[Bug] Closing a partially consumed streamify async stream raises BaseExceptionGroup[GeneratorExit]

What happened?

Closing a partially consumed async generator returned by dspy.streamify() raises a BaseExceptionGroup containing GeneratorExit, rather than closing normally.

This was reproduced after receiving an actual generated token from OpenAI gpt-4.1-mini, through native Chat, native Responses, and LiteLLM Chat. Fully consuming the streams succeeded. This affects explicit early-close cleanup, such as a consumer stopping generation; a resource leak has not been established.

Steps to reproduce

With OPENAI_API_KEY configured, run:

import asyncio
import dspy


async def main():
    lm = dspy.LM("openai/gpt-4.1-mini", cache=False, temperature=0, max_tokens=400)
    try:
        with dspy.context(lm=lm):
            stream = dspy.streamify(
                dspy.Predict("question -> answer"),
                stream_listeners=[
                    dspy.streaming.StreamListener(signature_field_name="answer")
                ],
            )(question="Write 100 numbered short sentences about paper airplanes.")
            async for item in stream:
                if isinstance(item, dspy.streaming.StreamResponse):
                    print("Received real token:", repr(item.chunk))
                    break
            await stream.aclose()
            print("Closed cleanly")
    finally:
        if hasattr(lm, "close"):
            lm.close()


asyncio.run(main())

Expected: aclose() returns normally and prints Closed cleanly.

Actual: process exits with status 1:

Received real token: '1'
BaseExceptionGroup: unhandled errors in a TaskGroup (1 sub-exception)
  ... dspy/streaming/streamify.py, in async_streamer
      yield output
  GeneratorExit

async_streamer yields inside an AnyIO task group. Closing the generator injects GeneratorExit, which reaches the task group's exceptional exit and is wrapped in an exception group.

DSPy version / environment

• Merged main at 453823209.
• Also reproduced with the complete 3.3.1 source tree in a separate worktree using the same installed dependencies. This is not a regression from the recently merged LocalInterpreter cancellation fix.
• Linux x86-64, Python 3.11.6; AnyIO 4.14.2, LiteLLM 1.96.0, OpenAI SDK 2.53.0.
• Real provider calls, caching disabled; no mocked LM responses.

AI assistance

Amp performed the reproductions and drafted this report during isaacbmiller's interactive release-robustness investigation. isaacbmiller explicitly authorized filing this issue after reviewing the finding. Relevant prompts: “Use a real LM, not just run pytests”, “what broke?”, and “I think its fine to create an issue for”.