.### Does the MCP server handle multiple repos?@@ -727,4 +604,4 @@ Yes. A single MCP server instance serves all workspace repos. It uses lazy-loadi### Can I use `repowise` with git worktrees?-Yes, and it's automatic. Running `repowise init` or `repowise update` inside a linked worktree detects the base checkout, seeds the worktree's index from it, and incrementally updates only the files that differ on your branch. No flags needed; `--seed-from <path>` and `--no-seed` exist as overrides. See [WORKTREES.md](WORKTREES.md).+Yes, automatically. Running `repowise init` or `repowise update` inside a linked worktree detects the base checkout, seeds the worktree's index from it, and incrementally updates only the files that differ on your branch. No flags needed; `--seed-from <path>` and `--no-seed` exist as overrides. See [WORKTREES.md](WORKTREES.md).diff --git a/docs/start/QUICKSTART.md b/docs/start/QUICKSTART.mdindex 0dd6d0ab5..eced955c1 100644--- a/docs/start/QUICKSTART.md+++ b/docs/start/QUICKSTART.md@@ -1,410 +1,236 @@# Quickstart-Index your repo and get your agent answering questions from it. The first three-steps need no API key and no configuration.+Index your repository and have your coding agent answer from that index. The+default setup needs no API key, makes no LLM calls and sends nothing off your+machine.-> Already indexed and looking for a specific command? See the-> [CLI Reference](../reference/CLI_REFERENCE.md). For everything else, the-> [User Guide](USER_GUIDE.md).+> Looking for one command's flags? See the+> [CLI Reference](../reference/CLI_REFERENCE.md). For everyday workflows, see+> the [User Guide](USER_GUIDE.md).----+## Fastest path: let your agent do it-## 1. Install--```bash-pip install repowise-```--That is the whole install. Every LLM provider SDK (Anthropic, OpenAI, Gemini,-LiteLLM) ships in the base package, so there is nothing to pick at install time-and no extras to remember. You choose a provider later, at index time, and you-can change it whenever you want.--**Requirements:** Python 3.11+ and Git. On Windows use `python -m pip install repowise`.--Check it landed:--```bash-repowise --version-```--## 2. Index your repo--```bash-cd /path/to/your-repo-repowise init-```--Bare `init` scans the repo and asks how to index it: everything (the wiki written-by a model), no prose (the same wiki rendered from your code's structure, no key-and no spend), or advanced, which walks through the indexing and generation knobs.-Nothing is spent before you see an estimate and confirm it, so it is safe to run-and read.--If you would rather not answer questions, name the mode and add `--yes`. This is-the form to use in a script, in CI, or when an agent is setting repowise up:--```bash-repowise init --yes --no-prose # free, no key, no questions-repowise init --yes --prose # model-written subsystem pages, cost pre-approved-```--Bare `repowise init --yes` needs no API key either: without a resolvable provider-it renders the whole wiki from structure.--The keyless run is the one worth doing first, because it costs nothing and answers-the question "is this useful on my codebase". It parses every file to an AST, builds-the dependency graph, reads your git history, scores every file for code health,-and finds dead code. It also renders a complete wiki from that structure: file,-symbol, layer and cycle pages, the architecture diagram, the repo overview, API-and infra pages, and the onboarding collection. Without a key the subsystem-(concept) pages, the repo overview, the architecture diagram and the onboarding-collection are structural stubs until you point a model at them. No LLM calls,-no key, no network.--Those pages are honest about where they came from. Each one ends with a footer-saying it was derived from structure, and the repo overview covers composition,-entry points, clusters and dependencies rather than what the project does end to-end, because no template can derive that.--When it finishes you have a working index. Try it:--```bash-repowise health # the lowest-scoring files, and why-repowise dead-code # what nothing references any more-repowise risk HEAD~5..HEAD # how risky your recent work looks-```--`repowise health` ends with a self-check against your own git history, something-like *"16 of the 20 lowest-health files had a bug fix in the last 6 months, 3.3x-the 24% baseline"*. That number is computed on your repo, not ours. If it looks-bad, the score is not working for your codebase and you should say so.--## 3. Connect your agent--This is the payoff: your agent reads the index instead of your codebase.--<details open><summary><b>Claude Code</b></summary>--**`repowise init` already did this.** Unless you passed `--no-editor-setup` or-set `REPOWISE_SKIP_EDITOR_SETUP=1`, it writes a repo-root `.mcp.json` and-registers repowise with `~/.claude/settings.json`. One switch covers both: the-flag used to gate only the machine-wide registration, so a run that opted out-of editor setup still wrote files into the working tree. `.repowise/mcp.json`-is the one deliberate exception — it is written either way, because it is what-`repowise mcp .` prints and is how you opt back in by hand.--A Claude Code session opened in this repo already sees the MCP server. Check-with `repowise agents` (Claude Code should show as wired) or `repowise doctor`.--Skipped editor setup, or setting up a machine where you did? Wire it up now:--```bash-repowise agents add --target=claude-code-```--**The plugin** additionally installs hooks and slash commands, which `init`-never writes because the plugin route is host-managed and only the plugin can-install it:+Open your coding agent (Claude Code, Codex, Cursor, VS Code Copilot, OpenCode,+Hermes or any MCP client) in the repository and paste this line:```text-/plugin marketplace add repowise-dev/repowise-/plugin install repowise@repowise-```--**A different MCP client, or wiring the server by hand:**--```bash-claude mcp add repowise -- repowise mcp+Read https://docs.repowise.dev/setup.md and set up Repowise in this repository.```-Or edit the project `.mcp.json` `init` already wrote (commit it if your team-should share it):+That page is a runbook written for agents. Following it, the agent will:-```json-{ "mcpServers": { "repowise": { "command": "repowise", "args": ["mcp"] } } }-```-</details>--<details><summary><b>Codex CLI</b></summary>--```bash-codex mcp add repowise -- repowise mcp-```+1. Install repowise with the first installer it finds: `uv tool install`, then+ `pipx`, then `pip`.+2. Index the repository keyless (`repowise init --yes --no-prose`).+3. Wire the MCP server into the host it is running in.+4. Ask you one question: build the documentation from the index for free, or+ have an LLM write it. If you pick an LLM, it asks which provider, tells you+ which key variable to set in `.repowise/.env`, and shows you a cost estimate.+ Nothing is spent until you approve that estimate.+5. Ask you to restart the host, then call `get_overview` to confirm the tools+ work.-Or add to `~/.codex/config.toml`:+The agent never asks you to paste a key into the chat. The rest of this page+does the same steps by hand.-```toml-[mcp_servers.repowise]-command = "repowise"-args = ["mcp"]-```+## By hand-`repowise init --codex` writes project-local `.codex/config.toml`,-`.codex/hooks.json` and a managed `AGENTS.md`. See-[Codex integration](../agent/CODEX.md).-</details>+### 1. Install-<details><summary><b>Cursor</b></summary>+Python 3.11+ and Git are required. Use the first one you have:```bash-repowise agents add --target=cursor+uv tool install repowise # uv brings its own Python+pipx install repowise+python3 -m pip install repowise # Windows: python -m pip install repowise```-Writes project-local `.cursor/mcp.json` and a managed `.cursor/rules/repowise.mdc`.-Cursor does not read `.vscode/mcp.json`, so this is separate from the VS Code-setup below. See [Agent integrations](../agent/INTEGRATIONS.md).-</details>+Check it with `repowise --version`. Every LLM provider SDK ships in the base+package, so there is nothing extra to pick at install time.-<details><summary><b>OpenCode</b></summary>+If the shell says `command not found`, the install directory is not on `PATH`:+run `uv tool update-shell` or `pipx ensurepath` and open a new shell.-```bash-repowise agents add --target=opencode # this repo and this machine-repowise agents add --target=opencode --scope=project # this repo only-```--Writes `opencode.jsonc` (or an existing `opencode.json`) and a managed section in-`AGENTS.md`, in both scopes by default. The repo-local pair sits at the repo root;-the per-machine pair goes to `$XDG_CONFIG_HOME/opencode`, falling back to-`~/.config/opencode`. That path is the same on Windows: OpenCode does not read-`%APPDATA%`. The repo-local entry names its repo outright, while the per-machine one-resolves whichever repo OpenCode was launched in.--OpenCode accepts comments in its config. Repowise does not rewrite a file it cannot-parse as strict JSON, so if yours has comments it prints the entry to paste instead:-`repowise agents print-config opencode`. See-[Agent integrations](../agent/INTEGRATIONS.md).-</details>--<details><summary><b>Hermes</b></summary>--```bash-repowise agents add --target=hermes # this repo and this machine-repowise agents add --target=hermes --scope=user # the MCP server only-```--Hermes reads one `config.yaml` per machine, so the MCP server is registered there-and serves every repo: `%LOCALAPPDATA%\hermes\config.yaml` on Windows,-`~/.hermes/config.yaml` elsewhere, or `$HERMES_HOME/config.yaml` when that is set.-Project scope writes a managed section in `AGENTS.md`, which Hermes loads per repo.--Repowise edits that config in place rather than rewriting it, so your comments, key-order and any anchors survive. If it cannot parse the file it leaves it alone and-prints the entry to paste instead: `repowise agents print-config hermes`.--`platform_toolsets.cli` is deliberately left alone. Hermes exposes every enabled MCP-server to the CLI by default, and that list only becomes an allowlist once it already-names one. So repowise adds itself there **only** when the list is already an-allowlist, and never converts a permissive config into a restrictive one. See-[Hermes](../agent/HERMES.md).-</details>+### 2. Index-<details><summary><b>Cline, Windsurf, and other MCP clients</b></summary>--Print the server entry and paste it into whatever config the host reads:+From the repository root:```bash-repowise agents print-config claude-code+repowise init # interactive+repowise init --yes --no-prose # scripted: no questions, no key, no spend```-Or point the client at `repowise mcp`, run from the repo directory, over stdio.-</details>--<details><summary><b>VS Code</b></summary>--Install the **Repowise** extension from the Marketplace or Open VSX, then run-**Repowise: Set Up This Repository**. It registers the MCP server with VS Code-too, so the same index serves you and your agent. See [VS Code](../agent/VSCODE.md).-</details>+Bare `repowise init` scans the repo and asks how to index it: **Everything**+(model-written docs), **No prose** (docs rendered from the code's structure, no+key) or **Advanced** (every indexing and generation option). It shows a cost+estimate and waits for your confirmation before any model call.-**Now ask your agent something it would normally grep for:**+`--yes` never prompts. Pair it with `--no-prose` in scripts, CI and agent+setups: `--no-prose` keeps the run keyless even when an API key happens to be+set in the environment. Without it, `--yes` writes model prose whenever a key+is available, with the cost pre-approved.-> *"Use repowise `get_overview` to summarize this repo."*->-> *"What's the blast radius if I change `src/auth.py`? Use `get_context` with-> `include: ["callers"]`."*->-> *"Review my branch with `get_change_risk` for `main..HEAD`."*+The keyless run is the one to start with. It parses every file, builds the+dependency graph, reads git history, scores code health, finds dead code, and+renders a complete wiki from that structure. A large repository takes a few+minutes.-You should get a graph-grounded answer immediately, instead of a run of greps and-file reads. That is the whole point.+### 3. Connect your agent-> **What works without a key:** `get_overview`, `get_context`, `get_symbol`,-> `get_risk`, `get_change_risk`, `get_dead_code` and `get_health` all synthesize-> from the graph, git and health layers. `search_codebase` and `get_answer` read-> the wiki, which step 2 already built, so they answer from pages rendered from-> structure. `search_codebase` is full-text only until you configure an embedder.+Unless you pass `--no-editor-setup` (or set `REPOWISE_SKIP_EDITOR_SETUP=1`),+`init` wires these up for you, with or without `--yes`:-## 4. Optional: write the subsystem pages as model prose+| What | Where |+|---|---|+| MCP server for Claude Code and other clients that read it | `.mcp.json` at the repo root |+| Claude Code context file | `.claude/CLAUDE.md` |+| VS Code MCP entry | `.vscode/mcp.json` |+| Claude Code and Claude Desktop MCP registration | `~/.claude/settings.json` and the Claude Desktop config |+| Claude Code hooks (session briefing, post-edit context) | `~/.claude/settings.json` |+| Post-commit hook that runs `repowise update` | the repo's git `post-commit` hook |-Every file page was already deterministic, and stays that way. The one layer a-model adds is the subsystem (concept) tree: the numbered pages that describe how-the codebase fits together above the file level. Without a key they are-structural stubs; `repowise generate` fills them with prose (and unlocks-architectural decision mining and chat). You do not have to redo anything: it-reuses the index you already built, so the graph is not re-parsed.+`--no-editor-setup` skips every row: only `.repowise/` is written.+`--no-hook` skips just the post-commit hook. The command-rewrite hook for+[Distill](../agent/DISTILL.md) is offered as a question in an interactive run+and is left off under `--yes` unless you pass `--distill-hook`.