Babysit a PR

Keep watching until the PR merges or closes, the user cancels, or another caretaker accepts the handoff. Readiness is a checkpoint, not the end of the watch. Never merge or enable auto-merge.

The caretaker needs a session that can run a background command, because the watcher below wakes it. One caretaker can own several related PRs. If another teammate's agent is still writing to a branch, don't push to it; send that agent the finding through Coordigent instead.

Assess and fix

Fetch the current PR state through gh. Use PR state queries for paginated review threads and required checks. Record the head SHA. CodeRabbit reviews each push in this repository; its finished review of the current head is the bot's completion signal. When it reports a rate limit or skips the review, don't wait for it.

Check conflicts against the actual PR base, diagnose failing checks, and read full review discussions. Validate each finding against current code and behavior before fixing it. Distinguish defects from stale findings, false positives and unrelated improvements. Keep uncertain findings visible.

Fix supported, in-scope defects, run affected checks, commit and push normally. Reply with the fix SHA or a concise reason for disagreement or deferral. Do not resolve threads merely to pass this workflow. Send substantial fixes to the existing implementer; caretaker edits require writer ownership.

After every push, refresh the head, required checks and review threads. Changed heads invalidate readiness. After three unsuccessful fix rounds on a PR, pause further fixes and report the blocker for a decision. Keep watching while awaiting that decision; unchanged feedback does not authorize another fix round.

Watch

Watch every assigned PR with one shared watcher, started through Bash with run_in_background:

node .claude/skills/babysit/scripts/pr-watch.mjs SamGu-NRX/spurs-hackathon#12,14

It makes one GitHub query for all targets every three minutes and exits only on a change you must act on: a check failing, all checks finishing, new reviews or comments, a changed unresolved-thread count, head, base, draft or merge state, a merge or closure. Its exit wakes the session, so an idle PR costs no turns. Do not watch PRs with Monitor, which stops after 30 minutes and wakes you to re-arm it, costing a full turn each time. Do not write your own watcher either.

End every turn with the watcher running; never sleep or poll inside a turn. After acting on its output, start the same command again. Its baseline persists per working directory, so a change that lands while you work is still reported. Your own push or reply also counts as a change: start with --rebaseline only immediately after you have re-read the PR state yourself.

Other exits: ALL_DONE means every target is merged or closed, so stop. NO_CHANGE after 23 hours means start it again. COVERAGE_LOST, exit 2, means GitHub failed for 30 minutes: report lost coverage, then start it again. A not-found target means fix the list.

Keep the watch active after readiness and while fixes await a decision. Remove merged or closed PRs from the targets. On cancellation or an accepted handoff, stop the background task and tell the new caretaker which targets it now owns.

Readiness and reports

When fixes change the PR's scope, behavior or evidence, update its description under pr-writing. Preserve the body's author ownership and bot-managed sections.

Recommend human review only when the PR is open and not draft, its current head passes required checks, the configured bot has finished for that head, no unresolved threads or actionable feedback remain, and the merge state is clean. Required human approval still applies. Pending or unknown evidence is not a pass; assess skipped checks under repository policy.

Report readiness once and continue watching. Report again only for a meaningful change, a decision needed or terminal state. Include relevant SHAs, finding dispositions and remaining blockers without repeating prior history. A readiness report must say the watch remains active; a coverage failure must say it does not.

Boundaries

A babysitting request authorizes in-scope fixes, commits, pushes and review replies. Ask only for a user-owned decision or an action outside existing authorization. Rebase or force-push requires explicit authorization; conflicts do not grant it.

Never bypass hooks or required checks, weaken tests to obtain a pass, mark a PR ready on your own, or reopen a closed PR.

ship-pr hands off open PRs when babysitting is assigned.

Adapted from Andrew Wang's astack babysit skill.