Add to the Python service

apps/ai is FastAPI plus a BullMQ worker; only apps/api calls it, through the generated client in packages/ai-client. docs/python-services.md describes the service.

1. Models. Request and response models in apps/ai/app/schemas.py, every field constrained (Field(min_length=…, max_length=…), Literal[...]). They become apps/ai/openapi.json and the TypeScript types.
2. Logic. A module in apps/ai/app/ (like summaries.py or documents.py). Data access only through tenant(org_id) from app/db/session.py; models through Pydantic AI, so budgets and AI_MODEL apply.
3. Route. In apps/ai/app/main.py, with operation_id="<name>" (the generated client's function name) and CallerDep (the API's signed token). Errors are AppError("<CODE>", status, params) with a code from packages/contracts/src/errors.ts. Blocking work in a plain def; I/O in async def.
4. Settings. A field in apps/ai/app/settings.py with alias="<ENV_NAME>", and the variable in .env.example and docs/environment.md.
5. bun run gen: regenerates apps/ai/openapi.json and packages/ai-client/src/generated/.
6. Call it. Add a method to createAiClient in packages/ai-client/src/index.ts (wraps the generated function with unwrap), then use it from apps/api/src/modules/ai/ai.service.ts behind a contract procedure (add-feature skill). Keep the API's input limits equal to the Pydantic ones; add the pair to apps/api/src/modules/ai/limits.test.ts.
7. Tests. Pure tests in apps/ai/tests/test_units.py (run by bun run test); anything needing Postgres or Redis in apps/ai/tests/test_service.py (pytestmark = pytest.mark.integration, run by bun run test:integration). bun run lint runs ruff and bun run check-types runs basedpyright on it.

Background work for the service: the add-job skill (Python section).

Finish with the verify skill, then the python-reviewer agent, the security-reviewer agent for a new route (a new entry point), and the reviewer agent for the API side.