services:
  postgres:
    image: pgvector/pgvector:pg17
    environment:
      POSTGRES_DB: osinara
      POSTGRES_PASSWORD: ${POSTGRES_PASSWORD:?POSTGRES_PASSWORD is required}
      POSTGRES_USER: osinara
    healthcheck:
      test: ["CMD-SHELL", "pg_isready -U osinara -d osinara"]
      interval: 5s
      timeout: 3s
      retries: 20
    restart: unless-stopped
    volumes:
      - postgres-data:/var/lib/postgresql/data

  agent:
    build:
      context: .
      target: runtime
    depends_on:
      migrate:
        condition: service_completed_successfully
      memory-embedding:
        condition: service_healthy
      sandbox-runner:
        condition: service_healthy
    environment:
      DATABASE_URL: postgresql://osinara:${POSTGRES_PASSWORD}@postgres:5432/osinara
      ELEVENLABS_API_KEY: ${ELEVENLABS_API_KEY-}
      GROQ_API_KEY: ${GROQ_API_KEY-}
      GOOGLE_OAUTH_CLIENT_ID: ${GOOGLE_OAUTH_CLIENT_ID-}
      GOOGLE_OAUTH_CLIENT_SECRET: ${GOOGLE_OAUTH_CLIENT_SECRET-}
      INTEGRATION_TOKEN_ENCRYPTION_KEY: ${INTEGRATION_TOKEN_ENCRYPTION_KEY-}
      INVITATION_SIGNING_SECRET: ${INVITATION_SIGNING_SECRET:?INVITATION_SIGNING_SECRET is required}
      KLIPY_API_KEY: ${KLIPY_API_KEY-}
      MEMORY_EMBEDDING_BASE_URL: http://memory-embedding:80
      MODEL_API_KEY: ${MODEL_API_KEY:?MODEL_API_KEY is required}
      PUBLIC_BASE_URL: ${PUBLIC_BASE_URL-}
      TELEGRAM_BOT_TOKEN: ${TELEGRAM_BOT_TOKEN:?TELEGRAM_BOT_TOKEN is required}
      TELEGRAM_BOT_USERNAME: ${TELEGRAM_BOT_USERNAME:?TELEGRAM_BOT_USERNAME is required}
      TELEGRAM_WEBHOOK_SECRET_TOKEN: ${TELEGRAM_WEBHOOK_SECRET_TOKEN:?TELEGRAM_WEBHOOK_SECRET_TOKEN is required}
    healthcheck:
      test: ["CMD", "node", "-e", "fetch('http://127.0.0.1:3000/v1/health').then(r=>{if(!r.ok)process.exit(1)}).catch(()=>process.exit(1))"]
      interval: 5s
      timeout: 3s
      retries: 72
    restart: unless-stopped
    networks:
      - default
      - sandbox-control
    volumes:
      - ./config/agent-model-providers.json:/app/config/agent-model-providers.json:ro
      - google-workspace-credentials:/app/google-workspace-credentials
      # A host directory, not a named volume: `npm run dev` on the host writes the same grant file.
      - ./network-access:/app/network-access
      - workspace-data:/app/workspaces

  migrate:
    build:
      context: .
      target: runtime
    depends_on:
      postgres:
        condition: service_healthy
    entrypoint: ["npm", "run", "migrate:runtime"]
    environment:
      DATABASE_URL: postgresql://osinara:${POSTGRES_PASSWORD}@postgres:5432/osinara
    restart: "no"

  sandbox-runtime-image:
    build:
      context: .
      target: sandbox-runtime
    command: ["true"]
    image: osinara-sandbox-runtime:local
    network_mode: none
    restart: "no"

  sandbox-runner:
    build:
      context: .
      target: sandbox-runner
    depends_on:
      sandbox-egress-proxy:
        condition: service_healthy
      sandbox-runtime-image:
        condition: service_completed_successfully
    healthcheck:
      test: ["CMD", "node", "-e", "fetch('http://127.0.0.1:8080/health').then(r=>{if(!r.ok)process.exit(1)})"]
      interval: 5s
      timeout: 3s
      retries: 20
    environment:
      SANDBOX_RUNTIME_IMAGE: osinara-sandbox-runtime:local
    networks:
      - sandbox-control
    read_only: true
    restart: unless-stopped
    security_opt:
      - no-new-privileges:true
    stop_grace_period: 30s
    tmpfs:
      - /tmp:size=64m,mode=1777
    volumes:
      - /var/run/docker.sock:/var/run/docker.sock
      - tool-environments:/runner/tools
      - workspace-data:/runner/workspaces

