fleet-health

Fleet Health

Fast, read-only health snapshot across all 5 machines in the personal fleet. Unlike spark-audit / jetson-audit (deep, single-machine, best-practices audits), this skill is a wide, shallow, fleet-wide pulse check — designed to answer "is everything up?" in well under a minute, with a final line a headless cron job can grep for.

This skill reads the live system. It never modifies it.

No paths: activation is configured — this skill runs only on explicit invocation ("fleet status", "morning brief", "is the spark available", etc.). Since it never auto-triggers, no loop guard is needed.

Fleet Config

ssh:
  key: "~/.ssh/id_claude_code"
  opts: "-o ConnectTimeout=5 -o BatchMode=yes"
  user: "claude"

curl:
  opts: "-sf -m 5"

fleet:
  - name: "DGX Spark"
    host: "spark.k4jda.net"
    lan_ip: "192.168.10.33"
    role: "LLM inference (vLLM) — chat, embeddings, NER"
    gpu: nvidia-smi
    key_ports:
      8000: { service: "spark-llm (Qwen3.6-35B-A3B-FP8)", check: "GET /v1/models, expect data[0].id == spark-llm" }
      8001: { service: "qwen3-embedding-4b",               check: "GET /v1/models, expect data[0].id == qwen3-embedding-4b" }
      8002: { service: "GLiNER NER",                       check: "GET /health, expect status == ok" }

  - name: "Jetson Orin Nano"
    host: "jetson.k4jda.net"
    lan_ip: "192.168.10.58"
    role: "Edge LLM inference (llama.cpp), single-process mode-switched"
    gpu: tegrastats   # NO nvidia-smi on this box — do not attempt it
    key_ports:
      8080: { service: "active llama-server mode (default qwen35)", check: "GET /health, expect status == ok" }
      8081: { service: "embedding mode (Qwen3-Embedding-4B)",        check: "see Known States — U2: expected DOWN unless mode.txt=embedding" }

  - name: "Homeserver"
    host: "homeserver.k4jda.net"
    role: "Docker host (Unraid) — ~55 containers, Prometheus/Grafana/Alertmanager stack"
    sudo_scoped: true   # `sudo -n docker ...` works; `sudo cat`/`sudo ls` do NOT — use plain cat/ls or docker's own output
    use_wget_locally: true  # curl is not reliably available ON the homeserver itself — use wget when SSH'd in

  - name: "Bond"
    host: "bond.k4jda.net"
    role: "General purpose"

  - name: "open-brain-vm (obvm)"
    host: "obvm.k4jda.net"
    role: "T2 Claude CLI batch, Python, ops"