Skill

/pr-review-loop — request and handle a Claude PR review, in-session

Automates the loop you'd otherwise do by hand: push → comment @claude → wait for CI → read the findings → fix → optionally re-review. The waiting is event-driven: a background script (watch-claude-review.sh, bundled next to this skill) blocks on gh run watch and, when it exits, the harness re-invokes this session with the result. You burn no turns polling.

Depends on the repo having a .github/workflows/claude.yml that fires on a comment containing @claude (the standard anthropics/claude-code-action workflow). That workflow is never modified by this skill. A repo that doesn't have it yet can scaffold one with npx github:adevra/claude-pr-review-loop init.

Step 1 — Preconditions (refuse early if unmet)

• git status --porcelain must be empty. If the tree is dirty, stop: "Commit or stash your changes first — /pr-review-loop reviews what's pushed."
• git branch --show-current must not be main (or master). If on the default branch, stop: "Switch to a feature branch first."

Step 2 — Push and ensure a PR exists

BRANCH=$(git branch --show-current)
git push -u origin "$BRANCH"
PR=$(gh pr view "$BRANCH" --json number --jq '.number' 2>/dev/null)
# If no PR yet, create one (fill title/body from the branch's commits):
#   gh pr create --base main --head "$BRANCH" --title "<title>" --body "<summary>"
# then re-read PR. Reuse an existing PR — never open a duplicate.

Step 3 — Request the review and launch the watcher

Post the trigger as a comment (the description does NOT trigger CI), capture the comment's GitHub createdAt as the discovery timestamp (server clock — no skew), then launch the watcher in the background:

gh pr comment "$PR" --body "@claude please review this PR."
SINCE=$(gh pr view "$PR" --json comments --jq '.comments[-1].createdAt')

Then run the watcher with run_in_background: true (this is the whole point — the session goes idle and is re-invoked when CI finishes):

bash "$HOME/.claude/skills/pr-review-loop/watch-claude-review.sh" "$PR" "$SINCE"

Tell the user: "Requested the review on PR #<n> and I'm waiting for CI — I'll pick this back up automatically when it finishes." Then end the turn (do not poll).

Step 4 — On wake: read the result

When the background command completes you'll see a WATCH_RESULT: line:

• WATCH_RESULT: no_run … → the @claude comment didn't trigger the workflow. Report that and stop (suggest checking the comment / that .github/workflows/claude.yml exists and is enabled).
• WATCH_RESULT: done … conclusion=success → fetch the newest bot output and continue to Step 5. Check both streams and use whichever is newest (robust to sticky-comment vs formal review):

  gh pr view "$PR" --json comments --jq '[.comments[] | select(.author.login | startswith("claude"))] | last | .body'
  gh api "repos/{owner}/{repo}/pulls/$PR/reviews" --jq '[.[] | select(.user.login | startswith("claude"))] | last | .body'
  

• WATCH_RESULT: done … conclusion=skipped or any non-success → surface the run URL; do not parse findings. Stop.

If the review has no actionable findings (approve/clean), report "the review came back clean" and stop — do not prompt.

Step 5 — Gate: fix or leave (Question A)

Summarize the findings, then ask with AskUserQuestion:

• Fix now → go to Step 6.
• Leave it for now → stop. The PR is left exactly as the reviewer saw it.

Step 6 — Fix, then gate again (Question B)

Apply superpowers:receiving-code-review discipline — do not implement verbatim:

• Fix the findings you judge valid for this codebase.
• For each finding you do not act on (false positive, technically wrong, or out-of-scope), name it and give a one-line reason in your summary, so the developer sees the full picture.

Commit the fixes. Then ask with AskUserQuestion (note: both options push — the only choice is whether to re-trigger a review):

• Push & request another review → push, post @claude please review this PR., capture a fresh SINCE, relaunch the watcher (Step 3's background command), end the turn → back to Step 4.
• Push, skip re-review → push the fixes and stop.

Round cap

At most 2 review cycles per invocation (round 1 = the initial review; round 2 = the one re-review reachable via Question B). After the round-2 review is handled, if the user again chooses to fix, push the fixes but do NOT trigger a third review — drop the "request another review" option, report, and hand back. This prevents runaway CI.

Edge cases

Situation	Behaviour
Working tree dirty, or on main/master	Refuse in Step 1; explain why.
PR already exists for the branch	Reuse it; don't open a duplicate.
WATCH_RESULT: no_run	Comment likely didn't fire the workflow (or no claude.yml); report and stop.
Run conclusion ≠ success	Surface the run URL; don't parse findings.
Review came back clean	Report "clean", skip Question A.

Notes

• The watcher runs through the Bash tool (Git Bash / POSIX) — POSIX shell + gh. Requires gh authenticated.
• The watcher lives next to this file: $HOME/.claude/skills/pr-review-loop/watch-claude-review.sh. It is launched as a background Bash command so the session sits idle until CI finishes.
• Run discovery is by timestamp, not headSha: every claude.yml run fires on issue_comment against the default branch, so all runs share main's headSha regardless of PR. Our run is the oldest claude.yml issue_comment run created at/after SINCE.