watch_pull_request never wakes when the remaining checks finish after a required check failed #25497

What happened

An agent was babysitting a stack of pull requests with watch_pull_request. On the first PR, two label checks failed right after the push and the agent was woken for them. In this repository those checks are "missing labels" checks. Some PRs must be checked by a human, who adds a label to let the PR continue, so the agent treats these failures as expected and waits for the real CI jobs. The real CI jobs finished green about six minutes later, but no wake ever arrived. The watch stays armed, the thread stays in Working, and the agent's last message is "still waiting for the main CI jobs". From the user's view the watch hangs.

Diagnosis

The server is healthy. PullRequestWatchReactor sweeps the PR every two minutes with no errors, the thread has no running turn, and the stored watch state matches GitHub.

evaluatePullRequestWatch in apps/server/src/orchestration-v2/pullRequestWatch.ts only produces four changes: checks-failed for newly failed checks, checks-passed when every gate check is neither pending nor failed, remarks, and conflicting. On this PR:

• "[C1] Check BC labels" and "[C2] Check labels" failed and were reported at 13:16. C2 is a required check.
• All other checks finished green by 13:22.
• No check is newly failed, so checks-failed does not fire again.
• The gate is the required checks. C2 is failed, so passedNow is false and checks-passed can never fire for this head.

So the completion of the remaining checks is not an event. The only things that could still wake the thread are a new comment, a conflict, or a push. The wake message tells the agent "T3 Code keeps watching and wakes you on the next change", so the agent reasonably ends its turn and waits forever.

main adds owedChecks to wake once non-required checks settle after the required gate passed. That path requires failed.length === 0, so it does not cover this case either.

Proposed behavior, in the order the reporting user prefers:

1. When every check has settled and the only failures were already reported, send a wake such as "All checks finished on df55d0b: N passed, M failed (already reported)". Since nothing on the PR can change until a human acts, this wake could also end the watch, like the comment limit does. A later label edit would then need a new watch_pull_request call.
2. At minimum, the checks-failed wake should state that no further wake will come for the other checks unless one of them fails or every gate check passes.

Steps to reproduce

1. Use a repository where a required check fails until a human adds a label, or any required check that fails fast.
2. Push a branch and open a PR. Make sure other, slower checks also run.
3. In a thread, have the agent call watch_pull_request on the PR and end its turn.
4. Wait for the fast required check to fail. The thread is woken with "Checks failed".
5. Let the agent end its turn and wait for the remaining checks.
6. Wait for every remaining check to finish green.

Expected: a wake telling the agent that all checks have finished.
Actual: no wake. The watch stays armed and the thread stays in Working indefinitely.

Version

0.1.2 (desktop app, local server reports serverVersion 0.1.2). Triage CLI was also 0.1.2.

Environment

macOS 27.0.0 arm64, Node v24.22.0, claude 2.1.310, gh 2.103.2. Host: GitHub.

Evidence

# stored watch on the thread (statev2.sqlite, orchestration_v2_projection_threads.payload_json)
"watch": {
  "startedAt": "2026-10-26T13:13:08.221Z",
  "headSha": "df55d0ba722263bb59fb80cac3d5ef40f60d3fa1",
  "failedChecks": ["[C1] Check BC labels", "[C2] Check labels"],
  "passed": false, "passedChecks": [], "conflicting": false, "wakes": 1
}
snapshot.checksState: "failing", mergeability: "mergeable", state: "open"

# gh pr checks --required
[C2] Check labels   fail   4s
[T1] Behat (1 8)    pass   7m57s
... 10 more required checks, all pass

# thread messages
13:13:08Z  thread.pull-request.watch dispatched (server:mcp-pr-watch:...)
13:16:02Z  wake "#11168: checks failed"  -> [C1] Check BC labels, [C2] Check labels
13:16:05Z  agent: "Both failures are the reviewer label checks ... I'll wait for the next wake."
13:20:07Z  wake "#11168: new comments" (socket-security)
13:20:14Z  agent: "Still waiting for the main CI jobs on PR 11168."  (last run, status completed)

# server.trace.ndjson, every sweep after that, all exit Success, no warnings
13:22:07Z PullRequestWatchReactor.sweep 1815ms / readGroup 1025ms
13:24:09Z PullRequestWatchReactor.sweep 2029ms / readGroup 1135ms
13:26:11Z PullRequestWatchReactor.sweep  758ms / readGroup 0ms
13:28:12Z PullRequestWatchReactor.sweep  691ms / readGroup 0ms

Related issues

#23701 (never wakes when a PR has no checks, fixed on main by the no-checks wake) and #24575 (flaps on action-required for fork PRs). Neither covers a failed required check blocking the completion wake.

Fix applied or workaround

None applied. The user can click "Stop watching" in the thread's pull request panel or send the thread a message to wake the agent.

Filed by

Claude Fable 5.1 (claude-fable-5-1) via Claude Code, through t3 triage.