Find code by asking what it does. A CLI for coding agents that uses Jev to discover relevant files and source context.
TypeScript ai-sdkclaude-codeclicode-searchcodexcoding-agentscontext-retrievaldeveloper-toolsjevsemantic-searchtypescriptvercel-ai-gateway
Separate completed parser restoration from active quality experiment
David Zhange04d7da97e83
diff --git a/specs/jevgrep/README.md b/specs/jevgrep/README.mdindex 2c9c065..0433205 100644--- a/specs/jevgrep/README.md+++ b/specs/jevgrep/README.md@@ -7,8 +7,10 @@ Jev classification through Vercel AI Gateway. This is the build plan; production## Next Agent PromptYou are implementing this spec. Status: **implementing**, updated **2026-09-26**.-Next: prepare and run Django first in the isolated-[mechanism-first skill experiment](assets/query-framing-study.md).+Next: follow the isolated `installed-jg-mechanism-query-v1b` plan in the+[mechanism-first skill experiment](assets/query-framing-study.md). Inspect the+retained Django attempt and its official grade/bill before applying the registered+continuation rule. Do not launch another attempt while its process is active.The frozen `installed-jg-cpython-parity-v1` cohort is complete and not accepted:7/10 solves versus 8/10 baseline, six cost wins, and 35.27% lower full Sol cost.[Confirmation evidence](assets/cpython-confirmation.md) owns all outcomes, trace@@ -25,23 +27,26 @@ exercised corpus; task-quality acceptance failed and remains required.`installed-jg-reference-parity-v1` is stopped and superseded. Requests andscikit-learn officially solved but exceeded baseline cost; Django was interruptedwith its patch and traces retained; seven cells never started. These superseded-studies remain stopped. The bundled-CPython cohort is terminal; no Sol run is-active. The completed query diagnostics made only bounded Jev calls. A separately-identified skill candidate is prepared for an actual Django task test.+studies remain stopped. The bundled-CPython cohort is terminal. The completed+query diagnostics made only bounded Jev calls; the separate mechanism-first plan+owns the new skill candidate and its Django-first evaluation.The changed-policy cohort `installed-jg-final-cohort-v1` was stopped and markedsuperseded, retaining completed outcomes and interrupted traces. It is not arelease acceptance cohort. Never rerun saved baselines or pool candidates.Report Jev observed API costs separately from scored full Sol cost.-Keep the restored expansion, discovery inputs/order, excerpt presentation, scoped-guidance and exact accepted skill (only rename executable to `jg`). Required+Keep the production incumbent's restored expansion, discovery inputs/order,+excerpt presentation, scoped guidance and accepted skill (only rename executable+to `jg`) unchanged. The isolated experiment changes only the recorded skill+instruction and does not promote that change into production. RequiredNode-only runtime, auth, cache, filesystem protections and truthful failurehandling remain, but must not silently change healthy retrieval decisions.The earlier deterministic and platform gates passed for the deviating port;they do not prove parity. See [parity restoration](assets/parity-restoration.md) and the[source-budget investigation](assets/source-budget-decision.md).-Remaining: frozen quality confirmation, whole-spec review and close-spec.+Remaining: a candidate that passes frozen quality confirmation, final choices+consolidation, whole-spec closeout review and close-spec.See [contracts](contracts.md) for the corrected contract.No user decision blocks starting. External verification dependencies are a working@@ -53,8 +58,8 @@ acceptance claim. Do not repair credentials or regenerate an existing baseline.Implement one focused slice at a time, running its narrow tests and recording itsartifact. Use write-tests for behavioral tests and review before completing animplementation pass. Update this section, slice status, decisions and unresolved-verification before ending each pass. These commands are targets to implement,-not claims about the current scaffold.+verification before ending each pass. The named commands are implemented; their+recorded evidence does not substitute for the open quality gate.- [x] [01 — Reference and real HTTP fixture](slices/01-reference.md)- [x] [02 — Bundled parser parity](slices/02-parser.md)@@ -143,7 +148,7 @@ The sample was tuned and Python-only. TS/JS conformance tests establish parserbehavior, not downstream solve generalization. Arbitrary roots are supported bycontract; whole-computer effectiveness is not established by synthetic scale tests.-## Draft synthesis and fog audit+## Planning rationale and remaining boundariesThree independent fresh-context drafts used the same interview brief: Codex withfewest-slices and seam-quality biases; Claude Opus/high with a risk-first bias.@@ -155,14 +160,16 @@ first task, and separates failures/cache/output evidence instead of bundling theinto a vague hardening phase.Claude proposed Pyodide for closer CPython parity. The initial smaller grammar-parser failed that requirement; the measured fallback is now being integrated.