What

PromptStreamConsumer launched its streaming work with bare asyncio.create_task(...) and threw away the returned task in three places (execute_template_async, execute_improve_prompt_async, execute_generate_prompt_async).

The event loop only keeps a weak reference to a task, so a fire-and-forget task can be garbage-collected before it finishes — silently dropping a user's prompt execution with no result and no error. Per the CPython docs: "Save a reference to the result of this function, to avoid a task disappearing mid-execution."

Closes #819.

How

• Added a self._background_tasks set and a small _spawn() helper that holds a strong reference to each task and removes it via add_done_callback once it completes (no leak).
• Routed the three create_task call sites through _spawn().
• Cancel any still-pending tasks in disconnect() so they don't outlive the connection.

Notes

• No behavior change for the happy path — same coroutines, same scheduling; this just prevents the GC race.
• Found while reviewing the async paths with a static analyzer (codehound). It flagged one more spot — model_hub/utils/kb_helpers.py uses a deprecated asyncio.get_event_loop() plus a fire-and-forget ensure_future — happy to send a separate small PR for that to keep this one focused.

Testing

• python -m py_compile clean; static analyzer reports the file clean after the change.