# PROD stack: Speaches server + this UI + Caddy reverse proxy.
#
# Single origin via Caddy so the phone browser only ever talks to one host
# (no CORS, and the realtime WebSocket works through the proxy):
#   /v1/*            -> speaches:8000   (REST + WS /v1/realtime)
#   everything else  -> ui:3000         (TanStack Start node server)
#
# Pick CPU or GPU by layering an override:
#   docker compose -f deploy/compose.yaml -f deploy/compose.cpu.yaml  up -d
#   docker compose -f deploy/compose.yaml -f deploy/compose.cuda.yaml up -d
#
# `.env` in the project root supplies SPEACHES_URL etc. for the ui service.
# Speaches model cache is persisted in the hf-hub-cache named volume.

services:
  speaches:
    # Image/flags chosen by the override (compose.cpu.yaml / compose.cuda.yaml).
    # Sensible CPU defaults live here so the base file alone is runnable.
    image: ghcr.io/speaches-ai/speaches:latest-cpu
    restart: unless-stopped
    ports:
      - "8000:8000"            # direct access for CLI / curl smoke tests
    volumes:
      - hf-hub-cache:/home/ubuntu/.cache/huggingface/hub
    healthcheck:
      test: ["CMD", "curl", "--fail", "http://0.0.0.0:8000/health"]
      interval: 30s
      timeout: 10s
      retries: 3
      start_period: 30s

  ui:
    build:
      context: ..
      dockerfile: deploy/ui.Dockerfile
    restart: unless-stopped
    depends_on:
      speaches:
        condition: service_healthy
    environment:
      # Server-side: the UI proxies to Speaches over the compose network.
      - SPEACHES_URL=http://speaches:8000
      - NODE_ENV=production
      - PORT=3000
      - HOST=0.0.0.0
    expose:
      - "3000"

  caddy:
    image: caddy:2-alpine
    restart: unless-stopped
    depends_on:
      - ui
      - speaches
    ports:
      - "8080:80"              # phone entrypoint. Front with Tailscale/Tunnel for TLS.
    volumes:
      - ./Caddyfile:/etc/caddy/Caddyfile:ro
      - caddy-data:/data
      - caddy-config:/config

volumes:
  hf-hub-cache:
  caddy-data:
  caddy-config: