python

Use asyncio.TaskGroup for related work so the scope owns completion and failure, and ensure every created task has an owner that retains it, observes exceptions, and handles cancellation. Avoid untracked fire-and-forget tasks and never swallow CancelledError.

Prefer uv for environments, dependencies, locking, and execution, Ruff for formatting and linting, and ty for type checking. Commit uv.lock for applications, and keep rule selection explicit with target-version matching the minimum supported Python.

Use asyncio for I/O-bound structured concurrency, threads for blocking I/O or GIL-releasing native code, and processes for isolation or CPU parallelism. Never treat the GIL as synchronization, and guard multiprocessing entrypoints with if __name__ == "__main__".

Yes. Free-threaded builds change context inheritance and warning-filter defaults, typically use more memory, and extensions may re-enable the GIL. Verify sys._is_gil_enabled() after loading production dependencies and test the full native dependency set before claiming support.

Pickle deserialization can execute arbitrary code, so untrusted or tamperable data must never be unpickled, including through shelve, multiprocessing, caches, or socket logging. Use safe serialization formats and validate external input at trust boundaries.

Only when a migration is explicitly requested. During scoped maintenance, preserve the existing package manager, formatter, linter, and type checker; a code fix does not authorize replacing the established toolchain.