ResponseReasoningItem' object has no attribute 'arguments'

[oegedijk]

• [x ] This is actually a bug report.
• I am not getting good LLM Results
• I have tried asking for help in the community on discord or discussions and have not received a response.
• I have tried searching the documentation and have not found an answer.

What Model are you using?

• gpt-3.5-turbo
• gpt-4-turbo
• gpt-4
• Other (please specify 03 with code interpreter)

Describe the bug
This is a bit of a stochastic bug, so find it hard to provide a lot of context, but just leaving it just in case.

I have a a structured extraction for which I am use o3 with code interpreter mode=Mode.RESPONSES_TOOLS_WITH_INBUILT_TOOLS

response, completion = await client.responses.create_with_completion(
                    model="03",
                    tools=[{"type": "code_interpreter", "container": {"type": "auto"}}],
                    instructions=system_prompt
                    + f"\n\nYou must call the {response_model.__name__} function to return your final extraction.",
                    input=inputs,
                    response_model=response_model,
                    reasoning={"effort": "high"},
                    max_retries=AsyncRetrying(
                        stop=stop_after_attempt(3),
                        after=lambda retry: logger.info(
                            f"Retrying code interpreter call after validation error: {retry}"
                        ),
                    ),
                )

And then every now and then I see this in the logs:

<RetryCallState 4455142688: attempt #1; slept for 0.0; last result: failed (AttributeError 'ResponseReasoningItem' object has no attribute 'arguments')>

I think calls to an .arguments attribute are mostly related at trying to extract the arguments of a function call?

So looking at this cookbook sometimes openai returns both a ResponseReasoningItem and the ResponseFunctionToolCall

https://cookbook.openai.com/examples/responses_api/reasoning_items

response.output

[ResponseReasoningItem(id='rs_68210c71a95c81919cc44afadb9d220400c77cc15fd2f785', summary=[], type='reasoning', status=None),
ResponseFunctionToolCall(arguments='{"latitude":48.8566,"longitude":2.3522}', call_id='call_9ylqPOZUyFEwhxvBwgpNDqPT', name='get_weather', type='function_call', id='fc_68210c78357c8191977197499d5de6ca00c77cc15fd2f785', status='completed')]

So seems sometimes we try to get the arguments from the reasoning instead of the function tool call?

Sorry cannot be more specific.