# SurakshaMCP — runs the NitroStack MCP server + the Python inference service together.
services:
  inference:
    build: ./inference-service
    ports:
      - "8000:8000"
    environment:
      - PPE_WEIGHTS=/models/ppe_classifier.tflite  # trained multi-label classifier (see ppe_detector.py)
      - LANGID_WEIGHTS=/models/langid.pt
      - TTS_BACKEND=gtts                          # gtts (implemented) | indic_tts (stub — see tts_engine.py)
    volumes:
      - ./models:/models
    healthcheck:
      test: ["CMD", "python", "-c", "import urllib.request,sys; sys.exit(0 if urllib.request.urlopen('http://localhost:8000/health').status==200 else 1)"]
      interval: 15s
      timeout: 5s
      retries: 5

  mcp-server:
    build: ./mcp-server
    ports:
      - "3000:3000"
    environment:
      - INFERENCE_URL=http://inference:8000
      - MCP_API_KEY=${MCP_API_KEY:-dev-key-change-me}
      - NODE_ENV=production
      - SURAKSHA_DB_PATH=/app/data/suraksha.db
      # WhatsApp alert channel (optional — tool falls back to dry-run without these)
      - TWILIO_ACCOUNT_SID=${TWILIO_ACCOUNT_SID:-}
      - TWILIO_AUTH_TOKEN=${TWILIO_AUTH_TOKEN:-}
      - TWILIO_WHATSAPP_FROM=${TWILIO_WHATSAPP_FROM:-}
    volumes:
      - ./data:/app/data          # persists the SQLite incident/training/alert DB
    depends_on:
      inference:
        condition: service_healthy