PR #23033 - manaflow-ai/cmux

Summary

nightly-next has been stuck at ff8d849 since 01:44Z on 2026-10-21, and every run involved stayed green:

No promotion was requested after 02:59Z. Every nightly-next tick (e.g. run 39550209508) walked the 120 newest feat-cmux-next commits and refused all of them. 87 had a published cmux-tui tree whose macOS daemon tests never passed (no daemon-tests.json), and 33 had no published tree. The feat-cmux-next daemon-test jobs sit queued on glaeda-side-std-xcode-26.6 (165 queued, 43 of 44 online runners busy at 05:50Z). The oldest still-queued job dates from 04:07Z. The pool fix is separate (huge-blue/release-rings owns it).
The 03:13 to 05:33 nightly.yml runs on nightly-next are only continuation ticks (CONTINUE_ONLY, "already published"). They are not promotions.
The feed watch opened #23013 only at 4h, and the issue said just "Nothing was promoted".

This PR makes that stall loud and gives its reason:

nightly-next-feed-watch: the default limit drops from 4h to 3h, and a stale check now exits 1, so the scheduled run is red every 30 minutes until the feed catches up. The issue still opens as before. A GitHub API error still only warns.
nightly-next-tick: when it requests nothing, it adds a nightly-next tick warning annotation with a tally of the refusal reasons (for example 120 refused: 87x the daemon tests have not passed for tree <sha> yet …; 33x the nightly cannot resolve a published cmux-tui tree for it yet). It also does this when the walk reaches the already-promoted commit, a path that used to exit 0 silently.
The feed watch reads that annotation from the latest tick run (three API calls, only during a promotion stall) and adds it to the stale reason, so the issue and the red run both say why.

This PR is against main because schedule and workflow_run workflows run only from the default branch. #22669 also edits nightly-next-tick.yml, but only the step if: lines, so the two should merge cleanly.

Testing

Observed

Ran the tick's tally pipeline on the real log of tick run 39550209508. It printed the warning quoted above.
actionlint passes on both workflows.

Not run

tests/test_nightly_next_feed_watch.py (new cases: tick refusal quoting, annotation lookup, exit 1 when stale and 0 when current, the 3h default) was not run locally because the cosmic-gray hook keeps tests off Leo's Mac. It runs in this PR's CI.

Changelog

none

Agent session: cosmic-gray

🤖 Generated with Claude Code

^{Need help on this PR? Tag @codesmith-bot with what you need.}

Summary by cubic

Makes the nightly-next stall alarm fail loudly and name the cause, instead of staying green with a vague "Nothing was promoted" issue.

The feed watch's stale limit drops from 4h to 3h, and a stale check now exits 1 so the scheduled run turns red.
When the tick requests nothing, it annotates a tally of refusal reasons, including when the walk reaches the already-promoted commit (a path that used to exit 0 silently). The feed watch reads that tally from the latest successful tick run and quotes it in the issue body and failing annotation, never in issue comments.
The repo variable check now fails when CI_NEXT_DAEMON_POOL equals CI_SIDE_LANE_RUNNER — the root cause of the stall, where daemon tests queued behind pull request side lanes.
The new feed-watch tests now run in the release-notary guard job.

Summary by CodeRabbit

Bug Fixes
- Stale nightly feed checks now fail when updates lag by more than three hours, rather than only issuing a warning.
- Stale issue handling retains a single issue to reduce duplicate reporting.
Improvements
- Promotion stalls now include clearer refusal counts, commit details, and relevant workflow information.
- Nightly promotion checks scan additional commits and stop promptly when promotion is requested or already completed.
- Warnings are more specific when no promotion can proceed.
- CI validation now detects when daemon tests share the pull-request runner, helping prevent queueing delays.