molecule-core/docker-compose.yml

# Include infra services (Temporal, Langfuse) so `docker compose up` starts the full stack.
include:
  - docker-compose.infra.yml

services:
  # --- Observability ---
  # digest-pinned 2026-05-10 (sha256:e7aafd3ccf721821b40f8b2251220b4bb8af5e4877b5c5a8846af5b3318aaf1d, linux/amd64)
  langfuse:
    image: langfuse/langfuse@sha256:e7aafd3ccf721821b40f8b2251220b4bb8af5e4877b5c5a8846af5b3318aaf1d
    depends_on:
      langfuse-clickhouse:
        condition: service_healthy
      langfuse-db-init:
        condition: service_completed_successfully
    environment:
      DATABASE_URL: postgres://${POSTGRES_USER:-dev}:${POSTGRES_PASSWORD:-dev}@postgres:5432/langfuse
      # Langfuse v2 expects the HTTP interface (port 8123). The previous
      # clickhouse://...:9000 native-protocol URL is rejected with
      # "ClickHouse URL protocol must be either http or https".
      CLICKHOUSE_URL: http://langfuse-clickhouse:8123
      CLICKHOUSE_MIGRATION_URL: clickhouse://langfuse-clickhouse:9000
      CLICKHOUSE_USER: langfuse
      CLICKHOUSE_PASSWORD: langfuse
      NEXTAUTH_SECRET: ${LANGFUSE_SECRET:-changeme-langfuse-secret}
      NEXTAUTH_URL: http://localhost:3001
      SALT: ${LANGFUSE_SALT:-changeme-langfuse-salt}
    ports:
      - "3001:3000"
    networks:
      - molecule-core-net
    healthcheck:
      test: ["CMD-SHELL", "wget --no-verbose --tries=1 --spider http://localhost:3000/api/public/health || exit 1"]
      interval: 10s
      timeout: 5s
      retries: 10