-This reopens the parser mechanism, not the Node-only user contract. Do not-normalize semantically meaningful source or question order. The skill-path ambiguity was checked: the confirmation-hash matches the retained ranked-leads skill; the current working skill has changed.+parser failed that requirement; bundled CPython now executes the unchanged+reference helpers. [Runtime evidence](assets/python-runtime.md) records the+completed integration, verification and interpreter-version limits. The Node-only+user contract is unchanged. Do not normalize meaningful source or question order.+The incumbent skill matches the retained accepted instructions apart from the+executable rename; the separate skill experiment is not a production replacement.Recursive fog audit: parser compatibility has its own artifact (02); eligibility-and snapshot integrity (03); service outcomes (05); freshness (06); numerical-source budget (07); actual task quality (08). Each open implementation freedom is+and snapshot integrity (03); service outcomes (05); freshness (06); preserved+output policy (07); actual task quality (08). Each open implementation freedom isnamed in its owning slice. New behavior decisions outside those delegations reopenthe spec instead of becoming silent defaults. No untested representation change,new recall heuristic or performance optimization hides in the integration slice.diff --git a/specs/jevgrep/assets/cpython-confirmation.md b/specs/jevgrep/assets/cpython-confirmation.mdindex b1843ac..a84cee1 100644--- a/specs/jevgrep/assets/cpython-confirmation.md+++ b/specs/jevgrep/assets/cpython-confirmation.md@@ -199,14 +199,15 @@ passing domain tests. Current also reproduces a broader warning-related failurein a pristine checkout. Its added regression covers fewer field variants than thespike, but no implementation defect was found. Evidence: `/tmp/jg-sphinx8638-paired/`.-## Next focused experiment--Do not rerun or replace this cohort. Preserve the CLI, parser, traversal, thresholds,-requests and skill as the recorded incumbent. Investigate query framing separately-on Django, where solving the retained regression is the most direct unresolved-quality question. The next bounded diagnostic holds a saved SQLite declaration-request constant and varies only its query; any later skill candidate must be-identified separately and validated on actual task outcomes. A classifier score-increase is insufficient for promotion. Sol baselines remain immutable, and any-new candidate must still preserve all eight baseline solves and achieve seven-successful lower-cost solves before final acceptance.+## Separate mechanism-first experiment++This cohort is terminal and failed acceptance; do not rerun or replace it. Keep+its CLI, parser, traversal, thresholds, requests and skill as the recorded+incumbent. The bounded Django query diagnostic is complete, and the separately+identified skill trial now belongs to the+[mechanism-first experiment](query-framing-study.md). That record owns the+candidate, preregistered Django-first continuation rule and subsequent outcomes.+A classifier score increase or one task solve is insufficient for promotion.+Sol baselines remain immutable, and any new candidate must still preserve all+eight baseline solves and achieve seven successful lower-cost solves before+final acceptance.diff --git a/specs/jevgrep/assets/parity-restoration.md b/specs/jevgrep/assets/parity-restoration.mdindex 799d33a..703ebea 100644--- a/specs/jevgrep/assets/parity-restoration.md+++ b/specs/jevgrep/assets/parity-restoration.md@@ -1,4 +1,12 @@-# Restore the accepted spike before further evaluation+# Historical restoration of the accepted spike++This record preserves the pre-CPython restoration work and its superseded+cohorts. The bundled-CPython port subsequently passed integration, platform and+whole-product checks; [runtime evidence](python-runtime.md) owns that coverage.+Its completed quality cohort failed acceptance, as recorded in+[confirmation evidence](cpython-confirmation.md). The separate+[mechanism-first skill experiment](query-framing-study.md) owns the active trial;+none of the stopped studies below is reactivated.The user rejected unvalidated architecture changes during production porting.The accepted reference is the source closure and skill bound by@@ -6,10 +14,12 @@ The accepted reference is the source closure and skill bound byThe required executable rename is `jevgrep` → `jg`; packaging, credentials,cache and Node-only runtime must preserve healthy retrieval behavior.-Restore computed discovery inputs and grouping, selection and bounded second-expansion, exact source evidence serialization, numbered stdout packets, scoped-guidance lookup, test suggestions, and the accepted skill. The benchmark prompt-returns to the accepted `$jevgrep` invocation plus the immutable baseline task.