babysit

Babysit

Keep an open PR moving toward merge-ready: watch its checks, review threads, and mergeability; act on what blocks them; and report evidence. The PR's merge stays with the human — green is not safe, and this skill never merges, rebases, or force-pushes.

When to use

• "get this PR green", "babysit PR #N", "address the review-bot comments"
• the user wants a PR they opened earlier driven to merge-ready while they are away

When not to use

• merely opening a PR (no monitoring is implied)
• the user asks to land or ship a green stack — that decision belongs to the user; report merge-readiness and stop

Loop

Declare the mode before the first poll: which PR, and whether bot comments are in scope.

1. Read state via gh: gh pr view <N> --json state,mergeable,mergeStateStatus,reviewDecision,isDraft,headRefOid,headRefName and gh pr checks <N> --json name,bucket (bucket is pass/fail/pending/skipping/cancel). Normalize before classifying: mergeable arrives uppercased (MERGEABLE/CONFLICTING, UNKNOWN or empty while GitHub computes it), and reviewDecision is an empty string when no review exists — not none. Exit codes: gh pr checks --json exits 0 even when checks fail — judge by bucket, never the exit code; with no checks configured it exits 1 printing "no checks reported", which is the no-checks path, not an error. Treat aggregate mergeStateStatus=BLOCKED as blocked and UNKNOWN as waiting before allowing ready. Confirm the first read reports the requested PR before acting on it.

2. Apply the watch-pr verdict policy (src/features/watch-pr/policy.ts semantics, embedded below). Do not eyeball; classify. Determine reviewRequired from the repository's branch-protection settings (for example, gh api repos/OWNER/REPO/branches/BRANCH/protection and the required_pull_request_reviews field); a ruleset that requires reviews independently of classic branch protection counts the same way. Pass that boolean into the policy; do not treat an empty review decision as a blocker when the branch does not require approval.

3. Act on the verdict — every lane gets the polled identity: pass the repo and PR number explicitly and verify the checkout sits on that PR's head before mutating anything. Bare gh pr checks/gh pr view select whatever PR owns the current branch, which need not be the one babysit polls.

   • failing check → fix-ci (classify real/flaky/environmental, then fix; pass the target repo, PR number, head SHA, and base branch)
   • requested changes or unresolved review threads → get-pr-comments on the polled repo and PR number — triage threads, then address each through the review lane; reply to threads you resolve; dismiss bot noise with a concrete reason instead of churning code. Read the threads surface (comments/reviews via the get-pr-comments fetch) even when reviewDecision is empty or approved — a non-blocking review or a comment after approval leaves the decision clean while real threads stand — and never report ready while an in-scope thread is untriaged
   • conflicts → fix-merge-conflicts in its no-rebase mode, run on the polled PR's owned checkout: resolve with a merge commit only. Never rebase and never force-push here — this constraint overrides that skill's own-branch rebase option, because babysit's contract forbids both. If the branch is not ours to resolve, report to the branch owner instead
   • mergeability unknown → wait one interval and re-read; it is never an error
   • draft or missing review → report what is missing; do not self-approve

4. Ledger discipline: watch verdicts (ready/waiting/blocked/gone) are gate states, not verification — never write them into the verification ledger, whose vocabulary is separate. Append a verification ledger row (the writer's canonical path under OMC's state root) only when you executed a check yourself — e.g. a fix-ci rerun went green (unit-test-verified, with the command as evidence). One row per executed check, never per poll. Write it through the repository writer so header creation, locking, and validation are enforced. Resolve the installed package root the figure-it-out way: plugin-root variables first, then walk the installed oh-my-claudecode package from the working directory — and report the resolution failure instead of skipping the write:

   node --input-type=module -e "const {createRequire}=await import('node:module'); const path=await import('node:path'); const {pathToFileURL}=await import('node:url'); let root=process.env.CLAUDE_PLUGIN_ROOT||process.env.OMC_PLUGIN_ROOT; if(!root){ const req=createRequire(process.cwd()+'/package.json'); for (const name of ['oh-my-claude-sisyphus','oh-my-claudecode']) { try { root=path.dirname(req.resolve(name+'/package.json')); break; } catch {} } } if(!root) throw new Error('cannot resolve the installed package root; set CLAUDE_PLUGIN_ROOT or OMC_PLUGIN_ROOT'); const {appendLedgerEntry,defaultLedgerPath}=await import(pathToFileURL(root+'/dist/features/verification-ledger/index.js').href); await appendLedgerEntry(defaultLedgerPath(),{ts:new Date().toISOString(),kind:'babysit',ref:'<N>',sha:'<HEAD_SHA>',verdict:'unit-test-verified',evidence:'<command>',verifier:'babysit'});"
   

   The writer owns the exact seven-column TSV layout; do not append raw rows.

5. Report: verdict, what changed since the last poll, what is still blocking, and the exact commands you ran. Every claim carries its evidence.

Poll with gh pr checks --watch or a sleep loop at a stated interval; state the interval. Stop when the PR is merge-ready, merged, closed, or the user says stop, and say which.

The verdict policy

Classification order, first match wins:

• conflicts, mergeStateStatus=BLOCKED, DIRTY, or BEHIND → blocked (merge conflicts or another aggregate gate; the branch owner decides the resolution)
• any failed check → blocked (fix lane)
• any canceled check → blocked (bucket cancel; rerun or clear it — a canceled check is not a green gate)
• changes requested → blocked (review lane)
• branch protection or rulesets require reviews, and approval is missing → blocked (needs a human reviewer; never self-approve)
• mergeability or mergeStateStatus=UNKNOWN or DRAFT → waiting (re-poll; not an error). HAS_HOOKS is not a wait: GitHub defines it as mergeable with passing commit status and pre-receive hooks — the hooks belong to the merge attempt, so an otherwise green PR with HAS_HOOKS classifies normally
• pending checks → waiting
• no checks configured → waiting until the same head SHA has shown no checks for a 60-second confirmation window; only then is the absence real. A first sighting of "no checks" is never readiness. Before serving that window, check whether the repo has PR-triggered workflows at all (.github/workflows/, or a run on another PR): when workflows exist but this head shows none — observed in the wild on path-filtered and skipped-event pushes — cross-check gh run list --branch <head>; a persistent zero-run anomaly is blocked (investigate: path filters, base mismatch, skipped events), not a no-checks repo.
• otherwise → ready, with the head SHA recorded in the proof

"Ready" means the gates you can see are green. It is not permission to merge, and it is not verification of the change itself — that evidence belongs to verify and the ledger.

Rules

• Never merge, never rebase, never force-push, never approve a review.
• A review-bot comment gets a skeptical read: real bugs are fixed, non-issues are dismissed with a concrete reason.
• One fix lane at a time; wait for its push before re-classifying.