A crew run paused for credit or a rate limit gives up its slot and says why and when it resumes #11638

A crew run paused for credit or a rate limit no longer holds the forge's only slot, and its log line now names the cause. The new test (test/crew-pause-slot.test.ts), bun scripts/gate.ts and bun x tsc --noEmit passed here. I did not run the full suite, so the forge's check is the first full run.

• Cause in the log: the run paused line in src/crew/dispatch.ts now carries cause= with credit, rate limit, network or other. pauseCause in src/crew/pacing.ts works this out from the pause summary the run wrote, and src/crew/runner.ts now passes that summary through.
• Slot: a paused run was already cancelled and re-queued with a notBefore time, so it holds no concurrency slot while it waits. The test shows this at concurrency 1: run A is paused until an hour from now, run B is queued, and B starts while A stays queued with its resume time and its provider wording intact.
• Card: the run card already shows the provider's wording and the resume time from notBefore. I only changed it so the summary survives the re-queue. I did not check how the card looks in a browser.
• Docs: I added a short note, docs/features/pause-gives-up-slot.md.

Earlier tries on this issue failed on the check's CPU limit across shards, not on a test. This push gives the check a new run.

I left out a change to how the resume time is worded on the card, and any retry or backoff policy.

To try it, run bun test test/crew-pause-slot.test.ts.

Closes #11564

A crew run paused for credit or a rate limit no longer holds the forge's only slot, and its log line now names the cause. The new test (test/crew-pause-slot.test.ts), bun scripts/gate.ts and bun x tsc --noEmit passed here. I did not run the full suite, so the forge's check is the first full run.

• Cause in the log: the run paused line in src/crew/dispatch.ts now carries cause= with credit, rate limit, network or other. pauseCause in src/crew/pacing.ts works this out from the pause summary the run wrote, and src/crew/runner.ts now passes that summary through.
• Slot: a paused run was already cancelled and re-queued with a notBefore time, so it holds no concurrency slot while it waits. The test shows this at concurrency 1: run A is paused until an hour from now, run B is queued, and B starts while A stays queued with its resume time and its provider wording intact.
• Card: the run card already shows the provider's wording and the resume time from notBefore. I only changed it so the summary survives the re-queue. I did not check how the card looks in a browser.
• Docs: I added a short note, docs/features/pause-gives-up-slot.md.

Earlier tries on this issue failed on the check's CPU limit across shards, not on a test. This push gives the check a new run.

I left out a change to how the resume time is worded on the card, and any retry or backoff policy.

To try it, run bun test test/crew-pause-slot.test.ts.

A paused crew run now gives up its slot, and its run paused log line names the cause (credit, rate limit, network or other). I removed docs/features/pause-gives-up-slot.md, which the "pipeline first" gate rejected. It was my own addition and the fix doesn't need it. bun scripts/gate.ts now passes here. I did not re-run tsc or the tests after removing the file, and the forge's check will run them on this push. The test passed on the previous push.

• Cause: pauseCause in src/crew/pacing.ts reads the cause from the pause summary the run wrote. src/crew/dispatch.ts adds it to the log line, and src/crew/runner.ts passes the summary through.
• Slot: a paused run is cancelled and re-queued with a notBefore time, so it holds no concurrency slot while it waits. test/crew-pause-slot.test.ts covers this at concurrency 1: run A is paused until later, run B starts now, and A stays queued with its resume time and wording.
• Card: the run card already shows the provider's words and the resume time. This change only keeps the summary through the re-queue. I did not check how the card looks in a browser.

I left out any change to how the resume time is worded and any retry or backoff policy.

To try it, run bun test test/crew-pause-slot.test.ts.

A paused crew run now gives up its slot and its run paused log line names the cause. The two outage tests in test/crew-run-timing.test.ts were red, and I fixed them. I did not run red-main.test.ts as a shard-mate of anything, so I can't say whether that failure is flaky. It passed here and failed in the check only beside other tests.

