Local API: cancelled HTTP requests keep running and block subsequent transcriptions

[harmonicholas]

I’m using Murmure 1.10.1 through its local API:
POST http://127.0.0.1:4800/api/transcribe
I noticed that closing the HTTP connection while a long transcription is running does not cancel the inference on the Murmure side.. The API is very nice.
Would it be possible to add proper cancellation support to the local API? For example:

• return a task or transcription ID when a request starts;
• provide an endpoint such as DELETE /api/transcriptions/{id};
• propagate client disconnection or cancellation to the inference engine;
• optionally return a clear cancelled status.
  Thank you for Murmure and for providing a local, privacy-friendly transcription API.

[Kieirra]

Hi, thank you for creating the issue and for your kind words.

I will add cancellation support in the upcoming API rework. Instead of implementing Task IDs and a DELETE endpoint, I am going with standard context cancellation (similar to how Ollama handles it).

If your client drops the HTTP connection, Murmure will detect it and automatically stop the underlying inference. A dropped request will also stop holding up the queue, so your next transcription starts right away instead of waiting for the abandoned one. This keeps the API stateless, much simpler to use, and removes the need to manage task IDs on your end in most cases.

This will be included in the next release.

[Kieirra]

The 1.11.0 beta is out. The transcription is now cancelled as soon as the client disconnects, and the queue is freed once the abandoned work has actually stopped.

#419