[Bug]: openhands:ERROR: loop.py:22 - RuntimeError: Error type: KeyError (KeyError: 'command' in fn calling)

[scosenza]

Is there an existing issue for the same bug?

• I have checked the existing issues.

Describe the bug and reproduction steps

poetry run python -m openhands.core.cli
...
...
Review the changes and make sure they are as expected. Edit the file again if necessary.
23:46:29 - openhands:ERROR: agent_controller.py:212 - [Agent Controller default] Error while running the agent (session ID: default): 'command'. Traceback: Traceback (most recent call last):
  File "/tmp/OpenHands/openhands/controller/agent_controller.py", line 210, in _step_with_exception_handling
    await self._step()
  File "/tmp/OpenHands/openhands/controller/agent_controller.py", line 555, in _step
    action = self.agent.step(self.state)
             ^^^^^^^^^^^^^^^^^^^^^^^^^^^
  File "/tmp/OpenHands/openhands/agenthub/codeact_agent/codeact_agent.py", line 390, in step
    actions = codeact_function_calling.response_to_actions(response)
              ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
  File "/tmp/OpenHands/openhands/agenthub/codeact_agent/function_calling.py", line 511, in response_to_actions
    if arguments['command'] == 'view':
       ~~~~~~~~~^^^^^^^^^^^
KeyError: 'command'

23:46:29 - openhands:ERROR: loop.py:22 - RuntimeError: There was an unexpected error while running the agent. Please report this error to the developers. Your session ID is default. Error type: KeyError

OpenHands Installation

Docker command in README

OpenHands Version

Built from source while testing #5284
SHA: baf0adb
Branch: add-local-runtime

Operating System

Linux

Logs, Errors, Screenshots, and Additional Context

No response

[sudo-vaibhav]

I am also facing similar error

[xingyaoww]

Could y'all share what model you are using?

[scosenza]

I'm using: litellm_proxy/bedrock/us.anthropic.claude-3-5-sonnet-20241022-v2:0

[enyst]

The 'command' argument is required for the str_replace_editor, and it looks like the LLM didn't send it. If that was the case, this should be rare, I assume.

I'll try to reproduce to see it. I wonder though, did you by any chance have DEBUG enabled or log_completions? Then we could double check for sanity, that the LLM response really didn't have it.