• Cause of the red tests: the queued re-run of a paused run now carries the provider's wording, so its summary starts with paused:. The tests' wait-for-settle filter matched on that prefix, picked the queued re-run, and waited 60 s for it to settle.
• Fix: the filter now also requires state === "cancelled", so it picks the paused run itself. This was a bug in the test, not the code.
• Checked: crew-run-timing and crew-pause-slot pass and tsc is clean. I did not run the full suite.
• Red-main test: test/red-main.test.ts ("files no sentinel issue for it…") failed in the check only beside other tests, and it passed here beside crew-run-timing. It looks unrelated to this change.

What changed

• src/crew/pacing.ts, src/crew/dispatch.ts and src/crew/runner.ts add cause= (credit, rate limit, network or other) to the run paused log line and keep the provider's words on the re-queued run.
• A paused run is re-queued with a resume time, so it holds no slot while it waits. test/crew-pause-slot.test.ts covers this at concurrency 1.
• The run's card already shows the provider's words and the resume time.

I left out any change to how the resume time is worded and any retry or backoff policy.

To try it, run bun test test/crew-pause-slot.test.ts test/crew-run-timing.test.ts.

All four points from the review are addressed, and the branch is pushed. I ran only the tests covering this change, not the full suite, and the forge's check runs everything on this push.

Changed

• Card (box 3): the card said only "paused, resumes soon". It now reads paused, resumes 07:21 UTC from the run's resume time, on the Runs list and the board. The change is in src/web/studio/words.ts and src/web/studio/lists.ts. I updated the runs-working-tab assertion to match.
• Cause matching: pauseCause now matches whole words only (\b…\b), so "generate" and "accurate" no longer read as a rate limit.

Tests test/crew-pause-slot.test.ts now drives a real pause. The stand-in provider turns run A away with its session limit, with concurrency 1.

• Log line: it captures the run paused line for run A and asserts cause="rate limit". The value is quoted because it has a space. Removing the cause: field from dispatch.ts fails the test.
• Provider words: the re-queued run's summary equals the paused run's summary, paused: the 5-hour window is full…. Removing the summary that runner.ts passes back fails the test.
• Slot: with A paused, B starts and finishes while A stays queued.
• Card: the signed-in Runs page shows paused, resumes <A's notBefore time> UTC.

I disagree on one point. The credit cause is covered only by the pauseCause unit assertion. The stand-in has no real path that produces a credit pause, and I did not add a stand-in branch just for that. The log line and summary tests use the rate-limit path.

The earlier timing-test edit (waiting on the cancelled run) is still needed and stays.

The pre-push hook asked for a "how to try it" commit, so there is an empty Try: commit on the branch.

Left out: any change to retry or backoff policy.

To try it: sign in and open /<owner>/<repo>/runs?state=queued. A run paused on a limit reads paused, resumes HH:MM UTC. Run bun test test/crew-pause-slot.test.ts.

A crew run held back by a credit or rate limit now says the date it resumes when that is not today. The terminal's run list also says the run is paused and when it resumes, in the same words as the web card. This covers all three review points.

• The day on the card: pausedNote moved to src/crew/pacing.ts. It uses a new resumeText(at, now) that gives 07:21 UTC when the resume falls on today's UTC date, and 2026-11-19 07:21 UTC when it does not. The shared whenText is unchanged, because other pages and many tests depend on its exact output.
• The terminal: a paused run's row in the run list (src/tui/screens/runlist.ts) now reads paused, resumes …, from the same pausedNote as the web card. A queued run with no resume time still shows its reason or summary as before.
• Formatting: the ternary branch in src/web/studio/lists.ts is now indented like the other branches.

Tests:

• The new test/paused-note.test.ts checks the time-only and with-date forms and the fallback for an unreadable time.
• It also checks that the terminal row shows the paused note with the date for a pause three days away, which would fail without this change.
• test/crew-pause-slot.test.ts now compares the live card against pausedNote, so it can't fail when the pause crosses midnight UTC.

Gate, type check, design lint and the run-list, pacing and pause tests all pass. The card was not checked on the preview.

To try: sign in and open /<owner>/<repo>/runs?state=queued, or the runs screen in the terminal. A run paused until a later day reads paused, resumes YYYY-MM-DD HH:MM UTC.

🤖 Generated with Claude Code