CrashLoopBackOff: how an AI agent diagnoses it with evidence

Start the investigation

/k8s-triage payment-service-7d9f is failing

What evidence gets gathered

The k8s-investigator agent reads pod status and events, the previous terminated container's exit code and reason, restart count, the last logs before the crash, current resource requests and limits, and — if the pod is GitOps-managed — the owning repository and manifest path via ArgoCD or Flux annotations.

A second agent checks the first

k8s-reviewer independently reaches its own conclusion from the same evidence, then compares notes with the investigator. The investigation can't validate itself — if the two disagree, that disagreement shows up in the output instead of being smoothed over.

When there isn't enough evidence, it says so

Each of OpsKit's 33 skills defines exactly what evidence is required before a conclusion, and what confidence level to report. If the logs were already rotated out or the exit code doesn't point anywhere conclusive, the investigation abstains explicitly rather than filling the gap with a plausible-sounding guess.

The fix, if there is one

For manifest-level root causes — wrong command, missing env var, bad image tag, an insufficient liveness probe timeout — k8s-fixer generates a minimal patch and opens a GitOps PR with the evidence summary, rollback plan, and verification checklist attached. Run /gitops-fix once the investigation has a hypothesis.