Skip to content

Run ARC CI when a PR is retargeted onto a release branch - #107

Merged
nullStack65 merged 2 commits into
release/native-0.1.10from
fix/arc-ci-retarget-trigger
Nov 19, 2026
Merged

nullStack65 merged 2 commits into
release/native-0.1.10from
fix/arc-ci-retarget-trigger

Conversation

@nullStack65

@nullStack65 nullStack65 commented Nov 19, 2026 •

Copy link
Copy Markdown
Owner

PR #100 (model-alert dedupe) was retargeted to release/native-0.1.10 after all five of its commits were pushed. The pull_request trigger in closura-arc-ci.yml used the default event types, which leave out edited, so ARC never ran and a real test regression merged (fixed later by #104).

This adds types: [opened, synchronize, reopened, edited, ready_for_review]. A title or body edit also reruns the full job. That's deliberate: skipping the job on non-base edits would post a skipped check on the same head, which could mask an earlier failure.

Not in this PR: making "Voice + founder inbox (Linux ARC)" a required check on release/native-* is a branch-protection setting, not a file change.

Deploy: none

🤖 Generated with Claude Code

PR #100 was retargeted to release/native-0.1.10 after all its commits were
pushed, so the pull_request trigger (default types, no edited) never ran ARC
and the regression merged unchecked.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
@nullStack65

Copy link
Copy Markdown
Owner Author

Automated technical review by claude-fable-5-2 (independent, one round; not a human review or approval)

Head reviewed: eaf21068d1bbc25e782c0d35d2ffaf450a2aa467 (base release/native-0.1.10)
Scope: the one-hunk diff to .github/workflows/closura-arc-ci.yml, the whole file at that head, .github/workflows/closura-deploy-line.yml at that head, the PR body, and run history for both workflows.

Result: 0 blocking findings, 3 non-blocking. The change does what it says.

What I checked and found correct

  • edited fires on a base change. The pull_request edited activity covers title, body and base-branch changes, and the branches: filter is evaluated against the PR's base at event time, so a retarget onto release/native-0.1.10 or feat/assemblyai-current now starts a run. A retarget away from both starts nothing. The retarget of the model-alert dedupe PR (feat(server): one Lead alert per usage-limit incident, auto-resume at reset, refuse launches on an exhausted provider #100) is recorded as base_ref_changed by nullStack65 at 2026-11-19T06:50:34Z, which matches the PR's account.
  • Stale runs are cancelled, not stacked. closura-arc-ci.yml:25-26 groups pull-request runs by PR number with cancel-in-progress: true, so a burst of title or body edits leaves one run on the ARC runners. Observed on this PR: the body edit at 19:38Z queued a second run (44600398941) on the same head after the 18:03Z run had succeeded.
  • No other job depends on the default types. The file has one job. "Deploy line declared" lives in closura-deploy-line.yml:12 and already lists its own types, including edited; this PR does not touch it.
  • Trusted-actor edits cannot produce a skipped check. The job if at closura-arc-ci.yml:33-48 does not look at github.event.action, so an edit by either founder account always runs the full job.

Findings

  1. Non-blocking: an edit by any non-founder actor cancels the real run and posts a skipped check. closura-arc-ci.yml:15 with :25-26 and :40-41.
    The concurrency group is at workflow level, so it applies before the job if. If an app or a collaborator outside the two founder logins edits the title or body (a review bot writing a summary into the description, for example), the new run cancels any in-progress ARC run for that PR, and then its only job is skipped because github.actor fails the allowlist. The head is left with a cancelled run and a skipped "Voice + founder inbox (Linux ARC)" check as the newest result, and if that check is later made required, skipped counts as passing. The same edit after a failed run puts a skipped check on top of the failure. This contradicts the comment at :13-14 ("a skipped check can never stand in for a real result on the same head"), which holds only for the two founder accounts.
    I found no case of it so far: all 40 runs of the deploy-line workflow, which already triggers on edited, have nullStack65 as actor. So it is latent, not live. Smallest fix: soften the comment to say what it covers, or keep non-founder runs out of the shared group, for example by adding github.actor to the group at :25.

  2. Non-blocking: ready_for_review is not needed for the stated goal. closura-arc-ci.yml:15.
    The job has no draft condition, so a draft PR already ran ARC on open and on every push. Adding ready_for_review only reruns the same head once more when the draft flag is cleared, at roughly 8 minutes of runner time each. The smaller sufficient change is types: [opened, synchronize, reopened, edited]. Harmless if kept for symmetry with the deploy-line workflow.

  3. Non-blocking: nothing enforces the result yet. The PR body says the required-check setting is out of scope, which is fair. For the record, release/native-0.1.10 currently has no branch protection at all (the protection API returns 404 "Branch not protected"), so after this merges a retargeted PR gets a run, but can still merge before it finishes or after it fails. The regression described in the PR body is prevented only if whoever merges waits for the check.

Complexity

A one-line trigger change is the smallest thing that works. on.pull_request cannot filter on "base changed", and the alternative of a job-level if on github.event.changes.base would add skipped checks on every founder title or body edit, which is worse than finding 1. The cost is a full rerun per title or body edit, bounded to one at a time by the concurrency group. I agree with that tradeoff.

Not verified

  • I did not retarget a live PR to watch the event fire; the base-change behaviour is from GitHub's documented edited semantics plus the body-edit rerun observed on this PR.
  • The second ARC run on this head (44600398941) was still queued when I reviewed, so I am not reporting its result.
  • A retarget that leaves the PR conflicted with its new base has no merge ref, and GitHub then starts no pull_request run until the conflict is resolved. That is existing platform behaviour, not something this PR changes.

Review round for this PR (Fable 5.2, non-blocking findings 1 and 2): an edit
by a non-founder actor such as CodeRabbit started a run that cancelled the
in-progress founder run and then skipped its only job. PR runs now also group
by triggering actor. ready_for_review is dropped because the job has no draft
condition, so it only added a rerun of the same head.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
@nullStack65

Copy link
Copy Markdown
Owner Author

Resolution of the Fable 5.2 review round (#107 (comment)), fixed in head 4ae4568d7319399a6f8e5384f542d935dd4229cc:

  1. Finding 1 (non-founder edits leave a skipped check): fixed. CodeRabbit is installed on this repo and edits PR descriptions, so this would happen in practice. PR runs now also group by github.triggering_actor. A bot's edit-triggered run lands in its own group, and its job still skips, but it can no longer cancel a founder's in-progress run. The workflow comment now describes this accurately.
  2. Finding 2 (ready_for_review unnecessary): fixed. It's dropped from types, because the job has no draft condition.
  3. Finding 3 (no branch protection on release/native-0.1.10): not changed here. It's a branch-protection setting, and the PR body already puts it out of scope.

actionlint passes apart from its usual warning about the custom t3code-ci-arc runner label, which was already there before this change. That closes the one bot review round.

@nullStack65
nullStack65 merged commit 7af284f into release/native-0.1.10 Nov 19, 2026
3 checks passed
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Labels

None yet

Projects

None yet

Development

Successfully merging this pull request may close these issues.

2 participants