  # --- Platform ---
  platform:
    build:
      # Build context MUST be repo root, not ./platform — the Dockerfile
      # COPYs `workspace-server/migrations`, `workspace-server/go.mod`,
      # `workspace-configs-templates/` etc. via repo-relative paths so it
      # can bake in templates + migrations alongside the platform binary.
      # When context was ./platform earlier, docker silently cached an
      # earlier image (the COPY workspace-server/migrations resolved to nothing
      # under ./workspace-server/, so layers stopped invalidating) — manifested
      # as migration 023 not landing after PR #417 merged. CI workflow
      # already uses context=. , this aligns local with CI.
      context: .
      dockerfile: workspace-server/Dockerfile
    depends_on:
      postgres:
        condition: service_healthy
      redis:
        condition: service_healthy
    environment:
      DATABASE_URL: postgres://${POSTGRES_USER:-dev}:${POSTGRES_PASSWORD:-dev}@postgres:5432/${POSTGRES_DB:-molecule}?sslmode=disable
      REDIS_URL: redis://redis:6379
      PORT: "${PLATFORM_PORT:-8080}"
      PLATFORM_URL: "http://platform:${PLATFORM_PORT:-8080}"
      # Container network namespace is already isolated; "all interfaces"
      # inside the container = the bridge interface only. The fail-open
      # default (127.0.0.1) would block host-to-container access.
      BIND_ADDR: "${BIND_ADDR:-0.0.0.0}"
      # Default MOLECULE_ENV=development so the WorkspaceAuth / AdminAuth
      # middleware fail-open path activates when ADMIN_TOKEN is unset —
      # otherwise the canvas (which runs without a bearer in pure local
      # dev) gets 401 "missing workspace auth token" on every request.
      # Override to "production" for SaaS/staged deploys; in those modes
      # ADMIN_TOKEN must also be set or every request rejects.
      MOLECULE_ENV: "${MOLECULE_ENV:-development}"
      CORS_ORIGINS: ${CORS_ORIGINS:-http://localhost:${CANVAS_PUBLISH_PORT:-3000},http://127.0.0.1:${CANVAS_PUBLISH_PORT:-3000},http://localhost:3001}
      RATE_LIMIT: "${RATE_LIMIT:-1000}"
      CONFIGS_DIR: /configs
      CONFIGS_HOST_DIR: "${CONFIGS_HOST_DIR:-${PWD}/workspace-configs-templates}"
      PLUGINS_HOST_DIR: "${PLUGINS_HOST_DIR:-${PWD}/plugins}"
      # github-app-auth plugin — injects GITHUB_TOKEN / GH_TOKEN into every
      # workspace env from the App installation token. Remap the host-side
      # path in GITHUB_APP_PRIVATE_KEY_FILE to /secrets/github-app.pem inside
      # the container (the private key is bind-mounted below read-only).
      # Soft-dep: skipped entirely when GITHUB_APP_ID is unset.
      GITHUB_APP_ID: "${GITHUB_APP_ID:-}"
      GITHUB_APP_INSTALLATION_ID: "${GITHUB_APP_INSTALLATION_ID:-}"
      GITHUB_APP_PRIVATE_KEY_FILE: "/secrets/github-app.pem"
      # ADMIN_TOKEN — required to fully close issue #684 (AdminAuth bearer bypass, PR #729).
      # When set, only this exact value is accepted on all /admin/* and /approvals/* routes;
      # workspace bearer tokens are no longer accepted as admin credentials.
      # Unset (default) → backward-compat fallback: any valid workspace token passes AdminAuth
      # (same behaviour as before PR #729, still vulnerable to #684).
      # Generate: openssl rand -base64 32
      # Store in fly secrets / deployment env — NEVER commit the actual value.
      ADMIN_TOKEN: "${ADMIN_TOKEN:-}"
      # Workspace hibernation default (issue #724 / PR #724). Sets platform-wide idle
      # threshold (minutes); per-workspace column takes precedence. Leave empty to
      # rely on per-workspace config only (current behaviour — global-default code pending).
      HIBERNATION_IDLE_MINUTES: "${HIBERNATION_IDLE_MINUTES:-}"
      # Plugin supply chain hardening (issue #768 / PR #775). Never set in production.
      PLUGIN_ALLOW_UNPINNED: "${PLUGIN_ALLOW_UNPINNED:-}"
      # Force ImagePull/ContainerCreate to request linux/amd64 manifests
      # for the workspace-template-* images. The templates ship single-arch
      # amd64 today; without this override, an arm64 host (Apple Silicon)
      # asks the daemon for linux/arm64/v8, which doesn't match the manifest
      # and the pull fails with "no matching manifest". Apple Silicon will
      # run the amd64 image under Rosetta — slower (~2-3×) but functional.
      # Override to "" or another platform when the templates start shipping
      # multi-arch (then this hardcoded amd64 becomes unnecessary).
      MOLECULE_IMAGE_PLATFORM: "${MOLECULE_IMAGE_PLATFORM:-linux/amd64}"
      # GHCR auth for the workspace-images refresh endpoint
      # (POST /admin/workspace-images/refresh). When set, the platform's
      # Docker SDK ImagePull on private workspace-template-* images
      # succeeds without per-host `docker login`. GHCR_USER is the GitHub
      # username; GHCR_TOKEN is a fine-grained PAT with `read:packages`
      # on the Molecule-AI org. Both unset → endpoint can only pull
      # public images (current state for all 8 templates).
      GHCR_USER: "${GHCR_USER:-}"
      GHCR_TOKEN: "${GHCR_TOKEN:-}"
      # Auto-refresh workspace-template-* images. The watcher polls GHCR
      # every 5 min; when a digest moves, it pulls and force-recreates any
      # matching ws-* containers (existing /admin/workspace-images/refresh
      # logic). Closes the runtime CD chain: merge → containers running
      # new code, no operator step. Default ON for local dev because that's
      # where the runtime → ws iteration loop is tightest. Set to "false"
      # if you don't want the platform to mutate ws-* containers behind
      # your back during a long-running test.
      IMAGE_AUTO_REFRESH: "${IMAGE_AUTO_REFRESH:-true}"
    volumes:
      - ./workspace-configs-templates:/configs
      - ./org-templates:/org-templates:ro
      - ./plugins:/plugins:ro
      - /var/run/docker.sock:/var/run/docker.sock
      # App private key — read-only bind-mount. The host-side path is
      # gitignored per .gitignore rules (/.secrets/ + *.pem).
      - ./.secrets/github-app.pem:/secrets/github-app.pem:ro
      # Per-role persona credentials (molecule-core#242 local surface).
      # Sourced at workspace creation time by org_import.go::loadPersonaEnvFile
      # when a workspace.yaml carries `role: <name>`. The host-side dir is
      # populated by the operator-host bootstrap kit (28 dev-tree personas);
      # /etc/molecule-bootstrap/personas is the in-container path the
      # platform expects (matches the prod tenant-EC2 path so the same code
      # works in both modes).
      #
      # Read-only mount — workspace-server only reads, never writes here.
      # If the host dir is empty/missing the platform's loadPersonaEnvFile
      # silently no-ops per its existing semantics, so this mount is safe
      # even on a fresh machine that hasn't run the bootstrap kit yet.
      - ${MOLECULE_PERSONA_ROOT_HOST:-${HOME}/.molecule-ai/personas}:/etc/molecule-bootstrap/personas:ro
    ports:
      - "${PLATFORM_PUBLISH_PORT:-8080}:${PLATFORM_PORT:-8080}"
    networks:
      - molecule-core-net
    restart: unless-stopped
    healthcheck:
      # Plain GET — `--spider` would issue HEAD, which returns 404 because
      # /health is registered as GET only.
      test: ["CMD-SHELL", "wget -qO /dev/null --tries=1 http://localhost:${PLATFORM_PORT:-8080}/health || exit 1"]
      interval: 5s
      timeout: 5s
      retries: 10

