feat(language_model): add LiteLLM provider for 100+ backings - #1386

The lock had drifted far behind: 30 of the 52 direct dependencies were
behind their latest release, some by a major version. Refreshing it moves
104 locked packages, adds 6 and removes 2, with no downgrades. Nine are
major bumps: complexipy 6 -> 8, cryptography 49 -> 50, fastmcp 3 -> 4 (and
fastmcp-slim), mcp 1 -> 2, sentence-transformers 5 -> 6, setuptools 83 ->
84, websockets 16 -> 17, xxhash 3 -> 4. Direct dependencies that moved
include fastapi 0.139.0 -> 0.141.1, instructor 1.15.4 -> 1.17.0, langchain-aws
1.6.2 -> 1.7.5, neo4j 6.2.0 -> 6.3.0, openai 2.45.0 -> 2.54.0, qdrant-client
1.18.0 -> 1.19.0, sqlalchemy 2.0.51 -> 2.0.52, ty 0.0.59 -> 0.0.79 and
uvicorn 0.51.0 -> 0.52.4.

Two specifiers had to change for the refresh to be honest:

- litellm was capped at <1.85 when the provider was added (MemMachine#1386, which
  records no reason for the bound), and dependabot later widened it to
  <1.86. That range ends inside a stretch of releases -- 1.83.8 through
  1.92.x -- that declare Requires-Python >=3.10,<3.14. uv ignores that
  upper bound and installed 1.85.7 anyway, but pip does not: on Python
  3.14, `pip install memmachine-server[litellm]` resolved to 1.83.7, not
  the 1.85.7 in the lock. 1.93.0 is the first release that supports 3.14,
  so that is the floor now, and the lock moves to 1.100.0. litellm also
  requires openai<3, which is what holds the OpenAI SDK on the 2.x line;
  without the floor the resolver satisfies that by walking litellm
  backwards to 1.83.0 instead.
- boto3-stubs was pinned exactly at 1.43.13 while boto3 resolved to
  1.43.90, so the stubs no longer described the installed SDK.

Two source changes the new versions require, and one cleanup:

- ty 0.0.79 rejects assigning over a bound method, so the metrics-factory
  wiring test installs its fake engine with monkeypatch.setattr.
- `list(vector.flat)` is a list of numpy scalars, not `list[float]`, which
  the newer numpy typing now reports at the assignment. The replay path
  converts explicitly.
- The dev group listed complexipy twice.

Left behind deliberately: nebula5-python (see the cap and its reason),
openai 3.x (litellm requires openai<3, and instructor's jiter<0.15 pin
holds openai below 3.11 regardless; 3.0 also swaps the transport to
httpx2), and ruff, still pinned at 0.15.14 -- 0.16 formats Markdown code
blocks, which would reformat six docs, and adds LOG004, which flags four
`logger.exception` calls, two of them genuine misuses and two in a helper
that is only ever called from an except block.

All six ty jobs pass, as do `ruff check`, `ruff format --check`,
`uv lock --check` and the unit suite.

Co-Authored-By: Claude Opus 5 (1M context) <[email protected]>
Claude-Session: https://claude.ai/code/session_01BcPSoGsQjnVJS8A5NkoGZN

The lock had drifted far behind: 30 of the 52 direct dependencies were
behind their latest release, some by a major version. Refreshing it moves
104 locked packages, adds 6 and removes 2, with no downgrades. Nine are
major bumps: complexipy 6 -> 8, cryptography 49 -> 50, fastmcp 3 -> 4 (and
fastmcp-slim), mcp 1 -> 2, sentence-transformers 5 -> 6, setuptools 83 ->
84, websockets 16 -> 17, xxhash 3 -> 4. Direct dependencies that moved
include fastapi 0.139.0 -> 0.141.1, instructor 1.15.4 -> 1.17.0, langchain-aws
1.6.2 -> 1.7.5, neo4j 6.2.0 -> 6.3.0, openai 2.45.0 -> 2.54.0, qdrant-client
1.18.0 -> 1.19.0, sqlalchemy 2.0.51 -> 2.0.52, ty 0.0.59 -> 0.0.79 and
uvicorn 0.51.0 -> 0.52.4.

Two specifiers had to change for the refresh to be honest:

- litellm was capped at <1.85 when the provider was added (MemMachine#1386, which
  records no reason for the bound), and dependabot later widened it to
  <1.86. That range ends inside a stretch of releases -- 1.83.8 through
  1.92.x -- that declare Requires-Python >=3.10,<3.14. uv ignores that
  upper bound and installed 1.85.7 anyway, but pip does not: on Python
  3.14, `pip install memmachine-server[litellm]` resolved to 1.83.7, not
  the 1.85.7 in the lock. 1.93.0 is the first release that supports 3.14,
  so that is the floor now, and the lock moves to 1.100.0. litellm also
  requires openai<3, which is what holds the OpenAI SDK on the 2.x line;
  without the floor the resolver satisfies that by walking litellm
  backwards to 1.83.0 instead.
- boto3-stubs was pinned exactly at 1.43.13 while boto3 resolved to
  1.43.90, so the stubs no longer described the installed SDK.

