services:
  backend:
    build:
      context: .
      dockerfile: backend/Dockerfile
      target: api
    env_file:
      - .env
      - .env.backend
    ports:
      - "${BACKEND_PORT}:${BACKEND_PORT}"
    volumes:
      - ./backend:/app
      # Artifacts and datasets are files. Without a volume they die with the
      # container, leaving rows in Postgres pointing at bytes that no longer
      # exist — which is exactly what happened to every run made before this.
      - storage:/var/lib/agentics/storage
    depends_on:
      postgres:
        condition: service_healthy
    networks:
      - app-network
    restart: unless-stopped

  # Consumes runs.fifo. Scaling this past a couple of replicas buys nothing:
  # FIFO keeps one message per group in flight, so useful concurrency here is
  # bounded by the number of runs in progress, and an advance is milliseconds.
  orchestrator:
    build:
      context: .
      dockerfile: backend/Dockerfile
      target: consumer
    command: python -m workers.orchestrator
    env_file:
      - .env
      - .env.backend
    volumes:
      - ./backend:/app
    depends_on:
      postgres:
        condition: service_healthy
      sqs:
        condition: service_started
    networks:
      - app-network
    # No restart policy yet: while the loop is new, a crash should stay visible
    # rather than being hidden by a restart loop.

  # Consumes tasks. This is the half worth scaling — `--scale worker=6` — since
  # the queue has no message groups and the work is sandboxed code and model
  # calls. It needs the storage volume for the same reason the API does.
  worker:
    build:
      context: .
      dockerfile: backend/Dockerfile
      target: consumer
    command: python -m workers.task
    env_file:
      - .env
      - .env.backend
    volumes:
      - ./backend:/app
      - storage:/var/lib/agentics/storage
    depends_on:
      postgres:
        condition: service_healthy
      sqs:
        condition: service_started
    networks:
      - app-network

  postgres:
    image: postgres:15
    container_name: postgres
    restart: unless-stopped
    environment:
      POSTGRES_USER: ${POSTGRES_USER}
      POSTGRES_PASSWORD: ${POSTGRES_PASSWORD}
      POSTGRES_DB: ${POSTGRES_DB}
    ports:
      - "${POSTGRES_PORT:-5432}:5432"
    volumes:
      - pgdata:/var/lib/postgresql/data
    networks:
      - app-network
    healthcheck:
      test: ["CMD-SHELL", "pg_isready -U ${POSTGRES_USER}"]
      interval: 5s
      timeout: 5s
      retries: 5

  # Local stand-in for SQS, speaking the real SQS wire protocol so the queue
  # semantics that matter — redelivery, visibility timeouts, FIFO groups — are
  # the same ones production will have. Deploying is an endpoint URL change.
  sqs:
    image: softwaremill/elasticmq-native
    restart: unless-stopped
    ports:
      - "9324:9324"   # SQS API
      - "9325:9325"   # built-in queue UI, handy for watching messages move
    volumes:
      - ./docker/elasticmq.conf:/opt/elasticmq.conf:ro
    networks:
      - app-network

  # dynamodb-local:
  #   image: amazon/dynamodb-local:latest
  #   command: "-jar DynamoDBLocal.jar -sharedDb -dbPath ./data"
  #   ports:
  #     - "8005:8000"
  #   volumes:
  #     - "./docker/dynamodb:/home/dynamodblocal/data"

  pgadmin:
    image: dpage/pgadmin4
    ports:
      - "5050:${PGADMIN_PORT}"
    environment:
      PGADMIN_DEFAULT_EMAIL: ${PGADMIN_EMAIL}
      PGADMIN_DEFAULT_PASSWORD: ${PGADMIN_PASSWORD}
    depends_on:
      - postgres
    networks:
      - app-network
    restart: unless-stopped

networks:
  app-network:
    driver: bridge

volumes:
  storage:
    driver: local
  pgdata: