Chat Completions model silently ignores reusable prompts

[Aphroq]

Please read this first

• Have you read the docs? Yes. The reusable prompt docs describe the prompt parameter for the Responses API.
• Have you searched for related issues? Yes.

Describe the bug

OpenAIChatCompletionsModel.get_response() and stream_response() accept a Responses reusable prompt, pass it through to _fetch_response(), and then silently drop it when building chat.completions.create() kwargs.

This makes prompt-managed configuration look accepted even though the Chat Completions backend never sends it to the provider. OpenAIResponsesModel does include prompt in its request path, so this is specific to the Chat Completions backend.

Related paths checked:

• Non-streaming and streaming Chat Completions both accept and drop prompt.
• prompt=None is the normal path and should stay unchanged.
• OpenAIResponsesModel already sends prompt explicitly.
• Chat Completions model should reject unsupported server-managed state #3275 / fix: #3275 reject chat completions server state #3279 cover unsupported server-managed state in Chat Completions, not reusable prompts.

Debug information

• Agents SDK version: main at 683b6e79 (latest release tag checked locally: v0.17.0)
• Python version: Python 3.12.1

Repro steps

Run this script against current main:

import asyncio
from typing import Any

import httpx
from openai.types.chat.chat_completion import ChatCompletion, Choice
from openai.types.chat.chat_completion_message import ChatCompletionMessage

from agents import ModelSettings, ModelTracing, OpenAIChatCompletionsModel


class DummyCompletions:
    def __init__(self) -> None:
        self.kwargs: dict[str, Any] = {}

    async def create(self, **kwargs: Any) -> ChatCompletion:
        self.kwargs = kwargs
        return ChatCompletion(
            id="resp-id",
            created=0,
            model="fake",
            object="chat.completion",
            choices=[
                Choice(
                    index=0,
                    finish_reason="stop",
                    message=ChatCompletionMessage(role="assistant", content="ok"),
                )
            ],
        )


class DummyClient:
    def __init__(self, completions: DummyCompletions) -> None:
        self.chat = type("_Chat", (), {"completions": completions})()
        self.base_url = httpx.URL("https://api.openai.com/v1/")


async def main() -> None:
    completions = DummyCompletions()
    model = OpenAIChatCompletionsModel(
        model="gpt-4",
        openai_client=DummyClient(completions),  # type: ignore[arg-type]
    )

    await model.get_response(
        system_instructions=None,
        input="hello",
        model_settings=ModelSettings(),
        tools=[],
        output_schema=None,
        handoffs=[],
        tracing=ModelTracing.DISABLED,
        previous_response_id=None,
        conversation_id=None,
        prompt={"id": "pmpt_123"},  # type: ignore[arg-type]
    )

    print("prompt" in completions.kwargs)
    print([key for key in completions.kwargs if "prompt" in key])


asyncio.run(main())

Actual output:

False
[]

Expected behavior

The Chat Completions backend should not silently accept and drop a reusable prompt. Since reusable prompts are a Responses API feature, Chat Completions should fail fast with a clear UserError, or otherwise explicitly map/send supported prompt fields if that behavior is intended.