Live demo

This is the real dashboard - the same page the doctor serves - replaying a recording of a made-up home server called atlas: 22 containers in eight stacks, a two-disk mirror and an NVMe system drive, and two weeks of history. Open the incidents, the containers, the storage and the model’s call log; it behaves as it would on your host.

Open the demo full screen

What is going on in it

Paperless is down: its Postgres database keeps failing its health check after hitting its memory limit, and a restart of it is waiting for approval. Sonarr’s indexer rejects its API key, /srv/data is full in about 17 days, and one disk has remapped sectors. The classifier runs in on: it hid Uptime Kuma’s errors about the sites it monitors as noise, raised the database’s failing health check to critical, and lowered paperless’s restart loop to a warning, since the database is the cause. Earlier in the week the doctor restarted a hung Jellyfin and a locked Uptime Kuma, and checked that both came back.

What is real and what is not

Everything you see was produced by the doctor’s own code: the rules opened the incidents, the policy proposed the restart, the briefing was written from the same facts, and the memory suggestions come from two weeks of samples. What stood in for the outside world: Docker, the host’s /proc and /sys, the SMART exporter, the registries, the network checks - and the models, the diagnosis’s and the classifier’s, whose answers are canned. The recipe is in the repository, under demo/.

Being a recording, it changes nothing: approving, dismissing, refreshing or saving a setting answers with a note that this is a demo, and times shift so the recording always looks recent.