Skip to content

A review-triggered run of « Test sheet and cycle » fails before the record is final and is never re-run: the merge stays blocked #394

Closed

Description

@napkinstack-agent

Module concerned

The skeleton's .github/workflows/pull-request.yml (the « Test sheet and cycle » job), on a project whose checks are attested, NapkinStack 0.27.0

Expected vs observed behaviour

Expected: once the head's record is complete and passed, the required check « Test sheet and cycle » no longer blocks the merge.

Observed: the workflow also runs on pull_request_review. If the person approves before the record is complete, which is the usual moment for a code-owner approval, that run fails (A2). Nothing re-runs it once the record is green, because pushing the notes is no event. Later pull_request runs (on a description edit) pass, but the ruleset still sees the review-triggered run as failing, and the merge is refused: « Required status check "Test sheet and cycle" is failing » (HTTP 405), with mergeStateStatus BLOCKED. Seen on NapkinStack/kitokey #809 on 2026-11-14: run 43587778954 (event pull_request_review) failed at 10:58, run 43587846142 (pull_request) passed at 10:59, and the record became 5/5 at 15:27. Agents' Apps cannot re-run a workflow, so the person has to.

Reproduction

  1. A PR on an attested project, its record still incomplete (sheet not filled, for example).
  2. A code owner approves → the pull_request_review run of « Test sheet and cycle » fails on A2.
  3. The sheet is filled and the head attested 5/5; the description is edited → the pull_request run passes.
  4. Merge → refused, since the review-triggered run is still failing.

Pull request that brought it

unknown: the workflow's pull_request_review trigger.

Impact

Every attested project where a code owner approves before the record is final. The merge waits for a person's manual re-run, and the cause is not obvious from the forge (« failing » while the latest run is green).

Smallest answer for the framing to weigh

Run the pull-request checks in one concurrency group per PR, and have the review-triggered run judge only the approval-dependent rules. Or drop pull_request_review from the jobs that do not read approvals, as « Test sheet and cycle » reads the sheet and the record, not reviews. Prior art: a check tied to the head commit, re-evaluated per head (GitHub's required checks, Gerrit's submit requirements), not per review event.

Feedback loop (skeleton/.nstack/handbook/10-measurement.md §4)

  • A test reproducing the bug will be added (it must fail BEFORE the fix)
  • I have identified why the system did not catch it: the record moves by a notes push, which is no forge event, so the review-triggered run is never re-run.

No local reproducer: this is the forge’s event ordering, which a throwaway clone cannot replay. The evidence is the two runs above; the trigger is still on main (skeleton/.github/workflows/pull-request.yml l.13).

Activity

  1. napkinstack-agent commented on Nov 14, 2026

    @napkinstack-agent
    ContributorAuthor

    Triage: generic, and it blocks merges, so it goes before #387. The cause is a run a later event cannot replace: GitHub keeps the review-triggered check run beside the pull_request one, and the notes push that completes the record is no event. A run nothing can re-trigger is a dead end for an agent, who cannot re-run a workflow. Direction: the required job runs only on pull_request events. A review is read as T2's hand at the next run; the verifier signs in the description, so that edit re-runs it, and anyone can edit, an agent included. Prior art: required checks bound to the head and re-evaluated per head event (GitHub, Gerrit's submit requirements). Slot D218; #387 moves to D219.

  2. napkinstack-agent commented on Nov 14, 2026

    @napkinstack-agent
    ContributorAuthor

    More evidence (kitokey #809): re-running the failed review-triggered job does not help. A re-run replays the original pull_request_review payload, including the PR description as it was then (an empty sheet), so it fails again on T2/T5 even though the record is 5/5 and the current description passes. Only a new review event re-judges the current description. So the person's re-run, the only gesture available to them, cannot clear it.

  3. napkinstack-agent commented on Nov 14, 2026

    @napkinstack-agent
    ContributorAuthor

    Thanks, taken into D218 (#395): the migration note and the register row now say a new push, or before the update a new review. A re-run replays the old payload and never clears the block.

  4. napkinstack-agent commented on Nov 14, 2026

    @napkinstack-agent
    ContributorAuthor

    D218 merged in 91c9a06 (#395). Unreleased, it goes in v0.29.0.

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

    bugSomething isn't working

    Type

    No type

    Projects

    No projects

      Milestone

      No milestone

      Relationships

      None yet

      Development

      No branches or pull requests

      Issue actions