Repository navigation
Description
Before submitting
- I searched existing issues and did not find a duplicate that covers this UI mismatch.
- I included enough detail to reproduce or investigate the problem.
Related: #37477 (same 2-minute PullRequestWatchReactor sweep lag; that report is GitHub-focused). This issue is the user-visible Waiting vs green-sidebar mismatch on GitLab, with measured lags on t3 v0.1.10.
Area
apps/server (also client status: Waiting · Watching pull request)
Summary
When an agent is babysitting a GitLab MR with watch_pull_request, the PR sidebar can show green checks / Merge as soon as CI finishes, while the thread footer stays Waiting · Watching pull request #N for up to ~2 minutes until the next PullRequestWatchReactor sweep. Only then does checksPending clear and the agent get a checks passed wake.
So from the user's point of view the thread looks stuck on finished CI. The watch does eventually wake; it is late relative to the sidebar.
Why (from local 0.1.10 binary + server.trace.ndjson)
PullRequestWatchReactor:SWEEP_MINUTES = 2. NoPROBE_SECONDSpending-check probe in this build.- Sidebar / PR detail updates via
PullRequestSyncReactor+pullRequests.detailon a faster path. - Thread
Waitingis driven by pending background work: a PR watch task is{ awaited: true }unlesslink.watch.checksPending === false. That flag only flips when the watch reactor records a read — same cadence as the agent wake.
Steps to reproduce
- On t3 v0.1.10, link/watch a GitLab MR whose pipeline takes several minutes.
- Have the agent push, call
watch_pull_request, and end the turn. - When the pipeline goes green, open the thread: note sidebar green check vs footer
Waiting · Watching…. - Compare GitLab
pipelines/:id.finished_atwith the thread'spr-watch/checks passednotification time.
Observed timelines (UTC, machine on)
| MR / pipeline | CI finished_at |
Sidebar-ish sync/detail | Agent wake (checks passed) |
Lag |
|---|---|---|---|---|
tax-office MR !46 / pipeline 4155 (6ac85c6) |
08:01:26 | PullRequestSyncReactor / pullRequests.detail around 08:01:22 |
08:03:22 | 115s |
same MR / pipeline 4151 (2a6ee60) |
07:41:59 | — | 07:42:58 | 59s |
keycloak MR !3 / pipeline 4156 (ce44dce) |
08:03:53 | — | 08:05:23 | 90s |
Screenshot case (4155) watch reactor sequence:
07:43— agent rebased/pushed6ac85c6, left watch armed, ended turn08:00:00—PullRequestWatchReactor.sweep(CI still running; this sweep ran long)08:01:26— GitLab pipeline success08:01:22+— sync/detail already reflecting green (sidebar)08:03:10— next watch sweep08:03:22—PullRequestWatchReactor.record+ wake
Expected behavior
Once CI is terminal on the watched head, either:
- the watch notices within seconds (pending-check probe / host events), and/or
- the thread leaves
Waitingas soon as the UI already knows checks are done (same freshness as the sidebar),
so the footer does not show Waiting · Watching… against a green Merge-ready PR for ~1–2 minutes.
Actual behavior
Sidebar: green. Footer: still Waiting. Agent idle until the next 2-minute watch sweep.
Impact
Looks like a hung babysit. Users interrupt or re-prompt while the watch is merely waiting for the sweep. On GitLab there is no GitHub event-stream shortcut, so every pass/fail pays the sweep lag (#37477).
Environment
- t3 v0.1.10 (WSL runtime)
- Host: GitLab (self-hosted),
glabauth - Threads observed: babysit / full-send style watches that push, wait for green, then merge/tag
Logs
server.trace.ndjson span names around the mismatch: PullRequestSyncReactor.sweep / getMergeRequestDetail / ws.rpc.pullRequests.detail vs later PullRequestWatchReactor.sweep → record → orchestrationV2.dispatch.pullRequestWatchSync.
Workaround
Wait up to ~2 minutes, or manually prompt the thread after CI is green.
Possible direction
Same as #37477 / #37492 / #37494: probe fingerprints while checksPending between sweeps. Optionally drive checksPending / Waiting from the sync path when it already sees a terminal pipeline, so the UI cannot disagree with itself.
Activity
tiborrr commentedon Nov 20, 2026
Screenshot note
I could not attach the PNG via
gh gist(binary rejected). The captured UI showed:Waiting · Watching pull request #46#46with Merge enabled6ac85c6after a prior green runHappy to attach manually if maintainers want the image on the issue.
Also linking the sweep-lag parent: #37477
tiborrr commentedon Nov 20, 2026
juliusmarminge commentedon Nov 20, 2026
This is the same problem as #37477, so I'm closing it in favour of that one. If it is different, say how and it will be reopened.