docker-compose.prod.yml

# AF Stack production docker-compose.
#
# What this gets you in 3 bullets:
#   - Runtime + dashboard + AgentField control plane behind Caddy with
#     automatic Let's Encrypt TLS on {$AF_STACK_DOMAIN}.
#   - NO Postgres or MinIO services — point at managed/external ones via
#     AF_STACK_DATABASE_URL + AF_STACK_S3_*.
#   - NO docker.sock mount on the runtime. Sandbox runs via e2b (default)
#     or gvisor adapter. Stronger isolation, no privileged container.
#
# Required env (set in .env next to this file):
#   AF_STACK_DOMAIN              public hostname for Caddy
#   ACME_EMAIL                   Let's Encrypt registration email
#   AF_STACK_DATABASE_URL        postgres://user:pass@host:5432/afstack?sslmode=require
#   AGENTFIELD_STORAGE_POSTGRES_URL  postgres://user:pass@host:5432/agentfield?sslmode=require
#   AF_STACK_KMS_KEY             openssl rand -hex 32
#   AF_STACK_AUTH_SECRET         openssl rand -hex 32
#   AF_STACK_S3_ENDPOINT         https://fly.storage.tigris.dev (or R2/AWS)
#   AF_STACK_S3_BUCKET           af-stack-prod
#   AF_STACK_S3_ACCESS_KEY       <s3 access key>
#   AF_STACK_S3_SECRET_KEY       <s3 secret key>
#   OPENROUTER_API_KEY           (or ANTHROPIC_API_KEY / OPENAI_API_KEY)
#   E2B_API_KEY                  required when AF_STACK_SANDBOX_ADAPTER=e2b
#
# Verify the file BEFORE deploying:
#   docker compose -f docker-compose.prod.yml config --quiet
#
# Bring up:
#   docker compose -f docker-compose.prod.yml up -d
#
# Notes on resource limits:
#   `deploy.resources.limits` are enforced ONLY in swarm mode. In plain
#   `docker compose up` they are advisory. To enforce on a single node,
#   either run `docker stack deploy -c docker-compose.prod.yml af-stack`
#   or migrate the limits to `mem_limit` / `cpus` (compose v2.x supports
#   both natively; we keep `deploy.resources.limits` so a swarm upgrade is
#   a one-liner).

services:
  caddy:
    image: caddy:2-alpine
    restart: unless-stopped
    ports:
      - "80:80"
      - "443:443"
      - "443:443/udp"  # HTTP/3
    environment:
      AF_STACK_DOMAIN: ${AF_STACK_DOMAIN}
      ACME_EMAIL: ${ACME_EMAIL}
      METRICS_ALLOW: ${METRICS_ALLOW:-10.0.0.0/8}
    volumes:
      - ./deploy/caddy/Caddyfile:/etc/caddy/Caddyfile:ro
      - caddy-data:/data
      - caddy-config:/config
    networks:
      - frontend
    depends_on:
      - runtime
      - dashboard
    deploy:
      resources:
        limits:
          cpus: "0.50"
          memory: "256M"

  agentfield:
    # Pinned off :latest for reproducible prod deploys. Override AGENTFIELD_VERSION
    # to move the AgentField control plane deliberately (published on Docker Hub).
    image: agentfield/control-plane:${AGENTFIELD_VERSION:-v0.1.96}
    restart: unless-stopped
    environment:
      AGENTFIELD_STORAGE_MODE: postgres
      AGENTFIELD_STORAGE_POSTGRES_URL: ${AGENTFIELD_STORAGE_POSTGRES_URL}
      AGENTFIELD_STORAGE_POSTGRES_ENABLE_MEMORY_FALLBACK: "false"
      AGENTFIELD_STORAGE_POSTGRES_ENABLE_DID_FALLBACK: "false"
      AGENTFIELD_STORAGE_POSTGRES_ENABLE_VC_FALLBACK: "false"
      AGENTFIELD_STORAGE_POSTGRES_ENABLE_AUTO_MIGRATION: "true"
      OPENROUTER_API_KEY: ${OPENROUTER_API_KEY:-}
      ANTHROPIC_API_KEY: ${ANTHROPIC_API_KEY:-}
      OPENAI_API_KEY: ${OPENAI_API_KEY:-}
    networks:
      - backend
    deploy:
      resources:
        limits:
          cpus: "1.0"
          memory: "512M"