-Set a key, preview the cost, then write:+Other hosts need one more command, run after `init`:-```bash-export ANTHROPIC_API_KEY="sk-ant-..." # or OPENAI_API_KEY / GEMINI_API_KEY--repowise generate # shows the wiki state, then writes the unwritten subsystem pages-```+| Host | Command |+|---|---|+| Claude Code | nothing more |+| VS Code (Copilot) | nothing more; the [extension](../agent/VSCODE.md) adds editor features |+| Codex | `repowise agents add --target=codex --yes`, or add `--codex` to `init` (writes `.codex/config.toml`, `.codex/hooks.json` and `AGENTS.md`) |+| Cursor | `repowise agents add --target=cursor --yes` |+| OpenCode | `repowise agents add --target=opencode --yes` |+| Hermes | `repowise agents add --target=hermes --yes` |+| Any other MCP client | `repowise agents print-config claude-code`, then paste the entry into the client's MCP config |-Bare `generate` prints the wiki's state and writes every unwritten subsystem-page behind a single cost estimate. On Windows PowerShell:-`$env:ANTHROPIC_API_KEY = "sk-ant-..."`+`repowise agents` lists every supported host and whether it is wired.+`--scope=project` or `--scope=user` limits `agents add` to repo-local or+per-machine config. Per-host details: [Agent integrations](../agent/INTEGRATIONS.md),+[Codex](../agent/CODEX.md), [OpenCode](../agent/OPENCODE.md),+[Hermes](../agent/HERMES.md).-You do not have to write them all at once. Restrict to an area, a single page, or-refresh what is stale, each run behind its own cost estimate:+On Claude Code, the plugin adds slash commands and skills on top:-```bash-repowise generate --path src/api # just the subsystem pages under one area-repowise generate --page module_page:src/api # or a single subsystem page-repowise generate --stale # refresh pages the last update marked stale-repowise generate --all # or rewrite every page, structural ones included+```text+/plugin marketplace add repowise-dev/repowise+/plugin install repowise@repowise```-Semantic search is a separate step: configure an embedder and run `repowise-reindex` to build the vector store.+**Restart the host.** Most agents read MCP config only at startup. Claude Code+may ask once to approve the project's `repowise` server; `/mcp` shows its+status. Then ask the agent to call `get_overview`. An architecture summary back+means the setup works.-Prefer to write the whole wiki as part of a fresh index instead? `repowise init`-on its own is interactive: it asks which provider to use, shows a cost estimate,-and waits for you to confirm before spending anything.+## First questions to ask-Three ways to avoid paying a provider at all:+Ask in plain language. The agent picks the tool.-- **Codex subscription:** `repowise init --provider codex_cli` uses your existing- Codex CLI login, no API key. Run `codex login` first.-- **Fully local:** point it at Ollama with a local embedding model for zero- external calls. See [Config](../reference/CONFIG.md).-- **Stay in index-only mode.** The graph, git, health, risk and dead-code layers- never needed a provider, and the wiki you already have was rendered without- one. Full-text search works on it; semantic search is the part that needs an- embedder, and Ollama is the keyless one.+| Ask | Tool it reaches for |+|---|---|+| "Give me a tour of this codebase." | `get_overview` |+| "What breaks if I change `src/auth.py`?" | `get_context` with callers, `get_risk` |+| "What code is dead here?" | `get_dead_code` |+| "What should we fix first?" | `get_health` |+| "Which tests should I run for my branch against `main`?" | `get_change_risk`, `get_risk` |-## 5. Keep it in sync+Every one of these works without an API key. What each tool returns:+[MCP Tools](../agent/MCP_TOOLS.md).-An index that drifts is worse than no index, because your agent will not know it-is stale. Every response carries the indexed commit and warns when it diverges-from your live `HEAD`, but the fix is cheap:+The same answers are on the command line:```bash-repowise update # incremental, seconds+repowise health # lowest-scoring files, and why+repowise dead-code # what nothing references+repowise risk main..HEAD # review priority for your branch+repowise impacted-tests main..HEAD # tests your branch's changes exercise```-Better, make it automatic:--```bash-repowise hook install # re-index on every commit-repowise watch # or run a file watcher while you work-```--Working with `git worktree`? A new worktree seeds its index from your main-checkout on the first `init` or `update`, so there is no second full index and-nothing to configure. See [Worktrees](../scale/WORKTREES.md).--All the sync options (hooks, watcher, GitHub/GitLab webhooks, polling):-[Auto-Sync](../scale/AUTO_SYNC.md).-------## See it+## Open the dashboard```bashrepowise serve```-Starts the API on `http://localhost:7337` and the web dashboard on-`http://localhost:3000`, alongside the MCP server. If Node.js 20+ is installed-the dashboard starts automatically; the frontend downloads once (~50 MB) and-caches in `~/.repowise/web/`. Use `repowise serve --no-ui` for the API alone, or-run the [Docker image](../../docker/README.md) if you would rather not install Node.+The API starts on `http://localhost:7337` and the dashboard on+`http://localhost:3000`. The dashboard needs Node.js 20+; it downloads once+(about 50 MB) and is cached in `~/.repowise/web/`. `--no-ui` starts the API+alone, and the [Docker image](../../docker/README.md) avoids installing Node.+View by view: [Dashboard](DASHBOARD.md).-Every view and what it answers: [Dashboard](DASHBOARD.md).+## Optional upgrades-## Spend fewer tokens on command output+**Model-written docs.** A keyless index renders the subsystem pages from+structure. `repowise generate` rewrites them as prose that explains how the+code fits together, reusing the index you already have. Put the key in+`.repowise/.env` (gitignored; loaded by `generate`, `update` and the MCP+server), preview, then write:```bash-repowise distill pytest -x # errors first, raw output recoverable via `repowise expand`-repowise saved # tokens and dollars saved so far+repowise generate --provider anthropic --dry-run # cost estimate only+repowise generate --provider anthropic # asks before spending```-Distill compresses noisy command output before your agent reads it, 60-90% fewer-tokens on noisy commands with no error lines dropped. Opt into the rewrite hook-during `init` (or `repowise hook rewrite install`) to have it applied-automatically. Rewrites run without a prompt and only ever wrap a recognized-command; set `permission: ask` to review each one.-See [Distill](../agent/DISTILL.md).+`--path <dir>` or `--page <id>` writes a subset. Providers that need no key:+`ollama` (local), `claude_cli` and `codex_cli` (your subscription login),+`opencode`. Full list and key variables: [Config](../reference/CONFIG.md).-## More than one repo+**Semantic search.** Full-text search works on any index. For semantic search,+configure an embedder (`REPOWISE_EMBEDDER`, or `--embedder` on `init`) and run+`repowise reindex`.-If your project spans several repositories, index the parent directory instead:+**Resuming.** If a run was interrupted, or a provider outage left some pages+missing, `repowise init --resume` writes only the pages that are absent.-```bash-cd my-workspace/ # contains backend/, frontend/, shared-libs/-repowise init .-```+## Keeping it fresh-repowise finds the git repos, asks which to index, and then runs the analysis that-only makes sense across repos: co-change pairs, API contracts between a producer-and its consumers, and package dependencies. One MCP server serves all of them.+The post-commit hook `init` installed runs `repowise update` after every+commit, so there is usually nothing to do. Other options:```bash-repowise workspace list # repos and their status-repowise workspace add ../new-svc # add one-repowise update --workspace # update every stale repo+repowise update # incremental catch-up, by hand+repowise watch # update continuously while you edit+repowise status # what is indexed and how far behind HEAD```-Full guide: [Workspaces](../scale/WORKSPACES.md).-------## If something looks wrong--```bash-repowise doctor # checks install, API keys, index drift, store health-repowise doctor --repair # fixes what it safely can-repowise status # what is indexed, and how stale it is-```--## Taking it back out--```bash-repowise uninstall # list what repowise wrote, then ask what to remove-repowise uninstall --all # everything: wiring, index, and machine-wide state-repowise uninstall --dry-run # the same list, changing nothing-```--It reports every path it removed and every path it left, with the reason on the-row. The index is not selected by default, because rebuilding it is the one-expensive thing here. Full flags in the-[CLI Reference](../reference/CLI_REFERENCE.md#repowise-uninstall-path).--## Environment variables--None of these are required. Every one of them is only needed for a model-written-wiki.--| Variable | When needed | Description |-|----------|-------------|-------------|-| `REPOWISE_PROVIDER` | Optional | Provider name. An empty value is treated as unset. |-| `ANTHROPIC_API_KEY` | Using Anthropic | Anthropic API key |-| `OPENAI_API_KEY` | Using OpenAI | OpenAI API key |-| `GEMINI_API_KEY` | Using Gemini | Google Gemini API key |-| `REPOWISE_EMBEDDER` | Semantic search | Embedder: `gemini`, `openai`, or `mock` (default) |-| `REPOWISE_DB_URL` | Custom database | SQLite/PostgreSQL connection string (default: `.repowise/wiki.db`) |-| `REPOWISE_API_URL` | Frontend only | Backend URL for the web UI (default: `http://localhost:7337`) |--Full list, plus `.repowise/config.yaml`: [Config](../reference/CONFIG.md).-----+`repowise hook uninstall` removes the post-commit hook. Webhooks, polling and+workspace-wide sync: [Auto-Sync](../scale/AUTO_SYNC.md). A new `git worktree`+seeds its index from the main checkout: [Worktrees](../scale/WORKTREES.md).## Where to go next-- **[User Guide](USER_GUIDE.md)** for the everyday workflows-- **[MCP Tools](../agent/MCP_TOOLS.md)** for what each tool answers, with worked examples-- **[Hooks](../agent/HOOKS.md)** to have context arrive without the agent asking for it-- **[Code Health](../layers/CODE_HEALTH.md)** for what the score measures and how it is validated-- **[Dashboard](DASHBOARD.md)** for the web UI, view by view+| Goal | Read |+|---|---|+| Everyday workflows: before a PR, in CI, changing provider | [User Guide](USER_GUIDE.md) |+| What each MCP tool answers | [MCP Tools](../agent/MCP_TOOLS.md) |+| Context that arrives without the agent asking | [Hooks](../agent/HOOKS.md) |+| Fewer tokens on test and build output | [Distill](../agent/DISTILL.md) |+| What the health score measures | [Code Health](../layers/CODE_HEALTH.md) |+| Unused files and exports | [Dead Code](../layers/DEAD_CODE.md) |+| Review priority for a change | [Change Risk](../layers/CHANGE_RISK.md) |+| Gates and annotations on pull requests | [CI](CI.md) |+| Several repositories as one | [Workspaces](../scale/WORKSPACES.md) |+| Every setting and environment variable | [Config](../reference/CONFIG.md) |+| Something is not working | [Troubleshooting](TROUBLESHOOTING.md) |++## Troubleshooting++Start with `repowise doctor`. It checks the install, API keys, index drift,+store health and each wired agent's MCP entry.++**`repowise: command not found` after install.** Run `uv tool update-shell`+(uv) or `pipx ensurepath` (pipx), open a new shell, and restart the agent+host, which starts `repowise` by name.++**The agent does not see the repowise tools.** Restart the host first. If they+are still missing, run `repowise doctor --repair`, which re-registers a stuck+Claude Code MCP entry and refreshes every wired agent's config. Failing that,+re-run `repowise agents add --target=<id> --yes`.++**The agent answers from old code.** `repowise status` shows drift and+`repowise update` catches up.++More, including large repositories, Windows encoding, empty results and+removing repowise: [Troubleshooting](TROUBLESHOOTING.md).diff --git a/docs/start/TROUBLESHOOTING.md b/docs/start/TROUBLESHOOTING.mdnew file mode 100644index 000000000..594232103--- /dev/null+++ b/docs/start/TROUBLESHOOTING.md@@ -0,0 +1,114 @@+# Troubleshooting++Start with:++```bash+repowise doctor # install, API keys, index drift, store health, wired agents+repowise doctor --repair # fix what it safely can+repowise status # what is indexed, and how far behind HEAD+```++`doctor --repair` reconciles drift between the database and the search stores,+re-registers a stuck Claude Code MCP entry, and refreshes the config of every+wired agent. It does not rewrite stale pages: `repowise update` does that.++## Install++**`repowise: command not found` after install.** The install directory is not+on `PATH`.++- uv: `uv tool update-shell`, then open a new shell. `uv tool dir --bin` prints+ the directory.+- pipx: `pipx ensurepath`, then open a new shell.+- pip: add the scripts directory under `python3 -m site --user-base` (`bin` on+ macOS and Linux, `Scripts` on Windows) to `PATH`.++Agent hosts start `repowise` by name, so restart the host after fixing `PATH`.++**"Provider X requires the 'Y' package".** Every provider SDK ships with+repowise, so this means a broken or partial install. `pip install <package>`+clears it at once; reinstalling repowise (`uv tool install --reinstall+repowise`, or `pip install --force-reinstall repowise`) fixes the cause.++**Garbled symbols or encoding errors on Windows.** The CLI switches its own+output to UTF-8 and substitutes a placeholder for any glyph the console cannot+draw, so a run should never stop on an encoding error. If a wrapper script or+an older shell still fails, set `PYTHONIOENCODING=utf-8` before running it+(`$env:PYTHONIOENCODING = "utf-8"` in PowerShell).++## Indexing++**init was interrupted, or finished with pages missing.** A provider outage+or rate limit can fail individual pages while the run completes.+`repowise init --resume` writes only the pages that are absent and makes no+model call for the ones you have, even if you switched provider.++**Doctor reports zero pages.** Even a keyless run writes pages, so an empty+wiki means the run did not finish. Run `repowise init --resume`.++**Memory or time on a very large repository.** Options, roughly in order of+effect:++```bash+repowise init --yes --no-prose --mode fast # graph and essential git only; backfill later+REPOWISE_PARSE_WORKERS=2 repowise init # fewer parse processes, less memory+repowise init -x vendor/ -x 'generated/**' # exclude what you never edit+repowise init --max-file-pages 2000 # cap file pages, highest importance first+```++The parse pool defaults to at most eight worker processes; each one holds its+own parser, so lowering the count is the most direct memory lever.