pr-review-sweep

name	pr-review-sweep
description	Nightly sweep of recently-merged PRs to address unhandled review comments via a Claude Code sub-agent.
version	0.1.0
license	MIT
metadata	{"hermes":{"tags":["github","code-review","automation","claude-code","cron"],"related_skills":["claude-code","github-pr-workflow"]}}

PR Review Sweep

Walks the last N days of merged PRs across configured GitHub orgs, finds review comments that were never addressed (no PR-author reply, no reactions), and dispatches a Claude Code sub-agent to fix them in a follow-up PR.

Designed to run nightly as a cron job. Reports a summary to the home channel at the end.

When to load

• Cron tick named pr-review-sweep
• Manual ask: "sweep recent PRs for unanswered review comments"
• Investigating whether a specific PR has unhandled feedback

Configuration

Edit these knobs at the top of your cron-job prompt (or pass via env):

Knob	Default	Notes
LOOKBACK_DAYS	7	How far back to scan merged PRs
ORGS	<your-orgs-comma-separated>	Passed to gh search prs --owner
MAX_PRS_PER_RUN	10	Cost ceiling — bump after a few clean runs
TIMEOUT_PER_PR_S	1200	Claude Code wall-clock per PR
EXCLUDE_REPOS	[]	Hard-skip list (archived/abandoned)
FOLLOWUP_LABELS	["review-sweep"]	Labels to add to the follow-up PR
ADMIN_TARGET	home channel	Where the summary report goes

Prerequisites

• gh CLI authenticated (gh auth status must pass)
• Claude Code CLI installed and an ai-coding-config plugin with an address-pr-comments skill (the agent your sub-process actually runs)
• terminal and delegation toolsets enabled
• Workspace provisioning script (e.g. workshop {repo} {branch} or gh repo clone ... ; cd ...)

How it thinks

1. Discover merged PRs in the window:

   gh search prs \
     --owner <org1> --owner <org2> \
     --state closed --merged \
     --merged-at ">$(date -d '${LOOKBACK_DAYS} days ago' +%Y-%m-%d)" \
     --json number,title,repository,author,headRepository \
     --limit 50
   

2. For each PR, fetch comments from both endpoints and filter out anything the PR author has already engaged with:

   # Line-level review comments. Three filters:
   #   1. Not authored by the PR author
   #   2. Top-level (no in_reply_to_id) — replies are handled via their parent
   #   3. Zero reactions
   # Then second pass: drop any top-level comment that has a child reply from the
   # PR author (they already responded inline).
   gh api repos/{owner}/{repo}/pulls/{pr}/comments > /tmp/line-comments.json
   
   AUTHOR="{author}"
   jq --arg author "$AUTHOR" '
     # Collect IDs of parent comments that the author has replied to
     (map(select(.user.login == $author and .in_reply_to_id != null) | .in_reply_to_id) | unique) as $replied_to
     | map(select(
         .user.login != $author
         and .in_reply_to_id == null
         and ((.reactions.total_count // 0) == 0)
         and (.id as $id | $replied_to | index($id) | not)
       ))
   ' /tmp/line-comments.json
   
   # Issue-level (general PR) comments. These are flat (no threading), so use a
   # timestamp heuristic: drop any comment if the PR author posted ANY issue comment
   # AFTER it. That's a reasonable proxy for "the author saw it and moved on."
   gh api repos/{owner}/{repo}/issues/{pr}/comments > /tmp/issue-comments.json
   
   jq --arg author "$AUTHOR" '
     # Latest timestamp of any author issue comment (null if none)
     (map(select(.user.login == $author) | .created_at) | sort | last // "0") as $author_last
     | map(select(
         .user.login != $author
         and ((.reactions.total_count // 0) == 0)
         and (.created_at > $author_last)
       ))
   ' /tmp/issue-comments.json
   

   "Unhandled" = not from PR author, zero reactions, AND no later author engagement (inline reply for line comments, later issue comment for issue comments). No bot allowlist — catches bot reviewers (CodeRabbit, claude-review) and humans alike.

3. Filter:

   • Skip dependabot/renovate PRs unless they have non-bot review comments
   • Skip any repo in EXCLUDE_REPOS
   • Stop after MAX_PRS_PER_RUN

4. For each remaining PR (serial, not parallel):

   • Provision a fresh workshop: workshop {repo} pr-sweep-{pr} (or git clone + git checkout -b pr-sweep-{pr})

   • Spawn Claude Code in ACP mode:

      ```python
      # Pseudocode — adapt to your ACP wrapper
      run_claude_code(
          agent="claude",
          mode="run",
          runtime="acp",
          cwd=workshop_path,
          timeout_s=TIMEOUT_PER_PR_S,
          task=f"""
     

     Use the Skill tool to invoke the address-pr-comments skill with argument {pr_number}.

This skill is provided by the ai-coding-config plugin (globally installed). It knows the project's triage conventions and review standards. Do NOT manually triage — you MUST use the Skill tool to load and execute the address-pr-comments skill.

If the skill creates fixes, they go in a new follow-up PR with labels: {FOLLOWUP_LABELS}.

After creating any follow-up PR, use the Skill tool to invoke address-pr-comments on the NEW PR number too (cascading sweep).

Do NOT merge anything. Leave merge for human review. """, ) ```

• Verify completion: When the sub-process returns, scan its output. If it did NOT use the Skill tool for address-pr-comments, send it back with:

  "You did not use the Skill tool. Invoke the address-pr-comments skill now using the Skill tool."

  Keep doing this until the skill is actually invoked. (Without this, Claude Code in ACP mode will silently freelance instead of running the skill — /slash prefixes are NOT auto-intercepted in ACP/--print mode.)

5. Cleanup workshop directory after each PR is done.

6. Report summary to ADMIN_TARGET covering: PRs scanned, PRs swept, fixes vs declined, PRs that need human attention, whether the skill was invoked per PR.

Pitfalls

• ACP slash commands are meaningless — claude --print "/address-pr-comments" does NOT invoke the skill. Always use the explicit "Use the Skill tool to invoke ..." prompt above. Verify in the output.
• Bot allowlist trap — early versions hardcoded bot logins to skip. Don't. Use the zero-reactions heuristic — it covers any reviewer category.
• Cost runaway — keep MAX_PRS_PER_RUN low (≤10) until you've watched a few successful runs. A bad sweep can spin up 50+ Claude Code processes.
• Serial, not parallel — Claude Code workshops grab the same ~/dev/{repo}* paths; parallel runs collide. Serial keeps state clean.
• Workshop cleanup must succeed — orphan workshop dirs accumulate fast. Always rm -rf after each PR, even on failure.

Cron setup

hermes cron add \
  --name "pr-review-sweep" \
  --schedule "0 6 * * *" \
  --tz "<your-timezone>" \
  --skill pr-review-sweep \
  --model "<work-tier-model>" \
  --timeout 7200

Daily at 06:00 local. Long timeout because Claude Code per-PR runs can be slow (MAX_PRS_PER_RUN * TIMEOUT_PER_PR_S upper bound).

Origin

Ported from an OpenClaw workflows/pr-review-sweep/ folder (AGENT.md + config.md) into a single Hermes skill. The config.md knobs are now the table at the top of this file; the AGENT.md procedure is the body. Cron scheduling lives separately in hermes cron, where it belongs.