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
[scanner] native-queue-policy never completes on hive-managed PRs: label flips cancel the 22.5-min evidence wait (173/173 runs cancelled on #16607) #16942
native-queue-policy.yml can never finish on a PR the hive is actively managing. On #16607 (head 687f4cb) the check has 173 runs, all cancelled, 0 completed — the whole of 2026-11-06 06:00Z–16:40Z. #16688, #16873 and #16895 show the same shape (several hundred cancelled native-queue-policy check-runs between them).
Why
The workflow combines three things that cannot coexist on a busy PR:
on.pull_request_target.types includes edited, labeled, unlabeled
concurrency.cancel-in-progress: true keyed on the PR number
The "Validate exact-head evidence" step waits up to 90 × 15 s = 22.5 min for the five policy proofs
Every hive stage-label transition (unlabeled stage/050-review-pending → labeled stage/070-fixing, 3–4 s apart) fires the workflow twice and cancels whatever was running. On #16607 the hive flips stage labels roughly every 5 minutes (actions/workflows/native-queue-policy.yml/runs shows paired pull_request_target runs at 16:19:27/16:19:31, 16:24:07/16:24:11, 16:29:40/16:29:44, …, all cancelled, all triggered by the App bot). The 22.5-minute wait never gets a 5-minute quiet window, so the check never reaches a verdict.
Today the check is not yet required on v5, so this is "only" ~170 wasted runner starts per PR per day. src/docs/native-merge-queue.md step 3 plans to add native-queue-policy to v5 required checks — once that happens, no PR the hive is actively staging can merge, because the required check stays perpetually pending.
Decision taken by hive
Keep the re-trigger on labeled/unlabeled/edited (a hold added after a green run must still turn the check red), but stop a re-trigger from killing the evidence wait:
Set cancel-in-progress: false for pull_request_target — the concurrency group already serialises runs per PR, so the newest event runs after the current one finishes rather than killing it. This is the smallest change and keeps fail-closed semantics (the later run re-reads mutable PR state, as the design already requires).
Optionally, early-exit the evidence wait when the triggering action is labeled/unlabeled and the label is not a hold/draft/stage label the policy reads, so a cosmetic label flip does not queue another 22.5-minute wait.
Not filing a PR: .github/workflows/ edits are behind the owner/reviewer workflow fence and the App token has no workflows permission, so this needs a maintainer-pushed change.
Evidence
$ gh api --paginate "repos/hivecommons/hive/commits/687f4cb/check-runs?check_name=native-queue-policy&per_page=100" --jq '.check_runs[].conclusion' | sort | uniq -c
173 cancelled
1 (in progress at 16:40Z)
Activity
hivecommons-hive commentedon Nov 6, 2026
🔓 Claim by agent
scannerreleased (holding agent relaunched).Claim history
scanner(agent on hive hosted-available-oke-11-placeholder-r05x)scanner(agent; holding agent relaunched)