Imported from jleechanorg/jleechan-skills (.claude/skills/pr-babysit/SKILL.md). Install upstream with npx skills add jleechanorg/jleechan-skills --skill pr-babysit. Copyright stays with the author.

/babysit — PR Green-Gate Babysitter

Purpose

Systematically drive all open PRs toward /green (CI green + no merge conflicts, see .claude/skills/pr-green-definition.md) plus draft-phase quality readiness (/es, /er, /advice, evidence, comment resolution). For each PR:

1. Audit /green (gates 1-2 below) and the draft-phase quality checks (gates 5-6 below)
2. Fix any blocking failures
3. Optionally surface CodeRabbit/Bugbot feedback (gates 3-4) for information — never wait on or require them
4. Once all blocking checks pass, run /smoke and report ready for human merge approval

Note (2026-07-07): the former Gate 7 (Skeptic PASS, an external bot VERDICT poll run by skeptic-cron.yml) was removed repo-wide — its async dispatch was unreliable and could block a PR's own CI for up to 30 minutes waiting on a comment that often never arrived (bead rev-p0cc4). skeptic-cron.yml was also the repo's only automated gh pr merge path; that capability is gone too. All merges now require explicit human "MERGE APPROVED" authorization per this repo's merge-safety policy — babysit's job stops at reporting readiness, not auto-merging.

Note (2026-07-28): /green itself is now ONLY gates 1-2 below (CI green + no merge conflicts). Gates 5-6 (comment resolution, evidence) are draft-phase quality gates that gate leaving DRAFT status, not /green — see .claude/skills/pr-green-definition.md. Gates 3-4 (CodeRabbit, Bugbot) are optional advisory reviewers — babysit may surface their feedback for information, but never waits on or requires them to consider a PR ready.

Draft-Phase + /green Criteria

Execution Protocol

Phase 1: Discover All Open PRs

gh pr list --state open --json number,title,headRefName,headRefOid,mergeable \
  --jq '.[] | "\(.number)|\(.title)|\(.headRefName)|\(.headRefOid[:12])|\(.mergeable)"'

Phase 2: Audit Each PR (/green + Draft-Phase Check)

For each open PR, check all 6 gates:

# Gate 1: CI status
gh pr view <NUM> --json statusCheckRollup \
  --jq '[.statusCheckRollup[] | select(.conclusion == "FAILURE")] | length'

# Gate 2: Mergeable
gh pr view <NUM> --json mergeable --jq '.mergeable'

# Gate 3: CodeRabbit review
gh api repos/$GITHUB_REPOSITORY/pulls/<NUM>/reviews \
  --jq '[.[] | select(.user.login=="coderabbitai[bot]")] | last | .state'

# Gate 4: Bugbot errors
gh api repos/$GITHUB_REPOSITORY/pulls/<NUM>/comments \
  --jq '[.[] | select(.user.login=="cursor[bot]" and (.body | test("error";"i")))] | length'

# Gate 5: Unresolved comments
gh api graphql -f query='
  query($owner:String!, $name:String!, $pr:Int!) {
    repository(owner:$owner, name:$name) {
      pullRequest(number:$pr) {
        reviewThreads(first:100) {
          nodes { isResolved }
        }
      }
    }
  }
' -f owner=jleechanorg -f name=your-project.com -F pr=<NUM> \
  --jq '[.data.repository.pullRequest.reviewThreads.nodes[] | select(.isResolved==false)] | length'

# Gate 6: Evidence in PR body
gh pr view <NUM> --json body --jq '.body' | grep -i -E "evidence|gist|video|mp4|N/A"

Phase 3: Categorize PRs

Group PRs by the four blocking gates (1, 2, 5, 6) — gates 3-4 (CodeRabbit, Bugbot) are informational only and never affect a PR's category:

• RED (2+ blocking gates failing): needs code fixes
• YELLOW (1 blocking gate failing): needs comment resolution or evidence
• GREEN (all 4 blocking checks passing — /green gates 1-2 plus draft-phase quality gates 5-6): run /smoke, then report ready for human merge approval