  # --- Canvas ---
  canvas:
    # The publish-canvas-image CI workflow pushes a fresh image to GHCR on
    # every canvas/** merge to main. To update the running container:
    #   docker compose pull canvas && docker compose up -d canvas
    # First-time local setup or testing unreleased changes — build from source:
    #   docker compose build canvas && docker compose up -d canvas
    # Note: ECR images require AWS auth — `aws ecr get-login-password --region us-east-2 | docker login --username AWS --password-stdin 153263036946.dkr.ecr.us-east-2.amazonaws.com` before pull.
    # Digest-pin requires: aws ecr describe-images --repository-name molecule-ai/canvas --image-tags latest --query 'imageDetails[0].imageDigest'
    # TODO: pin canvas ECR image digest once AWS creds are available in CI.
    image: 153263036946.dkr.ecr.us-east-2.amazonaws.com/molecule-ai/canvas:latest
    build:
      context: ./canvas
      dockerfile: Dockerfile
      args:
        NEXT_PUBLIC_PLATFORM_URL: ${NEXT_PUBLIC_PLATFORM_URL:-http://localhost:${PLATFORM_PUBLISH_PORT:-8080}}
        NEXT_PUBLIC_WS_URL: ${NEXT_PUBLIC_WS_URL:-ws://localhost:${PLATFORM_PUBLISH_PORT:-8080}/ws}
        NEXT_PUBLIC_ADMIN_TOKEN: ${ADMIN_TOKEN:-}
    depends_on:
      platform:
        condition: service_healthy
    environment:
      PORT: "${CANVAS_PORT:-3000}"
      # Local dev — relaxes CSP to allow cross-port fetches (canvas:3000 → platform:8080).
      CSP_DEV_MODE: "${CSP_DEV_MODE:-1}"
      # NOTE: NEXT_PUBLIC_* are baked into the JS bundle at `next build` time —
      # these runtime values are ignored by the standalone output. They're kept
      # here for documentation / override during `docker compose build`.
      NEXT_PUBLIC_PLATFORM_URL: ${NEXT_PUBLIC_PLATFORM_URL:-http://localhost:${PLATFORM_PUBLISH_PORT:-8080}}
      NEXT_PUBLIC_WS_URL: ${NEXT_PUBLIC_WS_URL:-ws://localhost:${PLATFORM_PUBLISH_PORT:-8080}/ws}
    ports:
      - "${CANVAS_PUBLISH_PORT:-3000}:${CANVAS_PORT:-3000}"
    networks:
      - molecule-core-net
    healthcheck:
      test: ["CMD-SHELL", "wget -qO /dev/null --tries=1 http://127.0.0.1:${CANVAS_PORT:-3000} || exit 1"]
      interval: 10s
      timeout: 5s
      retries: 10

