Codebase intelligence for AI and humans: code health scores, auto-generated docs, git analytics, dead code detection, and architectural decisions via MCP.
Changelog
All notable changes to repowise will be documented in this file.
The format is based on Keep a Changelog, and this project adheres to Semantic Versioning.
[0.59.0] - 2026-10-25
repowise 0.59 makes the first hour with a repository a guided one. repowise init ends on what stood out in your code and the exact model command for this machine, repowise serve opens the web UI in your browser, a three-step welcome greets a newly indexed repository, and a Setup page tracks what is set up and what is left, each row with one action and a prompt for your agent. Agents record, accept and merge architectural decisions as themselves, and a post-commit prompt asks them to record the decision a commit made.
Salesforce reaches the Full tier: Apex is parsed and resolved per sfdx project, and Lightning Web Components, Aura, Visualforce, flows and object and field metadata link into the same graph, with health, dead code, Extract Method and coverage. COBOL reads like a codebase: copybooks, CICS LINK and XCTL, JCL, BMS and DB2 includes resolve, paragraphs are scored for code health, and CSD-defined programs are entry points.
A folder of services gets a System Map that opens on groups of services and lets you focus one service or list every route between two, plus an estate report. Cross-repo links read many more HTTP clients and servers, gateways, brokers and gRPC shapes, and every link says how it was matched. Test reach, impacted tests and coverage say what they measured, with per-function coverage and an untested_complexity finding ranked by CRAP score.
Highlights
- A Setup page and a welcome for new repositories ·
repowise serveopens your browser, a three-step welcome shows what repowise found, and Setup lists what is set up and what is left, each with one action and a prompt for your agent. → /repos/:id/settings/setup - Salesforce and COBOL, mapped end to end · Apex, Lightning, Visualforce, flows and object metadata, and COBOL with copybooks, CICS, JCL and DB2 includes, link into one graph with code health, dead code and coverage. → /repos/:id/code-health
- A System Map for your services · Services grouped into systems, a focus view of who calls whom, every route between two services, and links read across many more frameworks, gateways and brokers. → /workspace/system-map
- Agents record and accept decisions · Your agent records the decision a commit made, accepts candidates signed as itself, and every accepted decision carries its own text. → /repos/:id/decisions
- Tests and coverage that say what they measured · Per-function coverage and an
untested_complexityfinding ranked by CRAP score, mocked modules kept apart from real reach, and impacted tests that count broadcast runs on their own line. → /repos/:id/tests
By the numbers
In the 4 days since 0.58.0:
- 251 pull requests merged, from 37 contributors, 26 of them first-time contributors
- 1,720 files changed, +117,289 / -16,849 lines
- 110 fixes, 116 features, 13 performance changes, 4 documentation changes
- 552 test files touched, 207 of them new
- 42 issues closed
Thank you to everyone who built this release with us.
Added
repowise initends on what stood out (#4677). The finish panel shows up to four findings (health score with its band and lowest file, unreachable files and safe-to-delete findings, the top hotspot, decision candidates), leaving out anything not measured. Documentation built from the index gets an "Add a model when you want more" block with the exactrepowise generatecommand for the model source found on this machine.--progress jsonadds afindingsobject and puts the model command innext.repowise serveopens the web UI in your browser (#4679) once it answers.--no-openorREPOWISE_NO_OPEN=1turns it off, and it opens only from a terminal, never in CI.- A three-step welcome for a newly indexed repository (#4682). It opens on the repository overview once per repository per machine (or any time with
?tour=1): what repowise measured, the agents on this machine with Connect and "Copy setup prompt", then the model and sync choices, finishing in Setup. - A Setup page and a setup pill (#4683, #4681, #4692, #4699).
/repos/:id/settings/setupgroups what is set up into Needs attention, Essentials (index, coding agent, sync) and Optional (model, coverage, shorter command output). Each row has a status, one action, its terminal command and a "Copy prompt" button, links to the settings page that holds its control, and a "Setup N of 3" pill sits in the sidebar. The page loads in about two seconds fromrepowise doctor --quickand refreshes in the background.GET /api/repos/{id}/setupserves the page, and its write routes accept only a loopback, same-origin request to a keyless server. - Settings on their own pages (#4694, #4683). Global settings open as Setup, Agents, Models, Sync, Server, Display and Privacy, and a repository's as Setup, General, Agents, Models, Sync and Coverage, each with a left rail and one row pattern. Repository General saves each field as you change it, rarely used options sit under Advanced, older
#group-links open the page they meant, and turning sync off asks first. - A model in two clicks (#4678). For a repository indexed without a model, the docs-mode badge ("Write with a model") and the chat "Add a model" state open one sheet that lists a signed-in Claude Code, Codex or OpenCode, an existing key, a local Ollama model or a pasted key, with the page count and price. "Write the docs" writes the pages as
repowise generatedoes. - Agents record and accept decisions as themselves (#4672, #4688, #4658).
repowise decision add --agent SLUG [--session ID]records a decision signed by the agent, accepted once it carries an evidence reference.decision confirm <id> --agent <slug>accepts a candidate on by default (decision config agent-acceptance --offkeeps acceptance for people),decision mergetakes--agentand--session, and the signer shows indecision show, the web UI and CLAUDE.md. Agent-recorded decisions rank beside session-mined ones and filter withdecision list --source agent. - A prompt to record the decision a commit made (#4689). On by default for Claude Code, a PostToolUse entry runs only when the shell command contains
git commit, adding about 35 ms to other shell calls on Windows and a few on Linux.repowise decision config capture-prompt --offturns it off per repository. - Decision review an agent can drive (#4702, #4704).
repowise decision reaffirm(CLI and REST) re-grants an accepted decision, and staleness counts from the latest person or agent grant, so a decision just accepted or reaffirmed is notneeds_review.confirmrefuses a decision that is already accepted,confirm,reaffirm,merge,deprecateanddismissall returnresults[]under--format json, retired records leave the candidate queue, and the confirm command in the generated How to work line passes--evidenceso it runs as copied. - Apex as a parsed language (#4404, #4417, #4421, #4429). Classes, interfaces, enums, triggers, methods, properties, annotations and ApexDoc are symbols. Calls, heritage and type references resolve in each sfdx project's namespace, then through declared package dependencies or a namespace prefix, with same-arity overloads chosen on exact argument types. Cross-file class recall is 99.9% to 100% across four public Salesforce projects, and 99.9% on a repository of 192 sfdx projects.
- Lightning, Aura, Visualforce, flows and metadata in the graph (#4411, #4405, #4415, #4414, #4418, #4441). LWC imports, template tags and
@salesforce/apeximports, Aura markup, controllers and helpers, Visualforce controllers, extensions,{!expr}members and pages, and flows with their Apex actions, subflows and screen components all link. Object and field metadata are symbols bound from SOQL, SOSL, DML and triggers, at 100% recall on two public projects, and record-triggered flows link their object. - Apex health, dead code and Extract Method (#4419, #4416, #4425, #4442). Methods, triggers and accessors are scored for complexity, classes for size and cohesion, and governor-limit work inside loops reports as
io_in_loop. Dead code roots triggers,@AuraEnabled,@InvocableMethod,@RestResource, platform interfaces and every class named by Lightning, Visualforce, flow or metadata files, and its findings are never deletion-ready. Extract Method plans render typed Apex helpers. Four advisory markers list without moving scores:apex_soql_injection,apex_missing_sharing,apex_hardcoded_idandapex_see_all_data. - Apex coverage and contracts (#4413, #4446, #4407). Coverage reads
sf apex run test --code-coverageoutput, legacysfdxoutput and deploy or validation reports, with a per-test map that letsimpacted-testsselect Apex test methods.@RestResourceclasses are HTTP providers at/services/apexrest/..., and literalHttpRequestcallouts are workspace consumers keyed by their Named Credential. - COBOL copybooks, CICS, JCL, BMS and DB2 (#4434, #4435, #4436, #4437, #4438, #4460, #4472). Fixed format, embedded
EXEC SQL/CICS/DLIblocks andCOPYstatements parse without losing paragraphs, including programs whosePROCEDURE DIVISIONheader comes from a copybook.COPYandEXEC SQL INCLUDEresolve to copybooks and DB2 DCLGEN members,EXEC CICS LINKandXCTLfollow literal constants to their programs,PERFORMstays inside its program, and JCL steps link to programs andSEND MAPto BMS mapsets. On a public CICS sample, program-to-copybook pairs resolve 246 of 246. - COBOL code health and entry points (#4439, #4461, #4462, #4459). Paragraphs and sections are scored for complexity, nesting and size, CICS CSD definitions are reachability roots, and an unused paragraph is reported at low confidence, never as deletion-ready. The advisory
cobol_unstructured_flowmarker lists paragraphs that jump withGO TOorALTER, and level-88 condition names sit under the data item they test. repowise grammars fetch(#4471, #4433). Stages the COBOL and Apex grammars for offline machines, andREPOWISE_GRAMMAR_DIRpoints at a pre-staged cache. Grammars load per language, a grammar that cannot load prints a warning with the file count, andrepowise doctoradds a row per downloaded grammar.- A System Map for a folder of services (#4621, #4600, #4624, #4623, #4591, #4622).
/workspace/system-mapopens on groups of services (agroup:tag, a shared name family, then linkage) with bands that carry link counts, and draws one level at a time with a details rail and a breadcrumb. Focus a service to see its callers and dependencies side by side, pick two services to list every route between them, search with/or Ctrl/Cmd+K, or read the map as a list. Gateway-routed links are labelled, and a callout says when evidence is partial.GET /api/workspace/system-graph/linkspages every link between two services (#4598). - An estate report (#4554, #4561, #4592).
repowise report estatereads the stored workspace and reports per-repository health and churn, the riskiest seams between services and where to start, and the same report opens at/workspace/estateandGET /api/workspace/estate. - Cross-repo links across many more stacks. HTTP calls from .NET
HttpClient, Refit and typed clients (#4538, #4550), SpringWebClient,RestClient, Feign and@HttpExchange(#4540, #4531), Apache and MicroProfile clients, Noderequest,superagentandneedle, AngularJS$httpand Locust (#4545). Routes from chi, fiber, gorilla/mux and Go 1.22 method patterns (#4539), Fastify, hapi, Koa and Next.jspages/api(#4558), Azure Functions (#4547) and Spring context paths (#4549). Gateway routes from Spring Cloud Gateway, Zuul (#4536), nginxproxy_pass(#4543), Ocelot and YARP (#4548). Kafka clients by declared type (#4534), JMS (#4557), Go AMQP (#4542), C# integration events (#4544) and Azure Service Bus queues and topics (#4637). gRPC from C#AddGrpcClient, JavaImplBaseand grpc-js servers, with vendored proto copies serving nothing (#4541, #4529, #4537). A Python repository links to the sibling that publishes its dependency (#4585), and JS template literals with nested templates read as URLs (#4655). - Services found by host name (#4495, #4656). A workspace builds a service-name map from compose, Kubernetes, Spring and Aspire config,
.envfiles and${VAR:-default}defaults, so a call tohttp://catalog-api/...or${CATALOGUE_URL}resolves with no manualservice_bases. - Every link says how it was made (#4486, #4587, #4588, #4580, #4551). Each contract names its origin and each link its matching strategy, shown in the web views and in
get_blast_radius,get_change_riskandget_risk. Ribbons andrepowise workspace diagnosticslead with the share of HTTP calls linked to another service, name dangling calls and separate a service's calls to itself, and diagnostics count messaging SDK calls with unresolved topic names (#4667). - Per-function coverage and
untested_complexity(#4326, #4330, #4335). Functions get coverage and a CRAP score, anduntested_complexityreports untested complex functions, high from CRAP 50 and critical from 100 on measured coverage. Measured findings appear everywhere, estimated ones withinclude=["unverified"], andget_health(include=["untested_complexity"])ranks functions by CRAP. Coverage findings carry an add-tests prompt naming the nearest test file and thehealth-rules.jsonrule that silences them. - One test-reach walk with typed states (#4329, #4338, #4341, #4337). Every file reads
reached,unreached,new,no_signal,type_onlyormocked_only. A Python test that swaps a module out, or a JS or TS test with a closedvi.mockorjest.mockfactory, does not count as reaching it, and files of only types are never called untested. - The CI action summarises every gate (#4313, #4642).
comment: trueposts one editable PR comment covering every gate, the job summary gets a gate table, and averdictoutput readspass,warnorfail. A failingrepowise health check --format githubends each annotation with a one-line fix and prints the exact local command that reproduces it. - An opt-in health gate on agent commits (#4371).
repowise hook health-gate installgates Claude Codegit commitandgit pushonrepowise health check, naming up to five failing findings, and fails open on a crash or a 20 s overrun. - Health scores name their analyzer (#4370).
analyzer_versionsits beside scores inrepowise healthJSON,health check,--trendrows andget_health. repowise doctorchecks carry an id, a level and a fix (#4675), in table and--format json, and caveats show aswarnwithout failing the run.doctor --quickskips the MCP launch and store reconciliation and answers in about two seconds (#4699).- Offline and telemetry controls (#4440, #4687, #4706).
REPOWISE_OFFLINE,DO_NOT_TRACKandREPOWISE_TELEMETRY_DISABLEDskip the release check, the highlights fetch and every agent CLI sign-in probe, so repowise makes no network call of its own. Anonymous telemetry carries fixed reason codes on failures and a host label (claude_code,codex,cursor,vscode,ci,terminal), never message text or paths, and sends nothing under a test run. The health headline names its coverage when fewer than half the code files were analysed. - Callers through an interface (#4506).
get_context(include=["callers"])andrepowise context --include callerslist callers that reach an implementation through its interface or base class. get_whyandget_contexton a directory (#4356, #4372).get_whybuilds an origin story from the directory's git history, andget_contextreads the files under a directory with their owners, last change and hotspot.get_overviewsays why an input is unavailable (#4649) innext_actions.unavailable_reasons.- Language reach. Ansible playbook and role dependency edges (#4399), Scala dataflow and Extract Method (#3580), Ruby gem public API with Ruby visibility forms (#4526), Dart closure parameters typed from the collection call (#4528), Gradle dependencies in the package inventory (#4488), a Spring Boot main as a declared entry point ranked above filename guesses (#4487), and .NET types loaded by MassTransit, NServiceBus, Azure Functions and WCF rooted for dead code (#4562).
Changed
- Settings live in
.repowise/config.yaml(#4676). A repository's excluded paths and wiki style are written there by the CLI, the server and the web settings form, which all read the same value. - Ink primary actions (#4680, #4692). Primary buttons are solid ink in light and dark, focus rings are neutral, and orange stays for links, selection and the brand. Every "Copy prompt" button is one quiet control that copies at once and previews the prompt on hover.
- A calmer docs bar (#4698). The bar holds the view switch, Search, a "Write N pages with AI" button while structural pages remain, and one More menu for export, present and doc freshness. A page's own write or rewrite action sits under its title, and an out-of-date page shows a warning there.
- What's new and the welcome share one look (#4693), open as a full-height sheet on a phone, and list connected agents first.
- Impacted tests count what they ran (#4324, #4325, #4632, #4321, #4333). Tests kept only by a conftest broadcast and tests reached through long import chains still run and are counted apart (
broadcast_total,import_closure_total), so a hub change's headline reads 308 tests where the run set holds 1,865. A package__init__.pyreaches every module under it, parametrizedgetfixturevalueis followed, cache-reset autouse fixtures select nothing, and reaching tests rank by distance before the per-target cap. - Coverage is stored when it maps (#4323, #4320, #4327, #4331, #4336, #4339).
coverage addstores nothing from a mostly unmapped report unless--allow-partial, and suggests the path prefix that maps it. A.coveragefile is explained as a per-test map with the export command for per-file numbers.updatere-reads a report written after the last ingest (coverage.reingest_on_update), checking in a few milliseconds, the per-test map dedupes and cuts whole tests, and the file page skips its line tint when the file changed after measurement. - One answer per figure (#4373, #4367, #4360, #4365, #4364, #4369, #4368, #4581). One file count and one symbol count across
status, the dashboard, workspace andget_overview, the index's last sync time on every surface, one severity per Do next row, a band beside every headline score (Healthy from 8, Warning from 4, Alert below), and bug-fix counts that name their 6-month window beside 90-day churn. - Change risk ranks against the commits before the change (#4503), however the change is spelled, and the checklist names that sample. Code moved between files gets no health before-and-after in the headline (#4314).
repowise health --scorematches the stored score (#4499) on an unedited file, scoring against the index's clone partners.- Honest workspace evidence (#4500, #4654, #4586, #4657). A workspace with too little cross-service evidence reads "not determined" or "partial" with reasons, a route served by two or more repositories links to none unless a host picks one (
ambiguous_provider), a generic path such as/healthneeds a host to link, and cycle counts past 20,000 read "at least N". - A workspace command that reads one repository names it (#4559), and
update --workspaceprints each repository's stored counts (#4555). - Cross-repo dead code on every surface (#4375, #4383, #4358). The CLI, MCP, REST, the file page and Do next share one cross-repo read, and a finding another repository may use is never deletion-ready.
repowise decision liststates its total (#4355) with--limitand--all, and decision staleness is measured from the commit that landed the decision (#4357).- Clearer MCP hints (#4366). A wrong selector id names the right one, and an argument that needs another names it in
ignored_arguments. - File classification fits the code (#4322, #4353, #4652, #4665, #3462, #4468). A
tests/folder inside a JS or TSsrctree is product code,*.test-utils.*is test support,tools/is tooling at the root or undereng/,hack/,ci/ordev/, Python_vendor/and Salesforce static resources are vendored, anddb/*_migrations/are migrations. - Faster workspaces and walks (#4627, #4616, #4651, #4498, #4569, #4485, #4505, #4560, #4556, #4577, #4359). The System Map lays out 300 services in 70 ms, with layout in a worker.
/api/workspaceat 207 repositories answers in 137 ms (from 1,786 ms) and/api/workspace/graphin 139 ms (from 4,786 ms). A no-changeupdate --workspaceon 12 repositories takes 2.6 to 4.5 s (from about 11 s), one changed repository rewrites only its own store and editor files, and contract extraction runs in 6 to 8 s on 8,141 files (from about 20 s). The pruned repository walk is about four times faster on Windows, persistence imports in 17 ms, and exact symbol betweenness skips symbols with no call or heritage edges, scores unchanged (#4602). - Version stamps.
HEALTH_ANALYZER_VERSIONmoved from 41 to 43, so health re-scores on the next update.PARSER_SCHEMA_VERSIONmoved from 9 to 10, so the parse cache refreshes once, automatically.SECURITY_SCANNER_VERSIONmoved from 4 to 5, so the next update rescans.STORE_FORMAT_VERSIONis unchanged and nothing needs re-indexing.
Fixed
- MCP and search.
search_codebasescans literal queries that contain a slash, such asapp/widget.py, when no file page matches (#4230).repowise mcpstarted in a directory with no index runs and every tool reports "no repowise index yet" with therepowise init --yes --no-proseremedy (#4670).get_execution_flowsresolves an entry point by bare name or partial id likeget_symbol(#4674). The server instructions nameget_architectureandget_blast_radiusonly when they are served (#4707). Concurrent Streamable HTTP sessions share one lifespan (#4398), andget_dead_codereturns its pooled connection (#4096). - Decisions. Mining keeps the verbatim decision sentence, and every acceptance requires decision text (#4686, #4671), with
decision confirm --decision TEXTto supply it. A mining batch cut off atmax_tokenscounts as failed (#4424), whole titles print indecision health(#4389), and backslashes in decision titles stay literal in CLAUDE.md and AGENTS.md (#4530). - Init, update and doctor.
updateseeds a linked worktree whose base synced past its fork point (#4673),init --no-editor-setupwrites only inside.repowise/(#4496), the real default branch and a credential-free remote url are recorded and kept (#4381, #4374, #4388), large repositories update without "too many SQL variables" (#4403), the per-file git-index timeout is 120 s and set withREPOWISE_GIT_INDEX_TIMEOUT_S(#1837), an invalidREPOWISE_FULL_RESCORE_INTERVAL_DAYSwarns once (#1844), a Claude Desktop registration is not a Claude Code duplicate indoctor(#4318),doctorcounts a provider key saved in.repowise/.env(#4708), a file parsed while its grammar was missing is not cached and re-parses once the grammar is available (#4703),--provider mockruns without a warning (#4287), notices from an olderinitdrop out of the degraded list on the next update (#4285), andrepowise serverebuilds a stale local web UI even when the checkout sits inside a folder nameddistornode_modules(#4085). - Graph. TypeScript path aliases resolve under solution-style
tsconfig.jsonreferences (#4519), a JS or TS bare call binds to a same-file function or an import (#4518), Dart dotted receivers and prefixed named constructors bind (#4114), Express mounts key correctly (#4571), PL/I.incmembers are not C++ (#4553), Apex parameters parse type-first (#4429), Apex callouts trace through wrapper methods (#4532), COBOL CSD files without definitions index cleanly (#4492), COBOL compiler-option cards that start in column 1 are skipped (#4491), and cycles list from the smallest node id (#4513). - Workspace. FastAPI routes with a comment in the decorator stack (#4583), SQLite readers linking within their own repository (#4584, #4653), NATS read only beside a NATS client (#4650), service names and gateway routes read from production deploy config only (#4648), Postgres catalog tables skipped (#4187), and components cover flat repositories and count contracts outside every component (#4552, #4494).
- Health and coverage. Java switch arms with comments (#4527),
brain_methodreasons with judged CCN (#4184),churn_percentilematching SQLPERCENT_RANK(#4189), on-disk coverage stamped at the commit it measured (#3935), per-test records resolving like the aggregate (#4328), nested checkouts naming changed files correctly (#4340), the Tests tab and file page sharing one reach definition (#4334), six performance findings with readable labels (#4354), schema statements kept out of N+1 (#4363), Fix first plans complete or withheld (#4378), a fragile file's lead simplified only when worth it (#4361), short-history files marked on the health file list (#4351), and a fragile file no test reaches in the code graph reads as "confirm it is tested", while one reached only through mocks still asks for tests (#4394). - Security and dead code. Empty
public_env_secrettemplate values (#4037) and Salesforce merge fields and regex ranges (#4468) are not secrets, and Python lazy-loader attribute tuples count as uses (#3958). - Web UI and reports. The overview's Architectural Clusters section renders (#4463), module and component file counts cover indexed files (#4376), the conformance page shows "at least N" cycles (#4666), a merge body's issue-keyed first line becomes the change title (#4469), and the welcome's dead-code card and the Commits score open their Code Health tabs (#4695).
Documentation
- The CLI, config, MCP, CI and health references have one heading per command, tool or block with flags in tables, matched to the code (#4342).
- Salesforce at the Full tier, with an Apex row in the benchmarks: 40,907 cross-file references across seven public projects at 0.999 recall (#4447).
- An architecture page for the multi-repo workspace layer (#4509), which names the file suffixes with a comment mask (#4615), and the dead-code confidence ladder matches the analyzer (#4362).
Dependencies
certifi2026.10.19, which distrusts the compromised emSign CA (#4352).- Security updates (#4696): GitPython 3.2.1 or newer (submodule path advisory), Pillow 13.0.0 and filelock 4.2.4 (transitive), and for the web UI
next15.5.28,vite7.3.7,dompurify3.4.17,undici7.30.1,markdown-it14.3.3,fast-uri3.1.9 andserialize-javascript7.1.4.
[0.58.0] - 2026-10-20
repowise 0.58 makes the first run read like a plan. repowise init opens on one plan screen with the scan, an ETA and every step it will run, shows progress as a checklist of finished steps with counts and times, and ends on one finish panel that leads with repowise serve. A logged-in Claude Code, Codex or OpenCode is offered with its quota, agents get --progress json, a folder of repositories gets the same screen with --repos, and init, status and update share one set of mode names, with repowise update --full taking a fast index to a full one.
The web UI is redesigned around what to do next: a short Overview, Commits led by each change's health delta and recent wins, an Ownership page built around knowledge coverage and off-boarding, a zoomable folder map with lenses, coupling by area pairs, and Tests and Security pages in a grouped sidebar.
repowise health check gates a change in CI on the code-health findings it introduces, using git alone, and flags added suppressions, weakened tests and test-only branches. Health scoring follows a saturating curve per category, names the production-code score beside the headline, and scores files from disk with repowise health --score and --against. Change risk follows code that moves between files, get_overview lays out every package and its front doors in a lean default, GitHub Copilot CLI joins the providers, and every release carries an openapi.json.
Highlights
- A guided
repowise init· One plan screen before the run, a readable checklist of every step while it runs, and a finish panel that opens withrepowise serve; agents follow along with--progress json. - A redesigned Code Health page · A zoomable folder map with lenses, a Performance tab that opens on the fixes worth making now, what each file's score lost to code shape, history and coverage, and quieter findings with a prompt ready for your agent. → /repos/:id/code-health
- An Overview that opens on what to do · One health reading, a folded Do next, questions drawn from your repository and Explore cards, in a grouped sidebar where Tests and Security have their own pages. → /repos/:id/overview
repowise health checkin CI · Fail a change on the code-health findings it writes, with git alone, and warn on added suppressions, weakened tests and test-only branches.- Commits led by code health · Every commit and change opens on its health delta, the page leads with the period's wins, and a commit sheet hands your agent a prompt for what got worse. → /repos/:id/commits
- Ownership and off-boarding · See how much of your recently changed code more than one person knows, which files one person holds, and what happens to the code if someone leaves. → /repos/:id/owners
By the numbers
In the 5 days since 0.57.0:
- 278 pull requests merged, from 35 contributors, 19 of them first-time contributors
- 1,609 files changed, +157,134 / -64,952 lines
- 92 fixes, 80 features, 9 performance changes, 79 refactors and 7 documentation changes
- 462 test files touched, 181 of them new
- 58 issues closed
Thank you to everyone who built this release with us.
Added
- One plan screen for
repowise init(#4253, #4256). An interactive init shows the repository scan, an ETA in seconds ("Under a minute" on a small repository) and the plan (index, decisions, documentation, search, connect, sync) on one screen. Enter runs the free run,mpicks a model,eedits the plan,skeeps the global MCP entry where it is, andqquits and removes the.repowise/the run just made. A model is a visible opt-in that names the detected key or the logged-in CLI, and the screen offers to keep machine-specific files out of git. - Progress you can read (#4251). Each finished step leaves a row with its count and time, running rows show the current file or page, a time-left figure appears once a step has a known total and a measured rate, and a Next line lists the stages still to run. A step running past 10 s says Ctrl-C is safe and names
repowise init --resume. One progress display serves init, workspace init,update,update --fullandreindex. - Short lines when piped, and
--progress jsonfor agents (#4254). Init exits 0 on a clean run, 2 when it finishes with gaps and 130 when interrupted (#4242, #4269). - One finish panel that leads with
repowise serve(#4248, #4263). It names what the run built and which agent is connected (with the Claude Code restart it needs), gives two example questions,repowise servewith its port, at most two more steps, the files init added and whether auto-sync is on. A workspace run ends on the same panel with a Scope row and a per-repository documentation list. - Logged-in Claude Code, Codex and OpenCode, offered with their quota (#4247). A logged-in subscription CLI appears in the provider picker and the generation plan with its cost as model calls, parallel processes and minutes, and it runs only when you pick it. The picker defaults to what
--yeswould run. - A brand theme with ASCII fallbacks (#4257). Init honours
NO_COLOR, every glyph has an ASCII form for non-UTF-8 terminals, and the mini owl ends the run showing how it went. - The same plan screen for a folder of repositories, and
init --repos(#4262, #4260, #4264). A workspace init lists its repositories andrchooses which to include.--repospicks them by name or path in scripts, a re-run that picks a subset keeps the others in.repowise-workspace.yaml, linked git worktrees are skipped, andworkspace addbuilds documentation from code by default. - Init asks before pointing the global MCP entry at this repository (#4246). It names where the Claude Code and Claude Desktop entry points today and where it will point.
--yesand scripted runs register and print both targets. repowise update --fulltakes a fast index to a full one (#4244), with no API key needed for a Built from code index, and every surface names that command. After a keyless init,generateandupdate --fullwith a model also read the decision sources the fast run skipped, priced on their own line (#4243).repowise statussays what the next update re-scores (#4301). One line names a pending health re-score or security rescan, andstatus --format jsonreportsrescore_pending.- A lean CLAUDE.md and AGENTS.md block (#4235, #4294). It carries the date and commit, the files that need care, one code-health line, the top three decisions, three working bullets and the tool table, and the sections you want can be added back.
- A redesigned web UI (#4163, #4157, #4138, #4164, #4142). The Overview, Code Health, Commits and Ownership pages open on what to do next, findings carry a prompt for your agent, and the sidebar groups every page. The entries below cover each page.
- A short Overview that opens on what to do (#4163, #3966). One health reading, a folded Do next with its count in the heading, an ask box with questions drawn from the repository, and six Explore cards whose figures match the pages they open.
- Commits and Changes lead with code health (#4137, #4141, #4154, #4133). Every commit row opens on its health delta (better, worse, mixed or neutral) and marks hotspot touches. The page opens on a period headline that counts improvements first, then recent wins, then filters for health impact, hotspots, time window, author and area kept in the URL. The commit sheet reads the verdict, Got better and Got worse grouped by file, with a prompt to fix what got worse, and Change Records lead with their health rollup and say where a moved finding came from.
GET /winslists resolved findings, files whose score rose and refactorings that landed. - Ownership around knowledge coverage and off-boarding (#4136, #4153). The page opens on knowledge coverage, the share of recently changed code held by two or more people, beside single-owner hotspots, quiet owners and the truck factor, with tabs for knowledge risks, knowledge loss and an off-boarding what-if (
?leave=&threshold=). A person's page leads with what depends on them and who could take it over.GET /owners/summaryandGET /owners/knowledge-matrixserve the figures. - A zoomable folder map with lenses (#4157, #4135). Code Health's Map shows each directory as a disc inside its parent and opens areas into sub-areas and files as you zoom, with
?area=in the URL, lenses including main author and knowledge risk, and a context rail with the area's score, 30-day delta, trend, worst files, hotspots, owners and top findings.GET /areas/treeandGET /areas/{path}serve directory rollups at any depth, and JSON responses over 1 KB are gzipped (the map payload goes out at 51 KB from 634 KB). - Hotspots searched and paged over the whole set (#4162), with server search by path and owner, filters for bus factor 1, accelerating and bug magnets, sorting on any column and facet counts.
- Coupling opens on area pairs (#4152). The tab leads with the area pairs that change together with no import to explain it, ranked by unexplained file pairs, and drills into one pair's files. Each row copies a decouple prompt, and
GET /coupling/areasserves the pairs. - Tests and Security pages, in a grouped sidebar (#4142, #4186). The repository sidebar groups pages into Understand, Code quality, People & History and Settings, Tests and Security have their own pages, and the sidebar stays open on Docs.
- What each file's score lost (#4124, #4132). File drawers and pages split the deduction under the score into code shape, change history and test coverage.
repowise health check: a CI gate on the findings a change introduces (#4302).repowise health check origin/main...HEADneeds git only: it analyses the changed files at base and head and matches findings by file, marker and symbol. It exits 1 on a code-shape finding written on the change's own lines at high severity or above (--fail-at medium|high|critical), and--fail-on worsenedalso fails findings the change made worse. Moved and renamed code is context, rules in.repowise/health-rules.jsonare read from the base, and output comes as a table, JSON, Markdown, GitHub annotations, SARIF or GitLab Code Quality. Exit codes are 0 passed, 1 failed, 2 could not evaluate.- Diff-risk rules in
health check(#4312).added-suppression(a new lint, type-check, coverage or repowise suppression),weakened-tests(a new skip or focus marker, fewer test cases or assertions, a looser coverage threshold) andtest-only-branch(a test-environment check in production code) warn by default.--diff-risk warn|error|off, orRULE=LEVEL, sets each level. - Baseline-growth gates for security and doc drift (#4299).
--fail-on-baseline-growth(action inputfail-on-baseline-growth) fails the gate when the committed baseline lists a finding the base's copy does not. A first baseline passes. repowise health --scoreand--against(#4139, #4155).--score <paths>scores files from disk with the index's git history, coverage, test pairing and dependents, so a file reads as the next update would score it.--against <rev>scores the same files at a revision and prints before, after and the weighted average.get_health(include=["what_if"])(#4130). For the worst files or the targets, each code-shape finding's points, what removing it gains the file, the headline and the hotspot average, the file's ceiling, and the best single fixes across them. Metric rows carrystructure_deductionandhistory_deduction(#4108).- Production code's score beside the headline (#4204).
repowise healthnames Production code beside Code health while tests are in the reading, andget_healthkpis carryproduction_average. Files with fewer than 5 commits reporthistory: "short"(#4117). - Change risk follows moved code (#4109, #4112, #4125, #4127). Symbols a change moves across files are mapped, findings pair across the move, a pure split reads as moved, and improved findings are counted and stored with
moved_from. health_deltaper changed file (#4123, #4128).get_change_risklists up to 10 analysed files with base and head score, structure deduction and size, largest move first, and both sides read the index's dependents.get_overviewlays out every package (#4214, #4207). The lean default carries the week's top three actions, the headline health with the production score and band counts, the five riskiest files and a compact scope digest, plus every package with its front doors, layer roots and build, test and lint commands.include=["tools"],["actions"]and["scope"]add the rest.- MCP tool annotations (#2762), built from each tool's safety class.
get_answercites the files its answer names (#4202), andget_whyanchors a question on the indexed file it names, reported inanchored_on(#4203).- A symptom for suspects over MCP (#4205).
get_contexttakessymptom(issue text, an error or a stack trace) forinclude=["suspects"],repowise context --symptomdoes the same, and path targets feed suspects their stem, symbol and directory names. - GitHub Copilot CLI as a provider (#3710).
openapi.jsonon every GitHub release (#3948). Each operation carriesx-repowise-stability: stablefor integration routes orinternalfor dashboard payloads.- PHP dataflow and Extract Method plans (#3466).
- Prior defects counted by conventional commit type (#4126), including an unknown type that names a fix.
- Performance loops sized by the tables they read (#4236). A loop that grows only by reading small tables is bounded with reason
small_table. - Language reach. Dart for-in loop variables are typed from the container they walk (#3853), and C# 14
extensionblock members are extension methods (#3920). - Old docs links on the website land on the matching docs.repowise.dev page (#4238).
Changed
- One set of mode names (#4249, #4284, #4309). Init, status, update and the editor files say Built from code, Written by and Index only, "Documentation" names the pages, and each mode names the same next command everywhere (
repowise update --full, thenrepowise generate). Flags, state and JSON values are unchanged. - Health scoring saturates per category (#4144, #4129). Each defect category follows
A * (1 - exp(-raw / A))withAat 3.09 times the category cap, findings on one function count the largest whole and the rest at 0.25, history findings are de-stacked the same way before their cap, and the score breakdown reports each category's ceiling. bumpy_roadisrepeated_nesting, andbrain_methodis advisory (#4113, #4115). Old ids keep working inhealth-rules.json, findings and filters, andbrain_methodis listed without a deduction since the size markers already charge that function. Five markers read as Deep nesting, Long function, Compound condition, Many authors and Fading ownership (#4119), and size markers read judged CCN and code lines (#4116).- Health findings fit the code they describe.
primitive_obsessionskips test material (#4118), coverage markers skip re-export-only files (#4110), a docs folder counts as docs only at the root or beside a docs-site config (#4200), and a critical code-shape marker ranks as high in Fix first (#4199). - One performance headline, led by the causes worth fixing now (#4231, #4227, #4138, #4165). CLAUDE.md, AGENTS.md,
get_health,repowise health,repowise status, the Overview and Code Health share one figure, Fix first and Do next name performance work once its cost is confirmed, and the Performance tab opens on the changes worth making now, grouped by area with a bulk agent prompt. - Do next acts on a week's regressions (#4198, #3966). Its Now group holds defect-class regressions with proven cost and lost measured coverage, the week's code-shape findings become one row to plan, and only a live credential is critical.
- A leaner
get_change_riskdefault (#4221, #4215, #4210). It keeps every fact an agent acts on, leads the headline, reasons, actions and findings with the same finding ("1 of 6 new findings needs review"), and trims a requested evidence checklist inside its rows so a 75-file commit returns one. - Quieter findings cards (#4164). Hairline outlines, severity as a dot and a word, the path printed once, and "Copy prompt for your agent" as a ghost button, with one filled prompt for the filtered rows.
- One
FilePatheverywhere (#4161). The file name is never cut, the directory shortens first, and the full path is the tooltip. - Dead code is tiered after the cross-repository cut (#4196), so
safe_onlyand the summary counts agree. - Workspace views state
code_filesfrom one population with a real summary per repository (#4216), and five workspace error replies are clearer (#3943). repowise reindexturns semantic search on after a keyless init (#4240) with a detected embedder, and every run names its embedder.- The knowledge graph's layer-naming prompt lists each layer's largest folders with file counts (#4206).
- Code health reads as technical debt in the UI (#4180), and clearer labels in
risk, dead code,whats-newand the generation report (#4197). - Faster updates and MCP calls. A prose-only update reuses stored clone and dead-code results, saving 17-32 s and 12-27 s (#4217), the commit-health scan per update is capped at 10 s and drains newest first (#4213), SQLite statistics refresh after each index write (#4229), health rows are replaced in one statement (#4226), and the architecture map classifies each path once (#4222). A repeated committed
get_change_riskrange is served from a per-process cache (#4209), test selection runs off the event loop and is reused per change (#4208), change health reuses parses by content (#4311), andget_whylooks up named files in one query (#4218). Repositories open through aRepothat closes without a full garbage collection, about 1 s per pass on Windows (#4212). - Behaviour pinned by golden snapshots (#3970, #3971, #3972, #3974) across scanning, dead code, health, the parser, persistence order and MCP payloads, while init, update, workspace, serve, doctor, the pipeline, git indexing, health, dead code, security scanning and the MCP tools are organised into small staged modules (#3951, #3973, #3976, #3977, #3979, #3980, #3981, #3982, #3983, #3985, #3986, #3987, #3988, #3989, #3997, #3999, #4000, #4001, #4002, #4003, #4004, #4005, #4006, #4007, #4008, #4010, #4011, #4013, #4016, #4017, #4018, #4020, #4021, #4022, #4025, #4038, #4039, #4040, #4041, #4042, #4043, #4044, #4045, #4046, #4047, #4048, #4049, #4050, #4051, #4052, #4053, #4054, #4055, #4056, #4057, #4058, #4059, #4060, #4062, #4063, #4064, #4065, #4066, #4067, #4068, #4069, #4070, #4071, #4072, #4073, #4074, #4075, #4076, #4077, #4078, #4080, #4081, #4082, #4265, #4268, #4270).
HEALTH_ANALYZER_VERSIONmoved from 40 to 41 with the scoring changes above, so existing indexes re-score on their next update.PARSER_SCHEMA_VERSIONmoved from 8 to 9 (Kotlin calls chained on calls), so the parse cache refreshes once, automatically.STORE_FORMAT_VERSIONis unchanged and nothing needs re-indexing.
Fixed
- Init and update.
--no-prosemakes no model calls with a key set (#4239). A re-run skips the files init wrote and a fast re-init keeps the documentation (#4245).statusandstate.jsondescribe the index the last command left, with "Index in progress" while init runs (#4241). Workspace init andworkspace addstamp each repository so the firstupdatehas nothing to redo (#4293, #4310).init --reposwith no names is a usage error (#4300), customize keeps-xexcludes (#4061),update --fullworks on an index copied to a new path (#4308),updaterecounts the file-page scope every run (#4220) and retires template pages the selection drops (#4211), and a docs-mode update records its repair marker (#4026).generatecounts model-free pages correctly in its summary (#3776). - Workspace. Repository aliases are kept and cancellation propagates (#4107), any
state.jsonshape reads (#4036), errors reading saved exclude patterns are logged (#3054), andworkspace listcounts likestatus(#3984). - Doctor and serve.
doctorexits 1 in table mode when a check fails (#4289) and runs in linked git worktrees (#1952).servekeeps web UI archive members inside its cache directory (#3975). - Editor files. Key modules and architecture skip structure-only preambles (#3261).
- Graph. Go calls through interface-typed variables bind to the interface's method specs (#3850). A Kotlin call chained on a call is typed (#3845), Kotlin and Swift properties initialised with a constructor are typed (#3789), a Kotlin typealias leads to the class it names (#3787), and top-level Kotlin declarations with two annotations are symbols (#3839). A Python name alias reaches its real definition (#2799), Rust
cfgbuild variants get their own ids (#3020), C# object-initializer locals bind (#3957), and React Native platform files resolve to import edges (#3711). - Ingestion and classification. Stored signatures and docstrings use LF line endings (#3929),
test-helpersandtest-utilsdirectories are test support (#3257), and PHP production classes namedTest.phpoutside a testsuite are production code (#3399). - Dead code. A Swift extension whose member name is declared elsewhere (#3876), C and C++ files listed in CMake,
.vcxprojor.rc(#4121), Kotlin@Suppress("unused")(#3873), C# classes with custom MEF export attributes (#4009), C# cmdlets, source generators and benchmarks (#3926), Spring servlet initializers (#3921) and@GlideModule(#3886) are all recognised as used. A module's dead-code penalty is normalised by its symbol count (#4083). - Health.
repowise health --format jsonkeeps withheld finding types (#4111), the headline shows Hotspot only when a figure exists (#4295), findings stored under a renamed marker id get their advice (#4305, #4177),.repowise/health-rules.jsonapplies when rescoring commits (#4179),score_breakdownlists defect findings only (#4178), active contributors count by canonical identity (#4086), cut-off signatures and stub bodies are declarations (#3950), and C++ condition declarations count as reads (#4122). - Performance and refactoring. A cause reached by a request, consumer or job role is entry reachable (#4234), queue rows are numbered by their place (#4233), the characterization step names the changed function and the tests that reach it (#4232), Java helper types and throws are kept (#3882), Extract Method helpers return what the caller reads (#3938), and dataflow builds try regions inside switch arms (#3694).
- Coverage and tests. The recommended per-test coverage command exists (#3676), the inferred basis reads the cached test-reach walk (#4158), test selection runs only the tests that read a changed fixture file (#4267), and test-reachability reads batch their
INlists (#3933). - Security. Allow markers work in prose without a comment opener (#4143), a truncated digest or a cache key's md5 reads as an id (#4201), and public API in published C# projects is flagged (#3579).
- MCP and suspects. Bare symbol names defined only in excluded files are gated (#4084), and suspects says when candidates reach the change index cap (#3930).
- Web UI. Code Health, Overview, Commits and Owners render payloads from the previous release's server (#4304, #4306, #4303), the Overview Tests card's route and command can be set by a host (#4307), and the Why this score arithmetic names the 1.0 floor (#4132).
- Events, decisions and doc drift. Tied owners at the ownership margin are ignored (#3870), a hand-recorded proposed decision reads as hand-recorded (#4015), doc-drift paths join by normalised segments (#4023), and a failing
on_errorleaves a guarded step's result intact (#4079). - VS Code. Doc drift diagnostics follow live edits (#2782).
- Cost. Gemini 3 preview models are priced at their own rates (#3967).
Documentation
- The quickstart has a human path and an agent path, serves first, and lists every mode with its upgrade command (#4250), and every page that teaches init says what the CLI does (#4255).
- Change evidence, components, policy and reports in the README and docs (#3947), the README's benchmark tile matches the benchmark method (#3969), and the untested hotspot conditions and complexity README are corrected (#3449, #4169).
Dependencies
- Security updates: aiohttp 3.14.4 and mcp 1.30.0 (declared floor
mcp>=1.30.0,<2) (#4292).
[0.57.0] - 2026-10-15
repowise 0.57 reads every merged change as a Change Record, rebuilt from git with an evidence checklist (owners, reaching tests, contracts, decisions crossed, provenance with coverage) and its outcomes, and the Commits page groups commits by change. On top of those records, repowise check evaluates a local change policy from .repowise/policies.yaml and writes SARIF or JUnit, repowise report writes assessment, change evidence and inventory packs, and repowise suspects ranks recent changes against a symptom with the evidence for each. Components follow the Backstage model, compare declared owners with recent activity, and export package usage as CycloneDX. AI attribution is three-state with coverage, with opt-in edit capture for Claude Code, Codex and Cursor.
Chat draws any chart of your repository in one call and checks every figure in an answer against the turn's tool results, and chat runs headless over a session API. Models and base URLs are set from Settings with live model lists, and the dashboard shows what's new after every upgrade. The server authenticates REST, MCP HTTP and webhooks to a principal with repository and tool scope, installed plugins extend named seams, and MCP routes workspace targets to the repository that owns them. Code Health serves every language, labelling the ones still in preview, and its map, file drawer and trend read clearly at any repository size. Astro is indexed at the Full tier, Gitea and Forgejo push webhooks sync, and MassTransit message contracts link C# and VB.NET services.
Highlights
- Commits grouped by change · Every merged change is a Change Record with its evidence checklist and outcomes, and the Commits page groups commits by the change that landed them. → /repos/:id/commits
- Change policy and evidence packs ·
repowise checkruns the rules in.repowise/policies.yamlagainst any change and writes SARIF or JUnit, andrepowise reportwrites an assessment, a change evidence pack for a period or a component inventory. - Charts in chat · Ask for any chart of your files, findings, changes, health trend, coverage, dead code, dependencies or layers, and every figure in the answer is checked against the data. → /repos/:id/chat
- Models from Settings, with live lists · Pick a model from the provider's live list or type any id, and point repowise at LM Studio or another local OpenAI-compatible server with no key. → /settings
- Suspects for a symptom · Paste an error, a stack trace or an issue and
repowise suspectsranks the recent changes to read first, with the evidence for each.
By the numbers
In the 3 days since 0.56.0:
- 264 pull requests merged, from 41 contributors, 25 of them first-time contributors
- 1,421 files changed, +123,474 / -15,118 lines
- 192 fixes, 51 features, 3 performance changes, 5 refactors and 7 documentation changes
- 504 test files touched, 168 of them new
- 53 issues closed
Thank you to everyone who built this release with us.
Added
- Change Records, rebuilt from git on every index (#3745). Each first-parent commit of HEAD is one change: a PR, MR, Gerrit change or lone commit, identified by its forge number, then
Change-Id, then its sha, with ids stable across updates. A merge carries the commits its branch brought in, and a maintainer merge gives up the PR merges and squashes inside it as their own changes. Each record carries its category and three-state provenance with coverage. Served throughrepowise changes list|show,GET /api/repos/{id}/changesand a change page in the web UI, andrepowise changes show --format in-totoexports an in-toto Statement v1. Forge metadata plugs in through aChangeMetadataSource; the default readsApproved-bytrailers and closing keywords from git. - An evidence checklist on every Change Record (#3791, #3814). Components touched, declared owners, suggested reviewers, reaching tests (measured or inferred), contracts and consumers in other repositories, decisions crossed, health delta, AI provenance with coverage and the diff size rank, each with a status (
ok,attention,unknown), the basis it rests on and its source. Each row reads as a plain sentence.repowise changes showprints it,repowise changes export --since --until --format json|csvstreams a window, andget_change_risk(include=["evidence"])and REST return the same record. - Outcomes on Change Records, from explicit links only (#3812). A record says whether a later commit reverted it, whether a later fix names it as the cause in its own words (quoted, with
Fixes: <sha>and "regression in" style links), and how many of its files received fixes within 30 days (changes.outcome_fix_days). The change page has an Outcomes section, and outcomes flow into the checklist, export and assessment report. - The Commits page groups commits by change, and shows events (#3803, #3872). A History view switch offers By commit, By change and Events. By change lists each merged change with its category, commits, files, lines, provenance and an Evidence column, with Category and Needs attention filters, and each row opens the change page with its full checklist. Events lists component events newest first.
repowise check: local change policy (#3779)..repowise/policies.yamlholds Kyverno-shaped rules that match components by name, tag or criticality, with conditions on attribution and agent tier, reaching tests, contract breaks, decisions crossed and bus factor, and actionswarn,require_owner_review,require_testsorfail. The security, coverage and doc-drift gates can be wrapped as rules.repowise check [REVSPEC]writes a table, JSON, SARIF 2.1.0 or JUnit XML (--sarif-fileand--junit-filefrom one run), exits 1 when afailrule fires, and reports a condition whose evidence is unavailable as not evaluated. The JSON carries the change's evidence for teams that feed OPA.repowise report assessment: an evidence pack per repository or workspace (#3805). Components and their posture, hotspots, bus-factor exposure, dead code by tier, test reach, contracts, governing decisions, provenance coverage and a same-stamp trend, each item in OpenSSF Scorecard v2 check shape plus amethodandbasis. JSON is canonical, with Markdown, self-contained HTML and CSV renderers, local paths scrubbed, and missing evidence reported as unresolved rather than negative. Sections and renderers are plugin seams.repowise report changes: a change evidence pack for a period (#3815, #3924). For every change merged in a window: what evidence exists that it was reviewed, tested and policy-checked, which teams are accountable, and whether it was later reverted or named by a fix, with a period summary and a method appendix that states what git cannot attest. Reviewers and individuals are counted, never named. With no--sinceit covers the 90 days before the newest indexed commit, and--since allreports every change.repowise suspectsranks recent changes against a symptom (#3813). Give it issue text, an error message, a stack trace or a list of files, and it ranks the Change Records merged in a window by text, graph distance, co-change, recency and stack-trace signals, combined by reciprocal rank fusion, showing which signals fired for each. The output is a reading order, and the docs give agit bisect runrecipe to confirm. Also served asPOST /api/repos/{id}/suspects,get_context(include=["suspects"])and a chat tool.- Components, Backstage-shaped (#3722, #3809). Declare
Component,SystemandGroupentities in.repowise/components.yamlorcatalog-info.yaml, withrepowise.dev/*annotations for paths, criticality, tags and capability; with no file, components are derived from workspace services, module pages or top-level directories, with readable names past route segments.repowise components check|list,get_context(targets=["component:<name>"])with health, hotspots, bus factor, test reach and dead code rolled up, andget_overview(include=["components"]). - Declared owners compared with recent activity (#3781). Owners from the catalog and CODEOWNERS (GitHub, and GitLab with sections) are matched to the last 90 days of commits, raising four component findings: a high or critical component with no active owner, an owner with no recent changes on its files, a bus factor of 1 on a critical component, and a catalog entry that claims no files.
.repowise/teams.yamldeclares team membership. Served byrepowise components ownershipandget_context(..., include=["ownership"]), naming teams only. - Package usage, component inventory and CycloneDX export (#3806).
repowise packages where <name>lists import sites with their component, owner teams and reachability.repowise packages import <file>joins a Syft, SPDX 2.x, CycloneDX or OSV file you supply to those sites, advisories first.repowise packages export --format cyclonedxwrites CycloneDX 1.6 withevidence.occurrencesfrom the import graph, andrepowise report inventorylists components with their owners, languages, entry points and packages. Alsoget_context(targets=["package:<name>"])and a REST usage route. - Component snapshots and typed events (#3799, #3861). Every
initandupdatestores a snapshot per component, and what moved since the previous run under the same analyzer stamp is stored as a typed event: ownership shift, owner inactive, bus factor drop, new hotspot, contract added, contract broken, decision stale, dead-code growth and test-reach drop. Events are written to.repowise/events.jsonlas CloudEvents 1.0.2, served byrepowise eventsandGET /api/repos/{id}/events, and delivered through a pluggableevent-sinkseam. Each kind is switchable underevents.kinds. - Three-state AI attribution with coverage (#3732). Commits are labelled agent, human-attested or unknown, and a commit without a marker is unknown. Every AI share is shown as the agent share of commits with a known source, with that coverage, and is withheld below 30% coverage. git-ai v3 notes are read in full, including human attestations. A new
attribution:block (enabled,capture,agents) in repo and user config, where the stricter value wins, andattribution.enabled: falsestores and shows nothing about AI.repowise attribution statusandrepowise attribution purge. - Opt-in edit capture for Claude Code, Codex and Cursor (#3744, #3765). With
attribution.capture: local|notes, the agent's edit hook records the lines it wrote as content hashes in.repowise/attribution/(never prompts, code or identities), andrepowise updatebinds them to commits by content. Atnotes, a git-ai v3 compatible note is written underrefs/notes/ai.--yesandinitnever turn capture on, andDO_NOT_TRACKkeeps it off. The integrations matrix gains an Attribution column. - Attribution backfill from local agent transcripts (#3792, #3804).
repowise attribution backfill --agent claude-code,codexbinds past commits to the edits recorded in this machine's Claude Code and Codex transcripts, keeping only line hashes and reading only what was appended since the last run.repowise attribution statusshows coverage before and after, and the nextupdatenames the backfill as the reason it re-reads history. - Change Records and attribution through MCP and chat (#3784, #3790).
get_contextacceptschange:<id>targets (or a commit sha) andinclude=["changes"]andinclude=["provenance"]on files,get_risk(include=["changes"])lists each file's recent changes, andget_overview(include=["changes"])gives changes by category and week.repowise context --includeoffers the same blocks. Default responses are unchanged. - Scoped authentication for REST, MCP HTTP and webhooks (#3725). Every authenticated surface resolves the caller to a principal with repository and tool scope. Optional scoped bearer tokens in
~/.repowise/tokens.yaml(stored as SHA-256 hashes, withexpires_at) can be limited to repositories and MCP tools, and the MCP streamable HTTP and SSE transports honour the same credentials, with an RFC 9728 protected-resource document. Chat runs tools as the request principal. With no key and no tokens file, every surface behaves as before. - Plugins from entry points, and named seams (#3717). An installed package extends repowise through a
repowise.pluginsentry point orREPOWISE_PLUGINS. Each extension point is a named seam with a local default that a plugin orREPOWISE_SEAM_<NAME>can switch, andrepowise pluginsshows what loaded, what failed and which implementation each seam uses. Plugin commands never replace a built-in, and a repository's.repowise/.envcannot select plugins. - Charts in chat (#3860, #3843, #3743, #3780, #3728, #3869, #3865). A
charttool draws any of eight repository datasets (files, findings, changes, health trend, coverage, dead code, dependencies, layers) in one call as bars, horizontal bars, lines, areas, donuts, treemaps, heatmaps, scatter or stacked charts.show_viewbuilds tables and charts on the server from a turn's tool results, so every plotted number comes from a result, and a small spec filters, groups, aggregates and sorts any result. Each chart keeps a text summary, a Data toggle and a basis line, chats print as documents, and empty chats and each page suggest a chart that fits. Plugins add datasets and views on thechat-datasetandchat-viewseams. - Chat checks every figure in an answer (#3793). Each number the answer states is looked up in that turn's tool results and views. A figure no result holds is marked with a dotted underline, a tooltip and screen-reader text, and counted in
diagnostics; a verified figure carries the id of the result it came from. - Evidence cards (#3718). A versioned, layout-free
EvidenceCardschema for answers, changes, components, scope, suspects and events, in which every fact says how it is known and where it came from, with a strict validator and a web renderer used by chat. - Headless chat sessions (#3746).
ChatSessionruns the chat loop without the web server, returning a structured turn result or stream events, overPOST /api/chat/sessionsand/sessions/{id}/messages(SSE or JSON). Each session carries a frozen policy (tool allowlist, read-only ceiling, loop cap, token budgets), sends accept an idempotency key, and long conversations are trimmed before each model call. The web chat runs on the same engine. - Model and base URL from the dashboard, with live model lists (#3849). Repo settings take a base URL with presets for local servers, Test connection, a searchable model field listing the provider's live models that also accepts any typed id, and the source of each value. An OpenAI-compatible server on this machine (LM Studio, Ollama, vLLM) needs no API key, and an unconfirmed
config.yamlbase URL offers "Trust this URL". The chat model menu loads live models too, and/settingsholds the global default. - What's new after every upgrade (#3874). After a minor upgrade the dashboard opens What's new once, with the release summary, a By the numbers ribbon, the highlights with links into the current repository, and every change by section; a patch shows a short note. Highlighted pages carry a New mark in the navigation, the upgrade banner names the next release's top highlights, and
repowise whats-newprints highlights and numbers withsummary,highlightsandstatsin its JSON. - Gitea and Forgejo push webhooks (#3797).
POST /api/webhooks/giteaverifies the HMAC-SHA256 signature againstREPOWISE_GITEA_WEBHOOK_SECRETand syncs on push for both forges. - MCP routes workspace targets to the repository that owns them (#3740, #3786). In workspace mode,
get_context,get_symbol,get_riskandget_healthsend a path orpath::Nameto the repository that has the file,<alias>:<target>names one outright, and an ambiguous path returnscandidates. A bare symbol name such asrenderorClient.sendroutes to the repository that defines it, andget_symbolresolves an omission ref from any workspace repository, backed by a derived.repowise-workspace/workspace.dbthat is always safe to delete. repo="all"fans out concurrently with partial results (#3742).search_codebase,get_dead_code,get_whyandget_overviewquery workspace repositories concurrently with a per-repository timeout (REPOWISE_WORKSPACE_CONCURRENCY,REPOWISE_WORKSPACE_REPO_TIMEOUT_S), naming any that failed inrepos_failedand any whose index is behind inrepos_stalewhile the rest answer.- Astro at the Full tier (#3164). Frontmatter and every
<script>body are parsed as TypeScript at their original offsets, frontmatter imports and exports become symbols and edges, PascalCase markup tags become component calls, and pages and endpoints undersrc/pagesare entry points. - MassTransit message contracts for C# and VB.NET (#3070). Publishers (
Publish,Send, request clients, scheduled sends, sagaInit<T>and wrappers found by shape) and consumers (IConsumer<T>, fault consumers, saga events, open-generic consumers) link across workspace repositories by message type. - Spring routes link to their handler methods (#3650). Each
@GetMapping,@PostMapping,@RequestMappingand related method gets aframework_bindsedge, across positional,value=andpath=arguments, arrays and Kotlin, and the HTTP contract dialect reads the same parser. On three Spring sample applications, route edges go from 0 to 33, all hand-checked correct. - Spring Data and MyBatis calls in Java loops are flagged as database I/O (#3615). A receiver counts only when its declared type is a Spring Data repository or MyBatis mapper; on one Spring Boot application the performance list goes from 5 to 16 rows, 10 of the 11 new rows judged real costs.
