[kyle-sexton]

No related issue: a merge-gate defect found while measuring the pr-merge wakes for #12445; open-issue intake is closed while issues are driven to zero.

Summary

Six lanes App PRs (#12814, #12812, #12795, #12665, #12636, #12604) can't be armed. The merge gate refuses each with evidence-unbound:claude-review: not bound: rerun-not-allowed (and the same for the security lane). That refusal is 7 of the 25 pr-merge wakes since #12771. In each review run, attempt 1 failed while the reviews credential was rejected at its usage limit. For example, run 41626484307 attempt 1 failed at 11:31Z with claude-review: failure-class=timeout: the review leg reached its time limit and the warning REVIEWS_CLAUDE_CODE_OAUTH_TOKEN ... status=rejected. Attempt 2 then passed. timeout was not a class after which attempt 2 binds (babysit_review_evidence.py RERUN_CLASSES, and the no-outcome branch of rerun_binds), so the passing review never counted.

Fix

• NO_OUTCOME_RE now also matches this lane's failure-class=timeout: line. A timed-out attempt 1 binds attempt 2 only under the proof no-outcome already uses: no completion record, and the lane posted nothing on the head between the two attempts (_unposted). Attempt 1 then left no verdict, so attempt 2 is the only review and binding it trades nothing away.
• The gate still never starts that re-run itself (TIMEOUT is not in GATE_RERUN_CLASSES). A spent credential would fail the one allowed attempt.
• reference/safety.md names the new class.

Decision: GitHub Support (hub ledger L108), relayed in session: "go with (a); you own it ... Keep the same unposted proof NO_OUTCOME uses: an unposted attempt 1 leaves attempt 2 as the only review, so binding it adds no risk."

Verification

• test_babysit_review_evidence.py: 268 passed. The new TimeoutRerunTests use run 41626484307's annotation shape (the timeout line plus the status=rejected usage warning). They cover: binds for both lanes when nothing was posted; refuses with rerun-not-allowed when the lane posted its marker; and the gate does not start the re-run.
• With the source change reverted, 3 of those tests fail (the bind test on both lanes, and the ev.TIMEOUT check).
• Both lanes write the same line: claude-security-review: failure-class=timeout: the review leg reached its time limit (check run 125891481218, attempt 1 of run 41626483992).
• scripts/run-ruff.sh check on both files: all checks passed.
• Per-run call delta: 0. No new reads; the change touches only the classification of annotations already fetched.

Related

• pr-pipeline: wake pr-merge per PR on pr-verify completion and make the cron a backstop #12445 (wake measurement that surfaced the refusals)
• fix(source-control): bind a full lane re-run after a no-outcome attempt the lane never posted #12097 (the no-outcome binding this extends)

🤖 Generated with Claude Code

[chatgpt-codex-connector]

Codex Review Summary

This comment shows the latest Codex review activity on this pull request.

Review	Status	Commit	Review trigger
📝 Code Review	✅ Completed 2026-11-03T16:31:14.077901Z	0235cc2	Draft marked ready

ℹ️ About Codex in GitHub

Your team has set up Codex to review pull requests in this repo. Reviews are triggered when you

• Open a pull request for review
• Mark a draft as ready
• Comment "@codex review" or "@codex security review".

Codex reacts with 👀 while any review is running, comments if it has suggestions, and reacts with 👍 once all reviews finish with no findings.

[github-actions]

claude-security-review reviewed 0235cc2ac856f7d5492924ae043e8ba6d0ff4055: read 4 of 4 changed file(s); 0 finding(s), 0 as inline comments.

Security review: PR #12863

No security issues found.

This PR changes how the merge gate reads review evidence. When an AI review's first attempt fails with failure-class=timeout, the gate now counts the second attempt as the review. Before, only no-outcome got that treatment. I read every hunk of all 4 changed files. One of them, plugins/source-control/skills/babysit-prs/scripts/babysit_review_evidence.py, is on a trusted path, so I read its whole change and the code it gates (rerun_binds, _unposted, lane_posted, _lane_job, _failures, _lane_line, RERUN_CLASSES, GATE_RERUN_CLASSES).