  sandbox-egress-proxy:
    build:
      context: .
      target: sandbox-egress-proxy
    cap_drop:
      - ALL
    healthcheck:
      test: ["CMD", "node", "-e", "require('node:net').connect(3128,'127.0.0.1').on('connect',function(){this.destroy();process.exit(0)}).on('error',()=>process.exit(1))"]
      interval: 5s
      timeout: 3s
      retries: 20
    networks:
      default:
      sandbox-egress:
        # Resolves only on the trusted sandbox network: the proxy finds its own address there and
        # opens owner-approved IP destinations only to requests that arrive on it.
        aliases:
          - sandbox-trusted-egress-proxy
    read_only: true
    restart: unless-stopped
    security_opt:
      - no-new-privileges:true
    tmpfs:
      - /tmp:size=32m,mode=1777
    volumes:
      - ./network-access:/app/network-access:ro

  memory-embedding:
    image: ghcr.io/huggingface/text-embeddings-inference:cpu-1.9@sha256:ad950d30878eceb72aaf32024d26fa2b1d04a75304fa0b4776b49aa1941fea07
    # Sized by measurement on a six-core host, not by guess. Memory is set by the math threads
    # alone: three cores cost eleven megabytes more than two, while a third math thread costs four
    # hundred and does not fit a gigabyte at all — the container is killed while loading the model.
    # Two threads peak at 870 MB under a saturating load. The CPU budget is separate and cheap:
    # tokenization and request handling use the third core without touching that peak, and it buys
    # 45% more indexing throughput and a third off the wait when a search queues behind a batch.
    cpus: 3.0
    mem_limit: 1024m
    command:
      - --model-id
      - intfloat/multilingual-e5-small
      - --revision
      - 614241f622f53c4eeff9890bdc4f31cfecc418b3
      - --tokenization-workers
      - "2"
      - --max-client-batch-size
      - "8"
      # One queue slot per input, so a client batch of eight fills a limit of eight and the search
      # arriving beside it is refused rather than queued — the turn silently loses its semantic
      # branch. Measured against a saturating indexer: at eight, 15 searches out of 60 were refused;
      # at sixteen, none, at no cost in memory.
      - --max-concurrent-requests
      - "16"
      - --max-batch-requests
      - "8"
      - --max-batch-tokens
      - "4096"
      - --auto-truncate=false
    environment:
      OMP_NUM_THREADS: "2"
    healthcheck:
      test: ["CMD", "curl", "--fail", "http://localhost/health"]
      interval: 5s
      timeout: 5s
      retries: 120
    restart: unless-stopped
    volumes:
      - memory-embedding-model-e5:/data

  memory-embedding-worker:
    build:
      context: .
      target: runtime
    depends_on:
      migrate:
        condition: service_completed_successfully
      memory-embedding:
        condition: service_healthy
    entrypoint: ["node", ".runtime/scripts/memory-embedding-worker.js"]
    environment:
      DATABASE_URL: postgresql://osinara:${POSTGRES_PASSWORD}@postgres:5432/osinara
      MEMORY_EMBEDDING_BASE_URL: http://memory-embedding:80
    restart: unless-stopped

  telegram-ingress-worker:
    build:
      context: .
      target: runtime
    depends_on:
      - agent
    entrypoint: ["node", ".runtime/scripts/telegram-ingress-worker.js"]
    environment:
      AGENT_INTERNAL_BASE_URL: http://agent:3000
      TELEGRAM_WEBHOOK_SECRET_TOKEN: ${TELEGRAM_WEBHOOK_SECRET_TOKEN:?TELEGRAM_WEBHOOK_SECRET_TOKEN is required}
    restart: unless-stopped

  edge:
    image: nginx:1.29-alpine
    depends_on:
      - agent
    ports:
      - "8080:80"
    restart: unless-stopped
    volumes:
      - ./infra/nginx.conf:/etc/nginx/nginx.conf:ro

volumes:
  google-workspace-credentials:
  postgres-data:
  memory-embedding-model-e5:
  tool-environments:
  workspace-data:

networks:
  sandbox-control:
    internal: true
  sandbox-egress:
    internal: true