Incident Response Skill

When to Trigger

• Production is down or severely degraded
• Error rate spikes significantly above baseline
• A customer reports a critical bug in production
• A security breach or data leak is suspected
• A deployment caused an unexpected regression

Step-by-Step Process

1. Declare and Classify the Incident

First: breathe. Then classify.

Open .agent/incidents/INC-<YYYY-MM-DD>-<slug>.md and write:

# Incident: <short title>
**Date:** <YYYY-MM-DD HH:MM UTC>
**Severity:** P<N>
**Status:** Investigating
**Incident Commander:** <name>

2. Triage — Gather Evidence Fast

Don't guess. Collect data first.

Checklist:

• What is the user-visible symptom?
• When did it start? (check monitoring, not just "someone reported it")
• What changed recently? (last deploy, config change, dependency update)
• What does the error log say? (get exact error messages + stack traces)
• Is it affecting all users or a subset?
• Is there an automated alert, or was this user-reported?

Commands to run immediately:

# Recent deploys
git log --oneline -20

# Check running process / container health
# (adjust for your stack)
docker ps
systemctl status <service>

# Tail live logs
docker logs -f <container> --tail 100
journalctl -u <service> -n 100 -f

Add findings to the incident file under ## Evidence.

3. Contain the Blast Radius

Before fixing, limit the damage:

• If a bad deploy caused it: roll back immediately
  bashCopy

  git revert HEAD
  # or redeploy previous known-good version
  

• If a specific feature is failing: toggle it off via feature flag
• If data is being corrupted: take the affected service offline temporarily
• If a secret was exposed: rotate it immediately, then investigate

Document every action taken in the incident file with timestamps.

4. Root Cause Analysis

Once contained, find the real cause — not just the symptom.

Use the 5 Whys technique:

Why did users see a 500 error?
→ Because the database query timed out.
Why did the query time out?
→ Because a new index was missing after the migration.
Why was the index missing?
→ Because the migration script didn't include CREATE INDEX.
Why didn't the migration script include it?
→ Because we don't have a convention for checking query plans before deploy.
→ Root cause: Missing pre-deploy query plan review process.

The root cause is almost never the immediate technical failure — it's the process gap that allowed it.

5. Fix and Deploy

Write the fix:

• Smallest possible change that resolves the issue
• Add a test that would have caught this
• Double-check the fix in staging before production

Deploy:

• Monitor metrics during deploy
• Keep someone watching logs for 15 minutes post-deploy
• Confirm symptom is resolved with a real user test

6. Write the Post-Mortem

Within 24h of resolution, write a post-mortem in .agent/incidents/INC-<date>-<slug>.md:

# Post-Mortem: <title>

## Summary
<2-3 sentences: what happened, how long, how many users affected>

## Timeline
| Time (UTC) | Event |
|------------|-------|
| HH:MM | Issue first occurred (inferred from logs) |
| HH:MM | First alert / user report |
| HH:MM | Incident declared |
| HH:MM | Mitigation applied |
| HH:MM | Root cause identified |
| HH:MM | Fix deployed |
| HH:MM | Incident resolved |

## Root Cause
<The actual root cause — not just the symptom>

## What Went Well
- <something that helped — good monitoring, fast rollback, etc.>

## What Went Poorly
- <something that made it worse — slow detection, unclear runbook, etc.>

## Action Items

| Action | Owner | Due Date |
|--------|-------|----------|
| Add missing index to migration convention | <name> | <date> |
| Add test for query performance | <name> | <date> |
| Set up alert for DB query p99 latency | <name> | <date> |

7. Update the Knowledge Base

After the post-mortem, use the knowledge-base-update skill to create a gotcha entry for the root cause so future agents and developers know about it.

Rules

• Contain first, diagnose second. Stopping the bleeding is more important than understanding the wound.
• No blame in post-mortems. Focus on process gaps, not people.
• Every incident generates at least one action item. If nothing changes, it will happen again.
• Never declare an incident resolved until a real user confirms the fix. Metrics can lie; users don't.
• All P1/P2 incidents require a post-mortem. No exceptions.