++**Indexing cost more than expected.** `repowise init --dry-run` shows the+estimate without writing a wiki, `--test-run` generates for the top 10 files+only, and `--skip-tests --skip-infra` narrows scope. Lower `--concurrency` if+you hit provider rate limits.++## Answers++**The agent does not see the repowise tools.** Restart the host; most read MCP+config only at startup. Then `repowise doctor --repair`, then re-run+`repowise agents add --target=<id> --yes` for your host.++**The agent answers from an old version of the code.** Every MCP response+carries the indexed commit and warns when it trails `HEAD`. Run+`repowise update`. To stop it recurring, keep the post-commit hook `init`+installs (`repowise hook install` puts it back) or run `repowise watch`. See+[Auto-Sync](../scale/AUTO_SYNC.md).++**An empty result.** Empty means "not found in what was analyzed", which is+not always "none exists". An empty caller list comes with a `*_basis` field+saying how much of that language's calls the graph resolved, and a response+whose `_meta` reports `degraded` failed to load part of the index. Read those+before concluding nothing calls a symbol or nothing is wrong.++**Semantic search returns nothing, or warns `embedder.mock_active`.** No real+embedder is configured, so search is full-text only. Set `REPOWISE_EMBEDDER`+(for example `gemini`, `openai` or `ollama`) and rebuild the vector store with+`repowise reindex`. `reindex` makes embedding calls only, no LLM calls, and+also repairs a corrupted vector store.++## FAQ++**Do I need an API key?** No. Without one, repowise builds the graph, git+history, code health, dead code, change risk and a full wiki rendered from+structure. Every MCP tool works; `get_answer` and `search_codebase` answer+from the structural pages. A key, or a keyless provider such as `ollama`,+`claude_cli` or `codex_cli`, adds model-written subsystem pages, decision+mining, `repowise ask` and dashboard chat.++**Where does the key go?** In `.repowise/.env`, which is gitignored.+`generate`, `update` and the MCP server load it. `repowise init` saves the key+it ran with there unless you pass `--no-save-key`.++**Will `--yes` spend money?** Only if a key is available and you did not pass+`--no-prose`. `repowise init --yes --no-prose` never calls a model.++**The dashboard shows no repositories.** The API and the dashboard must use+the same database. Check `REPOWISE_DB_URL` on the API and `REPOWISE_API_URL`+on the dashboard. `repowise serve` runs both together, so this only comes up+when you run them separately.++**How do I remove repowise?** `repowise uninstall --dry-run` lists everything+it wrote; `repowise uninstall` asks what to remove. The index is not selected+by default because rebuilding it is the expensive part.diff --git a/docs/start/USER_GUIDE.md b/docs/start/USER_GUIDE.mdindex 811e171a2..4dc59d805 100644--- a/docs/start/USER_GUIDE.md+++ b/docs/start/USER_GUIDE.md@@ -10,14 +10,13 @@ them, and it is the file that stays in sync with the code.| If you want | Go to ||---|---|| To get running in five minutes | [Quickstart](QUICKSTART.md) |+| To fix something that is not working | [Troubleshooting](TROUBLESHOOTING.md) || Every command and flag | [CLI Reference](../reference/CLI_REFERENCE.md) || Every config key and env var | [Config](../reference/CONFIG.md) || What each MCP tool answers | [MCP Tools](../agent/MCP_TOOLS.md) || The web dashboard, view by view | [Dashboard](DASHBOARD.md) || What a metric actually means | [Glossary](../reference/COMPUTED_GLOSSARY.md) |-----## Table of contents1. [Installation](#installation)@@ -31,27 +30,20 @@ them, and it is the file that stays in sync with the code.9. [Common workflows](#common-workflows)10. [Troubleshooting](#troubleshooting)-----## Installation-```bash-pip install repowise-```--That is the complete install. Every LLM provider SDK (Anthropic, OpenAI, Gemini,-and LiteLLM for 100+ others) ships in the base package, so there is no provider-extra to choose. You pick a provider at index time and can change it later.+Install with `uv tool install repowise`, `pipx install repowise` or+`pip install repowise` (Python 3.11+ and Git). Every LLM provider SDK ships in+the base package. Step by step, including letting your agent do it:+[Quickstart](QUICKSTART.md).-Two extras exist, and neither is about providers:+Two optional extras exist, and neither is about providers:```bashpip install "repowise[postgres]" # PostgreSQL + pgvector instead of SQLitepip install "repowise[graph-extra]" # optional graph algorithms via graspologic```-**Requirements:** Python 3.11+ and Git. On Windows, use `python -m pip install repowise`.-For local development against a clone:```bash@@ -61,17 +53,13 @@ uv sync --all-packagesuv run repowise --version```-Step-by-step first run: [Quickstart](QUICKSTART.md).------## The mental modelThree things happen, and only the first one is slow:1. **Index.** `repowise init` parses every file to an AST, builds the dependency- graph, reads git history, and scores code health. With a provider it also- generates a wiki page per file and module. This is the one expensive step.+ graph, reads git history, scores code health and renders the wiki. With a+ provider, a model also writes the subsystem pages. This is the one slow step.2. **Ask.** Everything after that is a read against the index: your agent through[MCP tools](../agent/MCP_TOOLS.md), you through the CLI or the dashboard.Nothing re-analyzes anything.@@ -85,29 +73,12 @@ it has diverged from your live `HEAD`, but the real fix is step 3.**Two modes, one upgrade path.** `--no-prose` gives you the graph, gitintelligence, code health, change risk and dead code, plus a complete wikirendered from the code's structure, with no LLM, no key and no network. Adding a-provider rewrites those pages as model-written prose and unlocks decision mining-and chat.--You do not have to choose up front, and you do not have to know the flag: bare-`repowise init` scans the repo and asks, with the keyless mode offered as one of-the three answers. Start keyless, then upgrade the wiki with-`repowise generate` when you are ready, a page, a directory, or the whole thing-at a time, each behind a cost estimate:--```bash-repowise generate # write the unwritten subsystem pages, behind one cost estimate-repowise generate --path src/api # or just the part you care about-repowise generate --all # or rewrite the prose on every subsystem page-```--Bare `generate` prints the wiki's state and writes the unwritten subsystem-pages behind a single cost estimate. `generate` reuses the persisted graph-instead of re-parsing it, so the upgrade is cheap. (`repowise update --full`-still does the whole wiki in one shot if you prefer.) Semantic search is-separate: it needs an embedder, and `repowise reindex` is what builds the vector-store.-----+provider rewrites the subsystem pages as model-written prose and unlocks+decision mining and chat. Start keyless and upgrade later with+`repowise generate`, which reuses the stored graph; see+[Writing the docs with a model](#writing-the-docs-with-a-model). Semantic+search is separate: it needs an embedder, and `repowise reindex` builds the+vector store.## What gets created@@ -118,25 +89,28 @@ your-repo/│ ├── state.json # sync metadata (last commit, pages, tokens used)│ ├── config.yaml # provider, model, embedder, excludes│ ├── .env # saved API keys (gitignored)-│ └── lancedb/ # vector store for semantic search-├── .mcp.json # MCP server entry, so Claude Code and other clients find it+│ └── ... # search indexes and caches+├── .mcp.json # MCP server entry for Claude Code and other clients├── .claude/CLAUDE.md # generated Claude Code context-├── AGENTS.md # generated Codex context, when enabled-└── .codex/ # project-local Codex MCP/hooks config, with --codex+├── .vscode/mcp.json # VS Code MCP entry+├── AGENTS.md # generated agent context, with --codex or --agents+└── .codex/ # project-local Codex MCP and hooks config, with --codex```-`init` also registers repowise machine-wide by default: the Claude Code-(`~/.claude/settings.json`) and Claude Desktop MCP entry, plus the Claude Code-PostToolUse/SessionStart hooks. `--no-editor-setup` (or setting-`REPOWISE_SKIP_EDITOR_SETUP=1`) skips that machine-wide registration; every-project-local file above is still written.+Outside the tree, `init` also registers the MCP server with Claude Code+(`~/.claude/settings.json`) and Claude Desktop, adds the Claude Code+PostToolUse and SessionStart hooks, and installs a git post-commit hook that+runs `repowise update`.++`--no-editor-setup` (or `REPOWISE_SKIP_EDITOR_SETUP=1`) skips all of it, the+project-local files and the machine-wide registration alike: only `.repowise/`+is written, and `repowise mcp .` prints the config to connect a client by hand.+`--no-hook` skips only the post-commit hook.`.repowise/` is safe to delete and rebuild, and safe to gitignore. Committing itis a reasonable choice for a team that wants everyone on the same index withouteach person paying to generate it.-----## The commands you will actually useGrouped by what you are trying to do. Every flag for every command lives in the@@ -146,11 +120,11 @@ Grouped by what you are trying to do. Every flag for every command lives in the| Command | What it is for ||---|---|-| `repowise init` | First index, and the one command to start with. Bare `init` scans the repo and asks how to index it: everything, no prose, or advanced (every indexing and generation knob). Nothing is spent before an estimate is shown and confirmed. `--no-prose -y` and `--prose -y` skip the questions for scripts and agents. |+| `repowise init` | First index, and the one command to start with. Bare `init` scans the repo and asks how to index it: everything, no prose, or advanced (every indexing and generation knob). Nothing is spent before an estimate is shown and confirmed. `--yes --no-prose` skips the questions and stays keyless, for scripts and agents. || `repowise generate` | Write wiki pages with a model, on demand. The upgrade path for a keyless repo: `--unwritten` (default) writes everything still on a template, `--path`/`--page` writes a subset, all behind a cost estimate. || `repowise update` | Incremental catch-up after pulling or committing. Seconds, not minutes. || `repowise watch` | File watcher that updates continuously while you work. |-| `repowise hook install` | Post-commit hook so syncing happens without you. |+| `repowise hook install` | Restore the post-commit hook `init` installs, if you removed or skipped it. || `repowise status` | What is indexed, and how far behind it is. || `repowise doctor` | Checks install, keys, index drift, store health. `--repair` fixes what it safely can. |@@ -205,22 +179,22 @@ Grouped by what you are trying to do. Every flag for every command lives in the| `repowise telemetry disable` | Turn off anonymous usage telemetry. || `repowise uninstall` | Remove what repowise wrote from this repo, and optionally this machine. It lists everything first, and says what it left and why. |-----## Working with your agentThis is the main event, and it has its own docs. The short version:-**Connect once.** `repowise mcp`, run from the repo directory, over stdio. For-Claude Code the plugin wires the MCP server, hooks and slash commands together;-for Codex, `repowise init --codex` writes project-local config. Setup per client-is in [Quickstart](QUICKSTART.md#3-connect-your-agent), and-[Codex](../agent/CODEX.md) / [opencode](../agent/OPENCODE.md) have their own guides.+**Connect once.** `repowise init` wires Claude Code (MCP server and hooks) and+VS Code itself. Codex needs `--codex` on `init`; Cursor, OpenCode and Hermes+need `repowise agents add --target=<id>`. The Claude Code plugin is optional+and adds slash commands and skills. The per-host table is in+[Quickstart](QUICKSTART.md#3-connect-your-agent); [Codex](../agent/CODEX.md),+[OpenCode](../agent/OPENCODE.md) and [Hermes](../agent/HERMES.md) have their+own guides.**Ten tools, task-shaped.** Your agent gets architecture summaries, per-filetriage cards with callers and ownership, symbol source with exact bounds, riskassessment for a set of changed files, decision lookups and health scores, each in-one call rather than a chain. What each one answers, and worked multi-tool+one call, not a chain of greps. What each one answers, and worked multi-toolexamples: [MCP Tools](../agent/MCP_TOOLS.md).**Context that arrives unasked.** Hooks push the relevant thing into the session@@ -232,16 +206,15 @@ Inventory and exact settings: [Hooks](../agent/HOOKS.md).**Agents that do not speak MCP** still benefit, because `init` generates`CLAUDE.md` and `AGENTS.md` from the real index.-----## Keeping the index fresh-Five ways, in rough order of how little thought they need. Full guide:+`init` installs the post-commit hook by default, so most repos need nothing+more. The options, in rough order of how little thought they need. Full guide:[Auto-Sync](../scale/AUTO_SYNC.md).| Method | Command | Best for ||--------|---------|----------|-| Post-commit hook | `repowise hook install` | Set-and-forget local dev |+| Post-commit hook | installed by `init`; `repowise hook install` restores it | Set-and-forget local dev || File watcher | `repowise watch` | Active development sessions || GitHub webhook | Server endpoint | Teams, CI/CD || GitLab webhook | Server endpoint | Teams, CI/CD |@@ -253,8 +226,6 @@ Working in a `git worktree`? A new worktree seeds its index from your maincheckout on the first `init` or `update`, so there is no second full index andnothing to configure. See [Worktrees](../scale/WORKTREES.md).-----## Spending fewer tokensMost of an agent's context goes to command output it never needed: 300 lines of@@ -265,16 +236,8 @@ repowise distill pytest -x # errors first, exit code preservedrepowise distill git log -50 # subjects and counts instead of full bodies```-Nothing is lost. Omissions leave a marker that is always recoverable:--```-[repowise#a1b2c3d4e5f6: 230 lines omitted (~6.1k tokens); restore: repowise expand a1b2c3d4e5f6]-```--```bash-repowise expand a1b2c3d4e5f6 # the full original output-repowise expand a1b2c3d4e5f6 -q "FAILED" # just the matching lines-```+Nothing is lost. Omissions leave a `[repowise#<ref>: N lines omitted]` marker;+`repowise expand <ref>` returns the full output, and `-q <regex>` filters it.