- Dart call receivers are typed from fields, parameters and locals (#3610). On a public Flutter sample repository, call edges go from 979 to 1,086 with no edge lost.
- Swift has a complexity map (#3450). Swift files get function rows (CCN, nesting, cognitive complexity, parameters) and class rows (methods, fields, LCOM4, TCC), and the markers derived from them.
- Pascal performance loops say whether they grow with the data (#3517), including
while not Q.Eofloops over a data set opened in place. - Refactoring plans carry test commands for Go, Rust, JVM, Ruby and C# (#3585, #3583, #3613).
go test,cargo test -p, Maven and Gradle module tests,bundle exec rspecorrake test, anddotnet teston the owning test project or solution, stored on the plan invalidation.test_runsalongside the JS/TS commands so batches rebuild them without a checkout.
Changed
- Code Health serves every language, with a Preview label where precision is still being measured (#3867, #2921, #3877). Every finding, refactoring plan, performance cause and dead-code row is served by default; on a language whose layer measured below the precision bar, it carries a preview note with its latest hand-judged measurement (a badge with the measurement in its tooltip on the web, a
previewfield in REST, one compact_meta.previewentry in MCP, one line in the CLI). Health findings ship normally on Java, C# and C, each measured at 92.5% or more correct or debatable on a blind sample. Queue counts report language-held rows in agatedfield. - History markers on production files are context, not defects (#3867, #3840). Git-history-only markers (ownership, churn and similar) leave the default defect lists and are counted under
history, while scores, KPIs,get_riskand the work queue still use them. Findings on vendored and generated files and history markers on test files are held at serving, andinclude=["unverified"]shows all of them. Long parameter lists leave out parameters bound by framework annotations such as@PathVariable. - The Code Health map, file drawer and trend read clearly at any size (#3879). The map is drawn on canvas with zoom, pan and a Fit control, files circle-packed by folder inside each module, dots in the five score bands, and "Draw more" up to 8,000 files with vendored and generated files behind a toggle. The non-modal drawer opens on the score, one plain sentence and the top three issues with vital signs, with one primary action. The trend chart has a hover and keyboard readout and a movers list.
- Charts draw from one palette (#3871). Every chart and share bar in the web UI draws from the same eight series colours, a commit type keeps one colour on every chart, health and risk charts keep their green, amber and red, heatmaps step through the accent orange, and chat charts sit on the conversation without a card around them.
- One provider order everywhere (#3847). The CLI,
repowise servechat and jobs, and the MCP server resolve provider, model and base URL as flag or UI choice, then.repowise/config.yaml, then the web UI's global default, thenREPOWISE_PROVIDER/REPOWISE_MODEL, soconfig.yamltakes precedence over those env vars in the CLI. Choosing a model or base URL for a repo in the web UI writes that repo'sconfig.yaml, which makes switching models on a local OpenAI-compatible server such as LM Studio a single edit. A base URL inconfig.yamltakes precedence overOPENAI_BASE_URLand the other*_BASE_URLvariables. The default provider and model saved in the web UI settings also apply to CLI runs in repos whoseconfig.yamlnames none. Aconfig.yamlbase URL receives an API key from outside that repo's.repowise/.envonly when it is a local address or was confirmed in Settings or withrepowise init. - Current model lists and prices (#3868). Pickers offer GPT-6 Luna and Sol, Claude Sonnet 5.5 and Opus 5.5, Gemini 3.8 Flash and Gemini 4 Argon. DeepSeek defaults to
deepseek-flashand LiteLLM togroq/openai/gpt-oss-120b, matching what those providers serve today. Cost estimates use current rates, and GPT-6 models get the same reasoning, temperature and chat tool handling as GPT-5.6. - Sync shows its progress and explains a failure (#3838). Sync and full re-index from the dashboard show progress, and a failed or partial run stays on screen with the reason (a rejected key, rate limit, unreachable base URL or missing model) and Retry, Provider settings and Dismiss actions.
- Git history reads the newest 5,000 commits by default (#3833), up from 500, so churn, hotspots, co-change, Change Records, suspects and 90-day counts cover a real span of history on active repositories. The co-change window follows it. An existing index keeps the
commit_limititsconfig.yamlrecords (initwrote 500 before 0.57), andupdatenames that value once; setcommit_limit: 5000there to read the wider window. - AI shares are reported with coverage (#3732).
/commits/agent-trendreturns per-label counts,coverage_pctandagent_pct_of_coveredin place ofagent_pct, the authorship filter gainsunknownandhumanmeans human-attested,get_contextownership carries oneagent_sharesummary in place ofagent_authored_pct, and AI figures are reported per file, component and repository. - Extract Method plans leave the helper's name to whoever applies them (#3588). The header, call and stage chaining use a placeholder (
_extracted_helper,extractedHelper, numbered for later stages) with one note to rename it. Judged blind on 107 real refactorings from 19 repositories, names chosen by the applying agent were accepted about 74% of the time. Plan ids are unchanged. - Fix first orders by the function's own change history (#3565). Inside a value band, items are weighted by how many commits changed the lead function, and the reason names that history ("function changed in 38 commits").
- Fix first leads with the stored plan its first step lifts (#3582, #3662), titled "Split X into N helpers" with "(more stages needed)" when the plan alone leaves the function above the break-up bar, and the dashboard card opens that plan. A Break up reason says what the split buys (#3581).
- Change risk leads with the size percentile (#3664).
repowise risk, the CI markdown andget_change_riskopen with "Size and spread: Nth percentile of recent commits (diff size, not a bug predictor)", then the touched files' bug-fix record. The plugin's risk command and change review skill follow the new wording. get_answergradesconfidenceon whether the top hit covers the question (#3716).retrieval_qualitymoves behindinclude=["diagnostics"].- Repository health headlines state their sample (#3641, #3653). A band such as "Excellent" is printed only when at least half of the code files were analysed, each headline says "analysed N of M code files" when some were left out, and every surface names the same worst file and worst test as
get_health. - The dead code tab leads with what is safe to delete (#3884), with the flagged count as a secondary sentence and
safe_to_delete_counton the summary endpoint. - Change risk and suspects answer faster on a running server (#3925, #3927, #3928). A
get_change_riskfinding drill-down returns in under a second (about 5 s before), test selection reuses its per-file parsing across calls (get_change_riskandget_riskPR mode about 25% faster), and suspects reads each change's git text once per server (a warm call drops from about 6 s to under 1 s). - Bug-fix commits are recognised in more forms (#3567).
Fixed,Fixes,Fixing,hotfix,bugfixandFixes #Ncount as bug fixes.
Fixed
- Change Records, events and components. The reaching-tests checklist item asks for a look only on a measured absence, and partly-known items carry counts (#3814). Component snapshots are taken on repositories with git owners (#3808), contract events are measured after the cross-repository pass extracts contracts (#3861), and derived components with colliding names are qualified by their parent (#3809).
- Chat. Charts and tables cite results from earlier turns, and setup or provider failures end as a readable chat error (#3834). A chart field that plays two roles draws one column (#3865), a new conversation keeps a bare URL (#3576), and taps pass through the hidden conversation row toolbar (#3764).
- Graph, Java. Same-arity overloads, enum-constant methods, anonymous-class methods and
@interfaceelements get their own nodes (#3616, #3624, #3670). Overloads are chosen by literal argument types, arity against inherited overloads and the declared types of named arguments, and a tie never lands on an overload that cannot take the call (#3672, #3704, #3625). Call chains, inherited fields,superand cast receivers are typed,S.super.m()resolves, a member type shadows a same-named import, and the same-package tier offers only the right class's method (#3623, #3671, #3669, #3620). - Graph, Kotlin. Overloads and same-named extensions get their own ids, and calls whose only argument is a trailing lambda are call sites, including a local assigned from one (#3766, #3783, #3785). Extension calls resolve on a typed receiver and only with the receiver in scope, operator
invokebinds, multiplatform calls bind only to declarations their source set sees, alias imports bind by alias, and imports of top-level functions and properties resolve (#3707, #3681, #3733, #3831, #3607, #3407, #3737). Constructor properties type receivers, parameters shadow a bare call, and a property is never a bare call's callee (#3609, #3679, #3680). - Graph, Swift, Dart and Python. Kotlin, Swift and Dart locals are typed from the call that assigns them (#3734). Swift stored properties type receivers, a Swift extension is not a constructible class, a subscript read is not a call, and optional-binding call assignments resolve (#3609, #3619, #3618, #3768). Dart named constructors bind, chained calls are typed by return type, and chains of three or more links resolve (#3701, #3794). Python receivers are typed through properties and returned objects, literal lazy exports are recognised, and test paths anchored at
Path(__file__)ancestors resolve (#3666, #3570, #3518). - Graph, general. A module-alias call resolves to the module's top-level name only (#3617), name-only call tiers stay inside the caller's build (#3682), and C++ project export macros are discovered in headers beside target sources (#3706).
- Ingestion. Every module of a Groovy
settings.gradleinclude is read and relative Gradle module paths resolve (#3626, #3587).!negation rules re-include pruned directories (#3563), Unity state directories are skipped only inside a Unity project (#3866), and C# files in test projects are test material (#3586). Shallow-clone boundary lines are left out of blame ownership (#3073), and suggested reviewers are weighted by each author's own recent commits (#3074). - Code origin. Dart SDK headers under
dart-sdk/are vendored, SwiftPM version-specific manifests are build files, Visual C++ resource headers and Dart codegen, minified and webpack output are generated (#3652, #3841, #3842, #3637). - Health metrics. Declared parameters are counted consistently across languages (#3848), a try block is not a nesting level (#3648), anonymous-class members roll up to their named owner (#3643), C# type arguments stay out of callee names (#3475), and the catch variable is a def in the Java, C# and C++ dataflow dialects (#3483). AVA's
t.*assertions count in JS/TS tests (#3674). - Health findings.
ownership_risksays which rule fired, files under 40 code lines are capped at low, and test-file findings rank after production ones (#3656). A flat boolean chain holds at low incomplex_conditional(#3665),developer_congestionrequires recent commits (#3675), sweep commits and inheritance stay out of hidden coupling (#3686), a keyed loop call clearskey_unusedfor its line (#3539), and a C#usingdeclaration stays with the code it guards (#3463). - Health surfaces. A repository with no hotspot files reports hotspot health as null in the trend, snapshots and an empty history, and the trend view says so (#3684, #3705, #3713).
get_healthfile rows account for every point their score lost,gap_analysisagrees with the gap, and plan steps carry no stray\r(#3683). Queue counts count a language-held unit once (#3667), every performance finding states its queue reason (#3612), and example-named packages with a production origin stay in Fix first (#3629). Fix first item detail stays about the item itself (#3693), and--refactoring-targets --recomputeplans carry their findings (#3521). - Refactoring and dataflow. Extract Method never plans a span over match arms and binds pattern captures (#3782). A write that may not run does not cover a later read (#3589), a write in a TypeScript, Go or Python switch arm is a may-def (#3522), a grouped Go
vardeclares its names and taking a Go variable's address is a write (#3472, #3474), and Rustusedeclarations are not variable reads (#3575). Performance rows that are the recommended idiom or stated intent are dropped (#3611). - Dead code. C# classes found by assembly scanning or MEF, Swift cross-file uses and Java classes named in resources are reachable (#3832). Kotlin operator functions are contract methods, Compose previews and Dagger/Hilt modules are framework-loaded, Gradle and Android convention plugins, test modules and manifests are entry roots, and a module with a program
mainis run (#3695, #3661, #3657, #3645). A co-change edge is not an importer in JVM, Go and C++ reachability (#3658), generated files read as unindexed users (#3639), packages that import their own directory are linked (#3884), and empty safe-only views and withheldunused_internalrows say why (#3646, #3668). - Security. History findings are deduplicated across file versions (#3631), the repository's
.gitleaks.tomlpath allowlist is honoured (#3709), other scanners' inline allow markers are honoured, generated, vendored and.d.tsfiles are scanned for secrets only, and sinks in scripts, tools and docs examples are low (#3685). Uppercase identifier values and__htmltype keywords in type positions stay quiet (#3420, #3687). - Public API. Exported Go identifiers outside
internalpackages and Java and Kotlin public API in published modules are flagged (#3537, #3577). - Change risk. Repowise's own output stays out of working-tree change risk and test selection, untracked files count the same everywhere, and text templates are data (#3712, #3770, #3767). Every per-file block is scoped to
include_paths, given as a list or a string (#3730, #3698), the indexed commit is read through one resolver (#3708), the health delta matches findings by identity across renames, line shifts and centrality changes (#3715), the overlap block counts only fixes before the change (#3632), MCP says which checkout and commit it read (#3719), and a missinggitis named (#3659). - Test selection. A conftest reading a literal constant or calling a self-contained function reaches only its module, and type-only conftest imports are ignored (#3731, #3772, #3633). A comment or docstring mention does not name a file (#3720), prose suffixes follow the documented extensions (#3729), tests reached through a package barrel are seen (#3651),
@SpringBootTestintegration tests run with every selection (#3636), only the base side of the index gap counts (#3628), and a second call shares a running selection (#3727). - Coverage. Files changed since a stale coverage report carry no figures (#3735), and any unapplied gate gives way to the project verdict (#3519).
- MCP. Freshness and live files are read from the served checkout (#3644). Caller definition lines are checked live and symbol search folds plurals (#3741), reference edit sets are complete for Go, Java and Kotlin packages (#3642),
Class.methodandClass::methodtargets resolve with no file (#3635), a missedget_contexttarget names the nearest symbols (#3638), tree-sitter query and history-tier files are classed as what they are (#3723),used_byand search notes agree with their payload (#3699), andget_riskstates its churn rank as a rank (#3560). get_answerand search. Retrieval reaches function bodies by rare terms and literals, a class's constructors count as part of the class, a project-name word or directory name carries no extra weight, and a body named only as context is served as evidence (#3736, #3773, #3771). A value is extracted only for the constant the question names, and each file appears once inbest_guesses(#3700, #3673). Hybrid search scores symbols on content words, and code-punctuation literals are scanned and kept out of path mode (#3677, #3721).get_why. A target is governed only by exact scope with the reason worded from its basis, rationale comments rank by weighted coverage, commits are deduplicated by full sha,git_logis capped at five, and an index with no accepted decisions or a missing path says so (#3769, #3724, #3726, #3655, #3654).- CLI.
repowise riskreports git's message for an unknown ref (#3622),repowise whyprints each history commit once (#3608),repowise doctornotes a Claude Code MCP entry that serves another repository (#3649), a keyless index reads as complete and init, status andget_answername the keyless upgrade commands (#3634, #3143), andgenerate --unwrittenseeds from model-written pages only (#3811). - Distill. Verbose
go testruns keep their failure messages (#3627), andgit log -nis honoured with each commit's author (#3660). - Providers and integrations. The default Ollama embedding timeout scales with the batch size (#3578), and OpenCode CLI v2 is supported (#3572).
- Decisions.
superseded_byis validated against the repository's own decisions (#3748). - Web UI. The docs Export menu opens above the header (#3778), MetricCard labels and descriptions render as blocks (#3875), UI copy uses correct singulars, ordinals and field names (#3465), charts share one median helper and one effort label map (#3561), and UI components fill in summaries and states when a host omits them (#3566).
- Upgrades. The chat sessions migration is chained after Change Records (#3763), REST responses carry
previewonly on preview cells (#3877), and one sharedgit cat-file --batchreader serves every analysis consumer (#3702). - Release-day hardening. Suspects runs its git reads off the MCP stdin pipe, so it answers promptly on every platform (#3888), and
get_change_riskaccepts only revisions, never git options (#3893). An oversized omission ref reads through page by page (#3899), the MCP handshake reports the repowise version (#3900),get_overviewnames unknownincludevalues inignored_arguments(#3916), and dead-code cross-repo notes list each consuming repository once (#3895). - Index accuracy. A symbol spotlight retires when its symbol leaves a live file (#3892), test-quality markers read the stamped test flag (#3898), replayed function hotspots are re-measured on upgrade (#3891), update keeps non-code co-change partners such as templates and stylesheets (#3896), git-ignored manifests stay out of package discovery (#3894), and src-layout packages are service boundaries (#3897).
- CLI.
repowise changes showaccepts the forge refchanges listprints (#3904),statussync totals agree with the pages table (#3889),doctorleaves tombstoned pages out of store reconciliation (#3907), distill runs a forward-slash program path on Windows (#3918), and SyntaxWarnings from analysed Python stay off the console (#3890). - Internal: the contract-facts refresh and the C# and Swift dead-code passes are split into helpers, and unused code is removed (#3863, #3846, #3538, #3883, #3417, #3339, #3795, #3774, #3696).
Documentation
- Layer pages for AI attribution and Change Records (#3775), and every layer page rewritten as a scannable present-tense guide with method detail in the architecture docs (#3777).
- The benchmarks page is rewritten for readers new to repowise, with October scores (#3640).
- Language support lists class metrics for Dart, F# and Objective-C (#3479), and the architecture docs say semantic search uses LanceDB on every SQL backend (#3155).
Dependencies
- Core:
tree-sitter-swiftpinned to>=0.7.4,<0.8(#3647). - Web: the unused
recharts2 dependency is removed; charts use recharts 3 from the UI package (#3780). - Security updates: fsspec 2026.9.0, Mako 1.4.3, multidict 6.9.1, source-map-js 1.2.2, postcss-selector-parser 7.1.6 and KaTeX 0.18.4 (#3887).
[0.56.0] - 2026-10-12
repowise 0.56 hands agents the exact thing they need to act on. Search returns matching source lines and finds every file in the repository, get_context lists every edit site of a symbol verified against live files, caller rows carry the calling line, and repowise update --working-tree brings uncommitted work into the index in seconds. repowise impacted-tests selects a targeted subset on far more changes, with --prioritize and --explain.
Refactoring plans can be applied as written: large functions get staged Extract Method plans with each helper's header and call, text anchors and the tests to run between batches, on a precise dataflow. Fix first ranks by complexity removed times reach and by the execution role of the code, every surface reports the same counts, and Code Health views are served from stored rows in milliseconds. repowise now works natively with GitLab, Azure DevOps and Bitbucket, sets up GitHub Copilot CLI and Kiro, and the web UI ships in Simplified Chinese.
Added
- GitLab, Azure DevOps and Bitbucket are first-class git hosts (#3350, #3361, #3362, #3367, #3368, #3381, #3382, #3386, #3351, #3355). Remotes, nested GitLab groups and Azure
org/projectpaths parse into one key. Merge request and PR numbers are read per host (!Ntrailers,Merged PR N:,(pull request #N)), so every change links to the right review. Noreply addresses and bot accounts are recognised per host. CI build facts are read for GitHub Actions, GitLab CI, Azure Pipelines, Bitbucket Pipelines and Jenkins, anddefault_revspecworks on Azure Pipelines PR builds. search_codebasereturns matching source lines (#3131, #3130, #3189, #3191, #3199, #3201, #3128). Identifier and literal queries returnlines(path,line,kind,text, definitions first) with acompleteflag andreasonswhen it is false. Prose queries carry up to 3matched_lineson the top file rows, and file rows name the matching symbols with their lines. Literal scans narrow withgit grepand read source before tests and docs, so on a 20k-file TypeScript repository a key is found in all 17 source files that use it. One test file's mocks are capped at 3 plain references.patternis accepted as an alias forquery.- Search and
get_answerreturn files that have no wiki page (#3210, #3214, #3194, #3180). A file found by the symbol or the new filename lookup gets a row even when the index has no page for it. On a 21k-file TypeScript repository three such targets now rank 1, 4 and 5, with search and answer MRR unchanged on a 247-question localization set. Prose queries surface symbols whose name words match, and path mode matches spaced queries against filename words. get_contextlists every edit site of a symbol (#3126, #3190, #3183).include=["references"]returns the definition, import, call and other live occurrence lines, checked against the files on disk, withcompleteandreasons. A local import alias is followed, and one file's plain references are capped at 3 with the rest counted insites_omitted_by_file.- Caller rows show the calling line, through forwarding wrappers (#3198, #3195, #3197, #3181).
get_contextcallers andget_answergraph_callerscarrycall_lineand its livetext. A thin wrapper that only forwards to the target is followed one hop to the real caller. Functions passed as values in TypeScript, JavaScript and Python (registered in a list, passed as a callback, returned, exported and registered) and names destructured from a dynamicimport()now have callers. get_answeranswers caller and trigger questions from the call graph (#3196, #3212, #3116). "Which file calls X" questions get agraph_callersblock, impact questions show who uses the lead file, and files named in a pasted stack trace (Python, JS/TS, Java/Kotlin, Go, Rust, C#) lead the results. On probe questions over a 20k-file TypeScript repository, keyless replies hold every expected file in 3 of 3 cases.repowise update --working-treeindexes uncommitted edits (#3095, #3107, #3100). Symbols, edges and pages for edited files are current without a commit, and a working-tree-only update skips the history and analysis phases (a 2-file edit on a small repository refreshes in 4.5 s, down from 12.5 s). MCP responses mark which served files carry uncommitted edits.get_contextskeleton+ view and focused symbol skeletons (#3097, #3135, #3136).include=["skeleton+"]keeps every line except function bodies, each replaced by a marker with its line range. A symbol target's skeleton keeps only that symbol's body.get_symbolon a large class returns an outline with its members' ids.impacted-tests --prioritizeand--explain(#3322, #3217, #3341). Selected tests are ordered likeliest to fail first, and--prioritizeemits the whole suite with the selection at the head.--explain <test>says why a test was or was not selected, and JSON output carries the same record per test. Tests that walk source directories or start the project's own command run with every subset.- Staged Extract Method plans for functions one helper cannot fix (#3387, #3392, #3433, #3470, #3468, #3471). The function keeps its setup and calls helpers in order, each stage at most CCN 10 and 60 lines, cut inside a loop body or
ifarm when that is where the complexity sits. Every stage is anchored by its first and last line text, and plans over 5 stages are served in batches with the tests to run between them. Applied from the plan text alone on 15 of the top 30 "Break up X" items across three repositories, every resulting CCN matched the estimate. - Extract Method plans write out the helper's header and call (#3352, #3330, #3254, #3369). Plan detail carries the helper signature and the replacing call in the file's language (Python, TS, JS, Go, Java, Rust, C/C++), whether it is a method, what fields it writes, and whether it awaits. A due Fix first item carries its first step and verify command inline.
- Refactoring plans name the tests that call their code (#3478, #3323, #3289, #3486). Plans quote up to 5 nearest test functions as runnable
path::testselectors, each step of a multi-step plan gets its own verify, tests that only pass through a hub module are dropped, and a JS/TS test command runs from the package that owns the test. - Refactoring plans carry decisions, consumers, dead code and recent edits (#3348, #3485, #3390, #3459). A plan says which accepted decision governs its file, whether other code imports what it moves, whether its target is unreachable, whether the function is being edited right now, and whether it changes a symbol a package exports. An Extract Helper group whose occurrence already is a function calls that function instead of proposing a new one.
repowise health --plan <id>prints a plan as an agent recipe (#3364). A plan exports as one recipe of preconditions, steps and checks, and one core renderer writes the agent prompt every surface shows.- Resolved plans record what happened (#3391, #3363). Each function's CCN, size, parameters, nesting, execution role and body facts are stored, and a plan that stops being detected records its outcome with before and after measurements.
- Performance causes carry the execution role of their loop (#3347, #3428). Request, event consumer, scheduled job, startup, CLI, UI, tooling or test, seeded from routes, job registrations, lifecycle hooks and command decorators in Python and TS/JS. Cold roles leave the default queue, and REST and the UI filter by role.
- Performance findings say whether the loop grows with the data in Go, Java, C# and Ruby (#3489). A database or network call over query rows is told apart from one over a fixed list. On a Ruby application the default performance list goes from 0 to 16 proven items.
- Performance findings say whether the loop grows with the data in Rust, Kotlin and C++ (#3515). Over 18,000 loops in ten Rust, Kotlin and C++ repositories, 6 to 15% of loops per language move from unknown to a proof, with no default list row lost or gained. A parameterised
LIMIT ?is a capped read in every language. - C# and C get Extract Method plans (#2919, #3441). C# (and Razor) has a def/use dataflow dialect, and C uses the C++ one.
- PHP code health (#3089). PHP has complexity, cohesion and assertion support and is scored like the other supported languages.
- The secret scanner detects GitLab and Azure DevOps tokens and credentials in git remotes (#3356). CLI warnings keep those tokens masked.
- GitHub Copilot CLI and Kiro targets (#3173, #3169).
--target=copilotwrites the Copilot CLI MCP config,--target=kirothe Kiro MCP config and a steering file, and--target=vscodeadds Copilot instructions in.github/copilot-instructions.md. The integration matrix gains a Basic tier and an Indexing column. - Cursor shell commands are rewritten through a
preToolUsehook (#3177), for command families configured asallow. repowise initdefaults the provider picker to the agent you use (#3171). When Claude Code, Codex or OpenCode is set up and signed in, indexing uses that login by default.- The web UI has a Simplified Chinese catalog and a language switcher (#3129, #3151, #3327).
get_healthopens a Do next action by its id (#3544). Theact_idsget_overviewhands out resolve withget_health(fix_id=...), so an agent can follow a Do next item straight to its findings.- The inline ignore marker also clears the stored scan (#3536). A line marked
repowise-security-ignoreleaves the dashboard,get_riskand REST as well asrepowise security check, which still reports it as suppressed. - Dead code and security report when they last ran (#3303). The UI tells "never ran" from "ran and found nothing".
Changed
- Fix first ranks by complexity removed times reach (#3366, #3422, #3484, #3482). Worth is complexity removed scaled by the file's importers, and Do next and Fix first lead with the same item. On a Python agent repository a CCN 963 function rises from #25 to #1. Terminal print and prompt flows, moderate functions in quiet files, and small closures credited with a whole file's importers leave the lead under named, counted exclusions. A plan that would leave a function above the break-up bar ranks after plans that finish the job and says so.
- Every surface reports the same counts (#3354, #3328, #3291, #3299, #3288, #3419). Each finding, performance cause and refactoring opportunity stores its queue verdict at index time, and CLI, MCP and REST report one count object (
inventory,in_scope,eligible,due,shown,excluded). The Do next sentence is written once in core. - The performance queue leads with loops that grow (#3276, #3258, #3395, #3253, #3250, #3443, #3425, #3447). A cause whose loop size is unmeasured never leads. Migrations, startup, shutdown, recovery, fire-and-forget tasks, inert call sites and functions gated off by a module flag stay out of the queue. Plans give one step per call site or per loop and variable, a chained query is one sink, and a TS/JS import's library is read from its whole module string.
impacted-testsselects a targeted subset on far more changes (#3219, #3243, #3379, #3262, #3216, #3220, #3218, #3221, #3325, #3353, #3241, #3358). When the index lags the change's base, files changed in between are compared and only the routes they rewired are added; replayed over 32 merged PRs, a trailing index is never the only reason to run every test. Manifests, lockfiles, production__init__.pyfiles and assets select the tests their reach implies (over 83 replayed failed PR runs, full-suite runs drop from 39 to 9 with zero new misses). A conftest reached only through its imports selects only the tests that can break, and setup files, path-named helpers, lazy string registries, Alembic migrations and VS Code extension entries are wired into the graph.get_change_riskandget_riskname tests with theimpacted-testsselection (#3296, #3101). Both tools use the same selection and order as the command, and the PR directive liststests_to_update, the test files a change will probably need to edit.get_riskreplies are leaner (#3124, #3117, #3118, #3123, #3139, #3152, #3120).structural_impact_score,overall_risk_scoreandmay_break_testsare dropped by default,pr_blast_radiusships only withinclude=["blast"]and detailed owner metrics only withinclude=["owners"]. PR summaries count totals, andget_change_riskscores a change that also touches docs or config files.- Plugin: the change review, code health and pre-modification skills and Lens follow the new
get_risk,get_change_riskandrepowise healthbehaviour, and the plugin passes the plugin directory's checks with an icon and links (#3296, #3118, #3139, #3152, #3280, #3043). - Bare
repowise healthreads the stored analysis (#3280). It writes nothing and returns in seconds;--recomputeruns the in-process analysis as before. - MCP server instructions say when to use each tool (#3134, #3125). Hosts that defer tool schemas see one "use X when" line per task.
- Keyless
get_answerreplies are compact and lead with the right file (#3108, #3137, #3184, #3102, #3186). Low and keyless replies serve a ranked shortlist with each file's functions, size and lines, and the answer sentence leads with the file retrieval ranked first. - Extract Method offers only spans worth making (#3252, #3274, #3287, #3421, #3427, #3277, #3410, #3415, #3488). A span must cover at most 60% of the function, hold at least 8 code lines and remove at least 2 decisions. Spans across which a closure shares a local, that lose a local import, that split a try from its handlers or change how an exception leaves, or that hold a Java switch-rule
yieldare refused. Flat record builders are refused, every helper is named from banners or labels in the file's casing, and a plan stays within one language. - Import cycles are found over runtime edges (#3259, #2803, #3077). Cycles count only imports that run: imports under
TYPE_CHECKING, TypeScript type-only imports, imports inside function bodies anddynamic_useshint edges are left out, andbreak_cycleplans are advisory. - Complexity reads nested functions and flat chains correctly (#3405, #3278, #3338). A function defined inside another gets its own complexity entry, and TS/JS ternary dispatch chains, JSX conditionals and Python conditional-expression chains read as the flat dispatch they are.
- Test support, bundles and pytest-uncollected modules are classified once (#3429, #3218, #2669, #3046, #3290).
*.test-harness.ts,*.test-support.mjsand*.bundle.jsare classed the same on every ranked list, a test-named module pytest does not collect is production code,__mocks__counts anywhere, and Perl-stylet/trees are Python tests. - Dead code: one gate decides
safe_to_delete(#3511). Docs and example code, export-only findings, and languages whose imports or calls are weakly resolved in the repository are never deletion-ready. - Dead code: an export only its own file uses is export-only (#3510). These rows are capped at 0.3, below the default floor, and are never deletion-ready. Reading
exports.xno longer makes a TypeScript file CommonJS. - Code Health pages tell all-clear from not-yet, filtered and error (#3308, #3310, #3306, #3304, #3300, #3309, #3307, #3313, #3326, #3301, #3314). Every health, refactoring, impact, coupling, decisions and owners view separates those four states. The dead code page is one figure, one list and one table, security leads with posture and scan freshness, chat is a single column with stage-aware progress, and proportion bars share one component and palette. The VS Code extension follows the native theme.
- Refactoring code drafts use the chat model (#3329, #3383). One switch on the repo settings page and in an open plan turns helper naming and code drafts on.
- Every LLM provider is declared once (#3168, #3167, #3172). Registry tables, pricing, the init picker and the settings page derive from one spec per provider.
anthropicandclaude_clidefault toclaude-haiku-5-5(#3093).- Faster reads on large repositories (#3311, #3315, #3423, #3279, #3340, #3389, #3255). Fix first and Do next are stored at index time (19.4 s to 0.01 s on a 20k-file repository), refactoring plan lists page from stored ranks (about 22 s to 0.2 s), and
get_healthlists andget_overviewskip whole-repository work (dashboard 8.6 s to 3.3 s). Read-only commands start without loading the parser or reconciling the schema, impacted-tests walks the graph in memory, and the health walk cache is saved after a parallel pass (a second pass on a 5,200-file repository takes 3.4 s instead of 106 s). updatechecks decision evolution in parallel, against this update's evidence (#3524). The judge calls share wiki generation's--concurrencysetting and read only the commits the update brought in, so the decisions phase takes seconds instead of minutes on an active repository. Its cost is booked as decision extraction.- The per-commit health scan runs after the update commits (#3525). The write lock is released before the scan starts, each update spends at most 60 s on it, and commits an earlier budget did not reach are picked up by the next update.
- Cohesion counts only groups of methods that share state (#3552). In Python, TypeScript and the other explicit-receiver languages, a method that touches no field no longer forms a group of its own; on our own packages flagged classes went from 87 to 15, and on five popular open-source libraries from 57 to 7, each checked by hand.
- A function worth breaking up leads its file (#3526). Fix first and the file card lead with the largest break-up candidate instead of a smaller finding with a higher local impact.
- Stored git history is re-read once when its rules change (#3386). The first update after upgrading walks the whole history once, with no page generation and no model calls.
- The CLI starts faster (#3438). The git indexer's exports load on first use, so commands that never read history skip importing GitPython.
- Blame ownership honours
.git-blame-ignore-revs(#3078), so formatting sweeps listed there leave line ownership with the real authors. search_codebaseserves one location per row (#3317). Rows carrypath; the deprecatedfilefield announced in 0.55.0 is removed, along with page ids and structural titles a row can rebuild.
Fixed
- Refactoring plans stay visible in the health views on SQLAlchemy 2.1, which a fresh install picks up (#3476).
- An index written by 0.55 or earlier re-renders its structural pages on its first update, with no model calls (about 140 pages a second, once), and files newly indexed in this release get their pages; prose pages are marked stale for
repowise generate --stale(#3557). initstamps the parser fingerprint, so a parser change always reaches an index on its next update, and renamed symbols no longer keep their old graph nodes (#3558).- An existing index is rescanned for security findings on its first update, so every scanner improvement in this release reaches unchanged files too (#3516), and every update rescans the files it changed (#3530).
- Recording a decision leaves the index's commit untouched, so test selection keeps its targeted subset afterwards (#3541).
- Do next judges a recent regression by the finding as it reads today, quoting its current severity, reason and line (#3543).
- The CLI, MCP and the overview report the same file and finding counts (#3532), rescored rows carry the commit they were scored against (#3531, #3540), and Fix first titles name a helper only when the plan has one (#3529).
- A type imported from an external package never resolves to a repository file by name, and TypeScript and C/C++ name lookups stay within their own language (#3527).
repowise dead-codeon an index behind HEAD builds the graph the wayupdatedoes and keeps its log lines off the table (#3520, #3528).- Extract Method on Java, C++, Rust and C# keeps a value a branch may have written (#3460).
- C# and C++ attributes after a comment are captured (#3461), and a deleted file's symbol pages are retired with it (#3464).
- repowise reads HEAD through a worktree's
.gitfile (#3293), finds its local index after a repository moves (#3240), and MCP finds a repository by any spelling of its path (#3275). - Every source walk uses the indexer's output-directory rule, so decision, vocabulary and workspace walks read source packages named
build,coverageordist(#3454). - Test selection orders the files that name a test deterministically (#3453), and communities hold only the repository's own nodes (#2611).
- Extract Method returns values that a later conditional write does not kill, gives plans only the inputs their span uses, and reads Python loop
elsebranches, Go writes through a field or index, comprehension scopes andexcept ... asnames correctly (#3384, #3357, #3424, #3418, #3413, #3349, #3294). - A docs-mode
updateprunes deleted files' findings, graph nodes, plans and embeddings (#3412). - Credentials in a git remote or a registered repository URL are stripped before they reach the overview or repo responses (#3343, #3346).
- The security scan reports a sink only where code can run it, and reads weak hashes, f-string SQL and TLS verification from the code rather than its text: a
hashlibcall markedusedforsecurity=Falseis a checksum,verify=Falsecounts on HTTP client calls, and symbol names and placeholders stay quiet. Every secret kind fires as before (#3534). - A database read filtered to a set of keys (
in_(keys),filter_by(id=...), PostgREST.in_/.eq) is bounded by the keys, so it no longer leads the performance queue (#3535). - The security scan honours
pragma: allowlist secret, and skips documented placeholders and values containing shell or template expansions (#3246, #3179, #3319). A scanner change now triggers a full rescan (#3086). - Dead code recognises as used a function stored by its own registering decorator, a VB.NET Module whose members are used, Grafana plugin modules and Storybook config files (#3490, #3509, #3080, #3087). Files a tool config passes to a plugin factory,
.vscode-test.*and scripts whose own usage line runs them are entry points (#3508). get_contextresolvespath::Class.methodlikeClass::method,get_symbolaccepts a bare unique name or aqueryargument as the id, anddays_agoand fix-history recency are measured to the newest indexed commit (#3099, #3140, #3153, #3145).get_healthandget_riskredirect a path that was renamed or split (#2922).only=["fix_first"]pages up to 25 items and the refactoring plan list returns what was asked for (#3255).- Searches answer immediately (#3115), and
repowise mcpin an uninitialised directory leaves it untouched (#3166). - On PostgreSQL, existing columns are loosened at startup so upgrades need no Alembic step (#3068), and the
vectortype warning is silenced (#3157). claude_cliparses the event arrayclaude -pprints with verbose output on (#3170). Gemini chat accepts every tool schema (#3064), and reasoning models get room to write the synthesized answer (#3094).- A scripted
--no-proseinit stays fully keyless (#3242), and size-skip notices leave the index marked healthy (#3272). repowise uninstalllists and removes the post-commit hook and lists a configured database as kept (#2655, #3397). Decision commands in an unindexed worktree use the base checkout's index (#3398). Init drops a stale embedding model pin (#2654).- TypeScript configs with comments resolve their aliases, unique ambient types resolve, and source packages named
build,coverage,distortargetare indexed (#3067, #2999, #3221). - Go generic calls with explicit type arguments have call sites (#2992), and contract impact keys symbols by id so same-named methods are kept (#3176).
- Clone detection stays within one language (#3414), Rust grouped
useleaves are classified on their own (#3076),# noqaon Pythonexceptlines and Javatry ... fail()are respected (#2853, #2996),knowledge_lossignores a rename-only window (#3075), and a repository with no scored file has no score instead of 10.0 (#3053). - Numbered ADR folders are found at any depth (#3061), and doc drift honours Markdown closing fences and GitHub's suffixed anchors (#3239, #3256).
- Java, C#, Kotlin and Scala test classes such as
FooTest.javaandFooSpec.ktpair with their source (#3096). - The web UI decodes file content as UTF-8 and numbers the performance queue from 1 (#3249, #3248). Wiki backlinks are ordered deterministically (#2881), and the branch-overlap summary says how many branches were not scanned (#2934).
- A workspace with a shared database queues its one-time reindex (#3044), and the MCP watchdog stops at a dead or unnamed ancestor (#3142).
- The decision summary separates kept decisions from candidates (#3088), and the CLAUDE.md and AGENTS.md header steers agents correctly on keyless indexes (#3109, #3144).
- Internal: the functions this cycle made most complex are split up across ingestion, dataflow, the health walker, test selection, the refactoring planner, the update pipeline and the web UI (#3542, #3545, #3546, #3547, #3548, #3549, #3550, #3551, #3553, #3554). The health analyzer, refactoring serving path and pytest edge linker are split into smaller units, unit tests run in parallel in CI, and unused code is removed (#3388, #3380, #3321, #3292, #3159, #3098, #3487, #3370, #3045, #3200, #3203, #3146, #3122, #3416, #3442, #3448, #3298).
Documentation
- A git hosts page covering GitHub, GitLab, Azure DevOps and Bitbucket, and a contributor page for the forges package (#3400).
- Performance has its own page, and the Code Health page is shorter (#3264).
- A contributor guide for adding agents and indexing backends (#3178).
- An API reference for what OpenAPI cannot carry (#2342).
- The Claude Code plugin README says what the plugin runs and what leaves the machine (#3041).
- VS Code extension 0.11.0 (#3042).
Dependencies
- Web:
next-intladded for translations (#3129). UI and web:thinking-orbsreplaces@lottiefiles/dotlottie-reactfor the loader (#3304). - Dev:
pytest-xdist(#3098).
[0.55.0] - 2026-10-04
This release is mostly about precision. The call graph resolves receivers, overloads and shadowed names the way each language's compiler does, measured against compiler and SCIP oracles. Dead code, health and refactoring advice stop reporting what the evidence does not support: files in languages health cannot analyse no longer score a perfect 10.0, an unused export reaches full confidence only when an importer names it, and performance and refactoring plans that would not help are withheld. Code Health now leads with one Fix first queue across defect, maintainability and performance.
Agents get leaner, more useful MCP responses, module wiki pages are written from structured facts with checked diagrams and citations, and indexing uses less time and memory. The Claude Code plugin ships Lens, a mod that shows Repowise's view of the repository while Claude works.
Added
- Plugin: Lens shows Repowise's view of the repo inside Claude Code (#3021, #3022, #3023, #3025). Lens ships in the Claude Code plugin as a mod: a setup and freshness band above the prompt, a post-turn change review, a
/lenspane with a per-turn Flow dashboard, and a repository map lit by what Claude reads and edits, which you can inspect and zoom from the keyboard. It adds no model calls, reads the local Repowise index, never blocks, rewrites or approves Claude's tool calls, and sends Claude nothing unless you press a button. It needs Claude Code 2.1.287 or later, an indexed repo, andrepowise serve --no-uifor the map. See docs/agent/LENS.md. - One Fix first queue across defect, maintainability and performance (#2821, #2817, #2839, #2841, #2820, #2825, #2831, #2838, #2832, #2859, #2843, #2848). Code Health leads with a single ranked list built from stored rows, with one lead item, its steps, a Verify block and the reason for its rank. A performance item points at the loop header line it changes, and performance opportunities are one per place you change. Code that does not ship (vendored, docs examples, generated, tooling, tests) stays out unless scope is
all. Suggested verification tests are ordered by measured coverage of the changed lines, then by tests that name the changed symbol.get_health(fix_id=...)opens an item, andfix_firstreplaces thedirective,refactoring_directiveandperformance_directivekeys inget_health. Do next and Fix first agent prompts are shown in the prompt modal with Ask in chat. - Files health cannot analyse show as "not analysed", not as a score (#2955, #2963, #3038). PHP, Swift, Elixir, Haskell and the other languages without a health dialect were stored at 10.0, so on
laravel/frameworkall 2,992 PHP files read as perfect. They now store no score and are left out of averages, hotspot health, band counts and trends. The code health map draws them grey, the file drawer says "Not analysed", andrepowise statusand the badge read "not analysed" or "no data" when nothing is scored.get_healthreportsunanalysed_file_countand alanguage_not_supportedreason. - Each file's code origin and function-level facts are stored (#2820). Every file is classified as production, test, vendored, docs example, generated or tooling by one shared rule, and complexity findings carry the share of CCN held by the largest switch or dispatch chain and a deprecation marker.
- Related work across lenses in the health drawers (#2847). A file's drawer lists the findings, Fix first items, refactoring opportunities, performance work and dead code held for it, capped per lens with totals.
get_health_findingsandget_dead_code_findingsaccept afile_pathsfilter. get_answernames the files to read at every confidence (#3012, #3009). The reply carries a ranked list of candidate files not already cited: three at high confidence, five otherwise. A long pasted question triggers a bounded search (12 content nouns, 4 numbers), and the recovery hint no longer repeats the question.search_codebasespends its result window on distinct files (#3011, #3014, #3019). A code-location query collapses several hits in one file into one row, and hybrid results keep only file-backed pages. Every row that names a file carries it aspath(filestays as a deprecated alias for one minor release). A capitalised English word such asAddorClientcounts as an identifier only inside backticks, next to::or when called.get_contextranks and caps the symbol card (#3010). A file card lists at most 15 symbols, types first, then functions and methods, withsymbols_truncatedpointing atinclude=["symbols"]for the full list. Markdown and reST previews report line counts, heading counts and the first three headings.get_overviewlists the heaviest package dependencies (#2749). The architecture block carries the ten heaviest package-to-package edges with their verb and weight. Edges that touch a test file are dropped from this and from the container view.get_change_riskcan be scoped to named paths (#2621).include_pathstakes gitignore-style paths, as a list or one comma-separated string, and the health comparison keeps only changes under them.- Signature changes are classified by what they do to callers (#2620). A contract change reads as none, compatible or breaking, so a reflow, a trailing comma or an appended optional argument is no longer treated like a removed required parameter.
- People, coding agents and bots are told apart (#2754). Identity classification returns human, agent or bot. Bots and agents never merge into a person, the active team size counts people only, and agent-authored commits stay eligible for decision extraction.
- The repository overview opens with a tiered system map (#2747). Boxes are containers from manifests, sorted into who uses the system, where it is entered and what does the work, with at most 7 boxes and a capped set of arrows. The map is built from structure, links to module pages and adds no model call.
- Module pages are written from structured facts, with checked diagrams and citations (#2740, #2739). Each module page prompt carries one JSON object of parts, import edges, key files, Public API, excerpts and the traced flow, and picks a sequence diagram, a flowchart or none. Every Mermaid block is validated before a page is stored, and a cited file that does not exist is dropped. Questions, the concept index and git signals move into a page digest that search and agents read beside the body.
- File pages open with sentences, not a stock line (#2741). A file page leads with the module docstring, or says what the file defines and who imports it, then where it sits in its layer. The reference below is compact: a Public API list with the first docstring sentence of each declaration, and one
Used byandDepends online per directory. - Present mode tells the repository as a short story (#2734). The deck runs from the overview through the architecture diagram, one slide per major part and one traced flow to where to start reading. The walkthrough mode is removed, and the dialog traps focus and respects reduced motion.
- The file page shows the owner's share of lines and the merge commit count (#2730, #2733). The header prints both "% of current lines" and "% of commits", and the Commits tile says how many merges were counted apart. The Code Health lede names the lowest-scoring test file, and provisional findings show an unverified mark.
- Lambda handlers named in serverless, SAM and Terraform config link to their code (#2834). A handler declared only as a string such as
handler: src/users.creategets an edge, so the function the runtime invokes is no longer reported as unused. repowise publishputs a local repository on repowise.dev, and the CLI, web UI and VS Code extension offer it with occasional dismissible tips (#3004, #3005, #3006, #3007, #2931).
Changed
get_riskcard responses are leaner (#3112).co_change_partnersrows keep onlyfile_path,support,conf_abandhas_import_link. Detailed owner metrics (owner_pct,owner_line_pct,recent_owner,recent_owner_pct,bus_factor,contributor_count) move behindinclude=["owners"], saving ~1.2 KB per target card.primary_ownerand the owner clause inrisk_summarystay unconditional.Call resolution is more precise for Go, Rust, Python, TypeScript, Java, C#, C++ and Objective-C (#2766, #2767, #2777, #2778, #2800, #2773, #2772, #2775, #2779, #2768, #2827, #2833, #2776, #2770, #2837). Go calls are typed through struct fields,
:=locals and range variables, and a chain on an external return type is no longer bound to a same-named repository function. Rust receivers are typed from parameters,letannotations and constructors. Pythonsuper()walks the MRO. A parameter or local that shadows a name no longer resolves to the module-level symbol, and a member call is resolved by its receiver first. Against compiler or SCIP oracles, precision moved on gitleaks from 0.9755 to 0.9908, on cobra from 0.9724 to 0.9944, on httpx from 0.9696 to 0.9831 and on click from 0.9358 to 0.9572.Java, C# and C++ overloads each get their own symbol (#2811, #2829). Overloads of different arity, and a C# generic type beside its non-generic sibling, are separate graph nodes, so a call draws an edge to the overload it names.
TypeScript object-literal methods and C/C++ functions returning pointers or references are symbols (#2772, #2774). Methods in mixin and builder objects get ids such as
ZodNumber::gt, which on zod moved recall from 0.7762 to 0.7959.Java records the access level it was written with (#2771, #2798). A declaration with no access keyword is package-private, not public, and a record's canonical constructor takes the record's own access level. On gson, unused-export findings dropped from 11 to 4.
The MCP tool result is compact JSON, and
_metais lean by default (#3018, #3017). The text block no longer pretty-prints JSON, and routine responses dropcontract_version,timing_ms, an untouchedresponse_budget, an uncappedcompletenessblock, a completeindex_scopedigest and the savings fields.get_overviewkeeps the full envelope andMCP_CONTRACT_VERSIONmoves to 3.REPOWISE_MCP_DEBUG_META=1restores the diagnostics andREPOWISE_MCP_PRETTY_JSON=1restores indented text.get_contextshares its budget across targets (#2727). Each target gets an equal share, unused share passes to the others, and an over-budget target degrades in place instead of being dropped. When a target is dropped the response leads withdropped_targetsand a recovery call.get_answerreads typed identifiers and sequence questions better (#2736, #2737). A question namingmodule_pagereaches the files defininggenerate_module_page, and a "what happens when" question is answered from the entry function's call order.Code Health leads with a short score headline (#3026, #2989). The headline holds the score, the band, one line on scope and the most-changed files. The band breakdown and the method notes sit behind two closed toggles, and Fix first opens closed beneath it with its item count.
Fix first leads with work that is worth doing first (#2965, #2950). A function-size item leads only when the function is far past a bar (CCN 40, 200 lines, or nesting 6 in 100 or more lines) and its branching is not one dispatch. Rust unwraps and panics, rated 1 useful in 19 across four Rust repos, no longer appear in Fix first, and unwraps that cannot panic are no longer findings.
Dead-code confidence reflects the evidence (#2903, #2910, #2912, #2967, #2990, #2750). An unused export reaches 1.0 only when an importer names one of the file's symbols, otherwise it scores 0.6 and is never deletion-ready, and C, C++ and Objective-C exports never reach 1.0. A C# file no edge reaches is capped at the review tier whatever its age, and a C# file is reachable when a type it declares is named by its project or a project that references it. On
dotnet/runtimethis removes 274 C# files that were reported unreachable at 1.0 on age alone.Dead code is quieter for entry points and tool-loaded files (#2916, #2982, #2975, #2979, #2917, #2986). Files with a shebang or a
__main__block, files a CI workflow, Makefile or manifest names by path, composerautoload.files, Django installed-app modules, tool config loads, C# polyfills, JMH and build-named JVM classes, and Kotlinexpectandactualpairs are not reported. A folder of Dockerfiles or shell scripts is not a zombie package.repowise dead-codeandhealth --refactoring-targetsread the stored results (#2769, #2813). With an index at HEAD they return what the index stored, in the same order as MCP and the web UI, and fall back to live analysis (--recomputefor refactoring targets) when there is no index, the index is behind HEAD, or a lower confidence floor is asked for.Cohesion, duplication and hotspot findings are narrower (#2908, #2764, #2951, #2907, #2906, #2905, #2918, #2915, #2255). Java, C#, C++ and Kotlin classes that read their own fields without
this.are no longer split into one component per method, and Rust builders, accessor types and enums are not called low-cohesion.dry_violationcounts clones with other files only, and a long parameter list is reported only when the author can change the signature. A function hotspot needs at least five changes. Goselectarms and TS and JS default arms count in complexity.Test, vendored, generated and build code is classified once (#2899, #2913, #2987, #2958, #2954, #2993, #2980, #2814).
generated/output,external/anddeps/C libraries, minified trees, Gradle build logic, benchmarks, Cargo examples, JVMqa/projects, schema migrations and a docs site's code samples take one shared classification that dead code, Fix first, performance and refactoring read.Performance findings stop advising against code that is correct (#2911, #2923, #2981, #2984, #2969, #2898, #2968, #2985, #2964). Lock-guarded one-time initialisation and file-backed logger writes no longer get a "move I/O out of the lock" plan, and a batch plan is offered only when the loop key reaches the call. String concatenation is reported only when it grows across iterations, and
hot_path_sync_iono longer claims a function is request-reachable.Refactoring suggestions skip spans and moves that would not help (#2914, #2960, #2973, #2952, #2956, #2970, #2976, #2991, #2930, #2932, #2816, #2751).
move_methodleaves overrides, factories, stateful methods and C# static classes alone,extract_classleaves data holders and trait impls alone, and an Extract Method span covering three quarters or more of a body is refused. Extract Method credits only the share of the finding that moves into the helper, and the refactoring queue never leads with a test file.Extract Method infers parameters and returns correctly (#2927, #2953, #2959). Reads inside a closure count as uses, a name the span declares stays out of the parameters, and a may-def's bookkeeping use is no longer a parameter.
Same-package JVM, C# and Rust cycles are labelled idiomatic (#2983). Such a cycle is still reported, at low confidence with
idiom=same_package, and no longer as a cut to make. Declaration-only edges no longer formbreak_cycleplans.Finding triage survives re-index (#2819). A re-detected finding that matches an acknowledged, false-positive or resolved row refreshes that row: acknowledged and false-positive stand, and resolved reopens. Finding ids are anchored on the symbol and the line offset into it, so edits above a function keep the id.
Next-call strings share one shape (#2849). Actions, Fix first, health summaries and
get_riskbuild their suggested calls from a tool and its arguments.Opportunity lookups report
found, notresolved(#2812), andget_dead_codesays what its summary counts (#2842).get_overviewand wiki ordering skip the test layer (#2726), and knowledge-graph layer names are unique (#2725).Provider SDK clients close on the loop that used them (#2949), which ends leaked-client errors at exit.
Shared UI components can hide CLI hints and the refactoring scope control (#3037) for an embedding that has no CLI, and the untested-hotspot rows highlight on hover again.
Serving logic for dead code, refactoring, performance and MCP tool selection moves into core (#2752, #2846, #2850, #2851, #2852, #2854). Output is unchanged. Tool registration gains an
available_whenpredicate.
Fixed
updatere-scores and re-parses an unchanged checkout after an upgrade (#3035). When the health analyzer or the parser changed since the index was built,updatesays so and refreshes scores, findings and the graph instead of printing "Already up to date", so an upgraded index matches a fresh one without waiting for a new commit. No model is called.- Fix first Verify names tests for findings with no plan (#3033). Such items showed an empty Verify block; they now list tests by measured coverage of the finding's lines, then tests that reach the file.
- No batching plan for a loop that already queries a chunk at a time (#3034). A call in a for loop's iterable is attributed to the loop around it, so a chunked query loop no longer gets a "batch the database calls" plan.
- Go builder chains keep their call edges (#3036). A chain such as
NewReport(...).WithPhase(...).Build()types each link through the repository return type, including methods declared in a sibling file of the same package. - Update and init write the same results as a fresh index (#2856, #2872, #2871, #2926, #2904, #2830, #2873).
split_fileplan ids, the lead finding of a tie, I/O import classification, governance findings kept across a re-score, opportunities left open by an older plan model, and the community label map no longer depend on which path wrote them. - The first update after init no longer rebuilds every page (#2765). The
hooks.*andeditor_files.*keys init writes were read as a traversal change. init --resumeregenerates pages and decisions that were left half done (#2883, #2482).- Search and index report a failed embed (#3015). A partial LanceDB install used to exit
initwith 0 and a healthy semantic index while every page embed failed.initandupdatenow mark semantic search unavailable, point atrepowise reindex, and exit 1 when a real embedder failed.doctorreports an unopenable store as a failing row. - Child processes stay off the MCP stdio transport (#3016). The git, ruff and ps calls reachable from MCP tools pass
stdin=DEVNULL, so a child never reads the JSON-RPC pipe. get_risk(changed_files=[...])works without targets (#2761, #2840). It failed schema validation before the tool ran. A call with neither argument returns an error naming both.- An MCP server whose loaded code predates the checkout is told to restart (#2844).
get_healthand the finding tools return therepo='all'error dict (#2728), and the chat dead-code card reads the tieredget_dead_codeshape (#2729). - Dead-code false positives per language (#2909, #2863, #2924, #2994, #2974, #2977, #2534, #2941). A C or C++ name written outside its declaration (a callback, a
#definealias, a P/Invoke export, a.defline) counts as a use, and a use from inside a symbol's own span does not. A module-path string,Outer.Inneron a Python nested class, Go package-scope and import-qualified names, and a Java static member import count as uses. - Graph and resolvers (#2877, #2857, #2855, #2961, #2978, #2943, #2475, #2940, #2971, #2972, #2884, #2824, #2835). Objective-C C calls in local initializers are kept, and VB.NET
MeandMyClasscalls resolve in the caller's own class. A C++ namespace or type opened after a macro is read as a scope. Rustself::,super::andmodpaths and brace-group imports resolve through the module tree, and amoddeclared inside a macro is a module import. PHP files loaded by a__DIR__path or a view name are linked, and PHP docblock class names no longer close an import cycle. Python__all__is a bounded visibility signal. - Health findings with wrong inputs (#2942, #2948, #2896, #2939, #2823).
defer_in_loopis suppressed when the next statement exits the loop, the Java package line stays out of the clone token stream, Rust localusepaths no longer count as I/O imports, a Go string reset no longer reads as a growing concatenation, and a C# task awaited earlier in the method is not flagged as blocking on.Result. - A data step that moves no rows no longer fails the schema upgrade (#2738). Opening a store that predated
primary_owner_line_pctwith nothing to move raised "Not an executable object: 0". - A failed tests-reaching request no longer reads as untested (#2632, #2732).
repowise generate --pagewrites the overview in a seeded worktree (#2748).- Workspace fixes (#2524, #2160, #2828, #2489).
serveno longer pins the workspace root's own database as the shared store, Postgres-backed watch and update route through the shared database,.Testsprojects are skipped in contract extraction, and extraction diagnostics are no longer dropped. - Smaller fixes (#2639, #2802, #2714, #2845, #2011, #2858).
rgoutput for a root-level file is kept in session rewriting, doc drift treats a reader-project path as uncheckable,get_change_riskrun lists expand a reachedconftest.pyto the tests under it, and logger levels are restored after a command silences them for machine output. - The web UI reads the server API key from a private variable only (#2721).
