The problem
Incidents scatter signals across Kubernetes state, metrics, logs, traces and recent changes. Temporal correlation does not prove causality. This project gathers evidence without promoting guesses to diagnoses.
Architecture and approach
The Next.js workspace talks to a FastAPI backend that enforces authentication, tenant/cluster scope and query budgets. Read-only adapters investigate authorized sources. PostgreSQL persists state, history and evidence.
Deliberate guardrails
Deterministic evaluation owns the RCA verdict, with evidence-integrity checks and Kubernetes precedence. Model output remains advisory and must cite authorized evidence. Remediation lives behind a separate boundary requiring human approval.
What has been validated
The repository documents synthetic benchmarks and controlled disposable Kind investigations, including independently verified multiple causes. Screenshots dated October 22, 2026 show the running Next.js frontend connected to demo adapters or verified replays, not production incidents.
What still needs care
Agentic execution is opt-in and certain worker modes remain experimental. Missing or ambiguous telemetry should yield inconclusive results rather than unwarranted confidence.
Repository and technical docs ↗
← Back to home