compose.yaml

name: self-improving
services:
  # The production to-do app. Runs a Next.js production build with no source
  # code, no agent, and no admin UI. Exactly one ingress port keeps it
  # stateless and scale-friendly (Cloud Run on GCP, Fargate on AWS).
  app:
    restart: unless-stopped
    build:
      context: ./todo-app
      dockerfile: Dockerfile
    ports:
      - target: 3000
        published: 3000
        mode: ingress
    environment:
      DATABASE_URL: postgres://postgres:${POSTGRES_PASSWORD}@db:5432/postgres?sslmode=no-verify
      BETTER_AUTH_SECRET:
      ADMIN_UI: "false"
    depends_on:
      db:
        condition: service_healthy
    healthcheck:
      test:
        - CMD
        - curl
        - -f
        - http://localhost:3000/api/health
      interval: 15s
      timeout: 5s
      retries: 10
      start_period: 20s
    deploy:
      resources:
        reservations:
          cpus: "0.5"
          memory: 1024M

  # The live "dev environment" service: Caddy in front of a Next.js dev server
  # plus a separate Hono server running the Mastra coding agent, which edits
  # the to-do app's source in place (hot reload). This service carries the full
  # sample source tree. replicas: 1 and the private host port pin it to a single
  # always-on instance so everyone previews the same edited copy (and so GCP
  # uses Compute Engine rather than a scale-to-zero Cloud Run service).
  dev:
    restart: unless-stopped
    build:
      context: .
      dockerfile: Dockerfile.dev
    ports:
      - target: 3000
        published: 3000
        mode: ingress
      - target: 8081
        mode: host
    environment:
      DATABASE_URL: postgres://postgres:${POSTGRES_PASSWORD}@db:5432/postgres?sslmode=no-verify
      BETTER_AUTH_SECRET:
      ADMIN_UI: "true"
      # Break-glass token for the admin console (served by the agent server,
      # outside the app the agent edits). Lets an admin reach the console even
      # if the main app is down and their login session has expired.
      ADMIN_TOKEN:
      # Enable the admin-triggered self-redeploy. Fabric auth is a per-publish
      # interactive `defang login` (no stored token). Defang supplies the
      # resolved provider and stack so self-redeploys target the same place.
      PUBLISH_ENABLED: "true"
      PUBLISH_PROVIDER: "${DEFANG_PROVIDER}"
      PUBLISH_STACK: "${DEFANG_STACK}"
      # CHAT_URL, CHAT_MODEL, and OPENAI_API_KEY (the gateway's master key) are
      # injected automatically by the `models:` binding below; the coding agent
      # reads them in agent/src/model.ts.
    models:
      chat:
        endpoint_var: CHAT_URL
        model_var: CHAT_MODEL
    depends_on:
      db:
        condition: service_healthy
    healthcheck:
      test:
        - CMD
        - curl
        - -f
        - http://localhost:3000/api/health
      interval: 15s
      timeout: 5s
      retries: 10
      start_period: 60s
    deploy:
      replicas: 1
      resources:
        reservations:
          cpus: "1.0"
          memory: 4096M
    # Self-publish uses the workload's ambient identity. The selected cloud
    # env file supplies either an AWS managed-policy ARN or a GCP project role.
    # The empty default is only for repository validation; every cloud deploy
    # must use a committed stack, which supplies the policy.
    x-defang-policies:
      - ${PUBLISH_POLICY:-}

  # Shared database: todos, users/sessions (Better Auth), feedback, and the
  # Mastra agent's storage. Both app and dev point here.
  db:
    image: postgres:16
    x-defang-postgres: true
    restart: always
    environment:
      POSTGRES_PASSWORD:
      POSTGRES_DB: postgres
      POSTGRES_USER: postgres
    ports:
      - mode: host
        target: 5432
    healthcheck:
      test:
        - CMD-SHELL
        - pg_isready -U postgres -d postgres
      interval: 10s
      timeout: 5s
      retries: 10
      start_period: 10s
    deploy:
      resources:
        reservations:
          cpus: "0.5"
          memory: 512M

# Managed LLM the coding agent talks to. Defang synthesizes a LiteLLM access
# gateway from this entry and injects CHAT_URL / CHAT_MODEL / OPENAI_API_KEY
# into any service that binds it (see `dev`). `chat-default` is a
# provider-portable alias the CLI maps to the cloud's managed inference
# (Amazon Nova on AWS Bedrock, Gemini on GCP Vertex AI). The gateway uses the
# CLI's default memory reservation.
models:
  chat:
    model: ${CHAT_MODEL:-chat-default}