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
qa-lead never advises a PR that was opened with merge conflicts — pull_request.opened does not run, and synchronize is deliberately not subscribed #2253
A PR that is opened while it conflicts with main never gets a qa-lead advisory.
GitHub does not run pull_request workflows for a PR it cannot build a merge commit for. So the opened event produces no run of .github/workflows/persona-mention.yml. When the conflict is later resolved, the push arrives as synchronize, which the stub does not subscribe (pull_request: [opened, ready_for_review, reopened]) and qa-lead does not declare (events: [opened, ready_for_review], with synchronize excluded on purpose so a PR gets one advisory, not one per push). Nothing else ever fires for that PR.
only pull_request_target workflows ran; no pull_request workflow ran at all (no CI, no CodeQL, no router)
00:30:14
Auto-rebase failed — merge conflict
01:38:04
dev-lead force-pushed a rebase; "PR is no longer conflicting"
01:38:07
the pull_request workflows ran for the first time, on synchronize — the router is not subscribed to it
No qa-lead advisory was posted. Of the 35 PRs opened in this repo between the gate fix (#1909, 2026-10-09 23:07Z) and 2026-10-11 02:20Z, 33 have exactly one advisory, one was correctly skipped as no-test-surface (#2225), and this is the one that was missed.
This is not new with the router: the retired qa-lead-pr-advisory.yml subscribed to the same two actions and had the same gap. It is visible now only because the surface works.
Why it needs a decision rather than a fix
The obvious change — subscribe and declare synchronize — reverses a deliberate choice recorded in personas/qa-lead/persona.yml and ADR-0009. The runner's already-advised check would keep it to one advisory per PR, so it would behave as "advise on the first event that can run", but every push to every PR in every repo carrying the stub would start a router run.
Options to weigh:
Declare synchronize for qa-lead and subscribe it in the stub. Simplest; relies on the already-advised check for idempotency. Cost: a router run per push, fleet-wide once the stub is deployed.
Catch-up on conflict resolution only. Have the path that resolves the conflict (dev-lead's rebase intent, or auto-rebase) signal the router for a PR with no advisory yet. Narrow, but couples two agents.
A scheduled catch-up sweep over open PRs with test surface and no <!-- persona:qa-lead --> comment. Covers this and any other missed opened event; adds a scheduled workload.
Accept the gap and document it in the persona's README.
Acceptance criteria
The solution-architect (or the maintainer) records which option is chosen and why, as an ADR-0009 consequence or a short new ADR.
Whatever is chosen is implemented or documented, with a test or a stated measurement that shows a PR opened in a conflicted state ends up either advised or explicitly, visibly skipped.
Thanks for the thorough write-up and the concrete timeline from PR #2238. The evidence is clear and the four options are well laid out. The question that most affects the choice is this: how often do you expect PRs to be opened in a conflicted state across the fleet, since that sets how much a per-push router run (option 1) would cost compared with a narrower catch-up (options 2 or 3)? If you have a count from the other repos carrying the stub, that would help the maintainer and solution-architect decide.
Activity
github-actions commentedon Oct 11, 2026
Thanks for the thorough write-up and the concrete timeline from PR #2238. The evidence is clear and the four options are well laid out. The question that most affects the choice is this: how often do you expect PRs to be opened in a conflicted state across the fleet, since that sets how much a per-push router run (option 1) would cost compared with a narrower catch-up (options 2 or 3)? If you have a count from the other repos carrying the stub, that would help the maintainer and solution-architect decide.