481 strip frequency penalty

[Omarhus01]

What does this PR do?

Fixes #481

Problem

When AdalFlow is used with OpenAI reasoning models (o1, o3, o3-mini, o4, gpt-5 family), every API call fails with:

Responses.create() got an unexpected keyword argument 'frequency_penalty'

This breaks the official quickstart Colab notebook and any user code that passes frequency_penalty (a common default) when using a reasoning model. OpenAI's Responses API rejects frequency_penalty for reasoning models, but AdalFlow forwards all kwargs unconditionally.

Fix

Adds a _strip_unsupported_reasoning_kwargs helper on OpenAIClient that:

• Detects reasoning models by model-name prefix (o1, o3, o4, gpt-5).
• Removes a configurable allowlist of unsupported kwargs (_UNSUPPORTED_REASONING_KWARGS).
• Returns a copy — never mutates the caller's dict.
• Logs stripped kwargs at debug level.

The helper is invoked at the top of the LLM/LLM_REASONING branch in both call and acall, so both sync and async paths are covered.

Scope

Conservative on purpose: only frequency_penalty is stripped right now, since that is the parameter confirmed broken in #481. The constant _UNSUPPORTED_REASONING_KWARGS is a tuple so additional params (presence_penalty, top_p, logit_bias, etc.) can be added in follow-up PRs as they are independently confirmed.

Tests

Added four unit tests in adalflow/tests/test_openai_client.py:

1. frequency_penalty is stripped for each reasoning prefix (o1, o3-mini, o4-turbo, gpt-5).
2. frequency_penalty is preserved for non-reasoning models (e.g. gpt-4o).
3. Unrelated kwargs (temperature, max_tokens) are left intact.
4. The helper does not mutate the input dict.

Before submitting

• Did you read the contributor guideline?
• Did you make sure your PR does only one thing, instead of bundling different changes together?
• Did you write any new necessary tests?
• Did you verify new and existing tests pass locally with your changes?

[Omarhus01]

For context on why I picked this fix: I was running through the AdalFlow quickstart while reading the LLM-AutoDiff paper, and the breakage in the Object Count training loop blocks exactly the demonstration the paper builds on — selective gradient computation on a teacher generator, which is o3-mini in this notebook.

Happy to extend this to other params (presence_penalty, top_p) in a follow-up if you'd like a broader sweep, but kept this PR scoped to the one parameter the traceback proved was rejected.