`reasoning_enabled` and `reasoning_budget` invalid kwargs in Prompt.build

[tylerganter]

Calling OpenAI's chat completions I get the following error when running this example straight from the docs.

Example Code

Non-streaming request with reasoning parameters

Error

Non-streaming Response:
Error during non-streaming request: Completions.create() got an unexpected keyword argument 'reasoning_enabled'. Did you mean 'reasoning_effort'?

Dependency Tree

braintrust-proxy v0.1.0
    ├── openai v1.91.0
    │   ├── anyio v4.9.0
    │   │   ├── idna v3.10
    │   │   └── sniffio v1.3.1
    │   ├── distro v1.9.0
    │   ├── httpx v0.28.1
    │   │   ├── anyio v4.9.0 (*)
    │   │   ├── certifi v2025.6.15
    │   │   ├── httpcore v1.0.9
    │   │   │   ├── certifi v2025.6.15
    │   │   │   └── h11 v0.16.0
    │   │   └── idna v3.10
    │   ├── jiter v0.10.0
    │   ├── pydantic v2.11.7
    │   │   ├── annotated-types v0.7.0
    │   │   ├── pydantic-core v2.33.2
    │   │   │   └── typing-extensions v4.14.0
    │   │   ├── typing-extensions v4.14.0
    │   │   └── typing-inspection v0.4.1
    │   │       └── typing-extensions v4.14.0
    │   ├── sniffio v1.3.1
    │   ├── tqdm v4.67.1
    │   └── typing-extensions v4.14.0
    └── pydantic v2.11.7 (*)

Context and Urgency

The call works if I replace reasoning_enabled and reasoning_budget with reasoning_effort but

1. lacks the granularity of control that Google offers for reasoning
2. this is particularly problematic when using prompts in code with the OpenAI client. load_prompt(...).build(...) returns a dict that includes reasoning_enabled key so prompts cannot be used with reasoning and the OpenAI SDK!

I would also note that if you modify and save the prompt multiple times it causes reasoning_enabled to appear, so this issue will not surface unless params are modified for the prompt.

1. create prompt, don't configure reasoning

{
    'temperature': 0,
    'model': 'gemini-2.5-flash',
    'span_info': {
        'metadata': {
            'prompt': {...}
        }
    },
    'messages': [{'content': 'You are a helpful assistant', 'role': 'system'}]
}

2. enable reasoning

{
    'temperature': 0,
    'reasoning_enabled': True,
    'model': 'gemini-2.5-flash',
    'span_info': {
        'metadata': {
            'prompt': {...}
        }
    },
    'messages': [{'content': 'You are a helpful assistant', 'role': 'system'}]
}

3. disable reasoning

{
    'temperature': 0,
    'reasoning_enabled': False, # <-- same functionality as (1) but this kwarg wasn't there before!
    'model': 'gemini-2.5-flash',
    'span_info': {
        'metadata': {
            'prompt': {...}
        }
    },
    'messages': [{'content': 'You are a helpful assistant', 'role': 'system'}]
}

[ibolmo]

@tylerganter did you happen to try the new braintrust_proxy pip library? It provides a few methods to help with the type errors: https://pypi.org/project/braintrust-proxy/

[ibolmo]

I'll try to repro as well. Perhaps OpenAI added new validation checks on the input arguments

[tylerganter]

@ibolmo yeah this is at a level higher than the braintrust_proxy library. the examples are wrapping individual messages but the issues is the kwargs going directly to client.chat.completions.create()

[ibolmo]

yep easy to repro. let me look into this. thanks for reporting!

[ibolmo]

@tylerganter could you try updating to use the extra_body argument?

In my testing this seems to work. I'll update the examples.

[tylerganter]

@ibolmo works like a charm! I would emphasize more than just the example that braintrust.Prompt.build() gets updated to reflect this. I could make an issue in that repo if you'd like

[ibolmo]

I'll review and patch that up as well

[ibolmo]

Closing the loop. The examples have been fixed, but I'll leave this open since we haven't fixed the Prompt.build(). I'll update the issue to reflect this.