[Bug]: Hanging process in SWE-Bench reproducing script

[kevin-support-bot]

All-Hands-AI#5952 Issue

---

@fishmingyu fake_user_response_fn is used here. Could you debug it's value?

[fishmingyu]

Sorry, the issue is about how to avoid the state: AgentState.AWAITING_USER_INPUT. Do you know how to control it when reproducing swebench? It is important to have a non-interaction process.

[SmartManoj]

Adding an instruction here may solve this. But that's not considered as reproducing.

---

If the LLM gives just a message, the agent will provide this default value in eval.

You could modify this line. Here only, it asks for the user input.

[SmartManoj]

Regression from All-Hands-AI#5868. Could you test on a2e9e20?

MRE

[fishmingyu]

Hi @SmartManoj , thank you so much for providing these hints. I actually tested on branch 2ec2f25, the newest one. Thus I think the problem is same for a2e9e20. I will revise your recommended line then.

[SmartManoj]

You may see the error.

/usr/lib/python3.12/concurrent/futures/thread.py:95: RuntimeWarning: coroutine 'run_controller.<locals>.on_event' was never awaited
  del work_item
RuntimeWarning: Enable tracemalloc to get the object allocation traceback

In the latest, the function is not called itself due to the regression.

MRE

tested on a2e9e20 and works perfectly.

---

[SmartManoj]

Here the async keyword is the problem. Now works on the HEAD f3885ca

[fishmingyu]

Thank you so much, it works now!