# Vui streaming voice assistant — docker-compose.
#
# Default: `vui-stream` + `claude-task` start. `vui-stream` uses host
# networking and points at `http://localhost:11434`, so it talks to whatever
# Ollama you have on the host (recommended — start with `ollama serve` or
# your distro's service).
#
# Quick start:
#     ollama pull qwen3.5:4b           # on the HOST
#     docker compose up -d
#     open http://localhost:8080
#
# Don't have Ollama on the host? Start the bundled containerised one:
#     docker compose --profile ollama up -d
#     docker compose exec ollama ollama pull qwen3.5:4b
#
# `claude-task` (Gmail / Calendar / web research delegation) starts by
# default. Without auth (`~/.claude` login or ANTHROPIC_API_KEY) it stays up
# but task requests fail at runtime; the voice loop is unaffected. See readme.

services:
  # Optional bundled Ollama. Off by default — assumes the user has Ollama on
  # the host. Enable with `--profile ollama` if you want it containerised.
  ollama:
    image: ollama/ollama:latest
    container_name: vui-ollama
    profiles: ["ollama"]
    ports:
      - "11434:11434"
    volumes:
      - ollama-models:/root/.ollama
    deploy:
      resources:
        reservations:
          devices:
            - driver: nvidia
              count: all
              capabilities: [gpu]
    restart: unless-stopped

  vui-stream:
    build:
      context: .
      dockerfile: docker/Dockerfile.stream
    container_name: vui-stream
    # Override default CMD to load a local checkpoint instead of fetching
    # vui-nano.safetensors from huggingface.co/fluxions/vui (not yet published).
    # Drop this line once the HF repo has the v2 weights.
    # Path is `<run>/<step>.pt` so `_ckpt_id` parses to `7qx553s3_0060000`
    # and matches the cached compiled-prompt filenames in `prompts/`.
    command: ["python", "-m", "vui.serving.stream", "/app/checkpoints/7qx553s3/0060000.pt"]
    # WebRTC needs the browser to reach the server's media UDP sockets
    # directly. With bridged networking, aiortc's ICE host candidates
    # advertise the container's bridge IP and the browser usually can't
    # reach them — connection stays in "connecting" forever. Host net
    # avoids the NAT/UDP problem entirely. Trade-off: `ports:` is ignored
    # and the container shares the host's network namespace, so we just
    # talk to the host's Ollama at localhost:11434 directly.
    network_mode: host
    volumes:
      # HuggingFace cache so the Qwen codec only downloads once.
      - hf-cache:/cache/huggingface
      # Local checkpoint mount (see `command:` above).
      - ./checkpoints:/app/checkpoints:ro
      # Persist saved voice prompts across restarts.
      - ./prompts:/app/prompts
      # Optional: write conversation logs to host for inspection.
      - ./debug_dump:/app/debug_dump
      # Thoughts-stream tools dir — shared rw with claude-task so codegen
      # writes (via the propose_tool action) land on the host filesystem
      # and are visible to vui-stream's /tools/reload.
      - ./src/vui/serving/stream/tools:/app/src/vui/serving/stream/tools
    environment:
      # localhost works for both the host Ollama and the optional bundled one
      # (which publishes :11434 on the host). `OLLAMA_URL` is kept as the
      # fallback name the MLX auto-setup helper still reads directly.
      OLLAMA_URL: "http://localhost:11434"
      VUI_OLLAMA_URL: "http://localhost:11434"
      VUI_OLLAMA_MODEL: "qwen3.5:4b"
      VUI_TASK_SERVER_URL: "http://localhost:8642"
      # Cap conversation context, in seconds of audio. Budget is shared
      # between audio (~12.5 tok/s) and text (~31 tok/s headroom).  Default
      # below fits one full turn (prompt ~10s + user ~30s + 2 min reply)
      # comfortably, with multi-turn history off (`keep_context=false`).
      # Trained ceiling is ~240s.
      VUI_MAX_CONTEXT_SECS: "180"
    deploy:
      resources:
        reservations:
          devices:
            - driver: nvidia
              count: all
              capabilities: [gpu]
    restart: unless-stopped

  # Claude task delegation — starts by default. Mounts ~/.claude for auth;
  # falls back to ANTHROPIC_API_KEY / CLAUDE_CODE_OAUTH_TOKEN env vars.
  # Host networking (matches vui-stream) so the result callback POST to
  # http://localhost:8080/task_done resolves to vui-stream on the host.
  # Without it, "localhost" inside this container loops back to itself and
  # task results never reach the voice server.
  claude-task:
    build:
      context: .
      dockerfile: docker/Dockerfile.claude
    container_name: vui-claude-task
    network_mode: host
    volumes:
      # Claude Code subscription credentials (after `claude` login on host).
      - "${HOME}/.claude:/root/.claude:ro"
      # Thoughts-stream tools dir — same mount as vui-stream so
      # propose_tool codegen writes show up live on both sides.
      - ./src/vui/serving/stream/tools:/app/src/vui/serving/stream/tools
    environment:
      ANTHROPIC_API_KEY: "${ANTHROPIC_API_KEY:-}"
      CLAUDE_CODE_OAUTH_TOKEN: "${CLAUDE_CODE_OAUTH_TOKEN:-}"
    restart: unless-stopped

volumes:
  ollama-models:
  hf-cache: