You signed in with another tab or window. Reload to refresh your session.You signed out in another tab or window. Reload to refresh your session.You switched accounts on another tab or window. Reload to refresh your session.Dismiss alert
{{ message }}
Repository navigation
A record made after the PR's last run never re-runs « Test sheet and cycle »: the merge stays blocked until the description is edited #399
nstack attest (the record, pushed as notes) and the skeleton's pull-request workflow (« Test sheet and cycle », A1/A2). NapkinStack 0.29.0.
Expected vs observed behaviour
Expected: once nstack attest records the head 5/5, the required check « Test sheet and cycle » reflects it, and the merge is no longer blocked.
Observed: the record moves by a notes push, which is no forge event. The PR's last « Test sheet and cycle » run predates the record, so it stays red and the merge stays BLOCKED. Since D218, an approval no longer re-runs the check either. The only gesture that worked was the author's edit of the description, an edited event.
Evidence on NapkinStack/kitokey (UTC):
PR
Last runs of « Test sheet and cycle »
Record of the head
Unblocked by
#834 (c95ff79)
17:14 cancelled; 17:14, 17:16, 17:17 failure
17:42:29, 5/5
the author's description edit → run 43623648786 at 17:54, success
18:44:51, 4/5 (the sheet still empty); re-attest 19:31:23, 5/5
the author's description edit → run 43632604637 at 19:32, success
The verifier fills the sheet after the first attest, and the record must come after the last edit, since it holds the description's sheet (D214). So this order happens on every PR whose sheet is filled after the evidence run.
Reproduction
Open a PR. CI runs « Test sheet and cycle », which fails A1, since there is no record yet.
nstack attest: the record is 5/5 and is pushed as notes.
The PR stays BLOCKED, and nothing re-runs the check.
Pull request that brought it
D218 (#395) removed the review trigger, which had hidden the problem: an approval started a run. The root is older: the record has never been a forge event.
Impact
Every attested project. On each PR, someone has to edit the description after the last attest, and agents learn this by being blocked. A person's re-run of a pull_request-triggered run should also work, but that is unverified here.
Feedback loop
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 notes, outside the forge's events.
No local reproducer: this is the forge's event order (see #394). The evidence is in the table above.
Triage: generic, accepted, D221, taken before #368's root.
Any attested project meets it. The record holds the description's sheet (D214), so it always comes after the last edit. A notes push is no forge event, so nothing re-evaluates the required check after it.
How others solve it: a result made outside CI reaches a ruleset as a commit status that the tool posts itself. That is how basecamp/gh-signoff (local CI), Buildkite and Jenkins report. The fix weighs two candidates, at the smallest that holds:
nstack attest posts a commit status on the head, and the ruleset requires it. The tool that makes the record also gives the event. This needs Commit statuses: write for the App, and the workflow's A1/A2 must not then contradict that status.
nstack attest re-runs the head's last « Test sheet and cycle » run. This needs Actions: write, which the App never has (07-governance §7), so it is the person's gesture only.
Until it is merged, the gesture is the one this issue names: the author edits the description after the last nstack attest.
D221 merged in 3fd9750 (#400). Where the base branch's ruleset requires nstack / local checks, the pull request's run leaves A1 and A2 to that status, which nstack attest posts with the record. It ships in the next release; then the ruleset needs the status added, and an older agent App needs Commit statuses: write.
Activity
napkinstack-agent commentedon Nov 14, 2026
Triage: generic, accepted, D221, taken before #368's root.
Any attested project meets it. The record holds the description's sheet (D214), so it always comes after the last edit. A notes push is no forge event, so nothing re-evaluates the required check after it.
How others solve it: a result made outside CI reaches a ruleset as a commit status that the tool posts itself. That is how basecamp/gh-signoff (local CI), Buildkite and Jenkins report. The fix weighs two candidates, at the smallest that holds:
nstack attestposts a commit status on the head, and the ruleset requires it. The tool that makes the record also gives the event. This needs Commit statuses: write for the App, and the workflow's A1/A2 must not then contradict that status.nstack attestre-runs the head's last « Test sheet and cycle » run. This needs Actions: write, which the App never has (07-governance §7), so it is the person's gesture only.Until it is merged, the gesture is the one this issue names: the author edits the description after the last
nstack attest.D221 — a required record status keeps an attested pull request moving (…
napkinstack-agent commentedon Nov 14, 2026
D221 merged in 3fd9750 (#400). Where the base branch's ruleset requires
nstack / local checks, the pull request's run leaves A1 and A2 to that status, whichnstack attestposts with the record. It ships in the next release; then the ruleset needs the status added, and an older agent App needs Commit statuses: write.