docker-compose.agent-loop.yml

# Local agent-loop overlay for docker-compose.yml. Opt in from the checkout's
# .env with COMPOSE_FILE=docker-compose.yml:docker-compose.agent-loop.yml and
# COMPOSE_PROFILES=backend,agent-loop. See docs/engineering/local-agent-loop.md.
#
# The Rust checker service starts checker images with
# `docker run --mount type=bind,src=<material path>`. The Docker daemon resolves
# that source inside the Docker VM (or Linux host), so the material root is the
# named volume's own daemon path, mounted at that same path in API, worker, beat
# and checker. This assumes the default /var/lib/docker data root.
#
# The socket lives on a second named volume. Docker Desktop host-shared paths
# refuse chmod on Unix sockets, and AF_UNIX paths must stay within 107 bytes.
# An optional singleton model-proxy worker uses its own private named volume.
# Invocation relay sockets use a fourth volume mounted in the checker at the
# volume's daemon-visible path so Docker can bind-mount an individual socket
# into an otherwise networkless checker image.

x-agent-loop-backend: &agent-loop-backend
  environment:
    WORKSTREAM_ARTIFACT_SCRATCH_ROOT: /var/lib/docker/volumes/${COMPOSE_PROJECT_NAME}-checker-material/_data
    WORKSTREAM_EXTERNAL_CHECKER_MATERIAL_ROOT: /var/lib/docker/volumes/${COMPOSE_PROJECT_NAME}-checker-material/_data
    WORKSTREAM_EXTERNAL_CHECKER_SERVICE_SOCKET: /run/ws-checker/checker.sock
  volumes:
    - checker_material:/var/lib/docker/volumes/${COMPOSE_PROJECT_NAME}-checker-material/_data
    - checker_socket:/run/ws-checker:ro
  depends_on:
    checker-volumes-init:
      condition: service_completed_successfully

services:
  backend:
    <<: *agent-loop-backend
  worker:
    <<: *agent-loop-backend
    environment:
      WORKSTREAM_ARTIFACT_SCRATCH_ROOT: /var/lib/docker/volumes/${COMPOSE_PROJECT_NAME}-checker-material/_data
      WORKSTREAM_EXTERNAL_CHECKER_MATERIAL_ROOT: /var/lib/docker/volumes/${COMPOSE_PROJECT_NAME}-checker-material/_data
      WORKSTREAM_EXTERNAL_CHECKER_SERVICE_SOCKET: /run/ws-checker/checker.sock
      WORKSTREAM_CHECKER_MODEL_PROXY_SOCKET: /run/ws-model-proxy/model-proxy.sock
    volumes:
      - checker_material:/var/lib/docker/volumes/${COMPOSE_PROJECT_NAME}-checker-material/_data
      - checker_socket:/run/ws-checker:ro
      - checker_model_proxy:/run/ws-model-proxy:ro
  beat: *agent-loop-backend

  model-proxy:
    environment:
      WORKSTREAM_CHECKER_MODEL_PROXY_SOCKET: /run/ws-model-proxy/model-proxy.sock
      WORKSTREAM_ARTIFACT_SCRATCH_ROOT: /var/lib/docker/volumes/${COMPOSE_PROJECT_NAME}-checker-material/_data
    volumes:
      - checker_model_proxy:/run/ws-model-proxy
      - checker_material:/var/lib/docker/volumes/${COMPOSE_PROJECT_NAME}-checker-material/_data
    depends_on:
      checker-volumes-init:
        condition: service_completed_successfully
    healthcheck:
      test: [CMD, python, -c, "import socket; s=socket.socket(socket.AF_UNIX); s.settimeout(2); s.connect('/run/ws-model-proxy/model-proxy.sock'); s.close()"]
      interval: 5s
      timeout: 3s
      retries: 12

