feat(triage): stamp current prompt provenance as v3 - #2787
Conversation
Codex Review SummaryThis comment shows the latest Codex review activity on this pull request.
ℹ️ About Codex in GitHubYour team has set up Codex to review pull requests in this repo. Reviews are triggered when you
Codex reacts with 👀 while any review is running, comments if it has suggestions, and reacts with 👍 once all reviews finish with no findings. |
|
Coordinator cleanup: author/reviewer released head 8e7fd25, tracked tree was clean and remote matched. Ignored inventory contained only disposable backend bin/obj, frontend dist/node_modules, and Python pycache; no evidence/runtime files required copying. Worktree removed normally without force. Full-suite counts and direct exit-status evidence remain in the PR; hosted packaged/live-provider acceptance is still unverified. |
|
Control-plane review correction (2026-09-08): I merged PR #2787 at 0cebd93 after independent Terra review and green hosted checks, but missed the ADR-0066 amendment's maintainer-review requirement. The September 6 SC-10 delegation named twelve PRs and did not include this one. This was my error; no acknowledgement or post-hoc approval is inferred. The change includes the Windows archive acceptance script and tests alongside the atomic prompt-v3 producer/frontend/docs update. Validation remains recorded on the PR; a live packaged Windows/OpenAI run was not claimed. Please review #2787 post hoc. I have preserved the merged work and recorded the open checkpoint in OUTSTANDING_TASKS.md via PR #2791. New CI-control PRs #2791 and the #2335 test slice will complete independent review and exact-head hosted checks, then remain open pending your review. No settings, release, or selective-execution approval is inferred. |
Fresh-context review of this PR found I had searched the wrong issue. The continuation train's tracker is CI-12 #2336, not #2337 or #2324, and #2336 records the session as maintainer-requested and, at 02:40:20Z on 2026-09-10, "the maintainer's 2026-09-10 request to finish implementation, merging, testing and CI improvements". So "no delegation is recorded" was wrong. The row now asks whether that request supplied the per-PR review the amendment wants, and infers nothing either way. Three further corrections the same review prompted: The divergence did not start 2026-09-09. J.1 (#2772) and J.2 (#2787) are both 2026-09-08, and 2026-09-09 is the day the rule was honoured, with the four-PR delegation. The pointer file said the opposite. The row omitted the train's quality record, which changes how question (a) reads. Every PR carried a Codex and an independent review with current-base full required qualification. One confirmed HIGH, a trusted verifier bypassed through a Windows directory junction, was refused at 09:44Z, fixed in cdc6d8e at 09:49:14Z and is an ancestor of #2871's 10:23:22Z merge; export.test.mjs passes 7/7 on current main. The gap is procedural, not a shipped known defect. #2878 is added: it is the eighth merge in the same train. #2868's ci-continuation-observe.yml and #2791's nightly observer are different PRs under different trackers, and #2791 was delegated on 2026-09-09. An unqualified ruling about "the observer" would hit the wrong one, so the row says which is which. Also: name J.2 as where a parked control-plane PR is recorded now that SC-10 is closed; stop hard-coding live connector status in a file whose own preamble says live GitHub outranks it, and restore the standing "a usage-limit notice is informational, not a finding" triage line; correct the CF subset sentence, which spans v0.4 as well; bump Last reviewed. Refs #2337, #2336.
Incremental edits left the packet contradicting itself, which is the mis-framing failure a review already caught in this row once. The title still named only the continuation train after four non-train PRs were added. Question (a) still asked for review of "these eight merged changes" when the row lists twelve. The arithmetic paragraph still read eleven PRs, eighteen total, thirteen awaiting. Correct figures: this row covers twelve; J.1 (#2772) and J.2 (#2787) are one each, so fourteen await a reply across the three rows; five more were acknowledged in the 2026-09-06 q-2 = A ruling, giving nineteen control-plane merges outside the per-PR review since the amendment was adopted on 2026-09-03. The train is now introduced as eight of the twelve rather than as the whole subject.
Codex P2 on #2919, and correct on the operational point: a coordinator reading this row as 'Closed 2026-09-06' could skip open human actions. SC-10's row is ticked, but section J.1 and J.2 are open rows carrying the #2772 and #2787 post-hoc merges, J.2 also holds parked #2838, and J.3 carries the 2026-09-08 to 2026-09-10 disclosure. The row now names all three and says the tick is not the end of it. The other half of the finding is a cross-PR artifact rather than a defect: J.3 and J.2's #2838 entry live in #2916, which is unmerged, so a review against main sees references to sections that do not exist yet. That is an ordering constraint, not a text problem - #2916 merges first.
1 participant
Problem
The capture-date prompt semantics changed, while provider-backed triage records still used
llm-triage.v2. That made current and historical provenance indistinguishable.Change
llm-triage.v3.scripts/ci, where its local harness is importable.Validation
dotnet test backend/Taskdeck.sln -c Release -m:1(green: 8,973 passed; 5 expected skips)npm run typecheck,npm run build, andnpx vitest run --maxWorkers=2withVITE_API_BASE_URLremoved (379 files, 6,119 passed, 3 skipped)Not verified
Refs #2211