- Commit ids read from
state.jsonare validated before they reach git (#3039), and the VS Code extension follows a serve lock file only when it points at this machine (#3040). - Code Health is quieter, and pages stop scrolling past their content (#3031). Hovering one Fix first row no longer lights every row, file paths are no longer accent-coloured, and the Coverage tab's chart and table colour only the files that need attention. The Code Health notice now explains this release's score changes. Pages with long tables no longer scroll a screen or more below their last row.
- The security scan stops reporting names and rule prose as findings (#3030). A kebab-case value such as
x-api-keyorrepowise-security-ignoreassigned to a token- or key-named variable is a name, not a secret, while values with digits or capitals still report.verify=FalseandrejectUnauthorized: falsewritten in strings or comments are no longer reported; real calls still are. - A damaged vector store no longer stops a resumed init (#1882). Reading its page ids fails with a warning that points at
repowise reindex, and the pages are regenerated. - Workspace init shows progress while it persists generated pages (#2158), and the server's staleness check counts stale pages without loading their content (#2257).
Performance
get_riskreads only the targets' graph neighbourhood (#2763). On a 4.4k-file index a call dropped from about 40 s to about 0.5 s, and PR mode from 41 s to 1.5 s, with identical responses.- Full-text and vector writes are batched (#2780, #2880, #2868). A full re-render of redis (2,314 pages) dropped from 124-127 s to 92.8 s and from 20.1 GB read to 1.84 GB. LanceDB's metadata cache is capped at 64 MiB instead of 1 GiB so generation memory stops climbing.
- Init and persist use less memory and time (#2864, #2874, #2866, #2867, #2869, #2870, #2865, #2878, #2875). Bulk upserts flush in batches, new rows go in as bulk inserts, and the security scan masks comments and strings once and only when a pattern can match. Blame entries are shared per commit, the symbol subgraph and per-file dataflow parses are released after use, complexity passes walk each tree once, and init takes three serial waits off its critical path.
get_answerbuilds graph projections once per index state (#2735), andrepowise updatereads the stored vector width from a record instead of opening LanceDB (#2861, about 1.4 s per update on a real embedder).
Documentation
- README, benchmarks and user docs are rewritten around each reader's task (#3002, #2400). Layer docs share one shape, internals move to
docs/architecture/and field dictionaries todocs/reference/, and new pages cover ownership, troubleshooting, MCP response fields, hooks and workspace contracts.docs/agent/LENS.mddocuments Lens. - Reference pages match the code (#2724, #2731, #2759, #2760, #2815).
Dependencies
- Open npm and Python security advisories patched (#2723). next 15.5.27, sharp 0.35.5, dompurify 3.4.16, js-yaml 4.3.2, undici 7.30.0, vitest 4.1.11; pyjwt 2.15.1, urllib3 2.8.0, anyio 4.14.2, and the minimum litellm moves to 1.90.7.
[0.54.0] - 2026-09-30
Most of this release is about getting the facts right. Tests showed up as entry points, a package's real CLI did not, a file renamed last month started its history from zero, one person counted as three contributors, and dead-code reports flagged config files that a tool loads by name. Each of those came from a rule that guessed where it could have read the answer from the repository itself: the manifest, git's rename records, the symbol table. This release reads it.
It also brings coverage into CI. repowise coverage check gates a change on its patch coverage with no index and no API key, repowise impacted-tests picks the tests a change needs and runs everything whenever that is not certain, and one GitHub Action or GitLab template runs the coverage, documentation drift and security gates together.
Added
- Patch coverage gates a change in CI (#2630, #2634, #2644, #2649, #2664).
repowise coverage check [REVSPEC]reports the share of a change's executable lines the tests ran and fails below--fail-under(orcoverage.fail_under). It needs git and a coverage report only. Report paths accept globs and per-report prefixes,coverage.ignoredrops generated or vendored files, and a change with too few measured lines is reported rather than failed.coverage.gatessets per-area thresholds, andrepowise coverage suggest-gatesproposes areas from owners, layout and the graph. Branches on changed lines are reported beside patch coverage, never blended into it, with an optional branch gate. - Coverage compared with the base, and a coverage trend (#2659, #2642, #2636).
--base-reportor the coverage ingested at the base commit gives the project delta, and--max-dropfails a change that lowers it. Each ingest is recorded with its freshness and how its paths matched, and the last 50 ingests form a history shown as a trend in the dashboard and inget_health. - Each uncovered change names the test to extend (#2650, #2645). Uncovered ranges carry the innermost symbol around them and up to three tests to extend, and each carries its file's risk, so attention rows sort by it.
get_change_riskwith no revspec covers everything a push would bring, uncommitted files included. - An opt-in coverage re-ingest for agents (#2658). With
hooks.coverage_reingest: true, a test run inside Claude Code, passing or failing, refreshes the stored coverage through a repository-local hook.coverage addaccepts globs. - Risk as a CI gate (#2645).
repowise risk --fail-above-percentile Pfails a change ranked above the P-th percentile of the repository's own history, never on the 0 to 10 score. - Run only the tests a change needs, safely (#2665).
repowise impacted-tests --format argsprints runner-ready arguments (pytest node ids, Go packages, Jest or Vitest paths) or:all, with every reason on stderr. Anything not positively known runs the whole suite: lockfiles, build and CI config, test data and helpers no test imports, changed files no test reaches, an index built at neither end of the change, or a failed graph read.conftest.pyselects the tests under its directory, imported helpers select their importers, a document that code or a doctest config reads runs everything, and tests the graph cannot see into run with every subset.tests.full_run_onandtests.always_runextend the rules, and the GitHub Action and GitLab template expose the selection so a missing or failed one runs everything. - One GitHub Action and one GitLab template for all three gates (#2641, #2643). They run the coverage, documentation drift and security gates in one step, keep going when one fails, and can upload SARIF.
repowise doc-driftandrepowise security checkalso write a GitLab Code Quality report with--format gitlab. A new CI guide documents all three. - Documentation drift catches renamed and deleted symbols (#2661, #2631, #2640, #2646, #2647). A backticked identifier that was a real definition when the doc line was written, and that nothing names at HEAD, is a finding, with the rename as the suggested fix when one is clear. Findings suggest a replacement wherever the tree says where a reference went, the CI gate keeps only the drift a change is answerable for (
--since REVSPEC|auto), findings are grouped by document and marked new since the last update, and drift is recorded in health history. The VS Code extension shows drift in the Problems panel for Markdown files (#2648). - Security gate for CI and pre-commit (#2635, #2657).
repowise security checkfails a change on the findings it adds. Arepowise-security-ignoremarker silences a line,security.patternsadds your own named patterns, andrepowise hook install --securityadds an opt-in pre-commit check. - "Do next" leads the Overview (#2651). A short list of actions built from what the index already stores (health, fix history, performance, security, doc drift, dead code, decisions, coverage), each with its reason, effort and what done looks like, and a
repowise nextcommand. Dismiss, snooze and done are remembered. get_whyanswers from code comments and history when no decision matches (#2692). A question that matches no decision record now gets the rationale comments or commit subjects that state the reason, labelled as such and never presented as decisions.- Decisions whose commits were reverted are retired, and decisions whose files are all gone read as stale (#2675, #2694). A revert is matched from git's own records, and a restored file clears the stale mark on the next run.
- Commit history for non-code files (#2690). Docs, config, workflows and data files now carry commit counts, windows and authors instead of an unknown count. They stay out of hotspot and percentile rankings, and an existing index gains them on its first update (#2719).
Changed
- Entry points come from the manifest first (#2698, #2704).
package.jsonbin,main,moduleandexports["."](with a builtdistpath mapped back to its source when the build is not checked in),[project.scripts]and each distribution's package__init__.pyrank ahead of every filename guess. Test files no longer have to claim to be entry points to be exempt from dead code: a separate reachability flag carries that exemption. - Author identities are merged with guardrails (#2684). One person's GitHub noreply addresses, handle and machine-local email are joined into one contributor, and bots are left out of people counts. Two real email addresses are never joined on a shared display name alone, so someone who commits from two personal addresses still counts twice. Owners, the truck factor and the contributor count all read the merged identity.
- "Commits" means non-merge commits everywhere (#2683). Repository totals, per-file counts and the stats sample all count the same way, merges are counted separately, and a complete history no longer reads as truncated.
- A file keeps its history across renames (#2569, #2690). Churn, age, authorship, co-change partners and prior-defect counts follow the file to its new path, and rename detection is explicit.
- History windows are measured from the indexed commit (#2568), so re-indexing an unchanged tree on a later day gives the same scores.
REPOWISE_GIT_WINDOW_ANCHOR=nowrestores the wall-clock behaviour. - Git history weighs less on a file with no code problems (#2566). A file with no code finding loses at most one point to git history, and the score breakdown shows the cap each file was scored under.
hidden_couplingis advisory (#2596). It is still detected and listed but no longer lowers the score; on its own it ranked defect-prone files near chance.- Search ranks same-named symbols by one key, and a name that does not exist returns no symbol (#2701, #2691). Exact case, then kind (class before function before variable), then code before tests and docs. A code-shaped query such as
AuthServiceXyzno longer returns its fuzzy neighbour as if it were the match, and question parsing recognises lowerCamel and acronym names by checking them against the symbol table. - The C4 views show the system you run (#2699, #2697, #2678). System Context lists runtime services rather than every library and the repository's own packages, test and example projects are no longer containers, a container is named by its manifest and labelled by its code language, and a container's own components stay in its component view.
- Stats and the overview agree (#2710). One rule for the language split (code languages, with docs and config listed apart) and one module count on every surface.
- Package pages and a computed Public API (#2712). A package whose files span several concept groups always gets its own page. What a package publishes is read from its manifest entries or its barrel, following re-exports, aliases and
__all__, and listed in its page prompt, so a class consumers import but nothing inside the repository does is no longer left out. - Module pages see the code they describe (#2713). Module page prompts carry the Public API, signatures and capped bodies of each file's public or documented symbols within a fixed token budget, and the model is told to state behaviour only when an excerpt shows it.
- Wiki pages stick to the subject (#2681, #2702, #2705, #2709, #2711). Template pages are written only for real API contracts and never for tests, package roots bound concept groups, the overview drops capabilities the repository never defines, a UI layer needs UI files, the module prompt no longer invites the model to discuss dead code, page footers say what a page was generated from, and a template page names no model.
- Findings shown on each surface come from one registry (#2687). Every surface that lists dead-code or health findings reads the same record of which finding types are shown, so a withheld type never appears in one list and not another.
HEALTH_ANALYZER_VERSIONmoved from 31 to 36 across the health and git changes above, so existing indexes re-score on their next update.PARSER_SCHEMA_VERSIONmoved from 3 to 5 (Python import resolution and overload declarations), so the parse cache refreshes automatically.STORE_FORMAT_VERSIONis unchanged and nothing needs re-indexing.
Fixed
- Python imports resolve by their full dotted path (#2674, #2612).
from pydantic import Xcan no longer land on a repository's ownutils/pydantic.py,import typinginside a package no longer resolves to a siblingtyping.py, and qualified names use the importable module. A bare import inside a package still finds a module that sits beside the package's top level, asfrom adafruit_ticks import ...does in a CircuitPython library (#2717). - Calls on a chained field receiver resolve (#2703).
this.store.users.find()andself.store.users.find()now reach the method they call. - Overload stubs are declarations (#2696). Python
@overloadstubs and TypeScript overload signatures no longer stand in for the implementation in lookups. - TypeScript methods keep their parameters (#2680). Methods with an access modifier were recorded with a bare-name signature, and module-level regex and ternary constants are now symbols.
- No file self-loops, and tests no longer dominate PageRank (#2677, #2695). A
package.jsonthat exports itself no longer points at itself, and test-to-test edges are left out of PageRank and betweenness, soconftest.pyis no longer a repository's most central file. - Workspace packages are found at any depth and checked against the declared members (#2685), and a workspace export that names build output absent from the checkout binds to its source file (#2706).
- Dead code stops flagging files something loads by name (#2700, #2707, #2673). Any tool's
*.config.*file,.<tool>rcfiles,.pnpmfile.cjsandsitecustomize.pyare exempt. Files shaped like path-loaded entries (a lone default export, a script with a main block) are capped to the review tier. Internals used elsewhere in their own file, files named by path, and functions registered on a receiver are no longer reported. - Dead-code line counts are measured, not estimated (#2682), and a whole unreachable file stays in the review tier (#2716).
- Full-text search covers each page's file vocabulary (#2718), so keyword searches on an index built without a model match the terms a file defines. Existing indexes are widened and refilled when they are next opened.
- Health false positives (#2693, #2688, #2679, #2574, #2573). Clone pairs are verified on identifiers, so import blocks and same-shaped data tables no longer read as duplication. An untested hotspot pairs with tests that import it directly or through a barrel, and a file with nothing to test is skipped. SDK
executecalls and re-raising exception handlers no longer raise findings. A file's dependents are counted from code edges, not co-change history, and a chain through a re-export barrel counts as a dependency. - The update's full re-score keeps history markers (#2614), and displayed scores are floored so a figure never rounds into the next band (#2587).
- Test detection by whole words (#2676).
e2e-tests/andpkg_tests/are test trees,latest-release/andcontest_results/are not, and filename rules such as.test.apply to source files only, sotsconfig.test.jsonis config. - Hand-written files that mention code generation are kept (#2689). A generated-file marker must be in a comment or docstring opener, and tracked files under ignored paths are indexed.
- Refactoring plans are more accurate (#2570, #2571, #2572).
extract_methodis credited only with the impact its span recovers, each step carries only the findings it answers, andmove_methodno longer targets result types or base classes. - Security scanning reports credentials, not examples of them (#2660, #2618). Keyword secrets in comments, docstrings and doctests are examples, and call-kind findings ignore comments, strings and prose files.
- Windows: no console window flashes when the telemetry flusher or another detached child starts, for example during VS Code index updates (#2380).
repowise update --fullfinishes when a few pages fail and points torepowise generateto refill them, and Ollama prompts are no longer truncated (#2564). The cascade-budget warning separates model and structural pages (#2616), andrepowise generateaccounts for every planned page and prints the actual cost (#2615).- Decisions: the id migration no longer copies a record onto an occupied id, which failed persistence on every update (#2613); conventional ADR directories stay ahead of loose name matches (#2586); both commit miners read significant-commit metadata the same way (#2605); and archaeology cross-references match whole tokens (#2588).
get_answerreads "the evidence does not identify" as a hedge and lowers its confidence (#2708).get_change_riskno longer fails when the index has no repository row (#2553),get_symbolchecks for a repository path first (#2554), andget_health's targeted mode keeps its reads inside its session (#2670).repowise initedits only the repowise entry in.mcp.json, leaving other servers byte-identical (#1836). A malformed manifest no longer crashes the tech-stack scan (#2608), and a corrupt omission store closes its connection (#2555).- Anthropic cached input tokens are counted (#2433),
.envand other config dotfiles are classified as config (#2381, #2516), non-executable post-commit hooks are reported (#2476), the hook filters workspace targets and reports failures (#2478), coverage-gap uncovered lines come from the percentage (#2259), and Codex session tool calls are attributed to the assistant (#2652). - Web UI: commit filters stay visible on empty results (#2487), the repo avatar no longer shows twice on phones (#2583), graph narrowing handles communities outside the top 20 (#2656), and
text-xs/text-smsit on the type ladder (#2619).
Documentation
- A CI guide covers the coverage, documentation drift and security gates and test selection in one place (#2641, #2665).
- Dead-code docs list every unused-export edge type and the dynamic-import language coverage (#2477, #2629), graph benchmark figures match BENCHMARKS.md (#2628), a stale edge-type reference is fixed (#2536), and
get_overview's ownership block is described as it is returned (#2610). - Plugin: the
coveragecommand documents the new gate options.
Internal
- Many long functions across the server, core and CLI were split into named steps with unchanged output (get_health, get_why, get_risk, get_overview, the call resolver, the parser, the decision miners, the job executor and others).
[0.53.0] - 2026-09-24
A workspace is only as useful as the contracts it can read, and outside Python and TypeScript servers it read most stacks one call at a time. A Laravel route inside Route::prefix('v1')->group(...) lost its prefix, routes/api.php lost its /api, and apiResource produced nothing. A NestJS controller produced no routes. An Angular service calling HttpClient on an injected receiver, or an axios instance created in one file and called from another, produced no calls. Queues, pub/sub and sockets linked only when both ends spelled the same literal name. This release reads those stacks the way their frameworks serve them, and the workspace pages that show the result were rebuilt around drawers, so a contract, a relationship between services or a co-changing file pair opens in place with the evidence behind it.
Performance plans said more than their evidence supported. A plan now claims proven only when the loop itself shows that the rewrite returns what the loop returned. Repetition with nothing to change, such as N reads of N files, leaves the default queue. Two new detectors cover shapes with no call in the loop body for the old ones to see: an ORM relationship lazily loaded once per row, and an unbounded read deduplicated in memory.
PHP gets a graph that reads it correctly. Grouped use imports were dropped, same-namespace references had no edge, and PHP framework conventions lived nowhere, so a Laravel app's route files, providers and base controller read as dead code. On laravel/framework, dead-code rows drop from 680 to 392. The Architecture page opens on a Files view that bundles the links between communities, and security finding snippets stop storing credential values in plain text.
Added
- Laravel is read the way the framework serves it (#2512). Route chains and nested closure groups keep their prefixes,
resource/apiResourceexpand (honouringonly,exceptandshallow), androutes/api.phpis served under/apiunlesswithRouting(apiPrefix:)or a route provider says otherwise. A prefix that cannot be read, or two declarations that disagree, refuse the route instead of guessing. Listeners, policies, commands, middleware aliases and scheduled jobs get edges from the file that registers them, so an unregistered listener reads as unused. Eloquent models andSchema::createmigrations become data contracts. On a five-service TypeScript and Laravel workspace, all 39 cross-service HTTP links now match exactly, against 7 before. - NestJS routes, Node HTTP clients and Node ORM tables (#2527). NestJS controllers are served under the global prefix, URI version and controller prefix, read from wherever the bootstrap settings are declared, per app in a monorepo. axios, ky, got and ofetch calls are read directly or through an instance created with
baseURL/prefixUrl, and an exported instance is followed into the files that import it. Tables are read from Prisma, TypeORM, Sequelize, Drizzle and Knex. - Angular
HttpClientcalls and Angular entry points (#2531). Calls on a receiver typed or injected asHttpClientare read for every verb, with the base folded from theenvironment.tsthe service imports and from class fields built on it. Filesangular.jsonor Nxproject.jsonnames as loaded are anchored, and names written only in router config or module metadata count as used. Over four Angular apps, 129 new consumer contracts, and dead-code rows drop from 315 to 304. - Queues, pub/sub and socket events across Laravel and Node (#2525). Laravel dispatch sites link to the
ShouldQueueclasses that consume them, joined by fully qualified class name. php-amqplib, BullMQ, SQS/SNS, Redis pub/sub and NestJS microservices are read; NATS, Redis and Kafka pattern subscriptions link every publisher they match; two ends link only when they can share a broker. socket.io, NestJS gateways,wsand Laravel broadcasting to Echo and pusher-js are read as socket contracts. - Topics follow RabbitMQ bindings (#2501). A
bindQueue/queue_bindlinks each consumer of the bound queue to the exchange's publishers whose routing key the binding accepts, wherever the binding is declared. Every contract type now runs on one dialect base and one string-expression layer, and topic extraction got faster in the process (307 to 222 ms on a 724-file Python repository). - Workspace pages open their evidence in a drawer. A contract opens its detail and consumer test impact in place, and the contract list pages on the server with search and a linked / unlinked filter; a Needs attention section groups unmatched calls by the matcher's reason (#2497). A co-changing file pair shows each file's recent history, the commit pairs that likely formed it, and whether a contract link explains it, and the page reports which cap trimmed the stored pairs (#2499). Selecting a service or relationship on the System Map opens a non-modal drawer, and blast radius, breaking changes, conformance and the cyclic core are one lens control that recolours the map and lists its findings (#2500). Each surface offers a copyable prompt for an agent.
lazy_load_in_loop, a new Python performance marker (#2528). A loop over ORM rows that reads a relationship the query did not load issues one SELECT per row, and with no call in the loop body,io_in_loopcould not see it. SQLAlchemy, Flask-SQLAlchemy and Django are covered, and the finding names the fix (selectinload,select_related,prefetch_related). Blind-labelled on held-out applications: 29 of 32 correct on a fresh Django repository, 20 of 24 on SQLAlchemy. Advisory, and never the lead of the performance directive.unbounded_read_reduced_in_memory, an advisory over-fetch marker (#2495). A query with no limit whose rows are then deduplicated per key in code; the fix is to select one row per key in the query.- An
expectedperformance actionability state (#2514). Filesystem and subprocess repetition has no batch operation to point at, and a loop that already walks its data a chunk at a time is already batched. Both sat in the queue asinvestigatewaiting on a prerequisite nothing could supply. On this repository that moved 523 of 629investigaterows toexpected, one filter away and still counted in every total. - Vendor credential shapes and PEM keys are detected by value (#2223). AWS access keys, GitHub tokens, Slack tokens, Google API keys, Stripe secret keys and PEM private keys are found by their shape, not only by the variable name they are assigned to, in both the file scan and
repowise security scan --history. - The stats page is one report with scoped figures and seven new records (#2526). Figures drawn from the indexed commit sample say what window they cover. New records: longest function, most patched function, longest name, most reused name, biggest cleanup commit, longest silence between commits, and the test-to-source line ratio.
Changed
- Performance plans claim only what they can back (#2495, #2504, #2513, #2521). Fan-out against a pooled database or network client is advisory with a
bounded_concurrencyprerequisite, since independence says nothing about what the pool can serve. A chunked loop reportsloop_already_chunkedinstead of being told to batch. A bulk rewrite is proven only when one bulk call touches exactly the rows the loop did, and not when the loop builds its result in key order, calls something for its effect, or reads throughsession.get, which answers from the identity map when a row is already loaded. Loops report their magnitude (grows_with_data,bounded,unknown), and a raw SQL read with aLIMITno longer claims to grow. Plans carry ordered steps bound to the evidence's locations, and plan validation, which wasunknownfor 330 of 330 plans resolved by id, now resolves for all but 5 of 287. - The Architecture page opens on the Files view (#2503). Each strong pair of communities is drawn as one band sized by its edge count, third-party modules and unconnected files are hidden and counted in the key, and hovering a file draws its real imports, importers and co-change partners in about 1 ms at the default 1,500 nodes. Community colours are keyed by label, so they survive a re-index.
- PHP has one composer reader and one framework-facts registry (#2505). Nested manifests, PSR-0 and classmaps are read, and PSR-4 falls back to shorter prefixes as composer's loader does. Grouped and comma-separated
usestatements expand per class, and same-namespace and fully qualified references become edges. A namespaced class no autoload prefix claims is external, soIlluminate\Http\Requestno longer binds to a localLoginRequest.php. Laravel, TYPO3 and Symfony entry points anchor dead-code liveness. Measured on five public PHP repositories: guzzle 10 to 1 dead-code rows, monica 555 to 148, laravel/framework 680 to 392, with graph build time on laravel/framework down from 3.4s to 2.4s. - Pascal gains loop, performance and assertion coverage (#2233). A
for x in collection doloop was skipped by the complexity walker, understating CCN and nesting. A new performance dialect covers filesystem and subprocess sinks by RTL/VCL/FPC name, theusesclause classifies database and network units soTDataSet.Openor an HTTP client's.Getin a loop is a finding, and a parenless call such asQ.Open;is recognised as a call. DUnit and DUnitX assertions are counted, including insidetry ... finallyblocks: on a real ~150-file Delphi test suite, assertion coverage moved from 66 of 120 files to 86 of 120. HEALTH_ANALYZER_VERSIONmoved from 21 to 31 across the performance and Pascal work above, so existing indexes re-score on their next update.STORE_FORMAT_VERSIONandPARSER_SCHEMA_VERSIONare unchanged.- Workspace contract maps moved from version 8 to 13, so the next
repowise update --workspacere-extracts them. - LiteLLM reasoning efforts come from the model's own metadata (#2240), so discovery and execution agree on which effort values a model accepts.
- Importing one server schema loads one module (#2507).
from repowise.server.schemas import Xran all 29 schema submodules; it now loads the one that definesX, taking the import from about 1.3s to 0.44s. - Several server-side folds moved into
repowise.coreas plain functions over rows, with unchanged output: the overview attention ranking (#2506), test impact assembly (#2508), decision acceptance requirements (#2510),search_codebasequery routing (#2511), knowledge silos and onboarding targets (#2517), owners, module health and reviewer suggestions (#2518), refactoring summary chips (#2519), the per-file health lead (#2494) and the workspace list reads and MCP tool payloads (#2530).
Fixed
- Security snippets never store a credential value (#2496). A value longer than the 120-character snippet cap was cut before masking, so its head was stored raw;
public_env_secretvalues were not masked at all; and any other finding on the same line showed the value unmasked. Every secret on a line is now masked in every snippet taken from it. Keyword matches skip key-name values (ACCESS_TOKEN = "access_token") and template placeholders, a vendor match suppresses the duplicate keyword finding for the same value, and publishable Stripe keys no longer fire. - A Rust
implblock has the visibility of the type it implements (#2427). The block cannot carry a modifier, so everyimplin every Rust repository was recorded as private. - A Python stdlib import no longer resolves to a local file with the same stem (#2526).
import jsonresolved to a repository'sjson.py, 527 false edges on this repository, which also fed the stats page a wrong most-imported file. - A community is never named after a third-party module (#2502). Labels come from the community's own production files.
- A contract link always names the provider in
contract_id(#2530). A link whose ends were spelled differently carried the consumer's id, and went missing from its provider's detail, tests to run and breaking-change consumers. REPOWISE_MODELis honoured in provider resolution (#2276), and theopencodeprovider's concurrency is configurable throughREPOWISE_OPENCODE_CONCURRENCYinstead of being fixed at one (#2485).- Uninstall removes what install added. The distill allow rules are removed from
~/.claude/settings.jsonand the leftover check looks for them (#2483), andrepowise hook rewrite uninstallclearsdistill.commands.enabledand removes theAGENTS.mdsection even when Codex is no longer installed (#2491). - A pathless symbol row no longer joins
get_answer's homonym union (#2391), an unknownsearch_codebasemode is reported inignored_argumentsinstead of silently coerced (#2392), and a failed head-commit read in test impact reports as degraded instead of absent (#2529). - A validation list keeps the test named for the file when it is capped (#2515),
ViewTabsscrolls only its own row instead of the page (#2498), and the health map and file drawer name theexpectedstate (#2520, #2522).
Documentation
- 0.52.0's detector count is corrected in nine places, documentation drift gets its own page and a README row, and the generated agent brief mentions it (#2480).
- The workspace documentation lists HTTP and database contract support per language (#2527, #2531).
[0.52.0] - 2026-09-20
A wide cycle. Documentation makes assertions about code, and this repository has the graph to check them, so a fourth deterministic analysis joins the pipeline: it reads the repository's own markdown, extracts the claims each document makes, resolves them against the real tree, and reports the ones the tree refutes. No model is involved, findings refresh on every update, and the reverse view answers which documents name a given file.
Savings accounting was badly wrong, and wrong in the direction of reporting far less than was actually saved. Nothing wrote pricing onto a canonical event, so a report over that ledger valued real savings at $0.00, and the readers that avoided it repriced all of history at whatever model the newest session used. Two counterfactual estimators still described a tool that no longer exists, so the tools they covered recorded themselves as a net cost on nearly every call. Each event is priced from its own evidence now, and the API, the overview and the CLI agree. The figure remains a deliberate floor: where the evidence is ambiguous it declines to claim a saving rather than estimating one, so the real number is very likely higher than the one shown.
The pages that tell you where to look were rebuilt. The overview's attention list read three detectors because those were the three that existed when it was written, while code-health findings, security, documentation drift and refactoring plans sat on a tab; they are one list now. The costs page stopped being a stack of nested cards at one visual weight. A commit sheet that could say how big a change was, but not whether it made anything worse, reports the health delta and leads with where the change lands rather than with a percentile it then explains.
Code health can see test quality for the first time — a test that checks nothing, a test that is mostly mock setup — with a per-language assertion vocabulary underneath, since the old rule counted 87 assertions across a 502-test-file Go repository. Decisions stop claiming what they cannot support: a ruling from a record nobody accepted, governance over every file in the commit a decision was mined from, transcript working agreements stored as though the codebase had decided them. And a long precision pass lands on Rust, C#, Go, Objective-C, Swift, Java and TypeScript extraction.
Added
- Documentation drift, a fourth deterministic analysis (#2290, #2307, #2313, #2356, #2370, #2378). The pass reads the repository's markdown, extracts the claims each document makes about the tree, resolves them against the graph, and files a finding for each claim the tree refutes. Findings are refreshed on every
update, not only on a clean index. They surface in the CLI, inget_health, and on a Code Health tab backed by a router, a canonical wire contract and shared view components.get_context(include=["doc_drift"])answers the reverse question — which documents name this file, where in each, and whether those documents carry drift of their own — from the references that resolved. A finding opens its own evidence: the line as the document wrote it, the heading trail locating the passage, and what the resolver checked, all of which previously arrived with the finding and was discarded. - An agent can be asked to record the choice it just committed (#2465). The layer's first write-side hook. After a successful
git commit, a PostToolUse hook offers thedecision addline and says what to put in it. Off unless the repository turns it on (decisions.capture_prompt; no preset sets it); at most once per session; only ongit commit, never a merge or a rebase; only on a message carrying two or more decision signals — measured over 137 commits, one keyword admits 58% of them and two admits 25%; and never for a commit a record already cites, with an unreadable store counting as cited so the failure mode is silence. It proposes only: nothing is recorded unless the agent runs the command, and what it records lands as a candidate for review, never as a rule. - A working agreement is told apart from a decision about the code (#2405, #2451). "Never commit to main" is a rule about how work is conducted; "
get_whynever calls an LLM" is a claim about the code that a diff can violate. Only the second can be checked against files, and the first could not be accepted at all, because acceptance requires a scope and a repo-wide rule names none. They are separate kinds now, and a measured capture tier can deliver an agreement without an acceptance. - A commit says what it did to code health (#2459, #2463).
GitCommitstores which files each commit touched, so a per-commit surface no longer shells out to git and a server with no checkout can answer at all. On top of that, the commit sheet reports the health delta a change produced — the base-versus-head comparison that previously needed both sides of every changed file and a working tree. - Test quality is measured (#2317, #2358, #2371, #2393, #2399, #2419, #2436).
assertion_free_testflags a test case whose assertion and verification counts are both zero — Meszaros's Unknown Test, and the one test-quality marker with a yes-or-no answer rather than a threshold.mock_saturated_testmeasures how much of a test is mock setup, advisory and unweighted on purpose: no defect corpus labels mock density, so weighting it would move a number that claims to predict bugs using evidence that says nothing about bugs. Java is recorded as blocked after measurement rather than guessed at. Underneath both, assertions are counted with a per-language vocabulary instead of theassert/expectprefix pair, a called helper's assertions reach the caller that delegates to it, and a test's oracle is found wherever it sits in a multi-itemwithheader. - One attention list on the overview (#2429). "Needs attention" read decision health, knowledge silos and dead code, because those were the three detectors that existed when it was written. Code-health findings, security, documentation drift and refactoring plans lived on a tab and never reached the page whose job is to say where to look.
- One composed change review (#2273, #2274, #2281). Risk scoring, the health delta, contract impact, test impact, independent grouping and branch overlap each had one owner, and every surface composed them itself and reached its own conclusions from the same facts. A versioned
ChangeReviewBundleand oneChangeReviewServicecompose them once, so a second surface cannot disagree with the first, and the expensive comparison can be handed in already computed rather than forcing a caller to choose between the shared composition and its own concurrency. It is documented for programs that want the whole review as one object, without a checkout. - Canonical savings accounting (#2248, #2401, #2440). Three surfaces recorded savings and each answered the same three questions its own way — which database to open, whether to create one that is missing, and what to do when the write fails. One canonical event per logical interaction replaces agreement by hand.
- Rust bodiless function declarations are symbols (#2382).
fn foo();parses asfunction_signature_item, andrust.scmhad no pattern for that node anywhere. Two shapes were invisible to the symbol graph: a trait's undefaulted methods — the trait's entire contract, everything an implementor must supply — and everyextern "C"declaration. - An embedding model's width can be declared without sending dimensions (#1857), and local-endpoint MCP budgets are configurable (#1856).
Changed
- A decision record is keyed, scoped and promoted on evidence (#2425, #2426, #2434, #2439, #2441). Identity keys on the record's body, file set and quote rather than on a model-generated title. A commit's file list stops doubling as the governance binding: for a record mined from a commit it is the provenance of the change, which is true, but as scope it was mostly false, since the
prandgit_archaeologyminers read one decision out of one commit body and then took that commit's entire file list. The miner is asked which files a commit's decision is actually about. The governing-scope cutoff drops from ten files to five: ten was chosen by looking at a labelled sample and then scored on that same sample, and the first out-of-sample measurement put it at 49% on topic with 26% outright noise, against the 54% baseline the scope work set out to fix. A candidate is promoted for stating a reason rather than for being seen twice. - Acceptance requires a reason, and records what signed it (#2279, #2445, #2457, #2458). The contract read the reason as
rationaleand fell through todecision, so a record whose only content restated its own title passed the "no rationale" check — the what stood in for the why. Separately, the identity check was satisfied by any non-blank string and the accepter resolver fell through togit config user.name, so an agent accepting a decision signed the maintainer's name into the audit log and nothing downstream could tell the two apart.CLAUDE.mdandAGENTS.mdnow say when a standing decision was not signed by a person, since an agent reads that block as instructions. - Session transcript mining is off unless asked for (#2286, #2387, #2395). A decision mined from a transcript took its scope from whatever files the surrounding tool calls named, and a transcript names plenty that is not this codebase — scratch scripts, plan documents, files belonging to a checkout sitting beside this one. Over staged candidates on a real machine only 24.7% of the bound paths were files the index holds, and 64.5% of rows carried a list in which nothing resolved at all. Codex transcripts are scoped to the repository that owns them; both raw-line gates had dropped the
session_metaline carrying the working directory, so every Codex event read as having no opinion rather than as a miss. - Decision confidence accounts for what a record states (#2328). Confidence read source rank, corroboration and verification but never whether the record said anything, so the ranking ran backwards: records filling one body field averaged 0.845 against 0.776 for records filling five, with 65.2% of ordered pairs inverted. Scaling the rank-earned credit by how much an extracted record states takes the inversion rate to 12.9% and makes the table monotonic.
get_change_riskreports what a change wrote, not what it weighs (#2452). A census of 135 recorded calls from Claude Code and Codex, 118 of them real engineering use, found the tool spending its payload on signals no agent acted on and burying the ones they did.- The commit sheet leads with where a change lands (#2453, #2455). It opened on a percentile and then spent its body explaining the number it had just printed — three of its four blocks were commentary about the model, and it never named a file. The queue's chips filter the repository now instead of the loaded page: risk order returns the top tercile first, so all 50 loaded rows were already "high", the chips read "All (50), High priority (50)", and clicking High priority removed nothing. On this repository the real figures are 911 / 302 / 458.
- The costs page is an evidence-led scan surface (#2432). It was a stack of nested cards at one visual weight throughout. A card should mean "an object you can act on", and a savings total is not that, so the boxes carried no information while removing the page's hierarchy.
- MCP responses stop repeating themselves (#2272, #2284, #2456). The canonical index scope is about 900 characters and rode on every response — roughly a third of a small
get_symbolreply, spent answering a question almost no call asks. A diagnostic that ships with every answer is a tax rather than a disclosure, so a digest rides instead and the full scope is one call away. It is also parsed once per index state rather than re-readingstate.jsonand re-parsingconfig.yamlon every call. Four further fields, each rebuildable from the row it shipped beside or from the key it sat under, are gone. timing_msis reported for every tool (#2271), rather than for the ones someone had remembered to thread it through.- History-based health gates are ranked against the repository (#2437, #2438). Six months is a fixed window, so a repository landing many commits a day accumulates more fix commits per file inside it than a quiet one; with an entry gate of one fix,
prior_defectfired on 43.5% of the files here, and a signal that flags nearly half a tree is not locating anything. Co-change breadth is measured before truncation rather than by counting a list capped at 25 partners, where above the cap every file looks identical. - An agent is defined once (#2394). The same agent was spelled four ways across four hand-maintained namespaces, and the savings domain had added a fifth and a sixth that agreed with the target registry only by hand.
- Plugin: the
decisioncommand says that acceptance now requires a stated reason, documents--kind agreementand--evidence-commit, and stops claiming a fixed capture-source count.
Fixed
get_whystops claiming a ruling the store does not hold (#2280, #2282, #2322). A record is a candidate until aDecisionAcceptancerow exists, and accepting is a deliberate manual step, so a store of nothing but candidates is the state of every repository whose maintainer has not worked through the acceptance UI. Separately, neither path-mode call shape declared its lanes as requested, so the shared response budget shed them first and a path query dropped its origin story whole. The health mode names the active, deprecated, superseded, dismissed and unscoped lanes it previously reported as a bare count.- A hand-authored decision reaches the surfaces that deliver it (#2464).
decision addand the HTTP create route wroteaffected_files_jsonand stopped, while the edit-time injection hook,get_governing_decisionsand theget_riskdirectives all readdecision_node_links, which only the mining pipeline ever wrote. A record added by hand had files and no links, so nothing could deliver it. get_riskstops answering an unresolved target with a confident zero (#2321). Amodule:id, a bare directory, an unindexed path and a typo all fell through to the same card, every numeric field of which was a structural zero. Measured here, a subsystem directory returneddependents_count: 0while a single file inside it returned 59, so a subsystem nothing could safely be changed in read as safe to change. The directory is the dangerous form, because it looks like an ordinary target.get_context'sused_bykeys on the symbol, not on its file (#2329). The query joined the edge target against the symbol's file path, but every symbol-level edge is emitted symbol-to-symbol with apath::Nametarget, so calls, references, extends, implements, method_implements, dispatches_to, framework_binds and reads could none of them ever match. The field answered "who imports this symbol's file" under a name that promises otherwise.get_symbol(depth=3)serves the hops it found (#2285), and a requested projection is kept ahead of the ones a caller did not ask for (#2270) —get_overview(include=["outline"])returned 0 of 40 outline sections while leaving 73% of its budget unspent.- A
.tsxfile is read with the JSX grammar (#2398). It arrives taggedtypescript, and that grammar fails on the first<Component />; everything after it landed in error recovery, so the health pass was reading a tree that does not describe the file. Assertions after the first element were never counted, the clone detector hashed a token stream recovery had partly invented, and the dataflow CFG was built over the same wreckage. - A file's same-named functions survive to the markers that score them (#2408). Eleven scoring markers read a per-file map keyed by function name, and a function name is not unique inside a file — every
it(...)callback in a spec file walks under the nameit callback, and__init__walks under one name on each class in a module. - Rust graph precision (#1568, #2362, #2376, #2386, #2409, #2412, #2415, #2418, #2423). Field and type-argument types are recorded as type references, so a struct used only as a field type or nested in
Option<T>no longer reads as an unused export; Rust reaches an intra-crate field type without ausestatement, so nothing else recorded the reference. Apub trait's items take the visibility the trait declares instead ofprivate, which had left downstream-callable API sitting in the narrow dead-code pool. A trait's defaulted methods are parented to the trait rather than emitted parentless as plain functions.type_usecaptures widen to wrapper shapes, a scoped call binds to its own type rather than the file's last, a re-export through a crate root carries the rest of the path, a trait bound is no longer read as inheritance, a std-library name is refused on theimport_mergedtier, and impl blocks are treated as uncallable and non-importable. - C#, Go, Objective-C, Swift, Java, GDScript and TypeScript extraction (#2309, #2312, #2316, #2318, #2323, #2324, #2330, #2336, #2361, #2363, #2364, #2402, #2414, #2428). C# null-conditional call sites (
a?.Foo()) are captured — they were not resolved-and-wrong and not unresolved either, but never captured, so nothing downstream knew they existed — and an extension-method call binds to the type it extends, withbuiltin_parentsaligned to the shared .NET BCL set. Go functions in composite literals mint reference edges. Objective-C.hand.mfiles are paired, and a call-shaped attribute macro trailing a declaration is blanked rather than left to wreck error recovery past the terminating semicolon. A Swift qualified extension no longer names its qualifier as a conformer in its own right. A bare@Immutablestops fabricating a generated Java class. TypeScript class fields holding a function are indexed. GDScriptuid://references resolve through the.uidsidecar, and a Godot scene connection resolves to a C# script method. A nested sub-package resolves from its own manifest, an inheritance edge anchors to the child's real symbol id, and inertreadsmembership is retired fromSYMBOL_USE_EDGE_TYPES. - Savings are priced from their own evidence (#2413, #2461, #2468). Nothing wrote pricing onto a canonical event, so any report over the canonical ledger would have valued real savings at $0.00; the legacy readers avoided that by resolving the coding agent's model at read time and repricing all of history at it, which moved past figures whenever the newest session changed model. Rates are no longer fabricated and stamped as measured, evidence on disk is no longer thrown away, and two counterfactual estimators describing a tool that no longer exists had been recording the tools they cover as a net cost on nearly every call. The API, the overview and the CLI report the same numbers. The reported figure stays a deliberate floor: where the evidence is ambiguous the ledger declines to claim a saving rather than estimating one, so the real figure is very likely higher than the one shown.
distillruns a wrapped command in the shell it was written for (#2448)._POSIX_HOST = os.name == "posix"gated both the metacharacter bail and safe-chain recognition, so on a Windows hostcd X && git difffrom an agent's Bash tool was refused before classification ever ran. Measured over 30 days on one machine: 2,203 commands bypassed distillation, worth 738,123 tokens, and that one gate explains 97.5% of them.- Secret values are masked in security finding snippets (#2375),
provider_configis stored0600, and keys are redacted in the validate response (#1798). - CORS is browser-compliant and configurable by environment (#1793), and repository paths are passed to local CLI providers (#2242).
- A configured database is treated as a store by
update(#2404) anddelete(#2338). Both guarded on the repo-localwiki.dbfile before building an engine fromresolve_db_url; under a shared database that file is absent by design, so the guard returned before the engine was ever built. init --forcein a worktree is no longer a paid upgrade (#2339), MCP tool updates route to the workspace repository database (#2241), and a missing workspace repository can be removed from the UI (#2253).- A whole prompt is recorded as input tokens (#2337). Claude Code splits one prompt across three disjoint usage fields, and only
input_tokens— the uncached remainder — was persisted, so a roughly 20k-token module page recordedinput_tokens=2andstatussummed a whole wiki to almost nothing. - DeepSeek and Kimi are priced (#1790) rather than $0 in both the estimator and the live tracker, and
codex_clisubprocess concurrency is configurable (#2334). - The root layout's repository list is not prerendered (#2260, #2261).
/settingsandnot-foundare static routes, sonext buildprerendered them where no API answers and baked "can't reach the API" into their payloads; since layouts do not re-render on soft navigation, the sidebar then stayed wrong on a healthy server. total_pagesis zero when a phase reports no total (#2308), a formatter comment stops prescribingruff format(#2444), the no-index MCP guidance names the capability rather than one host's tool names (#2326), and chat stops suggesting questions no tool can answer (#2315).- Decision plumbing (#2310, #2314, #2357, #2421, #2422, #2447, #2450). Paraphrases are deduplicated at the pass that creates the record, and stored duplicates are folded in one sweep.
last_code_changeis filled from the dates staleness already reads, having held a value on 0 of 357 records. Staged decision paths are stored in a form the index can resolve. A bundled claim is flagged on every capture path rather than on one lane. A starved token budget no longer silently loses whole batches, and one wordy candidate no longer silences its whole lane. - The contradiction rate counts rows, not decisions (#2479).
hook statsreported 181 followed against 20 contradicted, a 10.0% rate, where all 20 rows were one decision: the verdict was computed per decision and stored per row, so a single match settled every row that decision had ever been shown in. The true rate on that corpus is 0 of 201. Alongside it, a contradiction now requires two-sided evidence rather than a reversal keyword found in either text alone, opposing verbs match as whole words rather than inside longer ones (usewas matching insideuser), and injection rows resolve throughdecision_aliasesso a re-keyed record no longer reads as gone and drains its rows. - A failed HEAD read no longer strands a store's sync pointer (#2446), a churn rank is published beside the raw churn sum on direct risks (#2443), a zero-impact finding is visible to a surface that ranks nothing (#2430), a shared
warn()helper unifies the warning stream (#1843), and the credential-masking migration has its own revision (#2462). - An MCP session's end is named (#2283) instead of arriving as EOF, a broken pipe or a cancelled task that were all indistinguishable from a crash — the hang-up escaped as a traceback the host would respawn on, and a clean end produced no line at all.
Documentation
- The change-review engine is documented for programs that compose it directly (#2281). Contributor guidance makes a claim binding from the moment the comment is posted rather than from assignment, which can lag it by days (#2377), and corrects stale setup steps. The Rust
unused_internalskip is given its real reason (#2390): the stated one was that the graph emits no intra-file Rust call edges, and it does — same-file calls are the largest single resolution origin in every Rust repository measured, and bare same-file call sites resolve at 97.5% or better.
[0.51.0] - 2026-09-16
An index makes claims about a repository, and several of them were not being checked. A fast index and a full one described their scope differently in every place that reported it, and a configured cap could be read as achieved coverage. Git history sampling let a file's retained history depend on which of two internal lanes happened to serve it. Pages that fell outside a generation selection could be marked stale and stay stale through every later update, while doctor blamed the wrong cause and named a command that could not clear it. An update could call a repository current because its commit had not moved, even when the configuration deciding what to index had changed underneath it. This release makes each of those say what is actually true, and records the coverage a run achieved rather than the coverage it was allowed. Next to that, contract evidence stops depending on whether a handler signature can be recovered from code: a checked-in OpenAPI document is now a provider schema, compared directionally, and local Maven coordinates resolve across a workspace. COBOL is a parsed language.
Added
- COBOL is a parsed language (#1770).
.cbl,.cob,.coboland.cpyreach the Good tier: program IDs, sections, paragraphs and data-level symbols are extracted, and literalCALLand localPERFORMtargets resolve. Programs entered externally by JCL or a scheduler no longer produce unsound unreachable-file findings. DynamicCALL data-item, copybook resolution and dialect-specific source formats are deliberately outside this tier and stay silent rather than guess. A 30-file COBOL repository indexed in fast mode read 1,348 symbols and resolved 86 of 88 extracted calls in 6.8 seconds. - A checked-in OpenAPI document is a provider schema (#2227). Previously a provider's wire shape came only from code, so a repository could describe its API precisely in
openapi.yamland still produce no request or response evidence. The extractor joins the existing per-repository walk rather than adding a second one, and an operation emits the samehttp::<METHOD>::<normalized-path>identity as a framework route, so it flows through the matching, diagnostics and persistence already in place. Candidate filenames are narrow on purpose, and when exactly one code provider matches the same identity, the code-backed row keeps its file and symbol. - OpenAPI request and response schemas are compared directionally (#2229). Request rules follow what a provider will accept; response rules follow what a consumer may receive; the comparison recurses through object properties and array items for type, requiredness, nullability and enum changes. Comparison runs only when both sides are complete and share an extraction source and comparison key — an incomplete extraction or a source transition produces a stated uncertainty rather than a confident claim that a field was removed.
- Local Maven package dependencies resolve across a workspace (#2231). A bounded, filesystem-only effective model reads local parents, reactor modules, properties, dependency management and direct dependencies, then resolves unique coordinates between repositories, carrying manifest, coordinate, version, scope and resolution evidence through the system graph, REST, MCP and the UI. No Maven subprocess, network resolution, transitive solving, BOM import or profile activation is involved; an ambiguous or invalid model is refused with a diagnostic instead of being half-resolved.
index_scope, one persisted contract for what an index actually covers (#2246). Run mode and content provenance, the Git tier and both the configured commit cap and the coverage achieved against it, configured and effective file-page caps with eligible / generated / omitted counts, analysis that was unavailable as distinct from analysis intentionally skipped, provider and model details, full-text and semantic search availability with the exact command that restores them, and the state of any pending or resumable upgrade. CLI status,/api/repos, the generated AGENTS/CLAUDE guidance and MCP metadata all read that one projection. Older state degrades missing evidence tounknownrather than inferring completed coverage from configuration. The default commit limit is unchanged.- Onboarding pages are grounded in the repository's own setup evidence (#2228). Setup guidance is drawn from contributor documents and package manifests, including working directories and exact script invocations, and concept and execution-flow selection now prefers representative lifecycle entry points and subsystem breadth over graph-central internals. Vocabulary corroboration is separated from definition quality, so incidental prose is no longer published as a definition, and generated file references are canonicalized into repository-relative links that resolve to a file page.
Changed
- Git history sampling is per-file and monotonic (#2244). For every eligible rename-free file the sampler retains the newest
min(per_file_limit, available non-merge history)commits, satisfied through three bounded lanes — the recent repository-wide window, serial deep repository-wide chunks that top up underfilled files without retaining unrelated histories, and a per-file walk only for what remains. Retained history no longer depends on which lane served a file, Git concurrency is bounded by the machine's actual resources rather than a fixed number, and the coverage a run achieved is recorded. The default commit limit is not raised and a bounded index is not reported as complete repository history. - Change analysis is separated from git and from transport (#2263). Scoring composition, independent-change partitioning, review verdicts and symbol-contract classification each have one owner and can be answered without a checkout or a rendering layer; a review decision is carried as data and rendered at the edge. No current surface changes what it returns.
Fixed
- Stale-page recovery converges (#2145, #2254).
updatemarked file pages stale on every commit but never the module, SCC or repository-overview pages above them, so model-written summaries drifted from HEAD withdoctor,get_stale_pagesandgenerate --staleall reporting clean. Those dependants are now derived from the same scoped cascade duringupdate, seeded so they are marked without being regenerated, which keeps sync free of model spend. A successful full generation is authoritative and retires pages outside the current selection, cleaning their vector and full-text entries; full-text cleanup debt is persisted so a failed deletion converges on retry;doctornow classifies who owns a stale page and namesupdate --fullas the reconciliation path; and a page regenerated by an update is no longer marked stale again by that same update. updatereconciles configuration changes, not just new commits (#2235). A repository whose commit had not moved was treated as current even when indexing-relevant configuration had changed, leaving clean updates out of step with a fresh index. Configuration is now fingerprinted by the layer it affects — traversal, Git history, health, generation or state alone — and an update rebuilds that layer and its dependants while leaving the rest untouched. Unknown keys take the conservative full-invalidation path, a failed required rebuild keeps the prior fingerprint so the next run retries rather than accepting partial state, and search or vector cleanup that fails after the database transaction is recorded and retried later.- A failed embedding batch no longer multiplies across the corpus (#2251). Reindex retried whole batches blindly, so a provider outage could run for an hour. Failed chunk identities and the stage they failed at are preserved; individual items are retried only for input-specific 400/413/422 failures; two consecutive provider-wide batch failures stop the run; and only a confirmed single-input token-limit failure is adaptively truncated, with a bounded retry count. A restricted-network run now stops in seconds with a clear connection error. Reindex batch size is validated and the underlying failure cause is surfaced.
- Symbol complexity computed during health analysis is persisted (#2236), so clean, incremental, resumed and workspace indexing converge on the same symbol IDs and the same stored complexity.
- Workspace members register and search fans out on a shared database (#2138, #2195). When a workspace runs against a configured database URL, per-repo indexes live in that database and no
.repowise/wiki.dbfile exists — server startup and search fan-out both skipped every member on that basis, so semantic search and full-text search could not reach any non-primary repository. - A language query that fails to compile now warns (#2166, #2177). Compilation failures were silent, so a whole language could quietly extract nothing. Unaffected languages continue indexing normally.
- The zero-result search rescue stops riding on one common word (#2092, #2214). The rescue's full-text fallback returned the first row pointing at a code page with no quality check, so a multi-token query where only one ordinary token matched produced a confident suggestion for an unrelated file. A row must now cover at least two distinct query tokens across its title, path and snippet; single-token queries still fire, since the rescue is meant to allow generic names.
editor_files.<id>: falseopt-outs are honored duringinitandupdate(#2137, #2226),vscode_mcp: falseincluded. An explicit per-run override still wins over a persisted opt-out, and Codex MCP and hook registration stay independent of theagents_mdopt-out.- Windows users are warned to stop running server and MCP processes before upgrading (#2149, #2204), in both
doctorand the shared update advisory. The note also clarifies thatRepowise: Stop Serveronly stops a server the VS Code extension started. - Saving or deleting a provider API key no longer reports a false failure (#2237). The endpoints answer
204 No Contentand the shared API client tried to parse every success as JSON, so a completed mutation surfaced asUnexpected end of JSON inputand skipped its refresh. - DeepSeek reasoning controls match the documented contract (#2239).
reasoning_effortis sent at the top level rather than nested in the thinking payload, low effort no longer maps to high, reasoning support is restricted to model IDs that actually have it, and thinking is disabled for streaming chat because the conversation history cannot replay the reasoning content a later tool call would need. - Kimi K2 and K3 advertise the controls they have (#2238). K2 exposes only the disable-thinking control instead of graded choices that all collapsed to the same toggle, K3 exposes and forwards its
low,highandmaxeffort levels, andautoleaves the provider default alone. - Generation progress cannot read above 100% (#2176, #2224). A job reporting more completed pages than a stale total rendered
4820%, passed an invalid value to the progress primitive and droppedaria-valuenow. The shared primitive now bounds every finite input; non-finite input keeps the indeterminate state. - A select inside a dialog opens in front of it (#2172, #2173). Floating layers are portalled to the document body, so DOM nesting does not settle paint order and the raw z-index did — a dropdown opened behind the dialog overlay while still holding keyboard focus, so typeahead changed a value nobody could see. The
--z-*scale is documented and guarded by an ordering test. - A long AI fix prompt stays inside its modal (#2249). Unbroken paths and generated identifiers contributed their minimum content width to the dialog and could push the footer, and the Copy prompt action with it, past the right edge.
Documentation
- The supported COBOL surface and its deferred ceilings are documented (#1770), as is the supported OpenAPI subset, its static evidence boundary and the endpoint-exposure vocabulary (#2229).
[0.50.0] - 2026-09-11
Code health had six different vocabularies for one number, and roughly half of what it measured was change history, which rises as a file is worked on. A week of refactoring could drop the headline and print "Declining health" in error red at the person who did the work. This release rebuilds that reading end to end: one absolute band scale wherever a score appears, one figure at the top of the page, a control that says whether you are counting code shape alone or code shape plus history, and a decline that names which half moved and whether an edit can settle it. No score moves. What changes is what the product tells you a score means. Next to it, findings stop being a list you scroll and become a queue you work, with a disposition you record once and never have re-proposed at you. The other change you will notice first is chat: it is no longer a widget in a corner, it is reachable from the row you are already looking at, pointed at that row, with its first answer grounded in the page you asked from.
Added
- One health vocabulary, five absolute bands (#2180). Excellent 8.5+, Good 7.0 to 8.5, Fair 5.5 to 7.0, Needs work 4.0 to 5.5, At risk below 4.0, defined once and locked by a parity test across Python and TypeScript. Absolute rather than percentile, so a score means the same thing behind a firewall as it does against a public corpus. Previously the same file could read amber on a treemap tile and green in the tooltip attached to it. 4.0 keeps its long-standing meaning, so no repository's most severe classification changes.
- A Counts control on the health page (#2164). Everything is the calibrated score; Code shape only subtracts the git-derived half (churn, co-change, ownership, prior fixes), which reports what a repository has been through rather than what its code is like. On this repo that reads 7.1 against 8.5. It is a projection over numbers already stored per file, not a second scoring pass.
scope=all|productionon the health overview, files, map, trend and work-queue routes, onget_health, and onrepowise health(#2154). A narrowed response drops what it cannot honestly narrow rather than relabelling it.- The findings tab is one triage list that pages (#2200). Files ranked by leverage, expandable to their findings, server-paged with search, dimension, severity, status and hotspot / untested / below-green filters. The old queue fetched 200 rows, raised that to 500 once, and had no way to reach anything past it.
set_finding_status, the write half of findings triage (#2043, closes #1535). Record a disposition on a refactoring plan and it survives every later analysis run:false_positiveis never re-emitted,acknowledgedstays visible but stops counting as unheard,resolvedcloses it out. Opt-in, since it mutates.- Ask about this, from the row you are looking at (#2190). One control on the file page header, symbol page, health finding rows, risk table rows, decision detail, commit detail and the documentation header. It opens the composer already scoped to that thing and sends.
- Chat grounds its first answer in the page you asked from (#2188). A question from a file view no longer costs a full tool round trip before the model can cite the file, and a page whose subject is a symbol, decision or commit is no longer invisible to it. The transcript shows what was read.
- Composer suggestions built from what the page measured (#2205). "Explain the 28 bug fixes in incremental.py" rather than "Explain this file's responsibility", the top health finding and the function it sits in, an owner and their measured share. Follow-ups read the tools the finished turn called and propose the next read worth making. No model call is involved.
- Co-change rows say which side leads (#2202).
get_riskandget_change_riskhardcoded"direction": "undirected"while already holding the two numbers that decide it. A pair now readsa_to_borb_to_aby which file seldom moves without the other, with both confidences alongside. An exact tie stays undirected rather than having a lead broken arbitrarily. - The post-commit hook is installed by default (#2143), with an opt-out on the spot and
repowise hook uninstallnamed in the line that reports it. Auto-sync after each commit is what keeps an index from going stale between the day you set it up and the day you ask it something.--no-hookskips it, and--no-editor-setupkeeps it off the way it keeps every other out-of-.repowisewrite off.doctorgains a row for it. - Complexity coverage for F# and Objective-C (#2152, #2225). Both reach the health complexity walker now, so their files carry real cyclomatic complexity, nesting and maintainability instead of metrics computed against an empty node set. Elixir is deliberately left uncovered: its grammar gives
defmodule,def,defpand an ordinary call the same node, told apart only by the target's identifier text, which a map of node types cannot express, so a map there reports the module as the file's only function and hides every real one. No rows is the honest answer until the walker can take a text predicate.
