I'm convinced that such a change should go in.

However, what about adding this logic inside read_response itself?

Also, perhaps it makes sense to use Lwt.{on_termination,on_failure} rather than just Lwt.on_cancel?

The change could use Lwt.on_termination. It is already used here so that makes sense as an alternative.

Is it safe to call closefn multiple times? If so then it could reduce some code duplication to add Lwt.on_failure inside read_response.

I believe that Lwt.on_termination [1] only executes after the stream has been exhausted. You can cancel/raise without exhausting the stream which would prevent the termination callback from executing. I should verify this though.

I think closefn is idempotent, and if it's not we should consider making it like that.

[1] https://github.com/rgrinberg/lwt/blob/master/src/core/lwt_stream.mli#L227-L230

I've updated the change to wrap read_response instead.

Merged in #447

Thanks