Trial P-66, a queued pull request: once the base it waits on moves into conflict with it, the queue takes it out unmerged (#812)

Merged
DO
4d ago
+892-1

Part of #812 (trial P-66). This PR records the trial's second case, a base that moves while a pull request is queued. Its first case, an enqueue in conflict, is the companion pull request with trials/P-66.json. The section 6 row added here names both fixtures, so the two land back to back.

The trial

Question, as posted on #812 before the run: does the merge queue drop a queued pull request when its base moves into conflict with it, and what does that leave on its timeline, in mergeable, mergeStateStatus and the queue's entry?

Verdict: pass.

  • Setup. Canary pull requests C (116) and D (117) each add p66/two.md with different words. Both were enqueued at once: C at 12:56:16Z in position 1, D at 12:56:18Z in position 2.
  • What the queue did. The queue builds one entry at a time (maximumEntriesToBuild 1), so only C's merge group ran. D read MERGEABLE/CLEAN and QUEUED until C merged at 13:01:15Z.
  • D's removal. At 13:01:16Z D's timeline gained a RemovedFromMergeQueueEvent with reason merge_conflict, actor github-merge-queue and beforeCommit null. No merge group was ever built for D. D then read CONFLICTING/DIRTY with no entry, and canary main never held its change. REST gives the removal as removed_from_merge_queue with no reason.
  • Approvals. The Maintainer made the four approvals on canary 114–117 by hand, in the browser, as the machine user donpetry-bot. The machine-user-approves workflow did not make them.
  • PR 120. Canary pull request 120, which belongs to another trial, ran its own merge group from 13:11Z to 13:15Z, after this case ended. It had no bearing on it.
  • Not asked: a queue that builds a group on another's, and an enqueue made while the head still reads UNKNOWN or a stale MERGEABLE.

Grounding. sim/grounding.py names this fixture in a later pull request, together with P-67's, once the protected edit is ruled.

What was removed and what was undone

  • Captures and redaction. Every exchange was captured with gh api --include into a scratch directory outside the repo and redacted with tools.redact_trial before anyone read it. The fixture keeps the date header alone, and running the redaction over the drafted fixture removed nothing more.
  • The extra read. One read was made after the run: a read-only GraphQL read of the node ids, at 13:31:46Z. tools/check_trials.py needs that answer for each enqueue, and the fixture marks it as made after the run.
  • Undone in the canary:
    • 115 and 117 were closed unmerged.
    • Branches p66-a to p66-d were deleted.
  • Kept in the canary:
    • Issues 118 and 119 stay closed; trial P-67 closed them.
    • p66/one.md and p66/two.md stay on canary main as harmless text.

Second readers

ReaderHeadResult
Carbonb3dfdb67not run (quota), owed
Museb3dfdb67not run (quota: the window resets 2026-11-16), owed
Astrab3dfdb67no findings
Claude + securityb3dfdb67one minor: the row claimed more than the run showed. Narrowed in 241189ce, which is b3dfdb67 rebased onto the newer main with only the row changed

Checks

  • make checks passes.
  • tools.check_trials and tests.test_tools.Trials pass.

šŸ¤– Generated with Claude Code