Changed
HEALTH_ANALYZER_VERSIONis 11 (#2153, #2225). Markdown, JSON, YAML and TOML rows carried a mechanical 10.0 meaning "nothing looked at this", so a repository could raise its score by adding documentation; they are no longer scored, and the checks that cannot apply to a test file skip them. Every file now storesstructure_deductionandhistory_deduction, which is what makes the split reading above a subtraction rather than a guess. Existing indexes re-score on their next update. The stamp also carries the F# and Objective-C complexity maps below, which change stored complexity for files in those languages.STORE_FORMAT_VERSIONandPARSER_SCHEMA_VERSIONare unchanged, so nothing needs re-indexing.- A health decline names its cause (#2171). When the composite fell, history is the driver, and history is the only half that moved down, the trend says so and takes the neutral treatment history findings already have rather than error red: "Code health dropped 0.60 points (7.60 to 7.00), and code shape improved 0.30 over the same window. Change history is the only half that moved down, and no edit to these files settles it."
- The health page leads with one score (#2161, #2164). It previously led with Code health and Maintainability side by side, both out of ten and both largely about code shape, so the first screen answered "which number do I steer by?" twice. Maintainability moves to the ribbon beside the other cuts of the same scoring.
- Full-text search weights the path and title columns (#2191).
bm25()ran with uniform weights, so a page whose path is the answer ranked no higher than one mentioning the word once in passing. Measured on a 4,630-page index over 14 path-naming queries with 18 held-out prose queries as a control: MRR 0.643 to 0.708, top-1 hits 7 to 8, no control query losing a good result. Query-time only, no reindex. - MCP responses size themselves by content class (#2179). The budget converted characters to tokens at a flat 4 chars/token; measured against the real tokenizer over 24 recorded responses the rate runs 2.5 to 5.4 and the flat figure read low on 18 of them, which is the direction that costs a caller a read. Responses also stamp what the budget dropped.
- Repeated searches skip the embedding provider (#2189, #2192). Query embeddings are cached on both the MCP and HTTP paths.
- A leaf workspace import stops loading the analysis stack (#2178).
repowise.core.workspace/__init__eagerly imported all nine submodules, soimport repowise.core.workspace.config, the first thingrepowise mcptouches and a leaf whose only dependency is yaml, was charged sqlalchemy, networkx and the ingestion graph. repowise update --fullshows live progress and persists as it goes (#1709, #2038). It passedprogress=Noneinto generation, so a long run showed a dead screen, and buffered every page in memory to write once at the end, so a Ctrl-C at hour eight cost the whole run. Pages are now written the instant each completes.- The tree-sitter floor is 0.25, and CI compiles every language query (#2165, #2170).
Fixed
- Indexing uses the provider you picked, not a guess: the dashboard's choice reaches the indexer (#2198), and a workspace index phase keeps the provider it was launched with (#2197).
- Two silent embedder degradations now surface (#1370, #2107): a malformed
REPOWISE_FULL_RESCORE_INTERVAL_DAYSwarned nobody and defaulted to 7.0, and a vector-store build failure was swallowed by a bare except. Generation-phase embed failures also reachinit's degraded list (#1369, #2108). - Workspace mode stops dropping the repo you pointed at when it is a submodule (#2196). The submodule flag governs what a scan discovers; the root is not a discovery, it is the path you named.
- e2e spec files are indexed (#1842). They were dropped from traversal, so they never got a wiki page and
get_answercould not answer a test-infrastructure question, with no signal that a whole area was missing. - Credential findings gain a value gate, a placeholder filter and path awareness (#2118, #2121), which is what separates a real secret from a placeholder in an example.
repowise init --resumeskips the interactive questionnaire (#2098, #2106).list_reposreads index presence from the database rather than inferring it (#2187), and the primary vector store resolves for PostgreSQL databases (#2140, #2144).- Blocking git inspection is offloaded off the server's event loop (#2103), and the cost-tracker engine is disposed so a later loop does not reuse a dead connection (#2083).
doctorreports actual LLM provider availability (#2086) rather than a config file's opinion of it.- Boolean config values map to valid reasoning modes (#1990).
- The web app loads the active provider from the providers API instead of the health probe, and renders the provider picker from that same response rather than a hardcoded array (#2155, #2203).
codex_cliandopenrouterresolve on the server and could not be picked in settings. - Heading vocabulary harvesting skips fenced code blocks (#2150).
git_function_blame.symbol_idis widened toTexton PostgreSQL (#2169).DOC_EXTENSIONShad drifted into several copies; they are consolidated behind a drift guard (#1489, #2159).regeneratereceivesrepo_id, so Write-with-AI works in workspace mode (#1809).
Documentation
- Native VS Code Language Model tools and the
agentToolssetting are documented (#2096). - The quickstart notes that
--no-editor-setupcovers the repo-root.mcp.jsontoo (#2199). - The health docs say what the score counts, and the agent line no longer reads inverted (#2201).
Dependencies
- Dependabot updates are grouped, and
gitpythonis bumped (#2206).
[0.49.0] - 2026-09-05
The largest language release so far: Elixir, F#, Objective-C, VB.NET and GDScript all move to the Good tier with real ASTs behind them, QML arrives at the import tier, and Razor/Blazor markup enters the graph. Alongside that, the Architecture Map stops being a picture and becomes somewhere you can drill into, update learns to size itself by what actually changed and reports which stage took the time, and the MCP tools stop returning bare zeros: every empty answer now says what it was measured against. Workspace mode gains cross-repo test impact and consumer contracts for five more languages.
Added
- Five languages reach the Good tier. Elixir (#2127), F# (#2133) and Objective-C (#2135) each get a full AST parse rather than a regex skim. VB.NET arrives at Good with its own published grammar (#2048), and GDScript joins it (#1620). Symbol extraction and same-file call resolution work in all five.
- QML at the lightweight import tier (#2039), and Razor/Blazor markup (
.razor,.cshtml) is indexed into the graph (#1962). - Flutter widget trees are edges.
build()bodies emit widget-tree edges, so a widget's children are reachable from the graph rather than buried in a method body (#2049). - Godot scripts are reachable through the project. Scenes, autoloads and
class_namedeclarations connect scripts that no import statement ever linked (#1624). - The Architecture Map is a place you can drill into (#2104), with the map chrome moved off the canvas and given a ground to sit on (#2102). Communities count production files, read conductance, and name the group that was previously just "ungrouped" (#2112).
- File pages render the history they already carried. The data was indexed and never shown (#2082).
- Cross-repo test impact. Workspace mode works out which tests in other repos a change can break (#2008), and surfaces it on
get_change_risk,get_blast_radius, the API and the contract page (#2123). - HTTP consumer contracts for Go, Ruby, Java, Kotlin and PHP (#2114), provider contracts for Flask, Micronaut and Remix (#2115), and breaking changes on the contracts page behind a contract drawer (#2116).
- Branch overlap on
get_risk: the other open branches editing the files you are editing (#2131). get_change_risksays when a diff is several changes the index cannot connect to each other (#2130).- Conventions the import graph proves are proposed as decisions, rather than waiting for someone to write them down (#2141).
- Unresolved Python imports become external nodes.
resolve_python_importreturnedNonefor any absolute import no repo file defines, so a Python file's third-party and stdlib dependencies never becameexternal:nodes or import edges, the way every other language resolver already registered them (#2136). - A C complexity node map.
LANGUAGE_NODE_MAPShad nocentry, so C files got no complexity or maintainability signal at all (#2113).
Changed
HEALTH_ANALYZER_VERSIONis 9. The C node map above changesmax_ccnand the maintainability index for every C file, which were previously computed against an empty node set and persisted that way. A repo containing C re-scores on its next update; a repo without C is unaffected, and neither is asked to re-index.PARSER_SCHEMA_VERSIONis 3, so the Python external-node change reconciles every file's edges once instead of leaving an updated index with new edges on changed files only (#2136). The cost is one reparse per existing index.STORE_FORMAT_VERSIONis unchanged.updatesizes itself by what changed. A one-file commit no longer pays for a whole-repo shaped update (#2134).updaterecords a phase timing row per stage, the wayinithas since 0.47.0, so a slow update can say which stage was slow (#2128).- MCP responses account for themselves. Every zero carries the basis it was measured against, scope hints say what was left out, a completeness line states whether a body was served whole, and no source is served twice in one response (#2125). Tool telemetry carries response size, execution flows carry freshness, and a stale index with a low-confidence answer points at
update(#2122). get_answer,get_contextandget_symbolresponses are documented to match the envelope they actually return (#2126).
Fixed
- Five places the incremental path knew less than a full index (#2129), and stale
file_pagerows no longer surviverepowise update(#1744, #1906). - Page regeneration prioritizes the oldest stale pages when the budget is constrained, instead of whatever came first (#1845), and
updatewarns when the cascade budget truncated regeneration rather than silently skipping pages (#1720). doctorstops reporting excluded files as missing (#2085).- The MCP server drops to single-repo mode for a nested indexed repo that is not a workspace member (#2088).
repowise hookresolves the hooks directory git actually runs, including a husky-managed one, rather than assuming.git/hooks(#1511, #1512).- Graph hooks are declared above the loading early-return, which is a rule React does not let you break (#2111).
- Symbol-only headings no longer crash
init(#2101). - Boolean defaults compile per dialect during persistence (#2087).
REPOWISE_EMBEDDING_DIMSis validated where it is parsed, not where it fails (#1979).ingested_commit_shafor coverage is stamped from the live HEAD rather than a stale stored column (#1800).- The
temporal_hotspotrisk-semantics contract is corrected to unbounded (#1971).
Documentation
LANGUAGE_SUPPORT.mdkeeps to the support ladder; the parsing mechanics move to the architecture note (#2124).
Dependencies
- Adds
tree-sitter-gdscript,tree-sitter-vb-dotnet,tree-sitter-elixir,tree-sitter-fsharpandtree-sitter-objc. - Bumps
fast-urifrom 3.1.5 to 3.1.7 (#2081).
[0.48.0] — 2026-09-02
Architectural decisions stop being a passive record and become a review queue. A decision is captured under one policy, keyed by the identity it dedupes on, and acceptance is the only thing that grants it authority — recorded as an event, on a tracked manifest, so every read can say who vouched for what. On top of that sits a queue that leads with the work you can actually finish. The other change you will see first is the repo overview: the front page described a file tree because a file tree was all it was given, and it now reads like a description of a product. Elsewhere: doctor stops trusting config files and launches the MCP server for real, the keyless and degraded paths report why they degraded, and a cycle of persistence and generation performance work.
Added
- Decisions are a review queue, not a log. Capture runs under one policy with acceptance as the only authority (#2067), acceptance is a recorded event against a tracked manifest (#2069), and discovery is one broad grounded call per update instead of many narrow ones (#2068). Reads carry review lanes and a single vocabulary, with authority stated on every read (#2070). A decision's id is derived from the identity it dedupes on, so the same decision found twice is the same row (#2071).
decisionreports what capture actually did and reviews a backlog in one command (#2072), and the queue leads with work that can be finished now (#2073). - The repo overview reads like a product, not a directory. The front page every reader opens described a file tree:
assemble_repo_overviewhanded the model pagerank, communities, package file counts and entry-point paths with no natural-language input at all, so a page of paths was the specified output. The context gains one bounded prose keyhole — a README digest capped at 2,500 characters — and the page has a fixed shape: what this is, how the parts fit, then five to seven named concepts. Prose supplies vocabulary and framing; structure stays the sole authority on paths, counts and package names (#2078). doctorlaunches the MCP server rather than reading its config and assuming (#2065). A config that looks right and a server that will not start are different problems, and only one of them was previously detectable.- Failure and degradation are reportable.
initfailures carry a reason code (#2057), and telemetry reports the degrade reason, the keyless case and every crash leaf (#2056). - Five more JS/TS security patterns in the registry (#2047).
augmenttakes--verbose, wired throughconfigure_cli_logging(#2033).- The TypeScript HTTP contract is generated from the FastAPI schema rather than hand-maintained (#2035).
tech_stackdetects bun as the script runner (#1940).
Changed
HEALTH_ANALYZER_VERSIONis 8. Paired-test detection hadtest_<stem>.pyhardcoded, so the prefix layout only ever matched Python; it now follows the file's own suffix, and<stem>_specjoins the suffix forms (#1841). Files wrongly counted untested become tested, which moves untested-hotspot findings and the scores carrying them. An existing index re-scores on its next update; it does not force a re-index.STORE_FORMAT_VERSIONandPARSER_SCHEMA_VERSIONare unchanged.- Indexing and persistence are faster. Betweenness is reused inside a churn budget instead of rescored (#2066), the full-text index is written in one transaction (#2052), job checkpoints are buffered behind a bounded atomic flush (#2053), and the generation and persistence tail is timed (#2050).
- Repository-scale native memory is bounded during indexing (#1778).
- Every MCP tool shares one response-budget contract (#2051).
get_answergrades a degraded answer's confidence from retrieval quality instead of pinning it low (#2055).init --resumeis documented, with expanded help text (#113, #2030).
Fixed
- Dry runs are read-only.
update --full --dry-runcould create locks and mutate the index; it now reports the persisted upgrade plan instead (#2005).init --dry-runhad the same shape and was worse: only the branch that prices a model honoured the flag, so a keyless or--no-prosedry run rendered the whole template wiki and persisted it. Against a repo that already had a model-written wiki that rewrote every page from templates and downgradeddocs_mode, on the one command whose promise is that it does not act (#2075). It also no longer passes an LLM client, so a dry run cannot bill. - The configured embedder is the one
inituses.resolve_embedderauto-detected from environment variables and fell through tomock, never reading theembedderfield in~/.repowise/config.yaml— the fieldrepowise serve's prompt writes. A machine configured only there indexed with keyless vectors, pinnedmock, and got no semantic search, with nothing to say so (#2074). - Persistence: symbol names are stored as
Textrather thanVARCHAR(255)(#1705), the WAL switch no longer kills a read against a live writer (#2064), page snapshots are swept when a repository is deleted (#1945), and stale pages take oneUPDATErather than two copies (#1199). - Embedder:
--embedder autoresolves through the repo pin rather than the environment, so a pinned backend is the one that queries the table it wrote (#2032). - Health and risk:
analyzed_commitis stamped by every remaining metric writer (#1864, #2034), andget_risknormalizes thegit_metadatatarget path (#1849). - Ingestion: a package language tie resolves by name rather than walk order (#2028).
- CLI and server: the page listing is no longer capped at 10,000 (#2061), git logs work in worktrees (#2045),
doctorexplains stale pages and why--repaircannot clear them (#1710, #1956), and workspace contract links inCLAUDE.mdare ranked by confidence before the cap (#1588, #1957). - The provenance footer on structural pages is a horizontal rule again. It rendered as
---*Built from the code's structure..., with the rule and the text joined, which is a paragraph beginning with three dashes rather than a rule (#2080). - UI: performance filter counts are clarified (#2044) and the context "all" badge is scoped (#2042).
- Plugin: the
change-reviewskill renders from its shared source (#2023).
[0.47.0] — 2026-08-30
The headline this cycle is that Code Health stops handing you observations and starts handing you work. Performance findings compose into per-file opportunities with a real lifecycle, served from a materialized read model so the web page, the CLI and the agent surface cannot disagree about what one opportunity is. Alongside it, a contextual chat dock that can read the indexed repository and keep its artifacts, coupling that says whether two files should change together rather than just that they did, and an architecture view that finally does something useful with external dependencies.
Added
- Contextual repository chat. Ask a question about the repo from a persistent dock that follows you across pages, with a conversation lifecycle and a durable artifact workspace so the things a conversation produces outlive it (#1949, #1951, #1960, #1964). The dock can be dismissed and brought back from settings (#2017).
- Refactoring opportunities are composed, not listed. The Refactoring board showed detector outputs, so a file with fifteen of them filled the screen fifteen times and the reader reassembled "this file needs work". Board rows, the structural field, the drawer, the CLI and the VS Code list now read one composed opportunity per file: its steps in dependency-safe order, how many are mechanical, and whether they address what is actually costing the file most (#1988, #1998). Plans gained a versioned identity and a lifecycle, and triage is one request (#1986).
- The Performance tab is a queue of causes, not a list of observations: what to change, why it ranks there, and the evidence behind it, projected from a materialized read model with server-side filters, counts, ordering and paging (#1981, #1985). Production, Tooling, Test and Unclassified stay separate, each with its own count including zero, and the performance lens joined the one Galaxy view (#1987).
- Coupling says whether files should change together. A co-change pair is now graded by whether the graph can explain it, filterable by that verdict, and openable as a pair with the matching
edge_typeshown, so "explained" says whether an import or a framework binding accounts for it (#1972, #1973, #1975). - External dependencies are a real view. Bounded per-package summaries, a dependency explorer, and focused package relationships in the architecture view (#1942, #1943, #1946).
- Local OpenAI-compatible gateways are supported as a provider, for LM Studio, vLLM and friends (#1822).
Changed
get_change_riskleads with what the change made worse. The response now opens with adirective(status, headline, reasons, next actions) and ahealth_deltaof the findings the change introduced or resolved, ahead of the percentile and the diff-shape drivers (#1980). Every previous field is still there; the ordering and the lede changed. The Claude Code plugin'schange-reviewskill is updated to match.- A repeated-cost cause is named by the caller that repeats the work, not only the sink that pays for it, so a shared infrastructure helper no longer merges unrelated workflows, and execution context gained
unknowninstead of defaulting an unclassifiable file to production (#1978). The opportunity queue defaults to production context (#2003). HEALTH_ANALYZER_VERSIONis 7. An existing index re-scores health on its next update rather than waiting out the decay timer; it does not force a re-index.STORE_FORMAT_VERSIONandPARSER_SCHEMA_VERSIONare unchanged.repowisestarts faster. The per-call CLI startup floor is cut (#1833), and an update walks the repo once rather than once per worker (#1926).- A refactoring plan the gates cannot prove behaviour-preserving is suppressed rather than offered (#1984).
- Loading motion moved from the box to the region, the sidebar's scale and rhythm were reworked, the AI prompt modal uses hairlines instead of a filled well, and model work has its own accent instead of borrowing orange and green (#1993, #1994, #1995, #2015, #2016).
Fixed
- Performance analysis: the loop-carried must-def analysis is solved to a fixpoint rather than to one pass (#2009), and the hot-path gate requires centrality before it fires (#2007).
- Dead code: JSX prop-guard dead code is detected (#1570), and unused-export confidence is clamped when a dynamic use cannot be resolved (#1959).
- Ingestion: three-part qualified calls resolve (#1920), C++ export-macro forward declarations parse (#1913), and Pascal dead-code false positives are resolved with type kinds disambiguated (#1675).
- MCP:
get_whytrims its lanes before dropping them and names what the answer rests on (#1954, #1963), guided tour steps are renumbered after dedup (#1889), path search supports trailing globs (#1881), and a workspace query prefers the most-specific repo (#1865). - Security scanning no longer reports
RegExp.exec, and catches the constant secret form it was missing (#1947). - Coverage: the Tests tab answers for files a partial report never named (#1805), and severe path-mapping loss is flagged instead of being treated as success (#1955).
- Health persistence: a full index is stamped with its commit and composed after governance (#2001), and the derived queues are rebuilt whenever the findings are replaced (#2000).
- Config and workspace: broken repo config is surfaced instead of silently falling back to defaults (#1958), workspace updates are single-flight (#1834), DB identity resolves on delegated worktree updates (#841), and
codex_cliis in the server's provider catalog (#1953). - Chat: GPT-5.6 function tools work against OpenAI (#1950).
- The coupling tab no longer crashes the page and names files unambiguously (#1965), the workspace health score is banded on the same ladder the repo pages use (#2012), the chat dock lifts only where the graph controls are actually under it (#2013), the dark overlay plane and its nested wells settle (#2014), the opportunity row columns align below the
lgbreakpoint (#2010), a refactoring step whose plan falls outside the limit renders (#2004), and the coverage migration has its own revision id (#2019).
Documentation
- The security scan is documented, floor and all (#1948).
[0.46.0] — 2026-08-27
The headline this cycle is cross-repo contracts becoming a surface you can use rather than a graph the indexer knew about. A workspace now recognises routes once and serves them to both producers and consumers, binds each contract to a real symbol id, reads request schemas off handler signatures, and gives the whole thing a front door in the UI and over REST. Alongside it the MCP tool surface got a shared response ceiling with recoverable pagination, so no tool can silently truncate an answer with no way back, and the C++, Rust, Java and TypeScript call graphs each lost a class of wrong edge.
Added
- Cross-repo contracts have a front door. A workspace's contracts are browsable, open from the table into what they link to, and carry the line, symbol and schema over REST (#1890, #1899). One route recogniser now serves Go, Laravel and Axum, with Django, JAX-RS, Next.js and the Hono router DSL alongside them (#1867, #1860), and a library's public API counts as a contract in its own right.
- Python HTTP consumers resolve through their client bindings, so a call made through a session or a generated client still names the path and verb it actually reaches.
- A shared response ceiling across the MCP tools. The five uncapped tools now answer within a canonical budget enforced at delivery, and every cap is recoverable: relationship, history and truncated test lists all carry a way back to the rows they dropped (#1924, #1930).
- Canonical references that round-trip. A reference emitted by one tool can be passed to another, and
get_whyreturns evidence references pointing at what it read (#1927, #1932). claude_cliprovider, for indexing against a Claude Code subscription instead of an API key (#1514).- The knowledge graph says what it is not showing (#1847), its arrows mean execution rather than mere reference (#1832), and a folder card is named after the module page documenting it (#1828).
Changed
- BREAKING (MCP):
get_risk'swill_breakfield is nowmay_break(#1892). The old name asserted a certainty the structural heuristic behind it does not have. Any client readingwill_breakmust be updated. get_overview's onboarding blocks are opt-in, andget_answerandget_overviewreturn leaner payloads for the same content.- Retrieval precision improved for
get_answer, which can now see the call graph when expanding a question (#1933). HEALTH_ANALYZER_VERSIONis 6 andPARSER_SCHEMA_VERSIONis 2. Both are picked up automatically: an existing index re-scores health on its next update rather than waiting out the decay timer, and the parse cache re-parses once so the call-graph fixes below actually apply. Neither forces a re-index.
Fixed
- C++ call resolution: scoped
Ns::f()calls resolve against their qualifier rather than falling to a bare name (#1915), chained calls are constrained by the inner callee's return type (#1782), in-class method declarations and receivers typed from declarations are extracted, types used only inside their own translation unit are rescued, types behind export macros parse, and an overload set resolves instead of being refused. - A Rust macro invocation is no longer counted as a function call (#1885), a Java chain rooted at an external type is no longer read as a bare-name call, and the repo-wide receiver tier no longer answers for foreign types.
- Health scoring inputs: Python
matcharms count toward cyclomatic complexity (#1795), Pascal else-if chains flatten instead of nesting each arm (#1679), a comment inside a parameter list is no longer counted as a parameter (#1796), and two test-detection gaps closed, one a pairing the name heuristic could not see (#1707) and one Delphi'sTest<Stem>.dprconvention (#1706). is_hotspotis emitted by the backend rather than re-derived in the client (#1797), fix density ranks against commits rather than individual files, andtests_to_runranks by files reached rather than alphabetically.coverage addexits non-zero when it stores nothing (#1751), and config errors that cannot self-heal no longer promise a retry.- The file page renders again: a pure
?tab=helper had been exported from a client module and called from the server (#1936). - API timestamps render as UTC (#1835),
CLAUDE.mdis stamped with the indexed commit rather than liveHEAD, and aSELECT INTOtarget is no longer flagged as a cartesian join.
Documentation
- The benchmarks pages lead each section with its result, pair precision with recall so neither reads alone, and state both cost denominators (#1928).
- The MCP tool table's drifted rows and the README's tool-count claims are corrected and asserted by a test.
[0.45.0] — 2026-08-21
The headline this cycle is test intelligence that needs no coverage report. 0.44.0 made the call graph good enough to lead with; this release walks it to answer which tests reach a file, serves that through the API, and turns the Coverage tab into a Tests tab that says something useful on a repository that has never ingested a report. Around it, Code Health's performance findings become bounded opportunities that open the matching refactoring plan, Eden AI joins as an EU-hosted provider and embedder, and pinning BLAS threads before numpy loads halves peak memory on a cold index.
Added
A test-to-code map that needs no coverage report. The per-test map only ever existed if you ingested a coverage report with contexts, so on most repositories
tests_to_runwas empty,impacted_testssaid "run the full suite", anduntested_hotspotfell back to matching filenames. The call graph already records which test can execute into which source file, and that relation now answers when the measured one cannot. It is labelledinferredeverywhere, never blended with measured coverage, and never turned into a percentage. Nothing is stored: it is a bounded walk over rows already indexed, measured at 63 ms once per health run on a 3,700-file repository.The call graph is the primary signal and the import graph the weaker fallback, which is a measured choice rather than a preference. Dogfooded against a real
coverage run --contexts=testover a slice where per-test attribution is complete, with both sides seeing the same 37 test files: walking imports one hop scored 72.1% precision at 19.5% recall, walking calls three hops scored 91.7% at 27.7%, and dropping the one call-resolution strategy that only matches a name repo-wide (global_unique) took that to 95.7% at no cost to recall. Unioning the two is worse than either lead, so the import tier is spent only on the files the call graph says nothing about, where it answers one more of them at no loss of precision.New fields:
tests_to_run_basisonget_risk's directive (measured/inferred/none),basisand astatus: "inferred"onget_change_risk'simpacted_tests, and aviamarker on everyrepowise impacted-testscandidate saying which tier answered.Code Health performance opportunities now lead into the existing Refactoring workflow. A dedicated Performance tab groups raw findings into bounded causal opportunities with production/tooling versus test context, affected-call-site totals, confidence, provenance paths, and raw-evidence paging. Exact stable opportunity ids open matching
performance_fixplans; opportunities without a safe plan say so instead of selecting a nearby one. The Refactoring page adds a Performance filter and shared priority, validation, path, and pagination renderers. Its new server-owned page endpoint bounds initial payloads while preserving the legacy unpaged route. VS Code consumes the same contracts and renders priority and validation without web dependencies; older missing optional fields remain supported./health/tests-reaching?file_path=answers which tests reach one file, withviaseparating a test whose calls run into it from one that only imports it, and/health/coveragefalls back to the inferred map underbasis: "inferred"when no report was ingested. The two bases never share a field, so a consumer cannot render derived data through the measured code path by accident.include_inferred=falsedeclines the fallback for a caller that only wants the cheap badge number. (#1757)Eden AI is a first-class LLM provider and embedder. An EU-headquartered gateway exposing 700+ models behind an OpenAI-compatible endpoint, so it reuses the existing
openaidependency with a custom base URL and adds no new one.EDENAI_API_KEY,EDENAI_BASE_URLfor the EU endpoint, and an embedder defaulting to a 1024-dimension EU model. It is appended to the autodetect order rather than inserted, so an unrelated key in the environment cannot move anyone off the provider they were already resolving to. (#705)
Changed
The Coverage tab is now a Tests tab, and it answers without a report. The URL stays
/coverageand "coverage" keeps meaning the measured thing inside the tab; what changed is that the empty state no longer tells you to go ingest a report while claiming nothing is inferred. On the inferred basis the chart collapses to two positions rather than a 0-100 continuum, which is the honest statement about the evidence, and nothing is rendered as a percentage or a bar. The file page names the tests that reach it, withviaseparating a test whose calls run into the file from one that only imports it. (#1758)OpenAI-compatible local gateways expose their models. Custom endpoints configured with
OPENAI_BASE_URLcan now return namespaced model ids such as 9router'sag/...andds/...entries; loopback requests also bypass proxy environment parsing so a standard IPv6NO_PROXYentry cannot block them.Custom OpenAI-compatible setup is a first-class init flow. The provider menu separates official OpenAI from custom/local gateways, validates the endpoint, hides and optionally saves the key, verifies
/models, supports model search plus exact-id fallback, and offers retry/manual/back recovery.--no-save-keypersists the non-secret endpoint without persisting the key; CI,--yes, and non-TTY runs remain non-interactive.untested_hotspotstops accusing files that the tests run. With no coverage ingested it fired on any hotspot without a paired test file, which is a filename convention, so a suite that names its tests for behaviour satisfied nothing. A test whose calls reach the file now suppresses it too. On this repository five of the six worst bug-magnet files had no test named for them and read as untested; the sixth,analysis/health/engine.py, was called tested because the convention matcheddistill/test_engine.pyon basename alone. The same floor now runs underget_risk'smissing_tests, which had two separate filename heuristics that disagreed.Measured on repowise's own index, against the 80 standing
untested_hotspotfindings the filename convention leaves behind: the call graph clears 32 of them, 22 graded high and 3 critical, where a one-hop import walk clears 11. The persistedhas_test_filewidens to match, so the file table stops labelling those same files "untested" while the biomarker says nothing.Both stored values are wrong on an index built before this, so
HEALTH_ANALYZER_VERSIONmoves to 3 and the nextrepowise updatewith changed files re-scores health rather than waiting out the decay timer.repowise impacted-tests --format jsonrenamesguessed_teststoinferred_tests. The bucket no longer holds only filename guesses, so the old name described the wrong thing; each entry carriesvia(call-graph,import-graphorfilename-pattern) to say which tier answered.Refactoring recommendations have one contract. REST, MCP and the CLI each rebuilt their own ranking and payload shape, and the TypeScript contract was duplicated in the UI package. The wire contract moves to
@repowise-dev/typesand the Python side gets a single read model owning rehydration, enrichment, ranking, validation and serialization for every surface. Blast radius now raises cost and risk rather than benefit, so a wide change ranks as expensive instead of valuable. (#1784)The semantic leg of search gets a budget the cold path can clear. The first vector query in a process pays for the store open, the first embed and the first ANN probe. That was measured at 6.3s and 13.4s on a cold Windows index where a warm query takes 0.19s, against a hardcoded 8s bound inside a suppressed exception. The budget is now one shared 30s default, capped at 120s and overridable with
REPOWISE_VECTOR_SEARCH_TIMEOUT_S, and a timeout logs a warning naming the knob instead of silently degrading to full-text with nothing said. (#1678, #1685)Plugin:
/repowise:impacted-testsand thechange-reviewskill describe the inferred map as the call graph with an import-graph fallback, andinit/reindexlist the embedders the CLI actually accepts.Execution-graph analysis is unified behind one code path in health, with the legacy call fallback restricted rather than left answering alongside it. (#1773)
Fixed
- An incremental update stopped erasing ingested coverage. The incremental
pipeline built its analyzer without a coverage map, so every changed file was
re-scored as if no report had ever been ingested and the partial-health writer
upserted
line_coverage_pctback to NULL, eroding coverage one file per update, starting with the files under active development. Both callers now load and supply the persisted map. (#1739, #1806) repowise initno longer crashes in a git worktree..gitis a file there, not a directory, so building.git/hooksby hand raisedNotADirectoryErrorand took the whole command down. The hooks directory is resolved withgit rev-parse --git-path hooks, which answers correctly for a worktree and a normal clone alike. (#1609, #1808)- The CLI and the MCP server resolve the same embedder key. The server never
consulted the process environment, so on a machine with an exported provider
key different from the repository's
.repowise/.env, semantic search degraded through the CLI while the MCP tool answered. The server now reads the env tier first, matching the CLI's "an exported key is an explicit override" contract. (#1711, #1810) edenaiis selectable everywhere it is documented:init --embedder,reindex --embedder, both interactive prompts,EDENAI_API_KEYautodetection and the server's embedder build, four lists that were left on the older five-backend set. (#1815, #1820)repowise reindexexits non-zero when every embed failed, so an automated pipeline cannot treat an empty vector index as a successful build. An empty wiki still exits 0. (#1495, #1732)- The file page states how many tests reach a file, not how many it listed. The list is capped at 50 and the page rendered that cap as the measurement; it now leads with the true count and says what it is showing. (#1764)
- A generated-file banner has to be a banner. The marker is required within the first two lines instead of substring-matched across the whole header, which fixes the class of prose false positives across all six markers. (#1519, #1731)
- Python resolution:
from pkg import submodulepinned the binding at the package__init__.py, so a latersubmodule.symbol()call resolved against a file declaring nothing and dead code called the symbol safe to delete (#1193, #1733); aliased submodule bindings are pinned properly (#1255, #1785). - Go: a call through a field (
a.b.Method()) matched no pattern and was missing from the graph entirely rather than present and unresolved (#1727), and an embed ofio.Readerkeeps its package qualifier instead of binding to whatever repo-localReadershares the short name, which made the type inherit from itself when the enclosing type had that name (#1726). - TypeScript:
#privateclass members produce a symbol and a call site (#1715, #1721), a return type no longer renders with a stray colon (-> : string) (#1695, #1719), and a relative import probesindex.tsxandindex.jsx(#1725). - Java and C# method return types are preserved in signatures (#1713).
- The bare-name tier stops answering for the standard library (#1708), and
barrel-file definitions are aligned with
file_reachability(#1491, #1557). - The paired-test filename heuristic is shared rather than reimplemented per surface (#1750).
- Security scanning:
replace_findingsno longer loses rows on a duplicate key (#1523), and a multi-linesubprocess.*(... shell=True)call is flagged (#1522). - Agent CLIs no longer load the target repository's instruction files when invoked for generation (#1100).
workspace init --dry-runwrites nothing (#1526), andinit --test-runactually limits generation to the top 10 files (#1525).share_of_repo_gap_pctuses the gross gap as its denominator (#1452), and history fetches avoid the partial-clone path (#1608).- The
get_riskartifact renders in chat. The renderer mapped over a list and readpath/churn_percentile, but the tool returns a path-keyed dict ofhotspot_score/file_path, so opening the artifact threw or showed blank rows. Both shapes are normalized, and the 0-1 score is scaled for display rather than printed as though it were already a percentage. (#1480)
Performance
- Peak memory on a cold index is roughly halved. Importing numpy brings up a
BLAS runtime that commits a private per-thread workspace sized to the host's
core count: 746 MB on a 32-core machine for the import alone, committed once,
never returned, and invisible to
tracemallocbecause no Python object owns it. The pipeline's only numpy work is PageRank over a sparse graph, which runs single-threaded in well under a second, so the thread count is pinned before numpy is imported. A coldinit --no-proseon an 876-file repository went from 1,547 MB to 818 MB at no wall-clock cost.OMP_NUM_THREADSis deliberately left alone: igraph community detection really is OpenMP-parallel, and pinning it cost 18% wall clock for no further saving. (#1394, #1776)
Documentation
- A roadmap, an enterprise section, and a benchmarks page that leads with its
results.
ROADMAP.mdis new and states plainly that language support is never gated behind the licence; the README gains a real enterprise section covering deployment, data boundary, compliance and contract;BENCHMARKS.mdgoes from 777 visible lines to 456 with every caveat and every losing row preserved behind folds. Language support is now stated the same way in every document: 19 parsed to a full AST, 35 on the ladder. Source control beyond git (Perforce, Subversion, and the mainframe promotion hierarchies) is scoped to the one history-derived layer that actually needs it. (#1766) - The graph precision result is published, judged by a compiler across five
tools (#1738), and
COMMERCIAL.md's tool count matches the way the document counts (#1771).
[0.44.0] — 2026-08-18
The theme this cycle is the call graph knowing who the receiver is. A call written user.save() only becomes an edge if something can say what user is, and for most languages nothing could: the resolver matched the bare name against every symbol in the repo and guessed. Receiver typing now runs for Go, Kotlin, Swift, C#, Java, Python, PHP and Luau, and resolution follows re-export chains. On microdot, a small pure-Python repo with none of the languages that gained most, call edges went from 876 to 1,371. Dead code got the matching precision pass, and updates now carry both onto an index built by an older version, which they previously did not. Expect the first repowise update after upgrading to run long, once.
Added
- Receiver typing across eight languages. A call's receiver is typed from its declaration in Go (#1674), Kotlin (#1687), Swift (#1688) and C# (#1680), from the enclosing class for a bare call plus PHP and Luau receivers (#1630), from a field's declaring class (#1642), from a local's declaration (#1639), from assignments in a Python function body (#1643), from a Java call on its own field (#1658), and from a framework decorator that retyped it (#1684). Three more receiver shapes a bare-name match had been guessing at are now captured directly (#1686).
- Resolution follows re-exports. A call resolves through the name a module publishes (#1682), a method on a type imported through a re-export (#1664), and a namespace member through the whole re-export chain (#1672).
dispatches_toedges link a base method to the implementations that answer for it (#1649), and a framework-wired symbol to the symbol it is wired to (#1654).- Every call edge records the strategy that resolved it (#1628), and the web UI says how each edge got into the graph (#1652). An execution flow now says why it stopped instead of just stopping (#1650).
- Security findings carry a verified line and a commit date (#1668).
- Pascal is registered in the complexity, duplication and dataflow dialects, so health scores it like the other 18 languages (#1629).
Changed
- The file detail page is ported onto the design language (#1621) and gained inbound and outbound navigation (#1622).
- The workspace System Map moves its chrome off the canvas, and the per-repo view works (#1616).
- Dead code surfaces staleness on the web and drops the package column (#1669).
this.method()self-dispatch is recorded in six languages (#1617).- Export aliases are read from the parser rather than a second scan of the file (#1683).
Fixed
- An extraction change now reaches unchanged files. An incremental update rewrote only the git-changed files' rows in
graph_edges, which is right for a content change and wrong for a parser change: the latter alters every file's edges at once. The build that wrote a repo's edges is now recorded, and a mismatch widens the reconcile to the whole parsed set once before re-stamping. Without this, none of this release's graph work would reach an existing index short of a full re-index. (#1619) - Dead-code analysis no longer skips itself on update. Stored commit timestamps came back without a timezone while freshly read ones carried one, so ageing a package raised
TypeError, a broad catch turned it into a one-line warning, and every finding silently kept its previous verdict. Present since 0.40.0. (#1702) - Symbols reached by a framework or a container (#1673), wired in by a registration decorator (#1681), or named by a docs build or an API dump (#1677) are no longer reported as dead.
- Every use of a private symbol counts, not only a call (#1662). Only genuinely narrow scopes stay in the uncalled-symbol pool (#1646), and a top-tier confidence means the checks behind it ran (#1666).
- C and C++: a forward-declared type is not an unused export (#1700), nor is a template forward declaration (#1703); a call edge attaches to the definition rather than the header declaration (#1626); a function named but never called counts as a use (#1627);
.inl/.ipp/.tpp(#1625) and.hhare recognised as C++, and.incis reassigned from Pascal (#1693). - Every language gives one answer for a type's bare name (#1634). A struct field is not callable, so the bare-name tier no longer offers one (#1692), and Rust type positions are filed as
type_userather than calls (#1690). - Inheritance is emitted for JavaScript classes (#1636) and Go interface embedding (#1641). C# resolves inherited calls again (#1651), captures generic method calls (#1637), and records the visibility a declaration actually has (#1644).
- A third-party JVM import no longer takes a same-named repo class (#1659). A Go method whose receiver is unexported is typed (#1676). A workspace package that publishes only from a build directory binds correctly (#1670).
- A method passed as a value is separated from a method called (#1661).
- MCP: four tools answered a bad argument with a reassuring negative instead of an error (#1671);
get_contextcounted a subclass and a fixture as callers (#1663); raw doubles and two different scales shipped under one key (#1631). The symbol page made the same subclass mistake (#1660). - A failed framework pass is reported rather than swallowed (#1645).
repowise update --fullruns under the single-flight lock (#1529), and the incremental hotspot gate reuses the persisted function-modification p80 (#1532).- ADR discovery no longer mines paths git cannot see (#1614), the security history scanner actually scans (#1667), and fix history ignores the shallow-clone boundary (#1633).
Performance
- Graph build stops re-reading the repository (#1648), and each language gets a call-strategy seam that rejects unresolvable names first (#1632).
- The file page stops reading the whole repo to render one file (#1618).
- Workspace
.csprojscanning walks each repo once (#1615).
Documentation
- The language-support pages are rewritten, a graph layer page is added, and the claims that pointed at them are corrected (#1689).
[0.43.0] — 2026-08-15
The through-line this cycle is a number meaning what it says. Change risk scored the size of the diff and presented it as danger. get_answer could return confidence: high beside retrieval_quality: weak, with a note that asserted dominance and denied it in the same sentence. get_why answered every question, including the ones its store knew nothing about. The dashboard summed dead exports across a side project and a monorepo. Six surfaces each computed hotspot health their own way and none of them agreed. Each of those is fixed by giving the answer one owner and reporting the reason alongside it. repowise risk now leads with the bug-fix history of the files a change touches, an answer's confidence note quotes the test that earned it, and a question the decision store cannot answer gets a redirect instead of the three closest records.
The second theme is what a keyless install gets. get_answer with embeddings but no LLM key used to return a ranked file list and nothing else, though the evidence around it (live symbol bodies, candidate justifications, rationale comments) needs no provider at all. It builds all of that now, and rates its retrieval instead of only doubting itself. init also stopped losing an API key supplied through the environment, which is the path agents and CI take.
Underneath, the graph got a good deal more honest: the edge-type vocabulary is now true and enforced, dependency answers stopped including containment and co-change rows, and "can execution start here" has one owner instead of four.
Added
- Object Pascal / Delphi support (Good tier), taking the language count to 19. Symbols, imports, heritage, and a call graph that handles the parenless calls Pascal is written with. Verified against a 151-file real Delphi codebase. (#1353)
- Dead code detects deprecation annotations.
@Deprecated,[Obsolete],@deprecatedand their language siblings now feed a finding's confidence. (#1472) repowise decision addrecords from flags. A flag per field, and no prompting at all once--titleand--decisionare both present, so a script or an agent with no terminal can record one.--format jsonreturns the id to pass back toconfirmorshow. Flag-driven records landproposedwhere the prompts recordactive. (#1566)- Record a decision from the web UI. The decisions page had no way to start a record and an empty state that sent you to a terminal. It now has a form, reachable even on a repo with no decisions yet. (#1567)
GET /api/repos/summaryreturns every repository's headline figures in one call, replacing the2N+1requests the multi-repo dashboard used to make. Flat in the number of repos: six queries at 1 repo, six at 26. (#1577)repowise workspace scan --exclude, plus a bulk decline, so a scan of a large tree does not have to be answered repo by repo. (#1591)
Changed
repowise riskandget_change_risklead with fix history, not the score. The 0-10 score restates the size of the diff: scoring by lines added alone reproduces it to within 0.16 points, and on 47 within-repo pairs across four repos it ranked the dangerous change above the boring one in none of them. The response now leads withfix_history, the recency-weighted bug-fix record of the files the change touches, which files carry it, and where that sits among the repo's own fix-bearing files. Ranking by that alone gets 46 of the same 47 pairs. It comes from onegit logwalk, so it needs no index and no coverage data. The score is kept and reported as what it measures, diff size and spread, named inscore_measures. Model constants are unchanged. (#1593)- Change-risk numbers say what they are.
probabilityis gone (it wasscore / 10with extra decimals). The absolute band is nowfallback_bandand appears only when there was no baseline to rank against. A newscore_unitstates that the scale is calibrated on individual commits, so a squash-merged PR or abase..headrange reads high by construction.nf/nd/nsare no longer reported as drivers: their coefficients are collinearity with diff size, so as an explanation they contradicted themselves. (#1583) - Default OpenAI model is now
gpt-5.6-luna(wasgpt-5.4-nano). Same input price ($0.20/MTok), cheaper output ($1.20 vs $1.25), so it gets its own pricing rows rather than an alias, and nano stays selectable on its own rate. The 5.6 family also changed its reasoning ladder:minimalis gone andxhighis new, so--reasoning minimalis rejected up front for these models instead of failing as a 400 on the first live call. Despite what the model docs say, the API also rejectsmaxfor 5.6 (verified live against bothgpt-5.6-lunaandgpt-5.6-sol), so it is not offered. (#1594) get_answerserves real evidence with no LLM key. An install with embeddings and no provider got a ranked file list and nothing else, though none of what it lost needs a provider: bodies are read live off disk at the indexed anchors. It now buildssymbol_bodies,best_guessesandcode_rationale, cites the paths they came from, and names one next action. It also rates the retrieval rather than only reportingconfidence: low. On 26 paired questions,symbol_bodieswent from 2/26 to 11/26 keyless. (#1508, #1576)- The per-file symbol budget is spent on the question. When a retrieved file held more symbols than its budget, hydration kept a prefix sorted by line number, so the bigger the file the stronger the top-of-file bias and the less reachable the symbol you asked about. Symbols are now scored against the question's terms for selection. The kept slice is still served in document order. (#1597)
get_whyranks by question vocabulary and refuses when nothing matches. Search mode answered every question: over twenty questions on an indexed repo, none of the eight unanswerable ones was refused, each costing 11k to 15k characters to answer wrongly. Ranking is now the share of the question's vocabulary a record carries, weighted by term rarity, and a miss returns an empty list with the tool that fits the question's shape. Average response 12,492 characters down to 5,624. The right record now ranks first on all twelve answerable questions, up from none. (#1558)get_whysearch payloads roughly halved. Natural-language mode was the only one of the three with no caps at all, at 27k to 35k characters against a 32k transport budget, so two of five probe questions were over the ceiling outright, where the host rejects rather than truncates. One embedding per call instead of two,affected_filesclamped to a head plus a total, restatements collapsed on cited evidence, and three whole records rather than eight thinned ones. Now 12k to 16k characters, latency down 11% to 29%. (#1555)get_whyanswers about its--target/targets. It chose the health dashboard before looking at the targets, so naming a file and asking nothing returned the repo-wide summary. Targets are read first now: one target is path mode, several get a card each. (#1566)- The multi-repo dashboard puts the repositories back at the centre. Four metric cards over a repo list becomes a repository list as the subject, ordered by attention rather than last write, each row carrying its health band and a marker only when something needs doing. The aggregates that legitimately add moved to a stat ribbon.
file_countwas counting symbol nodes as files, about 10x too large, and is fixed at the source, so the chat empty state gets it too. (#1577, #1578) - The workspace pages are on the design language. All four are server-rendered again with filter and selection state in URL params, so a filtered view is linkable. Figures that summed things which do not add were deleted rather than restyled: an averaged coupling heuristic over a truncated list, an "Unmatched" count computed over one page of 200 rows and presented as a workspace total, a breakdown whose sum equalled the total it broke down, and an unweighted average of per-repo coverage percentages. Contract figures now come from the diagnostics endpoint, which knows the denominators, and the table states its bound. (#1582)
- The VS Code extension is on the current design language. Two local
scoreColorcopies called anything at or above 7.5 healthy-green while the canonical bands start at 8, so a file's score printed green beside a map colouring it amber. Both copies are gone. High-contrast themes now get a real high-contrast treatment instead of falling through to the ordinary ramps, and the health, settings, decisions and risk views take the shared components. (#1575) --no-editor-setupmeans no editor setup. It gated only machine-wide MCP registration, so there was no combination of flags that indexed a repo without writing.mcp.json,.claude/CLAUDE.md,.vscode/mcp.jsonand.vscode/extensions.jsoninto it. Now one switch means one thing: only.repowise/is touched.initalso ends by naming the files it actually wrote, sourced from the writers rather than a hardcoded list. Note for CI:REPOWISE_SKIP_EDITOR_SETUP=1is the same switch, so a harness exporting it to protect a global config will stop getting.mcp.json. (#1572)- The init screens are one system. The banner's wordmark ran a colour ramp whose jitter exceeded its own gradient, so what shipped was orange static in the first thing a user sees. It is one brand colour now. Four panels that are not objects you act on lose their border, the palette is closed to the declared colours, and the completion panel names two commands instead of up to seven. (#1571)
- Generation progress splits paid work from free. One bar counted 4,229 items where about 4,134 were template renders and 95 were model calls, so it reached 97% in minutes and then sat there, indistinguishable from a hang. Each cost tier gets its own total, the free bar finishes and hides, and persistence gets a bar instead of an indeterminate spinner. (#1599, #1607)
- Degraded phases are visible on a normal run. The CLI pins core logging to ERROR unless
--verbose, so whole phases could fail (the parse pool dying back to one process, a checkpoint not reaching disk) and print nothing. Those now render through the progress channel, including on a non-TTY, and are collected intostate.jsonasdegraded, the wayupdatealready reports one, with key-shaped text redacted on the way in. (#1599) - The Codex hook stopped waking on every prompt. It registered
UserPromptSubmitwith no matcher and returned the same static noteSessionStarthad already delivered, so an agentic turn paid one process start per prompt to repeat a block the agent was already holding. The event is retired, and existing installs are migrated rather than left behind, since the merge into.codex/hooks.jsonis additive and reinstalling would have repaired nothing. (#1549) - Hotspot health has one owner, computed live. Six surfaces answered the question and disagreed.
get_overviewandrepowise statusaveraged the top 25% of files by size rather than churn, differing from the persisted KPI on all 42 local indexes with a median gap of 2.67 points of 10. Two dashboard routes read a snapshot that goes stale afterrepowise update, andget_healthdid not return the number at all.Nonenow means "this repo has no hotspot files", kept distinct from a low score. (#1552) - The edge-type vocabulary is true and enforced. The
EdgeTypedeclaration was decorative: it listed three types nothing emits and omitted four that producers emit constantly (6,750 rows), and thirteen modules each kept a private set written against it. Shared families replace all of them, with an AST guard test over every call site. Cycle detection gains 6,153 edges it never matched, and a file reachable only through a dynamic import could previously be reported dead. (#1516, #1518) - Dependency answers read dependency edges.
graph_edgesalso holds containment (defines,has_method) and the temporalco_changesrelation, and several graph reads loaded every row and presented the result as dependencies. 43.7% of the rows the file-dependencies panel read were neither, and 34,570 of this repo's 34,808 symbols listed their own declaring file or class as a caller. Each site now names the view it wants, and degree counts are scoped to whatever is displayed beside them. (#1531) - "Can execution start here" has one owner, decided at ingestion. Four implementations disagreed, so one file could lead the orientation list and be absent from the curated one. The flag is also what exempts a file from dead-code detection, so a name-only guess was a file nothing could ever report: 1,151 flags are withdrawn across 42 indexes and exactly one file becomes a new dead-code candidate. Entry-point lists on three remaining surfaces now use the shared ranking rather than PageRank, which rewards fan-in and floats a package-root barrel above the real front door. (#1533, #1547, #1556, #1559)
- Workspace contracts are read from the parsed index. Extraction re-derived route tables with regexes after ingestion had already parsed the same files. Python HTTP providers now come from the symbols ingestion produced, so a route cannot be picked up from a comment or lost for having an empty path, and the regex dialects stay reachable for every language with no AST tier. Five extractors that each walked the repo now share one walk: 45.2s down to 14.7-17.9s, 13,889 file reads down to 3,216, contracts and links unchanged. (#1581)
- HTTP client wrappers are confirmed from the parse, not from their names. The JS dialect decided a call was a service call by matching the callee's name against
fetch|request|http|api|ajax|rest|rpc, which invents contracts for functions merely named that way and misses every wrapper that is not. A symbol is now confirmed by its body issuing a sink call, within a 2-hop budget. Measured on one frontend API layer: 12 contracts to 162. Unresolvable paths are counted and reported, never guessed. (#1585) - A workspace commit re-extracts only the repos that changed.
changed_reposwas accepted and never read. Reuse is validated rather than trusted (carried forward only when the alias is unchanged, HEAD matches the stamp, the tree is clean and the contract config is unchanged), and every way of disagreeing resolves toward re-extraction. Cross-repo analysis and contract extraction now run concurrently, and each of the five phases logs its duration. (#1596) - Truncated lists carry their true totals for cycles, co-change pairs and per-edge evidence, after the architecture score turned out to be computed from a capped cycle count, so 500 cycles scored the same as 50. Conformance and breaking-change reports are stamped when they run, so a report that never ran no longer renders as a clean bill of health. Contract extraction reports its coverage from counts it was already recording. (#1591)
- Dead-code confidence tiers have one owner. Eleven more literal
0.7/0.4comparisons and caps now read the shared constants, with a guard test outlawing the shapes that kept recurring. Behaviour-preserving, measured over 42 indexed repositories: 3,930 findings both sides, none moving. Risk factors also render as words rather than API slugs, so a tooltip reads "configuration, runtime-loaded web asset" instead of "config, asset". (#1561, #1574) get_answer's internals are split into named modules.answer.pyhad grown to 2,805 lines doing every job in the tool at once. Pure moves: no logic, ordering, threshold or payload change. (#1510, #1584)
Fixed
- Inline decision markers stopped matching ordinary prose. The keyword matched case-insensitively, so a wrapped sentence beginning "# decision: ..." and a test label reading "# Rejected: nothing to extract" became the only two "decisions" a fresh index of this repository produced, one of them a mid-sentence fragment attached to twenty files. The keyword is now capitalised like
TODO:. Markers past the fifth in a file were also silently dropped, and a source that failed was indistinguishable from a source that honestly found nothing. Failures are now reported as warnings naming the source and the error. (#1610) - A confidence note is written from the reason the grade was reached.
get_answercould returnconfidence: highbesideretrieval_quality: weak, with a note claiming "clearly dominates (dominance ratio 1.00x)", a ratio of 1.00x being a tie. Dominance had two owners that disagree in a real window. There is one now, soretrieval_quality: weakmeans exactly "not dominant", and each note quotes the test that actually earned the grade. Answer-grounding can no longer earnhighover a weak retrieval unless you opt in. (#1605) - A stray delimiter no longer hides the rest of a file. The masking walk behind
withheld_symbolsdiscarded and re-walked when a template literal was left open at end of file, but triple-quoted runs and block comments got no such treatment. Either one left open masked every line below it, and the note then described a smaller set of withheld symbols than the payload had withheld. A stored symbol end line is also clamped to the live file, so a body served whole can no longer be flagged as cut with a continuation pointing past the last line. (#1580) repowise searchandaskstopped falling back to keyless embeddings on a repo indexed with a real key in the same shell, reporting no semantic search against an index full of 1536-wide vectors. The CLI had no key resolver at all and never read<repo>/.repowise/.env, the fileinitwrites the key to, so the MCP server and the CLI disagreed about the same directory. The degraded message also names where it looked, instead of telling a user to set a variable whose value is already on disk. (#1599)initpersists an API key supplied through the environment. Key persistence hung off the interactive key prompt, sorepowise init --provider openai --yeswithOPENAI_API_KEYexported indexed fine but left no.repowise/.env, andrepowise mcpagainst that repo answereddegraded: "no-llm-provider". The key that indexed the repo is now saved, with a one-line notice naming the file.--no-save-key(orREPOWISE_NO_SAVE_KEY=1) opts out, and answering No at the prompt still wins. Coversinit, workspace init andworkspace add..gitignoreis written before the key, so a failure cannot leave a committable secret. (#1595)- The chat window never streams with nothing to show. A turn that ended without a terminal event left the composer spinning forever with no reason printed. The stream never settled when the reader ended early, an unknown-conversation failure was sent without the
typefield the client switches on, and the send never handed itsAbortSignaltofetch, so a cancelled turn left the POST running and a database session open. (#1586) - The
module_pageprompt stopped tripping its own validator. It asked the model to ground claims in "the supplied material", which is a literal match for one of the artifact rules, so a page that echoed the instruction's own wording was rejected and thrown away with its tokens already spent. Reworded, with a guard asserting no prompt or template matches any rule, plus one corrective re-ask so a paid-for page is not simply lost. Symbol spotlight pages also bound their importer list: three pages on this repo were about 46.6k characters, of which 45.7k was the list, so the tail fell outside the embedding window and the pages were searchable only by their first few hundred importers. (#1606) - The knowledge graph rebuilds when the builder changes, not only when the graph does. The fingerprint measured node, edge and community counts, which hold still while the builder changes underneath them, so a store kept serving an artifact built before six edge types were admitted. (#1592)
- Truncated agent-facing lists are ranked before they are cut. Stale decisions, proposals awaiting review and ungoverned hotspots were appended in scan order, so "your ten stale decisions" meant ten arbitrary ones. Six generation lists and the transitive-affected walk behind
get_risk'swill_breakhad the same shape. (#1590, #1598) - Change-risk correctness. Uncommitted work is scored when the tree is dirty instead of silently scoring the previous commit. Merge commits are read against their first parent, so
impacted_testsandprior_fixesare no longer empty for a merged PR the same response says touched several files. A failed author lookup reports unknown rather than0, which the model priced as risk. The 200-commit baseline walk is now cached whole rather than per scored commit, and self-exclusion works for the defaultHEADtarget, which had been ranking against itself. (#1583, #1589) - Cached pickles under
.repowise/are HMAC-sealed before unpickling, so a crafted cache file committed into a repository cannot execute code duringinitorupdate(CWE-502). (#1439) REPOWISE_EMBEDDING_TIMEOUTis honoured by every embedder. Only the Ollama one read it. openai, gemini and openrouter hardcoded 10s, and an expired batch is not retried: it is reported once as "N/N items failed to embed" while the run exits 0. A malformed value now warns and falls back instead of silently downgrading the run to a keyless 8-wide store. (#1475)kimi-for-codingmodels are senttemperature=1, the only value that endpoint accepts. Every request through the kimi provider failed validation, which surfaced inrepowise generateas every page silently falling back to a structural stub. (#1323)- A failed schema-reconcile statement stops being permanent. The walk aborted on the first raise, so every table ordered after the victim was never reached, on that call or any later one, and 13 read paths swallow the error, so a half-migrated store was reached silently. It continues past a failure now and re-raises once the walk is done. (#1509)
repowise update's up-to-date self-heal runs under the single-flight lock (#1528), and the lock never reportsacquiredwith no lock file on disk (#1524). The sync baseline is preserved when HEAD is unavailable (#1521).repowise dead-code --format jsonemits valid JSON. The banner and two notices went to stdout ahead of the payload. (#1441)- C# generic type arguments resolve (#1550), and literal dynamic-import namespace use is tracked (#1548).
