Skip to content

fix(source-control): rerun a publish rate-limit review failure after GitHub's hourly reset, not five hours #18037

Description

@kyle-sexton

Steps to reproduce

  1. A review lane's Publish Review job fails with claude-review: failure-class=publish-failed: posting the review failed (API rate limit exceeded for installation ...). Example: PR fix(source-control): cap pr-verify repeated FAILs only at a FAILed head #18028, runs 44942450656 and 44942450584, failed 2026-11-21T19:47Z when the lanes App installation pool ran out.
  2. The babysit merge gate is the only automation that reruns it (maintenance-reconcile-lost-runs skips review lanes, reconcile.mjs:155-160).

Expected behavior

The gate reruns a publish-rate-limit failure once the installation pool has reset. GitHub's installation REST pool resets every hour (the x-ratelimit-reset header the error came from), so the rerun is due within about an hour.

Actual behavior

plugins/source-control/skills/babysit-prs/reference/safety.md:877-879 applies one wait to every rate-limit class: "for a rate limit, its job finished at least five hours ago ... Five hours is the plan's session-limit window". That window fits a Claude plan limit (rate-limit class), not GitHub's hourly installation pool (publish-rate-limit class). A PR whose review failed to post therefore waits 5 hours for its lane rerun, past the 2-hour PR age target; #18028 needed a manual rerun at 20:25Z.

Severity

Medium.

Severity justification

No data loss or security effect, but every installation-pool exhaustion holds each affected PR about 4 hours longer than needed; the pool ran out three times on 2026-11-21.

Suggested fix location

  • plugins/source-control/skills/babysit-prs/scripts/babysit_review_evidence.py near GATE_RERUN_CLASSES (:1309) and the publish-rate-limit matcher (:1300): give publish-rate-limit its own wait, 1 hour (or the job's finish time past the next hourly reset), keeping rate-limit at 5 hours.
  • plugins/source-control/skills/babysit-prs/reference/safety.md:877-881: state the two waits and why.
  • The one-rerun-per-check-per-head and first-attempt-only rules stay.

Environment

claude-code-plugins main at 2026-11-21; lanes App installation token pool.

Activity

  1. added
    needs-triageNot yet classified. Floor until a type and one priority tier are set.
    on Nov 21, 2026
  2. github-actions commented on Nov 21, 2026

    @github-actions
    Contributor

    Size: Small
    Complexity: Low
    Uncertainty: Known
    Flags: none
    Blocked by: none
    Blast radius: the babysit-prs merge gate's rerun wait (babysit_merge.py), its tests, and safety.md; read by the babysit merge gate and its maintainers.
    Reversible: yes; revert the change and the single five-hour wait returns.
    Open questions: none
    Understood: a publish-rate-limit review failure (GitHub's hourly installation pool) should be rerun after about one hour, while a rate-limit failure (plan session window) keeps the five-hour wait. The refinement comment's acceptance criteria give an exact done-when, including the mixed-class and attempt-above-one cases. No duplicate or merged fix found among the open issues.

    Automated triage. A maintainer reviews it.

  3. removed
    needs-triageNot yet classified. Floor until a type and one priority tier are set.
    on Nov 21, 2026
  4. melodic-automation-lanes commented on Nov 21, 2026

    @melodic-automation-lanes
    Contributor

    Not admitted to the issue lane yet: no trusted-human acceptance criteria: a listed human must put the confirmed - [ ] criteria in the body or post them as their own comment, or, on an issue a listed human filed, intake's newest refinement comment must hold them unedited; intake's refinement of this body is due and may supply them.

  5. melodic-automation-lanes commented on Nov 21, 2026

    @melodic-automation-lanes
    Contributor

    Refinement. A read-only run explored this repository, searched its issues and pull requests, and read the allowed docs to draft what follows. On an issue a listed human filed, it counts as the issue's spec once posted, while unedited; it is still model text, so check it.

    Intended outcome: The babysit merge gate reruns a publish-rate-limit review failure one hour after its job finished, keeps the five-hour wait for rate-limit, and safety.md states both waits.

