P-2 run again for #659: the merge queue merges on its required checks alone, with no pass read

Merged
DO
8d ago
+3359-1

This answers #659, "Trial: the merge queue merges only after a pass has read the pull request", following the Factory Builder's rulings of 2026-11-09: option (B), the first leg only, and the verdict below. It is P-2 run again for its question of the poller, recorded as trials/P-2.no-pass.json.

Verdict: pass, by the Factory Builder's ruling. The pass condition stays as written before the run. The ruling is quoted in the fixture's final answer exchange and in P-2's row: the clause "no check run or status but p2-queue-check's" means no check or status from a Factory App.

  • CodeQL, SonarCloud (two Apps) and CodeRabbit posted checks beside p2-queue-check. None is a Factory App's and none is required.
  • The answer exchange says plainly that the clause did not hold as written.

What GitHub did

In combsmith/software-factory-canary, with p2-queue-check as main's one required check:

  • The Maintainer opened PR #100, one commit adding trial-notes/pass-queue.md.
  • He queued it at 13:41:54Z.
  • The merge group's p2-queue-check ran on the group's commit, 6cc643a, and passed at 13:48:12Z.
  • The queue merged it at 13:48:46Z. GitHub names the Maintainer, who enqueued it, as the merger, and github-merge-queue as what took it out of the queue.
  • No pass was read, and no check or status from any Factory App came on the head or the merge.

So the queue merges on the required checks alone. A person's enqueue was shown here; the queue App's is shown in trials/P-2.queue-app.json.

The fake

sim/grounding.py's a_pass_comes_before_the_merge claims something GitHub doesn't do. It is left unverified, with a comment, and #677, "The fake's merge queue merges on its required checks alone, as GitHub's does", is raised to change the fake (skill step 7). A required Factory check on the merge group is not the design: it would break §9.1's recovery rule (the Builder's ruling).

Section 6: P-2's row gains the run.

What the trial did, and what was undone

  • The ruleset, main. Set for the trial to ask for no approving review, no code owner's review and no approval of the last push, so that a pull request the Maintainer opened could be queued. It was restored afterwards by a PUT of the ruleset as read before. The read after the restore matches the read before in every field except updated_at.
  • Left on the canary's main: the merge of PR #100, trial-notes/pass-queue.md, which moved main from 7737f8f to 6cc643a, as P-2's earlier merges did.
  • Deleted: the branch trial/pass-queue.
  • Removed from the fixture: the authorization header of each gh request, all 12 listed in removed.

No Maintainer step was needed.

Checks

CheckResult
SIM_SEEDS=40 make verifyOK, 1726 tests (at 6380bb14)
python3 -m tools.check_trialsok
gitleaksNo leaks
Argon (6d35865), 3 partsClean
Muse (6d35865)Major: the verdict against the written clause, settled by the Builder's ruling, now stated in the answer exchange. Minor, fixed in 6380bb14: the merger is now as GitHub names it, and a CodeQL job that ran after the merge is no longer cited as one the queue skipped
Claude review and security (6d35865)The same merger fix (blocking) and CodeQL fix. Minor, fixed: the ruling moved into an answer exchange; the restore described; "whoever enqueued" cited. Nothing for security
AstraOnly where Muse gives no read

🤖 Generated with Claude Code