[lukemaj]

What: A job whose PR is already merged at completion now blocks once with pr_merged and the merge commit, instead of looping on pr_not_open.
Why: The runner never merges (RUNNER.md step 6), so resuming the dispatcher with "update the existing open PR" could never succeed and looped (job prism-mumm1j50-69984526, twice).
So what: Needs independent review, then the user's merge; the planner still decides whether a merged outcome counts as done.

Closes #143.

Change

• runner/core.py: default_pr_verifier reads mergeCommit; verify_pr_for_completion refuses MERGED as completion_refused: pr_merged (PR was merged, merge commit <sha12>; ...). CLOSED still refuses as pr_not_open.
• runner/controller.py: _handle_completion_refusal records a pr_merged refusal and blocks at once, with no dispatcher resume. The existing blocked path then sends the planner one terminal report carrying the reason.
• RUNNER.md step 6: one sentence on the new behavior.
• Version 0.42.1 -> 0.42.2 (patch).

This follows the Issue's correction comment, not its original acceptance criteria: the runner does not accept a merged PR as completion, because merge authority for Prism jobs is still an open design question (optional-dispatcher and planner-dispatch work). Stopping jobs from merging in the first place belongs to the dispatcher instructions (#147) and is not changed here.

Proof

• tests/test_issue143.py (5 tests): merged PR refuses as pr_merged with the merge commit; closed PR still pr_not_open; default_pr_verifier parses mergeCommit through a fake gh on PATH; controller.step on a merged PR ends blocked, sends nothing to the dispatch thread, sends one planner report naming pr_merged, and keeps no PR identity; a closed PR still resumes the dispatcher once.
• Failing before: with runner/ reverted, the new tests fail (2 failures, 1 error; the merged case reported pr_not_open (state=MERGED ...) and resumed the dispatcher).
• python3 scripts/test.py: Ran 379 tests ... OK at 32795fa.
• versionctl release-check: release candidate OK: v0.42.2 at 32795facd32ac8d0d1a73415fd1b220f4318dedb.

Elon record

• Requirements and who asked: the user's correction on Runner rejects completion after an authorized merge (pr_not_open) #143 (2026-09-29): a merge by the job must be prevented or reported clearly instead of looping on pr_not_open; the runner never merges by design.
• Deleted: accepting a merged PR as completion (the original acceptance criterion, withdrawn by the correction); any merge-authority task field; the dispatcher resume for a state no turn can fix.
• Bottleneck: the refusal path resumed the dispatcher for every pr_ refusal, including a merged PR it cannot reopen.
• Checked myself: read verify_pr_for_completion and _handle_completion_refusal, confirmed the new tests fail without the fix and pass with it, and ran the full suite.

🤖 Generated with Claude Code

[lukemaj]

APPROVE

Reviewed commit: 32795facd32ac8d0d1a73415fd1b220f4318dedb.

No blocking findings. The completion gate has one caller, and every verifier refusal flows through _handle_completion_refusal; MERGED is detected before the generic non-open refusal and is recorded then blocked without resuming the dispatcher. The same-sequence retry path already blocks, and ordinary closed PRs retain the existing dispatcher retry. The focused tests distinguish the merged and closed cases and assert no dispatch-thread call, one planner report, and no saved PR identity for the merged case. RUNNER.md accurately describes the runner's behavior and leaves the completion decision with the planner.

No findings.