**Getting your agent to use it.** `repowise init` adds a section to the managed`CLAUDE.md` so the agent reaches for it voluntarily, which works in any agent that@@ -285,95 +248,75 @@ hook, which rewrites noisy commands automatically:repowise hook rewrite install # or answer Yes at the init prompt```-It never rewrites compound commands, redirections or watch modes. The one pipe-shape it does handle (on macOS/Linux) is a single stage into `head`, `tail`,-`grep` or `rg`, which runs unchanged inside distill's own shell.+It never rewrites compound commands, redirections or watch modes.Track it with `repowise saved`, or the Costs page in the dashboard. Full guide:[Distill](../agent/DISTILL.md).-----## The dashboard```bashrepowise serve```-API on `http://localhost:7337`, dashboard on `http://localhost:3000`, MCP server-alongside both. With Node.js 20+ the frontend downloads once (~50 MB), caches in-`~/.repowise/web/` and starts automatically. Use `--no-ui` for the API alone, or-the [Docker image](../../docker/README.md) if you would rather not install Node.+API on `http://localhost:7337`, dashboard on `http://localhost:3000`. Node.js+20+, `--no-ui` and the Docker image are covered in+[Quickstart](QUICKSTART.md#open-the-dashboard).`Ctrl+K` / `Cmd+K` opens a command palette from any page, which is the fastest wayto move between views and repos.-An index-only repo has a full Docs section, rendered from structure rather than-written by a model, so the pages read as structural summaries. Run `repowise-generate` to write any of them as model prose later. Chat still needs a provider.-Everything else, including Architecture, Code Health, Files, Commits and-Contributors, works off the parsed graph and git history alone.+An index-only repo has a full Docs section rendered from structure; `repowise+generate` rewrites any of it as model prose later. Chat needs a provider.+Everything else works off the parsed graph and git history alone.Every view and what it answers: **[Dashboard](DASHBOARD.md)**.-----## Common workflows-### First index, single repo+### Writing the docs with a model-```bash-pip install repowise-cd /path/to/your-project-repowise init # asks how to index: everything, no prose, or advanced-repowise hook install # keep it current from here on-```--Scripting it, or running it as an agent? Name the mode and skip the questions:--```bash-repowise init --no-prose -y # free, no key, seconds-repowise init --prose -y # model-written subsystem pages, cost pre-approved-```--Write the pages with a model when you want them, all at once or a piece at a-time, each behind a cost estimate:+First-time setup is in the [Quickstart](QUICKSTART.md). When you want the+subsystem pages written as prose, all at once or a piece at a time, each run+behind a cost estimate, put the key in `.repowise/.env` and run:```bash-export ANTHROPIC_API_KEY="sk-ant-..."-repowise generate # write the unwritten subsystem pages, behind one cost estimate-repowise generate --path src/api # or just one area first-repowise generate --all # or rewrite the prose on every subsystem page+repowise generate # the unwritten subsystem pages, behind one estimate+repowise generate --path src/api # or one area first+repowise generate --stale # refresh pages the last update marked stale+repowise generate --all # or rewrite every page```-Then add semantic search with `repowise reindex` once an embedder is configured.+`repowise update --full` does the whole wiki in one pass instead. Add semantic+search with `repowise reindex` once an embedder is configured.### First index, multi-repo workspace```bashcd /path/to/workspace/ # parent dir containing backend/, frontend/, ...repowise init . # finds the repos, asks which to index-repowise hook install --workspace```+The post-commit hook choice applies to every selected repo.+Cross-repo contracts and co-change come out of this automatically. See[Workspaces](../scale/WORKSPACES.md).### Day to day-Pick one and stop thinking about it:+The post-commit hook `init` installed keeps the index current on every commit.+For changes you pull, or if you skipped the hook:```bash-repowise hook install # A: syncs on every commit, nothing else to do-repowise watch # B: syncs continuously while you code-git pull && repowise update # C: manual, when you prefer control+git pull && repowise update # catch up by hand+repowise watch # or sync continuously while you code```### Before you open a pull request```bashrepowise risk main..HEAD # repo-relative review priority and supporting evidence-repowise impacted-tests # the tests this diff actually exercises+repowise impacted-tests main..HEAD # the tests your changes exerciserepowise health --trend # did anything you touched get worse```@@ -388,8 +331,8 @@ repowise risk origin/main..their-branchrepowise health --file path/to/the/scariest/file.py```-Or let the [PR bot](https://github.com/apps/repowise-bot) post it automatically on-every pull request, with no LLM calls.+To run the same checks on every pull request, with gates, annotations and+SARIF, use the GitHub Action: see [CI](CI.md).### Onboarding someone new@@ -407,7 +350,8 @@ instead of them interrupting someone.### In CI```bash-repowise init --no-prose -y # free, no keys in CI, no questions+export REPOWISE_SKIP_EDITOR_SETUP=1 # no hooks or editor files on the runner+repowise init --yes --no-prose # free, no keys in CI, no questionsrepowise risk "$BASE_SHA..$HEAD_SHA" # gate or annotate on review priorityrepowise export --format markdown --output ./docs/wiki/ # static hosting```@@ -435,76 +379,16 @@ health layers are provider-independent, so switching only affects generated pros### Cutting cost on a large repo-```bash-repowise init --dry-run # see the estimate before spending-repowise init --test-run # generate the top 10 files only-repowise init --skip-tests --skip-infra # narrow the scope-repowise init --provider ollama # or spend nothing at all-```-----+`repowise init --dry-run` shows the estimate before anything is spent; more+levers, including memory on very large repositories, are in+[Troubleshooting](TROUBLESHOOTING.md#indexing).## Troubleshooting-Start with `repowise doctor`, which checks the install, API keys, index drift and-store health, and `repowise doctor --repair` to fix what it safely can.--**"Provider X requires the 'Y' package"**-Every provider SDK ships with `repowise`, so this points at a broken or partial-install rather than a missing extra. The error names the package it wants, so-`pip install <package>` clears it immediately, and-`pip install --force-reinstall repowise` fixes the underlying install.--**Empty results in semantic search mode**-An embedder is probably not configured. Set `REPOWISE_EMBEDDER=gemini` (or-`openai`) and rebuild the vector store with `repowise reindex --embedder gemini`.--**"embedder.mock_active" warning**-The mock embedder produces random vectors, so semantic search cannot work-meaningfully. Set a real embedder as above.--**Pages look stale after code changes**-`repowise update`. To stop it happening, `repowise hook install` or-`repowise watch`.--**The agent is answering from an old version of the code**-Check `repowise status` for drift. MCP responses carry the indexed commit and a-staleness warning, so if your agent is not surfacing that, it is worth asking it-to.--**Indexing cost more than expected**-Use `--dry-run` for the estimate before committing to a run, `--test-run` to-validate on 10 files, and `--skip-tests --skip-infra` to cut scope. Lower-`--concurrency` if you are hitting rate limits.--**init was interrupted**-`repowise init --resume` picks up from the last checkpoint.--**init finished but some pages are missing**-A provider outage or rate limit can fail individual pages while the run itself-completes. `repowise init --resume` writes only the pages that are absent and-skips every page you already have, with no model call for them. It matches on-the page, not on which model wrote it, so switching provider after an outage-still keeps the pages the old one produced.--**Vector store looks corrupted**-`repowise reindex` rebuilds it from the existing wiki pages, with no LLM calls.--**Doctor reports 0 pages**-init failed or was interrupted. Even `--no-prose` writes pages, so an empty-wiki means the run did not finish. Try `repowise init --resume`.--**Dashboard shows an empty repo list**-The backend and frontend have to point at the same database. Check `REPOWISE_DB_URL`-on the backend and `REPOWISE_API_URL` on the frontend.--**CORS errors in the browser**-`repowise serve` runs the API and the dashboard together, so this should not come-up in normal use; the backend allows all origins by default. If you see it, the-API is probably not up, or you are running a frontend separately from source and-pointing it somewhere else with `REPOWISE_API_URL`.-----+Start with `repowise doctor`, and `repowise doctor --repair` to fix what it+safely can. Install problems, interrupted runs, very large repositories,+empty results and semantic search are covered in+[Troubleshooting](TROUBLESHOOTING.md).## Where to go nextdiff --git a/packages/cli/README.md b/packages/cli/README.mdindex c58b742eb..2170a4ce5 100644--- a/packages/cli/README.md+++ b/packages/cli/README.md@@ -1,429 +1,34 @@# repowise-cli-Command-line interface for repowise — generate, maintain, search, and export AI-powered codebase documentation from your terminal.+The `repowise` command-line interface: `init`, `update`, `serve`, `mcp`, `risk`,+`health`, `dead-code`, `coverage`, `distill` and the rest of the commands.-**Python >= 3.11 · Apache-2.0**+**Python >= 3.11 · AGPL-3.0-or-later**----+## Where it sits-## Installation+This is one of the Python packages in the repowise monorepo. It drives the+pipelines in `packages/core` and starts the API and MCP servers from+`packages/server`. Commands live in `src/repowise/cli/commands/`, and the entry+point is `src/repowise/cli/main.py`.-```bash-pip install repowise-cli--# Recommended: use uv-uv pip install repowise-cli-```--`repowise-core` is installed automatically as a dependency. The `repowise` command is available immediately after installation.-------## Quick Start--```bash-# 1. Generate documentation for the first time-cd /path/to/your/repo-repowise init --provider anthropic--# 2. Check what was generated-repowise status--# 3. Keep docs in sync after commits-repowise update--# 4. Search the wiki-repowise search "authentication flow"--# 5. Start the web UI-repowise serve-```--All data is stored in a `.repowise/` directory at the root of your repository.-------## Commands--### `repowise init`--Generate full wiki documentation from scratch. This is the expensive, one-time operation — it traverses every file, builds the dependency graph, and calls the LLM to generate all wiki pages in dependency order. It is resumable: if interrupted, re-run with `--resume`.--```-repowise init [PATH] [OPTIONS]-```+It is not published to PyPI on its own. Users get it inside the single+`repowise` distribution built from the root `pyproject.toml`.-| Option | Default | Description |-|--------|---------|-------------|-| `--provider` | `anthropic` | LLM provider: `anthropic`, `openai`, `ollama`, `litellm` |-| `--model` | provider default | Model identifier (e.g., `claude-opus-4-6`, `gpt-4o`) |-| `--concurrency` | `10` | Max simultaneous LLM calls |-| `--skip-tests` | off | Skip test files during generation |-| `--resume` | off | Resume an interrupted init job from last checkpoint |-| `--dry-run` | off | Show generation plan and cost estimate without calling the LLM |+## Install the product```bash-# Current directory, Anthropic (default)+uv tool install repowise # or: pipx install repowise / pip install repowise+cd your-reporepowise init--# Specific path and provider-repowise init /path/to/repo --provider openai --model gpt-4o--# Fully offline with Ollama-repowise init --provider ollama --model llama3.2--# Preview cost before spending tokens-repowise init --dry-run--# Resume after an interruption-repowise init --resume-```--**Prompt caching:** The Anthropic provider caches the shared system prompt and repository context across all generation calls. On large repos, this typically cuts cost by 60–90%.--See [`docs/reference/CLI_REFERENCE.md`](../../docs/reference/CLI_REFERENCE.md) for the full current `init` flag set (`--prose` / `--no-prose`, `--index-only`, etc.).-------### `repowise update`--Incrementally regenerate wiki pages for files changed since the last sync. Uses git diff to detect what changed and propagates changes through the dependency graph — only pages that actually need updating are regenerated.--```-repowise update [PATH] [OPTIONS]-```--| Option | Default | Description |-|--------|---------|-------------|-| `--provider` | last used | LLM provider |-| `--model` | provider default | Model override |-| `--since` | last sync commit | Base git ref to diff from (overrides saved state) |-| `--cascade-budget` | `30` | Max pages to regenerate per run |-| `--dry-run` | off | Show affected pages without regenerating |--```bash-# Update current repo (diffs HEAD against last sync commit)-repowise update--# Preview what would be regenerated-repowise update --dry-run--# Diff from a specific branch or commit-repowise update --since main-repowise update --since a1b2c3d-```--`repowise update` requires that `repowise init` has been run first. It reads `last_sync_commit` from `.repowise/state.json`.-------### `repowise watch`--Watch a repository for file changes and automatically run `repowise update` after a debounce period. Useful during active development: it indexes the working tree, so staged, unstaged and untracked files reach the index without a commit.--```-repowise watch [PATH] [OPTIONS]-```--| Option | Default | Description |-|--------|---------|-------------|-| `--provider` | last used | LLM provider |-| `--model` | provider default | Model override |-| `--debounce` | `2000` | Milliseconds to wait after the last change before updating |-| `--index-only` | off | Skip LLM page regeneration on every trigger |--```bash-# Watch current directory-repowise watch--# Custom debounce (wait 5 seconds after last change)-repowise watch --debounce 5000--# Keep the index current without a model call per save-repowise watch --index-only-```--Press `Ctrl+C` to stop. Writes to `.repowise/`, `.git/`, build output, and the files repowise manages itself (`CLAUDE.md`, `AGENTS.md`, `.mcp.json`) never trigger an update.