Incident Response
FeaturedInvestigates a production incident, tests competing explanations in parallel, has the root cause challenged before reporting it, and proposes a bounded mitigation.
How it works
- Runs when: an incident is opened, or you start it with an alert or a question.
- Gathers evidence: error rates, latency, saturation and logs for the affected services, plus every deploy and config change in the two hours before onset.
- Tests explanations: sends two or three competing hypotheses to subagents in parallel, so no line of inquiry colours another.
- Challenges its answer: a verifier subagent argues against the surviving cause link by link. A refuted or weakened cause goes back for more work.
- You get: the root cause as a causal chain of at most five links, each citing its evidence, with a severity, a confidence and a proposed mitigation. It never runs the mitigation itself.
What it needs
- Observability (Datadog, Prometheus or Sentry): read metrics, logs and errors
- GitHub: read deploys and recent changes
- PagerDuty: read the incident