Bug Description

When using the OpenAI Responses API, function_call items from a previous response are duplicated in subsequent requests.

This results in the same tool call appearing multiple times in the OpenAI logs for a single turn.

Expected Behavior

• Each tool call should appear only once in OpenAI logs

Reproduction Steps

1. Configure the agent LLM to use openai response api:

llm = openai.responses.LLM(
    model="gpt-5.4",
    temperature=0.4,
)

2. Ask the agent anything that triggers a tool call.
3. Inspect the OpenAI logs for the request.

Operating System

macOS Tahoe

Models Used

GPT 5.4

Package Versions

livekit-agents[openai]==1.4.6

Session/Room/Call IDs

No response

Proposed Solution

Filter out previously emitted function_call items when previous_response_id is present:

    async def _run_impl(self) -> None:
        self._response_completed = False
        chat_ctx, _ = self._chat_ctx.to_provider_format(format="openai.responses")

        if "previous_response_id" in self._extra_kwargs and self._llm._pending_tool_calls:
            chat_ctx = [
                item for item in chat_ctx
                if not (
                    isinstance(item, dict)
                    and item.get("type") == "function_call"
                    and item.get("call_id") in self._llm._pending_tool_calls
                )
            ]

Additional Context

No response

Screenshots and Recordings

No response