-------### `repowise search`--Search wiki pages by keyword, meaning, or symbol name.--```-repowise search QUERY [PATH] [OPTIONS]-```--| Option | Default | Description |-|--------|---------|-------------|-| `--mode` | `fulltext` | Search mode: `fulltext`, `semantic`, `symbol` |-| `--limit` | `10` | Maximum results to return |--```bash-# Full-text search (SQLite FTS5)-repowise search "authentication"--# Semantic / vector search-repowise search "rate limiting strategy" --mode semantic--# Find a symbol by name-repowise search "AuthService" --mode symbol-```--`semantic` search requires the `[search]` extra (`pip install "repowise-core[search]"`).-------### `repowise export`--Export wiki pages to files on disk (Markdown, HTML, or JSON), or emit the-architecture model as [Structurizr DSL](https://structurizr.com).--```-repowise export [PATH] [OPTIONS]-```--| Option | Default | Description |-|--------|---------|-------------|-| `--format` | `markdown` | `markdown`, `html`, `json`, or `structurizr` |-| `--output`, `-o` | `.repowise/export` | Output directory (for `structurizr`, a path ending in `.dsl` names the file) |-| `--full` | off | JSON only: include decisions, dead code, hotspots, provenance |-| `--standalone` | off | Structurizr only: emit a complete workspace with default views |-| `--components` | off | Structurizr only: include the component level (one box per directory) |-| `--no-externals` | off | Structurizr only: omit third-party dependencies |--```bash-# Export as Markdown to .repowise/export/-repowise export--# Export as JSON to a custom directory-repowise export --format json -o ./wiki-export--# Export as HTML-repowise export --format html -o ./docs/wiki--# Architecture as Structurizr DSL (model fragment by default)-repowise export --format structurizr-repowise export --format structurizr --standalone -o arch/```-Markdown and HTML write one file per page. JSON writes a single-`wiki_pages.json`. Structurizr writes one `.dsl` file from the existing-graph (no LLM). See-[`docs/architecture/structurizr-export.md`](../../docs/architecture/structurizr-export.md)-and the full flag set in-[`docs/reference/CLI_REFERENCE.md`](../../docs/reference/CLI_REFERENCE.md).+## Develop on it----+From the repository root, `uv sync --all-packages` installs every package in+editable mode. See [CONTRIBUTING.md](../../.github/CONTRIBUTING.md).-### `repowise status`+## Documentation-Show sync state, page counts by type, and total token consumption.--```-repowise status [PATH]-```--Displays:--- Last sync commit hash-- Provider and model used for the last init/update-- Total token consumption (input + output)-- Page count broken down by type (`file_page`, `module_page`, `symbol_spotlight`, etc.)-------### `repowise doctor`--Run health checks on the wiki setup. Useful for debugging why `init` or `update` is failing.--```-repowise doctor [PATH]-```--Checks performed:--| Check | What it verifies |-|-------|-----------------|-| Git repository | The path is inside a git repo |-| `.repowise/` directory | The init directory exists |-| Database | `wiki.db` is connectable; shows page count |-| `state.json` | Valid JSON with a `last_sync_commit` entry |-| Providers | At least one LLM provider is importable |-| Stale pages | Number of pages whose source has changed since generation |--Exits with a summary of passed/failed checks.-------### `repowise dead-code`--Detect dead and unused code using the dependency graph. Does not call an LLM — purely graph analysis and git metadata.--```-repowise dead-code [PATH] [OPTIONS]-```--| Option | Default | Description |-|--------|---------|-------------|-| `--min-confidence` | `0.4` | Minimum confidence threshold (0.0–1.0) |-| `--safe-only` | off | Only show findings marked as safe to delete |-| `--kind` | all | Filter: `unreachable_file`, `unused_export`, `unused_internal`, `zombie_package` |-| `--format` | `table` | Output: `table`, `json`, `md` |--```bash-# All findings, table format-repowise dead-code--# Only safe-to-delete findings-repowise dead-code --safe-only--# Only unreachable files (no importers in the graph)-repowise dead-code --kind unreachable_file--# Export to JSON for CI integration-repowise dead-code --format json > dead-code.json--# High-confidence findings only-repowise dead-code --min-confidence 0.8-```--Finding kinds:--| Kind | Description |-|------|-------------|-| `unreachable_file` | File has no importers in the dependency graph |-| `unused_export` | Symbol is exported but not imported anywhere |-| `unused_internal` | Symbol is defined but never called within the repo |-| `zombie_package` | Package directory exists but has no live references |-------### `repowise serve`--Start the repowise REST API and web UI server. Requires `repowise-server` to be installed.--```-repowise serve [OPTIONS]-```--| Option | Default | Description |-|--------|---------|-------------|-| `--port` | `7337` | Port to listen on |-| `--host` | `127.0.0.1` | Host to bind to (`0.0.0.0` to expose externally) |-| `--workers` | `1` | Number of uvicorn workers |--```bash-# Start on localhost:7337-repowise serve--# Expose on all interfaces-repowise serve --host 0.0.0.0--# Custom port-repowise serve --port 8080-```--The API docs are available at `http://localhost:7337/docs`. Start the Next.js web UI separately and point `REPOWISE_API_URL` at this server.-------### `repowise mcp`--Start the MCP server for AI editor integration (Claude Code, Cursor, Cline).--```-repowise mcp [PATH] [OPTIONS]-```--| Option | Default | Description |-|--------|---------|-------------|-| `--transport` | `stdio` | Transport protocol: `stdio` (editors), `streamable-http` (HTTP clients), or `sse` (legacy) |-| `--port` | `7338` | Port for HTTP/SSE transports |--```bash-# stdio transport, current directory (typical editor integration)-repowise mcp--# Specific repository path-repowise mcp /path/to/repo--# Streamable HTTP transport for HTTP clients-repowise mcp --transport streamable-http--# Legacy SSE transport-repowise mcp --transport sse-```--Exposes 18 registered MCP tools (11 advertised by default in single-repo mode:-ten flagship tools plus `list_repos`) for querying wiki pages, symbols, the-dependency graph, git analytics, ownership data, hotspots, dead code findings,-code health, change risk, and decision intelligence. Full surface:-[`docs/agent/MCP_TOOLS.md`](../../docs/agent/MCP_TOOLS.md).--**Claude Code setup** — add to `~/.claude/settings.json`:--```json-{- "mcpServers": {- "repowise": {- "command": "repowise",- "args": ["mcp", "/path/to/your/repo"]- }- }-}-```--**Cursor / Cline setup** — add to your MCP configuration file:--```json-{- "mcpServers": {- "repowise": {- "command": "repowise",- "args": ["mcp", "/absolute/path/to/repo"]- }- }-}-```-------## Environment Variables--| Variable | Description |-|----------|-------------|-| `ANTHROPIC_API_KEY` | API key for the Anthropic provider |-| `OPENAI_API_KEY` | API key for the OpenAI provider |-| `REPOWISE_DB_URL` | Override the default SQLite path (e.g., `postgresql://user:pass@host/db`) |-| `REPOWISE_EMBEDDER` | Embedder backend: `mock` (default) or `gemini` |-------## The `.repowise/` Directory--After `repowise init`, your repo will contain:--```-.repowise/-├── wiki.db # SQLite database — all pages, symbols, jobs, git metadata-├── lancedb/ # LanceDB vector store (if [search] extra is installed)-├── graph.json # Serialized dependency graph (repos < 30K nodes)-├── state.json # Sync state: last_sync_commit, provider, model, token counts-└── export/ # Output directory for `repowise export`-```--Add `.repowise/` to your `.gitignore` to avoid committing generated data.-------## Development--```bash-# Install for development (from repo root)-uv pip install -e packages/cli -e packages/core--# Run tests-pytest tests/unit/cli/-```+- Every command and flag: [docs/reference/CLI_REFERENCE.md](../../docs/reference/CLI_REFERENCE.md)+- How the system is built: [docs/architecture/](../../docs/architecture/)diff --git a/packages/core/README.md b/packages/core/README.mdindex a7748d4b7..9600dc57b 100644--- a/packages/core/README.md+++ b/packages/core/README.md@@ -1,268 +1,37 @@# repowise-core-Core library for repowise — the ingestion pipeline, dependency graph engine, documentation generation system, and persistence layer. All other repowise packages depend on this.+The engine behind repowise: ingestion and parsing, the dependency graph, git+history, health and dead-code analysis, decisions, wiki generation, providers+and persistence. Every other Python package depends on it.-**Python >= 3.11 · Apache-2.0**+**Python >= 3.11 · AGPL-3.0-or-later**----+## Where it sits-## What's in this package+This is one of the Python packages in the repowise monorepo. Code lives under+`src/repowise/core/`; the main areas are `ingestion/`, `analysis/`,+`generation/`, `persistence/`, `pipeline/` and `providers/`.+`packages/cli` and `packages/server` call into it.-| Module | Purpose |-|--------|---------|-| `repowise.core.ingestion` | File traversal, AST parsing (tree-sitter), dependency graph, change detection |-| `repowise.core.generation` | Context assembly, page generation, Jinja2 prompt templates, job system |-| `repowise.core.persistence` | SQLAlchemy models, async CRUD, full-text search, vector store |-| `repowise.core.providers` | LLM provider abstraction (Anthropic, OpenAI, Ollama, LiteLLM) |-| `repowise.core.analysis` | Dead code detection (unreachable files, unused exports, zombie packages) |-| `repowise.core.rate_limiter` | Token-bucket rate limiter (RPM + TPM) for all LLM providers |+It is not published to PyPI on its own. Users get it inside the single+`repowise` distribution built from the root `pyproject.toml`.------## Installation+## Install the product```bash-# Core only (SQLite backend, no vector search)-pip install repowise-core--# With LanceDB vector search (recommended for local/single-server deployments)-pip install "repowise-core[search]"--# With pgvector (PostgreSQL deployments)-pip install "repowise-core[pgvector]"-```-------## Supported Languages--AST parsing uses [tree-sitter](https://tree-sitter.github.io/) with one `.scm` query file per language. Adding a new language means writing one `.scm` file — `ASTParser` itself does not change.--| Language | Extensions | Query file |-|----------|-----------|------------|-| Python | `.py` | `queries/python.scm` |-| TypeScript | `.ts`, `.tsx` | `queries/typescript.scm` |-| JavaScript | `.js`, `.jsx` | `queries/javascript.scm` |-| Go | `.go` | `queries/go.scm` |-| Rust | `.rs` | `queries/rust.scm` |-| Java | `.java` | `queries/java.scm` |-| C / C++ | `.c`, `.cpp`, `.h`, `.cc` | `queries/c.scm`, `queries/cpp.scm` |-| Ruby | `.rb` | `queries/ruby.scm` |-| Kotlin | `.kt` | `queries/kotlin.scm` |--**Special handlers** (purpose-built parsers, not tree-sitter):--- OpenAPI / Swagger specs (`.yaml`, `.json`)-- Protobuf (`.proto`)-- GraphQL schemas (`.graphql`)-- Dockerfiles-- GitHub Actions / CI YAML-------## LLM Providers--| Provider | Name | Notes |-|----------|------|-------|-| Anthropic | `anthropic` | Prompt caching on shared system/repo context (large-init cost reduction) |-| OpenAI | `openai` | OpenAI or any OpenAI-compatible endpoint, including local gateways |-| Ollama | `ollama` | Fully offline, no API key required |-| LiteLLM | `litellm` | 100+ providers through one interface (optional dependency) |-| Mock | `mock` | In-memory stub for tests — no network calls |--The `openai` adapter accepts custom `OPENAI_BASE_URL` values and discovers-model ids from compatible `/models` endpoints, so local gateways such as-9router can be used without a provider-specific adapter.--```python-from repowise.core.providers import get_provider--provider = get_provider("anthropic", api_key="sk-ant-...", model="claude-sonnet-4-6")-response = await provider.generate(system_prompt="...", user_prompt="...")-print(response.content)-print(f"{response.input_tokens} in / {response.output_tokens} out")-```--### Registering a custom provider--```python-from repowise.core.providers import register_provider-from my_package import MyProvider--register_provider("my_provider", lambda **kw: MyProvider(**kw))-provider = get_provider("my_provider", model="my-model-v1")-```-------## Key Modules--### Ingestion--`FileTraverser` walks a repository tree respecting `.gitignore`, `.repowiseIgnore`, a hardcoded blocklist (node_modules, `__pycache__`, build artifacts), auto-detected generated files, and binary files.--`ASTParser` is a single class for all languages. It loads the matching `.scm` query file, runs tree-sitter queries to extract symbols and imports, and returns a `ParsedFile` with a consistent shape regardless of language.--`GraphBuilder` converts all `ParsedFile` outputs into a `networkx.DiGraph` and computes PageRank, strongly connected components (SCCs), betweenness centrality, and community detection (Louvain). The resulting graph drives generation ordering, change propagation, and dead code detection.--`ChangeDetector` diffs two git refs with GitPython, identifies added/modified/deleted/renamed files, detects symbol renames using a signature-similarity heuristic, and determines which wiki pages need regeneration via cascade analysis.--```python-from repowise.core.ingestion import FileTraverser, ASTParser, GraphBuilder, ChangeDetector-from pathlib import Path--# Traverse-traverser = FileTraverser("/path/to/repo")-file_infos = list(traverser.traverse())--# Parse all files-parser = ASTParser()-graph_builder = GraphBuilder()-parsed_files = []--for fi in file_infos:- source = Path(fi.abs_path).read_bytes()- parsed = parser.parse_file(fi, source)- parsed_files.append(parsed)- graph_builder.add_file(parsed)--graph_builder.build() # PageRank, SCCs, centrality-graph = graph_builder.graph() # networkx.DiGraph--# Change detection-detector = ChangeDetector("/path/to/repo")-diffs = detector.get_changed_files("a1b2c3d", "HEAD")-# → [FileDiff(path="src/auth.py", status="modified"), ...]