Skip to content

Find code by asking what it does. A CLI for coding agents that uses Jev to discover relevant files and source context.

Stop paid confirmation until whole-product preservation review is clean

David Zhang76eec4ae55d0
diff --git a/specs/jevgrep/README.md b/specs/jevgrep/README.md
index d3cecf7..439549a 100644
--- a/specs/jevgrep/README.md
+++ b/specs/jevgrep/README.md
@@ -7,16 +7,17 @@ Jev classification through Vercel AI Gateway. This is the build plan; production
## Next Agent Prompt
You are implementing this spec. Status: **implementing**, updated **2026-09-26**.
-Next: finish and grade the frozen `installed-jg-reference-parity-v1` cohort,
-account for every attempt, and compare its traces with the fixed baselines.
-The user explicitly rejected architecture changes disguised as implementation
-cleanup. The deterministic parity and supported-runtime gates now pass.
-Selection, discovery, parser fallback, output, skill and guidance repairs are
-integrated. Matched HTTP comparisons exercise the production entry point against
-the frozen winner; the preservation matrix records their scope and limits.
-The exact archive passed macOS arm64 and Linux amd64/arm64 checks. Requests
-officially solved but exceeded baseline cost. The other nine cells are running
-sequentially, with no policy changes or baseline reruns.
+Next: close the whole-product review's two reproduced preservation defects:
+second-pass evidence must follow selection-completion order, and invalid TS/JS
+must retain the reference's comment ranges. Fixes are delegated in isolated
+worktrees. Then run the merged parity gate and finish whole-product review before
+freezing another package or starting paid confirmation. Passing the exercised
+fixtures was insufficient; the earlier paid start was premature.
+
+`installed-jg-reference-parity-v1` is stopped and superseded. Requests and
+scikit-learn officially solved but exceeded baseline cost; Django was interrupted
+with its patch and traces retained; seven cells never started. No paid run is
+active. The goal remains active and independent deterministic work can continue.
The changed-policy cohort `installed-jg-final-cohort-v1` was stopped and marked
superseded, retaining completed outcomes and interrupted traces. It is not a
@@ -46,12 +47,12 @@ verification before ending each pass. These commands are targets to implement,
not claims about the current scaffold.
- [x] [01 — Reference and real HTTP fixture](slices/01-reference.md)
-- [x] [02 — Bundled parser parity](slices/02-parser.md)
+- [ ] [02 — Bundled parser parity](slices/02-parser.md)
- [x] [03 — Filesystem eligibility and snapshots](slices/03-filesystem.md)
-- [x] [04 — Installed CLI + skill + first real task](slices/04-checkpoint.md)
+- [ ] [04 — Installed CLI + skill + first real task](slices/04-checkpoint.md)
- [x] [05 — Faults, cancellation and partial results](slices/05-failures.md)
- [x] [06 — Default-on fresh cache](slices/06-cache.md)
-- [x] [07 — Preserved output policy](slices/07-output-policy.md)
+- [ ] [07 — Preserved output policy](slices/07-output-policy.md)
- [ ] [08 — Frozen quality and release verification](slices/08-release.md)
## Product and scope
diff --git a/specs/jevgrep/assets/parity-restoration.md b/specs/jevgrep/assets/parity-restoration.md
index 8ab355a..54bd87c 100644
--- a/specs/jevgrep/assets/parity-restoration.md
+++ b/specs/jevgrep/assets/parity-restoration.md
@@ -92,13 +92,15 @@ policy. Treating the selection ceiling as permission to classify text fragments
instead was another unvalidated behavioral change; it can multiply requests and
remove the class anchor that reconsiders skipped directories.
-## Restored live confirmation
+## Superseded restored confirmation
-The new frozen study is
-`evals/runs/swebench/installed-jg-reference-parity-v1/plan.json`. All ten no-call
-validations passed before paid execution. Requests ran first; the remaining
-cells run sequentially under the same package and prompt. Completed attempts are
-retained, never rerun by continuation.
+The study at `evals/runs/swebench/installed-jg-reference-parity-v1/plan.json` is
+stopped and marked superseded. All ten no-call validations passed, but the later
