| name | ci-watch |
| description | Watch a PR's CI hands-free after a push. A persistent Monitor streams each check as it lands; the skill auto-handles the mechanical (rebase a clean conflict, fix lint) and STOPs on any judgment (failing tests, a stack restack, going ready). Triggers on "watch CI", "surveille la CI", "attends la CI", or right after /pr or a push. |
| argument-hint | [pr-number|url] |
| disable-model-invocation | true |
CI Watch
Watch the current PR's checks without babysitting them by hand. A persistent Monitor streams every state change while the session keeps working; you react per the classification table below.
The boundary, one line: act without asking only when the action is mechanical (reversible, no decision) — rebase a clean conflict, fix lint, push, gh pr ready. Every judgment (which finding to fix, a red test, restacking a stack, the final go) is a STOP: surface it and hand back.
CI has three states, and each routes differently: green (advance), red (classify the failure), conflict (rebase or restack).
Step 1 — Resolve
Find the PR to watch:
gh pr view --json number,url,baseRefName,isDraft
If there is no PR for the branch and no argument, stop and say so — there is nothing to watch.
The Monitor and notification tools are deferred; load them before arming:
ToolSearch "select:Monitor,PushNotification"
Done when: the PR is resolved and Monitor + PushNotification are loaded.
Step 2 — Arm
Arm a persistent Monitor that polls the PR's checks and mergeable state and emits one event per state change, so the session can keep working while checks land:
gh pr checks <pr-url>
gh pr view <pr-url> --json mergeable,mergeStateStatus
Poll no faster than every 30s (API rate limits). Read the loaded Monitor schema for how to phrase the until-condition and interval.
Done when: the Monitor is running and has emitted its first snapshot.
Step 3 — React
Route each event through the table. Do not declare green on the first passing check — wait until every required check is terminal (none pending / in_progress), or a premature "all clear" will fire while checks are still running.
| Signal | Action | Nature |
|---|
| all required checks green | PushNotification, go to Step 4 | mechanical |
| conflict · isolated branch (not in a stack) | git rebase <baseRefName> + git push --force-with-lease | mechanical |
conflict · branch in a git-spice stack (gs ls lists it) | propose gs stack restack + cascade push — one confirmation | judgment |
| conflict with leftover markers to resolve | STOP, hand back | judgment |
| red · lint / format only | deterministic fix (prettier --write / eslint --fix), commit, push | mechanical |
| red · test / typecheck | /pr-feedback to triage, surface the P1, STOP | judgment |
| red · a known-flaky check (list below) | re-run once, then escalate | waiting |
Known-flaky allow-list (the only checks a red state may auto-retry — everything else red stops):
- i18n
_latest in the merge queue — fix: run i18n:pull, then re-run the check.
Done when: every required check is terminal and its row has been actioned.
Step 4 — Green
The PR is green and ready for a human review.
PushNotification — "PR ready for human review."- If the PR is still draft, ask "mark it ready?" (STOP — this is a judgment). On yes:
gh pr ready <pr-url>.
- Never request a review on GitHub and never post to Slack. Soliciting the team is the user's call.
Done when: the notification is sent and, if approved, the PR is marked ready.