  # Named-volume roots start root-owned 0755. The scratch manager and checker
  # service both require a private 0700 root owned by the runtime UID.
  checker-volumes-init:
    profiles: ["backend"]
    image: ${COMPOSE_PROJECT_NAME}-backend:local
    pull_policy: never
    user: "0:0"
    network_mode: none
    environment:
      MATERIAL_ROOT: /var/lib/docker/volumes/${COMPOSE_PROJECT_NAME}-checker-material/_data
      RUNTIME_OWNER: "${LOCAL_UID:-1000}:${LOCAL_GID:-1000}"
    command:
      - /bin/sh
      - -c
      - |
        set -eu
        for directory in \
          "$$MATERIAL_ROOT" \
          /run/ws-checker \
          /run/ws-model-proxy \
          /var/lib/docker/volumes/${COMPOSE_PROJECT_NAME}-checker-relay/_data; do
          chown "$$RUNTIME_OWNER" "$$directory"
          chmod 0700 "$$directory"
        done
    volumes:
      - checker_material:/var/lib/docker/volumes/${COMPOSE_PROJECT_NAME}-checker-material/_data
      - checker_socket:/run/ws-checker
      - checker_model_proxy:/run/ws-model-proxy
      - checker_model_relay:/var/lib/docker/volumes/${COMPOSE_PROJECT_NAME}-checker-relay/_data

  # Loopback-only registry so the daemon records a platform-manifest digest.
  checker-registry:
    profiles: ["agent-loop"]
    image: registry:2.8.3@sha256:a3d8aaa63ed8681a604f1dea0aa03f100d5895b6a58ace528858a7b332415373
    ports:
      - "127.0.0.1:${WORKSTREAM_CHECKER_REGISTRY_HOST_PORT:-15000}:5000"
    volumes:
      - checker_registry:/var/lib/registry

  checker:
    profiles: ["agent-loop"]
    build:
      context: .
      dockerfile: docker/checker/Dockerfile.release
    image: ${COMPOSE_PROJECT_NAME}-checker:local
    init: true
    user: "${LOCAL_UID:-1000}:${LOCAL_GID:-1000}"
    # Docker socket access is root-equivalent on the daemon. Docker Desktop
    # exposes the socket as root:root 0660 (GID 0); Linux uses the docker group.
    group_add: ["${WORKSTREAM_DOCKER_SOCKET_GID:-0}"]
    command: [/etc/workstream/checker-service.json]
    network_mode: none
    read_only: true
    cap_drop: [ALL]
    security_opt: [no-new-privileges:true]
    restart: on-failure:5
    environment:
      DOCKER_CONFIG: /tmp/docker-config
    tmpfs:
      - /tmp:mode=0700,uid=${LOCAL_UID:-1000},gid=${LOCAL_GID:-1000}
    volumes:
      - type: bind
        source: ${WORKSTREAM_CHECKER_CONFIG_FILE:?set WORKSTREAM_CHECKER_CONFIG_FILE in .env}
        target: /etc/workstream/checker-service.json
        read_only: true
        bind:
          create_host_path: false
      - type: volume
        source: checker_material
        target: /var/lib/docker/volumes/${COMPOSE_PROJECT_NAME}-checker-material/_data
        read_only: true
      - checker_socket:/run/ws-checker
      - checker_model_proxy:/run/ws-model-proxy:ro
      - checker_model_relay:/var/lib/docker/volumes/${COMPOSE_PROJECT_NAME}-checker-relay/_data
      - /var/run/docker.sock:/var/run/docker.sock
    depends_on:
      checker-volumes-init:
        condition: service_completed_successfully
    healthcheck:
      # Connect rather than `test -S`: a socket left on the volume by a killed
      # run must not report health before the new process binds it.
      test:
        - CMD
        - python3
        - -c
        - "import socket; s = socket.socket(socket.AF_UNIX); s.settimeout(2); s.connect('/run/ws-checker/checker.sock'); s.close()"
      interval: 5s
      timeout: 5s
      retries: 12
      start_period: 10s

volumes:
  checker_material:
    name: ${COMPOSE_PROJECT_NAME}-checker-material
  checker_socket:
  checker_model_proxy:
  checker_model_relay:
    name: ${COMPOSE_PROJECT_NAME}-checker-relay
  checker_registry: