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
docs: record integrated freshness correction and verification boundary
David Zhang4a2f71959cab
diff --git a/specs/jevgrep/README.md b/specs/jevgrep/README.mdindex 6a4e8c0..867cdca 100644--- a/specs/jevgrep/README.md+++ b/specs/jevgrep/README.md@@ -12,15 +12,14 @@ its frozen artifact: eight baseline solves preserved, eight cost wins, 33.21%lower total Sol cost. [Confirmation evidence](assets/work-clock-confirmation.md)owns all ten results and their limits. No paid benchmark or grader remains active.-Next: integrate the filesystem-freshness fixes being implemented in the isolated-`codex/final-freshness` worktree. Final independent review found buffered navigation-source uploaded after ignore changes and stale non-Python excerpts returned after-role evaluation. Required final freshness checks and per-network-attempt validation-must fix those cases without changing healthy spike requests or output. Run their-red/green installed tests and healthy parity, then determine what remains necessary-to validate the corrected artifact. Do not transfer the frozen cohort's measured-results to changed code without evidence. The pre-fix full default gate passed;-its green result did not cover these newly reproduced cases.+Next: finish serial merged verification and no-call preparation, then validate+and run the separate `installed-jg-freshness-v1` cohort against saved baselines. The freshness+correction is integrated in `92ca7f9`; its isolated worktree is removed. Exact+archive checks passed on macOS and both Linux architectures. The first full gate+had one five-second timeout under concurrent platform testing; no test allowance+was changed. [Freshness evidence](assets/source-freshness.md) owns the corrected+archive identity, failure reproductions and verification status. Do not transfer+the previous cohort's measured results to the corrected archive.Preserve healthy retrieval semantics and the canonical skill while fixingfilesystem freshness. Keep the measured archive immutable; it matches the@@ -42,7 +41,7 @@ Evidence and boundaries:[presentation trial](assets/presentation-study.md): rejected experiments,source-parity probes and limits on causal explanations.-Priority: validate and integrate the freshness correction without weakening+Priority: validate the integrated freshness correction without weakeningacceptance; resolve the whole-product review; consolidate choices against thefinal code; then close-spec. Slice 08 remains open. No publication orrelease tag is authorized for the intentional `0.0.0` development checkpoint.diff --git a/specs/jevgrep/assets/source-freshness.md b/specs/jevgrep/assets/source-freshness.mdnew file mode 100644index 0000000..f4395c4--- /dev/null+++ b/specs/jevgrep/assets/source-freshness.md@@ -0,0 +1,50 @@+# Source freshness at deferred boundaries++The final review reproduced two gaps: queued navigation could upload buffered+content after an ignore change, and a non-Python file changed during role+evaluation could retain stale excerpts. Earlier green checks did not cover those+boundaries. The correction is committed in `92ca7f9`.++Retrieval now binds request inputs to their source snapshots without adding that+metadata to Jev's payload. It checks eligibility and content before evaluation,+again before each provider attempt after any retry delay, and before returning+the final result. The existing filesystem reader remains the policy owner.+Source invalidation removes stale evidence and marks discovery incomplete.++Two details preserve useful behavior. Preflight reads become ready in FIFO order,+so adding asynchronous validation does not reorder healthy requests. When a+buffered group contains an invalid donor, finite splitting retains healthy+siblings rather than discarding the entire group. Neither changes healthy+prompts, thresholds, source budgets, skill text or rendering.++This is revalidation, not a filesystem lock. Sequential checks cannot prevent a+file changing afterward, and already transmitted requests cannot be recalled.+The guarantee is that detected changes are not knowingly reused at subsequent+validation boundaries.++## Evidence and candidate identity++The new real-filesystem/HTTP regressions cover queued requests, provider retries,+relationship and cross-file donors, healthy siblings and final output. The packed+CLI also exercises queued uploads and final invalidation. Pre-fix reproductions+and corrected results are retained under `/tmp/jg-freshness-*-red*.log` and+`/tmp/jg-freshness-*-green.log`; the maintained tests are in+`test/retrieval-freshness.test.ts` and `test/installed.test.mjs`.++The corrected frozen archive is+`/tmp/jg-freshness-candidate/dzhng-jevgrep-0.0.0.tgz`, SHA-256+`4589072fc0ff44257599dddda8bc5807c9c5515bbe1dcaedfe729645d9c4622c`.+Its native macOS arm64 smoke passed two cases. Both Linux arm64 and amd64 passed+all seven selected installed cases, including source freshness and Python assets.+Logs: `/tmp/jg-freshness-native.log` and+`/tmp/jg-freshness-frozen-{arm64,amd64}.log`.++The first merged default verification passed 109 tests but timed out one existing+donor test at five seconds while both platform checks ran concurrently. This is+not a passing full gate. The serial rerun passed that donor test in 0.51 seconds+and all 110 Bun tests. The remaining full verification is running, with evidence+in `/tmp/jg-freshness-merged-verify-serial.log`; no test allowance was changed.++The accepted work-clock cohort measured the preceding archive. A separate frozen+ten-task treatment cohort must validate this corrected executable; saved baselines+remain immutable and earlier treatment cells cannot be pooled into its result.diff --git a/specs/jevgrep/choices.md b/specs/jevgrep/choices.mdindex 7047933..9d4d7aa 100644--- a/specs/jevgrep/choices.md+++ b/specs/jevgrep/choices.md@@ -221,6 +221,21 @@ Gap: artifact layout and execution mechanics were delegated. Reach: results fromdifferent candidates cannot be pooled into acceptance; failed attempts and unknownbills remain visible. Verdict: sound. Confidence: high.+### Recheck source when deferred work is ready to use it++When: final source-freshness correction. A queued request may wait while the user+edits or ignores a file. Bind its content to an in-memory snapshot, then use the+same filesystem policy to check that snapshot before evaluation, provider retries+and final output. Keep this bookkeeping out of Jev's input. A detected change+removes stale evidence and makes the result incomplete.++Gap: source freshness was required, but the deferred validation boundaries were+not fully specified. Reach: these checks are observations, not locks; changes+after a check remain possible. Validation readiness preserves request order, and+splitting invalid groups preserves their healthy siblings. This adds filesystem+reads without changing the healthy retrieval policy. Verdict: sound. Confidence:+high for detected-change handling; no claim of atomic filesystem consistency.+### Reuse installed command journeys for native macOS verificationWhen: supported-runtime verification. Install a tarball into a temporary npm