NEXUS — Autonomous Homelab SRE

Confidence is derived from observations, contradictions and required-evidence coverage — never invented by the model. When NEXUS cannot verify something, the honest output is “I could not verify disk usage because the host did not respond” — not “disk usage is normal.”

No arbitrary shell. Every capability is a typed tool with a permission class (READ_ONLY, LOW_RISK, REQUIRES_APPROVAL, FORBIDDEN). The model cannot bypass the permission layer because authorization is enforced outside the LLM.

REQUIRES_APPROVAL actions cannot run until a human says yes. Default mode is READ_ONLY — no destructive operations until the permission system is explicitly enabled.

The console shows reasoning summaries and the evidence collected at every step — never hidden chain-of-thought.

classify → context → evidence → hypothesis → test → root cause → remediation → approval → execute → verify → close. When evidence is insufficient the reasoning loop iterates instead of inventing a root cause just to finish.

FastAPI application factory with a live /api/v1/health database check, PostgreSQL 16 persistence (SQLAlchemy 2.0 async, Alembic migrations), Redis, structured logging with structlog, a sandbox Docker Compose stack, and CI running ruff, strict mypy and pytest against real PostgreSQL and Redis — with graceful degradation when a dependency is down.