-```--### Generation--`ContextAssembler` builds the prompt context for each page. For every file it assembles (in priority order, dropping lower-priority items when over the 12K-token budget): full source, symbol signatures, graph metrics, git ownership and commit history, import summaries from already-generated pages, RAG context from the vector store, co-change partner pages, dead code findings, and reverse import context.--`PageGenerator` orchestrates hierarchical generation across 9 levels (API contracts → symbol spotlights → file pages → SCC pages → module pages → cross-package pages → repo overview → infra pages → index pages). Within each level, up to `concurrent_jobs` tasks run in parallel via `asyncio.Semaphore`.--`JobSystem` checkpoints progress after every completed page so that a long init job can be resumed after an interruption.--```python-from repowise.core.generation import ContextAssembler, GenerationConfig, PageGenerator--config = GenerationConfig()-assembler = ContextAssembler(config)-generator = PageGenerator(provider, assembler, config)--pages = await generator.generate_all(- parsed_files, source_map, graph_builder, repo_structure, repo_name="my-repo"-)--for page in pages:- print(page.title, page.page_type, f"{page.confidence_score:.2f}")+uv tool install repowise # or: pipx install repowise / pip install repowise+cd your-repo+repowise init```-### Persistence--Two SQL backends (SQLite / PostgreSQL) with identical async SQLAlchemy models. Pass a plain `sqlite:///...` or `postgresql://...` URL — `get_db_url()` injects the correct async driver automatically.--```python-from repowise.core.persistence import (- create_engine, init_db, create_session_factory,- get_session, upsert_repository, upsert_page_from_generated,-)--engine = create_engine("sqlite:///path/to/wiki.db")-await init_db(engine) # creates tables + FTS5 index (idempotent)-sf = create_session_factory(engine)--async with get_session(sf) as session:- repo = await upsert_repository(session, name="my-repo", local_path="/path/to/repo")- for page in pages:- await upsert_page_from_generated(session, page, repo.id)-```--Key tables: `repos`, `wiki_pages`, `page_versions`, `wiki_symbols`, `generation_jobs`, `webhook_events`, `graph_nodes`, `graph_edges`, `git_metadata`, `dead_code_findings`.--### Search--```python-from repowise.core.persistence import FullTextSearch, InMemoryVectorStore, MockEmbedder--# Full-text search (SQLite FTS5 or PostgreSQL GIN index)-fts = FullTextSearch(engine)-await fts.ensure_index()-results = await fts.search("authentication middleware", limit=10)--# Semantic search — swap MockEmbedder for a real embedder in production-store = InMemoryVectorStore(embedder=MockEmbedder())-results = await store.search("how does rate limiting work?", limit=5)-```--Vector store backends:--| Backend | When to use |-|---------|-------------|-| `InMemoryVectorStore` | Development / testing (no persistence) |-| `LanceDBVectorStore` | SQLite mode — embedded, no separate server, stored in `.repowise/lancedb/` |-| `PgVectorStore` | PostgreSQL mode — HNSW index via the `pgvector` extension |+## Develop on it-### Dead Code Detection+From the repository root, `uv sync --all-packages` installs every package in+editable mode. See [CONTRIBUTING.md](../../.github/CONTRIBUTING.md).-Finds unreachable files (in-degree 0 in the dependency graph), unused exports, unused internal symbols, and zombie packages. No LLM calls — purely graph analysis and git metadata.--```python-from repowise.core.analysis.dead_code import DeadCodeAnalyzer--analyzer = DeadCodeAnalyzer(graph, git_meta_map)-report = analyzer.analyze({"min_confidence": 0.5})--for finding in report.findings:- print(finding.kind.value, finding.file_path, f"{finding.confidence:.0%}")- if finding.safe_to_delete:- print(" -> safe to remove")--print(f"Deletable lines: {report.deletable_lines:,}")-```-------## Page Types--repowise generates these page types in a strict dependency-aware order. The-subsystem-and-overview set (`module_page`, `repo_overview`,-`architecture_diagram`, `onboarding`) is the one model-written layer: model prose-when a provider is configured, structural stubs otherwise. Every other type is-always rendered from structure, with no model and no key.--| Level | Page Type | Rendered | Description |-|-------|-----------|----------|-------------|-| 0 | `api_contract` | structure | OpenAPI specs, Protobuf, GraphQL schemas |-| 1 | `symbol_spotlight` | structure | High-PageRank symbols (top 10% by centrality) |-| 2 | `file_page` | structure | One page per source file |-| 3 | `scc_page` | structure | Circular dependency cluster summary |-| 4 | `module_page` | **model or stub** | Subsystem (concept) tree above the file level |-| 5 | `layer_page` | structure | Architectural layer summary |-| 6 | `repo_overview` | model or stub | Repository overview |-| 6 | `architecture_diagram` | model or stub | Repo-level architecture diagram |-| 7 | `infra_page` | structure | Dockerfile, CI YAML, Makefile documentation |-| 8 | `onboarding` | model or stub | Curated onboarding collection |-------## Database Backends--| Backend | Config | Vector search |-|---------|--------|---------------|-| SQLite + LanceDB | `sqlite:///path/to/wiki.db` + `[search]` extra | LanceDB embedded in `.repowise/lancedb/` |-| PostgreSQL + pgvector | `postgresql://user:pass@host/db` + `[pgvector]` extra | HNSW index on `wiki_pages.embedding` column |-------## Development--```bash-# Install for development (from repo root)-uv pip install -e "packages/core[search]"--# Run tests-pytest tests/unit/core/-```+## Documentation-See [ARCHITECTURE.md](../../docs/architecture/ARCHITECTURE.md) for a deep dive into how the ingestion pipeline, generation engine, and three stores fit together.+- How the system is built: [docs/architecture/](../../docs/architecture/), starting with+ [ARCHITECTURE.md](../../docs/architecture/ARCHITECTURE.md)+- Adding a language: [docs/architecture/language-support.md](../../docs/architecture/language-support.md)+- CLI commands that drive it: [docs/reference/CLI_REFERENCE.md](../../docs/reference/CLI_REFERENCE.md)diff --git a/packages/core/src/repowise/core/analysis/health/asserts/lexicon.py b/packages/core/src/repowise/core/analysis/health/asserts/lexicon.pyindex 22a4ddbd2..670a35c33 100644--- a/packages/core/src/repowise/core/analysis/health/asserts/lexicon.py+++ b/packages/core/src/repowise/core/analysis/health/asserts/lexicon.py@@ -27,7 +27,7 @@ and a narrow one for the same reason we keep two, but its wide list matchesname prefixes; measured across three test corpora, the ambiguous English verbsin that shape (``check``, ``validate``, ``approve``, ``fail``) matchedproduction functions under test and no assertions at all. Measurements and-reasoning in LANGUAGE_SUPPORT.md#code-health-coverage.+reasoning in docs/architecture/language-support.md#test-quality-markers-per-language.Both directions of error are understood. Over-counting shrinks a saturationratio and can only suppress a finding; under-counting inflates it. A missingdiff --git a/packages/core/src/repowise/core/analysis/health/biomarkers/assertion_free_test.py b/packages/core/src/repowise/core/analysis/health/biomarkers/assertion_free_test.pyindex 3cb3a099c..906264b2b 100644--- a/packages/core/src/repowise/core/analysis/health/biomarkers/assertion_free_test.py+++ b/packages/core/src/repowise/core/analysis/health/biomarkers/assertion_free_test.py@@ -31,7 +31,7 @@ predicate; it is deliberately not folded into ``assertion_count``, which iscalibrated.**This marker is advisory and never deducts.** See ``ADVISORY_DIMENSION`` in-``scoring.py``, and LANGUAGE_SUPPORT.md#code-health-coverage for the measured+``scoring.py``, and docs/architecture/language-support.md#test-quality-markers-per-language for the measuredper-language precision and the false positives it does not separate."""@@ -51,7 +51,7 @@ from .base import BiomarkerResult, FileContext#: sitting outside that pass's admitted set. Re-admitting either language is a#: measurement of its own, on its own hand-labelled sample, not a consequence of#: this one.-#: Figures and the reasoning: LANGUAGE_SUPPORT.md#code-health-coverage.+#: Figures and the reasoning: docs/architecture/language-support.md#test-quality-markers-per-language.SHIPPING_LANGUAGES = frozenset({"javascript", "python", "tsx", "typescript"})diff --git a/packages/core/src/repowise/core/analysis/health/complexity/assertions.py b/packages/core/src/repowise/core/analysis/health/complexity/assertions.pyindex 53bb83009..7c6378d2b 100644--- a/packages/core/src/repowise/core/analysis/health/complexity/assertions.py+++ b/packages/core/src/repowise/core/analysis/health/complexity/assertions.py@@ -118,7 +118,7 @@ def _split_callee_matches_assert(call_node: Node) -> bool:rather than in the narrow one on purpose: the narrow tier is what thecalibrated block markers count, and admitting these names there was measuredto move assertion runs. Promoting it is a calibrated change needing its own- before/after. LANGUAGE_SUPPORT.md#code-health-coverage.+ before/after. docs/architecture/language-support.md#test-quality-markers-per-language."""if call_node.child_by_field_name("function") is not None:return Falsediff --git a/packages/core/src/repowise/core/analysis/health/mocks/lexicon.py b/packages/core/src/repowise/core/analysis/health/mocks/lexicon.pyindex 77c992fb6..86c0669af 100644--- a/packages/core/src/repowise/core/analysis/health/mocks/lexicon.py+++ b/packages/core/src/repowise/core/analysis/health/mocks/lexicon.py@@ -154,7 +154,7 @@ _JS_TS = MockDialect(# Keyed by ``LanguageTag`` (``ingestion/models.py``); one dialect may serve# several tags. Go and Java are absent deliberately: neither has a trustworthy# assertion count, and that is this marker's denominator. Measurements and-# reasoning in LANGUAGE_SUPPORT.md#code-health-coverage.+# reasoning in docs/architecture/language-support.md#test-quality-markers-per-language.MOCK_DIALECTS: dict[str, MockDialect] = {"python": _PYTHON,"javascript": _JS_TS,diff --git a/packages/core/src/repowise/core/generation/page_generator/README.md b/packages/core/src/repowise/core/generation/page_generator/README.mdindex b34e8872e..e4cc80bf5 100644--- a/packages/core/src/repowise/core/generation/page_generator/README.md+++ b/packages/core/src/repowise/core/generation/page_generator/README.md@@ -47,14 +47,16 @@ from repowise.core.generation.page_generator import PageGenerator, SYSTEM_PROMPT- `orchestrate.py` — `_GenerationRun` (per-call state, level runner) and`run_generate_all` (the entry point `generate_all` delegates to).- `levels.py` — pure per-level coroutine builders reading the run state.-- `tiering.py` — `partition_file_tiers` (tier-1/tier-2 split by PageRank).+- `structural.py` — renderers that build a page from structure with no model+ call: the sole renderers (file pages among them) and the keyless stubs.- `prompts.py`, `validation.py`, `helpers.py` — constants and pure helpers.## Extension points- New page type: add a `generate_*` method to `pertype.py`, a level builder to`levels.py`, and wire it into `_GenerationRun.execute()`.-- New tier policy: extend `partition_file_tiers` / add a renderer template.+- New structural page: add a renderer to `structural.py` and its template under+ `templates/` (or `templates/stub/` for a keyless stub).- High-level synthesis evidence: `context/evidence.py` owns the provider-neutral,bounded selection and provenance channel shared by explicit`generation_context.files` and automatically derived exact references. Andiff --git a/packages/core/src/repowise/core/generation/page_generator/__init__.py b/packages/core/src/repowise/core/generation/page_generator/__init__.pyindex 64e94fac3..4ab4e0693 100644--- a/packages/core/src/repowise/core/generation/page_generator/__init__.py+++ b/packages/core/src/repowise/core/generation/page_generator/__init__.py@@ -9,7 +9,7 @@ Internal layout:pertype.py — per-page-type generate_* methods (mixin)orchestrate.py — level-by-level generate_all orchestrationlevels.py — per-level coroutine builders- tiering.py — tier-1/tier-2 partition (large-repo doc generation)+ structural.py — renderers that build a page from structure, no modelprompts.py — SYSTEM_PROMPTS constants + language namesvalidation.py — hallucination detection for LLM outputhelpers.py — pure helpers (summaries, significance, clone dedupe)diff --git a/packages/server/README.md b/packages/server/README.mdindex 85dfe853f..7cd83ecaa 100644--- a/packages/server/README.md+++ b/packages/server/README.md@@ -1,322 +1,43 @@# repowise-server-FastAPI REST API, webhook handlers, MCP server, and background job scheduler for repowise. This package powers the `repowise serve` command and the Next.js web UI backend.+The FastAPI REST API, the MCP server, webhook handlers and the background+scheduler. `repowise serve` and `repowise mcp` start it, and the web UI reads+from its API (port 7337 by default).-**Python >= 3.11 · Apache-2.0**+**Python >= 3.11 · AGPL-3.0-or-later**----+## Where it sits-## What's in this package+This is one of the Python packages in the repowise monorepo. Code lives under+`src/repowise/server/`: `app.py` builds the FastAPI app, `routers/` holds the+REST routes, and `mcp_server/` holds the tools: 18 registered MCP tools+(10 advertised by default in single-repo mode). It reads indexes built by+`packages/core`.-| Component | Description |-|-----------|-------------|-| **REST API** | FastAPI application with full CRUD for repos, pages (with version history), symbols, jobs, git analytics, dead code, decisions, graph intelligence, blast radius, costs, knowledge map, security findings, providers, chat, CLAUDE.md generation, and multi-repo workspace |-| **MCP Server** | 18 registered MCP tools (10 advertised by default in single-repo mode: the canonical set) for AI coding assistants (Claude Code, Cursor, Cline) |-| **Webhooks** | GitHub and GitLab push event handlers — trigger sync jobs automatically on push |-| **Scheduler** | APScheduler background jobs — polling fallback (auto-syncs diverged repos), stale page detection |+Seen from an MCP client, repowise registers 18 MCP tools and advertises **10 by default**.+Workspace mode adds `list_repos`, and the other seven are opt-in through+`repowise mcp --tools` or the `mcp:` config block. See+[docs/agent/MCP_TOOLS.md](../../docs/agent/MCP_TOOLS.md).