+whole-product review reproduced two mismatches missed by the exercised parity
+fixtures. Starting paid confirmation before that review finished was premature.
+Requests and scikit-learn completed; Django was interrupted with its patch,
+native sessions and Jev traces retained; seven cells never started. None of these
+results establishes release acceptance or authorizes continuing this study.
Requests officially resolved. Full Sol cost was **$0.3331174** across 11 billed
generations versus **$0.2685004** baseline: a solve, but not a cost win. Observed
@@ -123,6 +125,24 @@ strictly contains the other. These observations identify downstream work; they
do not establish that file-list differences caused it. The saved generation
receipts fully account for the $0.064617 increase over baseline.
-The quality gate remains every baseline solve preserved and at least seven
-solved strict Sol-cost wins in this single frozen ten-task cohort. One restored
-solve and deterministic parity do not satisfy it.
+Scikit-learn officially solved at Sol **$0.3426444** versus **$0.2944360** baseline,
+also not a cost win. Jev observed cost was **$0.071465982**, separately reported.
+The accepted spike already lost this pair on cost ($0.3188932). The restored
+agent recovered from a patch-context failure and a rejected test command before
+running its full module once; those steps were downstream of retrieval. A
+passing official grade does not establish all behavioral edge cases equivalent.
+
+## Remaining review gate
+
+The whole-product review reproduced two ordinary preservation defects. The
+reference builds second-pass `selectedEvidence` in source-selection completion
+order; the port used admission order. Delaying one file reverses meaningful
+array order despite identical model scores. Separately, TS/JS syntax fallback
+returned before collecting comments, so a selected fragment inside a block
+comment ended inside it instead of expanding to the reference's full comment.
+Both require whole-CLI regressions and fixes before another frozen candidate.
+
+Finish whole-product review before paid confirmation, in addition to matched
+HTTP tests and supported-runtime gates. The quality target remains every fixed
+baseline solve preserved and at least seven solved strict Sol-cost wins in one
+prospective ten-task cohort. Historical or interrupted attempts cannot be pooled.
diff --git a/specs/jevgrep/slices/02-parser.md b/specs/jevgrep/slices/02-parser.md
index ce28223..d242f04 100644
--- a/specs/jevgrep/slices/02-parser.md
+++ b/specs/jevgrep/slices/02-parser.md
@@ -1,6 +1,7 @@
# Bundle source inspection without Python
-Status: verified against the frozen conformance cases (2026-09-26). Depends on: 01.
+Status: reopened for TS/JS syntax-fallback comment preservation (2026-09-26).
+Depends on: 01.
Read [contracts](../contracts.md) first. Recognized syntax outside the reference's
Python 3.11 boundary shares one fallback across inspection, previews and
neighboring-method selection. TypeScript uses the reference comment traversal.
diff --git a/specs/jevgrep/slices/04-checkpoint.md b/specs/jevgrep/slices/04-checkpoint.md
index 1d07d7d..1b8be2f 100644
--- a/specs/jevgrep/slices/04-checkpoint.md
+++ b/specs/jevgrep/slices/04-checkpoint.md
@@ -1,6 +1,7 @@
# Deliver the installed CLI and agent workflow
-Status: verified on the restored package (2026-09-26). Depends on: 01, 02, 03.
+Status: installed workflow verified; full-pipeline preservation reopened for
+selection-completion order (2026-09-26). Depends on: 01, 02, 03.
Read [contracts](../contracts.md) first.
## Contract and owner
diff --git a/specs/jevgrep/slices/07-output-policy.md b/specs/jevgrep/slices/07-output-policy.md
index 2a751f9..d0aa17f 100644
--- a/specs/jevgrep/slices/07-output-policy.md
+++ b/specs/jevgrep/slices/07-output-policy.md
@@ -1,6 +1,7 @@
# Preserve the accepted retrieval packet
-Status: deterministic parity verified (2026-09-26); live quality belongs to 08.
+Status: reopened after whole-product review reproduced preservation defects
+(2026-09-26); live quality belongs to 08.
Depends on: 04–06.
The default is the accepted spike's uncapped source packet: ranked file locations,