Thanks for contributing! I normally keep quiet about objc heavy macos issues/prs but since i spent the whole day on the startTask code yesterday i do have some opinions this time.

I personally would like to reduce our reliance on ivars where possible, even if that means having to use global variables (via once_cell::Lazy for example). To make this easier, i saw that others stored just the hash of the object, for example like here. I did not hear anything bad about this approach in this context so i assume it's safe enough.

Also, i wonder if we have to check this multiple times and not only once before it actually starts sending it. For that to make sense we'd probably have to chunk the didReceive call so maybe that's overkill 🤔I don't know how fast the didReceive call is, and how it behaves if stopTask is triggered just before we call didReceive (but after the check you added) 🤔

I made changes to use the tasks hash value in the store as you suggested, also the lock ist now released only after the call to task->didFinish.
still works fine in my tests.
Though I wonder the swift code you linked doesn't use any lock/sync for the dictionary at all..

I don't know what I'm doing wrong but as soon as I run tauri with this change it crashes.

I managed to solve the sync/lock problem by sending the task response on the main thread (via NSObject.performSelectorOnMainThread).

Won't this cause issues with large payloads?

It should not as the response data payload is already provided as byte array by the custom thread.
Calls to the task like didReceiveResponse and didReceiveData seem to run on the main thread anyways - thus my issues with deadlocks when I locked the main thread in the stop_task call.
I did some tests with large data (3 parallel requests with about 3.5GB response payload) and the performance seems not to be affected.
For streaming the Responder API would probably need to provide some way to send multiple smaller chunks of data - and then call didReceiveData for each chunk.