services:
  mcspeak:
    build: .
    container_name: mcspeak
    restart: unless-stopped
    # Must exceed: 3s handler drain + TTS_SHUTDOWN_TIMEOUT + 5s safety margin.
    # Default: 3 + 30 + 5 = 38s. Increase if TTS_SHUTDOWN_TIMEOUT > 30.
    stop_grace_period: 38s
    env_file: .env
    # Publish to localhost only — Claude Code (and any local MCP client) talks
    # to 127.0.0.1:8371. External HTTPS access still flows through the caddy
    # label below if caddy-docker-proxy is running.
    ports:
      - "127.0.0.1:8371:8371"
    environment:
      # Override for Docker networking (container DNS instead of IPs)
      TTS_PIPER_HOST: piper-tts
      TTS_ORPHEUS_URL: http://llama-server:8081
      # PipeWire client config
      XDG_RUNTIME_DIR: /run/user/1000
      # PulseAudio socket for media ducking (bypass XDG_RUNTIME_DIR ownership check)
      PULSE_SERVER: unix:/run/user/1000/pulse/native
      # Force unbuffered Python output so stderr shows up in docker logs immediately
      PYTHONUNBUFFERED: "1"
    volumes:
      # Kokoro ONNX models (read-only)
      - ./models:/app/models:ro
      # HuggingFace cache for SNAC model download (lazy-loaded on first Orpheus call)
      - hf-cache:/home/tts/.cache/huggingface
      # Persistent data (voice assignments, etc.)
      - tts-data:/data
      # PipeWire socket for audio playback through host speakers
      - /run/user/1000/pipewire-0:/run/user/1000/pipewire-0
      # PulseAudio compat socket for media ducking (volume control)
      - /run/user/1000/pulse:/run/user/1000/pulse
    depends_on:
      llama-server:
        condition: service_healthy
        # required: false lets mcspeak start without llama-server when the
        # with-orpheus profile isn't active (kokoro-only mode is the default).
        required: false
    networks:
      - caddy
      - mcspeak-internal
    labels:
      caddy: mctalkbox.l.supported.systems
      caddy.reverse_proxy: "{{upstreams 8371}}"

  llama-server:
    # Opt-in: only starts when `docker compose --profile with-orpheus up`.
    # Default `make up` runs kokoro-only without GPU dependencies.
    profiles: ["with-orpheus"]
    build:
      context: .
      dockerfile: llama-server.Dockerfile
    container_name: orpheus-llama-server
    restart: unless-stopped
    deploy:
      resources:
        reservations:
          devices:
            - driver: nvidia
              count: 1
              capabilities: [gpu]
    volumes:
      # GGUF model file (set ORPHEUS_GGUF_PATH in .env, e.g. from Ollama blob
      # storage). The /dev/null fallback lets compose parse cleanly when the
      # with-orpheus profile is inactive; activating the profile without
      # setting ORPHEUS_GGUF_PATH will produce a clear "model load failed" at
      # llama-server startup rather than a confusing compose parse error.
      - ${ORPHEUS_GGUF_PATH:-/dev/null}:/models/orpheus.gguf:ro
    command: >-
      --host 0.0.0.0 --port 8081
      --model /models/orpheus.gguf
      --n-gpu-layers 999 --ctx-size 4096
      --flash-attn --cont-batching
    networks:
      - mcspeak-internal
    healthcheck:
      test: ["CMD", "curl", "-sf", "http://127.0.0.1:8081/health"]
      interval: 15s
      timeout: 5s
      start_period: 120s
      retries: 5

volumes:
  hf-cache:
  tts-data:

networks:
  caddy:
    external: true
  # Private per-stack network for mcspeak ↔ llama-server. Auto-created by
  # compose, isolated from other stacks (no DNS leak via shared caddy).
  mcspeak-internal:
    internal: true