A builder run that declines as "already done on main in <commit>" posts that note on its issue and stops further runs on it #6592

A builder run that declines "already done on main in <ids>" after its pull was closed now keeps that decline, so the issue is noted and held instead of parked as "the pull closed". I did not run the end-to-end path; the new tests cover the pieces, not the runner. I ran the type check, design-lint, and the crew parks and already-done tests (35 pass).

What was already there. The runner already turns a decline whose ids are on the base into already done: <oids>. On an issue-level run, settleAgentRun then posts the "Done on main already" note naming the commits and clears the assignee. The done_on_main hold already keeps new build runs off the issue. An id not on the base already falls through to an ordinary decline.

Why #6467 still lost. I could not see the run records, so this is my reading of the code and of #6467's comments. The builder's push made the forge open a pull whose body was the decline text. That pull was closed while the run was still finishing. pull.closed cancelled the run with the summary "the pull closed", so the decline was never recorded. freeStalled then read that summary and parked the issue with "the pull closed" as the reason, twice.

What changed.

• src/core/runs.ts has a new answerCancelled. It replaces the state and summary of a run that was cancelled with "the pull closed" or "the pull landed". It touches no other run.
• The runner no longer discards a DECLINED answer just because its pull closed (the meter stopping it still counts). When the run's pull has already cancelled it, the runner records the decline or the already-done outcome through answerCancelled and announces it.
• An already-done outcome then goes through the existing note and hold. A plain decline is parked quoting declined: ….
• src/crew/dispatch.ts shares the two cancel summaries as PULL_STOPPED instead of repeating the strings.
• The new tests are in test/crew-parks-done.test.ts.

- Already-done: a run cancelled by its pull but answered already-done posts the note and holds the issue, with no strike. - Plain decline: a plain decline is quoted in the park. - Left alone: a run cancelled for another reason, or already answered, is left alone. - Not covered: the "id not on the base" case, which leaves already.ts unchanged and is already covered by the existing "not on the base is a decline" test. The runner change is also not covered end to end.

Left out.

• I did not change which pulls the forge opens from a builder's push, or the pull body.
• A decline that arrives after the pull was closed is now recorded as a decline, so a plain decline there parks the issue where before it would have been cancelled silently.

Closes #6573

A builder run that declines "already done on main in <ids>" after its pull was closed now keeps that decline, so the issue is noted and held instead of parked as "the pull closed". I did not run the end-to-end path; the new tests cover the pieces, not the runner. I ran the type check, design-lint, and the crew parks and already-done tests (35 pass).

What was already there. The runner already turns a decline whose ids are on the base into already done: <oids>. On an issue-level run, settleAgentRun then posts the "Done on main already" note naming the commits and clears the assignee. The done_on_main hold already keeps new build runs off the issue. An id not on the base already falls through to an ordinary decline.

Why #6467 still lost. I could not see the run records, so this is my reading of the code and of #6467's comments. The builder's push made the forge open a pull whose body was the decline text. That pull was closed while the run was still finishing. pull.closed cancelled the run with the summary "the pull closed", so the decline was never recorded. freeStalled then read that summary and parked the issue with "the pull closed" as the reason, twice.

What changed.

• src/core/runs.ts has a new answerCancelled. It replaces the state and summary of a run that was cancelled with "the pull closed" or "the pull landed". It touches no other run.
• The runner no longer discards a DECLINED answer just because its pull closed (the meter stopping it still counts). When the run's pull has already cancelled it, the runner records the decline or the already-done outcome through answerCancelled and announces it.
• An already-done outcome then goes through the existing note and hold. A plain decline is parked quoting declined: ….
• src/crew/dispatch.ts shares the two cancel summaries as PULL_STOPPED instead of repeating the strings.
• The new tests are in test/crew-parks-done.test.ts.

- Already-done: a run cancelled by its pull but answered already-done posts the note and holds the issue, with no strike. - Plain decline: a plain decline is quoted in the park. - Left alone: a run cancelled for another reason, or already answered, is left alone. - Not covered: the "id not on the base" case, which leaves already.ts unchanged and is already covered by the existing "not on the base is a decline" test. The runner change is also not covered end to end.

Left out.

• I did not change which pulls the forge opens from a builder's push, or the pull body.
• A decline that arrives after the pull was closed is now recorded as a decline, so a plain decline there parks the issue where before it would have been cancelled silently.

A builder run that declines "already done on main in <commit>" after its pull was closed now keeps that decline. The issue gets the "already done on main" note, loses its assignee and is held from new build runs. Before this change the run was cancelled with the summary "the pull closed", and freeStalled parked the issue with that as the reason, twice.

• Cause: with the early pull open, closing it cancelled the run mid-flight, and the runner discarded its DECLINED answer.
• Fix:

- The runner keeps a decline reached before the pull closed. A run the meter stopped keeps nothing. - A new answerCancelled in src/core/runs.ts rewrites only a run cancelled with "the pull closed" or "the pull landed". - An already-done decline now follows the existing note-and-hold path. A plain decline is parked quoting declined: ….

• Review points: the decline checks now come from one helper set in src/crew/already.ts (declineOf, answeredBeforeStop, isAnswer), and I dropped the redundant ok/failed expression.

Tests

• Runner test (test/crew-early.test.ts): a stand-in agent pushes, so the early pull opens, then answers DECLINED: already done on main in <oid> with PUSH-THEN-DONE-ON-MAIN (new in test/helpers/crew-standin.ts). The test closes the pull.

- The run ends ok as already done: <oid>, not cancelled. - The issue gets the "already done" note naming <oid>, no park, and no assignee. - With the signal.aborted guard reverted to a plain abort, this test fails. - The repo is switched to autonomous for the test and back off at the end.

• Meter case: it is a pure unit test of answeredBeforeStop, not a runner run. The stand-in cannot print a result and then be stopped by the meter. The existing grant test already covers the stop itself.
• Second runner edit: I did not revert the isAnswer branch and re-run, but the runner test depends on it to turn cancelled into ok.

Plain decline after the pull closes now parks the issue. It adds a strike where there used to be none (the run was simply cancelled). I kept this on purpose: the decline is quoted in the park, so the planner sees the reason. If you would rather it add no strike, say so; it is a small change in settleAgentRun. The README is not touched, so this is stated only here.

The type check, design-lint and the tests I touched pass: crew-early, crew-parks-done and crew-already-done (26 tests).

A builder run that says "already done on main in <commit>" now posts its note on the issue and stops further runs, even when its pull was closed while it finished. The two touched suites pass and the type check is clean. I did not run the whole check.

All three review points are fixed:

1. Scope: isAnswer now matches only already done: …. answeredBeforeStop also keeps only an already-done decline, so a plain decline whose pull closed stays cancelled with no strike, as before. I dropped the plain-decline park test. A test now checks that a plain decline is not kept, and another that answerCancelled only touches a run its pull cancelled.
2. End-to-end test:

- The whole test is wrapped in try/finally, so autonomy is switched off again even if an assertion fails midway. - The three 300×100 ms loops are one until helper that polls every 20 ms and gives up after 15 s. - It waits for the stand-in's answered marker and the open pull together. - The test budget is 30 s instead of 60 s.

1. README and tidiness: The runner's already.ts import is broken onto separate lines. I described the behaviour in docs/CREW.md under "Already done", because the crew behaviour is documented there rather than in the README. The README is untouched.

To try it: run bun test test/crew-early.test.ts test/crew-parks-done.test.ts.