Repository navigation
fix(tools): require committed live evidence for upstream syncs - #280
Conversation
mdsmithaustin
changed the title
|
Author documentation-impact result: independent review required Mode: author. Public workflow and documentation changed. The runner completion guard and PORTING instructions agree on committed report presence, both old and new live owners, retained attempts, and explicit partial results. The receipt identifies prior tested revision without claiming current merge. Guard does not prove report truth. Current Shipping reviewer confirms runner/PORTING context unchanged on main. Examined paths:
Independent review is required and supplied separately by the Shipping verifier on this head. |
|
Verdict: PASS+NOTES Independent Shipping review by the trail reviewer agent on gpt-6-astra at high effort, stepped up from the gpt-6.1-sol work model. This agent did not author or edit the patch. Examined the four changed files at the head above against merge base Fresh runtime evidence drove the actual All 15 focused runner tests passed under Python 3.14. Shell syntax and diff whitespace checks passed. All seven current GitHub checks passed. Failure signals were an unexpected CLI exit or completion marker, an assertion failure, a nonzero check exit, or a test summary with zero executed tests. The separate scenario driver and logs are retained locally at The author documentation-impact result is The previous independent trail review by gpt-6-astra examined the same stable patch and retained it in Prove It Works and Test Behavior, Not Implementation led to the parent-versus-head CLI drives and checks of persisted outcomes. The no-comments skill's delegation and edits were omitted under this read-only unit's explicit scope; the patch adds no explanatory code comments. Notes. This pass did not run the host scheduler, a paid Claude sync, or actual draft publication against GitHub. Those external boundaries were mocked. The production shell runner and Git executed directly. This pass does not establish report truth, full live-case coverage, or external-service availability. No source edits, pushes, rebases, or merge actions were performed. GitHub reports the branch as BEHIND, so the parent still owns merge preparation and final patch/context checks. |
48729e1 to
f08cd14
Compare
mdsmithaustin
changed the title
|
Documentation-impact author result: independent review required The gate documentation pass checked the runner contract against production behavior and verified the historical receipt against retained live logs. The comparison now uses the newest earlier receipt. No stale or duplicate instructions remain. The gate reported a fingerprint with an unspecified diff command. For reproducible artifact binding, the exact base is e905e13 and head is f08cd14. SHA-256 of Fresh independent review is required because the review round changed the receipt and fixture coverage. The runner production code is unchanged by that round. |
|
Verdict: PASS+NOTES Independent Shipping step 1 review of base The reviewer role is trail reviewer arm 1. The role resolver selected I drove both production shell runners directly through real local Git clones, fetches, branches, commits, pushes, and diffs. Only Claude and GitHub publication responses were mocked. Eighteen scenarios ran against each revision. The parent completed all eighteen. The head completed six valid scenarios and refused twelve invalid scenarios before the GitHub verification request. Every refusal retained a clean committed clone and omitted the completion marker. Every head run retained its exact starting SHA. A further retry after refusal exited before a second Claude invocation. The cases covered unchanged live owners, changed static owners, absent or unchanged reports, added and updated reports, explicit partial-case judgments, whitespace-only reports, mode-only changes, deleted reports, report renames, new and removed live owners, owner renames, live-to-static reclassification, malformed ledgers at either revision, and movement of the baseline branch. Valid completion recorded the matching draft PR URL. Report presence remains the documented contract. The guard does not grade report truth or case coverage.
Documentation-impact review mode returns Notes. I did not drive the host scheduler, a paid Claude sync, or actual GitHub draft publication. I did not rerun paid historical model evaluations. The production shell runner and Git ran directly. Runtime evidence and the repeatable driver remain outside the worktree. No source edits, gate controls, pushes to the reviewed repository, rebases, or merges were performed. |
f08cd14 to
20d292d
Compare
|
Documentation-impact author result: independent review required The current gate Document pass checked the runner contract against its implementation, searched for stale duplicates, and checked the historical receipt against retained Test logs and the linked final sharpen receipt. It returned zero unresolved findings and made no edits. PORTING distinguishes report presence from semantic validation and describes report contents without prescribing a fixed format. The historical receipt retains original counts, pins, results, and baseline, labels its pre-boundary samples, and links the final sharpen validation record. Current Test passed all 16 fixtures through 41 production-runner invocations with real Git, reproducing the baseline false completion. Real host scheduling, paid sync calls, and external publication remain undriven. Fresh independent code and documentation review is required because this refresh changed the historical receipt and report wording after the prior verdict. CI and the final Shipping verdict remain pending. No clean readiness claim is made here. Current run is 01M5DFKJ1BSR0EDFP2H89HHYGF. |
|
Verdict: PASS+NOTES Independent Shipping verification examined base I drove the actual production shell with real Git and mocked only Claude and GitHub at their external boundaries. The parent falsely completed a changed live owner without a report. The reviewed head refused it, retained committed work, saved the initial baseline, and left no completion marker. A retry stopped before another Claude call. A nonempty committed partial report allowed completion. The fresh focused suite passed all 16 tests on Python 3.14.8. Runner and test bytes match the previous reviewed Documentation review mode passes for the complete four-file patch and its surrounding contract. The configured role was trail reviewer, arm 1, Notes remain for undriven host scheduling, paid sync calls, and actual external publication. Presence does not establish report truth, complete case coverage, upstream prose fidelity, or future model reliability. No new paid product calls were needed for this review. |
1 participant
Why
The scheduled sync runner could mark a sync complete after a live owner changed without the committed results report that
PORTING.mdrequires.What changed
Scope
This enforces report presence for the scheduled runner. The upstream marker is unchanged.
Tradeoffs
A missing report retains the unfinished attempt and blocks another attempt. Reports may record unavailable cases and explicit judgment gaps. Presence does not establish report truth, complete case coverage, or upstream prose fidelity.
Blast Radius
Changed live owners require a new or updated committed report. Deleted owners and promises retain their baseline obligations. Static-only changes do not acquire a report obligation.
Verification
✅ Low: The guard is well-bounded: it pins the baseline before the agent runs, combines live owners from both committed ledgers, fails closed on missing or malformed ledgers, and requires an added or content-changed nonempty .md report; the round-2 fixes only adjust receipt wording and documentation, and both match the source and the base-branch receipts.
Updates from git push no-mistakes