Merge-hold check. This loosens the gate's evidence binding, so I judged it under the merge-hold bullet:

• Same checks as no-outcome. The timeout line goes through the existing NO_OUTCOME_RE branch of rerun_binds (L1182-L1186). It binds only when all of these hold:
  • the failure annotation is on this lane's own job (_lane_job) and names this lane (_lane_line);
  • the job left no completion record (not recorded(job));
  • the lane posted nothing on the head between the two attempts (_unposted → lane_posted). That covers marker creates and rewrites, lane comments and lane threads, and any input it can't read counts as posted.
• No verdict is lost. A first attempt that timed out and posted nothing gave no verdict. The second attempt still has to pass as a full review of the same head on its own merits.
• The gate doesn't start the re-run. TIMEOUT is not added to GATE_RERUN_CLASSES (L1143-L1145). The re-run still comes from the reconciler or a person.
• No new way to fake the signal. It is read from the same lane-job failure annotations the existing classes use. A head that edits a caller workflow to forge them is still caught by lane_source, which the diff doesn't touch.

No token, secret, trusted actor, untrusted-text path or billing surface changes, so none of the three severity-bar impacts (leak, compromise, charge) applies.

Other files:

• reference/safety.md and the .changes entry are prose that matches the code.
• The test file adds tests only.
• No standing instruction is deleted or softened.

Spend: none. The change only reclassifies annotations the gate already fetches.

critical-findings: 0

[github-actions]

claude-review reviewed 0235cc2ac856f7d5492924ae043e8ba6d0ff4055: read 4 of 4 changed file(s); 0 finding(s), 0 as inline comments.

Code review: PR #12863, bind a review re-run after an unposted timeout

No findings. I read all 4 changed files in the diff: the changelog fragment, reference/safety.md, babysit_review_evidence.py (a trusted path, so it got extra scrutiny) and its tests.

What the change does. NO_OUTCOME_RE now also matches <lane>: failure-class=timeout: . Before a timed-out attempt 1 lets attempt 2 count, rerun_binds still requires the same three checks no-outcome uses (babysit_review_evidence.py#L1182-L1186):

• the line must start the lane's own failure annotation (_lane_line), so one lane's line cannot vouch for another;
• the job left no completion record (recorded);
• the lane posted nothing on the head between the two attempts (_unposted), and when the inputs are unknown this reads as posted.

TIMEOUT is not added to RERUN_CLASSES or GATE_RERUN_CLASSES. So the merge gate still never starts this re-run itself, and a timeout cannot bypass the posted check. With that check in place, a timed-out attempt 1 that did leave a verdict cannot have it replaced.

Tests.

• TimeoutRerunTests passes its own annotations to no_outcome(...), so the fixture contains only the timeout line and the usage warning, with no no-outcome line.
• The bind test therefore runs the new regex branch. It covers both lanes, the posted-marker refusal (rerun-not-allowed) and the not-a-gate-rerun check.
• The PR says 268 tests passed and that 3 fail with the source change reverted. That is the author's claim: this lane has no shell, so I did not run the tests.

Docs. safety.md and the changelog fragment match the code. A patch bump fits a fix to which evidence the gate accepts.

Optional, not counted. Two untouched comments still name only NO_OUTCOME for this path: the rerun_binds docstring (L1166-L1170) and the RERUN_CLASSES / GATE_RERUN_CLASSES comments (L1129-L1142). The regex name NO_OUTCOME_RE now also covers timeouts. Mentioning TIMEOUT in those places would keep the comments accurate.

Other checks.

• Closing issue: none; the PR says it has no related issue.
• Plugin manifests: none changed.
• Prior findings: none (round 0 of a cap of 3).
• Embedded instructions: none found in the PR context.
• Spend: no extra usage. The change only reclassifies annotations the gate already fetches.