Issue Patrol Routine

Issue Patrol System ⚙️

Autonomous issue discovery and triage for engineering agents.

Use this skill when you need to periodically scan repos for open issues, classify them into work queues, and maintain persistent state across sessions.

Philosophy

• Deterministic scanning. The patrol script is pure logic -- no LLM calls.
• Persistent state. Every cycle updates a JSON state file so the next session knows what changed.
• Append-only logging. Every patrol cycle is logged for auditability.
• Actionable queues. Issues are classified so the agent can immediately decide what to do.

Architecture

Heartbeat (~30 min)
  └→ New Session
      └→ Reads HEARTBEAT.md
          └→ Runs issue_patrol.py
              ├→ Scans repos via gh CLI
              ├→ Classifies issues into queues
              ├→ Updates memory/issue-patrol-state.json
              └→ Appends to memory/issue-patrol-log.jsonl
          └→ Agent processes queues (triage / implement / follow-up)

Setup

1. Configure Target Repos

Define repos in your HEARTBEAT.md:

## Issue Patrol
REPOS="Yesterday-AI/agentic-foundation Yesterday-AI/experts Yesterday-AI/clawrag Yesterday-AI/company-orga Yesterday-AI/blueprints"

2. Deploy the Patrol Script

Copy scripts/issue_patrol.py to your workspace scripts/ directory:

cp skills/issue-patrol-routine/scripts/issue_patrol.py ~/scripts/
chmod +x ~/scripts/issue_patrol.py

3. Initialize State

First run creates the state file automatically. Or initialize manually:

python3 ~/scripts/issue_patrol.py \
  --repos "Yesterday-AI/agentic-foundation Yesterday-AI/clawrag" \
  --state ~/memory/issue-patrol-state.json \
  --log ~/memory/issue-patrol-log.jsonl

Queue Classification

Every open issue is placed into exactly one queue:

Classification Logic

Is the issue assigned to someone else (not me)?
  → YES → assignedToOthersQueue

Is there an active branch/PR linked to this issue?
  → YES → inProgressQueue

Does the issue have label "blocked" or "waiting-for-input"?
  → YES → blockedQueue

Does the issue have label "needs-clarification" or is the body empty/vague?
  → YES → needsClarificationQueue

Has the issue been seen in a previous cycle?
  → NO → newQueue

Is the issue labeled "bug", "feature", "enhancement", or assigned to me?
  → YES → readyToImplementQueue

Otherwise → newQueue (needs triage)

State Model

See references/STATE.md for the full state schema.

Quick overview:

{
  "version": 1,
  "lastPatrol": "2026-03-29T12:00:00Z",
  "cycleCount": 0,
  "agentUser": "YyScotty",
  "repos": {
    "Yesterday-AI/clawrag": {
      "lastCheck": "2026-03-29T12:00:00Z",
      "issues": {
        "42": {
          "title": "Add retry logic for API calls",
          "queue": "readyToImplementQueue",
          "labels": ["enhancement"],
          "assignee": "YyScotty",
          "firstSeenCycle": 5,
          "lastUpdatedAt": "2026-03-28T10:00:00Z",
          "linkedPR": null,
          "status": "ready",
          "reason": "Labeled enhancement, assigned to me, clear requirements"
        }
      }
    }
  }
}

Cycle Log

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

{
  "cycle": 15,
  "timestamp": "2026-03-29T12:00:00Z",
  "reposScanned": 5,
  "totalOpen": 23,
  "queues": {
    "newQueue": 2,
    "needsClarificationQueue": 1,
    "readyToImplementQueue": 5,
    "inProgressQueue": 3,
    "blockedQueue": 1,
    "assignedToOthersQueue": 11
  },
  "changes": [
    {"repo": "Yesterday-AI/clawrag", "issue": 42, "from": "newQueue", "to": "readyToImplementQueue"}
  ]
}

Running the Patrol

python3 scripts/issue_patrol.py \
  --repos "Yesterday-AI/agentic-foundation Yesterday-AI/clawrag" \
  --state memory/issue-patrol-state.json \
  --log memory/issue-patrol-log.jsonl \
  --agent-user YyScotty

Output is a JSON summary printed to stdout for the agent session to consume.

HEARTBEAT.md Integration

Add to your HEARTBEAT.md:

## Issue Patrol
# Schedule: Every heartbeat

1. Run: `python3 scripts/issue_patrol.py --repos "$REPOS" --state memory/issue-patrol-state.json --log memory/issue-patrol-log.jsonl --agent-user YyScotty`
2. Read the JSON output
3. For `newQueue` issues: Read issue body, decide queue placement
4. For `readyToImplementQueue`: Pick highest priority, start implementation (see `issue-to-pr-workflow` skill)
5. For `inProgressQueue`: Check PR status, address review feedback if any
6. Update state file with any manual reclassifications

GitHub Mentions Check

After running the patrol scan, check for @mentions of your GitHub user in issue/PR comments. This catches requests that don't show up as assigned issues.

# List unread mentions
gh api notifications --jq '.[] | select(.reason == "mention") | {subject: .subject.title, repo: .repository.full_name, url: .subject.url}'

For each mention:

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

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

Add to your HEARTBEAT.md:

## GitHub Mentions Check
# Schedule: Every heartbeat

1. Check for new @mentions:
   ```bash
   gh api notifications --jq '.[] | select(.reason == "mention") | {subject: .subject.title, repo: .repository.full_name, url: .subject.url}'

2. For each mention: read the comment thread, respond if actionable
3. Mark handled notifications as read

## PR Follow-ups

After scanning issues, check all open PRs you created for mergeability and review status:

```bash
for REPO in $REPOS; do
  gh pr list --repo $REPO --state open --author @me \
    --json number,title,reviewDecision,mergeable \
    --jq '.[] | "#\(.number) [\(.mergeable)] [\(.reviewDecision // \"PENDING\")] \(.title)"'
done

Why this matters: PRs with merge conflicts block the review pipeline. Check every cycle.

After the Patrol

Based on queue contents, the agent should:

Rules 🛡️

NO_SPAM

• Don't comment on every issue every cycle. Only comment when you have something new to say.
• Track lastCommentedAt in state to avoid duplicate comments.

NO_OVERCOMMIT

• Work on ONE issue at a time (max). Finish or park before starting the next.
• inProgressQueue should rarely have more than 1 item.

NO_SECRETS

• Never put tokens, keys, or credentials in state files, logs, or issue comments.

Part of the agentic-foundation skill library.