docker-compose.yml

services:
  postgres:
    # pgvector/pgvector ships the `vector` extension on top of postgres:16.
    # PR 2 (embed pipeline) introduces vector(1536) columns; the plain
    # postgres:16 image would fail `CREATE EXTENSION IF NOT EXISTS vector`
    # in schema.sql at startup. Cloud SQL needs the
    # `cloudsql.enable_pgvector` flag flipped at the instance level
    # (one-time `gcloud sql instances patch`).
    image: pgvector/pgvector:pg16
    # Durability off for LOCAL DEV ONLY (prod is Cloud SQL): the test
    # suite TRUNCATEs + reseeds per test, so it is fsync-bound on
    # Docker-for-Mac. Worst case on host crash is losing local dev
    # data, which is disposable. Roughly halves the pytest wall clock.
    command: postgres -c fsync=off -c synchronous_commit=off -c full_page_writes=off
    environment:
      POSTGRES_USER: agentdrive
      POSTGRES_PASSWORD: dev
      POSTGRES_DB: agentdrive
    ports: ["5432:5432"]
    volumes:
      # NO schema init-script mount. `apply_schema` is the SINGLE schema
      # authority — it must see an empty DB so it takes the fresh-baseline
      # path (apply schema.sql once, mark every migration embodied). A
      # docker-entrypoint init-script that pre-applied schema.sql made the
      # DB non-fresh, so apply_schema replayed historical migrations against
      # a baseline that had already folded their effects in — e.g. 0008
      # UPDATE-ing the `visibility` column that current schema.sql no longer
      # creates (UndefinedColumnError). Every consumer (scripts/dev.sh,
      # feedback-fix-verify-setup.sh) runs apply_schema after `compose up`.
      - pgdata:/var/lib/postgresql/data
    healthcheck:
      test: ["CMD-SHELL", "pg_isready -U agentdrive -d agentdrive"]
      interval: 3s
      timeout: 3s
      retries: 10

  gcs:
    image: fsouza/fake-gcs-server:latest
    command: -scheme http -port 4443 -public-host localhost:4443
    ports: ["4443:4443"]
    volumes:
      - gcsdata:/storage

volumes:
  pgdata:
  gcsdata: