How should an asyncio task clean up when it is cancelled?

Cancellation is delivered cooperatively at an await point. Swallowing it can break timeouts and structured-concurrency shutdown. Put resource cleanup in finally; if you catch asyncio.CancelledError to record a cancellation, re-raise it after cleanup unless the design explicitly requires suppression. Test cancellation while the task is waiting, while it owns a resource and immediately after successful completion. Cleanup that itself awaits may need a carefully bounded design. A cancellation request is not proof that remote work was rolled back. For network writes, retain an operation identifier and reconcile the remote result instead of blindly retrying.