Babysit

Babysit a PR (nubjs/nub)

Loop: pull reviews → verify each finding → fix → re-verify locally → push → repeat. CI comes ONCE, at the end: PR CI is opt-in, so a push starts nothing and the run is requested with gh pr edit <n> --add-label ci when the head is final.

Babysit ends at merge-ready. Report that state and stop; the merge is the maintainer's. Only an instruction to babysit it to merge authorizes one — and that phrasing delegates the readiness judgment too, so decide against the criteria below instead of asking.

Reviews

Two bots review within minutes of every push: pullfrog, which posts a top-level review plus inline threads with a <details> write-up per finding, and copilot-pull-request-reviewer. The maintainer outranks both. A PR carrying no review was merged before they ran rather than approved by them, and only the review list distinguishes those.

The gh pr view --json reviews,comments form silently omits inline pulls/{n}/comments, which is where the actionable findings are. Use scripts/pr-reviews.ts — one GraphQL round trip for reviews, inline threads (id, resolved/outdated state, file:line) and PR conversation:

nub scripts/pr-reviews.ts <pr> | jq '.reviewThreads[] | select(.isResolved | not)'

Triage

Skip resolved and outdated threads first — an outdated one anchors a line a later push already rewrote.

Every finding is a hypothesis. Verify it against the code before acting; a confident write-up is the one that gets acted on unchecked. Null guards, "just in case" branches, and try/catch around code that cannot throw are the slop shape AGENTS.md names — reject those by default.

Reject with one factual line on the thread saying what you checked. Never ignore silently, and never reply conversationally to a bot. Load the prose-writing skill before writing any comment. Resolve each thread you address:

gh api graphql -f query='mutation($id:ID!){resolveReviewThread(input:{threadId:$id}){thread{isResolved}}}' -F id=<thread-id>

Fix rounds

A review-driven fix earns the same gate as the original change. Run make verify. For a behavior change, sweep adversarial fixtures against the branch's MERGE-BASE build (ad-hoc-test) — a run that only re-confirms the reported case tests your intent, not your fix. Escalate to the impact-analysis skill when the fix widened the blast radius. A one-line typo fix skips both.

Then push. Pushing is free — PR CI is opt-in and nothing fires until it is asked for — so push each round as it lands and hold the CI request until the review loop is done. A requested run is roughly five workflows across eight platforms, and asking round after round is what starves the shared runner pool.

CI

Look at gh pr checks <N> before asking. The label is stripped within a minute of being added, so an empty label list says nothing about whether a run was requested, and a second request for the same head cancels the run already in flight — a full matrix thrown away (2026-09-19, #959: the maintainer had labelled it half an hour earlier, and the babysitter's request cancelled fifteen jobs). Checks already pending or running on the current head mean the run exists; watch that one.

Ask for the run, then watch it. Both steps — the request is not optional, and until it lands the pull request has no checks at all:

gh pr edit <N> --add-label ci      # requests CI against the current head; the label is removed again immediately
nub scripts/ci-watch.ts --pr <N> --required "CI gate" --timeout 90

Push a further commit after that and the run you were watching is stale — its checks belong to the old head, so ask again. ci-watch will not cover for a forgotten request: an empty or all-skipped rollup reports as pending, never green.