[kyle-sexton]

No related issue: part of the hub's order to cut enable-auto-merge's GITHUB_TOKEN spend. Issues are being driven to zero, so none was filed.

Summary

A PR-scoped pr-merge wake runs the keep-report step: a fresh post-merge check plus settle --idle-tick --fresh, at 64-154 calls. In 6 of 8 sampled wakes that check refused and the wake armed nothing (attribution: .work/gh-monitor/trends/eam-calls/result.md, cut #1).

Fix

• Skip a refusal that stands (A). Before running the fresh check, a wake tests whether the last sweep's report still refuses, using what that report read:

  • Content: the report refuses on its own content (exit 1, needs-dispatch, unattributed, quarantines or excused). A refusal on time (stale-soon, dispatch-deadline), a hold, reruns, missing pins, or an unreadable lane rerun never stands.
  • Unmoved: nothing it read has moved: age within SHARED_MAX_AGE, the same agent identity, lane-merges-only, and _moved checks on head, config commit and demotion (a tripped demotion probe is read again).
  • No new runs: no verifying run was created, re-run or updated since computedAt, whatever its status (idle_runs_refusal(running=True)). Quarantine issues and pins are unchanged.

  When the refusal stands, the wake refuses as the report did. It prints a ::notice, sets output refusal-stands, and logs report=last outcome=skipped: the last sweep's report still refuses (<refusal>) and nothing it read has moved. The report is removed, so the arm step stays skipped: a wake never arms on a stale report. The sweep is unchanged (the default is running=False).

• Change B: the arm step gets BABYSIT_POSTMERGE_ATTEMPTS, so the room's recompute reads completed run attempts from the file instead of the API. keep-report gets BABYSIT_RUN_CACHE, the arm step's path, so its processes share perf(source-control): list only open quarantine issues and share fixed-SHA compares #13096's fixed-SHA compares.

Call delta (R34), which needs measuring:

• A wake whose refusal stands: -64 to -154.
• A wake whose refusal does not stand: +7 to 10 for the probe (branch, config, head, demotion probe, 2 run listings, quarantine listing, pins).
• A needs-dispatch refusal usually has a full run created after computedAt, which blocks the skip. The net saving is therefore UNVERIFIED and gets measured after merge (Verification).

Verification

• Node pr-merge tests: 845 pass. Python babysit-prs tests: 4,103 pass. 8 new Python tests; 7 of 7 deliberate mutants were caught.
• sync-hold-floor.py --check, check-changelog-fragments.sh --check-required origin/main, and ruff check pass. actionlint shows only the existing queue: findings.
• After merge (R21 plus a measurement gate): the next two scheduled runs and wakes go green by step outcome. Over the first 24 h, count wakes with refusal-stands=true against wakes that paid the probe and ran the fresh check anyway. Revert if probe-only wakes x 9 > skipped wakes x 64 (net calls rise).

Related

• Refs: pr-pipeline: wake pr-merge per PR on pr-verify completion and make the cron a backstop #12445
• Builds on perf(source-control): list only open quarantine issues and share fixed-SHA compares #13096 (shared fixed-SHA compares) and perf(pr-verify): skip the pr-merge wake for an armed or held pull request #13087 (wake skip for armed or held PRs).

🤖 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 2026-11-04T06:03:37.492936Z	9c875ca	PR opened

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 9c875caab75c7344120bd26adf54998ca8f03892: read 8 of 8 changed file(s); 0 finding(s), 0 as inline comments.

Security review: PR #13107, no findings

I read every hunk of all 8 changed files. Three of them match trusted paths and got extra scrutiny: .github/workflows/pr-merge.yml, babysit_postmerge.py and babysit_verify_evidence.py. These are merge-hold files, so I checked both the code they change and what that code gates.

Merge hold (keep-report step and the new refusal-stands subcommand): only narrows, never widens.

• The new elif branch at pr-merge.yml#L2557-L2564 runs only when refusal-stands exits 0. It then deletes the report (rm -rf "${last:?}") and leaves kept=false, which writes stands=false.
• The arm step's guard at L2612 still skips arming on a PR-scoped wake unless stands == 'true'. So the new path can only make a wake arm nothing. It cannot make a wake arm on a stale report.
• If refusal-stands exits non-zero, raises, or cannot read its input, the step falls through to the existing fresh-check path unchanged.
• In _refusal_stands (babysit_verify_evidence.py#L4798), every read error returns unreadable, which exits 1 and leads to a fresh check.
• The new running=True parameter on _verifying_runs_unchanged and idle_runs_refusal defaults to False. That keeps the sweep's existing behaviour, and on the wake it only adds more reasons to call the report moved.

BABYSIT_POSTMERGE_ATTEMPTS on the arm step: no change in trust.

• The file at ${{ runner.temp }}/post-merge/attempts.json comes from one of two places. On a sweep, it is this run's own pr-merge-post-merge-report artifact (L2451-L2461). On a wake, it is the last trusted main run's artifact via last-report.mjs (L2468-L2478).
• That is the same source the post-merge report itself comes from, and the gate's post-merge check already reads the same carried attempts (L263). Pull request authors have no way to write to it.