  runtime:
    # Pinned off :latest. AF_STACK_VERSION selects the runtime + dashboard image
    # tag together; set it to the release CI publishes (e.g. a git SHA or vX.Y.Z)
    # so prod pulls a reproducible build instead of a moving :latest.
    image: ghcr.io/agent-field/af-stack-runtime:${AF_STACK_VERSION:-0.1.0}
    restart: unless-stopped
    environment:
      AF_STACK_HTTP_ADDR: ":8080"
      AF_STACK_METRICS_ADDR: ":9090"
      AF_STACK_DATABASE_URL: ${AF_STACK_DATABASE_URL}
      AF_STACK_AGENTFIELD_URL: http://agentfield:8080
      AF_STACK_KMS_KEY: ${AF_STACK_KMS_KEY}
      AF_STACK_MODULE_MULTI_TENANCY: ${AF_STACK_MODULE_MULTI_TENANCY:-true}
      OPENROUTER_API_KEY: ${OPENROUTER_API_KEY:-}
      ANTHROPIC_API_KEY: ${ANTHROPIC_API_KEY:-}
      OPENAI_API_KEY: ${OPENAI_API_KEY:-}
      # External S3 (Tigris/R2/AWS) — NOT MinIO.
      AF_STACK_S3_ADAPTER: s3
      AF_STACK_S3_ENDPOINT: ${AF_STACK_S3_ENDPOINT}
      AF_STACK_S3_BUCKET: ${AF_STACK_S3_BUCKET}
      AF_STACK_S3_ACCESS_KEY: ${AF_STACK_S3_ACCESS_KEY}
      AF_STACK_S3_SECRET_KEY: ${AF_STACK_S3_SECRET_KEY}
      AF_STACK_S3_REGION: ${AF_STACK_S3_REGION:-us-east-1}
      # Sandbox: e2b by default (hosted, strong isolation, no docker.sock).
      # Trade-off vs docker adapter:
      #   - e2b:    no privileged mount, no host kernel exposure, costs per
      #             sandbox-second, requires E2B_API_KEY.
      #   - gvisor: stays on-box, user-space kernel, free, requires gvisor
      #             installed on host (Phase 9.2).
      #   - docker: easiest in dev, NEVER in prod — full host kernel exposed
      #             via /var/run/docker.sock.
      AF_STACK_SANDBOX_ADAPTER: ${AF_STACK_SANDBOX_ADAPTER:-e2b}
      E2B_API_KEY: ${E2B_API_KEY:-}
      AF_STACK_E2B_BASE_URL: ${AF_STACK_E2B_BASE_URL:-https://api.e2b.dev}
      LOG_FORMAT: json
      LOG_LEVEL: ${LOG_LEVEL:-info}
      OTEL_EXPORTER_OTLP_ENDPOINT: ${OTEL_EXPORTER_OTLP_ENDPOINT:-}
      OTEL_SERVICE_NAME: af-stack-runtime
    networks:
      - frontend
      - backend
    depends_on:
      - agentfield
    deploy:
      replicas: 2
      resources:
        limits:
          cpus: "2.0"
          memory: "1G"
    # NOTE: NO `user: "0:0"` override here. The distroless image's nonroot
    # user (uid 65532) is used. NO docker.sock mount.

  dashboard:
    image: ghcr.io/agent-field/af-stack-dashboard:${AF_STACK_VERSION:-0.1.0}
    restart: unless-stopped
    environment:
      DATABASE_URL: ${AF_STACK_DATABASE_URL}
      AF_STACK_AUTH_SECRET: ${AF_STACK_AUTH_SECRET}
      BETTER_AUTH_URL: https://${AF_STACK_DOMAIN}
      BETTER_AUTH_TRUSTED_ORIGINS: ${BETTER_AUTH_TRUSTED_ORIGINS:-}
      RUNTIME_URL: http://runtime:8080
      NEXT_PUBLIC_RUNTIME_URL: https://${AF_STACK_DOMAIN}
      GOOGLE_CLIENT_ID: ${GOOGLE_CLIENT_ID:-}
      GOOGLE_CLIENT_SECRET: ${GOOGLE_CLIENT_SECRET:-}
      NODE_ENV: production
      NEXT_TELEMETRY_DISABLED: "1"
    networks:
      - frontend
    depends_on:
      - runtime
    deploy:
      replicas: 1
      resources:
        limits:
          cpus: "1.0"
          memory: "512M"

networks:
  # Caddy <-> runtime/dashboard. Public-facing edge.
  frontend:
    driver: bridge
  # runtime <-> agentfield (+ external Postgres / S3 over the host network).
  # Dashboard is NOT on this network.
  backend:
    driver: bridge

volumes:
  caddy-data:
  caddy-config: