Agent skill · mia-cx
babysit
Use when the user asks to babysit a pull request: watch it for review comments from bots and human reviewers, fix what is real, resolve the threads, and keep watching until it is green and mergeable. A request for a fresh review pass or a review loop is review-relay, not babysitting.
What it needs
About 4k tokens when loaded.
What this skill does
Babysit a pull request Fix review findings until the current head is clean, green, and mergeable, then stop. The reviewers are whoever already posts on it: Pullfrog, other review bots, and humans. You are the fixer. Babysit runs no reviewer of its own and never invokes review-relay. The relay recruits a fresh reviewer every lap; babysitting reacts to the ones already on the PR. If the PR needs another opinion, the user asks for it separately. Setup 1. Identify the PR: the number the user gave, else gh pr view --json number,url,headRefName,baseRefName,headRefOid from the current branch. gh repo view --json nameWithOwner gives OWNER/REPO. 2. Baseline: run the project's checks (tests, typecheck, lint, build) and record what passes and what already fails. A fix is only verifiable against a recorded baseline. 3. Read the verification bar once: the sections "What counts as real", "The trigger test", and "Fixing" in review-relay. Those govern which comments become code changes. The rest of that skill (lineup, legs, relay log, reviewer prompt) stays closed. 4. Run one tick over the backlog: every unresolved review thread and every unanswered conversation comment already on the PR. 5. Check Done before starting the watch. If already done, report once and end the turn. The watch Report findings as currently known, because later reviews can raise more after every fix. Say "one finding is currently open"; never say "one remaining" or imply the next fix ends review. Claim completion only when Done is verified. One line per event: a new conversation comment, inline comment, or review; CI changes; the head moving. green <sha>, merged, and closed are terminal events: the script exits and clears its state. On green, finish the current tick, verify Done, report once, and end the turn. Bodies carrying the gh-comment attribution header are skipped, so your own replies never re-trigger a tick. The script decides when to look. …
How to use it
Reference it in AdaL, Claude Code, Cursor or any coding agent — nothing to install:
@skills mia-cx/babysit