Phase 4: Fix RED PRs

For each RED PR, use subagents (Agent tool, subagent_type: copilot-fixpr) to:

1. Read the Green Gate workflow log to identify the exact failing gate
2. Fix the root cause:
   • CI test failures: read test output, fix code, push
   • Design Doc Gate: add design doc reference to PR body
   • Lint/type errors: fix code, push
3. Dispatch agents in parallel — one per PR or group of related PRs

Common fixes:

• Test assertion mismatch after rebase → rebase onto origin/main
• Missing evidence → create evidence bundle or add N/A justification
• Stale review threads → resolve via GitHub API

CodeRabbit/Bugbot feedback (gates 3-4) can be read for useful signal, but triggering or waiting on a re-review is never required to unblock a RED PR.

Phase 5: Fix YELLOW PRs

For each YELLOW PR:

1. Gate 5 (unresolved comments): Resolve review threads
2. Gate 6 (no evidence): Add evidence or N/A to PR body
3. Optionally comment @coderabbitai all good? to refresh CR feedback — informational only, not required

Phase 6: Run Smoke on GREEN PRs

For PRs at all 4 blocking checks (/green gates 1-2 plus draft-phase quality gates 5-6):

# Trigger Green Gate workflow if no recent run
gh workflow run green-gate.yml --ref <branch> -f pr_number=<NUM>

# Trigger smoke tests via PR comment
gh pr comment <NUM> --body "/smoke"

Once smoke passes, the PR is /green plus draft-phase-quality-complete and ready for merge — but babysit itself never merges. Report the PR as ready and wait for a human to give explicit "MERGE APPROVED" in the live conversation (see Anti-Patterns below).

Phase 7: Report Final Status

Print a summary table:

PR #<N> — age: <Xh Ym> — status: <red|yellow|green|/green>
  Blocking gates: 1=✓ 2=✓ 5=✓ 6=✓ | Advisory (informational only): CR=✓ Bugbot=✓
  Action: /green, ready for human MERGE APPROVED

PR Green Loop Protocol (MANDATORY)

• Batch all fixes into one commit, not one per finding
• Push once per batch of fixes
• CodeRabbit/Bugbot: optional advisory reviewers — read their feedback, take what's useful; never a gate, never a wait. Dismissing or ignoring their CHANGES_REQUESTED reviews is allowed; they never block a merge.

Subagent Dispatch Strategy

• Independent PRs: Dispatch one agent per PR in parallel
• Related PRs (same feature area): Group under one agent
• GREEN PRs: One agent to trigger smoke on all, then report readiness

Agent type: copilot-fixpr for fixing, general-purpose for auditing/triggering

Phase 8: Hold Loop — Stay Alive Until All PRs Merge or Are Reported Ready

After Phase 7, if any PRs are still open, do not exit. Schedule a wakeup to re-check:

# In /loop mode (Claude Code interactive) — fires only when REPL is idle, never interrupts work
ScheduleWakeup(
    delaySeconds=180,   # 3 min — within 5-min cache window, catches Design Doc bot commits
    prompt="<<autonomous-loop-dynamic>>",
    reason="babysit hold: re-checking gate status after Design Doc bot commits"
)

On each wakeup iteration, run a lightweight status sweep (not a full audit):

# Re-check gate status for each open PR (no skeptic dispatch — nothing to re-dispatch,
# skeptic-cron.yml was removed repo-wide; see rev-p0cc4).
for pr in $(gh pr list --repo $GITHUB_REPOSITORY --state open --json number --jq '.[].number'); do
  fail=$(gh pr view $pr --repo $GITHUB_REPOSITORY --json statusCheckRollup \
    --jq '[.statusCheckRollup[] | select((.__typename=="StatusContext" and (.state=="FAILURE" or .state=="ERROR")) or (.__typename=="CheckRun" and (.conclusion//"" | ascii_upcase | (.=="FAILURE" or .=="ERROR" or .=="TIMED_OUT"))))] | length')
  pending=$(gh pr view $pr --repo $GITHUB_REPOSITORY --json statusCheckRollup \
    --jq '[.statusCheckRollup[] | select(.__typename=="CheckRun" and (.conclusion==null or .conclusion==""))] | length')
  mergeable=$(gh pr view $pr --repo $GITHUB_REPOSITORY --json mergeable --jq '.mergeable')
  if [[ "$fail" == "0" && "$pending" == "0" && "$mergeable" == "MERGEABLE" ]]; then
    echo "PR #$pr is /green and ready — surface to the user for explicit MERGE APPROVED (babysit never merges on its own)."
  else
    echo "PR #$pr not ready yet (failures=$fail pending=$pending mergeable=$mergeable)"
  fi
done

