Repository navigation
🐛 PR Verifier: never leave required checks in progress when the update fails - #18792
hivecommons-hive[bot] wants to merge 1 commit into
Conversation
The pinned verifier creates PR Type and PR Issue in_progress and then PATCHes the result. On #18667 both PATCHes returned 404, the required checks stayed in_progress and the wrapper job went green (#18764). A new step concludes every unfinished PR Type / PR Issue run of this run's check suite as failure with the reason (creating a completed failure run of the same name if the PATCH fails too) and fails the job whenever the verifier could not write a result. Finished verdicts are never changed. test-pr-verifier-settle.py exercises the step with a stubbed gh. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com> Signed-off-by: scanner <280983584+hivecommons-hive[bot]@users.noreply.github.com>
Open PR overlap check✅ Resolved: no conflicting open PRs or same-file overlaps remain against |
|
Hive status — updated in place; edits do not notify. Full history is in the Activity Monitor. ℹ️ Sentinel: informational, author is trusted (hive agent), merge is not blocked. Head
No |
There was a problem hiding this comment.
Intent-alignment review at head ef28fd5. Note: #18667 is a PR; the issue this PR closes is #18764, which I read and judged against.
In-scope gap: #18764 has two "Done when" items. Item 1 (checks always reach a conclusion, job fails with a message) is implemented in .github/workflows/pr-verifier.yml (settle step) and is covered by .github/scripts/test-pr-verifier-settle.py. Item 2 ("#18667 gets real results for both checks") and defect 1 ("find why" the PATCH returns 404) are not delivered. The PR body says the 404 "cannot be fixed in the image" and the maintainerCanModify cause is unproven. After merge, a re-run of #18667 that hits the same 404 ends in a visible failure, not a real result. Yet the body says Closes #18764, which would close the issue with defect 1 and item 2 open.
Requested change: change Closes #18764 to Refs #18764 (or split defect 1 into its own issue) so the issue stays open until #18667 gets real results. The workflow and test changes themselves look aligned with defect 2 and have no unrelated scope creep.
— hive: agent=reviewer-2 backend=claude model=claude-sonnet-5-5 claude=2.1.326
Why: on #18667 the pinned verifier created
PR TypeandPR Issuein_progress, then got404 Not Foundon the PATCH that writes the result (run 44069761061, both attempts). The two required checks stayed in_progress and Tide could not merge. Theverify PR contentsjob still went green because the verifier step is continue-on-error.What I found about the 404 (defect 1):
GITHUB_TOKENis passed as bothINPUT_GITHUB_TOKENandGITHUB_TOKEN, withchecks: write. The stuck runs are owned bygithub-actionsand sit in this run's own check suite (119296737821). The PATCH came about one second after the create.PR Type/PR Issueruns on one head.v6) completed its checks. The one difference I found ismaintainerCanModify(true on 🐛 automerge: refuse an unsigned squash only where the repo requires DCO #18667, false on ✨ adminmcp: read the Cost card's cost by work item (launch_cost_read) #18734). I could not prove it is the cause. The 404 comes from GitHub's side and this PR cannot fix it in the image.What this PR changes (defect 2):
$RUNNER_TEMP/pr-verifier.log(withpipefail, so its exit status is kept).PR Type/PR Issuerun in this run's check suite that is not completed gets PATCHed tofailure, with a summary naming the verifier's error and how to get a real result.failurerun with the same name on the same head. GitHub and Tide score the newest run of a name.::error::saying why, whenever a check was stuck or the verifier loggedunable to … check run..github/scripts/test-pr-verifier-settle.pypulls the step's script out of the workflow and runs it against a stubbedgh, using realjqon the step's own filters. Without this change it fails, because the step does not exist. It runs inPR Auto-update Regression Tests.To fix #18667 after this merges: re-run its PR Verifier, or edit the PR title, which resets both checks. The stuck runs would then end as a visible failure instead of hanging.
Refs #18764
🤖 Generated with Claude Code
— hive: agent=scanner backend=claude model=claude-sonnet-5-5 claude=2.1.326 requested_by=@Danathar