Review Patrol Routine

Review Patrol System

A review policy is not a system. This skill turns PR patrol into a system.

Design Principle: Deterministic First, LLM Second

The patrol script (review_patrol.py) is pure Python + gh CLI -- no LLM tokens consumed. It runs deterministically: same input (GitHub API state) → same output (queue classification). This means:

• Every 30-min heartbeat cycle costs ~0 tokens for the scan itself
• The LLM is only needed after the scan, when there's actually something to review
• If the scan finds new=0, reReview=0 → the heartbeat can return HEARTBEAT_OK without burning tokens on review work
• State persistence means no redundant re-discovery of already-reviewed PRs

Rule: Everything that can be deterministic (scan, classify, log) is deterministic. Only the actual review judgment requires LLM reasoning.

Use this skill when you need periodic PR review patrol with:

• explicit repo scope
• persistent queue state
• deterministic new-vs-rereview classification
• heartbeat-safe reporting
• direct integration with GitHub review workflows

When to Use

Activate this skill when:

• a human asks for regular PR patrol
• an agent keeps forgetting to run review cycles consistently
• you need to survive resets/restarts without losing review state
• you want a lightweight ops layer without a full external service

Do not use this skill for one-off PR reviews. For single PR review execution, use github-pr-review.

What This Skill Provides

Three layers:

1. Trigger layer -- heartbeat, cron, or scheduler invokes the patrol regularly
2. State layer -- memory/heartbeat-state.json stores repo/PR review state
3. Classification layer -- patrol script classifies every open PR into a queue

Files

skills/review-patrol-routine/
├── SKILL.md                          # This file
├── scripts/
│   └── review_patrol.py              # Deterministic scan + classify script
└── references/
    └── STATE.md                      # State schema documentation

Queue Semantics

newReviewQueue

• PR is open, not draft, not authored by the patrol agent
• No prior review recorded in patrol state

reReviewQueue

• PR is open, not draft
• Last verdict was REQUEST_CHANGES
• Current headSha differs from lastReviewedHeadSha (new commits pushed)

readyToMerge

• PR is open, not draft
• Last verdict was APPROVE
• Current headSha matches lastReviewedHeadSha (no new commits)
• GitHub mergeStateStatus is CLEAN or HAS_HOOKS

staleApproved

• PR is open, not draft
• Last verdict was APPROVE
• Current headSha differs from lastReviewedHeadSha (branch moved since approval)

blockedByReview

• PR is open, previously received REQUEST_CHANGES
• No new commit since the last review

blockedByMergeability

• GitHub reports mergeStateStatus as DIRTY, BLOCKED, UNKNOWN, or UNSTABLE

A PR can be blocked by review, blocked by mergeability, both, or neither.

drafts

Observe only. Do not review unless explicitly asked.

Per-PR status + reason

Every tracked PR gets a status and reason field for transparency:

Patrol State Model

Canonical state file: memory/heartbeat-state.json

{
  "version": 2,
  "lastPatrol": "2026-03-29T11:00:00Z",
  "cycleCount": 290,
  "repos": {
    "org/repo": {
      "lastCheck": "2026-03-29T11:00:00Z",
      "prs": {
        "42": {
          "title": "feat: something",
          "url": "https://github.com/org/repo/pull/42",
          "author": "someone",
          "headSha": "abc123",
          "lastReviewedAt": "2026-03-28T...",
          "lastReviewedHeadSha": "abc123",
          "lastVerdict": "APPROVE",
          "lastMergedAt": null,
          "open": true
        }
      }
    }
  },
  "lastSummary": { ... }
}

Cycle Log

Every patrol run appends one JSON line to memory/patrol-log.jsonl:

{"cycle":290,"at":"2026-03-29T11:00:00Z","scanned":19,"open":11,"new":2,"reReview":0,"readyToMerge":1,"staleApproved":0,"blockedByReview":5,"blockedByMergeability":3,"drafts":1,"errors":0,"items":[{"pr":"org/repo#42","status":"new","reason":"unreviewed"}]}

This log is append-only and never truncated. Use it for:

• Debugging why a PR was or wasn't reviewed
• Auditing patrol coverage over time
• Answering "when did the patrol last see PR #X?"

Setup

1. Configure repos and identity

The script loads configuration in this precedence order:

Option A: Environment variables (recommended for CI/cron)

export PATROL_REPOS="org/repo-a,org/repo-b,org/repo-c"
export PATROL_SELF_LOGINS="YyMyAgent,myagent"
export PATROL_WORKSPACE="/path/to/agent/workspace"  # optional, defaults to cwd

Option B: Config file (recommended for agents) Create patrol-config.json in your workspace root:

{
  "repos": [
    "org/repo-a",
    "org/repo-b"
  ],
  "selfLogins": ["YyMyAgent", "myagent"]
}

Option C: Inline edit (last resort) Edit _FALLBACK_REPOS and _FALLBACK_SELF_LOGINS directly in the script.