+The restoration required computed discovery inputs and grouping, selection and+bounded second expansion, exact source evidence serialization, numbered stdout+packets, scoped+guidance lookup, test suggestions, and the accepted skill to match the reference.+The benchmark prompt returned to the accepted `$jevgrep` invocation plus the+immutable baseline task.Do not use matching prompt templates as proof that their computed inputs match.Use the frozen executable/selection block with fixed HTTP answers and compareactual native request objects, ranges and complete output. Keep array ordering;@@ -39,7 +49,7 @@ solve rate and full task cost; timing is diagnostic. Any subsequent cachemaintenance change must preserve answer/cache behavior and identify its artifactseparately from this confirmation.-## Preservation evidence+## Preservation evidence and historical checkpointThe retained winning executable is`evals/runs/swebench/auto-research-80/unit-locators.mjs`, SHA-256@@ -65,12 +75,12 @@ it does not prove all possible parser inputs equivalent or establish live taskcost. Historical live Jev answers were not retained, so reconstructing selectedRequests ranges does not prove the cause of every historical cost difference.-The merged integration checkpoint passes 87 core/reference tests, 16 parser+The pre-CPython integration checkpoint passed 87 core/reference tests, 16 parsertests, three release tests, 14 benchmark-harness tests and all 20 installedjourneys, including discovery'sdeclaration index, large-test suggestions, TypeScript context expansion and thelarge-file anchor/selection boundary. Native and alternate-architecture-verification uses the final restored archive. SHA-256+verification used that checkpoint archive. SHA-256`10cc842f25a11ff1c62cb109df27cd08b9875892907f2e89eb7bac26665a1e85`passed native macOS arm64 on Node 24.14.0 and both Linux amd64 and arm64 installedsmokes on Node 22.23.3. These are package/runtime checks, not live quality results.@@ -132,7 +142,7 @@ agent recovered from a patch-context failure and a rejected test command beforerunning its full module once; those steps were downstream of retrieval. Apassing official grade does not establish all behavioral edge cases equivalent.-## Remaining review gate+## Historical review correctionsThe whole-product review reproduced two ordinary preservation defects. Thereference builds second-pass `selectedEvidence` in source-selection completion@@ -142,25 +152,28 @@ returned before collecting comments, so a selected fragment inside a blockcomment ended inside it instead of expanding to the reference's full comment.Commits `a7a28bb` and `bcf3253` fix these defects with whole-CLI regressions thatfailed before the corrections and now compare complete native requests and stdout-against the frozen executable. The merged gate and whole-product review must-finish before another frozen candidate.+against the frozen executable. The later bundled-CPython integration completed+the merged gate and whole-product review; these were prerequisites for its now+terminal confirmation, not outstanding parser gates.The interrupted Django attempt has a known Sol subtotal of $0.9671056 and knownJev subtotal of $0.277447296. Neither is a complete total: the final coding-agentresponse was interrupted, and five of 1,267 Jev calls lack complete cost responses.These costs remain retained even though the attempt cannot count as a solve.-Finish whole-product review before paid confirmation, in addition to matched-HTTP tests and supported-runtime gates. The quality target remains every fixed+Whole-product review, matched HTTP tests and supported-runtime checks remain+prerequisites for candidate confirmation. The quality target remains every fixedbaseline solve preserved and at least seven solved strict Sol-cost wins in oneprospective ten-task cohort. Historical or interrupted attempts cannot be pooled.-## Python runtime parity remains open+## Python runtime replacement completedFurther whole-product review found Python decorator, line-ending and invalid-AST-mismatches beyond the already corrected TS/JS comments and selection ordering.-The grammar approximation is being replaced with unchanged reference helpers-executed in bundled CPython. [Runtime evidence](python-runtime.md) owns the bounded-probe and remaining integration requirements. Neither stopped paid cohort is-reactivated by this change; merged verification and a clean full review must-precede a new prospective confirmation.+mismatches beyond the corrected TS/JS comments and selection ordering. The grammar+approximation was replaced with unchanged reference helpers executed in bundled+CPython. [Runtime evidence](python-runtime.md) owns the completed probe,+integration and supported-runtime checks, including the finite interpreter-version+coverage. Those checks do not prove every possible parser input equivalent or+establish task quality. Neither stopped cohort was reactivated; the subsequent+CPython cohort was separately frozen and its failed quality result remains+recorded in [confirmation evidence](cpython-confirmation.md).