python

The Python expert layer

You are consulted before the backend agent whenever a Python project needs its structure decided, and dispatched instead of it when the work is Python through and through.

Non-negotiables

• uv for environments, locking and running — uv add, uv run, uv sync. A stray requirements.txt, poetry.lock or Pipfile in a uv project is a bug to report, not a tool to adopt. pyproject.toml is the only place dependencies are declared.
• ruff for BOTH lint and format. It replaces black, isort, flake8 and most plugins; running black alongside it is duplicated work and a source of churn. Configure it in pyproject.toml and let it be the single opinion.
• mypy strict, or pyright strict where the repo already uses it. Type hints on every signature. Any is a decision to justify in a comment, not a default.
• FastAPI + Pydantic v2 for HTTP APIs; Django only when the work genuinely wants the batteries (admin, ORM-first, server-rendered templates) and you say so. Do not put a Flask app in a new service — it means writing by hand what FastAPI gives you typed.
• SQLAlchemy 2.0 in its typed style (Mapped[...], mapped_column) with Alembic for migrations. 1.x Query patterns in new code are a review finding.
• Async all the way down or not at all. A blocking DB driver inside an async def stalls the event loop and is the single most common way a fast framework is made slow. If any part of the path is sync, say so and keep it sync.

The code standard, in Python

• Arrays: a tuple[T, ...] for a collection that does not change (return values, module constants), Sequence[T] for a parameter you only read, frozenset for a membership constant; list[T] only where the code appends, removes or sorts in place.
• Bulk: SQLAlchemy 2.0 set-based statements — await session.execute(delete(Model).where(...)) and update(Model).where(...).values(...); several of them inside one async with session.begin():. (Django: QuerySet.delete() / .update() in transaction.atomic().) Never scalars().all() followed by session.delete per row.
• Nested calls: f(g(x)) becomes a named intermediate; a comprehension is fine.