  # --- Optional: LiteLLM Proxy (unified OpenAI-compatible API for all providers) ---
  # Start with: docker compose --profile multi-provider up
  #
  # Workspace agents then set:
  #   OPENAI_BASE_URL=http://litellm:4000
  #   OPENAI_API_KEY=${LITELLM_MASTER_KEY:-sk-molecule}
  #
  # And use model names from infra/litellm_config.yml (e.g. "claude-opus-4-5",
  # "gpt-4o", "openrouter/deepseek-r1", "ollama/llama3.2").
  # Edit infra/litellm_config.yml to add/remove providers and models.
  # digest-pinned 2026-05-10 (sha256:7c311546c25e7bb6e8cafede9fcd3d0d622ac636b5c9418befaa32e85dfb0186)
  # Refresh: curl -sI https://ghcr.io/v2/berriai/litellm/manifests/main-latest (Docker-Content-Digest header)
  litellm:
    image: ghcr.io/berriai/litellm/main-latest@sha256:7c311546c25e7bb6e8cafede9fcd3d0d622ac636b5c9418befaa32e85dfb0186
    profiles:
      - multi-provider
    ports:
      - "4000:4000"
    volumes:
      - ./infra/litellm_config.yml:/app/config.yaml:ro
    command: ["--config", "/app/config.yaml", "--port", "4000", "--num_workers", "4"]
    environment:
      # Pass provider API keys through — only the ones you have are needed
      ANTHROPIC_API_KEY: ${ANTHROPIC_API_KEY:-}
      OPENAI_API_KEY: ${OPENAI_API_KEY:-}
      OPENROUTER_API_KEY: ${OPENROUTER_API_KEY:-}
      LITELLM_MASTER_KEY: ${LITELLM_MASTER_KEY:-sk-molecule}
    networks:
      - molecule-core-net
    restart: unless-stopped
    healthcheck:
      test: ["CMD-SHELL", "wget --no-verbose --tries=1 --spider http://localhost:4000/health || exit 1"]
      interval: 10s
      timeout: 5s
      retries: 5
      start_period: 15s

  # --- Optional: Local LLM Models via Ollama ---
  # Start with: docker compose --profile local-models up
  # After first start, pull a model:
  #   docker compose exec ollama ollama pull llama3.2
  #   docker compose exec ollama ollama pull qwen2.5-coder:7b
  # Then set MODEL_PROVIDER=ollama:llama3.2 in your workspace config.yaml
  # Workspace agents reach Ollama at http://ollama:11434 (internal Docker network).
  # digest-pinned 2026-05-10 (sha256:90bd8ed1ad1853fbfb1ef5835f9d7a24fe890e05ace521e2d8d7a6f56bb667dd, linux/amd64)
  # Refresh: curl -s https://hub.docker.com/v2/repositories/ollama/ollama/tags/latest | python3 -c "import json,sys; ..."
  ollama:
    image: ollama/ollama@sha256:90bd8ed1ad1853fbfb1ef5835f9d7a24fe890e05ace521e2d8d7a6f56bb667dd
    profiles:
      - local-models
    ports:
      - "11434:11434"
    volumes:
      - ollamadata:/root/.ollama
    networks:
      - molecule-core-net
    restart: unless-stopped
    healthcheck:
      test: ["CMD-SHELL", "ollama list || exit 1"]
      interval: 10s
      timeout: 5s
      retries: 5
      start_period: 20s

networks:
  molecule-core-net:
    name: molecule-core-net

volumes:
  pgdata:
  redisdata:
  clickhousedata:
  ollamadata: