Fix vision model sending max_tokens and max_completion_tokens together

[wgnrai]

Problem

Setting max_completion_tokens in a vision preset's kwargs breaks every vision_load call with:

litellm.BadRequestError: OpenAIException - Setting 'max_tokens' and 'max_completion_tokens' at the same time is not supported.

The cause is in build_vision_model() (plugins/_model_config/helpers/model_config.py). It always injects max_tokens, either from the preset or the default of 2000, into the model kwargs. When the preset also sets max_completion_tokens under kwargs, both parameters end up in the same request and the provider rejects it.

The Responses API path already handles this collision (helpers/litellm_transport.py pops both and maps them to max_output_tokens), but the vision side-car builds its call through unified_call and never goes through that normalization.

Fix

Skip the max_tokens injection in build_vision_model() when max_completion_tokens is already present in the merged kwargs. Preset intent wins, which matches the precedence used on the Responses path.

Testing

Added plugins/_model_config/tests/test_build_vision_model_max_tokens.py with four cases:

1. no max_completion_tokens anywhere: max_tokens still injected (default behavior unchanged)
2. max_completion_tokens in kwargs: max_tokens not injected (the bug)
3. explicit slot-level max_tokens=4096 with no collision: injected as 4096
4. slot-level max_tokens and kwargs max_completion_tokens both set: only max_completion_tokens survives

All four pass with python -m pytest plugins/_model_config/tests/test_build_vision_model_max_tokens.py -v.

Also confirmed on a live instance: with the patch, vision_load succeeds again where it previously returned the BadRequestError above.

[chatgpt-codex-connector]

You have reached your Codex usage limits for code reviews. You can see your limits in the Codex usage dashboard.

[wgnrai]

This one needed a rebase after some time on the back burner; it applied cleanly. The underlying issue is unchanged: main still sends both max_tokens and max_completion_tokens for vision models.