----+It is not published to PyPI on its own. Users get it inside the single+`repowise` distribution built from the root `pyproject.toml`.-## Installation+## Install the product```bash-pip install repowise-server--# Recommended: use uv-uv pip install repowise-server-```--Installs `fastapi`, `uvicorn[standard]`, `mcp`, `apscheduler`, `cryptography`, and `repowise-core` automatically.-------## Running the Server--```bash-# Via the CLI (recommended)-repowise serve # localhost:7337-repowise serve --host 0.0.0.0 --port 8080--# Directly with uvicorn-uvicorn repowise.server.app:create_app --factory --port 7337--# With hot reload (development)-uvicorn repowise.server.app:create_app --factory --reload --port 7337-```--Interactive API docs are available at `http://localhost:7337/docs` once the server is running.-------## Configuration--All configuration is via environment variables:--| Variable | Default | Description |-|----------|---------|-------------|-| `REPOWISE_DB_URL` | `sqlite+aiosqlite:///.repowise/wiki.db` | Database URL (SQLite or PostgreSQL) |-| `REPOWISE_API_KEY` | _(none)_ | Bearer token required on all API requests (leave unset to disable auth) |-| `REPOWISE_EMBEDDER` | `mock` | Embedder backend: `mock` (FTS only) or `gemini` (real semantic search) |-| `REPOWISE_WEBHOOK_SECRET` | _(none)_ | HMAC-SHA256 secret for verifying GitHub/GitLab webhook signatures |-| `ANTHROPIC_API_KEY` | _(none)_ | Anthropic API key (required for Anthropic provider jobs) |-| `OPENAI_API_KEY` | _(none)_ | OpenAI API key (required for OpenAI provider jobs) |-------## REST API--All endpoints are prefixed with `/api/`. When `REPOWISE_API_KEY` is set, every request must include an `Authorization: Bearer <key>` header.--### Repositories--| Method | Path | Description |-|--------|------|-------------|-| `GET` | `/api/repos` | List all registered repositories |-| `POST` | `/api/repos` | Register a new repository |-| `GET` | `/api/repos/{id}` | Get repository details and sync state |-| `PATCH` | `/api/repos/{id}` | Update repository settings (name, branch, provider) |-| `POST` | `/api/repos/{id}/sync` | Re-index and regenerate only affected pages (graph, git, dead code, decisions + incremental page regen) |-| `POST` | `/api/repos/{id}/full-resync` | Full re-generation of all pages from scratch (equivalent to `repowise init`) |--### Pages--| Method | Path | Description |-|--------|------|-------------|-| `GET` | `/api/pages` | List wiki pages, with optional `repo_id` and `page_type` filters |-| `GET` | `/api/pages/lookup` | Look up a page by `repo_id` + `page_id` query params |-| `POST` | `/api/pages/lookup/regenerate` | Regenerate a single page on demand |-| `GET` | `/api/pages/{id}/versions` | Full version history for a page |--### Jobs--| Method | Path | Description |-|--------|------|-------------|-| `GET` | `/api/jobs` | List generation jobs (supports `repo_id` filter) |-| `GET` | `/api/jobs/{id}` | Get job status and progress |-| `GET` | `/api/jobs/{id}/stream` | SSE stream of real-time job progress events |--Job progress events (`JobProgressEvent`) carry: `event` type, `file` currently being processed, `level` in the generation hierarchy, `completed` and `total` counts, elapsed time, and a final summary on completion.--### Symbols--| Method | Path | Description |-|--------|------|-------------|-| `GET` | `/api/symbols` | List/search symbols with `repo_id`, `kind`, `language`, `query` filters |-| `GET` | `/api/symbols/{id}` | Get symbol details including signature and source location |--### Graph--| Method | Path | Description |-|--------|------|-------------|-| `GET` | `/api/graph/{repo_id}` | Export dependency graph as nodes + edges (supports language and test filters) |-| `GET` | `/api/graph/{repo_id}/path` | Shortest dependency path between two modules |-| `GET` | `/api/graph/{repo_id}/modules` | Collapsed directory-level module graph with doc coverage |-| `GET` | `/api/graph/{repo_id}/ego` | N-hop neighborhood of a given node with git metadata |-| `GET` | `/api/graph/{repo_id}/entry-points` | Subgraph reachable within 3 hops from entry-point nodes |-| `GET` | `/api/graph/{repo_id}/dead-nodes` | High-confidence unreachable files with 1-hop neighbors |-| `GET` | `/api/graph/{repo_id}/hot-files` | Most-committed files with 1-hop outgoing neighbors |-| `GET` | `/api/graph/{repo_id}/nodes/search` | Full-text search over node IDs |-| `GET` | `/api/graph/{repo_id}/communities` | List all communities with labels, cohesion scores, member counts |-| `GET` | `/api/graph/{repo_id}/communities/{id}` | Community detail: members, neighboring communities, cross-edge counts |-| `GET` | `/api/graph/{repo_id}/metrics` | Graph metrics with percentile ranks for any file or symbol |-| `GET` | `/api/graph/{repo_id}/callers-callees` | Callers and callees for a symbol, with heritage support |-| `GET` | `/api/graph/{repo_id}/execution-flows` | Top entry points with BFS call-path traces |--### Git Analytics--| Method | Path | Description |-|--------|------|-------------|-| `GET` | `/api/git/{repo_id}/metadata` | Per-file git metadata (ownership, commit count, co-change partners) |-| `GET` | `/api/git/{repo_id}/hotspots` | Files ranked by churn rate (most frequently changed) |-| `GET` | `/api/git/{repo_id}/ownership` | Code ownership breakdown by file or module granularity |-| `GET` | `/api/git/{repo_id}/co-changes` | Files that frequently change together (temporal coupling) |-| `GET` | `/api/git/{repo_id}/summary` | Recent commit activity summary |--### Dead Code--| Method | Path | Description |-|--------|------|-------------|-| `GET` | `/api/dead-code/{repo_id}` | List dead code findings (supports `kind` and `safe_only` filters) |-| `POST` | `/api/dead-code/{repo_id}/analyze` | Trigger dead code analysis |-| `GET` | `/api/dead-code/{repo_id}/summary` | Summary stats (total findings, deletable lines, by kind) |-| `PATCH` | `/api/dead-code/{id}` | Update a finding's status: `resolved`, `acknowledged`, or `dismissed` |--### Decisions--| Method | Path | Description |-|--------|------|-------------|-| `GET` | `/api/repos/{id}/decisions` | List decision records (supports `status`, `source`, `tag`, `module` filters) |-| `GET` | `/api/repos/{id}/decisions/health` | Decision health summary — active, proposed, stale counts, ungoverned hotspots |-| `GET` | `/api/repos/{id}/decisions/{decisionId}` | Get a single decision record with full context |-| `POST` | `/api/repos/{id}/decisions` | Create a decision record manually |-| `PATCH` | `/api/repos/{id}/decisions/{decisionId}` | Update decision status (active, deprecated, superseded) |--### Search--| Method | Path | Description |-|--------|------|-------------|-| `POST` | `/api/search` | Search wiki pages. Body: `{"query": "...", "repo_id": "...", "mode": "fts|semantic|hybrid"}` |--### Blast Radius--| Method | Path | Description |-|--------|------|-------------|-| `POST` | `/api/repos/{repo_id}/blast-radius` | Structural PR impact for changed files — returns an uncalibrated 0-10 structural heuristic, raw direct weights, transitive reach, historical co-change, reviewers, and typed test state |--### Costs--| Method | Path | Description |-|--------|------|-------------|-| `GET` | `/api/repos/{repo_id}/costs/summary` | All-time totals: cost (USD), call count, input/output token counts |-| `GET` | `/api/repos/{repo_id}/costs` | Per-group breakdown of LLM usage with `by=day|model|operation` |--### Knowledge Map--| Method | Path | Description |-|--------|------|-------------|-| `GET` | `/api/repos/{repo_id}/knowledge-map` | Bus-factor / silo / freshness intelligence — files owned by single contributors, knowledge concentration risk |--### Security--| Method | Path | Description |-|--------|------|-------------|-| `GET` | `/api/repos/{repo_id}/security` | List security findings (filterable by kind/severity) |--### Chat--| Method | Path | Description |-|--------|------|-------------|-| `POST` | `/api/repos/{repo_id}/chat/messages` | Send a chat message; SSE streams the assistant reply with tool calls and source citations |-| `GET` | `/api/repos/{repo_id}/chat/conversations` | List saved conversations for a repo |-| `GET` | `/api/repos/{repo_id}/chat/conversations/{conversation_id}` | Fetch a full conversation with its messages |-| `DELETE` | `/api/repos/{repo_id}/chat/conversations/{conversation_id}` | Delete a conversation |--### CLAUDE.md Generation--| Method | Path | Description |-|--------|------|-------------|-| `GET` | `/api/repos/{repo_id}/claude-md` | Fetch the latest generated CLAUDE.md content for the repo |-| `POST` | `/api/repos/{repo_id}/claude-md/generate` | Trigger a fresh CLAUDE.md generation job (returns 202 with job id) |--### Providers--| Method | Path | Description |-|--------|------|-------------|-| `GET` | `/api/providers` | List available LLM providers and which one is active |-| `PATCH` | `/api/providers/active` | Switch the active provider |-| `POST` | `/api/providers/{provider_id}/key` | Store an encrypted API key for the provider |-| `DELETE` | `/api/providers/{provider_id}/key` | Remove the stored API key |--### Workspace (Multi-Repo)--| Method | Path | Description |-|--------|------|-------------|-| `GET` | `/api/workspace` | Workspace summary — registered repos with per-repo stats, cross-repo and contract summaries |-| `GET` | `/api/workspace/co-changes` | Cross-repo co-change pairs (files in different repos that frequently change together) |-| `GET` | `/api/workspace/contracts` | Detected API contracts (HTTP/gRPC/topic) and their provider/consumer file links |--### Webhooks--| Method | Path | Description |-|--------|------|-------------|-| `POST` | `/api/webhooks/github` | GitHub push event handler |-| `POST` | `/api/webhooks/gitlab` | GitLab push event handler |--### Health--| Method | Path | Description |-|--------|------|-------------|-| `GET` | `/health` | Server liveness/readiness, version, and database connectivity |-------## MCP Server--repowise registers 18 MCP tools and advertises **10 by default** in single-repo-mode (10 advertised by default, exactly the canonical set). Workspace mode adds-`list_repos`; seven specialist tools are opt-in where eligible. See-[`docs/agent/MCP_TOOLS.md`](../../docs/agent/MCP_TOOLS.md).-Start the MCP server via:--```bash-repowise mcp # stdio transport (Claude Code, Cursor, Cline)-repowise mcp --transport streamable-http # HTTP transport on port 7338-repowise mcp --transport sse # legacy SSE transport on port 7338+uv tool install repowise # or: pipx install repowise / pip install repowise+cd your-repo+repowise init+repowise serve # API on :7337, web UI on :3000```-| Tool | What It Answers | When to Call |-|------|----------------|-------------|-| `get_overview` | Architecture summary, module map, entry points, community summary | First call when exploring an unfamiliar codebase |-| `get_answer` | One-call RAG with confidence gating and caching | First call on any code question |-| `get_context(targets)` | Docs, ownership, history, decisions, freshness for files/modules/symbols | When you need to understand specific code before reading or modifying it |-| `get_symbol(symbol_id)` | Source body, signature, docstring for a qualified symbol | When the question names a specific function or class |-| `search_codebase(query, mode?)` | Hybrid symbol / path / concept search (auto-routes by query shape) | When locating a symbol, a file, or code by topic |-| `get_risk(targets?, changed_files?)` | Hotspot score, dependents, co-change partners, risk summary | Before modifying files — assess what could break |-| `get_change_risk(revspec?)` | What the change newly made worse, with attribution, plus a supporting diff-shape percentile | Before merging a commit or PR range |-| `get_why(query?)` | Architectural decisions, rationale, constraints | Before making architectural changes — understand existing intent |-| `get_dead_code` | Unused/unreachable code sorted by cleanup impact | Before cleanup tasks |-| `get_health` | 1–10 code-health scores and marker findings | Before refactoring — find the worst files |+## Develop on it-Also on by default: `list_repos`. Opt-in / workspace-only tools are documented in `MCP_TOOLS.md`.+From the repository root, `uv sync --all-packages` installs every package in+editable mode. See [CONTRIBUTING.md](../../.github/CONTRIBUTING.md).-**Claude Code / Cursor / Cline setup** — add to your MCP config:+## Documentation-```json-{- "mcpServers": {- "repowise": {- "command": "repowise",- "args": ["mcp", "/absolute/path/to/your/repo"]- }- }-}-```-------## Webhooks--Register the webhook URL with GitHub or GitLab so `repowise update` runs automatically on every push.--**GitHub setup:**--1. Go to `Settings → Webhooks → Add webhook` in your GitHub repository-2. Set **Payload URL** to `https://your-server.example.com/api/webhooks/github`-3. Set **Content type** to `application/json`-4. Set **Secret** to the value of `REPOWISE_WEBHOOK_SECRET`-5. Select **Just the push event**--**GitLab setup:**--1. Go to `Settings → Webhooks` in your GitLab project-2. Set **URL** to `https://your-server.example.com/api/webhooks/gitlab`-3. Set **Secret token** to the value of `REPOWISE_WEBHOOK_SECRET`-4. Enable **Push events**--The server verifies HMAC-SHA256 signatures, deduplicates events (stored in the `webhook_events` table), and launches a background sync job that re-indexes the repo and regenerates only affected wiki pages.-------## Background Jobs--The APScheduler instance manages the following recurring tasks:--| Job | Schedule | Description |-|-----|----------|-------------|-| Staleness check | Every 15 minutes | Finds stale/expired wiki pages and logs them for monitoring |-| Polling fallback | Every 15 minutes | Compares git HEAD against `state.json` for all repos; enqueues and launches sync jobs for any that have diverged (catches missed webhooks) |-------## Development--```bash-# Install for development (from repo root)-uv pip install -e packages/server -e packages/core--# Start with hot reload-uvicorn repowise.server.app:create_app --factory --reload --port 7337--# Run tests-pytest tests/unit/server/ tests/integration/-```+- MCP tools: [docs/agent/MCP_TOOLS.md](../../docs/agent/MCP_TOOLS.md)+- CLI commands, including `serve` and `mcp`: [docs/reference/CLI_REFERENCE.md](../../docs/reference/CLI_REFERENCE.md)+- How the system is built: [docs/architecture/](../../docs/architecture/)diff --git a/packages/web/README.md b/packages/web/README.mdindex f8b7eb315..30c032f80 100644--- a/packages/web/README.md+++ b/packages/web/README.md@@ -1,210 +1,39 @@-# repowise Web UI+# repowise web UI-Next.js 15 frontend for the repowise codebase documentation engine. Provides an interactive interface for exploring AI-generated wiki pages, dependency graphs, code analytics, and architectural insights for any indexed repository.