Babysit

Stay on the PR after handoff until it's actually done. Being asked to babysit IS the ask to poll — but report state changes only, no heartbeats, no per-poll narration.

Loop

1. Snapshot: gh pr view <N> --json state,mergeable,mergeStateStatus,reviewDecision,statusCheckRollup,autoMergeRequest plus gh pr checks <N>.
2. Poll inside bounded shell loops (a single for with sleep 30–60 between iterations), not one message per check.
3. React by cause:
   • CI never started → kick it: git commit --allow-empty -m "chore: kick CI" && git push.
   • Red caused by this branch → fix it (the pr-ready loop), push, resume watching on the new SHA.
   • Flake / infra red → rerun the failed jobs; repeat until green. Note the flake in one line with the run link.
   • New review threads → handle per pr-ready (fix or reply, verify resolved via the forge API), resume.
4. Repeat until done. A push or rerun is progress, not an exit.

Merging

Arm automerge only when the ask says so — "set to automerge", "make CI green", or any mention of merging. A plain "babysit" / "watch it" ends at green with threads resolved: report, don't merge.

CI quirks

• gh run rerun refusing with "already running" → cancel the run first, then rerun.
• A cancelled run showing "no checks reported" is pending, not red — wait for it to re-report before calling it either way.
• Rollup gates ("All required checks pass") go red because a leaf did — diagnose the leaf failure, never the gate.

Done when

Re-query the forge — don't trust that a push went green or a rerun stuck. Merged, or every check green with no unresolved threads. Fail closed: any red or open thread means keep watching. Wrap up with the PR/MR link and what the watch consumed — reruns used, kicks, fixes pushed.

Scope

• In: watching an open PR, kicking stalled CI, rerunning flakes, small branch-caused fixes, late review threads.
• Out: opening or refreshing the PR (pr-update, cpr), full rebase/red-CI overhauls (pr-ready), other people's PRs (pr-triage), merging unasked.