itil-ops（信息操作）

ITIL Ops — IT Service Management for AI Agents

Structured incident, problem, and change management adapted from ITIL 4 for autonomous agent operations.

Core Concepts

Severity Levels

Incident vs Problem vs Change

• Incident: Something broke. Restore service ASAP. (reactive)
• Problem: Pattern of incidents. Find and fix root cause. (proactive)
• Change: Planned modification. Assess risk before executing. (controlled)

Incident Management

Detection Sources

Scan these in order of criticality:

1. Service crashes — journalctl --user -u SERVICE --since "12 hours ago" for watchdog timeouts, SIGABRT, SIGSEGV, core dumps
2. Cron failures — consecutive error count > 2 in job state files
3. Health endpoints — HTTP health checks returning non-200
4. Resource pressure — disk > 80%, RAM > 80%, swap active
5. Data integrity — schema validation failures, broken files, load errors

Detection Script

Run scripts/itil-review.sh to scan all sources. It outputs:

• ITIL_CLEAR if nothing found (reply HEARTBEAT_OK)
• Formatted report with incidents and problems if issues detected

Incident Lifecycle

DETECTED → CLASSIFIED (P1-P4) → DIAGNOSED → RESOLVED → CLOSED
                                      ↓
                              (3+ occurrences)
                                      ↓
                              ESCALATE TO PROBLEM

Auto-Classification Rules

# P1 — Critical
- Service crash count >= 3 in 12h (crash loop)
- Disk usage >= 90%
- RAM usage >= 90%
- Data loss detected

# P2 — High
- Service crashed 1-2 times
- 3+ services down simultaneously
- Auth/token failures affecting operations
- Cron job with 5+ consecutive failures

# P3 — Medium
- Broken data files (schema violations)
- Memory load errors > 10 in 12h
- Cron job with 3-4 consecutive failures
- Disk usage 80-89%

# P4 — Low
- 1 service down (non-critical)
- Config warnings
- Log noise

Creating Incident Tickets

When incidents are found, create coordination tasks:

Title: [ITIL-INC] <brief description>
Body:
- Severity: P1/P2/P3/P4
- Category: service|cron|memory|disk|security
- Detected: <timestamp>
- Detail: <what happened>
- Impact: <what's affected>
- Action: <what to do>

Problem Management

Pattern Detection

An incident becomes a problem when:

• Same error occurs 3+ times in 24h
• Same incident type recurs across 2+ review cycles
• Multiple related incidents share a common root cause

Root Cause Analysis (RCA)

When a problem is identified:

1. Gather evidence — journal logs, error messages, state files, recent changes
2. Timeline — reconstruct the sequence of events
3. 5 Whys — ask why iteratively until you reach the actual root cause
4. Fix classification:

- Quick fix — config change, file repair, timeout bump - Code fix — bug in script or daemon, needs PR - Architecture fix — design flaw, needs redesign

Problem Ticket Format

Title: [ITIL-PRB] <root cause description>
Body:
- Related incidents: <list>
- Root cause: <what's actually broken>
- Evidence: <logs, patterns, data>
- Fix applied: <immediate remediation>
- Fix needed: <permanent solution>
- Prevention: <how to prevent recurrence>

Known Error Database

Track resolved problems in state file (itil-state.json):

{
  "last_review": "2026-03-22T04:19:50Z",
  "last_incident_count": 2,
  "last_problem_count": 1,
  "known_errors": {
    "memory-content-dict": {
      "description": "Scripts writing content as dict instead of string",
      "root_cause": "Missing json.dumps() in memory file writers",
      "fix": "Wrap content in json.dumps() before saving",
      "fixed_date": "2026-03-22"
    }
  }
}

Change Management

Pre-Change Checklist

Before modifying services, configs, or infrastructure:

1. What's changing? — specific files, services, configs
2. Why? — linked incident/problem ticket
3. Risk? — what could go wrong
4. Rollback plan? — how to undo if it breaks
5. Test? — how to verify it worked
6. Notify? — does the human need to know

Change Categories

Post-Change Verification

After any change:

1. Check service status — systemctl --user status SERVICE
2. Watch logs for 60s — journalctl --user -u SERVICE -f --since "now"
3. Run health check — scripts/itil-review.sh
4. Verify no new errors in first 5 minutes

Event Management

Log Monitoring Patterns

# Service crashes
journalctl --user -u SERVICE --since "12h ago" | grep -ciE "watchdog timeout|killed|SIGABRT|SIGSEGV|failed with"

# Memory/resource issues
journalctl --user -u SERVICE --since "12h ago" | grep -c "Failed to load"

# Auth failures
journalctl --user -u SERVICE --since "12h ago" | grep -ciE "unauthorized|403|token expired|auth fail"

Health Check Endpoints

Check services with curl:

curl -sf --max-time 5 "$URL" >/dev/null 2>&1 || echo "DOWN"

Configure endpoints in the review script for your environment.

Continual Improvement

Review Cadence

KPIs to Track

• MTTR (Mean Time to Resolve) — how fast do we fix incidents?
• Incident recurrence rate — are the same things breaking?
• False positive rate — are we alerting on non-issues?
• Known error resolution — are problems getting permanent fixes?

State Tracking

The review script maintains itil-state.json with:

• Last review timestamp and results
• Incident/problem counts per review
• System metrics (disk, RAM, restart count)
• Cross-review pattern detection data

Cron Setup

Recommended Schedule

# Incident review — every 12 hours
openclaw cron add --name "itil-review" --every "12h" \
  --model "anthropic/claude-sonnet-4-6" --timeout-seconds 180 \
  --session isolated \
  --message "Run ITIL review: bash ~/.skcapstone/agents/lumina/scripts/itil-review.sh"

# Weekly problem review (Sunday 9 AM)
# Analyze the week's incidents, identify patterns, suggest improvements

File Structure

itil-ops/
├── SKILL.md              # This file
├── scripts/
│   └── itil-review.sh    # Main review script (scan + classify + report)
└── references/
    └── itil4-agent-mapping.md  # ITIL 4 → Agent operations reference

Integration Points

• Coordination tasks — skcapstone coord create for incident/problem tickets
• Memory snapshots — skmemory_snapshot to record resolutions for future reference
• Heartbeat — integrate with existing heartbeat to run lightweight checks
• Cron — scheduled reviews via OpenClaw cron system
• Alerting — Telegram/Discord delivery for P1/P2 issues