Issue with `finish_reason` validation for GPT-OSS models (20B & 120B) causing errors in pipeline integration

[Gautam8387]

Encountering multiple issues when integrating the GPT-OSS models (20B and 120B) into the current pipeline. The error is specifically related to the finish_reason field in the response from the OpenAI chat completions endpoint.

Error:

ERROR: Agent 'CellDescriptionAgent' failed after 1 attempts. Last error: Invalid response from OpenAI chat completions endpoint: 1 validation error for ChatCompletion
choices.0.finish_reason
  Input should be 'stop', 'length', 'tool_calls', 'content_filter', or 'function_call' [type=literal_error, input_value=None, input_type=NoneType]
    For further information, visit https://errors.pydantic.dev/2.11/v/literal_error

Observations:

• This error only occurs when using the GPT-OSS models (20B and 120B).
• Other OpenAI models (GPT-5, GPT-5.1) are working as expected without issues.
• The error specifically pertains to the finish_reason being returned as None, which is rejected by the pydantic model validation.

Providers tried:

• Groq
• ollama
• OpenRouter

Possible Causes:

• finish_reason validation: It seems that the finish_reason value is either missing or set to None in the response from GPT-OSS models. OpenAI’s API expects the finish_reason to be one of the following values: 'stop', 'length', 'tool_calls', 'content_filter', or 'function_call'. The GPT-OSS models are returning None for finish_reason, which is causing the validation error.
• Provider behavior: The different providers return null (or None) for finish_reason, while other OpenAI models handle this properly.

Attempts to resolve:

• Multiple providers tested: The issue persists across Groq, Ollama, and OpenRouter
• Monkeypatching pydantic_ai: Attempted to modify pydantic_ai to accept None for finish_reason, but this resulted in empty responses instead of the expected output.

Potential Next Steps:

• Custom middleware: A custom HTTP client middleware that intercepts the OpenRouter response and injects finish_reason where missing.
• Priority: This issue is not high priority at the moment due to the time spent troubleshooting so far, but it will be revisited.

Any insights or suggestions would be appreciated.