Bug Description

When calling litellm.acompletion() with tool_choice and a model using the openai/responses/ prefix (e.g., openai/responses/gpt-5.4), the tool_choice parameter is silently dropped and never reaches the underlying API call.

Steps to Reproduce

import litellm

response = await litellm.acompletion(
    model="openai/responses/gpt-5.4",
    messages=[{"role": "user", "content": "Search the web for today's news"}],
    tools=[{
        "type": "function",
        "function": {
            "name": "web_search",
            "description": "Search the web",
            "parameters": {
                "type": "object",
                "properties": {"query": {"type": "string"}},
                "required": ["query"],
            },
        },
    }],
    tool_choice={"type": "function", "function": {"name": "web_search"}},
    stream=True,
)

Expected: The model is forced to call web_search.
Actual: The model ignores tool_choice and may call other tools or generate text directly.

Root Cause

tool_choice is missing from the supported params list for the openai/responses/ chat completion bridge path.

In get_optional_params(), get_supported_openai_params(model="openai/responses/gpt-5.4", custom_llm_provider="openai") returns a list that includes tools but not tool_choice. This causes _check_valid_arg to either raise UnsupportedParamsError or silently drop the param (when drop_params=True).

The Responses API itself does support tool_choice — it's defined in ResponsesAPIOptionalRequestParams and handled correctly by LiteLLMCompletionResponsesConfig._transform_tool_choice(). The issue is that the chat completions → responses bridge validation rejects it before it can reach that code.

Verified on litellm 1.82.0:

from litellm.utils import get_supported_openai_params

params = get_supported_openai_params(
    model="openai/responses/gpt-5.4", 
    custom_llm_provider="openai"
)
print("tool_choice" in params)  # False
print("tools" in params)        # True

Suggested Fix

Add tool_choice to the supported params list returned for models using the openai/responses/ bridge path, so that it flows through optional_params → _map_optional_params_to_responses_api_request → responses_api_request["tool_choice"] → the actual API call.

Environment

• litellm version: 1.82.0
• Python: 3.12