| name | pr-monitor |
| description | Watch a GitHub PR asynchronously until CI is green and review threads are
resolved (or through merge), using the harness Monitor tool so the main
session keeps working. Optional auto-fix mode reacts to CI failures by
fixing and pushing on the PR branch.
Use when (1) the user asks to watch a PR or CI (「PR見てて」「CI監視して」),
(2) changes were just pushed and CI needs confirmation, (3) another skill
(e.g. issue-ship) needs CI watched through merge. Self-contained — no
plugin dependency, so it works on every machine.
|
| argument-hint | [pr-number] [--until merged] [--auto-fix] [--exclude name1,name2] |
PR Monitor
/pr-monitor # watch current branch's PR until green
/pr-monitor 123 # watch PR #123
/pr-monitor 123 --until merged # keep watching through merge/close
/pr-monitor 123 --auto-fix # fix CI failures on the PR branch as they appear
/pr-monitor 123 --exclude e2e # exclude additional check names
Steps
1. Resolve PR, owner, repo
If no PR number was given, detect the current branch's PR with
gh pr view --json number,url,title. No PR found → report and stop.
Resolve the repo with gh repo view --json owner,name.
2. Build the exclude list
Start from require-approval,check-approval (approval-gated checks only go
green when a human clicks Approve — waiting on them is pointless). Append any
--exclude values. Validate each name: ASCII alphanumerics, hyphens,
underscores, dots, slashes, and spaces only — reject anything with shell
metacharacters instead of quoting it.
3. Launch the Monitor (once)
Monitor:
persistent: true
description: "PR #<N> — CI + review threads"
command: bash "${CLAUDE_SKILL_DIR}/scripts/pr-monitor.sh" <PR> <OWNER> <REPO> "<EXCLUDED>" <green|merged>
The 5th argument is merged when --until merged was passed, otherwise omit
it (defaults to green). Print a one-line confirmation with the PR URL and
end the turn — events arrive as notifications; do not poll or block.
Event format
The script emits one line per state change (silence = nothing changed) and
preserves the last known state across transient fetch failures:
CI {"total":7,"pending":3,"failed":0,"failures":[]} | Reviews {"total":4,"unresolved":2} | PR OPEN
CI {"total":7,"pending":0,"failed":1,"failures":["test-unit"]} | Reviews {"total":4,"unresolved":2} | PR OPEN
DONE: all checks green and threads resolved
DONE: merged
DONE: closed without merge
[POLL_ERROR] 3 consecutive fetch failures — check gh auth / network
failed covers every terminal non-success conclusion (FAILURE, CANCELLED,
TIMED_OUT, ACTION_REQUIRED, STALE, ERROR, STARTUP_FAILURE) — silence is never
a missed crash.
Reacting to events
Default mode is watch-and-report: surface failures with check names to the
user; decisions (rerun, fix, ignore) are theirs. On DONE: all checks green,
tell the user the PR is ready.
Auto-fix mode (--auto-fix)
Invoking with --auto-fix is the user's standing approval to push follow-up
fixes to this PR's branch — and only this branch. On each failure event:
- Fetch the failing run log (
gh run view --log-failed) and reproduce
locally on the PR branch.
- Fix, verify locally, push. Keep the Monitor running — the next event
confirms the outcome.
- Hard limits: max 3 fix attempts per check name, then stop and report
(oscillation guard). Never force-push. Never change scope beyond making
the named checks pass; anything structural goes back to the user.
[POLL_ERROR] repeating → stop the Monitor (TaskStop) and tell the user to
check gh auth status / network. After DONE: merged or DONE: closed,
no further pushes — the approval expires with the PR.
