Sometimes Gemini via LiteLLM will return a response with no `choice`s

[handrew]

openai-agents-python/src/agents/extensions/models/litellm_model.py

Line 107 in e382ec0

assert isinstance(response.choices[0], litellm.types.utils.Choices)

The above is the offending line. When Gemini via LiteLLM sometimes returns a response with no choices, it will throw an IndexError. I suspect that sometimes it'll just call a tool and call it a day. I'm not sure if we want to deal with this somehow. It is breaking my things :(

[seratch]

Thanks for reporting this. If the choices array is empty, the method is unable to return anything meaningful. So, just removing the assertion may not be a right solution.

[handrew]

Let me know if I can help - would love to figure out a way to solve this

[seratch]

Can you enable debug-level logging and share the outputs without sensitive/confidential information? You can add these lines of code before running your agent code:

import logging
logging.basicConfig(level=logging.DEBUG)

If their data structure provides something useful for proceeding the agentic workflow, we may adjust our logic for it.

[handrew]

I just set a breakpoint right before the assertion.

python ModelResponse(id='removed', created=1755967895, model='gemini-2.5-flash', object='chat.completion', system_fingerprint=None, choices=[], usage=Usage(completion_tokens=0, prompt_tokens=354, total_tokens=354, completion_tokens_details=None, prompt_tokens_details=PromptTokensDetailsWrapper(audio_tokens=None, cached_tokens=None, text_tokens=354, image_tokens=None)), vertex_ai_grounding_metadata=[], vertex_ai_url_context_metadata=[], vertex_ai_safety_results=[], vertex_ai_citation_metadata=[])

The logging is a bit less interpretable to me but I think the below is what you're looking for.

python model_info: {'key': 'gemini/gemini-2.5-flash', 'max_tokens': 65535, 'max_input_tokens': 1048576, 'max_output_tokens': 65535, 'input_cost_per_token': 3e-07, 'cache_creation_input_token_cost': None, 'cache_read_input_token_cost': 7.5e-08, 'input_cost_per_character': None, 'input_cost_per_token_above_128k_tokens': None, 'input_cost_per_token_above_200k_tokens': None, 'input_cost_per_query': None, 'input_cost_per_second': None, 'input_cost_per_audio_token': 1e-06, 'input_cost_per_token_batches': None, 'output_cost_per_token_batches': None, 'output_cost_per_token': 2.5e-06, 'output_cost_per_audio_token': None, 'output_cost_per_character': None, 'output_cost_per_reasoning_token': 2.5e-06, 'output_cost_per_token_above_128k_tokens': None, 'output_cost_per_character_above_128k_tokens': None, 'output_cost_per_token_above_200k_tokens': None, 'output_cost_per_second': None, 'output_cost_per_image': None, 'output_vector_size': None, 'citation_cost_per_token': None, 'litellm_provider': 'gemini', 'mode': 'chat', 'supports_system_messages': True, 'supports_response_schema': True, 'supports_vision': True, 'supports_function_calling': True, 'supports_tool_choice': True, 'supports_assistant_prefill': None, 'supports_prompt_caching': True, 'supports_audio_input': None, 'supports_audio_output': False, 'supports_pdf_input': True, 'supports_embedding_image_input': None, 'supports_native_streaming': None, 'supports_web_search': True, 'supports_url_context': True, 'supports_reasoning': True, 'supports_computer_use': None, 'search_context_cost_per_query': None, 'tpm': 8000000, 'rpm': 100000}

[seratch]

Thanks for taking the time on this. While I have other priorities next week, I will do my best to check this one. But if their output data structure does not fit OpenAI Agents SDK's requirements, we may not be able to provide full support for some use cases.

[handrew]

Thanks @seratch. I just switched off of LiteLLM in favor of just providing an api_base and api_key to the AsyncOpenAI client thinking it might fix the problem, and unfortunately it is still an issue with Gemini. So while LiteLLM might be a bit outside of your scope, I hope that supporting other popular models like Gemini might be within scope for this package. Thanks so much for your work.

[handrew]

Actually you know what I can actually control the outputs myself a bit better if I go the client path and just check if the response.output from Runner.run_sync is empty. 🎉

[github-actions]

This issue is stale because it has been open for 7 days with no activity.

[github-actions]

This issue was closed because it has been inactive for 3 days since being marked as stale.