- Chat tooling matches the real surface. The
get_contextinclude schema listed blocks that do not exist and omitted two that do (#1477), andget_whyresults render in chat artifacts instead of coming back blank (#1478). - Dismiss dismisses. The button sent
deprecated, which is re-derived on every re-extraction, so dismissing a wrong proposal did not stick, which is the one thing the button exists to prevent. The shared status union gaineddismissed, and the table gained the filter without which a dismissal is a one-way door. (#1567) - Docker standalone output path corrected for the monorepo layout. (#1562)
python -m repowise.cli.mainruns instead of exiting 0 with no output. (#1591)
Documentation
docs/layers/CHANGE_RISK.mdrewritten around fix history, including what the keyword-based fix classifier misses and why the ranking set is a falsification test rather than an accuracy estimate. (#1593)- Dead-code docs corrected against the code: the git-age ladder has six rungs and was documented as three,
safe_to_deleteomitted its runtime-load condition, the risk-factor token list read as complete while being partial, andARCHITECTURE.mdpromised a whitelist file no code reads. (#1559, #1574) - Workspace docs drop a pom.xml scanning claim and an OpenAPI field-level breaking-change claim, neither of which is implemented, and document the socket extractor. (#1591)
- Object Pascal added to
docs/layers/LANGUAGE_SUPPORT.mdwith its known gaps. (#1353) - A dead-code CLI walkthrough, corrected hook event lists, and the editor-files architecture doc brought back in line with the current queries. (#1549, #1556)
Dependencies
- Adds
tree-sitter-pascalfor Object Pascal support. (#1353)
[0.42.0] — 2026-08-13
repowise shipped real integrations for exactly three agents — Claude Code, Codex and VS Code — with no cheap way to add a fourth. The wiring lived in four unrelated places, each carrying its own copy of what a given agent needs, and the two plugins had already forked from one another with nothing detecting the skew. This cycle puts every agent behind one descriptor seam and then proves it by adding three: Cursor, OpenCode and Hermes. A new repowise agents command group lists what is wired, adds and removes targets, and prints a config snippet for hosts we write nothing for. A support tier is derived from what a target actually wires rather than declared, so the docs cannot claim more than the code delivers.
The seam paid for itself immediately in a place that was not the point of it. Removal was never really implemented: init wrote to six surfaces and nothing took them away, so uninstalling repowise left MCP servers registered, hooks firing, and instruction blocks in files repowise did not own. There is a repowise uninstall now, and every removal path reports what it actually removed instead of calling a half-refused removal clean.
The other theme is a group of correctness fixes in ingestion and retrieval that all have the same shape: a signal was being read as something it is not. Co-change partners were being served as dependency edges, so "what does this import" answered with files that merely change alongside it. A TS/JS binding was classified by its name rather than its value. Large source files were dropped silently. Dead code called runtime-loaded web assets safe to delete.
Added
repowise agents. Lists every agent repowise knows, with its support tier, whether it looks installed, and every place it is currently wired — including duplicates, because "configured" is the wrong answer when the truth is "configured three times". Subcommandsadd,remove,refreshandprint-config, each with--format jsonthat reports what changed, so an agent can wire itself up and read back the result. (#1448)- Cursor, OpenCode and Hermes as agent targets. Each is a descriptor file and a registry line. Cursor writes
.cursor/mcp.json(it does not read.vscode/mcp.json) plus a rules file; OpenCode and Hermes read the same host-neutralAGENTS.mdthat Codex does, so the instruction block is shared rather than owned. Hermes also writes YAML and registers into its toolset list. (#1457, #1459, #1460) repowise uninstall. Removes repowise from everything it wired: MCP registrations, hooks, instruction-file blocks, and the local index, in either scope. Shared surfaces are left alone while another agent still reads them, and the user is told who. (#1467)- Codex slash commands, built from the same shared source as the skills rather than forked a second time. A Codex plugin manifest has no slot for commands, so they install from package data into
~/.codex/prompts/; Claude Code gets its commands from the plugin and never frominit. Both hosts' skills now render from one body, with a drift report when they diverge. (#1450) - A generated agent support matrix in the docs, built from the registry so it cannot drift from the code. The published MCP tool count is generated from the same place — six artifacts had been asserting a number that was wrong. (#1455)
Changed
- The three existing integrations were rewritten onto the new seam, not extended. Claude Code, Codex and VS Code all resolve through one
AgentTargetprotocol with composable format helpers rather than a base class, because the targets genuinely differ: Codex writes a TOML server table, a TOML feature flag and a JSON hooks file for one install; VS Code writes two JSON files, one of which may carry comments. Every file the old path wrote is reproduced byte for byte. (#1446) - When a host-managed plugin is already registered, direct wiring stands down. Claude Code is reachable both through its plugin and by repowise writing the config, and the host merges both without complaint — leaving two process spawns per matched tool call and a duplicate set of tool schemas resident in every session. Measured on a live machine: three repowise MCP servers at once, roughly 36 tool schemas for one product. (#1446)
- A dead-code CLI walkthrough was added to the examples. (#1384)
Fixed
get_riskandget_contextno longer serve co-change partners as dependencies. Both read a graph edge set that mixed the two, so a file that merely changes alongside another was reported as importing it. Dependency edges are read on their own now. (#1462, #1470)- Large source files are indexed instead of silently dropped, and when one genuinely is too large to parse, the run says so and counts it rather than leaving a gap the user cannot see. (#1443, #1237)
- TS/JS bindings are classified by their value, not their name. A
constwhose value is a call was not indexed at all, and a binding was typed from its identifier, so factory results and configured instances went missing from the graph. (#1468, #1469) - TypeScript workspace members resolve from
pnpm-workspace.yaml, not only from thepackage.jsonworkspacesfield, so pnpm monorepos link across packages. (#1454) - Files are decoded as UTF-8 rather than the platform locale, which on Windows silently mangled any source with non-ASCII bytes. (#1466)
get_answercaps its confidence when a cited symbol body was truncated, instead of reporting high confidence on a partial read, and validates the withheldsymbol_idit hands back. It also stops reading backtick strings as code. (#1444, #1445, #1451, #1461)- Reachability is answered in one place. Two callers each had their own notion of "is this file reachable", so the overview page could contradict the dead-code pass about the same file. Barrel files also rank down rather than dominating. (#1487, #1464)
- Dead code stops calling runtime-loaded web assets safe to delete, and the CLI now indicates when low-confidence findings are being hidden rather than presenting a filtered list as the whole answer. (#1463, #1434)
- Deterministic structural wiki pages are localized. With
--languageset, the model-written pages were translated while every structural page stayed English, so a non-English wiki was mixed. (#1092, #1102) - Orientation entry points rank the same way on every surface — the CLI, the MCP tools and the web UI each had their own ordering. (#1488)
- Concept pages stay local and are named for the directory they are actually under. Unrelated top-level directories could be merged into one concept group, which then took a name from neither. (#1280, #1465)
repowise mcpkeeps an older index readable instead of failing on the first query, and caps confidence on a truncated symbol lookup. (#1458)- MCP tools emit
index_behindandembedder_degradedwhen false, not only when true. A caller cannot distinguish "not degraded" from "this build does not report it" when the key is simply absent. (#1449) - A failing CLI command names the real error rather than a generic wrapper, and pytest runs are flagged. (#1447)
- Webhooks match repositories by normalized URL instead of a
containsprefix, which matched the wrong repo when one name was a prefix of another. (#1440) - The test suite no longer repoints the developer's global editor config. Running the suite could overwrite the real
~/.claudeMCP registration, breaking the developer's own tooling. (#1481)
Documentation
- The documentation layer has its own reference, with the wiki style guidance folded into it. (#1493)
- Claims that contradicted the code were corrected across the reference docs, and the long ones were made navigable. (#1479)
Plugins
- Claude Code plugin at 0.42.0; Codex plugin to 0.5.0 — both skill sets now render from one shared source (#1450).
Internal
- The internal prerelease publish is idempotent, so a re-run no longer fails on an already-published version. (#1494)
[0.41.0] — 2026-08-11
This cycle closes the gap between what an agent can ask over MCP and what it can ask from a terminal. get_answer, get_context, get_symbol and get_why had no CLI form at all, so anyone driving repowise from a shell or a CI job could search the wiki but could not ask it a question, read a triage card, pull one verified symbol body, or ask why the code is shaped the way it is. Those four are commands now. search and risk were rebuilt as thin adapters over the same tool functions, so a CLI answer and an MCP answer are the same answer, and thirteen more commands gained a machine-readable mode under one flag spelling.
The other theme is get_health, taken end to end by following its own advice as a sequence rather than testing calls one at a time. That found seventeen defects across three passes. The single call directive.plan_via told an agent to make returned 60k characters and failed the MCP token cap outright; only projected away the unresolved block, so a mistyped target went silent again; and one documented call shape took thirteen seconds.
Alongside those: importing one MCP tool no longer costs all seventeen, four tool blocks stopped returning payload the caller already had, and two memory regressions on the indexing and generation paths are fixed. Three authentication holes in the local server are closed, including one on a route that had none.
Added
repowise ask,context,symbolandwhy. The four highest-value MCP tools now have CLI forms. They are not reimplementations: a registered tool is a plainasync def f() -> dict, so each command awaits the same function the MCP server serves, over resources built by the existing tool bridge. (#1411)--format table|jsonacross the agent-facing commands. Four flag names used to mean the same thing (--format,--json,--output,--progress) and 54 commands had no machine-readable mode at all. Thirteen gain one:status(single-repo and workspace),costs,saved,corrections,whats-new,decision list/show/health, andcoverage status. Three that already had a machine mode under another name fold onto--format, with the old flag kept but hidden so existing scripts keep working.--progressis an event stream rather than a result payload and is untouched. (#1407, #1402)- Five plugin slash commands for the new adapters:
/repowise:ask,/repowise:context,/repowise:symbol,/repowise:why,/repowise:export. (#1428)
Changed
repowise searchandrepowise riskare adapters oversearch_codebaseandget_risk.searchcarried its own retrieval: an FTS query, a LanceDB query, a LIKE overwiki_symbols, and a workspace fan-out that fused them. The tool already does that and does it better, fusing the full-text and vector legs through RRF instead of picking one, honouring per-repo excludes and tombstones, and routing an identifier-shaped query to the scored symbol index. Both halves of the public contract are preserved, including--mode fulltext|semantic|symbol. (#1412)get_symbolis presented as a follow-up rather than a place to start. Where a short capability list names it besideget_answer, agents spend most of their retrieval calls on it and finish having made more tool calls than an agent with no tools at all, because a per-symbol tool supplements navigation rather than replacing it. Given the same surface with no such list, they reach forget_answerinstead. The tool stays registered and stays in the lean profile; what changed is what four surfaces say about it. (#1427)- Importing one MCP tool no longer imports seventeen. The package imported all 17 tool modules at package import, and through them the health analysis, onboarding generation, FastMCP and sqlalchemy. A leaf import initialises its parent package first, so importing any one tool paid for all of them: about 2.5s, against about 290ms for the entire CLI. Tool modules resolve lazily now, through the module's existing PEP 562
__getattr__. (#1409) - Four tool blocks stopped returning payload the caller already had. A tool result is new text entering an already-cached conversation, so it is billed at the cache-write rate and paid again on every result.
get_contextserves the symbol card again instead of auto-upgrading file targets to a skeleton, which measured 2,171 characters against 6,585 on one mid-size library. Each cut is conditional, so nothing is lost. (#1426) - The parse pool is capped instead of sized from the host's cores. It asked for one spawned worker per CPU. Every worker is a fresh interpreter that imports the repowise stack and builds tree-sitter
Language/Queryobjects, a flat cost of about 50 MB of private memory each, measured at 51.0 MB/worker on PowerToys and 49.2 MB/worker on hugo. Peak memory was set by the machine rather than the work: 32 workers held 1.57 GB where 8 hold 0.46 GB. Nothing was bought with it. With 16 cores free, 8 workers parsed PowerToys faster than 16 did. (#1410)
Fixed
- Seventeen
get_healthdefects, in three passes.directive.plan_vianamed a call that returns 60,296 characters and fails the MCP token cap, and now names the projected form at 13,503.onlyprojected awayunresolvedandknown_modules, so a typo'd target returned an empty list that read as healthy. The performance dimension had no ranking key, the refactoring surface offered clone plans that could not be extracted, and plan ranking clustered on one file. On the cost side,graph_nodeshad no index coveringnode_type, its most-issued predicate, so every "all the file nodes" read scanned 36,480 rows to return 3,449 (new composite index, Alembic 0051). (#1413, #1414, #1415) get_contextresolves the symbol idsget_symbolhands out.get_symbolnormalises the separator between qualified-name segments across.,::and/, since which one a caller writes is a fact about their language.get_contextsplit on::only to gate the file path and then matched verbatim, so the dot form returned "Target not found" and the::form failed worse: it matched the graph-node rung, was typed as a file, and returned a card describing the method as empty. (#1435)- Generation no longer holds a second copy of the repository's source for the whole run. Every
FilePageContextcarriedfile_source_snippet, the decoded source trimmed to a 48k-token budget, which for almost every file is the entire file. No template, prompt or caller ever read it. Contexts stay alive until the run ends, so a large repository could exhaust memory during generation. (#1403, #1394) - Piped and CI output stopped truncating result paths. Rich sizes a non-terminal console at 80 columns, so a pipe got the most aggressive truncation of all while a human in a wide terminal saw everything.
searchrendered paths aspackages/core..., which cannot be opened or grepped. Width is now resolved once where the shared consoles are built.COLUMNSstill wins when set. (#1402) export --format json --fullcrashed with anAttributeError, reading afinding_typeattribute that does not exist onDeadCodeFinding. (#1416)- The chat tool surface matches the registry it actually has. Labels, suggestions, system prompt and artifact contracts still described tools that were removed or were never wired, including
get_architecture_diagramandget_dependency_path, andget_change_riskresults were not summarised. (#1417, #1418, #1419, #1420, #1421, #1423, #1424, #1425)
Security
- Webhook authentication fails closed.
REPOWISE_GITHUB_WEBHOOK_SECRETis required rather than recommended, the secret is read per request, and a request without a signature is refused with 403. (#1408) - The MCP server binds explicitly on the HTTP and SSE transports and warns when an exposed bind has no
REPOWISE_API_KEY. (#1406) - The repo health coordinator route requires an API key. It was reachable off-host without one. (#1405)
Documentation
- Command walkthroughs for
search,decision,wiki-exportand hooks sync. (#1385, #1386, #1387, #1388) - The architecture docs' MCP inventory matches the 11-tool default surface, and the chat tool registry matches the live 7-tool set. (#1173, #1174)
Dependencies
- mermaid 11.16.0 to 11.16.1 clears three advisories at once: prototype pollution in architecture diagrams, an infinite-loop DoS in XY charts, and a DoS in radar diagrams. The declared floor in
packages/uimoves with it so a later resolve cannot land back on a vulnerable build. nanoid and brace-expansion move too, closing two more. (#1433)
[0.40.0] — 2026-08-10
The big one this cycle is a new kind of record. Repowise already told you what the code is and what history says about it; it had no way to say what has happened to a repository: that this directory is a separate checkout, that a console script is shadowed, that this bug was fixed in this commit and nothing in its scope has moved since. Episodes are those records: dated, bound to the files they touch, derived from the walk and the git pass that already run, and checkable against git rather than decayed by a curve. They are read by get_answer, get_why, get_risk and get_context, and served over HTTP for the dashboard.
Alongside it: a keyless index stops ranking on vectors that carry no signal, which was roughly 40% of every search window; the health dashboard ranks by how much a file actually lost rather than by a score that clamps; repowise watch indexes the working tree, so it finally updates anything it watches; an incremental update prunes deleted files and can no longer strand the commits a failed step skipped; and the update path got about ten seconds faster on a large checkout.
Two authentication and file-access holes in the local server are closed. If you have ever run repowise serve on anything other than the default loopback bind, read the Security section first.
Security
repowise serve --host 0.0.0.0no longer serves an unauthenticated API to the network. The auth layer decided whether a request could skip the API key by readingREPOWISE_HOSTto tell a loopback bind from an exposed one, and nothing on the CLI path ever set it. So the server bound the world while auth read "127.0.0.1" and let every request through, and the fail-closed branch for an exposed bind was unreachable code. The decision now comes from the request's peer address, so an exposed server fails closed however it was started and whatever the environment claims. Separately, the bundled web UI bound the wildcard while the API stayed on loopback: the UI rewrites/api/*to the API over localhost, so a UI on0.0.0.0handed the whole API to the network on the UI's port with every proxied request looking local, even at the default--host 127.0.0.1. The UI now takes the API's bind. One residual is known and warned about at startup: with an explicit--host 0.0.0.0and no key configured, that proxy still fronts the API. Setting a key makes the proxied path 401. (#1391)/file-contentserves only files the indexer recorded. Its one guard was containment in the repository root, and.repowise/.env(your provider API keys) and.git/configsit inside that root, so any caller who could reach the endpoint could read them by path. It now applies the same three-table existence testGET /files/{path}already used, and denies.git/and.repowise/outright. Only those two directories are denied by name, because the traverser does walk other dot-paths and.github/workflows/ci.ymlis a legitimately indexed file. Both callers, the web file page and the VS Code architecture view, only ever requested paths from the indexed file list. (#1393)
Added
- Episodes: a dated, scoped record of what has happened to a repository. Three tiers. Structural facts are derived at index time from the walk that already runs, with no history, no transcripts and no API key: nested git repositories, an editable install shadowing a console script, configuration that changes what a command does, a tree that is not formatter-clean (#1317). Git episodes are bug-fix commits whose diff changes production code, each carrying the author's own account, the files that moved together, and a birth commit that makes "does this still hold" a question git can answer (#1324). Session episodes record each agent session in its own words, bound to the files its tool calls touched; they are per-machine and never leave it (#1325). A session outlives the transcript it came from, which agents prune on a schedule nobody sets, and episode bodies are full-text searchable (#1331). (#1325, #1331)
get_answerappends a recorded fact that disagrees with its answer. It reasons from the code as it is now, so asked whether to run the declared formatter before committing, it reads a format target out of the build files and says yes, while the checkout's own record says the tree is not formatter-clean and a repo-wide run would produce a large unrelated diff. A scoped, still-current record now rides alongside the answer as a dated quotation. Three conditions, all required, or the payload is byte-identical to before. (#1318)get_whyanswers "what happened here" from episodes, beside decisions, git archaeology and rationale comments.get_riskandget_contextcarry anepisodescount on their cards. Both surfaces serve only the tiers that describe the repository, so two people asking one question of one repo get one answer. (#1333)- Episodes over HTTP. Four routes under
/api/repos/{repo_id}/episodes(a paged list, grouped counts, a per-file lookup and a detail route), plus the TypeScript contract and API client. Only shareable tiers cross the wire, reading never creates the store, and no list route shells out to git. (#1346) - A decision is bound to the directories it governs, and its staleness is measured rather than modelled. Module linkage was the first path segment of each affected file, so in a
packages/layout almost every record claimed "packages" or "tests". Staleness is now the fraction of a record's files that have been committed to since it was recorded, not a formula fitted to one repository that needed 15 commits in 90 days before it moved at all. Records naming no file are counted separately as unscoped, because they cannot be asked the question. (#1322) - Unchanged re-reads collapse to a notice. On PostToolUse Read, when the same range was served this session, no Edit came between and the bytes hash the same, the payload is replaced by a short line naming the earlier read. Opt-in via
hooks.read_reread, default off. A file whose content differs is served in full plus a line saying it changed on disk and not through an edit in this session, which is something nothing else in the session can discover. (#1362) - A failed path is answered with the file the index says you meant. On PostToolUseFailure, when the basename resolves to exactly one indexed file still on disk, repowise names it. Everything else is silence: an ambiguous basename, a directory target, a path in another checkout, a failure the agent already answered itself. (#1336)
repowise savedreports the net trade, and can report a loss. It counted credits only and structurally could not show one. Debits are priced as they are actually paid: the CLAUDE.md block and the MCP schema are resident, so they are re-read on every API call, and the amplification is measured from real cache reads rather than assumed. Costs that local data cannot compute are named in the output rather than dropped. (#1362)- A Codex session transcript adapter, on the harness-agnostic transcript layer the shared session code was refactored onto. (#1277, #1316)
- Every embedding provider verifies the width of the vectors it gets back. (#1305)
/repowise:securityin the Claude Code plugin, exposing full-history secret scanning. (#1107)repowise watch --index-only, so a watched docs repo does not mean a model call per save. (#1363)
Changed
- The generated CLAUDE.md / AGENTS.md block is about 30% smaller. It is resident in the prompt prefix, so it is re-read on every API call rather than paid once; measured on one local corpus that amplification is 50.4x at a median 299 calls, which made the block the largest single cost repowise imposes on a session. Only the static half is cut. 5,788 of its 10,350 characters were identical in every repository repowise has ever indexed. Every section derived from your tree is untouched, all ten tool rows stay, and so do the argument shapes and the response fields an agent needs to call a tool correctly. (#1382)
- The augment hook no longer fires on shell commands. Across one 287-session corpus, Bash and PowerShell were 8,942 of 17,441 hook invocations (51%) and produced 64 emissions (0.7%). The cost is process start, paid before repowise reads the payload, so no gate inside the handler could avoid it. SessionStart already carries the freshness line; what is given up is a mid-session commit going unflagged until the next session. Installed machines narrow themselves without a re-init. (#1382)
- The skeleton-nudge Read hook is retired. It was the loudest surface in the system and it never changed behaviour: replaying 516 firings across 203 sessions, a structure call followed 11.4% of nudges against an 11.9% unconditioned base rate, and a second exposure read the same as the first. It was also answering the wrong reads. 53.3% of firings followed a ranged Read, which is advice to be more targeted handed to a read that already was. The replacement surface covers the case by serving the skeleton instead of describing it. (#1339)
- The Decisions page says what it actually knows. It led with the size of its review queue, so the largest number on screen pointed at unreviewed guesses. The confirmed count leads now. Conflicts is gone (it rendered from a table holding zero rows on every repository). Most-governed-files is gone (81% of the links it ranked belonged to proposed records, under a heading that read like established governance). Staleness is a count rather than a percentage, and a new "About this checkout" section carries the structural facts, which is the only part of the page that says anything on a repository indexed an hour ago. (#1352)
- Health's
moduleis the package boundary, not a graph community label. The column held two vocabularies with nothing saying which one a row carried, and which one you got depended on which code path last wrote it. 41.5% of this repo's rows named a directory the file is not in. Go, Maven, Ruby, sbt, Clojure and Erlang monorepos get real package attribution for the first time, corrected in place onrepowise updaterather than requiring a re-index. (#1359, #1366) - Health findings split test material into its own bucket. At the default limit a quarter of the most-read list described the test suite.
top_findingsandfindingscarry production findings,test_findingscarries the rest, and the two totals partition the whole open set. Bucketed rather than excluded: a thrashing test suite is a real signal, just not the same question as where the defect risk is. KPIs deliberately keep test files, since tests score better than production code and dropping them would lower every repo's headline with no defect having been found. (#1361) get_health's directive stops promising a fix it does not have. It names one file, states the dominant reason, and pointed at the refactoring block as the way to fix it. Nothing emits a plan for a coverage gap, which is the dominant cause on 10 of the 10 worst files here, so the directive named coverage and returned plans for complexity and duplication. It now carriesplan_addresses_reasonand, when false, a note naming the unaddressed cause. (#1358)repowise healthhonours the config's exclude patterns. (#1366)
Fixed
- A keyless index stops ranking on its own vectors. Keyless mode is documented as full-text-only in four places and the code never implemented it: the no-key embedder's 8-dimensional vectors were fused into rankings at full weight. They carry no signal: every component is non-negative, so two unrelated strings score 0.750 cosine on average and never below 0.207. The consequences were measured: the relevance floor written to stop unrelated pages surfacing was cleared by 100% of them and never fired, about 40% of every
search_codebasetop-5 was noise,get_answer's confidence was corrupted along with its ranking, and decision dedup merged an unrelated new decision into an existing one 100% of the time past 50 records. (#1378) - Two more keyless vector reads, and the confidence grade underneath them.
repowise search --mode semanticbuilt its own store inline, so the first sweep never reached it and it rendered a window of nearest-neighbour noise with the full-text fallback sitting unused below. Separately,get_answer's agreement lift required the full-text and vector legs to concur, which a keyless index can never do, so every keyless answer was graded by the gate that signal exists to correct. The symbol leg becomes the second opinion, on a stricter tie. Also:--embedder ollamaagainst a stopped Ollama now says so instead of indexing silently without semantic retrieval. (#1380) - A keyless wiki stops warning about every page it has. One confidence value meant two things, a page whose provider call failed and the deterministic page a keyless or
--no-proserun renders on purpose, and the reader turned it into a banner telling you to verify the content against the source. On a keyless index that banner landed on the repository overview and every subsystem page, about pages assembled entirely from the parse, the import graph and git history. The two cases are separated, and the reader keys on the failure marker rather than the number, so an already-published wiki reads correctly without a re-index. (#1383) repowise watchindexes the working tree, so it actually updates. It fired an update on every save, and that update diffedlast_sync_commit..HEAD, and on a repo you are actively editing there are no new commits, so the diff was empty by definition and every run printed "Already up to date". The watcher has therefore never indexed anything it watched. Also: events are filtered through the traversal blocklist so a build no longer means thousands of triggers, the files repowise itself rewrites can no longer make each update schedule the next, both ends of an atomic-save rename are considered, updates are serialised, the lock is released between runs, and reverted edits are undone in the index. (#1363)- A config change no longer advances the sync pointer past commits it never indexed.
updatetook the config-changed branch, re-scored health and returned early, skipping the git re-index, then savedlast_sync_commit: head. The next run reported "Already up to date" and those commits' churn, ownership and co-change were lost until a manual--full. (#1360) - An incremental
updateprunes deleted files. Deleting a file tombstoned its wiki page and left everything else: graph nodes and edges, metrics, symbols, health findings, dead-code and security findings, git metadata. MCP, search and the graph kept serving symbols for a file that no longer existed. Liveness is asked of things independent of the parse, so a transient read failure or an antivirus scan cannot look like a deletion, and a prune that would take more than half a table is refused and reported instead of applied. (#1377) - A degraded update no longer strands the commits it skipped. Every persist step degrades into a report, then the pointer moved to head regardless, so a failed step took its commit range with it permanently. A failure now records a repair marker holding the previous pointer, and the next update widens its diff base back to it. (#1377)
- Dead code persists the repo-wide verdict an update already computed. The analysis was always repo-wide; its result was then filtered to the change set, discarding exactly the findings the change had just caused. Drop the last import of a module and that module becomes dead, and it is not in the change set. Widening the write alone would have made things worse, because the analyzer saw no git metadata for unchanged files and an empty record reads as "no commits in 90 days", so the fix also feeds it the stored per-file git fields. A dismissed finding is no longer resurrected. (#1389)
- A cohesive Go package is no longer reported as a circular dependency. Files in one Go package reference each other with no import statement, so several passes synthesise file-level edges to keep reachability honest, and cycle detection read those as dependencies. Never Go-only: every
foo.h/foo.cpair was a guaranteed two-file cycle and C# partial classes formed an all-pairs ring. Cycles before to after: gitleaks 1 to 0, osv-scanner 16 to 0, syft 16 to 0, hugo 10 (largest 571 files) to 0. Existing indexes converge onrepowise update. (#1354) - The worst-files lists rank by how much a file lost, not by a score that clamps at 1.0. Thirty files sit at exactly the floor here, so every list sorted on score came back in path order and this repo's actual worst file landed at position 27, invisible at the default limit. Fixed once in the crud layer, so REST, MCP and hosted inherit one answer. A payload's headline and the list printed directly beneath it also disagreed about which file was worst; both now read one ordering. (#1342)
- A file at the score floor can show movement, and the trend stops hiding improvements. Snapshots persist the clamped value, so 28 of 29 floored files had a flat 1.0 series while carrying 9 to 13 points of deduction. The per-file movement list also sorted ascending and sliced to 50, making it the 50 biggest drops. Also repairs a migration chain that had two revisions declaring the same id. (#1364)
- The health file drawer shows the line on file-level findings. The
function:lineanchor was gated on a function name, so all 1,057error_handlingfindings here rendered with no line and nothing to click: 34 rows on one file, identical in every visible field, because the drawer was withholding the one field that told them apart. A file-level biomarker that repeats now gets its own group, with acappedchip when the scorer is holding that category at its ceiling. (#1344) get_health'sonlyprojection stops dropping the totals.only=["modules"]atlimit=50returned 50 of 116 with nothing saying so, breaking the tool's own promise that truncation is never silent.includeandonlywere also two vocabularies for the same blocks, soonly=["biomarkers"]projected away the block it had just switched on; those are aliases now.limit=0means none rather than clamping up to one. (#1361)get_healthstops returning 4.7MB.include=["biomarkers"]with no targets was the one ranked list with no cap, serving all 10,349 open findings, enough to overflow an agent's context and return nothing usable. Capped, withfindings_totalalongside. The dimension filter also ran over the finished response, soinclude=["biomarkers","performance"]came back empty while the total reported everything, which reads as "no performance risk in this repo". (#1337)- The coverage tab's badge no longer reports one module. A single
limitwas capping two unrelated things, a page of files and the repo's covered-directory count, so the badge, which asks for one row because it wants a cheap response, was told the repo had one module. A?file_path=request had the same defect from the other side. (#1356) - The code-health map's churn lens paints the data it has. It requested churn with no limit, so the server's default of 300 applied and about 1,700 of 2,000 nodes rendered as the no-data swatch. (#1342)
hook stats,hook rewrite status,doctorand thesavedtip report what a surface is doing, not what it once did. Retired surfaces kept computing rates over denominators that had stopped growing and rendered identically to live ones; they are labelled and dimmed, with aretiredflag in--json. "Installed" was keyed on the hook command, so an entry whose matcher names a tool the agent has since renamed was reported as working; presence and reach are separate answers now and the verdict is three-state.repowise savedpriced tokens at a hardcoded model while the costs endpoint priced the same ledger at the detected one, understating an Opus session by two thirds. (#1343)- An injected decision nothing could disagree with no longer counts as followed. "Followed" was the else branch of the contradiction test, so a session that mined no correction handed out a positive verdict for free: 100 followed, 0 contradicted, and none of the 100 from a session holding any correction. Those rows settle with no verdict now. The supply side had the matching hole. The correction gate only fired on a message that opened with a pushback lead, finding 62 corrections across 436 transcripts against 260 for the same words at sentence start. (#1335)
- A marker decision is verified against its own span. Every marker in a file had its context joined into one blob handed to every decision drawn from that file, so a decision from marker 1 could be stamped
exactagainst marker 3's words, and all of them inherited the first marker's line. Staleness was also scored against one run's change set, so on an incremental update decisions over untouched files went maximally stale for not having been touched: 13 of 19 rows here read 1.00 and now 1 does. Existing rows correct themselves on the next update. (#1315) - Three refactoring-plan fields stop reporting constants as measurements.
suggested_namewas unconditionally null on all 854 storedextract_methodplans while a sibling detector computed one;blast_radiuswas a hardcoded{"callers_count": 0}indistinguishable from a measured zero; andsuggested_sitemixed a graph community label with a filesystem path under generic keys, which is how a helper got named after the repository itself. (#1374) get_whyno longer wedges the session. It never returned on a repo with git-tier episodes: 300s with no response and onegit rev-liststill running, with the agent's whole MCP session stuck behind it. Three git spawns did not passstdin=DEVNULL, so under the stdio transport the child inherits the server's JSON-RPC stdin and can hold it open indefinitely, and thetimeout=on each call is not the backstop it looks like, since on Windows the re-enteredcommunicate()blocks on threads the killed child still holds. (#1395)doctorstops reporting resume stubs as drift. A stub awaiting--resumeis held out of the vector store deliberately, because the vector is what tells the nextinit --resumethe page is done, so counting it as a missing vector pointed people at--repair, which embedded the stub and silently burned the retry. It is surfaced on the row as "N stub(s) awaiting --resume" rather than hidden behind "in sync". Alongside it:initsuggestedrepowise decisions, which is not a command, and blamed "provider failures" for pages its own artifact check had rejected;dead-codeprinted a raw dict; andreindexrelabelled its progress bar "Indexing decisions..." on a repo with no decisions. (#1395)statusanddoctorquery the configured database when there is no localwiki.db. (#1320)- A transient PyPI failure no longer pins the update check to "unknown" for 24 hours. The failed result was written into the TTL cache, and the next call inside the window saw a cached
Noneand skipped the live fetch. (#1330) - Resume no longer redoes finished phases on a repo with more than 100 pipeline jobs. The completed-phase readers took the default row limit, so the set was silently truncated and prior git-history, centrality and decision-extraction work was repeated. (#1329)
- The Read line count is correct for output ending in a newline, which had been overstating by one and feeding both the nudge gate and the served-read windows. (#1328)
get_riskandsearch_codebasefit their schema budget again. (#1334)
Performance
- The update prologue stops paying four repeated costs. Assessing the store no longer imports lancedb on a keyless repo (1.517s to 0.036s), the tech-stack scan is memoized on the root manifests' stat signature instead of running twice per update, the deterministic page persist no longer rebuilds a page tree its caller rebuilds immediately after, and
initstamps the re-score cadence it satisfied so the first update does not score every file again. A post-commit hook runs this path on every commit, so all of it was charged per commit. (#1376) - Lifetime churn folds onto an anchor instead of re-walking history. Every update ran
git log --shortstatover the entire history to move a number by one commit's worth of lines: 2.5s on a 2,216-commit checkout, 13.3s on a 9,335-commit one, growing with every commit anyone makes. Now 0.34s and 0.19s, byte-identical output, with a full re-walk every 100 commits to bound the drift that no ancestry check can see. Also fixes a timestamp parser that made a related backfill re-select the same rows forever, because strict ISO 8601 spells a zero offsetZand only+00:00was understood. (#1379) - An update stops rewriting the graph rows it did not change. A one-file change arrived as tens of thousands of rewrites of rows that had not moved: 1.9s on a 15k-node checkout, 11.3s on a 63k-node one. Centrality values compare with a tolerance, because the kernels are not bit-stable across processes and exact equality would rewrite every node to store noise. End to end on a real store, 2.10s to 0.48s and 12.73s to 8.81s, with row-for-row parity. (#1381)
- Dead-code analysis is four times faster. All 579 never-flag globs were joined into one alternation with every branch tried at position 0, most beginning
.*, which was 12.8s of an 18.1s cold analysis. Bucketing patterns by their end-anchored tail takes hugo from 3.10s to 0.72s and PowerToys from 25.29s to 5.71s, with identical findings. (#1389) - The public health badge stops reading the whole dataset to render one float: 358.2ms to 19.8ms.
wiki_symbolsalso gained an index able to seek by file, taking a 400-path lookup from 33.3ms to 11.7ms. (#1351) /health/overviewstops reading both health tables twice (554.3ms to 376.4ms, and 2.2MB of snapshot JSON to zero), and/overview-summarystops loading the full retained snapshot history as entities to read three scalars off it. Both fixed a real defect on the way: the overview summary reported one recent fix for every flagged file, disagreeing with the health dashboard on the same page. (#1350, #1356)/health/refactoring-targetsstops shipping every file's findings. It built and serialized findings for all 2,411 files before slicing to the limit, then discarded about 90%: 1.16MB to 190KB at the limit the UI sends. Both consumers sit behind a click and now fetch on demand. (#1345)- Query-time exclusion is compiled once per repo instead of per row, 1,442ms to 3ms warm on a dashboard call, on a path shared by every MCP tool.
health_findingsgained the index it never had, taking a file-scoped lookup from 8.8ms to 0.1ms. (#1337) - A Glob that ripgrep gave up on is answered from the index instead of costing twenty seconds and returning nothing. (#1362)
- Read state is partitioned per session, which fixes every Read hook surface being near-useless in subagents: a sidechain runs under a different session id from its parent, and the per-repo state file was keyed on session id and reset whenever it changed, so parent and subagent wiped each other's state on every Read. (#1362)
Documentation
- The README leads with the numbers. Its strongest evidence was in its smallest text, under a headline that was a feature list. It now opens with the three results at size: first of six on finding the right files (0.876 against the next tool's 0.610 on a sealed 42-instance split), 31.6% fewer output tokens from the agent, and 97% fewer tokens to load a commit, plus the line that keeps what the strip cannot hold, including that we are the slowest indexer in the field. BENCHMARKS.md opens with a scoreboard and charts instead of forty-five lines of essay about why the page is built the way it is; the essay moves to the end as the closing argument. Nothing that runs against us was cut to make room. (#1365)
- The benchmarks are rerun on a second and third agent harness, and one claim is retracted. The page read the "agent used it" column as a design result about how well we had named our tools. It does not support that: repeating an identical Claude Code run with nothing changed on anyone's side moved us from 15 of 15 to 3 of 15, and a competitor from 13 of 15 to 2 of 14, while switching harness moved every tool in the field to 15 of 15. Adoption is a joint property of tool, harness and day, so every figure now travels with its harness and its date. Section 2 leads with the full 48-question Codex run, where every tool actually gets called. (#1319, #1321, #1340)
- The code-health marker taxonomy is corrected. The docs said 25 markers; the registry has 49 detectors across 52 marker ids, of which 26 feed the defect score, 20 the performance pillar, 3 maintainability only and 3 move no score at all. CODE_HEALTH.md is restructured from a reference manual into an argument, with a dozen figures corrected against live code and the bench data. (#1338)
- The website's CLI and Claude-plugin pages match the shipped surface, and a change-risk CLI walkthrough is added to the examples. (#1106, #1177)
[0.39.0] — 2026-08-05
A small release. get_answer stops handing back bare file paths, the docs sidebar reads as an outline instead of a directory listing, and a store upgraded from 0.37.0 opens again instead of taking the whole wiki down with its search index.
Added
get_answernames what each candidate file defines, not just its path. Of 499 paths served across the 25 flow questions, 65 carried any content at all, and 89% of the agent's post-answer searches were the same shape: take a name the payload gave, go fetch the substance it did not attach. Each candidate now carries its declarations asname:linepairs, question-named symbols first, imports and private names dropped. Substance per served path goes from 0.130 to 0.864, for 10% of the response and no change to which paths are served or in what order. (#1306)
Changed
- The docs sidebar reads as a table of contents. Every group on the top rung opens on load and nothing below it does, so a chapter that parents sub-chapters no longer expands wherever it sits and makes the tree read as a filesystem. The four-digit file corpus closes the tree instead of fronting it, selecting a page opens the chapters above it, and outline rows drop the folder glyph the no-icons rule had never reached. (#1312)
- Mermaid diagrams scale to the column they are read in. They rendered at natural size inside a scroll box, so the architecture map arrived clipped mid-subgraph behind two scrollbars. Width only, never above 1:1, with the scaled height reserved so a fitted diagram does not sit in a pool of empty space. Maximize still pans and zooms for the dense ones. (#1312)
Fixed
- A store upgraded from 0.37.0 no longer refuses to open. 0.38.0 widened
page_ftsfrom three columns to five, and FTS5 cannot be altered, so the index is dropped and refilled fromwiki_pagesthe first time an upgraded install opens the store. That rebuild refused whenever the index held more rows thanwiki_pagescould account for, and the error it raised told the reader to runrepowise doctor --repair— which opens the store the same way, hit the same refusal, and died before repairing anything.serveand the MCP server died there too, so a search index took the entire wiki down with it. The excess rows are orphans: pages swept from SQL whose index delete never ran, because it runs after the commit, outside the transaction, on a best-effort path, and nothing ever reconciled them. An orphan answers a query in full and 404s when the reader opens it, so the rebuild discards it and reports the count rather than refusing. (#1309) - An interrupted sweep heals itself.
ensure_indexprunes orphaned index rows on every open, which is the only thing that ever reconciled the two halves of the store; the residue could otherwise only grow. The scan is read-only and takes the write lock only when there is something to delete. doctor --repairfinishes what it was called for. A failing full-text schema upgrade is reported and stepped over instead of aborting the repair, and orphans are deleted in one transaction rather than one per id.serveand the MCP server start on a store whose index cannot be prepared, falling back to whatever shape the index is in plus the vector arm, with a warning. Every other surface — the wiki, the graph, code health — was unreachable over a keyword index.
Documentation
- The benchmarks are rescored on the sealed half after the retrieval gate fixes, and the token-efficiency figures are refreshed. The agent-loop result is reported as its own number rather than folded into the payload one, which measured a different thing. (#1307, #1308, #1310)
Dependencies
- aiohttp 3.14.1 → 3.14.3, cryptography 48.0.1 → 50.0.0, gitpython 3.1.55 → 3.1.58, postcss 8.5.22 → 8.5.25, and the dev-only undici 7.28.0 → 7.29.0 and fast-uri 3.1.4 → 3.1.5. Clears the open security advisories against the tree. The cryptography floor moves rather than only its ceiling, so a fresh install of the published wheel cannot resolve back into the vulnerable range. (#1313)
[0.38.0] — 2026-08-04
Four new languages, an orientation set rebuilt around what a reader actually needs, and a wiki that draws on the repository's own vocabulary instead of writing generic prose about it. Svelte and Vue reach the Full tier through a byte-preserving projection into TypeScript; HTML lands at the import tier; reStructuredText documents are read as reStructuredText rather than silently yielding nothing. Onboarding is six pages now, ending in a glossary built entirely from mined terms, and a directory that heads a subsystem gets a chapter even when it also holds files of its own. Full-text search was rebuilt on both backends after the query shape turned out to match 65% of the corpus on a median question. The agent hooks got measurement first and then acted on it: the Grep flood is replaced by its digest instead of ranked next to it, triage ranks the files the search actually matched, and the three hot index lookups dropped an ORM import that cost a second per hook fire.
Added
- Svelte and Vue at the Full tier. A markup grammar locates the JavaScript-bearing regions of a single-file component, every other byte is blanked to a space with newlines preserved, and each markup expression is fenced by rewriting its two surrounding delimiter bytes. The result is valid TypeScript at byte-identical offsets, so the TypeScript queries, language config and all three health dialects apply unchanged; only region location differs per language. Components reach the file tree, git history, wiki and health for the first time. (#1221, #1232)
- HTML at the import tier, via
script srcandlink href. HTML has no functions, classes or calls, so this ships an import-only tier and says so on every surface rather than minting symbols that do not exist. Template dialects are out of scope and that is stated:{% extends %}is plain text to an HTML grammar, and 744 of 749 measured dialect files produce no edges. (#1235) - reStructuredText is read as reStructuredText. Every pattern in the document miner was markdown-only, so a repository that documents itself in
.rstlooked identical to one that says nothing. flask, requests and django were all in that position. Underline-length, directive bodies and roles are each handled, and.rstprose written under a.txtextension is picked up too. (#1238, #1247) - A glossary page, written entirely in the repository's own words. House vocabulary (blast radius, change risk, co-change, distill) had no definition anywhere a reader could reach. Each row is a term, the repository's own defining sentence, where it is used and which document it was written in. The page has no model in its path at all: every cell is a fact the run already holds, so it costs no tokens and cannot hallucinate. (#1276)
- Subject chapters. A parent directory was disqualified from heading its children whenever it also held loose files of its own, because both pages would collide on one page id. That excluded exactly the directories a reader most wants a chapter for, and did it silently: nine of thirteen chapters suppressed on this repository, five of six on django. (#1282)
- Perf dialects for Kotlin and C++, and a dataflow def/use dialect for C++, which moves C++ to the Full tier with intra-procedural CFG, reaching definitions and Extract Method suggestions. C++ deliberately omits three markers other languages carry, each of which would be a guaranteed false positive there. (#1224, #1225)
- A page says how far it can be trusted.
confidencehad been a constant 1.0 on every page ever generated, so the reader's low-confidence banner had never rendered for anyone and retrieval could not weight by it. A wiki where a provider outage left hundreds of structural stubs looked exactly as trustworthy as a complete one. (#1213) - File pages and symbol spotlights name the questions they answer, built from structure rather than prose, and emitted only where the page can actually answer them. (#1209, #1210)
- Hook efficacy measurement. Four of five hook surfaces wrote rows nobody read, and the Read surface wrote none at all, so a nudge could fire 500 times at a 0.2% action rate with nothing in the product noticing. Transcripts are replayed to pair each emission with the tool calls that followed it. (#1272)
- Vocabulary mining with provenance. Mined terms now carry their defining sentence, the document they were read from and every document that names them, which is what the overview, onboarding, key concepts and the glossary are built on. (#1233, #1241, #1242, #1249)
get_answermeters what one synthesis call costs, and configurable synthesis evidence lets a deployment choose how much the model is shown. (#1180, #1227)- Retired page ids keep resolving, and a retired page can hand off to the repository overview instead of dead-ending. (#1163, #1166)
- The OpenAI embedder honours a configured output width, for deployments pinning a narrower vector than the model's default. (#1254)
updatenames the orientation pages an index has never been offered and points at--full, which is the path that can actually deliver them. An incremental run reads a changed-file slice, and every onboarding gate reads whole-repo signals. (#1288)
Changed
- The docs tree opens on the shape of the repository, not its contents. Layers used to open by default, which was fine when a layer held a handful of children and wrong once layers grouped every module: roughly ninety module rows on the first screen, burying the layer names and the chapters alike. Layers start closed, the file corpus sits above the layer outline rather than under it, and an unclaimed module no longer poses as a layer. (#1184, #1185, #1186, #1187)
- A search hit's snippet is centred on what the query matched. It used to be the first 200 characters of the page, which on a generated page is the same
## Overviewopener every time: identical across thousands of hits, and never the passage that matched. (#1191) - The Grep flood is served as its digest, not ranked beside it. Ranking a flood you also keep is a lens, not a saving; the digest was being added next to output the agent had already been billed for. It replaces the flood now, the same trade
distillmakes for shell output. Measured over real Grep payloads in 25 transcripts, the digest is 0.30 of the flood. (#1283) - Grep triage ranks the files the search actually matched. It built candidates from name and path matches ranked by PageRank without ever reading the grep output; replayed over 1,899 real Grep calls, 83 of the 111 files it named were not in the grep results at all. (#1296)
- The Read hook serves the skeleton instead of recommending it. (#1275)
distillrewrites safe command chains instead of bailing on their shape.repowise saved --missedreported 138,827 tokens over 478 runs in 7 days that never reached distill. The binding gate was a re-quoting rule that refused any command containing a quote, a dollar sign or a backslash, sogrep -n "a\|b" f.py | headbailed on the quotes it obviously contains. What--missedcounts is corrected alongside. (#1291)- Module pages name their own symbols and stop opening with the same sentence as every other module page. (#1211, #1212)
- The overview carries the architecture map, counts its packages rather than describing them, and says what the repository does in the repository's own words. (#1164, #1246, #1249)
- One recipe for every page vector, with a page below the information floor getting no vector at all and pages that lose text at the embedding cap named rather than silently truncated. (#1200, #1192, #1203)
- Running one CLI command stops paying for the whole import graph. Three module-level imports were charging their dependency tree to every invocation, including every hook fire:
repowise --versiondrops from 1,240ms to 150ms, and a silent hook from 965ms to 167ms. (#1273) - The three hot index lookups read through stdlib
sqlite3. Reaching the index from a hook cost a second, 95% of it a single import that none of the three plain SELECTs needed. (#1297) - A command no longer waits out its own telemetry POST. (#1286)
- Decisions rank a person above a document, two sources that never landed are retired, and
get_whypath mode is bounded. (#1290, #1293) - Layers group the docs tree without needing a page to hang off, and pages carry the provenance of the layer that groups them. (#1165, #1170, #1171)
Fixed
- Full-text search was asking for most of the corpus. Both backends built a MATCH expression that could not retrieve, failing in opposite directions: SQLite OR-ed every token with a prefix wildcard, so on a 3,678-page corpus the median question matched 65% of it and one matched everything; PostgreSQL handed the raw question to
plainto_tsquery, which ANDs every lexeme and therefore matched almost nothing. A tombstoned page is also dropped from the index now, and a page can be too thin to be worth indexing. (#1188, #1190, #1198) - The served skeleton never reached the agent.
updatedToolOutputis validated against the schema of the tool being replaced; the hook emitted a bare string where Read's output is an object, so Claude Code rejected it, used the original file, and the hook went on recording a saving. Every firing since the feature landed was a no-op that reported success. (#1278) - The answer prompt formatter halved every page excerpt it fetched, page content is attached on every retrieval rather than only the weak ones, and decision vectors and pageless ids stay out of the answer. (#1168, #1169, #1183)
- The first MCP tool call raced the lancedb import, and the embedder API key is resolved from persisted config rather than the environment alone. (#1230, #1231)
- The data-shape fast path reads the question, not the whole paste, and the early returns hand back the ranked pool they already hold. (#1284, #1289)
- Generation rejects pages that talk to the prompter, refuses a structurally-keyed page with no key, drops sections a page cannot fill, tombstones a page whose file is gone, grounds flow narratives in exact source, and reads execution flows by the field names they have. A definition is taken to be prose the author wrote, not the markup below it. (#1172, #1201, #1202, #1206, #1207, #1208, #1229, #1245, #1248)
- JS/TS extraction picks up unparenthesized single-parameter arrow functions, and HTML intrinsic elements no longer pose as JSX component call targets. (#1215, #1217)
- Python aliased imports parse correctly, and a Node.js package
exportswildcard may cross directory boundaries. (#1243, #1256) doctorreconciles the store against the database again. (#1196)- The Python perf dialect knows
pathlibis filesystem I/O. (#1269) - The betweenness pool is bounded, live update locks are kept, and a failed page is counted once. (#1262)
- Generation checks reach a normal run of
initandupdate. (#1178) - Docker:
/datais created beforechown, the image moves tonode:20-bookworm-slimfor glibc compatibility, and.gitattributesenforces LF on.shfiles so the entrypoint runs on a Windows checkout. (#1266, #1268, #1270) - The Stats punch card's UTC footnote describes rather than prescribes. (#1240)
Documentation
- The README carries media that renders on GitHub, a section and a picture for the PR bot, and cites the 21-repo health validation. (#1197, #1205, #1218)
- Published measured results with a comparison against real peers. (#1287)
- A
structurizrexport walkthrough in the examples and in the CLI package README. (#1175, #1176) - CONTRIBUTING documents how to claim an issue. (#1263)
- The vocabulary overlap thresholds record what they actually measure. (#1181)
Dependencies
tree-sitter-svelteandtree-sitter-htmlare new, for Svelte and Vue respectively. There is notree-sitter-vueon PyPI, andtree-sitter-htmlparses a Vue single-file component cleanly because<template>,<script>and<style>are ordinary elements to it.
[0.37.0] — 2026-07-29
The release where the web UI got taken apart and put back together. Fourteen surfaces moved onto one design language: sections and hairlines instead of a grid of near-identical bordered cards, a sentence above every figure saying what the figure means, and a header row plus a key row on the pages whose canvas is the page. Overview, Docs, Commits, Contributors, Code Health, Coverage, Dead code, Chat, Settings, Decisions, Stats, Files, Refactoring, Knowledge Graph and Architecture all changed shape. The Architecture and Knowledge Graph canvases also got their marks named and their per-frame cost cut, and the docs page stopped downloading 38 MB of page bodies to draw a tree. Away from the UI there is a Structurizr DSL export, a first-run pass over interactive init, refreshed provider model defaults, and three MCP response fixes.
Added
- A Structurizr DSL export.
repowise export --format structurizrandGET /api/graph/{id}/c4/structurizremit a Structurizr DSL model of the repository, with a download button on the Knowledge Graph page.--standalonewraps the model in aworkspaceblock so the file opens on structurizr.com as-is, and--componentsincludes the component level. Built on a new typed node-id module so ids carry their kind instead of being parsed by convention. (#1143, #1144) repowise init --no-editor-setupskips global MCP and hook registration, for CI, agents and throwaway installs that should not rewrite~/.claude/settings.json.REPOWISE_SKIP_EDITOR_SETUP=1is the same switch. (#1086)fields=summaryonGET /api/pages, which dropscontentandmetadataand returns acontent_charscount in their place.fields=fullremains the default, so existing callers are unaffected. (#1120)- A key row on the Knowledge Graph. The arrows, the role dot and the health dot were three kinds of mark the surface never explained. All three are named now, the two dots no longer collide on colour, and a mislabelled verb in the backend map is corrected. (#1145)
--verboseand richer failure reporting oninit. A run that generated some pages and failed others said it succeeded; it now names the failures and points at--resume. (#1094, #1098)
Changed
- The web UI runs on one design language. Cards at uniform weight are gone from Overview, Docs, Commits and Contributors, the contributor profile and commit detail, Code Health, Coverage, Dead code, Chat, Settings and Decisions. Sections and hairlines carry the hierarchy, a lede sentence carries the meaning of the number under it, and the dark surface ramp starts lower so reading columns sit on the base plane instead of in a trench between two lighter walls. (#1114, #1118, #1121, #1125, #1129, #1132, #1134)
- The Architecture page steers one axis from one control. Scope used to live in both the tab strip and a floating pill cluster, with "Communities" appearing twice on one screen. Tabs are datasets now (Map, Coupling, Packages, Symbols) and scope is a single labelled control in the section header. (#1146, #1150)
- The Refactoring page leads with what the pile is. The priority-by-effort quadrant plotted 120 of 1,819 plans, because 89% of plans rate small effort and the X axis was one column. It is replaced with charts whose axes actually spread, under a sentence that says what the plans are. (#1151)
- The Files page opens on the map. The treemap is the subject, so it goes first, with the four figures riding in the sentence beneath it rather than in a strip of equal-weight stat boxes above. (#1152)
- The Stats page shows what no other page shows, and the coding-rhythm punch card is reworked: marginal totals, a colour scale that means something, and no dead gutter on a wide viewport. On this repo the old chart made 853 commits look idle because every cell it could draw capped in the low tens. (#1091, #1096)
- Provider model defaults refreshed. Several shipped model ids had drifted from what the vendors serve, including a default for the recommended provider that was not a real id and only worked via the live
/modelsfallback. The Anthropic and OpenRouter defaults were flagship-priced, contradicting the calibration guidance. Cost and temperature handling is corrected alongside. (#1137) - A first-run pass over interactive
init. A read-only walkthrough found 52 places where the terminal was confusing, contradictory or invisible; 49 are applied, plus five defects found reviewing the result. Phase reporting is one system now rather than two glued together. (#1139) - Documentation generation enforces completion and filters the file dependency context it passes to the model. (#1010, #1011)
- Python lint is gated in CI. Ruff is configured to match how the repo is actually written, 508 accumulated violations are cleared, and lint runs on every PR. Types and UI tests do too. (#1101, #1161)
Fixed
- The docs page loaded every page's body to draw a tree. Opening
/repos/[id]/docson this repo meant twelve sequential round trips and 38.6 MB across 5,485 pages, none of whose text the tree renders. The four things that read bodies off that list are each served directly now. (#1120) - The architecture graph spent seconds arriving and kept paying per frame. The minimap redrew every node on
afterRender;getComputedStyleran 3,022 times on the mount colour pass and again on every theme flip; ELK wrote 3,000 graphology events for one position update. Measured on this repo's capped export (1,500 nodes, 4,107 edges), those are 0, 36, 12 and 1. Dimming also blended toward the wrong colour. (#1149) - The graph stopped blaming the backend mid-load. An error state rendered while data was still arriving. (#1146)
get_healthdropped targets it could not resolve and repeated itself, so an agent could not tell "not indexed" from "no findings". Nothing about scoring changed; this is the MCP response surface. (#1142)get_answersynthesis was capped at a flat 30 seconds, a budget sized for a remote API. Providers that shell out to a coding-agent CLI or drive a local model need 40 to 120 seconds for one turn, and those are exactly the providersinitoffers users with no API key. The budget is per provider now. (#1119, #1124)get_answerreports an empty completion instead of shipping a blank answer as if it were one. (#1126)- A
--resumerun deleted pages it had just protected. Pages skipped because they already existed were absent from the produced set, which is what the stale sweep treats as deleted. (#1089, #1097) - A page whose provider call failed is kept rather than dropped from the wiki. (#1104)
- The augment hook is silent when its console script is missing instead of printing an error into every agent turn. (#1141)
repowise --helpno longer imports the generation pipeline. One module-scope import in the generate engine pulled the orchestrator and persist layer into every invocation, including post-commit hooks. (#1140)- Malformed
.gitignorepatterns are tolerated rather than failing ingestion. (#1095) - Command palette answers a single keypress. It used to need two. (#1157)
- Nav links land where their label says. (#1158)
- The docs reader shows one breadcrumb, computes its height correctly, and wraps long backlinks. (#1156)
- Dead code: the default
min_confidencefloor aligns withRISK_CAP_CONFIDENCE, and the zombie-package detector reads git metadata withdict.get()instead of assuming the key. (#1084, #1087, #1128) - Coupling keeps
?focus=in sync with the pinned file. (#1088) - Pages parented to the repo root use the stored concept tree. (#1081, #1085)
- Server jobs fail explicitly on an unknown execution mode rather than silently. (#881, #911)
Documentation
- Contributors are onboarded through repowise itself, and the README badge row is trimmed. (#1110)
- Website pages resynced with the shipped surface: the MCP pages match the live default tool set, the language list matches the 16 AST languages, and the obsolete Anthropic Message Batches /
--no-batchmaterial is gone. (#1122, #1135, #1136) - New walkthroughs for
distill/expand/ saved output, the security scan, health coverage, and OpenCode provider setup. (#1020, #1021, #1109, #1123) - The commercial page acknowledges the OSS full-history secret scan. (#1105)
- Plugin: Claude Code and Codex skills and commands updated for the
--no-editor-setupflag, the dead-code confidence floor, theget_healthresponse shape and the augment hook change.
