Incident Fupan (事故复盘)

Structured root cause analysis and prevention protocol for production incidents.

Language Rule

Report language matches the user's language. Chinese request → Chinese report. English → English.

When to Trigger

• Production failure, outage, or unexpected loss
• Agent mistake with real consequences (wrong deployment, fabricated output acted upon)
• Near-miss that could have caused damage
• Periodic review of past incidents to extract patterns

Postmortem Flow

Step 1: Gather Evidence

Before writing anything, collect raw evidence. Do NOT rely on memory or assumptions.

Required evidence (use tools to retrieve):

• Logs: exec("grep -i 'error\|fatal\|exception' {logfile} | tail -50")
• Git state: exec("git log --oneline -10"), exec("git diff HEAD~1 --stat")
• Service state: exec("systemctl status {service}"), exec("ps aux | grep {process}")
• Data files: read any CSVs, configs, or state files involved

Rule: Every factual claim in the postmortem must cite a source. Format: [Source: {filepath}:{line} or {command output}]

If evidence is unavailable (logs rotated, service restarted), explicitly mark: [Evidence unavailable: {reason}]

Step 2: Build Timeline

Construct a minute-by-minute (or hour-by-hour) timeline from evidence:

HH:MM UTC — {what happened} [Source: {evidence}]
HH:MM UTC — {what happened} [Source: {evidence}]
...
HH:MM UTC — Incident resolved / mitigated

Include: first symptom, detection, escalation, diagnosis, fix, verification.

Mark the detection gap: time between first symptom and human awareness. This is often the real problem.

Step 3: 5 Whys Root Cause

Drill down from symptom to root cause:

Why 1: {symptom happened} → Because {direct cause}
Why 2: {direct cause} → Because {deeper cause}
Why 3: {deeper cause} → Because {systemic issue}
Why 4: {systemic issue} → Because {process/design gap}
Why 5: {process/design gap} → Because {root cause}

Stop when you reach something you can change. If you reach "the model hallucinated" — that's not actionable. Go deeper: why was the output trusted without verification? Why was there no checkpoint?

Step 4: Write Report

Save to: ~/incidents/INC{NNN}_{TOPIC}_{YYYYMMDD}.md

Create ~/incidents/ if it doesn't exist.

See references/report-template.md for the full template with all required sections.

Report must include all 8 sections:

1. Header (ID, severity, date, duration, status)
2. Executive Summary (3 sentences max)
3. Timeline with sources
4. Impact Assessment (quantified: time lost, money lost, trust lost)
5. 5 Whys Root Cause
6. Fix (what was done immediately)
7. Prevention (new rules, checks, or automation to prevent recurrence)
8. Action Items (P0/P1/P2 with owners and deadlines)

Step 5: Extract Defensive Rules

The most valuable output of a postmortem is new rules that prevent recurrence.

Pattern: Incident → Rule → Enforcement mechanism

Examples from real incidents:

• Unauthorized deployment → "All deploys require explicit human approval" → Gate in CI/CD
• Fabricated data report → "All data claims must cite source file + row count" → L1 in Hallucination Guard
• Session bloat causing errors → "Compact at 80% context usage" → Automated monitoring

See references/patterns.md for a library of incident-to-rule patterns.

Step 6: Reply and Store

1. Save report file to ~/incidents/
2. Reply to user with the full report
3. Store key lessons to long-term memory
4. If applicable: update AGENTS.md, TOOLS.md, or relevant skill with new rules

Severity Levels

Integration

• Hallucination Guard: Incidents caused by agent fabrication → add to L1/L3 detection rules
• War Room: After postmortem, use War Room to evaluate proposed prevention strategy
• Postmortem → new rules → enforcement → fewer incidents → feedback loop

References

• references/report-template.md — Full report template
• references/patterns.md — Incident-to-rule pattern library (sanitized real examples)