node:
  host_id: "singlebox-llamafile"
  display_name: "SingleBox llamafile"
  listen_host: "127.0.0.1"
  listen_port: 8891
  address: "127.0.0.1"
  labels:
    topology: "singlebox"
    runtime: "llamafile"

control_plane:
  base_url: "http://127.0.0.1:8800"
  node_api_key: "change-me-node-key"
  heartbeat_interval_s: 5

inventory:
  model_roots:
    - "/path/to/models"
  cpu_threads: 24
  ram_gb: 64
  capabilities:
    cpu: true
    cuda: true

managed_runtimes:
  enabled: false

services:
  - service_id: "singlebox/chat/qwen3-8b"
    kind: "chat"
    endpoint: "http://127.0.0.1:18091"
    runtime:
      engine: "llamafile"
      launcher: "external"
    assets:
      - asset_id: "qwen3-8b-q4_k_m"
        loaded: true
    state:
      health: "healthy"
      load_state: "loaded"
      accept_requests: true
    observed:
      p50_latency_ms: 900