Internal
- One home for "is this path a test?" The last inline test-path checks in core and the server are converted to the shared rules, and
.mts/.ctstests are classified correctly incoverage_gradientandmove_method. (#1103, #1111, #1112, #1113, #1115) - CLI test coverage for
repowise decision list/show/confirm/dismiss/healthand thesecurity scanstub. (#1108, #1138)
[0.36.0] — 2026-07-25
A smaller release that settles what 0.35.0 started. Wikis on very large repositories now bound their own file-page volume by default instead of emitting one page per file, distill learns a real shell lexer and gains install and infra command families, and a run of contributor fixes lands across serve, the job scheduler and dead code.
Added
- Install and infra command families for
distill.pip,uv,poetry,npm,pnpm,yarn,cargoandbrewoutput, plusterraform,tofuandhelmplans, are now compacted by data-driven TOML filters. A new family is a data file rather than a Python module, and the errors-first invariant is enforced mechanically so no filter can drop an error line. (#1067) - A shared shell lexer behind the rewrite hook. The hook used to scan for shell characters and could not tell a pipe from the same character inside a quoted string, so it refused anything with both a quote and punctuation. A single-pass tokenizer replaces that scan, which makes the bailouts structural and widens the safe final pipeline stage beyond
headandtail. (#1070)
Changed
- File pages are rationed on very large repositories, by default. A 10.8k-file monorepo produced 8,756 file pages inside a 14,027-page wiki, roughly 77 MB whose tail mostly restates the concept page above it.
max_file_pagesnow scales with repo size using two thresholds derived from the size distribution of indexed repositories: an interactive run offers a leaner wiki past 2,000 files, and an automatic ceiling of 4,500 applies to runs that never get asked, such as--yes, agent and CI runs. Smaller repos are unaffected. (#1074, #1076) - Symbol spotlight pages take the top decile, not the top fifth. That bucket restates what a symbol's file page already renders, and at 4,996 pages it buried the pages that say something new. The bucket still floors at one, so small repos keep a spotlight. (#1069)
- A tidier docs nav tree. Repeated folder glyphs are gone from concept rows where indentation already said it, top-level sections carry more weight, and a leading heading that just repeats the page title is stripped in the reader. (#1066)
- The health files API reports a repository's real file total rather than however many rows the current page happened to carry, while still accepting the older bare-array response. (#1080)
Removed
--safe-onlyonrepowise health, which never did anything. The flag remains live and functional onrepowise dead-code. (#1027)
Fixed
- Command injection when
distillrejoined argv on Windows. The rendered command was quoted for the C runtime's argv parser rather than forcmd.exe, so a token carrying a metacharacter and no space, such as--grep=a&whoami, came back unquoted and the shell read the&as a separator. Verified by execution before the fix:&and|both ran injected commands and>created a file. Every"&|<>^()is now caret-escaped. (#1071) - The getting-started page crashed on any repo with a dependency. The context dict never carried the
versionkey its template reads, and page rendering runs underStrictUndefined, so a missing key raised instead of being falsy. (#1068) repowise servereclaims its preferred ports on restart. The pre-flight probe bound withoutSO_REUSEADDR, so a port still inTIME_WAITfrom the just-exited instance read as busy and the API silently moved to 7338 while the UI kept proxying to 7337. The result was a healthy server rendering as empty panels. The probe now matches the bind semantics uvicorn and Next.js actually use, on POSIX only, since Windows treats the option differently. (#840, #845)- Persisted polling sync jobs never launched. The fallback path discarded the job returned by
upsert_generation_joband then referenced an undefined name, and the resulting error was swallowed by an outer handler, so the failure was silent. (#831) - Subsystem pages no longer print "Questions this page answers" twice. The requirement was stated in both the system prompt and the template, and under the doubled instruction the model emitted the heading twice on 86 of 92 measured pages. (#1073)
- Top-level
export constin TypeScript and JavaScript is evaluated for dead code instead of being exempted the way class members are. (#1065) - Alias resolution in
generate_refactoring_code, which passed a redundant argument inside an already alias-scoped session. (#884)
Documentation
- The docs lead with interactive
init. Getting-started paths handed new usersrepowise init --index-only -yand never mentioned that a bareinitscans the repo and offers a choice before spending anything. Every path now says--no-prose, since--index-onlyhas been a deprecated hidden alias since #1032. (#1075) - Live badges on the README. We ask other projects to carry a Repowise badge and had none of our own. Both read live endpoints, so they follow each index rather than freezing at whatever was true when someone pasted them in. (#1079)
[0.35.0] — 2026-07-24
This release rebuilds how wikis are generated. Every page now renders from structure with no key and no spend, except a single model-written subsystem layer; a new repowise generate command fills that layer on demand, and the web UI can generate AI docs page by page. Stores built before this change are recommended (never forced) to re-index for the newer navigable wiki.
Added
repowise generate, scoped and cost-gated wiki generation. A new command writes or refills the model-written subsystem pages on demand, gated behind a single cost question, and onboards a provider the same wayinitdoes when a docs run needs one. It is also exposed over HTTP for the web UI. (#978, #982, #983, #1060)- Keyless, deterministic wikis.
init(andinit --index-only/--no-prose) produces a complete, honest wiki with no API key and no spend: file, symbol, API, infra, cycle and layer pages all render from structure.init --docsopts into the model-written subsystem prose, and a template wiki renders when no provider is configured. (#972, #975, #976, #999, #1001, #1003) - AI docs from the web reader. The web UI gained provider-key settings and per-page and bulk AI documentation generation from the reader, with a coverage picker and provenance surfaces, streamed over an authenticated job channel. (#988, #991, #994, #984)
- Full git-history secret scanning.
repowise security scan --historyscans the entire git history for secrets, not just the working tree. (#821) - Hybrid
search_codebase. Full-text and vector results are now fused with reciprocal rank fusion, so search blends exact-term and semantic matches instead of picking one. (#1057) - Interactive change-coupling page, surfaced on the Overview. (#986)
- Fix history in the file UI. The symbols view marks which symbols in a file keep getting fixed, and a file's governing decisions moved into their own tab. (#960, #966, #968, #977)
- Dead-code page rebuilt on a shared table primitive with one confidence scale, a reopen path, and per-finding AI prompts, links and reasons. (#990, #992, #993)
- Disable reasoning on Ollama models for providers that support it. (#1009)
- Adaptive TSX grammar fallback so
.tsfiles that contain JSX still parse. (#967)
Changed
- One wiki, one renderer per page type. The old template-vs-model split is gone. Every page except the subsystem layer is rendered from structure, always, with no key and no spend: file, symbol, API, infra, cycle and layer pages. Above them sits a derived, numbered concept tree (the
module_pagetype) that is the one model-written layer: model prose when a provider is configured, structural stubs otherwise.repowise generatefills or refills that subsystem prose. (#1023, #1024, #1025, #1032) - A re-index is recommended after upgrading. Stores built before the subsystem-page restructure carry per-directory pages and no navigable tree, which no automatic reconcile can rebuild. Such stores are now told (never forced) that a full re-index would give them the newer wiki; ordinary updates keep working and the notice surfaces at most once. (#1035)
--prose/--no-proseis the single wiki-spend switch.--prosewrites the subsystem pages as model prose (needs a key);--no-proserenders the whole wiki from structure with no model and no cost.--index-onlyand--docs llm|deterministicremain as deprecated hidden aliases. The interactive coverage chooser and the--coverage/--topranked-generation flags are removed;initandgeneratenow ask a single cost question. (#1032, #1036)- Hierarchy lives in the data model.
wiki_pagescarriesparent_page_id,display_order,section_numberand a stablestructural_key, so MCP (get_overview), the web docs tree and the breadcrumbs all read one stored tree instead of each deriving their own. (#1014, #1022) - Subsystem pages read like subsystem docs. The concept tree's
module_pages open by situating themselves against their siblings, weave in git-derived health and history signals, and close with a "Questions this page answers" section; a parent directory that spans several concept pages gets its own overview page. (#1043) get_answerleads subsystem questions with their concept page. A question about a subsystem as a whole now surfaces that subsystem's page above its individual file and child pages, instead of returning the parts without the whole. (#1046)- Retrieval demotes decision records and test pages, and
get_contextreports where a file sits in the concept tree. (#1040, #1056) - Dead-code confidence default is aligned across the CLI, REST and MCP surfaces, so the same threshold applies everywhere. (#1030)
Removed
- The template-vs-AI provenance axis.
is_deterministicanddoc_tierare gone from the API responses and every TypeScript type; the Auto / AI / Mixed page badges, the "Auto-documented" partition, and the per-page "Write with AI" affordance are retired. The pages router's?deterministic=filter becomes?has_prose=, scoped to the model-written page types. (#1037)
Fixed
- Idle files recover their time-decayed health on
updateinstead of staying stuck. (#728, #1061) - The embedder is resolved from the store rather than the environment, so upgrades no longer switch models under you. (#1007)
- OpenAI temperature is clamped to 1 for GPT-5-and-later reasoning models. (#989)
- Ollama timeout and connection errors are wrapped so they retry. (#829)
- Workspace repos resolve correctly in chat tool calls. (#971)
- Stale
.update.pendingmarkers are cleaned up, and deferred workspace docs runs are reported honestly. (#1004, #1000) - The overview summary self-heals a missing indexed commit, and repository sort timestamps are normalized. (#1005, #1058)
- Tombstoned pages no longer count in the docs tree or appear in status and export listings, and table-of-contents keys stay unique across repeated headings. (#1049, #1050, #1053)
- Workspace file counts and top-language queries are scoped to file nodes. (#952)
Documentation
- Plugin: added
coverageand impacted-tests slash commands, and synced the MCP tool surface and hooks with reality. The Codex plugin was brought to parity (code-health skill and skill guidance). (#1029, #1017, #962, #1016) repowise securitydocumented in the CLI reference, an examples index added, and the README leads with the numbers and restores the five-layer table. (#1026, #1019, #961)
Dependencies
- Consolidated security bumps across the dependency tree; lockfile packages pruned by the refresh were restored. (#1054, #1059)
[0.34.0] — 2026-07-20
Added
- Bug-fix history is now a first-class signal. repowise counts the bug fixes that landed on each file over the trailing six months, traces each fix back to the commit that introduced it, and attributes it down to the symbol. A file that keeps getting fixed is flagged a bug magnet, and that flag now leads everywhere risk is shown:
get_riskandget_change_riskreturn adefect_profile(fix_count,last_fix_days_ago,bug_magnet,top_symbols), the generatedCLAUDE.mdattention list ranks on fix history instead of raw churn,get_context's triage does the same, and the CLI warns at edit time when you touch a file with a recent run of fixes. Only fixes that change production code are counted, so a test-only touch-up no longer inflates the number. (#931, #939, #940, #946, #947, #954, #956) - Bug-fix history in the UI. The health drawer, panel and hover show per-file fix history; the symbols view shows and filters by per-symbol fix counts; and the commits view replaces the old file-risk bars with commit-level distribution views. (#948, #949, #950)
- A zoomable Knowledge Graph. The zoom map is now the Knowledge Graph at
/knowledge-graph, with per-node code health rendered on the map. (#918) - Present mode for docs. Wiki pages can be presented as a slide deck with a guided walkthrough, and architecture diagrams in wiki pages are now deterministic rather than LLM-drawn. (#914, #915)
- Costs page tells the whole story. Spend is labelled by operation, local runs are recorded at $0 instead of being dropped, ROI framing was added, and the agent savings the ledger was silently discarding are now counted. (#925, #927)
--verbose/-vacross the CLI.initandupdateare quiet by default and show per-phase internals plus debug logs under--verbose; the same flag was added tohealth,watch,restyle,generate-claude-md,coverage add, andworkspace add/workspace scan. (#929, #936, #937, #941, #942, #943, #944)
Changed
get_answeralways synthesizes. Confidence is graded after synthesis on how well the answer is grounded in retrieved content, rather than inferred beforehand from the shape of the retrieval. A leaner high-confidence payload is available behindREPOWISE_ANSWER_LEAN_HIGH. (#919, #923, #938)- Update keeps more of the index fresh. A workspace update regenerates docs per repo, onboards a provider when a docs update needs one, and refreshes external systems (C4 L1) when the manifest changes. (#916, #917, #921)
- Faster generation. Knowledge-graph enrichment overlaps page generation, and the inline-marker scan for decisions reuses ingestion's source map instead of re-reading files. (#912, #913)
- The SZZ blame pass is gone from the git indexer; fix attribution is derived without it. (#951)
- Feedback CTA sharpened and the recalibration banner dropped from the web UI. (#920)
Fixed
- Author experience is counted over the whole history, not just the current update batch, so contributor stats no longer collapse on incremental runs. (#953)
- Update progress counts onboarding pages in the generation total, so the bar no longer overshoots. (#928)
- Headline stats stop assuming a Sat/Sun weekend and are given room to breathe. (#926)
Documentation
- README rewritten and the
docs/tree restructured intostart/,reference/, andlayers/; the quickstart was rewritten and the user guide collapsed to a guide. Provider extras that don't exist are no longer documented. (#957) - Plugin: version bump plus
--verboseon theinit,update, andhealthcommands, and the pre-modification skill now readsdefect_profile.
Dependencies
mcp1.26.0 → 1.28.1. (#958)
[0.33.0] — 2026-07-18
Added
- Line-level AI authorship from your agent sessions. repowise reads agent-trace records to attribute commits (and now individual lines) to the agent and model that wrote them, and stamps a per-commit model id, so the contributors and provenance views can tell human-written code from agent-written code down to the line. (#861, #866)
- Docs that repair themselves. Generation now self-repairs hallucinated symbol references, resolves cross-page links across the whole repo on updates, and backfills graph-derived "related pages" with a self-healing pass, so wiki pages link to real symbols and stay cross-linked as the code moves. (#871, #872)
get_change_riskgrew up. The change-risk score gained a-x/--excludeflag with.riskignoresupport to drop vendored or generated paths from the diff, and now lists the line-level tests a change actually impacts. (#867, #903)- Coverage-backed tests in PR risk.
get_riskPR mode now surfacestests_to_runproven by the per-test coverage map, so a reviewer sees exactly which tests exercise the changed files. (#859) - Fewer dead-code false positives. Whole classes of systematic false positives (re-exports, dynamic references) are eliminated, so
get_dead_codereports far less that is actually live. (#886) - Better flow answers.
get_answerrescues un-named flow endpoints by re-ranking on the graph neighborhood, so "how does X get to Y" answers land on real endpoints more often. (#864) - Leaner MCP entry point.
get_whyjoins the lean MCP profile andget_answeris the advertised entry point;get_conformanceandgenerate_refactoring_codeare now opt-in.AGENTS.mdwas brought to parity with theCLAUDE.mdtool surface. (#889, #875, #900)
Changed
- Faster init and update. A broad performance pass across indexing: analysis phases run concurrently and incremental re-ingest runs in parallel, init- and update-built graphs converge so the centrality cache hits on updates, self-call resolution and pagerank are cached for cascade ordering, pages reused verbatim skip re-embedding, end-of-run wiki-page upserts are batched, refactoring detectors drop repo-sized per-file work, and coverage discovery plus the repo pre-scan share the pruned file walk. (#894, #895, #891, #892, #904, #893, #898, #897)
- Cleaner UI. The Overview got a visual-hierarchy pass and a denser health card with docs counts split by AI vs auto-generated, and "Attention Needed" is a triage panel again. (#882, #879, #883)
- TypeScript path resolution in the CLI.
TsconfigResolveris now wired into the CLI commands, sotsconfigpath aliases resolve during indexing. (#675)
Fixed
- Docs pointer no longer drifts on index-only updates. The docs commit pointer is preserved during index-only updates and kept separate from the sync pointer, fixing
update --docsdrift. (#849, #878) - Wildcard re-exports are followed in call resolution.
from x import *re-exports (and the JS equivalent) are now traced, recovering call edges that were dropped. (#905) get_answerstops dropping substance. Synthesis is no longer silently skipped when a usable API key exists, the gated low-confidence path serves real substance and unpins cached misses, gated-path excerpts no longer die to a swallowed error, and the live-grep fallback respects gitignore and exclude patterns. (#888, #887, #896, #856)- Sturdier change-risk on the CLI. A friendly revspec error is restored in
changed_lines, and subprocess handling is hardened and aligned with the MCP tool conventions. (#902, #899) - Security scan reads real source and guards repo-root detection, instead of scanning stale or wrong content. (#890)
- Smaller fixes. Unknown
restylestyle arguments raise a usage error; CLI usage errors are classified separately from failures in telemetry; decision staleness timestamps are normalized; webhook jobs run in sync mode; MCP health resolves the repository from the scoped session; and the code-rationale git-grep is hardened for macOS and color configs. (#908, #907, #877, #848, #865, #825)
Documentation
- README now counts ten MCP tools, fixes the coverage example, and tightens the Code Health section. (#901)
- Plugin: both the Claude Code and Codex plugins ship a
change-reviewskill; the risk command documents the new--excludeflag.
[0.32.0] — 2026-07-15
Added
- Every file is now retrievable. After the LLM budget picks its file pages, every remaining parsed code file gets a zero-LLM deterministic page, so concept search reaches the whole codebase instead of only the ~20% the budget covers. These tail pages are down-weighted in
search_codebaseandget_answerso a thin template page never displaces a rich page on a tie, and they carryis_deterministic/doc_tierin the API responses. (#817, #819) - More signals on the stats page. "By the Numbers" gained a coding-rhythm punch card (weekday × hour commit heatmap, UTC), 90-day momentum against the prior window, the wiki's own build cost stats, a change-risk mix, and a truck factor — all derived from data the index already holds, with no new indexing. (#828)
- Kimi is a first-class provider. (#824)
- Decisions follow you into Codex. The Codex SessionStart hook now delivers relevance-ranked standing decisions, matching what Claude Code already did. (#788)
- Test breakage is called out separately in risk.
get_riskPR mode splits test files into a newwill_break_testsfield, so a burst of broken tests can't crowd real structural impact out of the cappedwill_breaklist. (#739) - Richer anonymous telemetry. One event per MCP tool call, an interrupted status, and index-shape outcomes, so the collected numbers answer real questions. Consent and env vars are unchanged. (#820)
Fixed
- Search scores the identifier, not the whole question. (#853)
- A generic method mention no longer hijacks the answer. A prose
get_answerquestion that merely mentioned a common method (to_dict,provider_name) was routed into the exact-name union path and returned the whole definition set instead of answering. Large unions on prose questions now defer to synthesis; small unions and bare symbol lookups are untouched. (#823) - Key Concepts picks real concepts. Ranking was by file PageRank, so every public symbol in a high-ranked file inherited that rank — surfacing five methods of one registry class as the codebase's core concepts, and leaking pytest fixtures into the prose. Symbols are now ranked directly and spread across clusters. (#813)
- Custom
serveports work again. Next.js standalone compiled rewrites with build-time env vars, breaking custom port arguments;/api/*,/health, and/metricsnow proxy through runtime middleware. (#838, #839) - Persisted vector indexes load on startup. (#835)
- Go dead-code false positives. Function references in Go are now detected, rescuing live functions that were reported as dead. (#815)
- Dataflow recurses into
with-statement bodies. (#836) - Stats report true project age, commit, and contributor counts, with compact k/m formatting and years/months age in the hero. (#730, #827, #798)
- Source citations show normalized confidence, not the raw score. (#846)
- Anthropic thinking blocks are handled. (#787)
- Add Repository dialog no longer overflows its inputs. (#843, #844)
- Friendlier empty state for documents. (#822)
[0.31.0] — 2026-07-12
Added
- Coverage-backed test intelligence. repowise now ingests per-test coverage to build a test-to-code map, and puts it to work two ways: a missing-test signal proven by that map (real gaps, not just heuristics), and a new
repowise impacted-testscommand that takes a diff and lists only the tests that actually exercise the changed code, so you can run the smallest relevant subset instead of the whole suite. (#789, #790, #792, #793) - Decisions mined from your agent sessions. A shared agent-transcript layer lets repowise mine durable architectural decisions out of Claude Code / agent session transcripts, then deliver them relevance-ranked at session start and at edit time with a usage-feedback loop. Agent co-authorship is detected and decision scope is derived from it. (#774, #775, #776, #803)
- Flow-shaped answers.
get_answercan now trace how something gets from one named endpoint to another by walking the graph path between them and narrating the flow. (#804) - Shell parsed to a real AST. Shell scripts now parse into functions, source edges, and complexity like every other supported language, instead of being treated as opaque text. (#768)
- Lean MCP-only Docker image. A new
Dockerfile.mcpbuilds a slim MCP-server-only image for MCP hosts and server directories. (#800)
Changed
get_answeris grounded in what it serves. Synthesis is grounded in the exact body the answer returns, data-shape questions are grounded in the field set mined from source (including alias keys the shape omits), homonym symbols are answered by union across their definitions, and low-confidence misses are slimmed.get_symbol(id=...)is forgiven and a hedged answer is re-grounded on the served body. (#778, #779, #780, #772, #783)- Performance findings move the health score. Open performance-risk findings are now deducted from the performance score, so the number reflects outstanding perf debt. (#801)
- Decision sources gated and stickier. The code-comment decision harvest is gone, decision sources are gated, and dismissals stick so a dismissed suggestion stays dismissed. (#773)
- FAQ-weighted docs budget. Generation tilts documentation depth toward the modules people actually ask about. (#781)
- Hardier MCP surface. Session-survival hardening for the tool surface, plus a response budget clamped under the live MCP host token cap, keeps the server responsive under strict hosts. (#767, #771)
- Prose search leads with concept pages. Hybrid search now leads prose queries with concept and wiki pages instead of raw symbol hits. (#797)
- Contributors strip explains agent authorship. The contributors view now spells out the agent-authored share of commits. (#799)
Fixed
- Incremental updates persist graph and symbol data.
repowise updatenow persistsgraph_edgesandwiki_symbolson incremental runs (both were dropped before), and edge upserts keep the maximum confidence instead of overwriting it. (#795, #806, #807) get_answerverifies live slices. Symbol bounds are verified before serving a live slice, and non-path node ids are kept out offallback_targets. (#796, #769)- Honest search and symbol signals. Zero-exact search and filename-only
get_symbolcarry honesty signals instead of implying a confident hit. (#770) - Execution-flow picker on the module overview. The module overview gained an execution-flow picker with hierarchical-layout guardrails so large graphs stay readable. (#805)
Documentation
- New hooks guide plus a "learns from your usage" section explaining the adoption feedback loop. (#784)
- Full AGPL-3.0 license text so GitHub detects the license correctly. (#794)
[0.30.0] — 2026-07-10
Added
- First-class output-language support.
repowise init --language <code>(15 languages: en, ru, es, fr, de, zh, ja, ko, it, pt, nl, pl, tr, ar, hi) sets the natural language for generated wiki pages; advanced interactive mode now asks for it too (English default). The choice persists to.repowise/config.yamlsoupdateregenerates changed pages in the same language, and the workspace init, workspace generate, and server regenerate paths now honor it instead of silently defaulting to English. Code, file paths, and symbol names stay untranslated. Previously this was config-file-only (#99). (#756) - Worktrees just work.
repowise initandrepowise updateinside a linked git worktree now auto-detect the base checkout and seed the worktree's index from it, then catch up incrementally; no flags needed.--seed-from <base>remains as an explicit override and--no-seedforces a cold init. See WORKTREES.md. (#655 introduced the manual flag; #747 makes it automatic.) - Ruby and Scala promoted to Full health tier. Both languages now get the full complexity node map and performance-risk dialect, so health scores and perf findings match the depth Python/TypeScript/Go already had. (#745, #749)
- Git stats awards. The contributor stats view gained biggest-commit, longest-streak, most-imported-file, and dependency awards. (#752)
- Workspace socket-contract detection. Cross-repo analysis now detects socket-based service contracts alongside HTTP ones. (#710)
- Anonymous feedback panel. The dashboard gained a lightweight feedback panel — no account or email required. (#763)
- Claude Code integration: session context + safer distill. The augment hook now emits a per-session context block with live index freshness on session start, the distill rewrite hook learned two safe shell-syntax carve-outs, and
repowise doctorchecks hook registration. (#743)
Changed
- Unchanged wiki pages are reused across runs. Cross-run page reuse now keys on the documented file's content plus the generation settings (template, language, style) instead of the exact rendered prompt — which embedded per-run retrieval context and practically never matched. A re-run over a repo where a handful of files changed now regenerates only those pages; changing the model, output language, or wiki style still regenerates everything it should. (#757)
- One dataflow parse per file. The health pass's dataflow consumers (the Extract Method detector and the perf advisory-to-asserted promotion) now share a single lazily parsed per-file analysis instead of each re-reading and re-parsing the file, making the health pass measurably faster on large repos with identical output. The new per-function dataflow summary and point-lookup APIs this introduces are the foundation for upcoming dataflow-backed surfaces. (#755)
- Leaner MCP surface for agents. An agent-lean tool profile with tool-search gating and registry-pinned docs, plus payload-precision trims across tools, cut the token cost of the MCP surface without removing capability. (#742, #744)
- Graph view cleanups. The module view no longer collapses into a single blob, external nodes are hidden by default, and node info panels are layer-aware. (#754)
Fixed
- Pages skipped by the update budget are labelled stale.
repowise updatemarks weakly-affected pages it chose not to regenerate as stale instead of leaving them claiming to be fresh, andtotal_pagesis recomputed from the store. (#748) - Dataflow CFG no longer flags reachable code as unreachable. (#760)
- TypeScript object-literal shorthand properties count as reads, fixing spurious unused-symbol findings. (#759)
- Cross-repo co-change is a bounded session-share signal, preventing one shared session from dominating workspace coupling scores. (#758)
- Seeded worktree indexes no longer split in two. Seeding now re-points the copied index at the worktree, so the first update reuses the seeded pages instead of minting a second repository entry and regenerating everything under it. (#747)
[workspace]notices render again. The one-line workspace auto-detect notice was being swallowed by console markup and printed without its prefix. (#747)
Documentation
- User-facing docs restructured and refreshed (quickstart, user guide, CLI reference, config). (#741)
- Plugin: the Claude Code plugin bundles the new SessionStart context hook. (#743)
[0.29.0] — 2026-07-09
Added
- Leverage-weighted health signals.
get_healthnow surfaces NLOC-weighting and per-file leverage, so the score points at where a fix buys the most. (#719) - Wider agent-authorship detection. Git indexing recognizes agent-written commits across more provenance channels. (#731)
Changed
- Leaner
get_overviewby default. Theget_overviewMCP payload is compact by default, cutting the token cost of orienting an agent. (#729) - Health scores reflect corrected findings. Removing the false findings below raises some file and repo health scores — this is expected and desirable, not a regression. Broad
except Exceptioncatches are now detected completely (they were previously under-counted), so a few files surface additional — but honest — findings.
Fixed
- Honest exception-handling rationale. A broad
except Exceptionis no longer described as catchingKeyboardInterrupt/SystemExit— only a genuine bareexcept:orexcept BaseException:carries that warning, andexcept Exceptiongets its own rationale. Go blank-identifier discards are only flagged when the discarded value sits in the error position, and Rust panic macros and.unwrap()/.expect()inside test code are no longer flagged as recoverable-error crashes. (#733) - Accurate complexity counting.
elif/else ifchains no longer read as deep nesting (a flat guard chain is flat); parameter counts ignore the bare*//separators; comprehension filters now count toward complexity; and docstring- and comment-only lines no longer inflate a function's or class's measured length. (#734) - Fewer false performance findings. I/O-in-loop, defer-in-loop, blocking-I/O-under-lock and related markers no longer fire on a closure that is merely defined — not run — inside a loop or lock; a parenthesized
awaitis recognized as awaited;deque.insert(0, …)is no longer flagged as quadratic; name-shadowing collisions in the Python and TypeScript detectors are resolved; and a semaphore-bounded goroutine worker pool is no longer called unbounded. (#735) - Truthful structural findings. A god-class finding cites the complexity of the actual brain method rather than the class-wide maximum; an
UPDATE/DELETEbounded by aLIMITis no longer said to touch every row; a flatmatch/casedispatch table is no longer flagged as a large method; and hidden-coupling severity is no longer overstated from a handful of shared commits. (#736) - Cleaner, more private MCP output. Contributor email addresses are never shown in overview output (display names only); extreme churn renders as a multiplier instead of a runaway percentage; search snippets no longer repeat a title-only decision; comment-derived decisions rank below real architecture decisions; file counts are labeled; and decision titles truncate on a word boundary with an ellipsis. (#737)
- Atomic update lock.
repowise update's lock file is now written atomically with its contents, closing a creation race. (#720)
[0.28.1] — 2026-07-08
Changed
- Performance map colors by findings. The performance code map now colors by open findings and detector coverage instead of the bounded score, so hot spots read at a glance. (#716)
Fixed
- Graph overlays show their nodes. The full dependency graph reserves part of its node budget for dead-code files, hotspots, and execution-flow members instead of selecting purely by PageRank — the Dead/Hot overlays and flow highlighting no longer come up empty on large repos. The view says how many flagged files are in view ("12 of 37 dead files"), and empty overlays explain whether the repo has none or they fell outside the loaded set. (#714)
- Graph controls explain themselves. The hierarchical layout says why it won't run above 500 nodes instead of silently doing nothing; the Execution Flows panel gained a close button and Escape handling, and warns when a selected flow has out-of-view nodes; the Dead/Hot toggle pair became an exclusive All / Hot / Dead control. (#714)
- Honest health drawer and stats labels. Missing structural metrics render "not measured" instead of 0; score-breakdown bars scale against real category caps with a tooltip explaining the cap; the lines-of-code and agent-authorship stats now say exactly what they measure. (#715)
- Decisions page polish. Missing decision dates render a dash instead of "Invalid Date"; the decision graph bounds its layout so large decision sets can't hang the page; Confirm/Dismiss/Deprecate explain what they do; and a new "Enforce this decision" button generates a paste-ready agent prompt that audits governed code for compliance. (#715)
- Honest performance coverage. Dead performance markers are wired up and performance coverage is reported honestly. (#711)
[0.28.0] — 2026-07-07
Added
- Lean 4 support. A lightweight regex tier brings symbol extraction to Lean 4 codebases. (#600)
Changed
- Reliable incremental updates.
repowise updatewas reworked this cycle. Incremental runs now rebuild the knowledge graph, so an updated index stays as fresh as a full one instead of serving a stale graph (#702). The store is persisted and locked atomically with honest failure reporting, so an interrupted update rolls back cleanly rather than leaving a torn store behind (#706). The workspace and single-repo update paths were reconciled onto one code path (#703).
Fixed
- Cleaner code-health flagging. Resolved false-flag presentation in Code Health across grouping, dominant-cause attribution, and floor magnitude. (#700)
- Contributors counted once. GitHub noreply emails are folded together, so one person no longer shows up as two contributors. (#701)
- Fewer Go dead-code false positives. Same-file type references are rescued from unused-export false positives in Go. (#629)
Documentation
- No-key quickstart. Published a verified no-API-key quickstart in the README and on repowise.dev. (#627)
[0.27.0] — 2026-07-05
Added
- VS Code extension 0.3.0. The editor experience grew up this cycle. Refactoring plans can now be handed straight to an AI agent, and the extension exposes native chat tools so agents can query Repowise from inside the editor (#694). The SCM view gained change intelligence, per-file change risk, and symbol hover detail, alongside more reliable server discovery (#664, #691). The listing now leads with a hero walkthrough GIF and screenshots, and the extension icon reads correctly on dark themes (#696, #697).
- Dart support. A Dart AST tier brings symbol extraction plus health and performance markers to Dart codebases. (#689)
- SQL and dbt intelligence. Indexing now extracts SQL DDL symbols and dbt lineage (#683), models app-to-database contracts, and surfaces SQL-specific health markers (#687).
- Java and Rust dataflow. The Extract Method dataflow layer now understands def/use chains in Java and Rust, extending refactoring analysis to those languages. (#686)
repowise login,logout, andwhoami. New CLI commands to connect the CLI to your Repowise account. (#690)- Storage footprint in
status.repowise statusnow reports the on-disk size of the index. (#681) - Add-repo wizard. The web app gained a guided add-repo flow with a cost preflight and a live first-index experience, plus first-run polish across the app icon, explore cards, and a collapsible workspace nav (#692, #685).
- Accurate coverage tab. The Coverage tab now paginates, sorts, and reports coverage accurately on large repos. (#665)
Changed
- One consistent UI. A design pass unified how the dashboard renders tables, stat tiles, row banding, loading skeletons, error states, and tooltips, so every view behaves the same way. (#695)
- Dataflow-verified performance findings. Advisory performance findings are now verified against the dataflow layer, cutting false positives, with refactoring config wired through. (#684)
Fixed
- Health sync honors repo excludes. Workspace health sync now respects the repo's configured excludes. (#638)
- External dependencies no longer masquerade as files. The Files view stops linking external dependency nodes as if they were source files. (#673)
- More robust parallelism. Parse and betweenness process pools now use
spawn, avoiding fork-related instability on some platforms. (#679)
Dependencies
- Added
sqlglot(SQL parsing) andtree-sitter-dart(Dart grammar).
[0.26.0] — 2026-07-03
Added
- VS Code extension. Repowise now runs inside your editor. The extension manages the local server lifecycle, walks you through install to first insight, and registers the Repowise MCP tools for agent mode (#643, #644). Low-health files surface as diagnostics with gutter heat, and editor-native signals include live range risk scoring, a refactoring lens, dead-code line spans, and inline docs (#642, #644). A sidebar Home dashboard shows index freshness, a theme switcher, and consolidated trees (#650), and the shared visualization panels (graph, C4, health, blast radius) render directly in webviews (#649, #653). A settings panel configures editor signals and the server connection (#654), and the latest pass adds panel navigation and quieter defaults for an editor-native feel (#660). Install it from the VS Code Marketplace or Open VSX.
- Continuous-zoom architecture view. The server builds a zoom-map artifact that drives a smooth, execution-aware zoom across the architecture graph. (#626)
- Configurable Ollama embedding timeout. The Ollama embedding request timeout can now be set via environment variable for slower local models. (#656)
Changed
- Sharper
get_answergrounding. Theget_answerMCP tool gained a frame-grounding gate and anchors rationale to in-code comments, with retrieval tuning acrossget_answerandget_context. (#621, #622) - Faster decision embeddings. Decision embeddings are batched during persistence and reindex, cutting indexing work on decision-heavy repos. (#641)
Fixed
- Config languages no longer inflate language usage. Configuration-file languages are hidden from the language-usage breakdown. (#623)
- Index freshness stamp stays current on no-op syncs. An
updatethat finds no changes still refreshes the freshness stamp, so agents don't distrust a current index. (#652)
Dependencies
- Cleared high and critical CVEs across the npm and Python dependency trees. (#645)
[0.25.0] — 2026-06-27
Added
- Split File refactoring. Code Health now detects files that should be decomposed into smaller modules and proposes a concrete split. A new detector identifies low-cohesion modules and groups their members into coherent target files (#607), with richer cohesion signals driving the grouping (#614). Each plan is browsable in the web Refactoring tab and can be turned into real code via the deterministic code-gen path (#608).
- Extract Method refactoring. Long, complex functions get an Extract Method suggestion computed over a real dataflow layer: an intra-procedural control-flow graph for flagged functions (#612), def/use chains and reaching definitions over that CFG (#613), and the Extract Method planner built on top (#615). The refactoring is available for Python, Go, and TypeScript/JavaScript (#616).
- Coverage report ingestion. Indexing can now ingest test-coverage reports, folding coverage into the code-health picture during a run. (#604)
Changed
- "Biomarker" is now "marker" in the UI. Code Health display copy renames the user-facing "biomarker" term to "marker" across the web app and plugin surfaces; internal identifiers are unchanged. (#619)
Fixed
.works as a glob pattern on Python 3.14+. Passing.as a path/glob no longer errors on newer Python. (#609)- Decision harvest skips title-only records. The decision harvester no longer emits empty records that carry only a title. (#605)
Documentation
- Strengthened the code-health validation story and fixed stale references across the docs. (#617)
- Tightened the code-health docs, named CodeScene explicitly, and moved deeper internals into the architecture doc. (#618)
[0.24.1] — 2026-06-25
Changed
- Workspace tables and the dependency-structure matrix stay responsive on large repos. The co-change, repo-pair, contract-links, and package-deps tables now use the windowed virtualized table, and the dependency-structure-matrix grid is capped to the top-60 services by connectivity so a large workspace can't render tens of thousands of cells; counts still reflect the full matrix. (#602)
[0.24.0] — 2026-06-25
Added
- Refactoring intelligence: deterministic, graph-aware refactoring plans. Code Health now derives concrete refactoring suggestions from the dependency graph and health biomarkers, with detectors for Extract Class, Extract Helper, Move Method, and Break Cycle (#586, #587, #588). Each suggestion is a ranked plan card carrying impact, effort, blast radius, and evidence, browsable in a new web Refactoring tab with file-first cards, a visual plan modal, and one-click agent export (#589, #590). Plans can optionally be turned into real code: opt-in LLM code generation produces a change from a deterministic plan, viewable in a side-by-side diff viewer (#592, #594).
- Airy Code Health overview with a Findings workbench. The Code Health page was redesigned around a calmer overview and a dedicated Findings workbench for triaging biomarkers. (#593)
- Browsable Files page. A new Files page lets you browse the repo's files directly, with a restyled table and dark-mode polish. (#591)
initAdvanced options.repowise initgained an Advanced section with a docs toggle and a configurable index-only mode, and raised the commit-history cap. (#599)
Changed
- Large tables are virtualized. A shared windowing primitive virtualizes large tables across the dashboard, keeping big repos responsive. (#598)
Fixed
- Execution-flow entry-point scores survive updates. Incremental updates no longer wipe entry-point scores on the execution-flow graph. (#585)
Documentation
- README and docs now lead with code health as a measure-locate-fix loop and document refactoring intelligence. (#595)
- Plugin: version bump to 0.24.0.
[0.23.0] — 2026-06-23
Added
- Repo-understanding overhaul: knowledge graph, C4, and guided tour. The C4 model now derives real actors, true coupling, and accurate containers instead of approximations (#576). The guided tour ranks orientation entry points by execution-start order and surfaces churn hotspots (#574), and docs/tooling files are routed out of the layer catch-all with an invariant reviewer gate to keep the layering honest (#575). The knowledge-graph info panel is now collapsible (#579).
- Enriched
get_healthfor pre-PR self-check. Theget_healthMCP tool returns a richer payload so an agent can read the same signals the merge gate judges a change on before opening a PR. (#571, #572) init --no-workspaceand a fully non-interactive--yes.repowise initcan skip workspace setup, and--yesis now fully non-interactive for scripted/CI use. (#573)- Rust performance dialect. Performance-risk detection gained a Rust dialect for I/O-in-loop / N+1 shapes. (#581)
- Configurable health rules.
health-rules.jsonnow supports severity overrides and a small-team profile. (#569) - Better cross-repo contract matching. HTTP contract extraction resolves router mount prefixes (#567), and consumer-side matching gained hygiene filtering and base-URL service resolution (#568).
Fixed
GraphBuilderis picklable across a process boundary. Fixes a failure when the graph builder is handed to a worker process. (#583)- Parse cache versioned by
ParsedFileschema shape. The parse cache now keys on the parsed-file schema rather than the package version, so unrelated releases keep the cache warm and a schema change invalidates it automatically. (#582) - Knowledge-graph panel guards
matchMedia. The KG panel mount effect no longer assumesmatchMediais present, fixing a crash in environments without it. (#580)
Documentation
- Plugin: version bump to 0.23.0.
[0.22.0] — 2026-06-22
Added
- Three-signal code health. The single code-health score is now split into three co-equal signals: defect risk (the headline score), maintainability (smells that raise change-cost without predicting bugs), and performance risk (static I/O-in-loop / N+1 shapes). Maintainability and performance are surfaced as their own pillars across the dashboard and the
get_healthMCP tool rather than being blended into the defect headline. (#528, #531, #533, #544) - Performance-risk detection across languages. A new performance detector finds I/O-in-loop and N+1 shapes, including cross-function cases resolved through call-graph reachability, with language-specific markers and dialects for Python, Java, Go, and C# (loop-level markers,
pandas_iterrows_in_loop, centrality gating, and a reusable severity ranker). Detection runs through aPerfDialectplugin registry. (#530, #532, #536, #537, #538, #539, #541, #542) - Dependencies classified by I/O boundary. External systems are now classified by the kind of I/O boundary they sit on, feeding the performance and architecture views. (#529)
- Reindex-free upgrades. The on-disk store now carries a format version separate from the package version, so upgrading repowise no longer forces a reindex. Upgrades show a release advisory and a "what's new" panel on the CLI, and the dashboard surfaces available upgrades with release info shared through core. (#553, #554, #556)
- Hybrid symbol and path search in
search_codebase.search_codebasenow searches repowise's own structural index for identifier- and path-shaped queries instead of only running wiki-semantic search. A newmodeparameter (auto/concept/symbol/path/hybrid) controls routing;autopicks by query shape and returns symbol IDs, file/line bounds, and signatures for identifier queries. (#558) - Anonymous, opt-out CLI telemetry. The CLI now collects anonymous usage telemetry behind a central platform layer; it is opt-out and collects no source code. (#555)
- AI prompt actions across the dashboard. Health findings now carry MCP-native "fix this" AI prompt actions, rolled across the dashboard, with a finding-count cap to keep output bounded. (#546, #547)
- Commits and stats redesigns. The commits page leads with a Code Evolution timeline, the blast-radius impact tab was redesigned, the owners view leads with a knowledge-distribution headline, and a new repo Stats "By the Numbers" page was added. Co-changes gained a repo-pair summary drill-down. (#543, #548, #549, #550, #551)
Changed
- Distill: seamless rewrite permissions + re-read savings.
repowise distillgained smoother rewrite-permission handling and additional re-read token savings. (#559)
Fixed
- Overview "Last synced" reflects CLI auto-syncs. The overview page now reflects auto-syncs triggered by the CLI in its "Last synced" timestamp. (#564)
- Dead-code analyzer keeps same-file Python symbols. Python symbols referenced only within their own file (callable-as-argument, annotation-only) are no longer flagged as dead code. (#563)
- Index freshness stays current. Updates now keep the
CLAUDE.mdstamp and the indexed commit current so agents don't distrust a fresh index. (#524) - Commits-page follow-ups. Full-width agent strip, collapsible risk panel, and repo-wide stat cards on the commits page. (#552)
- Fewer performance false positives. Eliminated three perf-detector false-positive classes across C#/Go/Python and now skips
for...ofover constant collections in the TS/JS detector. (#540, #545)
Documentation
- README refreshed: banner, numbers-first lead, combined demo GIF, three-signal code-health framing, and star CTAs. (#534, #535, #562)
- Plugin: version bump to 0.22.0;
search_codebaseskill docs updated for hybrid symbol/path search.
[0.21.0] — 2026-06-19
Added
- Cross-repo workspace intelligence. Workspace mode gained a live system map of cross-repo services, backed by a service-granular system graph with extraction diagnostics, cross-repo blast radius and change risk, and a breaking-change guard that flags edits to contracts other repos depend on. (#511, #512, #513, #514)
- Architecture analysis. New architecture conformance checks, dependency-cycle detection, a design-structure-matrix (DSM) view, and architecture metrics (propagation cost, core/periphery roles, and a 1-10 architecture score). (#515, #517)
- Repo-wide change-coupling graph. A new graph surfaces files that tend to change together across the whole repo. (#497)
- Wider cross-repo contract extraction. HTTP contract extraction now spans more languages and frameworks: Rust HTTP route providers and reqwest consumers, C#/Unity consumers, and JS wrapper / variable-URL consumers. Extractors were split into per-framework dialects for maintainability. (#505, #506, #507, #508, #510)
- Configurable MCP tool surface. The set of tools the MCP server advertises is now configurable. Workspace-only tools (
get_blast_radius,get_conformance,get_architecture) are advertised only in workspace mode instead of always, and two extra tools (get_dependency_path,get_execution_flows) can be opted in. Configure it with anmcp.toolsblock in.repowise/config.yaml(+/-deltas, an explicit allowlist, orall) or per launch withrepowise mcp --tools/--all. (#520) - MCP tool surface editor in the dashboard. The Settings page now lists every tool with its description and lets you toggle the surface per repo, writing the selection back to that repo's
mcp.toolsconfig. Backed byGET/PATCH /api/mcp/tools. (#521)
Changed
- Code-health-first repo overview. The repo overview page was rebuilt around code health. (#501)
- Airier, diagram-first web UI. A UX overhaul restyles the dashboard on a shared composition backbone, with more whitespace and diagram-forward layouts. (#504)
- Consolidated the MCP tool surface. Removed six redundant MCP tools (
annotate_file,get_callers_callees,get_community,get_graph_metrics,get_architecture_diagram,update_decision_records) whose capabilities are already covered byget_context(include=[...])andget_why. The MCP server exposes 13 tools: 10 in single-repo mode plus three workspace-only tools (get_blast_radius,get_conformance,get_architecture). Documentation and tool counts across the project were reconciled to match. (#519)
Fixed
- Contract extraction no longer scans nested repos. Workspace contract extraction could hang scanning up to a million files when a repo contained nested checkouts; it now uses the shared file traverser and skips nested repos. (#516)
- C# gRPC consumer extraction requires gRPC context before treating a client call as a cross-repo consumer, removing false positives. (#509)
- Skip Unity-generated dotnet scan paths during ingestion. (#499)
Documentation
- Plugin: version bump to 0.21.0; MCP tool surface docs reconciled to the consolidated, configurable set.
[0.20.0] — 2026-06-16
Added
- Churn x complexity quadrant on the hotspots tab. The hotspots view now plots files on a churn-versus-complexity quadrant, making it easy to spot the high-churn, high-complexity files that warrant attention first. (#491)
- Per-file process, people, and topology signals. The file page now surfaces per-file process signals (how the file changes over time), people signals (ownership and contributor spread), and topology signals (how connected the file is), each already computed during indexing. (#490)
- Per-file health score over time. The file page, the file drawer, and the MCP
get_healthsurface now show a file's health score history, so you can see whether a file is trending better or worse. (#489) - Health bands, repo distribution, and a README badge. Code health is now bucketed into named bands with a repo-wide distribution view, and a health badge can be embedded in your README. (#485)
Changed
- Quieter, polished
updateCLI UX.repowise updatenow uses the same calm, panel-based progress output asinit, with a-vflag for verbose detail. The previously monolithic update command was also split into a package for maintainability. (#476, #477) - Co-change page reframed as a temporal hint. The cross-repo co-change page now presents its data as a temporal hint rather than an authoritative dependency, and the average-strength figure is displayed correctly. (#481)
Fixed
- Workspace job progress stays accurate. In multi-repo workspace mode, job listing and progress now read from the correct per-repo database, stale jobs left running after a server restart are reset, and the progress timer and phase labels reflect persisted state instead of component mount time. (#487)
- Co-change noise filtering and cross-repo strength normalization. Noise files are filtered out of co-change analysis and cross-repo co-change strength is normalized so the signal is comparable across repos. (#480)
- Like-with-like population comparison in coordinator health. Coordinator health now compares files against like-sized populations rather than mixing dissimilar groups. (#479)
Documentation
- README dual-audience positioning. The README was reworked for a dual-audience frame and now surfaces change-risk, agent provenance, and wiki styles. (#478)
- Plugin: version bump to 0.20.0 (no command/skill/hook/MCP-surface changes).
Dependencies
- Bumped
pyjwtfrom 2.12.1 to 2.13.0, a security release bundling five advisory fixes. (#488)
[0.19.1] — 2026-06-13
Fixed
repowise serveweb UI failed to build for the release tarball. The bundled web dashboard could not be compiled once a workspace-package barrel entry was imported as a value (introduced by the wiki-styles constants), because Webpack could not resolve the ESM.jsre-export specifiers in@repowise-dev/types/@repowise-dev/uiback to their.tssources. Added an extension alias to the Next.js build so.jsspecifiers map to.ts/.tsx. This affected only the publishedrepowise-web.tar.gz; the Python wheel was unaffected. (#471)
[0.19.0] — 2026-06-13
Added
- Wiki styles: selectable documentation voice. Generated pages can now be produced in one of four styles:
comprehensive(the default, unchanged),caveman(token-condensed, AI-first),reference(API-manual), ortutorial(beginner-friendly), plus user-defined custom styles. Styles change only the prose voice and density, not the document structure: headings, sections, table of contents, search, and cross-links are unaffected. (#468) - OpenCode CLI provider. A new local OpenCode LLM provider runs documentation generation through the local OpenCode CLI via
opencode run --format json. Usesasyncio.create_subprocess_exec(no shell), parses JSONL output, validates model names against a safe character set, and treatsopencode/*cost as$0.00. No API keys are stored; OpenCode manages its own auth and model selection through its provider system. Interactive selection detects the OpenCode CLI onPATHand shows helpful install/setup instructions when it's missing. (#436) - Health score self-validation. The code-health surface now validates the score against each repo's own bug history: it ranks files by health, takes the 20 lowest, and reports how many were touched by a
fix:commit in the trailing ~6 months versus the repo-wide baseline rate (the lift), e.g. "16/20 lowest-health files had a bug fix in the last 6 months, 3.3x the 24% baseline". Stays silent when there is too little history to be honest. (#438) - Error-handling maintainability biomarker. A new biomarker surfaces swallowed-exception and unsafe-unwrap anti-patterns as a bounded maintainability finding: empty or trivial
catch/exceptbodies across Python, JS/TS, Java, Kotlin, C#, and C++, plus Python catch-allexcept:/except Exception:/except BaseException:. (#453) - MCP streamable HTTP transport. The MCP server can now serve over a streamable HTTP transport in addition to stdio. (#444)
REPOWISE_PORTenv var.repowise servenow honours theREPOWISE_PORTenvironment variable for the server port. (#455)
Changed
- Web dashboard UX overhaul. End-to-end rework of the web UI: a slimmer six-group sidebar (Overview, Docs, Architecture, Code Health, People and History, Chat) shared across desktop and mobile, Overview as the repo landing page, canonical entity pages, a single unified architecture destination, and surfacing of git/agent provenance data that was already persisted but previously invisible. Theme unchanged; design tokens were only added, never renamed. Retired pages (hotspots, ownership, dead-code, blast-radius) and stub routes redirect into their new homes, so every old URL still resolves. (#466)
- Hybrid MCP improvements. A batch of MCP server improvements centered on making a tool response trustworthy enough that the agent never re-reads the source it just paid for: a verified trust contract, an honest savings ledger, indexing of module-level constants, and trimmed per-call token overhead. Net additive to the tool surface, with no breaking changes to existing tool contracts. (#467)
- Change-risk clarity. Change-risk now prioritises repo-relative signals and uses honest driver labels instead of misleading absolute framing. (#465, #469)
- Dead-code framing. Findings are now framed as cleanup candidates rather than safe-to-delete, reflecting that static reachability can't prove a symbol is unused. (#433)
Fixed
- Co-change strength display. Co-change strength now shows the raw score instead of a misleading percentage. (#439)
- Health scoring of module-level JS callbacks. Module-level JavaScript callbacks are now scored correctly. (#456)
- Health trend wording. Clarified how health-trend score changes are presented in the web UI. (#457)
- CLI model selection. The CLI now honours the
config.yamlmodel when a provider is set via env var or flag. (#442) - Chat config inheritance. Chat now inherits the per-repo provider, model, and key from the init config. (#434)
Performance
- XL-repo indexing pass. Faster indexing on very large repos via cpp hint regex tuning, git deep-walk improvements, and XAML index reuse. (#459)
- Large-repo indexing pass. Indexing and update-path improvements covering type references, health, dynamic hints, and dead-code analysis. (#450)
- Incremental duplication splice. Update runs now splice duplication pairs incrementally instead of recomputing them wholesale. (#460)
Documentation
- Plugin: version bump to 0.19.0 (no command/skill/hook/MCP-surface changes).
[0.18.0] — 2026-06-08
Added
- MCP counterfactual token savings. Every MCP tool call now records what its
curated answer replaced — the raw file exploration the agent would have done
otherwise — into the unified savings ledger as a
mcp:<tool>row.get_symbolreports the whole file it sliced one symbol from,get_contextthe full files its skeletons stood in for,search_codebasea conservative floor per cited file; estimates undersell by design. The Costs hero now reads "N MCP queries answered" and grows per call, andrepowise saved --by sourcesurfaces the per-toolmcp:*breakdown. Recording is best-effort and never alters a tool's user-facing response. - Costs page savings hero. The Costs page now leads with a results card
showing every token and dollar repowise saved your coding agent, combining
the
repowise distillledger with MCP tool-response savings that were already on disk but previously invisible (source='mcp:*'in the omission store). The dollar estimate is priced at the agent's actual model — detected from your local Claude Code / Codex transcripts (read-only, on-machine), falling back to a default rate when undetectable — because saved tokens are input the agent never had to read. Missed savings now read as an "unlock more" prompt rather than a footnote.
Changed
repowise initalways renders the compact banner. The init splash now uses the compact owl variant (~60% of the old full-art width) on every terminal, so narrow shells no longer wrap it. (#423)
Internal
- Shared UI/server building blocks consolidated so the web UI and downstream
consumers stop duplicating code:
OwlLoaderand the design-token gate scripts move into the@repowise-dev/uipackage, a dependency-free@repowise-dev/ui/brandconstants export is added, and the community/ architecture view builders are extracted from the server routers into FastAPI-freeservices/functions. No change to the install/serve UX or any endpoint's response shape. (#423)
[0.17.1] — 2026-06-07
Added
- Official MCP Registry listing. repowise is published to the
MCP Registry as
dev.repowise/repowise(PyPI package, stdio transport), so MCP clients can discover and install the server from the registry directly. - Distill: stat-only diff filter.
git diff --statoutput gets its own filter — the roll-up line plus the top-20 rows by churn — instead of slipping past the hunk-based diff filter raw (#414).
Changed
- Skeleton is the default context card for files.
get_contexton file targets above ~80 lines now returns the smart skeleton (every signature, central bodies inlined) instead of the bare symbol list — measured strictly better per token.compact=Falseopts out; amostly_fullflag marks small files where a directReadcosts little more (#414). repowise initdefaults tuned. LLM concurrency defaults to 10 (tiny repos 12, huge repos 5) acrossinit,update, andworkspace add; the LLM cost-gate confirm defaults to yes (the cost was already shown beside the coverage tier); page generation prints a hint that runs are resumable withinit --resume(#412).
Fixed
- Semantic search lost embeddings on mid-size repos. A whole generation
level was embedded in one API request; file pages routinely blew the
provider's per-request token cap, failed 400, and the failure was swallowed
at debug level — fresh inits silently lost all file-page embeddings.
embed_batchnow chunks requests with failure isolation per chunk, and the loss (if any) surfaces as a warning with arepowise reindexrepair hint (#414). repowise updateevicted pages from semantic search. Regenerated pages were persisted to SQLite but never re-embedded, so every update drifted the vector corpus away from file pages. Updates now embed regenerated pages into the vector store; existing repos repair withrepowise reindex(#414).search_codebaseranking. Decision records (short title-statements) no longer dominate design-noun queries — they're down-weighted unless the query is why-shaped; retrieval over-fetches before re-ranking; thekindfilter runs before the limit cut sokind="implementation"can't return an empty list; pages without a backing file classify as"doc"(#413, #414).get_riskcalibration. The 0–10 score no longer pins at 10.0 from transitive-dependent breadth alone (exponential file term + capped breadth term); co-change partners survive incremental updates instead of being wiped; files excluded via.git/info/excludeno longer leak intowill_break;directive.missing_testsis scoped to the PR's changed files (#413, #414).get_contextcontract. Docs + freshness defaults are always returned —include=["skeleton"]no longer drops the summary and freshness card; signatures collapse onto one line (no leaked\r\nfrom CRLF files); module cards describe child files with their indexed summaries (#413).- Generated CLAUDE.md quality. Word-boundary truncation in tables (no more
mid-word chops), prose-only sentence extraction (list items and table rows
no longer jam onto the architecture paragraph), guided-tour steps carry file
paths again, the Owner column drops when no module has owner data, and tech
stack detection ignores test fixtures / vendored repos and finds
tsconfig.jsonin workspace packages (#413). repowise inithealth-phase progress bar moved from the first completed AST walk instead of sitting at 0/N through the pre-walk, and the duplication scan overlaps the pre-walk (#412).- Distill on Windows:
cmd /cwrappers are stripped during command normalization so native listings route to the file-listing filter, which now also accepts absolute Windows paths (#413).
