Babysit Open PRs

Watch CI on every PR in scope until each one is either green-and-mergeable or has a real (non-flaky) failure that needs human attention.

PURPOSE

When several PRs are open simultaneously (the harness work that landed in PRs #19/#20/#21/#23, the four-PR sweep on 2026-04-26, etc.) the agent otherwise wastes context polling each one's checks individually. This skill consolidates the watch loop into one event stream: one punchy line per state change, log tails only on real failures, automatic handoff to /review-pr when comments arrive.

CRITICAL

• Don't reinvent watching. Use gh pr checks <N> --watch --interval 15 per PR; it already streams check transitions over a long-poll connection. Aggressive sleep N; gh pr checks chains burn context and rate-limit budget.
• Don't trigger on the first red. A single transient failure (network blip, flaky integration test) should not page the human. Re-poll once before alerting — see "Flaky vs. real failure" below.
• Don't fix anything from this skill. This is a watch-and-report loop. When a PR has real CI failures or new review comments, invoke /review-pr <N> rather than fixing inline.
• Stop when scope is empty. A PR that gets merged, closed, or marked mergeable: CONFLICTING while we watch leaves the watch set; if the set becomes empty, exit cleanly and summarize.

ASSUMES

• gh auth status is authenticated (skill will check up front).
• The user is OK with watching for as long as it takes — typical CI on this repo is ~5-8 min per PR, and this skill runs until every watched PR settles.

STANDARD PATH

Phase 1: Discover the watch set

If $ARGUMENTS lists explicit PR numbers, use them. Otherwise default to every open PR authored by the current user:

gh pr list --state open --author '@me' \
  --json number,title,headRefName,baseRefName,mergeable,mergeStateStatus

For each PR, also fetch the initial check rollup so we know the starting state (don't alert on a failure that was already there before we started watching — note it as "pre-existing" in the summary):

gh pr view <N> --json number,statusCheckRollup,reviews,mergeable

# Cache initial review-comment + issue-comment counts to compare against later.
# `--paginate` matters: without it, `length` plateaus at the page size (30) and
# new comments past that boundary will be missed. Aggregate per-page lengths:
review_comment_count=$(gh api --paginate \
  repos/{owner}/{repo}/pulls/<N>/comments \
  --jq 'length' | awk '{sum+=$1} END {print sum+0}')
issue_comment_count=$(gh api --paginate \
  repos/{owner}/{repo}/issues/<N>/comments \
  --jq 'length' | awk '{sum+=$1} END {print sum+0}')

Print the watch set as a table. If empty, stop and tell the user.

Phase 2: Launch the watch loop

For each PR in the watch set, start gh pr checks --watch as a background job so the streams run in parallel rather than serially:

gh pr checks <N> --watch --interval 15 --fail-fast

Use Bash with run_in_background: true, one call per PR. The harness will deliver each line of output as a notification — let Monitor surface them rather than polling BashOutput in a tight loop.

In parallel, every ~60s poll for new review activity (gh pr checks --watch won't catch comments). Watch both review-thread comments and the PR discussion (issue-level) comments, since either kind can carry a request that needs /review-pr follow-up:

# Same `--paginate` + sum dance as Phase 1 — `length` on the first page alone
# stops growing past 30 items.
gh api --paginate repos/{owner}/{repo}/pulls/<N>/comments \
  --jq 'length' | awk '{sum+=$1} END {print sum+0}'
gh api --paginate repos/{owner}/{repo}/issues/<N>/comments \
  --jq 'length' | awk '{sum+=$1} END {print sum+0}'
gh pr view <N> --json reviews --jq '.reviews | length'

Cache the previous counts; emit a new-comment event when any of the three goes up.

Phase 3: Triage each event

Per PR, classify what just changed:

• All checks green + no new comments → emit PR #N all green, ready to merge and remove from watch set.
• One check failed, others still pending → wait. Don't alert yet; the remaining checks may pass or also fail (a real bug usually trips multiple checks). Note the failing check name internally.
• All checks complete, ≥1 failed → fetch the failure log tail and triage flaky vs. real (next section).
• New review comments arrived → emit PR #N has N new comment(s) — invoke /review-pr <N> and remove from watch set (the loop's job is done; the human or /review-pr takes over).
• PR closed / merged / went CONFLICTING → emit one-liner and remove from watch set.

Each event is one line, prefixed with PR #N. Keep it skim-friendly.

Phase 4: Flaky vs. real failure

When a PR has finished CI with ≥1 red check:

1. Fetch the failure log tail:

   gh run view <run-id> --log-failed | tail -60

2. Apply the heuristic:

   • Flaky if the tail contains a network error (ECONNRESET, dial tcp, 503 Service Unavailable), a timeout from a known-flaky integration test, or a transient action error like The runner has received a shutdown signal. Action: re-trigger the job once via gh run rerun <run-id> --failed, log PR #N <job> flaky-retry triggered, and keep watching.
   • Real otherwise (assertion error, type-check error, lint diff, compile failure). Action: emit PR #N <job-name> failed: <one-liner from log tail> and remove from watch set. The human or a follow-up /review-pr invocation handles the fix.
3. Cap retries at one per job. A second flaky-looking failure after re-trigger is treated as real — back-to-back transient failures on the same job are rare enough that human eyes are warranted.

Phase 5: Termination + summary

The loop terminates when the watch set is empty. Print a final summary:

If a PR ended in "real failure" or "new comments", surface that PR number in the final line so the user can copy-paste into /review-pr.

EDGE CASES

• A PR is freshly pushed and CI hasn't started yet. statusCheckRollup is empty for ~30s. Treat empty rollup as PENDING; don't classify it as green.
• gh pr checks --watch exits with a non-zero code on first red because of --fail-fast. That's expected — fold it into the triage path; don't treat the exit code as a tool error.
• A PR's base is another feature branch (stacked PR). Watch it the same way, but in the summary flag it as stacked on <base> so the human knows the merge order matters. Do NOT auto-rebase; that's the user's call.
• Rate limits. Each gh pr checks --watch holds one long-poll connection; watching ~10 PRs in parallel is fine. The 60s comment-polling loop adds ~1 req/min/PR, also fine. If gh returns API rate limit exceeded, back off the comment poll to 5min and continue.
• The user's gh CLI is not authenticated. gh auth status fails — stop immediately and tell the user.
• A PR was created by Dependabot / renovate (not the user). The default scope (--author '@me') excludes them. If the user wants those watched too, they need to pass the PR numbers explicitly.

After

When the loop ends with PRs in "new comments" or "real failure" state, the natural follow-up is /review-pr <N> for each. Surface those PR numbers in the final summary so the user can chain without re-discovering them.