    Acceptance criteria

    • babysit_merge.py defines a separate one-hour delay for class publish-rate-limit next to RATE_LIMIT_RERUN_DELAY (:345), which stays timedelta(hours=5) for class rate-limit
    • The due-time computation at babysit_merge.py:461-463 uses the one-hour delay when the classes are only publish-rate-limit, and zero for NO_WAIT_RERUN_CLASSES as now
    • When a check carries both rate-limit and publish-rate-limit, or either beside a no-wait class, the gate waits the longest delay that applies to the classes present
    • The waiting reason string at babysit_merge.py:466 no longer says only 'usage limit' for a publish-rate-limit wait, and still ends with the re-run due time
    • A test in scripts/tests/test_babysit_merge_rate_limit_rerun.py shows a publish-rate-limit failure finished 61 minutes ago is rerun, and one finished 30 minutes ago is held with a due time
    • A test shows a rate-limit failure finished 2 hours ago is still held, and one finished 5 hours ago is rerun
    • A test covers the mixed-class case and that a run at attempt above one is still never rerun
    • reference/safety.md:874-884 states the two waits (one hour for GitHub's hourly installation pool on publish-rate-limit, five hours for the plan session window on rate-limit) and why, and keeps the one-rerun-per-check-per-head and first-attempt-only rules
    • The comment block at babysit_merge.py:336-345 and the docstring at babysit_merge.py:18 match the new waits
    • The babysit-prs script test suite passes

    External state: none

    Related work

    Sources read

    • plugins/source-control/skills/babysit-prs/scripts/babysit_merge.py:336-345: comment and RATE_LIMIT_RERUN_DELAY = timedelta(hours=5), applied to every class outside NO_WAIT_RERUN_CLASSES
    • plugins/source-control/skills/babysit-prs/scripts/babysit_merge.py:461-469: due = completed + (0 if classes <= NO_WAIT_RERUN_CLASSES else RATE_LIMIT_RERUN_DELAY); reason says 'waiting for the usage limit to reset'
    • plugins/source-control/skills/babysit-prs/scripts/babysit_review_evidence.py:1300-1310: publish-rate-limit matcher and GATE_RERUN_CLASSES; the wait itself is not defined here, so the fix goes in babysit_merge.py not where the issue suggests
    • plugins/source-control/skills/babysit-prs/reference/safety.md:874-884: states one five-hour wait for every rate limit
    • plugins/source-control/skills/babysit-prs/scripts/tests/test_babysit_merge_rate_limit_rerun.py:254-258: existing tests on the rerun classes and NO_WAIT_RERUN_CLASSES

    Automated refinement by the lanes App. Intake triages the issue again with it.

  6. melodic-automation-lanes commented on Nov 21, 2026

    @melodic-automation-lanes
    Contributor

    {"v":1,"actor_id":337595936,"title_sha256":"c83bbb45bbe460bfffd820ba9673b8d993ca4dd1a5e29b819437254874682d48","body_sha256":"9b70f10c93cbec0a7a4a132eaf5fd64a9ee31f83bd46c00e54130630991c3793"}

  7. melodic-automation-lanes commented on Nov 21, 2026

    @melodic-automation-lanes
    Contributor

    This records a re-fire of an earlier stop, not a new one. The start gate stopped run 44949368467, the work-resolve-issue run that agent-ready started, before any model turn, because the GITHUB_TOKEN budget was under its start-gate share, or the window was projected to have under its start-project share of the header's limit left by the run's expected end (github-token start-gate, start-project and start-run-minutes in rate-limits.json; the run's start-gate notice names which); that run ended at 2026-11-21T21:18:41Z. work-unblock has now removed and added agent-ready back to start a fresh run. A budget stop is the lanes' load, not a fault in this issue: it never counts toward limits.limit-stops-per-issue or the kill-switch wait.

  8. github-actions commented on Nov 21, 2026

    @github-actions
    Contributor
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Metadata

Metadata

Assignees

No one assigned

    Labels

    No labels
    No labels

    Type

    Fields

    Work status

    agent-ready

    Priority

    medium

    Projects

    No projects

      Milestone

      No milestone

      Relationships

      None yet

      Development

      No branches or pull requests

      Issue actions