---
name: babysit
description: Monitor a PR until ready to merge. Use when asked to babysit a PR, watch CI, or keep handling review feedback.
---

Poll CI, comments, and unresolved review threads together; handle review feedback without waiting for CI to finish. A push is never the end of the task; resume watching.
Keep review fixes within the PR's original scope. Treat suggestions to add features or unrelated improvements critically; even if worthwhile, get explicit user approval before implementing them.
Verify findings against the latest code. Fix real issues within scope, including cheap nits; run relevant tests, commit, push.
Before every push, fetch origin and rebase onto origin/main; resolve conflicts and run all required checks on the rebased code before pushing.
For false positives, reply with a one-line reason and resolve the thread. Escalate valid findings you would rather defer, and subjective or architectural feedback, instead of dismissing them.
Resolve threads only after verifying they are addressed; outdated does not mean resolved.
Retry confirmed infra flakes up to twice, then report the blocker; if a check also fails on main, report it as a blocker.
Prefix comments you post with "<model slug> on behalf of <name>".
Before declaring readiness, freshly confirm that CI passes, required reviews are satisfied, no automated reviewer is still running (check their status comments and reactions), the PR is mergeable, and no comments or threads remain unaddressed. Check recent merged PRs for the review bots this repo uses; a bot that has not posted yet has not started, which is not a pass.
Stop when ready, merged, closed, or blocked on a human decision; briefly report the outcome or blocker.
Do not merge unless explicitly asked.
