Skip to content

Releases

server v0.1.0

Choose a tag to compare

@github-actions github-actions released this 28 Oct 07:20
Immutable release. Only release title and notes can be modified.
27c7004

Added

  • Health check at /health, JSON errors, X-Request-ID on every response, 5 MB request body limit.
  • Container image for linux/amd64 and linux/arm64 on GitHub Container Registry: ghcr.io/cerebellum-labs/cerebellum-server.
  • PostgreSQL 18 database, required: DATABASE_URL. The server applies its migrations on start.
  • Agent API under /api/v1, see AGENT_API.md: /decide asks the cloud model for irreversible decisions, /sync takes everything else and returns corrections. Devices sign in with their token. Uncertain decisions are reviewed in the background, and a review that disagrees becomes a correction. /decide asks without reasoning for a quick answer; the review lets the model reason. Replies are requested in JSON mode. The cloud model is set by NEBIUS_API_KEY, with NEBIUS_BASE_URL, MODEL_ID and the MODEL_PRICE_*_PER_M prices defaulting to Nemotron 3.5 Super VL on Nebius Token Factory.
  • Dashboard API under /api/dashboard: devices, the journal of decisions with frames, statistics and a WebSocket stream of changes. Changing devices takes ADMIN_PASSWORD; the session is a cookie signed with SESSION_SECRET. DASHBOARD_ORIGIN lets a dashboard on a host of its own call the API.
  • local_latency_ms in the decision the agent sends: how long the local model took. It comes back as latency_ms of a device decision, and the dashboard's statistics average it.
  • The journal filters by source, and changed=true keeps the decisions where the cloud disagreed with the robot.
  • cloud_cost_usd in every decision: all it cost in the cloud, retries, failed calls and the review included.
  • A decision's detail in the dashboard API lists every call made to the cloud model for it, with the raw reply or the error.

Changed

  • AGENT_API.md: a timeout, a 5xx or no network is no longer a failed attempt of an irreversible decision; it waits for the network, and only a retry: true answer counts. Documentation only: the server keeps no attempts.

Image: ghcr.io/cerebellum-labs/cerebellum-server:0.1.0