+The local dashboard: a Next.js app that renders the wiki, graph, health, dead+code, decisions and the other views for an indexed repository. `repowise serve`+downloads a prebuilt copy on first run, caches it, and runs it on port 3000.-## Tech Stack+## Where it sits-| Layer | Technology |-|-------|------------|-| Framework | Next.js 15 (App Router), React 19, TypeScript 5.7 |-| Styling | Tailwind CSS v4 (CSS-first, no config file), custom design tokens in `globals.css` |-| Components | Radix UI primitives, class-variance-authority for variants |-| Data Fetching | SWR (stale-while-revalidate) with 30s polling, Server-Sent Events for live job progress |-| Graph Viz | @xyflow/react (React Flow) with ELK layout (elkjs) |-| Charts | Recharts (bar/donut charts), D3.js (force-directed minimap, treemap) |-| Diagrams | Mermaid v11 (embedded in wiki pages) |-| Code Rendering | Shiki (syntax highlighting), next-mdx-remote (wiki content) |-| Search | cmdk command palette (Cmd+K), Fuse.js client-side fuzzy search |-| URL State | nuqs (type-safe URL search params for bookmarkable views) |-| Animation | Framer Motion |-| Icons | Lucide React |-| Fonts | Geist Sans + Geist Mono |+This is the `repowise-web` npm workspace in the repowise monorepo. Pages live in+`src/app/`, app-specific components in `src/components/`, and the API client in+`src/lib/api/`. Shared components come from `packages/ui`, shared types from+`packages/types`, and the typed client from `packages/api-client`. It reads+everything from the API in `packages/server`.-## Pages & Features--### Dashboard (`/`)--The home page. Shows aggregate stats across all registered repositories (total files, symbols, entry points, dead code findings), a list of repos with quick-action buttons, and recent indexing jobs with their status.--### Repository Overview (`/repos/[id]`)--Landing page for a single repo. Displays health score, git insights (commit activity, churn distribution, bus factor), documentation coverage, and quick links to all sub-pages. Includes a **Graph Intelligence** section with expandable architecture communities (cohesion scores, member lists, neighboring communities) and execution flows panel (entry point scoring, BFS call traces, cross-community classification). Operations panel for triggering sync and full-resync jobs.--### Repository Overview — Standalone (`/repos/[id]/overview`)--Dedicated overview surface separate from the repo landing page. Aggregates 15 parallel data fetches and degrades gracefully on any failure. Sections include the key metrics strip, attention panel (auto-derived issues), language donut, ownership treemap, hotspots mini-leaderboard, decisions timeline, module minimap, dependency heatmap, community summary grid, execution flows panel, and Git Insights row (churn histogram, commit-category donut + sparkline, bus-factor panel).--### Wiki Pages (`/repos/[id]/wiki/[...slug]`)--The core documentation viewer. Renders AI-generated wiki pages as MDX with:-- Shiki syntax-highlighted code blocks-- Embedded Mermaid diagrams (flowcharts, sequence diagrams, etc.)-- Auto-generated sticky table of contents with anchor-linked headings (scroll-to-heading)-- Confidence badge showing freshness (fresh/stale/outdated) with last-updated tooltip-- Git history panel showing the file's change timeline-- Regenerate button to force a fresh AI pass on stale pages-- Graph intelligence sidebar (XL screens) with PageRank/betweenness percentiles, community label, degree counts, entry point badge-- Hallucination warnings — surfaces LLM validation findings (symbol references not found in source) as amber banners-- Version history with diff view — browse previous page versions and compare changes inline--### Dependency Graph (`/repos/[id]/graph`)--Interactive React Flow graph with ELK layout engine. Supports five view modes:-- **Module view** — hierarchical module organization with drill-down-- **Architecture view** — entry point reachability analysis-- **Dead code view** — highlights unreachable files with confidence groups-- **Hot files view** — commit activity heatmap overlay-- **Full graph** — complete dependency graph--Three color modes: **Language** (per-language coloring), **Community** (Leiden community clusters with real labels from analysis), **Risk** (churn-based).--Features pan/zoom, path finder for dependency chains, community detail panel (click any node in community mode to see members, cohesion, and neighboring communities with cross-edge counts), and URL param support for deep-linking color modes.--### Search (`/repos/[id]/search`)--Full-text and semantic search across all wiki pages. Includes a type toggle (FTS / semantic / hybrid), debounced input, and result cards with highlighted snippets. Also powers the global command palette (Cmd+K) accessible from anywhere.--### Symbol Index (`/repos/[id]/symbols`)--Sortable, filterable table of all extracted symbols (functions, classes, constants, etc.). Filter by kind and language. Click any row to open a two-panel detail drawer: left side shows the symbol's signature, docstring, and location; right side shows **graph intelligence** — PageRank/betweenness percentile badges, in/out degree, community membership, callers and callees with confidence scoring, and class heritage (extends/implements relationships).--### Documentation Coverage (`/repos/[id]/coverage`)--Donut chart of overall documentation freshness plus a per-file table showing each page's confidence score, status (fresh/stale/outdated), and a per-row regenerate button for targeted updates.--### Code Ownership (`/repos/[id]/ownership`)--Contributor attribution view with granularity toggle (file/directory). Shows ownership percentages, knowledge silo badges (files owned by a single contributor), and a contributor activity chart.--### Hotspots (`/repos/[id]/hotspots`)--Ranked table of high-churn files with churn bar charts and an owner leaderboard. Helps identify files that change frequently and may need architectural attention.--### Dead Code (`/repos/[id]/dead-code`)--Tabbed view of dead code findings (unreachable files, unused exports). Supports row-level actions (ignore, mark as false positive), bulk selection, and a trigger to run fresh analysis. Summary bar shows counts by category.--### Documentation Explorer (`/repos/[id]/docs`)--Split-pane documentation browser (VS Code-style). Left panel is a searchable file tree with type/freshness filters and colored freshness dots. Right panel renders full wiki content with:-- Mermaid diagram rendering and syntax-highlighted code blocks with copy buttons-- Graph intelligence sidebar (PageRank/betweenness percentiles, community, callers/callees)-- Version history browser with LCS-based inline diff view-- Hallucination warning banners for pages with detected inaccuracies-- Deep-linkable page selection via URL search params (`?page=...`)-- Export All (single `.md`) and Download ZIP buttons--### Architectural Decisions (`/repos/[id]/decisions`)--Lists extracted architectural decision records with health metrics. Each decision has a detail page (`/repos/[id]/decisions/[decisionId]`) showing context, rationale, alternatives, and consequences rendered as markdown, with Confirm/Dismiss/Deprecate actions.--### Blast Radius (`/repos/[id]/blast-radius`)--Pre-PR structural-impact estimator. Paste changed file paths (one per line) — or click hotspot suggestions — and get an uncalibrated structural-impact score (0–10 normalized points), raw direct structural weights, transitive affected files with depth, historical co-change warnings, recommended reviewers, and explicit test-analysis state. The structural score is not a runtime-breakage probability or the authoritative live change-risk signal.--### Cost Tracking (`/repos/[id]/costs`)--LLM token usage and spend dashboard. Shows total cost, total calls, input/output token counts (all-time), a daily-spend bar chart, and a breakdown table grouped by day, model, or operation.--### Workspace Dashboard (`/workspace`)--Multi-repo workspace surface. Aggregates totals (files, symbols, average doc coverage, hotspot count, page count) across all registered repos, renders a RepoCard grid with per-repo stats, surfaces cross-repo intelligence (workspace-wide co-change summary, API contract counts by type with provider/consumer split), and shows a top cross-repo co-changes table with deep links into the full views.--### Cross-Repo Co-Changes (`/workspace/co-changes`)--Full browser for files across repositories that change together based on git history. Filterable by source/target repo with summary stats (total pairs, distinct repo pairs).--### API Contracts (`/workspace/contracts`)--All detected API contracts (HTTP, gRPC, topic) and the provider/consumer file links between repos. Filter by contract type, repo, or role. Surfaces unmatched contracts (providers without consumers, or vice versa).--### Settings (`/settings` and `/repos/[id]/settings`)--Global settings page with sections for API connection, LLM provider/model selection, webhook configuration, and MCP integration. Per-repo settings available under each repository.--## Architecture--### Data Flow+## Install the product+```bash+uv tool install repowise # or: pipx install repowise / pip install repowise+cd your-repo+repowise init+repowise serve # API on :7337, web UI on :3000```-Server Components (RSC) Client Components-───────────────────────── ──────────────────────-Fetch data via apiGet() ──> SWR hooks for polling/revalidation-Render static content useSSE() for live job progress-Pass props to client nuqs for URL state persistence- useState for transient UI-```--Most pages are **React Server Components** that fetch data server-side. Client components are used only where interactivity is needed (D3 canvas, command palette, SSE streams, form state). The `"use client"` boundary is kept as narrow as possible.--### API Proxy--`next.config.ts` rewrites `/api/*` to `REPOWISE_API_URL/api/*`, so the frontend never hardcodes the backend URL and CORS is handled at the proxy layer.--### API Client (`src/lib/api/`)--Organized by domain — `repos.ts`, `pages.ts`, `graph.ts`, `search.ts`, `symbols.ts`, `jobs.ts`, `git.ts`, `dead-code.ts`, `decisions.ts`, `health.ts`. Each module exports typed async functions wrapping `apiGet`/`apiPost`/`apiPatch` from `client.ts`. Auth is handled via Bearer token from `localStorage` (browser) or `REPOWISE_API_KEY` env var (server).--### Custom Hooks (`src/lib/hooks/`)-| Hook | Purpose |-|------|---------|-| `useRepo` | SWR wrapper for repository data with 30s refresh |-| `usePage`, `usePageVersions` | Page content and version history |-| `useSearch` | Debounced search (300ms, min 2 chars) |-| `useGraph` (+ variants) | Graph data with stable caching (no revalidate on focus) |-| `useCommunities`, `useCommunityDetail` | Community summaries and detail data |-| `useGraphMetrics` | Per-node PageRank, betweenness, percentiles |-| `useCallersCallees` | Symbol call graph with confidence and heritage |-| `useExecutionFlows` | Entry point traces with community crossing |-| `useSSE` | Generic SSE hook with reconnection, exponential backoff, named events |-| `useJob` | Combines SWR polling + SSE streaming for real-time job monitoring |-| `useDebounce` | Generic value debounce |+## Develop on it-### Component Organization (`src/components/`)--```-components/- ui/ Radix-based primitives (button, card, dialog, tabs, tooltip, etc.)- layout/ Sidebar, mobile nav- wiki/ Wiki renderer (RSC + client), code blocks, Mermaid, ToC, confidence badge, version history with diff- graph/ React Flow canvas, toolbar, legend, tooltip, community panel, path finder- search/ Command palette, search bar, result cards- jobs/ Generation progress (SSE), job logs- repos/ Add repo dialog, operations panel, run config form- coverage/ Coverage donut chart, freshness table- dead-code/ Findings table, row actions, summary bar- decisions/ Decisions table, detail view, health widget- git/ Churn bars, contributor charts, hotspot/ownership tables- symbols/ Symbol table, symbol drawer, symbol graph panel (metrics + callers)- dashboard/ Health ring, attention panel, quick actions, language donut,- ownership treemap, community summary grid, execution flows panel,- module minimap, dependency heatmap, hotspots mini, decisions timeline- chat/ Streaming chat interface, message bubbles, model selector,- tool-call blocks, source citations, conversation history, artifact panel- workspace/ Repo cards, cross-repo summary, co-change table,- contract links table, contract type / role badges- settings/ Connection, MCP integration, webhook, provider sections- shared/ Stat cards, empty states+```bash+npm install # from the repository root (npm workspaces)+repowise serve --no-ui # start only the API on :7337+npm run dev # web UI with hot reload on :3000```-### Design System+The server-side API address defaults to `http://localhost:7337`; set+`REPOWISE_API_URL` to point elsewhere. See+[CONTRIBUTING.md](../../.github/CONTRIBUTING.md).-Dark-mode only. All visual tokens are CSS custom properties in `src/styles/globals.css` — surfaces, borders, text colors, accent blue (`#5B9CF6`), confidence colors (green/yellow/red), language colors for graph nodes, edge type colors, typography scale, spacing grid, radii, and z-index layers. Changing the design means editing one file.+## Documentation-## Development--```powershell-# From repo root — requires the repowise API server on port 7337-$env:REPOWISE_API_URL = "http://localhost:7337"-npm run dev --workspace packages/web-# Open http://localhost:3000-```--```powershell-# Type check-npm run type-check --workspace packages/web--# Lint-npm run lint --workspace packages/web-```+- CLI commands, including `serve`: [docs/reference/CLI_REFERENCE.md](../../docs/reference/CLI_REFERENCE.md)+- How the system is built: [docs/architecture/](../../docs/architecture/)