[0.17.0] — 2026-06-06
Added
- Distill — index-aware output distillation. A new capability that
compresses noisy command output before the agent reads it, errors-first and
fully reversible.
repowise distill <cmd>runs a command and prints a compact rendering (exit code preserved, every error line kept, raw output stashed behind an inline[repowise#<ref>]marker);repowise expand <ref>restores it, optionally filtered with-q. Eight content filters ship (test/build output, git status/log/diff, search floods, file listings, generic logs), measured at 60–90% token reduction on noisy commands with zero error-line loss. An opt-in Claude Code PreToolUse hook (repowise hook rewrite install, or therepowise initprompt) rewrites noisy agent commands torepowise distill <cmd>pending approval — ask-by-default, with per-repo / per-familyallow/denyconfig; pipes, redirects, and compound commands are never rewritten.repowise savedreports tokens and estimated dollars saved (per-filter / per-day / per-source rollups), mirrored by a Distill savings card on the dashboard's Costs page. - Read intelligence: skeletons, stale-read notices, search digests.
get_context(..., include=["skeleton"])returns an indexed file with bodies elided — every signature plus the bodies of the most central symbols, sliced from persisted symbol bounds with zero query-time parsing (~15% of full-file tokens). The PostToolUse hook nudges once per file per session when a largeReadcould have been a skeleton, warns when a re-read follows anEdit/Write(excerpts predate the edit), and renders grep floods as a compact grouped-by-file digest ordered by graph centrality. - Reversible MCP truncation. Tool responses were always token-budgeted;
truncation is no longer silent. Dropped content is stored in the omission
store and surfaced via a
_meta.omittedenvelope (refs,tokens,restore);get_symbolresolvesrepowise#<ref>omission refs (with an optionalqueryparameter) alongside symbol ids — the tool count stays at nine. One durable store (.repowise/omissions/, TTL + size-cap pruned) serves the CLI, the hook, and MCP. - Distill config & doctor checks. A
distill:block in.repowise/config.yaml(master switch, hook permission posture, per-family overrides, disabled filters, omission-store TTL/size).repowise doctorvalidates the block, reports omission-store size against its cap, and shows rewrite-hook install state. - Distill on Codex CLI. The rewrite hook now supports Codex CLI alongside
Claude Code, with repowise-command corrections (#391). A lint filter joins
the filter set,
repowise saveddiscovers missed savings, and the ledger tags savings per surface (#390). - Multi-language import resolution. Lightweight per-language import resolvers with same-scope linking and spec metadata sharpen the dependency graph across the language registry (#392).
- Light-default design-token theme system. The web UI moves to a design-token theme with light as the default and a dual-theme component sweep (#405).
- Owl mascot init banner.
repowise initopens with the owl mascot and a repo-seeded heatmap wordmark (#364), plus a clearer mode panel and searchable model selection (#379). - Agent-provenance layer in the git indexer. Commit indexing records a deterministic agent-provenance layer (#366).
- Claude Code plugin. A
repowiseClaude Code plugin and root marketplace, refreshed to the current command/skill/MCP surface (#356).
Changed
repowise initdefaults the distill rewrite-hook prompt to yes (#409), and every init flow records the verdict (#382).- Indexing and incremental updates scale with change size. Parsing and betweenness results are cached across incremental updates (#369, #368), workspace indexing routes already-indexed repos through the incremental path (#384), and filesystem walks prune nested repos and junk trees (#380).
Fixed
get_contexthardened — segment-boundary partial matching, git-file fall-through, and batch isolation (#401).- Distill correctness: Grep-rescue fixes, PowerShell hook coverage, a nudge floor, and allowlist seeding (#389).
- Health scoring:
duplication_pctcomputed from the union of clone ranges (#388), whole-file NLOC for file metrics instead of function-body sums (#387), and hotspot/ownership signals calibrated for small teams and quiet repos (#363). - Submodules: persisted include-submodules flags are honored in health, dead-code, incremental updates, and upgrades (#383, #381).
- Dead code: local Express route middleware rescued from unused-export detection (#386); explicit relative JS imports resolve (#376).
- Process hygiene: MCP orphan watchdog, live-PID update locks, and PATH-hijack-proof registration (#385).
- Generation: never-started page coroutines are closed on cancellation (#365).
- Server: jobs honor
exclude_patternsand prune stale rows (#354); breadcrumb path labels are decoded in the web UI (#359).
Dependencies
- starlette 0.52.1 → 1.0.1 (#367).
[0.16.0] — 2026-06-03
Added
- Codex CLI provider and project integration. A new local Codex CLI LLM provider runs documentation generation through authenticated
codex execsessions (argv-based subprocess, JSONL parsing tolerant of non-JSON noise, async concurrency cap, exec timeout, and zero-cost subscription usage tracking). Adds project-local Codex setup — a.codex/config.tomlMCP server,.codexlifecycle hooks,.codex-pluginmetadata and marketplace entry, and managedAGENTS.mdgeneration. Reasoning effort is now wired across all LLM providers with per-provider model discovery and supported reasoning modes (#348). - Native Ollama embedder. Semantic indexing can now embed through a local Ollama instance directly, without routing through an OpenAI-compatible shim (#331).
repowise init --resumeactually resumes. Persistence is split into per-phase persisters (ingestion, git, analysis, generation) so a re-run skips phases that already completed instead of redoing the whole pipeline. Public API and end-state are unchanged (#343).- Advisory CLI-version check in
repowise doctor.doctornow shows current vs. latest published version and the exact install-method-aware upgrade command (uv tool / pipx / pip / editable). Advisory only — it never auto-updates, never flips doctor's pass/fail, and swallows network errors (#346, closes #338).
Fixed
- Owner "last touched" reflects your own last commit. Previously a teammate's commit to a file you co-own bumped your "last touched" timestamp. Each author's own first/last commit timestamps are now recorded and aggregated, and author identity is read through git's
.mailmapso one person's multiple names/emails fold into a single contributor (#349). repowise updatere-runs health when config changes even if git is unchanged. Editingexclude_patternsorhealth-rules.jsonused to have no effect until a code change touched each file.updatenow fingerprints the config and triggers a full health rescore when it changes (#337).- Minified/generated bundles can no longer wedge
init. Duplication detection's O(k²) window comparison could explode on checked-in minified chunks, leavinginitstuck athealth 0/N. New layered resource guards skip minified files, cap per-file tokens and the repo-wide window budget, and drop degenerate hash buckets (#342, closes #341). exclude_patternsare enforced in MCP tool responses at query time. Rows that predate anexclude_patternschange are now filtered out of every structured tool (context, answer, search, health, overview, dead_code, risk, and the rest) and out of aggregate KPIs, so excluded files never leak back into results or numbers (#339, #340).- User-added MCP
envsurvives re-registration.repowise init/updatere-registration did a shallow replace of therepowiseMCP server entry, silently wiping any user-addedenvblock (BYOK provider/embedder keys) and degrading semantic search to the mock embedder. Server definitions are now deep-merged (#336, fixes #307). - Cost tracking no longer wedges
repowise updateondatabase is locked. A second cost-tracking engine inserting per LLM call lost WAL's single-writer race against the doc-generation writer, blocking the full busy-timeout per call and turningupdateinto an effectively non-terminating run. Persistence is now best-effort, plus a--no-cost-trackingflag andREPOWISE_NO_COST_TRACKINGenv var to opt out entirely (#330, closes #326). - Rust sibling test modules are kept live in dead-code reachability instead of being flagged as unused (#332).
reindexuses the shared database engine rather than opening its own (#333).
Documentation
- README and linked docs revamped for accuracy, with a sharpened tagline, reframed layer positioning, and fixed README/CONTRIBUTING links (#334, #335, #345).
[0.15.2] — 2026-05-31
Added
on_page_readystreaming callback.run_pipeline/run_generation/PageGenerator.generate_allnow accept an optionalon_page_readycallback, invoked with each page the moment it is generated (alongside the existingon_page_done, which only receives the page type). Lets callers persist or stream pages incrementally — e.g. flush pages to storage per page so a generation cut-off yields a partial-but-usable set rather than nothing. Additive and backward-compatible; best-effort (a sink error is logged and never aborts a run) (#328).
[0.15.1] — 2026-05-30
Fixed
- Misconfigured embedders no longer fail silently. When an embedder is explicitly configured (
REPOWISE_EMBEDDERor.repowise/config.yaml) but can't initialise — most often a missing API key — the MCP server used to fall back to the mock embedder with only aWARNING, then report healthy while semantic search (search_codebase,get_answer) ran on vectors that can't match the real index. The failure is now logged atERRORwith the missing key and remediation, and surfaced in every tool's_metaenvelope (embedder_degraded: true) so it's detectable instead of masquerading as a healthy server. Embedder resolution also goes through the shared registry, soopenrouterand custom-registered embedders are honoured — not justopenai/gemini(#324). - Indexing artifacts serialize reliably. A transient blame index is dropped before artifact serialization, fixing a failure that could corrupt published index artifacts (#323).
[0.15.0] — 2026-05-30
Added
- Code-health biomarkers overhaul — calibrated, multi-language, process-aware. The health model is reworked into a broad biomarker suite whose weights are calibrated against real defect data (#305): class-level cohesion and god-class detection (#302), test-quality smells with hardened size sensitivity (#303), change-entropy and co-change scatter signals (#301), ownership and relative-churn process signals (#300), and a prior-defect process signal (#312). Coverage deductions now scale by the uncovered fraction rather than a flat penalty (#314), primitive-obsession is suppressed in tiny modules to cut false positives (#313), and the biomarkers extend to Kotlin, C++, and C# (#316).
repowise risk— just-in-time change-risk scoring. A new command and scoring pass estimate the risk of changing a file from churn, complexity, ownership, and defect history, surfaced both in the CLI and the dashboard (#315).- Commits change-risk page with per-function blame. A new commits explorer surfaces per-commit change-risk history (#317), change-complexity and defect-history signals (#318), and a per-function blame view with a coverage-gradient breakdown (#319).
- Coverage ingestion.
repowiseingests normalized-JSON coverage reports and surfaces coverage metrics across the health surfaces (#309). .mts/.ctsare treated as TypeScript for indexing and language detection (#310).
Fixed
exclude_patternsare enforced in the git indexer and dynamic-edge passes — excluded paths no longer leak back in through commit history or dynamic edges (#308). Index-only runs now persistexclude_patternsviasave_config_partialso subsequent updates honour them (#297).- Single-file re-index no longer wipes all dead-code findings — an incremental re-index of one file previously cleared the whole findings set (#298).
- CommonJS
require()is resolved so property-access calls on a required module are no longer mis-flagged as dead (#299). @repowise-dev/uiuses extensionless relative imports package-wide, fixing a web build failure where.js-suffixed relative specifiers failed to resolve (#320).
Documentation
- README header refreshed with a banner and badges (#311).
[0.14.0] — 2026-05-28
Added
- JVM (Java + Kotlin) indexing brought up to C# / Rust / Go parity — 5 PRs. A
JvmPackageIndexworkspace model + Maven / Gradle root recognition treats Java/Kotlin packages as the unit of reachability (so helpers next to a@SpringBootApplicationare no longer mis-flagged), annotation processors generate symbol stubs for@Generatedcompanions, type-ref resolution mirrors the Go/Rust pattern (field/parameter/return types resolve against imports), and JVM dead-code hardening recognises framework annotations (@Component,@RestController,@Configuration, JPA@Entity,@SpringBootTest) as live-from-framework entries. Spring expansion, Jakarta, Quarkus, Micronaut, and Android-component edges are emitted along with dynamic-hints for reflective lookups, JNDI, and serialization (#273, #274, #275, #279, #280). - C++ indexing hardened across symbol graph, dead-code, and framework recognition — 6 PRs. Workspace-aware include resolution walks the CMake/Bazel project tree to find headers (with a public-header rescue pass for
include/-style layouts), the symbol graph captures lambda captures + type references + synthesized destructors and entry markers, dead-code reachability respects contracts ([[nodiscard]],[[maybe_unused]], virtual overrides) and never-flagsWebAssembly/Emscriptenexports, dynamic markers cover function pointers and reflective lookups, tests / benchmarks / fuzzers are recognised via gtest / Catch2 / Google Benchmark / LibFuzzer entry points (broadtests/glob, pluralbenchmarks/, embedded pybind11 modules), and compiler-builtin macros (__has_attribute,__has_builtin) no longer mask reachable code (#281, #282, #283, #284, #285, #286). - JS/TS indexing brought up to C# / Rust / Go parity. Closes the gap with the strongly-typed languages — package-aware workspace model, type-ref resolution against
import typestatements, and dead-code never-flags for the standard JS entry points (#272).
Fixed
repowise initno longer crashes mid-pipeline on Windows shells defaulting to cp1252. Rich's legacy Windows renderer encodes every printed line through the active code page, so the first↳or✓glyph in the progress UI raisedUnicodeEncodeErrorand aborted the run, leaving partial state behind.sys.stdoutandsys.stderrare now reconfigured to UTF-8 witherrors="replace"before any Rich Console is constructed, so oldcmd.exeand PowerShell sessions render cleanly without needingPYTHONIOENCODING=utf-8(#290, closes #271).- Upgrading repowise no longer crashes with
no such column: decision_records.verification(or similar).Base.metadata.create_allonly creates missing tables — it never ALTERs existing ones, so any user who indexed a repo with an older release and thenpip install --upgrade'd would hit crypticOperationalErrors the moment a code path queried a column added by a later release.init_dbnow reconciles additive schema drift generically: walks every table inBase.metadata, adds any model-declared columns + indexes missing from the live schema, and synthesizes a DDLDEFAULTfrom a static Pythondefault=value so NOT NULL back-fills onto populated tables don't fail. Any future column/index added to the model is picked up automatically with zero code touch (#292). repowise serveno longer crashes when port 3000 is already taken. If the user's project already binds the default web-UI port (3000) or API port (7337), the server now probes for the next available port within a 20-port scan window and prints a yellow "using N instead" notice, instead of aborting after a clean API startup withEADDRINUSE. Falls back to an OS-assigned ephemeral port if the whole window is busy (#287, closes #232).repowise servedetects old Node.js and falls back gracefully. Previously the only check was whethernodewas on PATH, so a stale Node.js (e.g. 12 on a WSL setup) slipped past the gate and then crashed the bundled Next.js 15 runtime withSyntaxError: Unexpected token '?'. The serve command now parsesnode --version, compares it to the minimum required by the web UI (Node 20+), and falls back to API-only mode with a clear "upgrade Node.js" message when too old (#289, closes #276).- MCP registration is workspace-aware.
repowise initagainst a sibling repo in a multi-repo workspace used to overwrite~/.claude/settings.jsonwith the per-repo path, silently breaking workspace mode the moment a second repo was indexed. Registration now targets the workspace root when.repowise-workspace.yamlis found in any ancestor, so subsequent inits converge on the same entry andrepo="<alias>"queries work across all repos (#278, closes #277). repowise init"What's next" panel — spacing + headline command. Long commands likerepowise init --provider gemini(>28 chars) used to run straight into their description (geminigenerate full documentation) because the format spec only left-padded short commands. Always inserts at least one space now. For the index-only path, the headline next-step isrepowise serve(which actually launches the dashboard) rather thanrepowise mcp .(whichinithas already auto-registered) (#291).
[0.13.0] — 2026-05-27
Added
- Rust indexing brought up to C# parity — 26 fixes across 8 waves. Cargo workspace member globs (
crates/*) now resolve, visibility modifiers are captured for structs / enums / traits / consts / types / modules (distinguishingpub(crate)/pub(super)frompub),mod foo;declarations register as import edges, andself::/super::/crate::prefixes are recognised as relative imports. Rust never-flag patterns coverbuild.rs,examples/,benches/,tests/,src/bin/, and fuzz targets, plus dynamic-import markers for trait objects, FFI, serde, and conditional compilation. On Typst, dead-code findings dropped 1,013 → 244 and unreachable files 92 → 40 (#251). - Go indexing brought up to C#/Rust parity. A new
GoPackageIndexworkspace model and package-index warmup phase resolve package imports to all files in the package, and a package-aware reachability pass computes reachability at package (directory) granularity rather than per file — so helpers sitting next tomain.goare no longer mis-flagged. Go never-flag patterns cover test files,cmd/*/main.goand rootmain.goentry points,doc.gopackage-doc stubs,magefile.go, and generated code (*.pb.go,*_string.go,*zz_generated*.go,*bindata.go,*_gen.go,*.gen.go);init/TestMainare entry symbols (#267). - Wiki information-architecture upgrade. The docs hub gains a semantic "By domain" tree (Guided Tour → Architecture → Modules → Reference) as the default with a "By folder" toggle, hierarchical breadcrumbs with sibling prev/next, an in-page table of contents, and a "Start here" front-door panel promoted from a collapsed section. Inline backtick references render as clickable internal links, a ⌘K / Ctrl-K full-text command palette searches the loaded page list, and a sidebar surfaces related pages (forward links) and backlinks ("Linked by"). Pages carry page-type and "in {module}" zoom-out chips and an inline low-confidence banner. All of it lives in the shared
packages/uisurface (#238). - Decision layer overhaul — provenance, more sources, surfaced everywhere. Decision extraction now mines four new sources (deterministic ADR auto-discovery, CHANGELOG mining, PR/squash-body mining, centrality-bounded comment archaeology) in addition to commits, whose multi-line bodies are now captured. A new
decision_evidencetable records one-decision-to-many-evidence with a source-ranking ladder and verification status; matching extractions merge into a single record that accretes evidence instead of discarding duplicates, and confidence is derived from source rank, corroboration, and verification. A post-extraction substring gate drops LLM fields not grounded in their verbatim source span (#239).
Changed
- Dead-code accuracy for embedded JS and C/WASM. C functions exported across the JS↔WASM boundary via
EMSCRIPTEN_KEEPALIVE/WASM_EXPORTmacros or__attribute__((export_name(...)))/((used))are recognised as exported rather than dead (their caller is the host runtime). C/C++ type-reference resolution mirrors the Go/Rust strategy: struct / typedef / class types used as field, parameter, or return types now resolve against#included headers, so header structs are no longer read as unused exports. Also fixes a C/C++ include resolver bug where a repo-relative path was resolved against the process CWD, causing it to miss and fall back to a wrong stem match (#268). - Web navigation restructure and docs performance pass. Trend / Coverage / Refactoring fold back into Health, Security moves from a sidebar item to a tab under Risk (old route redirected), Docs is renamed Wiki and surfaced near the top of the nav, and the docs landing opens the repo overview by default with a collapsible "Start here". A global SWR config disables revalidate-on-focus/reconnect and dedupes requests, the full page list is fetched once instead of twice, and rendered wiki markdown is memoized so persona/sidebar toggles no longer re-parse the document (#240).
- Internal restructuring across CLI, ingestion, generation, persistence, server, MCP, and web. A sweep of large modules was decomposed into focused packages —
crud.py,schemas.py,tool_context.py,tool_answer.py,framework_edges.py,context_assembler,routers/graph.py, the pipeline phase functions, the CLIui.py/init/updatecommands, the language specs, and the webapi/types.ts— with no change to install or serve behaviour (#246–#265).
Fixed
- TS/JS re-export barrels no longer flag forwarded symbols as dead.
export { X } from "./x"andexport * from "./y"now produce graph edges, so a component reached only through anindex.tsre-export chain (the standard component-library barrel) is correctly seen as used.index.*/__init__.pybarrels are skipped in the unreachable-file pass without affecting the unused-export pass, so a real symbol defined in a barrel is still flagged (#245). - Python imports that hid live code as dead are now resolved. A source-root-aware module index resolves absolute imports under nested or namespace source roots (src layout,
packages/*/src, PEP 420), aliased namespace imports (from . import levels as _levels) record the module name so bare-relative expansion works, and a dynamic-import hints extractor emits edges forimportlib/string-based plugin registries (#244). --format json/mdoutput is no longer polluted by logs.repowise healthandrepowise dead-codeemitted structlog/stdlib info/debug lines on stdout, making the output unparseable byjqand other consumers; logs are now silenced before the ingestion pipeline starts whenever the format is nottable(#242).- Changelog decision source discovered under
docs/,doc/,.github/. Decision extraction only globbed the repository root, so projects keeping their changelog underdocs/had that entire source silently skipped; the conventional documentation subdirectories are now searched (root still first) (#243). - No more "Event loop is closed" traceback at the end of
repowise init. All five LLM providers recorded per-generation cost via a fire-and-forget task that could outlive theasyncio.runloop; cost recording is now awaited inline (#241). - Reindexing after an embedder change no longer fails with an opaque LanceDB error. Switching embedders (e.g. a mock dim 8 → OpenAI dim 1536) left the
wiki_pagestable on the old vector schema, so every write failed deep inside LanceDB with an IO error that never mentioned dimensions; the table is now dropped and recreated when the stored vector dimension differs from the current embedder's output (#266).
Chore
- Added
.well-known/funding-manifest-urlsso thefunding.jsonmanifest can verify repository ownership for the FLOSS/fund directory (#237).
[0.12.0] — 2026-05-25
Added
- Knowledge Graph visualization — full C4 revamp. The C4 diagram page is renamed to "Knowledge Graph" across all user-facing surfaces (sidebar, breadcrumb, page headers). Edges now resolve to distinct warm-palette colors by relationship type with animated flowing dashes, relationship labels ("imports", "depends on"), and arrowhead markers. Nodes are larger with colored left accent borders via the tone system and complexity badges on layer cards. Selecting a node dims unrelated nodes to 25% opacity. Backend adds an architecture view API, DB-first layer/tour loading, a KG enrichment pipeline with fingerprint-based skip logic, and file-level health scoring (#235).
Changed
- Near-linear scaling restored for .NET import resolver. Three independent algorithmic fixes — bucketing files by project reduced from O(N × M × depth) to a single parent-chain walk against a precomputed dict; type-ref ranking memoises
Path.resolve()per source file; using-directive resolution caches the importer path. Combined dotnet phases drop ~70% wall-clock on a 2000-file synthetic repo, and scaling at 4× file count goes from ~20× to ~7.4× (#233).
Documentation
- Condensed benchmark section to a single paragraph (#231).
- Refreshed README, added COMMERCIAL.md, fixed layer/tool/biomarker counts (#230).
[0.11.0] — 2026-05-24
Added
- Fast index mode + incremental fast→full upgrade.
repowise init --mode fastdoes a quick first pass on very large repos — builds the dependency graph and indexes only the essential git tier (last commits, no per-file blame or co-change), and skips LLM doc generation.repowise update --fullthen upgrades that index to a full one incrementally: it backfills the git tier to FULL (per-file blame + repo-wide co-change) via a resumable, checkpointed worker, rehydrates the persisted graph from SQL rather than re-parsing and re-resolving it, and generates the docs fast mode skipped. Because the expensive import/call/heritage resolution and centrality computation are reused, the upgrade is measurably cheaper than re-running a fullinit(~14× faster on the avoided structural work at 2k files). The backfill is resumable — an interruptedupdate --fullpicks up where it left off (#220, #224). - Four new code-health biomarkers.
hidden_couplingflags pairs of files that consistently change in the same commits without an explicit import/dependency edge — behavioral coupling static analysis can't see.complex_conditionalcatches branch/loop guards combining three or more boolean operators (severity grows with operator count).function_hotspotflags functions that are both structurally complex and frequently modified, andcode_age_volatilityflags old, settled functions that are suddenly being edited — both computed from a per-line blame index built once per file. The three git-derived biomarkers are tier-aware: they no-op on an ESSENTIAL-tier (fast) index and light up once the FULL git tier is present (#221, #222). - Pluggable storage seams + capability registries. New async
IndexStore/GraphStore/JobStoreinterfaces with SQL/in-process default implementations, and process-widecli_registry/mcp_tool_registry/pipeline_hooksregistries so downstream packages can extend the CLI, MCP tool list, and pipeline phases without monkey-patching internals. Behavior is unchanged for OSS users — same CLI commands, same MCP tools, same default storage (#219).
Changed
- Code-health scoring recalibrated. The organizational category cap is lifted from −1.0 to −3.5 so the strongest empirical predictors (
developer_congestion,untested_hotspot,hidden_coupling) are no longer suppressed, and a per-biomarker weight multiplier was added (developer_congestion×1.5,untested_hotspot×1.3,function_hotspot×1.2).knowledge_lossis de-rated to ×0.4 per OSS calibration (legacy code that works gets handed off) — enterprise users can raise it back via per-repo overrides. See the updated category-cap table indocs/layers/CODE_HEALTH.md(#221). - Health web UI surfaces the new biomarkers. Glossary entries and biomarker-specific detail views for all four new biomarkers (partner-file chip for
hidden_coupling, operator count forcomplex_conditional, mod/p80 ratio forfunction_hotspot, median age + recent edits forcode_age_volatility), recalibrated category caps in the score breakdown (now sorted by applied deduction), and a clickablefunction:linedeep-link in the file drawer. The new biomarker details also flow into the AI refactor prompt (#223). - Doc generation scales to very large repos. Batch embedding (one model call per generation level instead of N upserts), pipeline checkpoint/resume over the new JobStore seam, and graph metrics computed in SQL. Default behavior is unchanged when the new knobs aren't set (#220).
Fixed
repowise update --fullnow recomputes code health at the FULL tier. The upgrade backfilled the git tier and regenerated docs but never re-ran the health analysis, so the persisted health tables stayed frozen at the fast index's ESSENTIAL-tier findings — the blame/co-change biomarkers stayed invisible after an upgrade. The upgrade now runs a full-repo health pass against the rehydrated graph and persists findings/metrics/snapshot, matching whatinitand the normalupdatepath do (#225).
Documentation
- README mentions the Repowise PR Bot in the hosted-version section (#217).
[0.10.0] — 2026-05-18
Added
- Code health layer — a fifth intelligence layer alongside graph / git / docs / decisions. New
repowise healthcommand,health_*SQLite tables, biomarker engine, and/repos/[id]/healthweb UI surface what hotspots actually cost to maintain. Tree-sitter complexity walker feeds 10+ biomarkers across structural (large_method,nested_complexity,repeated_nesting,complex_method,brain_method,primitive_obsession), organizational (developer_congestion,knowledge_loss), test (untested_hotspot,coverage_gap), and DRY (dry_violation, backed by a tokenizer + Rabin-Karp clone detector with co-change correlation) categories. A composite 0–10 score is rolled up per file, per module (NLOC-weighted), and per repo. Findings persist with deterministic refactoring suggestions and anacknowledged | resolved | false_positivelifecycle.HealthSnapshotwriter + trend detector surface declining-health alerts..repowise/health-rules.jsonsupports per-file overrides.repowise updateruns an incremental upsert so the dashboard stays fresh without re-running the full pipeline. Newrepowise health --trend,--refactoring-targets,--module,--file, and--coverageflags;repowise statusprints a one-line health digest. Parallel biomarker analysis viaasyncio.gather. Full architecture deep-dive indocs/layers/CODE_HEALTH.md(#212). - Coverage ingestion (LCOV / Cobertura / Clover).
repowise health --coverage report.lcovparses one or more reports (format auto-detected, override with--coverage-format), persists per-file line + branch + covered-line sets to a newcoverage_filestable, and feeds two coverage-aware biomarkers —untested_hotspot(hotspot files with low coverage or, when no coverage is ingested, no paired test file) andcoverage_gap(significant uncovered surface area on non-test files). New/api/repos/{id}/health/coverageendpoint +/repos/[id]/health/coveragepage with risk × coverage scatter, module rollup, and per-file drilldown (#212). - Code-health web UX overhaul — tabbed chrome, trend, scatter quadrants, file drawer.
/healthis now a four-tab surface (Overview / Trend / Coverage / Refactoring) with sharedHealthPageChrome, sparklines pulled from/health/trend, a 5th "Hotspot Health" KPI card, server-side paginated file table with sortable headers + filter chips (Hotspots / Untested / Failing) + path search, biomarker glossary tooltips, severity-distribution bars, slide-overHealthFileDrawerwith score breakdown by category, impact × effort quadrant on the refactoring page, and one-click status mutation (Acknowledge / Resolved / False positive) wired to a newPATCH /health/findings/{id}endpoint. InlineHealthBadgechips appear on the hotspots / ownership / graph views so health context follows you across the app (#212). - AI fix / test prompts on the refactoring and coverage pages. Per-row
AI fix prompt/AI test promptbuttons open a modal that picks a target agent (Generic / Claude Code / Cursor), previews the generated prompt (with biomarkers, line ranges, severities, score deductions, suggested directions, hard constraints, completion contract — and the explicit "verify each finding against the real code; treat analyzer output as leads, not ground truth" preamble), and copies to clipboard. Prompt builder is generic (buildAiPrompt/buildCoverageAiPrompt) so future surfaces can reuse it (#212). - New
get_healthMCP tool + health enrichment on existing tools.get_health(include=['coverage'])returns the score + biomarker breakdown + coverage summary;get_context(include=['health'])surfaces per-file score, top two biomarkers, and linked coverage row;get_riskrows gaincoverage_pct/branch_coverage_pct;get_overviewexposes hotspot-health KPIs. Auto-generatedCLAUDE.mdgains a Code Health section listing critical biomarkers so agents see the health context on every invocation (#212). - C4 architecture diagrams (L1 System Context / L2 Containers / L3 Components). New
/repos/[id]/c4page (React Flow + ELK layout, URL-synced via?level=and?container=with nuqs), backed byservices/c4_builder/and three endpoints under/api/graph/{repo_id}/c4/{l1,l2,l3}. Container detection re-uses manifest paths from a newexternal_systemstable (manifest parsers for npm / PyPI / Cargo / Go / NuGet that capture name + version + ecosystem + heuristic category). Containers fall back to top-level dirs on repos without manifests. L3 components are subdirs inside a container. Inspector panel surfaces module-health context per component. SVG / PNG / Mermaid export menu with a/c4/mermaidserver endpoint. Shared UI lives inpackages/ui/src/c4/so the hosted product can reuse it (#204, closes #203). - MCP tool surface bumped 7 → 8 with
get_symbolexposed.get_symbol("path::Name")returns raw source bytes for one indexed symbol with exact line bounds — cheaper and safer thanRead+ offset math.get_contextwas trimmed to a triage card (title, summary, signatures,hotspotbit,decision_recordstitles,symbol_idpointers) — theinclude=["source"]mode was removed; agents should pipesymbol_idintoget_symbolinstead.get_riskPR mode now emits a structureddirectiveblock (will_break/missing_cochanges/missing_tests) with capped co-change / transitive lists.search_codebasegains akindfilter and a per-resultsearch_method(embeddingvsbm25fallback) plus a bareword-identifier grep hint. Every response carries an_metaenvelope (index_age_days,indexed_commit); astale_warningfires only when the indexed HEAD actually diverges from.git/HEAD, so silence is trustworthy (#210). get_answerrewritten as a hybrid retrieval pipeline. FTS + vector store run in parallel, merged via reciprocal-rank fusion, PageRank-biased, expanded one graph hop to rescue near-misses, fused with decision records on "why"-shaped questions, and prepended with a structured prelude (top symbols, recent significant commits, decision titles). Confidence andretrieval_qualityreport independently so synthesis quality and retrieval quality don't get conflated. Low-confidence returns now shipbest_guesseswith one-line justifications instead of an empty answer. Schema-versioned cache auto-invalidates earlier-pipeline payloads (#210).- Doc-generation upgrade — enforced coverage budget, faster runs, wiki interlinking. New
generation/selection/package is the single source of truth for which pages get emitted;PageGeneratorandcost_estimatorboth consume it, so the pre-run estimate can no longer drift from the actual run.GenerationConfig.coverage_pct(default 0.20) is the user-facing knob, with per-bucket shares acrossfile_page/symbol_spotlight/module_page/api_contract/infra_page/scc_page— no more bypass paths around the budget. New interactive coverage chooser shows per-bucket counts and a cost range (10 / 15 / 20 / 30 / 40 / 50 %) computed from live ingestion data, with self-calibration from prior runs'wiki_pages;--coverageCLI flag for non-TTY use. Prompt caching via provider-agnosticCacheHint(Anthropic emitscache_control, OpenAI surfacescached_tokens). Persistent cross-run page cache short-circuits the LLM call whensource_hash + modelmatch. Module pages now group by graph community (defaultmin_module_size=3) instead of top-level dir — handful of generic per-directory pages → 30–80 genuinely scoped module pages on large repos. Dead-code findings, decisions, and external systems now flow intofile_page/module_page/repo_overviewcontexts. Newapi_contract_detectorroutes FastAPI routers and ASP.NET controllers through the dedicatedapi_contracttemplate (#208). - Onboarding documentation collection — 8 curated pages, default-on at
repowise init. Newgeneration/onboarding/subpackage emits up to eight gated subkinds (codebase_mapalways;getting_started,key_concepts,how_it_works,development_guide,active_landscapegated on signals like manifest presence, PageRank P90 symbols, execution-flow depth, suffix patterns, recent commit volume) plus two promoted slots (project_overview,architecture_guide) that reuse the existingrepo_overview/architecture_diagrampages viametadata.onboarding_slot. UI renders an "Onboarding" folder at the top of the docs tree (Compass icon, auto-expanded, canonical slot order).--onboarding/--no-onboardingflag onrepowise init, persisted toconfig.yaml(#208). - Wiki interlinking. Post-gen regex scan resolves backtick refs to other pages'
page_ids and populatesmetadata.wiki_links+ reverse-indexmetadata.backlinks. NewWikiLinkMDX component renders resolved refs as clickable anchors; newBacklinksPanelin the wiki sidebar surfaces pages linking into the current one (#208). - Pipeline progress for previously silent phases. Added phase events around
tsconfig(TS path-alias resolver init),dynamic_hints(HintRegistry edge extraction), andexternal_systems(manifest parsing). The two graph aggregations (graph.metrics,graph.communities) now emit per-algorithm completion lines as eachasyncio.gathertask finishes — output makes it obvious which algorithm is the bottleneck (almost always betweenness on the symbol graph for medium+ repos) (#206).
Changed
- Dead-code analyzer cuts ~390 false positives across resolver, parser, and analyzer. Alembic migration scripts under
*/alembic/versions/*.pyare never-flagged (reflectively loaded). Click / Typer decorators on locally-named groups (@my_cli.command("add")) are now recognised via suffix matching.unused_internalcounts an incomingimportsedge whoseimported_nameslists the symbol — catches dispatch-table patterns (HANDLERS = {"python": _extract_python, ...}). Entry-point allowlist extends to WSGI / ASGI / Flask / FastAPI factory conventions (create_app,make_app,application,get_asgi_application, …). Bare relative Python imports (from . import a, b) now expand into per-nameImportobjects so plugin barrels resolve. Symbol extraction skips any AST node nested inside a callable, so React handler closures and async-method-local coroutines no longer hoist to the top-level symbol list..tsxfiles now parse with tree-sitter's JSX-aware grammar. JSX elements (<Component />) register as call sites for the named component. Public symbols in files imported as namespaces (from . import cargo,import * as cargo from "./cargo") are rescued — the static graph can't tell which attribute is being dispatched, so flagging individuals yields guaranteed false positives. TS workspace resolver honourspackage.json#exports(conditional, wildcard, longest-prefix) so turborepo / nx / pnpm monorepos resolve through subpath exports — verified−23 %dead-code findings on the dogfooded monorepo. Win32 entry points (wWinMain,WinMain,wmain,ServiceMain,LowLevelKeyboardProc, MSTest macro family) and never-flag globs for precompiled-header anchors (pch.h/stdafx.cpp), COM*ClassFactory.cpp, and broader test-project conventions skipped on C++ codebases — roughly 520 high-confidence findings cleared on PowerToys (#194, #207). - Ingestion treats nested git repos as traversal boundaries by default. When a working tree physically contains other independent git repositories as subdirectories (workspace roots that are themselves versioned, sibling repos checked out under a parent), the traverser walked into them and pulled in their entire file trees. Now a
.gitentry (directory, submodule file, or external gitdir pointer) is a hard traversal boundary. Opt-ininclude_nested_repos=Truepreserves the old behaviour. Newskipped_nested_repocounter surfaced in the filtering summary (#205). - CLI editor setup refactored into an integrations package with per-editor strategy classes (Claude / Cursor / generic) so adding a new editor is a new file rather than edits across the suite (#199).
- Dynamic-hint extraction is no longer the wall-clock-stall phase. The 13 dynamic-hint extractors used to call
repo_root.rglob(pattern)independently — each descended intonode_modules,.venv,.next,__pycache__, and on Windows followed directory junctions into infinite loops. New_walk.iter_globhelper doesos.walkwith in-place dirname pruning,followlinks=False, realpath-based cycle detection, and a hard depth cap of 64.HintRegistry.extract_allruns the 13 extractors in aThreadPoolExecutor. Walk completes in ~2 min on polyrepos with recursive junctions vs. the prior indefinite stall (#208). - Embedder no longer blocks the LLM critical path. Post-LLM embed-and-upsert spawned as a background task so the next wave's LLM calls start immediately; the level still drains pending embeds before advancing so the next level's RAG search sees a fully-indexed store. New
enable_rag_contextandrag_min_store_sizeconfig knobs short-circuit the RAG search on cheap models and on early pages before the store has enough indexed material to return useful hits (#208). - OpenAI default model bumped from
gpt-4.1togpt-5.4-nanoin both the interactive provider picker and the web settings placeholder, to match the in-app cost-tier recommendation (#208). - Tighter README tagline and corpus framing. New tagline: "The codebase intelligence layer for your AI coding agent." Drops the misleading "500 commits" references (the cap is per-file, not a global corpus cap) and softens engineering-team-only framing so solo devs see themselves in the README (#196). Refreshed
webui.gif, compressed 16.6 MB → 8.5 MB (#198).
Fixed
repowise updatepost-commit auto-sync rewrite. The hook was racing with itself — concurrent invocations from rapid commits all started from the same stale base, took 12+ minutes each, never converged, and discarded output to/dev/null.repowise updatenow enforces single-flight via.repowise/.update.lock; if another update is running, the new invocation writes.update.pendingwith the current HEAD and exits, and the running update rolls forward to it. Hook pre-writes.update.queuedsynchronously before backgrounding so the augment hook sees an in-flight marker during the start-up window. Augment hook emits "Wiki update in background — started Ns ago, target X" instead of "Wiki is stale" when a marker is present. Stdout / stderr captured to.repowise/.update.log(rotated to 64 KB tail when it exceeds 256 KB) so silent failures are diagnosable. Hook installer upgrades the marker block in place when the body differs (previously bailed with "already installed" and left users stuck on the buggy hook after a repowise upgrade). Cross-platform — git always runs hooks under POSIX sh, so the same script body works on Linux / macOS / Windows (#211).- CI integration test failures introduced by the doc-generation PR.
_TrackingProvidermock didn't acceptcache_hints;_SlowVectorStoremock didn't implementget_page_summaries_by_paths;test_level_values_in_rangeasserted level ≤ 7 but onboarding has been level 8 since phase 3. Also:select_pagesnow allocates all candidates when total supply ≤ budget (socoverage_pct=1.0on tiny repos returns pages instead of zero), andscore_fileapplies a tiny per-symbol floor so leaf modules with zero PageRank still enter the candidate pool (#209). - Packaging: three
__init__.pyfiles silently dropped by a local_*.pyexclude rule in.git/info/exclude—selection,cost_estimator,onboarding,external_systems, anddynamic_hints/_walk.pyall hit "module has no attribute" import errors on CI before this was caught. Thec4_builder__init__.pywas hit by the same rule and force-added in #204.
Documentation
CODE_HEALTH.mduser guide + CLI reference entry + READMEs touched up to mention the fifth intelligence layer and the 8-tool MCP surface. The MCP-tools table now leads each row with what only that tool answers and surfaces the new signals (retrieval_quality,best_guesses,search_method,hotspotbit,decision_recordspointer, PR-mode directive block,_meta.stale_warning) (#210, #212).- Removed
AUDIT_NOTES.md— internal scratchpad, not intended to ship in the public repo (#197).
[0.9.1] — 2026-05-13
Fixed
repowise serve404 on the web tarball for v0.9.0. The v0.9.0 publish workflow failed during the web build: twopackages/uicomponents introduced this release (attention-panel.tsx,co-change-list.tsx) importeduseStatewithout a"use client"directive, so Next.js' RSC compiler rejected them whenpackages/webpulled them transitively via the overview page and wiki git-history panel. The Python wheel published to PyPI but norepowise-web.tar.gzwas attached to the v0.9.0 GitHub release, so end-userrepowise servefalls through to "API only". v0.9.1 adds the missing"use client"directive at the top of both files. Anyone who installed 0.9.0 should upgrade to 0.9.1 to get a working web UI fromrepowise serve.
[0.9.0] — 2026-05-13
Added
- Contributor profiles, module health, and reviewer suggestions. New engineering-leader views composed from existing git metadata + dead-code rows. New endpoints
/api/repos/{id}/owners,/owners/{key:path},/modules/health(list + detail), and/reviewer-suggestions?paths=return paginated contributor directory, full per-owner profile (files owned, hotspots, dead-code burden, bus-factor risk, top files, co-authors, commit mix), composite 0–100 module-health scores, and confidence-ranked reviewers. New/repos/[id]/ownersdirectory,/owners/[owner]profile, and/repos/[id]/modules/[path]pages in the web UI. Risk page gains a Modules tab, blast-radius results page renders ReviewerSuggestions in a side rail, ownership-treemap shows bus-factor borders (red ≤1, amber 2, green ≥3), sidebar gets a Contributors entry. New@repowise-dev/uisubpackages:owners/,modules/(#188). - Hotspot drill-down to top symbols. Hotspot rows in the risk view now expand inline to show the importance-ranked top symbols in that file; clicking a symbol opens the existing SymbolDrawer. Backed by a new
file_pathfilter on/api/symbols(#191). - Writable decision ↔ module linkage. Decision detail page replaces the read-only affected-files block with a
ModuleLinkEditor— module-path autocomplete pulls from/modules/health.PATCH /decisions/{id}accepts optionalaffected_modules/affected_filesalongsidestatusso editor saves don't force a status change (#191). - Truthful pagination across risk + symbols + git surfaces. Hotspot, ownership, and symbol list endpoints now return a stable
{items, total, has_more, next_offset}envelope; new sharedResultsFooterrenders "showing N of M / load more" instead of silent client-side slicing. Hotspot cap raised from 100 to 500.HotspotResponsenewly surfacescommit_count_total,primary_owner_commit_pct,recent_owner_name/pct,merge_commit_count_90d,commit_count_capped,age_days,last_commit_at.git_indexerno longer capstop_authorsat 5 orsignificant_commitsat 10 — both lifted to 50, exposed via the per-file git-metadata endpoint (#187). - Importance-ranked symbols workspace. Server-side composite score combines file PageRank, visibility, complexity (log-normalised), kind, and entry-point status; transparent per-symbol component breakdown returned alongside each row. New filter facets —
visibility,in_hot_files,in_entry_points. Per-row signal chips (visibility, entry-point, hot-file, complexity) and a file-context panel in the SymbolDrawer (owner, bus factor, churn state, co-changes, overlapping dead-code findings, blast-radius shortcut) (#187). - Graph signal enrichment + architecture endpoint. New
/api/graph/{repo}/architecturereturns community super-nodes with per-cluster hotspot / dead / decision counts, doc coverage, and top languages. Full-graph export now capped to top-N by PageRank (configurable via?limit=, default 5000) withtruncated+total_node_countin the response. Every graph response (full, architecture, module, ego, dead, hot) carries the same hotspot / dead-code / decision / docs signals. Module nodes aggregatehotspot_count,dead_count,has_decision,primary_ownerfrom underlying files. Toolbar split into orthogonal Scope (Architecture / Modules / Full) × Overlays (Dead / Hot). New shared UI:NodeBadges,GraphContextDrawer,GraphTruncationBanner(#183). repowise deletecommand + DELETE endpoint + dashboard button. New CLI command lists repos in a numbered table, prompts for confirmation, then cleans FTS and CASCADE-deletes the repository and all child rows. NewDELETE /api/repos/{repo_id}endpoint. Trash icon appears on hover in the dashboard repo list with a confirmation dialog. Supports--forceand--path/-pon the CLI. Settings page redirects to/after delete instead of refreshing the now-404 route (#42).- Reasoning mode configuration for LLM providers (#175).
- Constructor + method parameter type-use edges for C#.
csharp.scmcaptures@param.typeon ctor / method / delegate / record-primary declarations; a newtype_ref_resolutionmodule dispatches per-language strategies. C# strategy resolves names throughDotNetProjectIndex.type_map, ranked by project enclosure. With these edges present, the universalinterfaceskip in dead-code analysis narrows — now only excluded for Java / Kotlin / Scala (#180). - XAML dynamic-hint extractor. Regex-parses
.xaml/.axamlacross WPF (clr-namespace:), WinUI / UWP / MAUI (using:), and Avalonia dialects. Resolvesx:DataTypeandDataContextbindings againstDotNetProjectIndex.type_mapso ViewModels reached only via{Binding}no longer read as orphans. Also emitsdynamic_usesedges for<ResourceDictionary Source="..."/>andMergedDictionariesentries (pack://, ms-appx:///, repo-rooted, and relative URIs) (#180, #186). - C# member-read resolution. New
languages/csharp_member_reads.pyresolvesvar x = new T()andthis.Proptoreadsedges on the defining file.nameof(Type)emitsdynamic_usesedges viaDotNetProjectIndex.type_map(#184). - ASP.NET host-builder extension-method resolution.
app.MapCatalogApi()/services.AddXxx()resolve to their defining file by scanning.csfiles forpublic static T MapX(this <HostType> ...)signatures against an allowlist of ASP.NET host types (IEndpointRouteBuilder,IServiceCollection, etc.). Host-builder extension scan now runs on any C# repo, not just ASP.NET Core (#182, #184). - CommunityToolkit MVVM synthetic symbols. New pluggable
extractors/synthetic_symbols.pyper-language registry. C# entry synthesises[ObservableProperty]fields → PascalCase property symbols and[RelayCommand]methods →<Name>Commandsymbols (#186). - C++ qualified method definitions (
void Foo::method() { … }) now extract withparent_name=Fooandkind=method. New cpp.scm pattern for two-levelNS::Foo::methoddeclarations plus a parser helper that walks thequalified_identifierscope to recover the immediate enclosing type (#190). - C++ visibility refinement. New
refine_cpp_visibilityreads access specifiers, file-scopestatic, and export attributes (__declspec(dllexport),__attribute__((visibility("default")))); the latter sets a newSymbol.is_exported_symbolflag the dead-code pass uses to skip language-level exports. C++ heritage now classifies I-prefixed bases asimplementsand concrete bases asextends(#186). - Phase timings persisted to
state.json. NewPhaseTimingRecorderProgressCallbackwrapper persists per-phase wall-clock durations tostate.json["phase_timings"]so before/after perf comparisons no longer require external instrumentation (#182). type_useedge provenance persisted as its own edge type (was previously a NetworkX-onlyviaattribute that the SQLite layer dropped) (#181)..xaml/.axamlingestion as a passthrough LanguageTag so the traverser produces file nodes that XamlDynamicHints can attach edges to. The extractor was previously emitting edges that GraphBuilder silently dropped because the source path was not a known graph node (#181).- Detected tech_stack persisted to
repositories.settings_json. Adds generic WPF / WinUI 3 / Windows Forms detection from canonical SDK indicators.tech_stack.pyreplaces the root+1-level.csprojglob with a bounded depth-first walk (≤5 levels, ≤200 projects,bin/objpruned) so monorepos withsrc/<area>/<module>/<Project>/<Project>.csprojlayouts register correctly (#180, #181).
Changed
- Parse progress ticks fire per worker.
asyncio.gatherheld everyon_item_doneevent until the last parse task completed, so the bar sat at 0/N for many minutes on large repos. Tick is now a done-callback on each task future — fires on the event loop thread as each worker returns (#183). - Co-change window widened from 500 → 2000 commits;
min_countdropped from 3 → 2, so low-churn repos surface pairs at all. Added funnel-stage debug log.co_changephase timer now closes the moment accumulation finishes via a newon_co_change_donecallback (#180, #184). - Churn percentile contract normalised at the HTTP boundary to 0–100 (was 0–1 in DB; UI consumers all assumed 0–100).
HotspotResponse,GitMetadataResponse, symbolfile_churn_percentile, and git-summary average all render correctly without per-component workarounds. Scatter defensively accepts the legacy 0–1 shape too (#187). - Churn × bus-factor scatter surfaces a danger-zone count badge and a clickable legend of the riskiest files; shows an explanation when churn is uniform across the repo instead of a degenerate vertical strip (#187).
- Hot Symbols Board collapsed by default on
/symbols— was a preview competing with the ranked table below (#187). - Single repo-wide
git log --numstatpass replaces O(files) per-file subprocesses; each worker reads commits from a shared in-memory dict. Per-language timing logged inside thegraph.importsloop. 200-file cap in_flush_commitguards against mass-edit OOM (#184). DotNetProjectIndex.build_indexcollapses three overlapping*.cswalks into one master rglob with cached file texts shared between the namespace-map and global-usings passes. Expected speedup: ~40 min → ~5–8 min on Windows for the import-resolution phase on monorepos. No data quality loss (#181).- Dead-code never-flag patterns picked up
Generated/output,*NativeMethods.csP/Invoke surfaces, ETW Telemetry/Events folders, merged-resource XAML dictionaries (Themes/Styles/Resources), standard test-project globs (*Tests/*.cs,*.UnitTests/*.cs,*FuzzTests/*.cs,*UITest*/*.cs,*Tests.cs), and*/unittests/*.cpp|.h(#180, #190). - Dead-code unused_export pass treats an incoming
calls/method_implements/readsedge on the symbol itself as evidence of life (was checking only file-levelimported_names, which missed intra-file C++ helpers and qualifiedFoo::methoddefinitions) (#190). dynamic_hints/dotnet.pylearnedtypeof(TypeName)— catches[JsonConverter(typeof(X))],[TypeConverter(...)],DataTemplate.DataType = typeof(X), and manual DI registration (#190).
Fixed
ResultsFooteroptional props widened toboolean | undefinedsoexactOptionalPropertyTypes: truein thepackages/uitsc pass accepts call sites that passloading={maybeUndef}. Unblockspublish-internal.ymlwhich had been failing on the last two main pushes (#189).- Graph community panel uncloseable.
CentralityLeaderboard(z-10) was rendered unconditionally on top of the community panel (also z-10) and physically covered the X. Auto-mounted leaderboard dropped entirely — the inspection panel already shows pagerank / betweenness / degree percentiles for the selected node in-context. Community panel bumped to z-20; newonCommunityPanelOpencallback fires when the legend triggers the community panel so doc + community never stack (#183). - Cap unused-export confidence to 0.4 for
kind=interfacesymbols with no incomingimplements/extendsedges. Implementor detection is heuristic across all statically-typed languages; absence is evidence-missing, not evidence-of-absence. Generic across C#, Java, Kotlin, Scala, Swift protocols, TS interfaces (#181). - Cap dead-code confidence to ≤0.4 for COM / IUnknown contract methods (
QueryInterface/AddRef/Release/IDispatch) in C++/C#/Rust — dispatched through native vtables and never observable via static call edges (#182). - Add Windows DLL entry points (
DllMain,DllGetClassObject,DllCanUnloadNow,DllRegisterServer,DllUnregisterServer,DllGetActivationFactory) to the never-flag list (#186). - Phase 1b progress bar.
graph.type_refswrapped withon_phase_start/doneso the CLI bar no longer appears frozen between import resolution and heritage resolution on large .NET repos (#181). - Module clicks across the app (Risk → Modules tab card, Heatmap treemap, Owner profile module rollup) now route to
/repos/[id]/modules/[path]instead of/ownership?module=…which the old page ignored (#188). SafeToDeletePilepreview groups findings byfile_pathwith a finding count, so files with multiple dead-code findings no longer appear multiple times in the top-5 strip (#187).- Top Contributors card removed from the Hotspots tab; it duplicated the Heatmap tab's contributor surface (#187).
- Symbol bloat warning. Parser logs a
parser.symbol_bloatwarning when a single file emits more than 500 symbols (#186).
Dependencies
urllib32.6.3 → 2.7.0 (#176)mermaid11.13.0 → 11.15.0 (#177)next15.5.15 → 15.5.18 (#178)
[0.8.0] — 2026-05-11
Added
- Workspace mode is now first-class across the CLI. Every relevant command auto-detects whether it's running inside a workspace root and routes accordingly, with a one-line
[workspace] …notice when it does. New flags--no-workspace(force single-repo) and--repo <alias>(scope to one repo) onupdate,status,watch,doctor,costs,search,dead-code,decision,generate-claude-md,hook install/status/uninstall.costsandsearchalso gained--allfor explicit workspace-wide fan-out. NewWorkspace auto-detectsection in CLI Reference (#173). repowise update --workspacenow first-time-indexes previously-skipped repos. Workspace entries without.repowise/no longer short-circuit with"not_indexed"— the full index pipeline runs (no LLM cost),state.jsonis written with adocs_skip_reasonmarker, and subsequentupdate --repo <alias> --docscleanly picks up doc generation (#173).repowise workspace adddefaults to full index + LLM doc generation when a provider is configured. Inherits provider, model, embedder, and exclude patterns from the primary repo's.repowise/config.yaml.--no-docs/--no-indexopt out. Cost-gate prompt still runs before any tokens are spent (#173).repowise doctor --workspacevalidates every workspace entry: directory exists, has.git/, state.json ↔ workspace config drift, MCP registration.--repairsyncs drifted entries from disk and drops dead entries whose directory no longer exists (#173).- Honest completion summaries.
initandstatusnow print a per-repo Docs status block listing whether docs were generated, the skip reason (cost gate declined,provider failure,index-only, …), and the exact remediation command. No more empty docs pages in the UI without context (#173). - Workspace-aware web UI. Sidebar now shows every workspace repo including unindexed ones (rendered as disabled
needs index/missingrows linking to the Workspace dashboard). Workspace dashboard has a top-level Sync workspace button plus per-repo Sync / Index now actions wired to the newPOST /api/workspace/syncendpoint.RepoCardsurfacesdocs_skip_reasonunder each card's stats (#173). - Per-repo search scope toggle on
/repos/<id>/search— switch between this repo and workspace. Syntheticws:<alias>IDs automatically fall back to workspace scope (#173). /api/workspace/syncendpoint fans out the existing job executor across every workspace repo (or a single one withrepo_alias). Returns one{alias, repo_id, status, reason}per repo so the UI can render granular feedback (#173)./api/searchaccepts arepo_idquery param in workspace mode. Returns[]for syntheticws:<alias>IDs (the corresponding repo isn't indexed) and fans out across every loaded FTS / vector store when omitted (#173)./api/reposreturns workspace metadata per row —workspace_alias,workspace_status(indexed|needs_index|missing_dir),is_primary,docs_enabled,docs_skip_reason. Unindexed entries appear as synthetic rows withid="ws:<alias>"so frontends can render a "Needs index" CTA card instead of silently dropping them (#173).- Shiki syntax highlighting in wiki page code blocks — client-side, lazy-loaded with the Vesper theme, falls back to plain text on failure (#171).
- Centrality Leaderboard right-rail on the graph view (PageRank / Betweenness / Degree) and Hot Symbols Board with score-driven intensity bars on the symbols table (#171).
Changed
- Dependency heatmap rewritten from canvas to CSS Grid — adds hover tooltips, row/column highlighting, a legend, an
external:-prefix stripper fordisplayLabel, and caps the rendered grid at the 15 most-connected modules (#171). - Docs filter panel defaults to expanded on first render of the Docs page (#171).
- MCP server is workspace-aware end-to-end.
get_overview(repo="all")returns a workspace summary with cross-repo topology;search_codebase(repo="all")runs Reciprocal Rank Fusion across every repo; tools that can't meaningfully fan out return_unsupported_repo_all()with the available aliases. (Pre-existing scaffolding; this release adds tests + audit confirmation.) (#173)
Fixed
repowise updatefrom a workspace root no longer errors with "No previous sync found". Auto-detection routes the command to workspace mode and prints[workspace] running across N repos; the helper performs all detection beforeensure_repowise_diris called, so stray.repowise/directories no longer get created at the workspace root. Original Discord report that motivated the overhaul (#173).repowise servein workspace mode no longer drops unindexed repos from the sidebar. Server lifespan now buildsapp.state.workspace_fts: dict[repo_id, FullTextSearch](per-repo, includes the primary) and lazily rehydrates each workspace LanceDB store viaresolve_workspace_vector_store()with anasyncio.Lockper repo so concurrent searches don't double-open. Reuses the primary store's embedder so workspaces built with gemini/openai stay embedding-compatible across fan-out (#173)..repowise-workspace.yamlno longer drifts when a child repo is updated outside the orchestrator. Newsync_workspace_state_from_disk()reads each repo'sstate.jsonat the start of everyupdate_workspaceand refresheslast_commit_at_indexso workspace-level decisions never operate on stale info (#173).
Documentation
- CLI Reference rewritten for workspace mode — new cross-cutting auto-detect section, per-command flag tables updated for
update,watch,search,status,dead-code,costs,workspace add,doctor(#173).
Dependencies
shiki^4.0.0 added as apackages/uidependency for client-side wiki code highlighting (#171).
[0.7.1] — 2026-05-10
Fixed
repowise serve404 on the web tarball for v0.7.0. The v0.7.0 publish workflow failed during the web build:useSearchParams()inside the newContextDrawerShell(mounted in the root layout for the?drawer=URL sync added in #168) tripped Next.js' static prerender of/settingswith a missing-Suspense bailout. The Python wheel published to PyPI but norepowise-web.tar.gzwas attached to the v0.7.0 GitHub release, so end-userrepowise servefalls through to "API only". v0.7.1 wrapsContextDrawerShellin a<Suspense>boundary inpackages/web/src/app/layout.tsxso the layout no longer blocks static prerendering. Anyone who installed 0.7.0 should upgrade to 0.7.1 to get a working web UI fromrepowise serve.
[0.7.0] — 2026-05-10
Added
- Risk page (consolidated). New
/repos/<id>/riskroute brings the Heatmap, Hotspots, Dead Code, and Impact views under a single page with a persistent summary strip across the top. Hotspot rows are now clickable and open the universal File Card (#168). - Security page. New
/repos/<id>/securityroute renders severity distribution, findings-by-directory, and a clickable findings table over the existing security signals (#168). - Costs reorganization.