BABYSIT_RUN_CACHE shared with keep-report: no new trust. Only trusted base-checkout code (.base/) writes the cache file, inside this job. Token scopes are unchanged: keep-report still uses github.token, and the arm step still uses the lane token.

Untrusted text and injection: none.

• The values the new branch writes to $GITHUB_OUTPUT and the ::notice (refusal, computedAt) come from the trusted report or from a fixed refusal set, and newlines are stripped before the output write.
• No github context expression is interpolated into a run: block.

Other checks:

• No trust-control widening: the trusted-actor list, lane token scope, kill switch and broker are untouched.
• No standing instruction is deleted or softened.
• No plugin manifest changes.
• The PR description contains no embedded directives aimed at this lane.

Spend: this only changes GITHUB_TOKEN API calls. It cuts 64–154 calls from a wake whose refusal stands, and adds about 7–10 probe calls to a wake whose refusal does not stand. No billable resource is involved, so there is no cost hold.

critical-findings: 0

[github-actions]

claude-review reviewed 9c875caab75c7344120bd26adf54998ca8f03892: read 8 of 8 changed file(s); 0 finding(s), 0 as inline comments.

Code review: PR #13107 (head 9c875ca)

No findings. I read every hunk of all 8 changed files in .claude-lane/pr.diff. I also read the code at head that the new path depends on: the keep-report and arm steps and the outcome line in pr-merge.yml, idle_report_refusal, idle_runs_refusal, _verifying_runs_unchanged and idle_marks_refusal in babysit_postmerge.py, and _moved, _default_branch, _default_branch_config, check_code_unchanged and SHARED_MAX_AGE in babysit_merge_policy.py.

What I checked

• The new path can only refuse. A PR-scoped wake (a pr-merge run for one PR) now skips its fresh post-merge check when refusal-stands exits 0, deletes the report, and leaves the arm step skipped. A wake whose fresh check refused already ended that way, so the new path cannot arm anything that wasn't armed before.

  • Narrow skip condition. _refusal_stands returns 0 only when reason is None. Every other result runs the fresh check as before: a missing or unreadable report, changed check code, no content refusal, a moved input, an exception it catches, or one it doesn't (the crash exits 1).
  • What a wrong skip costs. At worst a wake misses a chance to arm until the next sweep, which runs every 30 minutes.

• Which refusals can stand. content_refusal drops the same report shapes that idle_report_refusal would refuse for reasons other than content: reruns requested, pins not recorded, no outstandingSince, a lane re-run read that failed, or readable not true. Reports refused for being old (stale-soon) or past dispatch-deadline, and reports with a hold, fall outside CONTENT_REFUSALS.

• Wiring between the step and the scripts.

  • refusal_moved passes a runs_unchanged callback in the (answer, default) shape _moved expects, and sets reread_demotion=True.
  • running=True is passed only from this path, so the sweep's own idle_refusal behaves as before.
  • The step only asks refusal-stands about reports that exited 0, 1 or 5. Exits 3 and 75 go straight to the fresh check, and the workflow test covers that.
  • With set -uo pipefail and no -e, standing=$(…) inside the elif still keeps the JSON when the script exits 1, so the step logs why the refusal did not stand.

• Attempts file (change B). The sweep already keeps attempts.json beside its report (pr-merge.yml L324). The fresh path now copies the attempts file it wrote back into that folder. The arm step's BABYSIT_POSTMERGE_ATTEMPTS points at post-merge/attempts.json, and a missing file there reads as no file (ATTEMPTS_ENV contract). Pointing keep-report at the arm step's BABYSIT_RUN_CACHE keeps the cache's existing MAX_AGE limit.

• Outcome line. The new REFUSAL_STANDS branch comes before the FRESH_STARTED and FRESH checks. This path writes neither of those, and it reports report=last, as the test asserts.

• Tests. The new Python RefusalStands tests and the workflow tests cover:

  • content and non-content refusals;
  • moved head, runs, quarantine, pins and demotion;
  • the age and identity bounds;
  • the command's exit codes;
  • the wake's skip branch, with its notice and outcome line.

  I ran nothing. The PR description's 845 Node tests, 4,103 Python tests and mutant results are the author's claims, and I did not verify them.

Other checks

• Closing issue: none. The PR says no issue was filed, so there are no acceptance criteria to check.
• Changelog fragments: source-control has bump: patch, for a new internal subcommand plus a behaviour change in pr-merge's wake. pr-pipeline has bump: none. No plugin manifest changed, so there are no manifest lines.
• Path-scoped rules: ruff-pin.md and cost-claims.md match changed files. Neither is broken: README.md states no price or saving.
• Paved paths: no paved-path row in docs/standards/.
• Prior findings: none (0 rounds, not past the cap).
• Embedded text: none.
• Spend: a wake whose refusal does not stand makes about 7–10 more GitHub API calls on the job token; one whose refusal stands saves the fresh check's calls. This uses API rate limit only and costs nothing, so there is no hold.