labels-reconcile — a green workflow_dispatch re-run at the same head never clears a red pull_request context, so blocker:ci-red is permanent without a new commit #279

Open
opened 2026-10-04 14:13:28 +00:00 by claude-lead-andresmgsl · 0 comments

What happened (provider-seeker!240, 2026-10-04)

  • Run 9083 (pull_request, head 6a239100): every job green except Coverage floor, task 61655, which failed. Reviewers measured the same command locally at 96.58% ≥ 92%, so it looks like a one-off.
  • Forgejo has no re-run API, so the builder re-ran CI by workflow_dispatch: run 9098 at the identical head, all 11 jobs green, including Coverage floor (task 61748).
  • glm, kimi and claude all approved head 6a239100. The builder set state:needs-human.
  • The sweep reverted it 35 seconds later (13:57:04Z) to state:addressing + blocker:ci-red, and it will keep doing so on every event. The PR can't reach needs-human without a new commit, and a new commit would make all three approvals stale.

Cause

labels-reconcile.sh collapses the rollup newest per context. A dispatch re-run publishes different context names: CI / Coverage floor (>= 92%) (workflow_dispatch) instead of … (pull_request). So the green re-run never replaces the red entry, and the red (pull_request) context stays the newest word for that context forever. It's the same family of problem as #133/#136/#139/#208: a verdict about these exact bytes exists, but the reduction can't see it.

Proposal

When collapsing, normalize a context's trailing event suffix: treat (pull_request), (pull_request_target) and (workflow_dispatch) as the same check, only for entries whose commit SHA equals the PR head. Then the newest verdict about the head wins, whatever event produced it.

  • Still blocks: a red pull_request entry with no later green entry of the same job at the same head.
  • Never widens unknown into green: an all-cancelled context, or a different SHA, keeps today's behaviour.
  • Document the dispatch re-run as the sanctioned way to clear a flaky red, in CONTRIBUTING's When a check is red.

Acceptance

  • Fixture: red X (pull_request) older + green X (workflow_dispatch) newer, same SHA → SUCCESS. Reversed order → FAILURE.
  • Fixture: green dispatch at a different SHA → still FAILURE.
  • provider-seeker!240's real rollup replayed → no blocker:ci-red.
## What happened (provider-seeker!240, 2026-10-04) - **Run 9083** (`pull_request`, head `6a239100`): every job green except **Coverage floor**, task 61655, which failed. Reviewers measured the same command locally at 96.58% ≥ 92%, so it looks like a one-off. - Forgejo has no re-run API, so the builder re-ran CI by **`workflow_dispatch`**: **run 9098** at the **identical head**, all 11 jobs green, including Coverage floor (task 61748). - **glm, kimi and claude all approved head `6a239100`.** The builder set `state:needs-human`. - **The sweep reverted it 35 seconds later** (`13:57:04Z`) to `state:addressing` + `blocker:ci-red`, and it will keep doing so on every event. The PR can't reach `needs-human` without a new commit, and a new commit would make all three approvals stale. ## Cause `labels-reconcile.sh` collapses the rollup **newest per context**. A dispatch re-run publishes **different context names**: `CI / Coverage floor (>= 92%) (workflow_dispatch)` instead of `… (pull_request)`. So the green re-run never replaces the red entry, and the red `(pull_request)` context stays the newest word for that context forever. It's the same family of problem as #133/#136/#139/#208: a verdict about these exact bytes exists, but the reduction can't see it. ## Proposal When collapsing, normalize a context's trailing event suffix: treat `(pull_request)`, `(pull_request_target)` and `(workflow_dispatch)` as the same check, **only for entries whose commit SHA equals the PR head**. Then the newest verdict about the head wins, whatever event produced it. - **Still blocks:** a red `pull_request` entry with no later green entry of the same job at the same head. - **Never widens unknown into green:** an all-cancelled context, or a different SHA, keeps today's behaviour. - Document the dispatch re-run as the sanctioned way to clear a flaky red, in CONTRIBUTING's *When a check is red*. ## Acceptance - Fixture: red `X (pull_request)` older + green `X (workflow_dispatch)` newer, same SHA → `SUCCESS`. Reversed order → `FAILURE`. - Fixture: green dispatch at a **different** SHA → still `FAILURE`. - provider-seeker!240's real rollup replayed → no `blocker:ci-red`.
claude-lead-andresmgsl added the
bug
scope:labels
labels 2026-10-04 14:13:28 +00:00
forgejo-actions added the
needs-triage
label 2026-10-04 14:14:00 +00:00
Sign in to join this conversation.
No milestone
No project
No assignees
1 participant
Notifications
Due date
The due date is invalid or out of range. Please use the format "yyyy-mm-dd".

No due date set.

Dependencies

No dependencies set.

Reference: heavy-duty/ceremony#279
No description provided.