/repos/<id>/costsnow splits into five tabs — Daily, Cache, Hotspots, Providers, Operations — backed by newcache-hit-ratio-card,cost-heatmap,operation-breakdown, andprovider-comparisoncomponents (#168). - Universal File Card. New
FileCard+FileCardDialog(@repowise-dev/ui/shared/file-card) shows a unified overview of any file — git signals, docs, symbols, dead-code findings, decisions, security issues — with sections that render only when the underlying data exists. Wired into Risk and Symbols (#168). - Hot Symbols Board with score-driven intensity bars over the symbols table; Centrality Leaderboard as a collapsible right-rail panel on the graph view (PageRank / Betweenness / Degree) (#168).
- Docs onboarding.
FirstFiveFiles"Start here" card on the Docs page (collapsed by default) links to/docs?page=<id>. NewDriftBannerandConfidenceVsFreshnessMatrixon docs/coverage. Mermaid diagrams now have a maximize button with a zoom/pan modal and a neutral brand theme (#168). - Cross-page surfacing. Shared
RelatedAcrossRepowisecollapsible footer plus newEntityLink/EntityHoverCardprimitives and aContextDrawerscaffold with URL sync (#168). packages/uiexports. New entry points:./costs,./security,./onboarding,./shared/file-card,./shared/related(#168).
Fixed
- SymbolDrawer right-rail text cutoff — drawer widened and
ScrollAreapadding adjusted (#168). - Risk Hotspots table overflow — long rows now truncate cleanly and open the File Card on click instead of pushing the layout (#168).
- Health-score ring skipped doc components for index-only repos so the score reflects what was actually computed (#168).
- Chat
ToolCallBlockhydration error — split a button nested inside another button into sibling elements (#168). - Docs explorer sidebar toggle anchored to its
relativeparent instead of falling back to the viewport, so it no longer overlaps the header (#168). - Docs "Start here" links now route to
/docs?page=<id>instead of broken wiki slugs (#168).
[0.6.2] — 2026-05-10
Fixed
- Dead-code analyzer flagged DI-injected and convention-loaded code as unused. On real .NET solutions (e.g. eShop) the analyzer surfaced ~1,350 false positives — gRPC services, EF
DbContexts, MAUI entry points, mock services bound throughAddSingleton<TService, TImpl>(), and most public interfaces. Three classes of fix landed: (a)_NON_IMPORTABLE_SYMBOL_KINDSnow skipsmethod,variable,field,property,enum_member,constant,type_alias,namespace,module, andinterfacefrom the unused-export pass — these aren't importable by name in any language, so absence of animportsedge isn't evidence of unreachability. (b) The.NETdynamic-hint extractor (packages/core/src/repowise/core/ingestion/dynamic_hints/dotnet.py) now matches the full DI surface:Add|Map|Use×Scoped|Singleton|Transient|HostedService|DbContext(Pool|Factory)?|HttpClient|Options|GrpcService|GrpcClient|Hub|SignalR|Controllers?|Middleware, plusConfigure<T>, integration-event subscriptions, and class-name collision via atype → list[file]map so two classes namedBasketServicein different microservices both receive the synthetic edge. (c)_detect_zombie_packagesnow skips dot-dirs and code-less directories. eShop dead-code findings dropped 2,483 → 459 (−81%); safe-to-delete 1,354 → 339 (−75%) (#164, #166). - Symbol-level PageRank and betweenness were always 0. Centrality only ran on the file subgraph, so the symbol detail panel showed 0 for every symbol regardless of how heavily it was called or referenced.
GraphBuildernow exposessymbol_subgraph()(calls + heritage edges between symbol nodes) plussymbol_pagerank()andsymbol_betweenness_centrality()with caches;compute_metrics_parallel()includes them;persist_graph_nodes()writes them tograph_nodes.pagerank/betweennessfor symbol rows. On the local repo: 0/3,747 → 3,753/3,753 symbols with non-zero centrality (#164). - CLAUDE.md tech-stack inferred Node.js for any repo with a
package.json. Apackage.jsoncontaining only dev dependencies (Prettier, ESLint, Husky) was enough to brand a Python or .NET repo as "Node.js". Detection is now gated on real runtime evidence (runtime_deps,main/bin/module/exports/engines.node, or a framework dep). Added .NET / ASP.NET Core / EF Core / Aspire / gRPC / MAUI detection from.csproj/.sln/Directory.Build.props(#164). repowise update --index-onlycrashed withNameError: cannot access free variable 'dead_code_report'. Pre-existing bug:dead_code_reportwas defined inside the docs-generation branch but referenced after it. Moved dead-code analysis above theif index_only:early return; both index-only and full update paths now re-persistgraph_nodesso symbol metrics stay current on incremental refresh (#164).persist_pipeline_resultraisedNameError: name 'nodes' is not definedin CI integration tests after the persistence refactor extractedpersist_graph_nodes. The finallogger.infosummary still referencedlen(nodes)from the removed loop. Now reads node count from the graph builder (#166).- C# entry-point detection missed MAUI / WPF / WinUI starts.
MauiProgram.cs,Main.cs, andApp.xaml.csare now recognised entry points alongsideProgram.cs(#164). - Embedding latency serialised LLM throughput in
PageGenerator.generate_all(). The page-generation semaphore was held whileembed_and_upsert()ran, so a slow vector-store endpoint reduced effective generation concurrency to whatever embedding could keep up with. The LLM semaphore is now released as soon as a page is generated and embedding runs behind a separateembed_concurrencysemaphore (defaults tomax_concurrency). NewGenerationConfig.embed_concurrencyfield (#163).
Added
AUDIT_NOTES.mdat the repo root tracks deferred proper fixes from the May 2026 .NET audit (constructor-parameter type-use edges, XAML/Razor binding-path resolution, minimal-API extension-method resolution, member-access "uses" edges, co-change pair extraction returning 0 on real repos, hotspot ranking usingtemporal_hotspot_score, symbol metrics inget_context, language-aware never-flag patterns, graph-driven tech-stack inference, narrowing thekind=interfaceskip once ctor-param edges land). Each item has root cause, proper fix, touch points, and an estimate so future sessions can pick them up cold (#166).
Changed
repowise serverebuilds the local web bundle when source is newer. Previouslyservewould launch the cached UI even aftergit pullhad updatedpackages/web/. Now compares mtimes and rebuilds when stale; new--refresh-uiflag forces a rebuild. Affects local-monorepo dev only — end-user installs continue to downloadrepowise-web.tar.gzmatched to the wheel version (#165).- Smarter Claude Code augment hook. PostToolUse enrichment now runs against
Bash/Edit/Writeonly and skips noisyGrep/GlobPreToolUse, with self-healing migration for legacy hook entries on upgrade (#162).
[0.6.1] — 2026-05-10
Added
- DeepSeek provider —
deepseek-v4-flash(default) anddeepseek-v4-proare now first-class LLM providers via DeepSeek's OpenAI-compatible API atapi.deepseek.com. Implementation mirrors the OpenRouter pattern (openai SDK + custombase_url), withgenerate()andstream_chat()(incl. tool calling), 3-attempt exponential-backoff retries on rate limits, dedicated rate-limit defaults (60 RPM / 200K TPM), per-model pricing in the cost tracker, and full plumbing through CLI provider resolution, MCPget_answerauto-detection, the run-config form, and the settings-page provider list. New env vars:DEEPSEEK_API_KEY(required),DEEPSEEK_BASE_URL(optional override) (#159).
Fixed
- Claude Code hook crashes when the active venv is broken. PreToolUse (
Grep/Glob) and PostToolUse (Bash) hooks invoked the fullrepowise augmentClick command, whose import chain pullscli.main→init_cmd→cost_estimator→core.ingestion.graph→networkx/scipy. A single missing dependency in the user's environment caused every tool call to surface anImportErrortraceback and non-zero exit, because the in-handlertry/exceptcould not catch failures during module loading. Hooks are now wired to a newrepowise-augmentconsole script (repowise.cli.augment_hook:main) that imports only the augment handler — module-level imports are stdlib-only — and wraps the entire run, including the lazy import of the handler, in a last-ditchexcept BaseExceptionso any failure exits 0 silently. Existing users upgrading from any prior version are migrated automatically: everyrepowise <command>invocation, plus the hook itself on first firing, idempotently rewrites legacyrepowise augmententries in~/.claude/settings.jsontorepowise-augment—pip install -U repowiseis the only step needed (#160). repowise mcpcouldn't reach the LLM thatinithad configured. The MCP server didn't load.repowise/.envat startup, soget_answerfell back to retrieval-only withconfidence=loweven wheninithad completed cleanly with a real provider. The resolver now readsstate.json(provider + model) and.env(API keys) as a fallback layer behind process env, and themcpcommand itself loads.repowise/.envon startup. Samerepowise initconfiguration is now reused end-to-end without re-exporting anything (#158, #159).get_overviewcrash on legacy databases. The repo-overview query usedscalar_one_or_none()while older indexes left a staletarget_path="repo"row alongside the canonicaltarget_path=<repo_name>row, raisingMultipleResultsFoundon the documented "best first call". Switched to a deterministic ordered.first(): prefer the row matching the repository name, fall back to most recently updated. Same fix in the workspace overview path (#158).get_whyrouting natural-language questions tomode=path. The_is_pathheuristic returned True for any query containing/, so questions like "why does init use a two-phase plan/apply flow" dispatched to the path branch and returned empty results. Heuristic now recognises NL up front (trailing?, leading question word, or 4+ tokens including a question word route to search; whitespace anywhere disqualifies a path); genuine paths likesrc/auth/service.pystill route tomode=path(#158).- README marker examples leaking into decision records. The inline-marker scanner walked the whole tree, including
repowise.egg-info/PKG-INFO, where setuptools embedsREADME.mdverbatim — so example# WHY:/# DECISION:/# TRADEOFF:lines from the README surfaced as real architectural decisions inget_why's health dashboard. Walker now excludes*.egg-infoand*.dist-info(#158). .envparser handled only the simplest format.load_dotenvnow correctly handlesexport KEY=value, single- and double-quoted values, and inline# comments, fixing a common 401 cause where quoted API keys were treated literally and whereexport-prefixed entries were silently ignored (#159).- Provider import /
get_providerfailures logged at debug level. A user-visible failure (no provider available forget_answerto synthesize an answer) was hidden in debug logs; now logged atwarningso the cause is discoverable without debug logging enabled (#159).
Changed
geminire-added to the run-config form provider list — was inadvertently dropped in the v0.6.0 frontend reshuffle (#159).litellmAPI-key resolution in CLI provider plumbing —LITELLM_API_KEYis now picked up alongside the existingLITELLM_BASE_URL/LITELLM_API_BASE(#159).
[0.6.0] — 2026-05-09
Added
- Sigma.js graph renderer replaces React Flow as the primary graph view. ForceAtlas2 web-worker layout for the
forcemode and ELK-driven hierarchical layout share a single canvas. Inspection panel, search, community dimming, execution-flow highlighting, and legend counts all reach parity via Graphology adapters; double-click drills into modules and rebuilds the graph with child file nodes inline. Signal overlays (dead-code desaturation, hotspot tint, entry-point size boost) live in the SigmanodeReducer(#148). - Per-phase progress for graph build and metrics —
GraphBuilder.build()now reports imports/heritage/calls as sub-phases through the existingProgressCallback, and the orchestrator drives metrics, communities, and flows as their own sub-phases by priming the lazy caches.repowise initnow shows six indented bars under the graph phase instead of a single opaque spinner that previously sat at "0/1" for 5–10 minutes (#150). - Dashboard
EmptyStateguards — every dashboard panel now renders a labelled empty state instead of going blank when its data slice is missing (#148).
Changed
repowise updatedefaults to the modeinitwas run with.repowise initnow persistsdocs_enabledto.repowise/state.json(true for full init, false for--index-only), andrepowise updatereads that field so the post-commit hook does the right thing without extra knobs. New--docs/--no-docsflags override per run;--index-onlystill wins. Index-only init now also writesstate.json, so the post-commit hook has a baseline to diff against (#155).- Cost-gate persistence. Declining the cost gate now produces a clean index-only outcome instead of an aborted half-state — ingestion, graph, git, and dead-code work is persisted,
state.jsonlands withdocs_enabled=False, and subsequentrepowise updateruns default to index-only so there are no surprise LLM charges later (#156). - Cost-gate prompt is now visually separated from the Rich progress output above it (blank line + horizontal rule before the
[y/N]), preventing the prompt from being missed mid-output (#156). - Stale-wiki warning is much quieter. The
repowise augmentPostToolUse hook used to fire on every Bash tool call after a commit until an update completed; now it suppresses whilerepowise updateholds.repowise/.update.lock, and after warning once for a given HEAD it skips further warnings until HEAD moves. The hook installer also detects and excises legacy non-marker bodies before appending the marker block (#155).
Fixed
- Python relative imports drop their first imported name.
from .X import Yandfrom .X import A, B, Cwere being parsed withY/Adiscarded because the binding extractor's "skip the first dotted_name" heuristic, correct for absolutefrom foo.bar import X, also fired on tree-sitter'srelative_importwrapper. The graph storedimported_names_json: []for affected edges, which propagated into massive dead-code false positives on Repowise's own source (e.g.GraphBuilder,DeadCodeAnalyzer,CallResolverflagged at confidence 1.0). The extractor now detectsrelative_importand skips the heuristic, with regression coverage for both relative and absolute shapes (#149). - Dead-code unused-export false positives. Symbol decorators are now persisted on graph nodes (the framework-decorator whitelist was previously running against an empty list),
@-prefixed decorator names are matched against the bare prefixes in_FRAMEWORK_DECORATORS(so@router.get,@asynccontextmanager, etc. are recognised), and nested function definitions are skipped from unused-export detection — closures and inner generators can't be imported by name and were being flagged spuriously (#153). - State migration for legacy index-only installs.
state.jsonfiles written before #155 lackdocs_enabled. The previous default would have charged a full LLM regen on the first upgrade-and-commit for users who had originally runinit --index-only. The resolver now infersdocs_enabled=Falsewhenprovider/modelare also absent (the legacy shape of an index-only state file), backfills the explicit value intostate.jsonon first update, and preserves the existing default for full-init users (#156). - Workspace-mode chat 404.
POST /api/repos/{repo_id}/chat/messageswas the only/api/repos/{repo_id}/...endpoint not honouringapp.state.workspace_sessions, so every non-primary repo's chat returned404 Repository <id> not founddespite appearing inGET /api/repos. Factory-resolution logic is now lifted intoresolve_session_factory/resolve_request_session_factoryhelpers indeps.py, the chat router uses the request-scoped helper, and the duplicate helper inrouters/repos.pyis now a one-line alias (#146). - Init progress rendering cleanup — phase labels and indentation alignment fixes (#151).
Performance
- Graph metrics fan out in parallel. PageRank, betweenness, file/symbol community detection, and execution-flow tracing previously ran serially across persist + generation, with PageRank and betweenness recomputing from scratch on each call.
GraphBuildernow caches all four kernels on the instance (invalidated onbuild()), and a newcompute_metrics_parallel()runs them viaasyncio.gather+asyncio.to_threadso subsequent lazy callers hit warm caches. Betweenness dominates worst-case wall-time (O(VE)); fanning it out alongside PageRank and community detection meaningfully shortens the metrics phase. Falls back to lazy computation ifcompute_metrics_parallel()is never called (#152). - Tree-sitter query cache promoted to module-level
@lru_cachekeyed by language tag. Process-pool parse workers each held their own per-instance cache and recompiled every.scmquery on first use; now each worker compiles each grammar's query exactly once for its lifetime (#154). - Per-file
Compiled query language=...debug log dropped. It fired once per parser-instance × language during ingestion and was the single noisiest source of unfiltered stdout duringrepowise init(#149).
Dependencies
gitpython3.1.47 → 3.1.50 — security release: rejects out-of-repo reference manipulation (3.1.48) and rejects control characters in config writes (3.1.49) (#147).
[0.5.1] — 2026-05-07
Added
- TYPO3 framework edges — composer-based extension discovery (
"type": "typo3-cms-extension", canonical for v11–v14) with legacyext_emconf.phpfallback and project-modevendor/<vendor>/<package>/walking. Convention-loaded files (ext_localconf.php,ext_tables.php/.sql,Configuration/TCA/*.php,Configuration/TCA/Overrides/*.php,Configuration/Backend/*.php,Configuration/Services.{php,yaml,yml},JavaScriptModules.php,ContentSecurityPolicies.php,RequestMiddlewares.php,Icons.php,RTE/*.{yaml,yml}) now receive incoming edges from a syntheticframework:typo3-coreanchor and are no longer flagged as unreachable.Configuration/JavaScriptModules.phpis parsed forEXT:<key>/...jsreferences and edges are added to the registered JS modules.tech_stack.detect_tech_stackrecognisestypo3/cms-core,symfony/framework-bundle, andlaravel/frameworkfromcomposer.json(#114). framework:synthetic-node prefix in dead-code analysis — distinguishes framework-mediated wiring from third-partyexternal:imports.framework:predecessors count as cross-package importers (preventing legitimate convention dirs likeConfiguration/from showing as zombie packages);external:predecessors do not (#114).
Fixed
repowise dead-codenow invokesadd_framework_edges— the CLI previously skipped framework-aware edge synthesis, so even Django/Laravel/Rails repos showed convention files as false-positive unreachable findings. The dead-code command now callsdetect_tech_stackand adds framework edges before running the analyzer (#114).
Dependencies
cryptography43.0.3 → 46.0.7 (#130).lodash4.17.23 → 4.18.1 (#131).lodash-esandlangiumtransitive bumps (#129).esbuild,vitest, andvitedev tooling bumps (#134).
[0.5.0] — 2026-05-03
Changed
- Build packaging hardened —
pyproject.tomlnow uses[tool.setuptools.packages.find]to auto-discover allrepowise.*subpackages acrosspackages/{core,cli,server}/src, replacing the hand-maintained explicit list. Eliminates the missing-subpackage drift class that previously required hotfixes (#97, #110, #115). - Frontend monorepo restructure — visualization, dashboard, chat, wiki, graph, and workspace components extracted from
packages/webinto shared@repowise-dev/uiand@repowise-dev/typesworkspace packages (~50 components). Fully transparent topip install repowiseusers — the publishedrepowise-web.tar.gzstandalone bundle is unchanged in shape and behaviour. OSS contributors benefit from clearer module boundaries; both packages resolve via npm workspace symlinks with no extra auth required. packages/webdeclares its workspace dependencies explicitly —@repowise-dev/uiand@repowise-dev/typesare now listed inpackages/web/package.jsonso isolated installs (cd packages/web && npm install) no longer fail with module-not-found.
Fixed
- Jobs reliability pass — cancel endpoint added; progress hydration covers all phases; stuck-job detection on startup resets stale
pending/runningrows; SQLite WAL contention reduced during sync; per-repo DB used in workspace mode (#117). repowise updatenow persists LLM costs — costs were being computed but not written to thellm_coststable during incremental updates; cost dashboards underreported spend (#108).- Workspace dashboard — contract summary now renders when contracts exist but no cross-repo links have been detected, instead of showing an empty state (#111).
Documentation
- Computed glossary —
docs/reference/COMPUTED_GLOSSARY.mddocuments every derived metric, score, and signal Repowise computes (PageRank, hotspot score, freshness, confidence tiers, etc.) so the surface vocabulary is discoverable in one place (#127). - README + UI/UX audit fixes — confirmation dialogs, mobile responsiveness, accessibility, and empty/error/loading states across the dashboard (#117).
Dependencies
python-dotenv1.0.1 → 1.2.2 (#98).
[0.4.1] — 2026-04-30
Fixed
- Wheel packaging —
pyproject.toml[tool.setuptools] packageslist extended to include subpackages omitted in 0.4.0; some installs were missing modules at runtime (#110). get_answerMCP tool — citation format and confidence gating fixes (#107).
[0.4.0] — 2026-04-26
Added
C# Full tier
- MSBuild-aware import resolver — new
resolvers/dotnet/subpackage parses every.csprojand.slnin the repo, builds a namespace → file map across projects, walksDirectory.Build.propsancestry, and resolvesusingdirectives by ranking candidates: same project → directly-referenced project → anywhere. NuGet<PackageReference>ids are emitted asexternal:nuget:<id>nodes. Falls back to legacy stem-match for repos without.csproj. - Modern C# language features —
csharp.scmnow capturesrecord_declaration,delegate_declaration,event_declaration/event_field_declaration,field_declaration,enum_member_declaration, and both block-form and file-scopednamespace_declaration.LANGUAGE_CONFIGSand the registry'sheritage_node_typesare extended accordingly. global using/using static/using aliaspropagation —NamedBindinggainsis_globalandis_static_importflags;extract_csharp_bindingsdistinguishes all four flavours ofusingdirective. Default<ImplicitUsings>set (with Web SDK extras) andglobal usinglines are merged into a per-project implicit-usings set used by the resolver.- XML doc parsing — module-level and symbol-level
///runs are extracted,<summary>content is unwrapped as the rendered docstring, structural tags (<param>,<returns>,<see/>) are stripped, and<inheritdoc/>emits a{inheritdoc}marker. - Heritage for records —
record User(...) : Base(args), IInterfacenow produces bothextendsandimplementsedges; primary-constructor argument lists are skipped. - ASP.NET / .NET framework edges —
_add_aspnet_edges()runs whenever the tech stack mentions ASP.NET or any.csfile importsMicrosoft.AspNetCore.*. Adds edges fromProgram.cs/Startup.csto every[ApiController]file,app.MapGet/...handler classes,app.UseMiddleware<T>()middleware, and from eachDbContextto entity files referenced viaDbSet<T>. - .NET dynamic hints — new
DotNetDynamicHintsextractor (registered inHintRegistry) records DI registrations (AddScoped/AddSingleton/AddTransient/AddHostedService), reflection (Activator.CreateInstance,Type.GetType,Assembly.Load*),[assembly: InternalsVisibleTo], and MEF[Export]/[ImportMany]as graph edges. - Workspace contract extraction for ASP.NET and gRPC-dotnet —
http_extractor.pylearns[HttpGet/Post/...]attribute routing with class-level[Route]prefix stitching, parameterless[HttpVerb]attributes, minimal API (app.MapGet/...), and HttpClient consumers (*Async).grpc_extractor.pyrecognisesapp.MapGrpcService<T>(),class X : Service.ServiceBase, andnew ServiceClient(channel). - Cross-repo
<ProjectReference>and internal NuGet —cross_repo._scan_csprojwalks every.csprojin every workspace repo and emitsdotnet_project_reffor cross-repo project references anddotnet_nuget_internalwhen a<PackageReference>id matches a sibling repo's<AssemblyName>. - Dead-code dynamic markers for C# —
_DYNAMIC_IMPORT_MARKERSlearns reflection / DI / MEF /InternalsVisibleTopatterns so the dead-code analyser doesn't flag types only loaded by the framework at runtime. - Multi-project test fixtures —
tests/fixtures/dotnet_solution/(Api / Domain / Infrastructure with EF Core, controllers, minimal API, GlobalUsings) andtests/fixtures/dotnet_workspace/(3 repos demonstrating cross-repo<ProjectReference>+ internal-NuGet patterns), with end-to-end coverage intests/integration/test_dotnet_solution.py.
Dead-code accuracy
- Dynamic-edge consumption in dead-code analysis — graph edges of type
dynamic/dynamic_*(emitted by every dynamic-hint extractor) now suppress dead-code findings automatically.find_dynamic_edge_files()enumerates files involved in those edges and unions the result with the existing source-text_DYNAMIC_IMPORT_MARKERSscan. Sub-types (dynamic_uses,dynamic_imports) are preserved on the graph edge instead of being squashed. - Per-language dynamic-import markers —
_DYNAMIC_IMPORT_MARKERSextends to Go (reflect.TypeOf/reflect.ValueOf), Ruby (Object.send,Kernel.const_get,.public_send), PHP (call_user_func*,new $class,ReflectionClass), Kotlin (KClass.createInstance,::class.java), Swift (NSClassFromString,Selector,#selector,NSStringFromClass), and Scala (Class.forName,runtimeMirror,reflect.runtime). detect_unused_internalsenabled by default — private-symbol findings now surface in the standard dead-code report at confidence 0.65 withsafe_to_delete=False. CLI defaults stay explicit-False sorepowise dead-codeis unchanged unless--include-internalsis passed.
Workspace-aware resolvers across the Good tier
- PHP composer PSR-4 —
resolvers/php_composer.pyreadsautoload.psr-4andautoload-dev.psr-4fromcomposer.json, builds a longest-prefix-wins namespace → directory map, and is consulted before stem fallback. Real Laravel/Symfony apps with"App\\": "src/"style maps now resolve. - Go multi-module monorepos —
resolve_go_importwalks everygo.modin the repo (skippingvendor/node_modules), records(module_dir, module_path)tuples on the resolver context, and matches imports by longest module prefix. Single-module back-compat preserved. - TypeScript SFC + workspace package resolution —
.vue,.svelte, and.astroextensions probed only when the repo actually contains SFC files. npm/yarn/pnpmworkspaces(array or object form, with glob expansion) are parsed from rootpackage.jsonso@scope/pkgand@scope/pkg/sub/pathresolve to the sibling workspace dir before falling back toexternal:. - Kotlin Gradle subprojects —
resolvers/kotlin_gradle.pyparsessettings.gradle(.kts)include(...)declarations plus per-modulesrcDirs(...)overrides (defaultssrc/main/kotlin,src/main/java), then walks each source root recordingpackagedeclarations into apackage_to_filesmap. - Ruby Rails / Zeitwerk autoloading — gated on
config/application.rb,resolvers/ruby_rails.pybuilds bare-name and namespaced-name maps over standard autoload roots (app/*,lib/).ResolverContext.rails_lookupexposes the index for callers (heritage, call resolution, framework edges). - Swift SPM target → directory mapping —
resolvers/swift_spm.pyregex-parses.target(name: "X", path: "Y"),.executableTarget, and.testTargetdeclarations across allPackage.swiftfiles in the repo (defaultsSources/<Name>for code,Tests/<Name>for tests). - Scala SBT / Mill multi-project —
resolvers/scala_build.pyautodetects the build tool (build.sbtvsbuild.sc) and parses subprojects (SBTlazy val core = project.in(file("core")), Millobject Foo extends ScalaModule). Walks each project'ssrc/main/scala(orsrc/) recording packages intopackage_to_files. - Cargo workspace crate resolution —
resolvers/rust_workspace.pyparses rootCargo.toml[workspace] members = [...]plus each member's[package] name.resolve_rust_importconsults the index after the same-crate probe souse sibling_crate::moduleresolves to the sibling crate'ssrc/. Cargo's-→_import-identifier rewrite is honoured.
Framework-aware edges (every major web framework)
- Spring Boot (Java/Kotlin) —
@Component/@Service/@Repository/@Controller/@RestController/@Configurationbean classes wire to their injection sites via@Autowiredfield/constructor analysis. Interface-typed dependencies fall back toparsed.heritageto find implementing classes.@Beanfactory methods in@Configurationclasses link to their return-type files. - Rails (Ruby) —
config/routes.rbis line-walked with namespace-stack tracking:resources :users,get "/foo", to: "users#index", and nestednamespace :admin do … endall resolve to controller files via the Zeitwerk autoload index. ActiveRecordbelongs_to/has_many/has_onerelationships link model files (with simple inflector-style singularisation). - Laravel (PHP) —
routes/web.phpandroutes/api.phpparse modern[Foo::class, 'method']and legacy'Foo@method'syntaxes, plusRoute::resource. Service-providerbind/singleton/instancecalls link providers to bound classes. EloquenthasMany/belongsTo/hasOnelink models. Class resolution uses the composer PSR-4 map first, falling back to stem. - TYPO3 (PHP) — extension discovery via
composer.json"type": "typo3-cms-extension"(canonical for v11–v14) with legacy fallback toext_emconf.php; project-mode (vendor/<vendor>/<pkg>/composer.json) is also walked. Convention-loaded files (ext_localconf.php,ext_emconf.php,ext_tables.sql,Configuration/TCA/*.php,Configuration/Backend/*.php,Configuration/JavaScriptModules.php,Configuration/ContentSecurityPolicies.php,Configuration/RequestMiddlewares.php,Configuration/Services.php,Configuration/Icons.php) get incoming edges from a syntheticframework:typo3-coreanchor, so they are no longer flagged as unreachable.Configuration/JavaScriptModules.phpis parsed forEXT:<key>/...jsreferences and edges are added to the registered JS modules.tech_stack.detect_tech_stackrecognisestypo3/cms-coreandsymfony/framework-bundle/laravel/frameworkfromcomposer.json. framework:synthetic-node prefix in dead-code analysis — distinguishes framework-mediated wiring from third-partyexternal:imports.framework:predecessors do count as cross-package importers (preventing legitimate convention dirs likeConfiguration/from showing up as zombie packages);external:predecessors do not.repowise dead-codenow invokesadd_framework_edges— the CLI previously skipped framework-aware edge synthesis, so even Django/Laravel/Rails repos showed false positives. The dead-code command now callsdetect_tech_stackand adds framework edges before running the analyzer.- Express / NestJS (TS/JS) — Express
app.use(routerVar)mirrors the FastAPI router-var pattern (resolves imported names ending inRouter/routerto source file). NestJS@Module({ controllers: [...], providers: [...], imports: [...] })arrays parse into module → target edges using a class-name → file map. - Gin / Echo / Chi (Go) —
r.GET("/p", users.Index)style handler references resolve via the Go import list (using the multi-module resolver) for package-qualified handlers, or via a function-name → file map for receiver methods. Lambda handlers are accepted as missed. - Axum / Actix (Rust) — Axum
Router::new().route("/p", get(handler)), Actixweb::resource("/p").route(web::get().to(handler))/.service(handler)/.configure(routes::register)all resolve to handler files via a function-name → file map.
Per-language dynamic-hint extractors
- Spring (JVM) —
applicationContext.getBean(Foo.class)and named-bean lookups, plus@Beanfactory return-types. - Ruby —
Object.send(:method)/.public_send,Kernel.const_get,define_method, ActiveSupportdelegate :foo, to: :bar. - PHP —
call_user_func/call_user_func_array,new ReflectionClass(Foo::class), containerget/app/resolve/makewith::classarguments,new $variableinstantiation markers. - Scala —
Class.forName(...),runtimeMirror/reflect.runtimemarkers, namedgiven foo: Bar = ???andimplicit val foo: Bar = ???declarations. - Swift —
NSClassFromString("Foo"),NSStringFromClass(Foo),Selector("name"),#selector(name), KVCvalue(forKey: "key"). - C — function-pointer assignment (
fp = some_function;where the right-hand side is a known function name),dlopen("./libfoo.so"),dlsym(handle, "name"). - Luau —
game:GetService("Name"),setmetatable(t, {__index = Other}),require(game.Service.Path)markers. - Go —
reflect.TypeOf(Foo{}),plugin.Open(...),plugin.Lookup(...).
Symbol-extraction coverage
- Java records —
record Point(double x, double y) {}now captured as a class-kind symbol with optional modifiers. - Kotlin —
typealias Foo = Barand top-level / class-levelval/varproperties (locals inside function bodies remain excluded). - Scala 3 —
enum_definition,given_definition(named givens), andvar_definitionare now captured.class_definitionandfunction_definitionalso capture leading annotations (@deprecated,@tailrec). - Swift —
subscript_declarationcaptured as a method-kind symbol. - Ruby — top-level / class-level constant assignments (
MAX_RETRIES = 3). - PHP —
const_declarationandproperty_declaration(with or without explicit visibility) at both file and class scope. - C —
typedef int MyInt;andtypedef struct Foo Bar;aliases now produce symbols. - Java class/interface/record annotations —
(modifiers) @symbol.modifierscapture extended toclass_declaration,interface_declaration, andrecord_declarationso framework decorators surface in the symbol view.
Documentation extraction
- Java module-level Javadoc —
extract_module_docstringgains a Java branch that picks up a leading/** ... */block before the package/import declarations. - Luau docstrings — both
--[[ block comment ]]and runs of---triple-dash lines are extracted at module and symbol scope.
Fixed
- Java interface inheritance —
interface IFoo extends IBasenow produces a heritage relation; the extractor previously only recognised theinterfacesfield onclass_declarationand missedextends_interfacesoninterface_declaration. - Go struct embedding —
type Foo struct { Base }correctly emits a heritage edge fromFootoBase. The Go heritage extractor now traverses thefield_declaration_listchild when nobodyfield is present (matches the actual tree-sitter-go grammar layout). - Swift
extension_declarationheritage — extension conformance declarations now contribute heritage relations (extension_declarationwas missing from Swift'sheritage_node_types).
Changed
- Language tier promotion — C# moves from "Good" to "Full" in
README.mdanddocs/layers/LANGUAGE_SUPPORT.md. Eight languages now sit at Full tier (was: seven). - Heritage / bindings / dead-code internals refactored into per-language subpackages —
extractors/heritage.pyandextractors/bindings.py(previously 600+ LOC monoliths) andanalysis/dead_code.pyare now subpackages with one file per language plus a re-export shim. Public API (extract_heritage,extract_import_bindings,DeadCodeAnalyzer, etc.) is unchanged.
Tests
- +90 unit tests covering workspace-aware resolvers (PHP, Go, TypeScript, Swift, Kotlin, Scala, Ruby, Rust), framework-edge extraction (Spring, Rails, Laravel, Express/NestJS, Gin/Echo/Chi, Axum/Actix), per-language dynamic-hint extractors, and Java/Ruby/Scala/PHP/Go heritage + binding extractors.
0.3.1 — 2026-04-26
Added
- Output language for generated wiki content (#99) — set
language: ru(or any ofen,es,fr,de,zh,ja,ko,it,pt,nl,pl,tr,ar,hi) in.repowise/config.yamlto have the LLM produce documentation in that language. Code, paths, and symbol names stay untranslated. Cache keys include the language so different output languages do not collide. Closes #64. - Luau / Roblox language support (#89) — promotes the existing git-blame-only
luaLanguageSpec to a full AST-parsedluautier covering both.luaand.luau. Includes a dedicated resolver for string-literalrequireplusscript.Parentinstance paths and the:WaitForChild/:FindFirstChildRojo-safe idioms. Closes #52. - OpenRouter provider (#56) — new
openrouterLLM provider with fullstream_chatplus tool-call support, plus anOpenRouterEmbedderdefaulting togoogle/gemini-embedding-001. Sends OpenRouter's recommendedHTTP-RefererandX-Titleheaders. base_urlplus per-provider env vars (#85) — OpenAI, Anthropic, Gemini, Ollama, and LiteLLM all accept abase_url(withOPENAI_BASE_URL,ANTHROPIC_BASE_URL,GEMINI_BASE_URL,OLLAMA_BASE_URL,LITELLM_BASE_URLenv fallbacks) so users can route requests through proxies and self-hosted OpenAI-compatible endpoints.
Fixed
database is lockedon concurrentrepowise update(#101) — every SQLite connection now opens withjournal_mode=WAL,synchronous=NORMAL,busy_timeout=5000, andforeign_keys=ON. Two concurrent writers against the same workspace no longer collide; PostgreSQL is unchanged. Closes #95.- CLAUDE.md opt-out ignored in full mode (#102) — the "Generate .claude/CLAUDE.md? [Y/n]" prompt was nested inside the advanced-config flow, so users in full mode were never asked and the writer always created the file. Prompt is now extracted into a standalone helper and asked in both modes. Closes #81.
repowise initcould overwrite an unparseable user JSON config (#94) — when.mcp.jsonor.claude/settings.jsonexists but is not valid JSON, init now aborts with a clear error instead of silently treating the file as empty and overwriting the user's contents.- Editable installs and CI builds were broken (#97) —
[tool.setuptools].packagesreferencedrepowise.core.ingestion.parsers(no longer exists) and was missingextractors,languages, andresolvers(added during the language-support refactor). Resyncing the list unblockspip install -e .and every PR's CI. repowise servepointed at the wrong GitHub release —_GITHUB_REPOflipped fromRaghavChamadiya/repowisetorepowise-dev/repowiseso the web UI tarball downloads from the correct release URL. Project URLs on PyPI updated to match.
Changed
- PreToolUse hook — replaced FTS-only file retrieval with multi-signal ranking: symbol name match (highest weight), file path match, then FTS on wiki content. Returns top 3 files instead of 5. Removed git signals (HOTSPOT, bus-factor, owner) from enrichment output — use
get_riskfor that. Removed Bash command interception. Dependencies shown as "Uses" (2 per file) alongside symbols (3) and importers (3). - uv workflow documented and dev deps migrated to PEP 735 (#100) — README and USER_GUIDE document
uv tool install repowiseanduv sync --all-packages. Replaces the deprecated[tool.uv] dev-dependenciestable with[dependency-groups] dev, silencing thetool.uv.dev-dependencies is deprecatedwarning everyuv pip installwas emitting.
Security
- Bumps
dompurify3.3.3 → 3.4.1 (prototype-pollution + mXSS sanitizer-bypass fixes). - Bumps
gitpython3.1.46 → 3.1.47 (argument injection via underscored kwargs). - Bumps
mako1.3.10 → 1.3.11 (TemplateLookuppath traversal). - Bumps
litellm1.83.0 → 1.83.7 (routine patches). - Bumps
python-multipart0.0.22 → 0.0.26 (case-insensitive headers, MIME info).
0.3.0 — 2026-04-13
Added
Multi-repo workspaces
- Workspace support —
repowise init .from a parent directory scans for git repos (up to 3 levels deep), prompts for selection, and indexes each repo with cross-repo analysis. Config stored in.repowise-workspace.yaml. - Workspace CLI commands —
repowise workspace list,workspace add <path>,workspace remove <alias>,workspace scan,workspace set-default <alias>for managing repos in a workspace. - Workspace-aware MCP server — a single MCP server instance serves all workspace repos. Tools accept an optional
repoparameter to target a specific repo or"all"to query across the workspace. Lazy-loading with LRU eviction (max 5 repos loaded simultaneously). - Cross-repo co-change detection — analyzes git history across repos to find files that frequently change in the same time window.
- API contract extraction — scans for HTTP route handlers (Express, FastAPI, Spring, Go), gRPC service definitions, and message topic publishers/subscribers. Matches providers with consumers across repos.
- Package dependency scanning — reads package manifests (
package.json,pyproject.toml,go.mod,pom.xml) to detect inter-repo package dependencies. - Workspace CLAUDE.md — auto-generated context file at the workspace root covering all repos, their relationships, cross-repo signals, and contract links.
- Workspace web UI — workspace dashboard (
/workspace) with aggregate stats and repo cards, contracts view (/workspace/contracts) with provider/consumer matching, and co-changes view (/workspace/co-changes) with cross-repo file pairs ranked by strength. - Workspace update —
repowise update --workspaceupdates all stale repos in parallel (up to 4 concurrent) then re-runs cross-repo analysis.--repo <alias>targets a single repo. - Workspace watch —
repowise watch --workspaceauto-updates all workspace repos on file change.
Auto-sync hooks
repowise hookCLI —repowise hook installinstalls a marker-delimited post-commit git hook that runsrepowise updatein the background after every commit.hook install --workspaceinstalls for all workspace repos.hook statusandhook uninstallfor management.- Proactive context enrichment via Claude Code hooks —
repowise initregisters PreToolUse and PostToolUse hooks in~/.claude/settings.json. PreToolUse enriches everyGrep/Globcall with graph context (importers, dependencies, symbols, git signals) at ~24ms latency. PostToolUse detects git commits and notifies the agent when the wiki is stale. - Polling scheduler — when the server is running, a background job polls registered repositories every 15 minutes and triggers updates for new commits missed by webhooks.
Graph intelligence
- Symbol-level dependency graph — the dependency graph is now two-tier: file nodes for module-level relationships and symbol nodes (functions, classes, methods) for fine-grained call resolution. Call edges carry confidence scores (0.0–1.0).
- 3-tier call resolution — Tier 1: same-file targets (confidence 0.95). Tier 2: import-scoped targets via named bindings (0.85–0.93). Tier 3: global unique match (0.50). Extracted by tree-sitter for Python, TypeScript, JavaScript, Go, Rust, Java, and C++.
- Named binding resolution — tracks import aliases, barrel re-exports (
__init__.py,index.ts), and namespace imports across all 7 full-tier languages. - Heritage extraction — class inheritance and interface implementation for 11 languages (Python, TypeScript, JavaScript, Java, Go, Rust, C++, Kotlin, Ruby, C#, C) with
extends/implementsgraph edges. - Leiden community detection — two-level community detection (file communities from import edges, symbol communities from call/heritage edges) with cohesion scoring and heuristic labeling. Falls back to NetworkX Louvain when graspologic is unavailable.
- Execution flow tracing — 5-signal entry point scoring (fan-out ratio, in-degree, visibility, name pattern, framework hint) with BFS call-path discovery and cross-community classification.
- Graph query indexes (migration
0017) — composite indexes for sub-millisecond graph queries.
Web UI
- Graph Intelligence on Overview — expandable community list (labels, cohesion, member counts) and execution flows panel with call-path traces on the overview dashboard.
- Wiki sidebar — new collapsible section showing PageRank and betweenness percentile bars, community label, and in/out degree for the current file.
- Symbols drawer — right panel with graph metrics, callers/callees (with confidence scores), and heritage (extends/implements) for class nodes.
- Graph page — community color mode uses real community labels from Leiden detection. Clicking a node opens a community detail panel. Active color mode preserved as a URL parameter.
- Contributor network, hotspot, and ownership pages — new dedicated pages for git intelligence.
- Docs viewer — enriched with graph intelligence sidebar, version history, and improved markdown rendering.
- 5 new graph REST API endpoints — communities list, community detail, node metrics, callers/callees, and execution flows.
Other
- Improved init UX — pre-scan phase shows repo size and language breakdown before confirming. Advanced config options grouped logically with live insights during indexing.
- Doc generation enriched with graph intelligence — wiki page generation prompts now include community context, caller/callee information, and heritage relationships.
Changed
get_overviewnow includescommunity_summary— top communities by size with labels and cohesion scores.get_contextnow includescommunityblock per file target with community ID and label (whencompact=False). In workspace mode, enriched with cross-repo co-change and contract data.get_riskenriched with cross-repo signals in workspace mode — co-change partners from other repos and contract dependencies.search_codebasein workspace mode searches across all repos and merges results.- Job executor — improved progress tracking, concurrent run detection (HTTP 409), and crash recovery for stale running jobs on server startup.
0.2.3 — 2026-04-11
Added
annotate_fileMCP tool — attach human-authored notes to any wiki page. Notes survive LLM-driven re-generation and appear inget_contextresponses and the web UI. Pass an empty string to clear notes.repowise export --full— full JSON export now includes decision records, dead code findings, git hotspots, and per-page provenance metadata (confidence, freshness, model, provider).- Rust import resolution —
use crate::,super::, andself::imports now resolve to local files via crate root detection (lib.rs/main.rs). External crates mapped toexternal:nodes. - Go import resolution —
go.modmodule path parsing enables accurate local vs external package classification. Local imports resolve by suffix matching against the module path. - C/C++ parser improvements — added captures for
template_declaration,type_definition(typedef struct/enum),preproc_def(#define),preproc_function_def, and forward declarations. - Go parser — added
const_specandvar_speccaptures for package-level constants and variables. - Rust parser — added
macro_definitioncapture formacro_rules!macros. - Dynamic import detection — dead code analysis now scans for
importlib.import_module()and__import__()calls; files in the same package receive reduced confidence (capped at 0.4). - Framework decorator awareness — Flask, FastAPI, and Django route/endpoint decorators added to
_FRAMEWORK_DECORATORS. Decorated functions are never flagged as dead code. human_notescolumn on wiki pages — persists across re-indexing. Alembic migration0014_page_human_notes.- Decision staleness scoring during ingestion —
compute_staleness()now runs duringrepowise init, not justrepowise update.
Changed
- CLAUDE.md template — replaced imperative "MUST use" / "CRITICAL" language with advisory framing. Added
indexed_commitdisplay. Madeupdate_decision_recordsoptional ("SHOULD for architectural changes"). get_contextfreshness — freshness data now included by default instead of requiring explicitinclude=["freshness"].get_answerdocstring — removed "do NOT verify by Read" instruction. High-confidence note changed to "verify cited file paths exist before acting on them".- Token budget caps —
get_overviewcaps knowledge_silos (30), module_pages (20), entry_points (15).get_whycaps file_commits (10). - Dead code patterns — expanded
_DEFAULT_DYNAMIC_PATTERNSwith*Mixin,*Command,*_view,*_endpoint,*_route,*_callback,*_signal,*_task.
Docs
- README — tool count updated to 11,
annotate_fileadded to MCP tools table,--fullexport flag documented, dynamic import detection noted, comparison table updated. - Supported languages — tiered table with accurate "What works" descriptions per language.
- Updated USER_GUIDE.md, ARCHITECTURE.md, and deep-dives.md to reflect all changes.
0.2.2 — 2026-04-11
Added
- tsconfig/jsconfig path alias resolution (#40) — new
TsconfigResolverdiscovers alltsconfig.json/jsconfig.jsonfiles, resolvesextendschains (with circular detection), and maps path aliases (e.g.@/*->src/*) to real files during graph construction. Non-relative TS/JS imports that match a path alias now create proper internal edges instead of phantomexternal:nodes. Fixes broken dependency graph, PageRank, dead code false positives, and change propagation for any TS/JS project using path aliases (Next.js, Vite, Angular, Nuxt, CRA). - Traversal stats (#57) —
FileTraversernow tracks skip reasons (.gitignore, blocked extension, binary, oversized, generated,--exclude,.repowiseIgnore, unknown language) via a newTraversalStatsdataclass. Stats are surfaced after traversal as a filtering summary showing how many files were included vs excluded and why. - Submodule handling (#57) — git submodule directories (parsed from
.gitmodules) are now excluded by default during traversal. Added--include-submodulesflag torepowise initto opt in. - Language breakdown (#57) — generation plan table now shows language distribution (e.g. "Languages: python 79%, typescript 14%"). Completion panel shows top languages with percentages instead of just a count.
- Multi-line exclude input — interactive advanced mode now prompts for exclude patterns one per line instead of comma-separated on a single line.
- 38 new unit tests covering tsconfig resolver, traversal stats, and submodule handling.
Changed
- Traverse progress bar uses spinner mode instead of showing misleading pre-filter totals (e.g. "2132/83601").
- Traverse phase label changed from "Traversing files..." to "Scanning & filtering files...".
Fixed
- Server tests now use real temp directories with
.gitfolders for path validation (#69 compatibility).
Docs
- Updated README CLI reference with
--index-only,-x, and--include-submodulesexamples. - Updated website docs (
cli-reference.md,configuration.md,getting-started.md) with submodule handling,.gitignoredocumentation, and new output examples. - Reorganized
docs/directory: architecture docs intodocs/architecture/, internals intodocs/internals/. - Removed stale one-time documents (PHASE_5_5_IMPLEMENTATION, GIT_INTELLIGENCE_AUDIT, MCP_AND_STATE_REVIEW, MCP_TOOLS_TEST_REPORT).
0.2.1 — 2026-04-10
Added
get_answerMCP tool (tool_answer.py) — single-call RAG over the wiki layer. Runs retrieval, gates synthesis on top-hit dominance ratio, and returns a 2–5 sentence answer with concrete file/symbol citations plus aconfidencelabel. High-confidence responses can be cited directly without verification reads. Backed by anAnswerCachetable so repeated questions on the same repository cost nothing on the second call.get_symbolMCP tool (tool_symbol.py) — resolves a fully-qualified symbol id (path::Class::method, also acceptsClass.method) to its source body, signature, file location, line range, and docstring. Returns the rich source-line signature (with base classes, decorators, and full type annotations preserved) instead of the stripped DB form.Page.summarycolumn — short LLM-extracted summary (1–3 sentences) attached to every wiki page during generation. Used byget_contextto keep context payloads bounded on dense files. Added by alembic migration0012_page_summary.AnswerCachetable — memoisedget_answerresponses keyed by(repository_id, question_hash)plus the provider/model used. Added by alembic migration0013_answer_cache. Cache entries are repository-scoped and invalidated by re-indexing.- Test files in the wiki —
page_generator._is_significant_file()now treats any file taggedis_test=True(with at least one extracted symbol) as significant, regardless of PageRank. Test files have near-zero centrality because nothing imports them back, but they answer "what test exercises X" / "where is Y verified" questions; the doc layer is the right place to surface those. Filtering remains available via--skip-tests. - Overview dashboard (
/repos/[id]/overview) — new landing page for each repository with:- Health score ring (composite of doc coverage, freshness, dead code, hotspot density, silo risk)
- Attention panel highlighting items needing action (stale docs, high-risk hotspots, dead code)
- Language donut chart, ownership treemap, hotspots mini-list
- Decisions timeline, module minimap (interactive graph summary)
- Quick actions panel (sync, full re-index, generate CLAUDE.md, export)
- Active job banner with live progress polling
- Background pipeline execution —
POST /api/repos/{id}/syncandPOST /api/repos/{id}/full-resyncnow launch the full pipeline in the background instead of only creating a pending job. Concurrent runs on the same repo return HTTP 409. - Shared persistence layer (
core/pipeline/persist.py) —persist_pipeline_result()extracted from CLI, reused by both CLI and server job executor - Job executor (
server/job_executor.py) — background task that runsrun_pipeline(), writes progress to theGenerationJobtable, and persists all results - Server crash recovery — stale
runningjobs are reset tofailedon server startup - Async pipeline improvements —
asyncio.wrap_futurefor file I/O,asyncio.to_threadfor graph building and thread pool shutdown, periodicasyncio.sleep(0)yields during parsing - Health score utility (
web/src/lib/utils/health-score.ts) — composite health score computation, attention item builder, and language aggregation for the overview dashboard
Changed
get_contextdefault is nowcompact=True— drops thestructureblock, theimported_bylist, and per-symbol docstring/end-line fields to keep the response under ~10K characters. Passcompact=Falsefor the full payload (e.g. when you specifically need import-graph dependents on a large file).init_cmd.pyrefactored to use sharedpersist_pipeline_result()instead of inline persistence logicPipeline orchestrator uses async-friendly patterns to keep the event loop responsive during ingestion
Sidebar and mobile nav updated to include "Overview" link
Monorepo scaffold: uv workspace with
packages/core,packages/cli,packages/server,packages/webProvider abstraction layer:
BaseProvider,GeneratedResponse,ProviderError,RateLimitErrorAnthropicProviderwith prompt caching supportOpenAIProviderwith OpenAI Chat Completions APIOllamaProviderfor local offline inference (OpenAI-compatible endpoint)LiteLLMProviderfor 100+ models via LiteLLM proxyMockProviderfor testing without API keysRateLimiter: async sliding-window RPM + TPM limits with exponential backoffProviderRegistry: dynamic provider loading with custom provider registrationCI pipeline: GitHub Actions matrix on Python 3.11, 3.12, 3.13
Pre-commit hooks: ruff lint + format, mypy, standard file checks
Folder exclusion — three-layer system for skipping paths during ingestion:
FileTraverser(extra_exclude_patterns=[...])— pass gitignore-style patterns at construction time; applied to both directory pruning and file-level filtering- Per-directory
.repowiseIgnore— traverser loads one from each visited directory (like git's per-directory.gitignore); patterns are relative to that directory and cached for efficiency repowise init --exclude/-x PATTERN— repeatable CLI flag; patterns are merged withexclude_patternsfromconfig.yamland persisted back to.repowise/config.yamlrepowise updatereadsexclude_patternsfromconfig.yamlautomatically- Web UI Excluded Paths section on
/repos/[id]/settings: chip editor, Enter-to-add input, six quick-add suggestions (vendor/,dist/,build/,node_modules/,*.generated.*,**/fixtures/**), empty-state message, gitignore-syntax tooltip; saved viaPATCH /api/repos/{id}assettings.exclude_patterns helpers.save_config()now round-tripsconfig.yamlto preserve all existing keys when updating provider/model/embedder; accepts optionalexclude_patternskeyword argumentscheduler.pylogsrepo.settings.exclude_patternsin polling fallback as preparation for future full-sync wiring
13 new unit tests in
tests/unit/ingestion/test_traverser.pycoveringextra_exclude_patternsand per-directory.repowiseIgnorebehaviour
0.2.0 — 2026-04-07
A large overhaul: faster indexing, smarter doc generation, transactional storage, new analysis capabilities, and a completely revamped web UI that surfaces every new signal — all without changing the eight MCP tool surface.
Added
Pipeline & ingestion
- Parallel indexing. AST parsing now runs across all CPU cores via
ProcessPoolExecutor. Graph construction and git history indexing run concurrently withasyncio.gather. Per-file git history fetched through a thread executor with a semaphore. - RAG-aware doc generation. Pages are generated in topological order; each generation prompt now includes summaries of the file's direct dependencies, pulled from the vector store of already-generated pages.
- Atomic three-store coordinator. New
AtomicStorageCoordinatorbuffers writes across SQL, the in-memory dependency graph, and the vector store, then flushes them as a single transaction. Failure in any store rolls back all three. - Dynamic import hint extractors. The dependency graph now captures edges
that pure AST parsing misses: Django
INSTALLED_APPS/ROOT_URLCONF/MIDDLEWARE, pytestconftest.pyfixture wiring, and Node/TS path aliases fromtsconfig.jsonandpackage.jsonexports.
Analysis
- Temporal hotspot decay. New
temporal_hotspot_scorecolumn ongit_metadata, computed asΣ exp(-ln2 · age_days / 180) · min(lines/100, 3)per commit. Hotspot ranking now uses this score; commits from a year ago contribute ~25% as much as commits from today. - Percentile ranks via SQL window function.
recompute_git_percentiles()is now a singlePERCENT_RANK() OVER (PARTITION BY repo ORDER BY ...)UPDATE instead of an in-Python sort. Faster and correct on large repos. - PR blast radius analyzer. New
PRBlastRadiusAnalyzerreturns direct risks, transitive affected files, co-change warnings, recommended reviewers, test gaps, and an overall 0-10 risk score. Surfaced viaget_risk(changed_files=...)and a new web page. - Security pattern scanner. Indexing now runs
SecurityScannerover each file. Findings (eval/exec, weak crypto, raw SQL string construction, hardcoded secrets,pickle.loads, etc.) are stored in a newsecurity_findingstable. - Knowledge map. Top owners, "bus factor 1" knowledge silos (>80% single
owner), and high-centrality "onboarding targets" with thin documentation --
surfaced in
get_overviewand the web overview page.
LLM cost tracking
- New
llm_coststable records every LLM call (model, tokens, USD cost). CostTrackeraggregates session totals; pricing covers Claude 4.6 family, GPT-4.1 family, and Gemini.- New
repowise costsCLI:--since,--by operation|model|day. - Indexing progress bar shows a live
Cost: $X.XXXXcounter.
MCP tool enhancements (still 8 tools -- strictly more capable)
get_risk(targets, changed_files=None)-- whenchanged_filesis provided, returns the full PR blast-radius report (transitive affected, co-change warnings, recommended reviewers, test gaps, overall 0-10 score). Per-file responses now includetest_gap: boolandsecurity_signals: list.get_overview()-- now includes aknowledge_mapblock (top owners, silos, onboarding targets).get_dead_code(min_confidence?, include_internals?, include_zombie_packages?)-- sensitivity controls for false positives in framework-heavy code.
REST endpoints (new)
GET /api/repos/{id}/costsand/costs/summary-- grouped LLM spend.GET /api/repos/{id}/security-- security findings, filterable by file/severity.POST /api/repos/{id}/blast-radius-- PR impact analysis.GET /api/repos/{id}/knowledge-map-- owners / silos / onboarding targets.GET /api/repos/{id}/health/coordinator-- three-store drift status.GET /api/repos/{id}/hotspotsnow returnstemporal_hotspot_scoreand is ordered by it.GET /api/repos/{id}/git-metadatanow returnstest_gap.- Job SSE stream now emits
actual_cost_usd(running cost since job start).
Web UI (new pages and components)
- Costs page -- daily bar chart, grouped tables by operation/model/day.
- Blast Radius page -- paste files (or click hotspot suggestion chips) to see risk gauge, transitive impact, co-change warnings, reviewers, test gaps.
- Knowledge Map card on the overview dashboard.
- Trend column on the hotspots table with flame indicator (default sort).
- Security Panel in the wiki page right sidebar.
- "No tests" badge on wiki pages with no detected test file.
- System Health card on the settings page (SQL / Vector / Graph counts + drift % + status).
- Live cost indicator on the generation progress bar.
CLI
repowise costs [--since DATE] [--by operation|model|day]-- new command.repowise dead-code-- new flags--min-confidence,--include-internals,--include-zombie-packages,--no-unreachable,--no-unused-exports.repowise doctor-- new Check #10 reports coordinator drift across all three stores.--repairdeletes orphaned vectors and rebuilds missing graph nodes from SQL.
Fixed
- C++ dependency resolution edge cases.
- Decision extraction timeout on very large histories.
- Resume / progress bar visibility for oversized files.
- Coordinator
health_checkfalsely reporting 100% drift on LanceDB / Pg vector stores (was returning -1 for the count). Now useslist_page_ids(). - Coordinator
health_checkreturningnullgraph node count when no in-memoryGraphBuilderis supplied. Now falls back to SQLCOUNT(*).
Internal
- Three new Alembic migrations:
0009_llm_costs,0010_temporal_hotspot_score,0011_security_findings.
Compatibility
- Existing repositories must run migrations:
repowise doctorwill detect the missing tables and prompt; alternatively re-runrepowise initto rebuild from scratch. - The eight MCP tool names and signatures are backwards compatible -- new parameters are all optional.
[0.1.31] — earlier
See git history for releases prior to 0.2.0.