Exit condition: Stop scheduling wakeups when gh pr list --repo $GITHUB_REPOSITORY --state open --json number --jq 'length' returns 0 (all merged), OR when all remaining open PRs have been reported ready and are waiting on human merge decisions (don't loop forever polling PRs that are already fully reported — surface them once, then hold at a longer interval or hand off).

CronCreate alternative (non-interactive / cron-based sessions):

# Guard: create at most ONE cron job per babysit session. Check CronList first.
jobs = CronList()
babysit_jobs = [j for j in jobs if "babysit" in j.get("prompt", "").lower()]
if not babysit_jobs:
    job_id = CronCreate(
        cron="*/5 * * * *",  # every 5 minutes
        # Sweep-only prompt — do NOT use "/babysit" here (that runs full 6-phase audit).
        # This fires the Phase 8 status sweep: report which open PRs are /green and
        # ready for human MERGE APPROVED. No skeptic re-dispatch (mechanism removed).
        prompt="Run /babysit Phase 8 status sweep only: for each open PR in $GITHUB_REPOSITORY, check /green (CI + no-conflicts) and draft-phase quality status and report any that are ready for human MERGE APPROVED. Skip full Phase 1-7 audit.",
        recurring=True
    )

Use CronCreate only when NOT in an interactive /loop session. Delete the job with CronDelete once all PRs merge. The guard above prevents duplicate cron jobs from accumulating across repeated babysit invocations.

Key rule: never sleep-poll inline. The wakeup fires while the REPL is idle — it cannot interrupt a user conversation or an in-progress tool call. This is safe to leave running.

Evidence Monitoring Protocol

When babysitting active testing or evidence generation:

1. Poll evidence traces: Continuously verify that the testing framework writes traces to /tmp/.../iteration_XXX. Confirm .jsonl files are non-empty.
2. Copy to persistent path: After iteration completes, copy to docs/evidence/pr-<number>/ (e.g., docs/evidence/pr-6851/).
3. Testing Gap Close Integration: If evidence bundles fail to generate (empty .jsonl files, missing checksums, server timeout errors, or EvidenceSignatureGuard rejections), immediately invoke the /testing-gap-close skill to harden the server lifecycle and resolve the failure.
4. Nudge stale processes: If a background daemon hangs on testing or evidence compilation, use scoped shutdown: (1) SIGTERM to the test-managed PID file, (2) lsof -ti :<port> | xargs kill for the bound port, (3) kill -- -$(ps -o pgid= -p <pid>) for the process group. Only fall back to pkill -9 -f gunicorn as a last resort when all scoped methods fail.

Anti-Patterns (BANNED)

• Polling CI status in a sleep loop — check once
• Waiting for or requiring CodeRabbit/Bugbot before pushing or reporting a PR ready — they are optional advisory reviewers only
• Resolving review threads to trigger auto-approve
• Declaring /green without Green Gate log verification for gates 1-2
• Reporting a PR as "looking good" after checking only mergeable — always run statusCheckRollup check first
• Running gh pr merge without explicit human "MERGE APPROVED" in the live conversation — babysit itself never merges, it only reports readiness
• Calling sleep inline to wait for bots — use ScheduleWakeup instead
• Exiting babysit after Phase 7 while PRs are still failing/pending gates — must hold and keep working them; PRs that have reached /green plus draft-phase-quality-complete and been reported ready are a hold-and-surface case (Phase 8), not a "keep polling forever" case — babysit cannot force a merge to happen, only a human can