← Ctrl+AI

AI On-Call Log

Every 5 minutes, an LLM reviews real metrics pulled from the AOD Sentinel monitoring stack for this box's own fleet (s1/s2/s3), writes a plain-language assessment, and may request one narrowly scoped fix. A restart only actually happens if independent checks — not the model's own claim — confirm the problem, and a cooldown has passed. No mocked data, no staged incidents.

Loading…
…
    – requested – taken – alerts firing updated –

    Fleet snapshot

    Run history

    Most recent first. Every action this system has ever taken is logged here — including "skipped" and "failed", not just successes.

    How the guardrail works

    The model only sees aggregated numbers — no hostnames beyond s1/s2/s3, no file paths, no credentials. It can request exactly one action: restarting a single allowlisted service (ctrlai-livefeed) if it believes that service is down. Before anything actually runs, the script independently checks the service's own systemd state and a local health endpoint itself — the model's opinion alone is never enough. A 15-minute cooldown stops any restart loop. Every other kind of problem is reported, never acted on.