Workspace resolution: The script uses PATROL_WORKSPACE env var if set, otherwise cwd. State and log files are written relative to this workspace root. This means agents can use the script directly from agentic-foundation without copying -- just configure via env or config file and run from the workspace directory.

Only include repos the human explicitly approved. No scope drift.

2. Initialize state

mkdir -p memory
echo '{"version":2,"lastPatrol":null,"cycleCount":0,"repos":{},"lastSummary":null}' > memory/heartbeat-state.json

3. Add to HEARTBEAT.md

## PR Review Patrol

```bash
python3 skills/review-patrol-routine/scripts/review_patrol.py markdown

### 4. Seed known review history

If the agent already reviewed active PRs before adopting this skill, write those
PRs into the state file so they don't appear as false-new.

## Operating Procedure

### Per heartbeat / cycle

1. Run patrol script
2. Check GitHub notifications for @mentions and review requests (see below)
3. Review every PR in `newReviewQueue`
4. Re-review every PR in `reReviewQueue`
5. **Do NOT merge** -- merging is the PM's responsibility. Leave `readyToMerge` items as-is and report them.
6. Skip drafts
7. After each review, update the matching PR entry in state:
   - `lastReviewedAt`, `lastReviewedHeadSha`, `lastVerdict`
8. Emit a compact cycle summary

> ⚠️ **Merge Policy:** The review agent approves, the PM (project manager) merges. Never auto-merge after approval unless explicitly configured for a specific repo by the repo owner. When in doubt: don't merge.

## GitHub Mentions & Review Requests Check

After the patrol scan, check for notifications that don't show up as open PRs in the patrol scope -- @mentions in comments, explicit review requests, etc.

### Notifications API (recommended -- catches everything)

```bash
gh api notifications --jq '.[] | select(.reason == "mention" or .reason == "review_requested") | {reason, repo: .repository.full_name, title: .subject.title, type: .subject.type, url: .subject.url}'

Mentions via Search (alternative)

gh api "search/issues?q=mentions:YOUR_GITHUB_LOGIN+updated:>2026-03-30&sort=updated&per_page=10" --jq '.items[] | {title, number}'

Review Requests (explicit reviewer assignment)

gh api "search/issues?q=review-requested:YOUR_GITHUB_LOGIN+is:open+is:pr&sort=updated&per_page=10" --jq '.items[] | {title, number, html_url}'

Handling

For each notification:

1. Read the comment/thread to understand what's being asked
2. Respond if actionable (review the PR, answer the question, comment on the issue)
3. Mark the notification as read:

   gh api notifications/threads/{thread_id} -X PATCH
   

Why this matters

The patrol script only sees open PRs in the configured repo scope. It misses:

• @mentions in issue comments or PR discussions
• Explicit review requests via GitHub's "Request review" button
• Mentions in repos outside the patrol scope
• Comments on already-reviewed/merged PRs asking for follow-up

The notifications check closes this gap.

Required state write-back after review

{
  "lastReviewedAt": "2026-03-29T...",
  "lastReviewedHeadSha": "abc123",
  "lastVerdict": "REQUEST_CHANGES"
}

If merged: add "lastMergedAt": "...", "open": false

Triggering Strategy

This skill does not create time itself. You need an external trigger:

• OpenClaw heartbeat (recommended)
• OpenClaw cron
• Host-level crontab
• project-engine scheduled workflow

Recommended cadence: every 30 min for new + re-review patrol.

Token budget: The scan itself costs 0 LLM tokens. Only when newReviewQueue or reReviewQueue is non-empty does the agent need to invoke the LLM for actual review work. On quiet cycles (nothing new), the entire heartbeat completes without any LLM call.

Gotchas

Memory without trigger is fake periodicity

Files on disk but no scheduler → the system is persistent but not periodic.

Trigger without state causes duplicate work

No heartbeat-state.json → every open PR looks new every time.

Review submitted, state not updated

Queue stays wrong. The skill depends on explicit write-back.

Bad repo scope pollutes the patrol

Remove dead/unresolvable repos quickly.

"Docs-only = zero risk" bias

Security checks (especially secret detection) apply to every line in the diff, including docs, examples, and markdown. See github-pr-review skill for details.

Relationship to Other Skills

Anti-Patterns

A patrol that depends on mood is not a patrol. It is luck.