Two source changes the new versions require, and one cleanup:

- ty 0.0.79 rejects assigning over a bound method, so the metrics-factory
  wiring test installs its fake engine with monkeypatch.setattr.
- `list(vector.flat)` is a list of numpy scalars, not `list[float]`, which
  the newer numpy typing now reports at the assignment. The replay path
  converts explicitly.
- The dev group listed complexipy twice.

Left behind deliberately: nebula5-python (see the cap and its reason),
openai 3.x (litellm requires openai<3, and instructor's jiter<0.15 pin
holds openai below 3.11 regardless; 3.0 also swaps the transport to
httpx2), and ruff, still pinned at 0.15.14 -- 0.16 formats Markdown code
blocks, which would reformat six docs, and adds LOG004, which flags four
`logger.exception` calls, two of them genuine misuses and two in a helper
that is only ever called from an except block.

All six ty jobs pass, as do `ruff check`, `ruff format --check`,
`uv lock --check` and the unit suite.

Co-Authored-By: Claude Opus 5 (1M context) <[email protected]>
Claude-Session: https://claude.ai/code/session_01BcPSoGsQjnVJS8A5NkoGZN

The lock had drifted far behind: 30 of the 52 direct dependencies were
behind their latest release, some by a major version. Refreshing it moves
104 locked packages, adds 6 and removes 2, with no downgrades. Nine are
major bumps: complexipy 6 -> 8, cryptography 49 -> 50, fastmcp 3 -> 4 (and
fastmcp-slim), mcp 1 -> 2, sentence-transformers 5 -> 6, setuptools 83 ->
84, websockets 16 -> 17, xxhash 3 -> 4. Direct dependencies that moved
include fastapi 0.139.0 -> 0.141.1, instructor 1.15.4 -> 1.17.0, langchain-aws
1.6.2 -> 1.7.5, neo4j 6.2.0 -> 6.3.0, openai 2.45.0 -> 2.54.0, qdrant-client
1.18.0 -> 1.19.0, sqlalchemy 2.0.51 -> 2.0.52, ty 0.0.59 -> 0.0.79 and
uvicorn 0.51.0 -> 0.52.4.

Two specifiers had to change for the refresh to be honest:

- litellm was capped at <1.85 when the provider was added (MemMachine#1386, which
  records no reason for the bound), and dependabot later widened it to
  <1.86. That range ends inside a stretch of releases -- 1.83.8 through
  1.92.x -- that declare Requires-Python >=3.10,<3.14. uv ignores that
  upper bound and installed 1.85.7 anyway, but pip does not: on Python
  3.14, `pip install memmachine-server[litellm]` resolved to 1.83.7, not
  the 1.85.7 in the lock. 1.93.0 is the first release that supports 3.14,
  so that is the floor now, and the lock moves to 1.100.0. litellm also
  requires openai<3, which is what holds the OpenAI SDK on the 2.x line;
  without the floor the resolver satisfies that by walking litellm
  backwards to 1.83.0 instead.
- boto3-stubs was pinned exactly at 1.43.13 while boto3 resolved to
  1.43.90, so the stubs no longer described the installed SDK.

Two source changes the new versions require, and one cleanup:

- ty 0.0.79 rejects assigning over a bound method, so the metrics-factory
  wiring test installs its fake engine with monkeypatch.setattr.
- `list(vector.flat)` is a list of numpy scalars, not `list[float]`, which
  the newer numpy typing now reports at the assignment. The replay path
  converts explicitly.
- The dev group listed complexipy twice.

Left behind deliberately: nebula5-python (see the cap and its reason),
openai 3.x (litellm requires openai<3, and instructor's jiter<0.15 pin
holds openai below 3.11 regardless; 3.0 also swaps the transport to
httpx2), and ruff, still pinned at 0.15.14 -- 0.16 formats Markdown code
blocks, which would reformat six docs, and adds LOG004, which flags four
`logger.exception` calls, two of them genuine misuses and two in a helper
that is only ever called from an except block.

All six ty jobs pass, as do `ruff check`, `ruff format --check`,
`uv lock --check` and the unit suite.

Co-Authored-By: Claude Opus 5 (1M context) <[email protected]>
Claude-Session: https://claude.ai/code/session_01BcPSoGsQjnVJS8A5NkoGZN

…edkick) (#1596)

* Bring the dependency lock up to date

The lock had drifted far behind: 30 of the 52 direct dependencies were
behind their latest release, some by a major version. Refreshing it moves
104 locked packages, adds 6 and removes 2, with no downgrades. Nine are
major bumps: complexipy 6 -> 8, cryptography 49 -> 50, fastmcp 3 -> 4 (and
fastmcp-slim), mcp 1 -> 2, sentence-transformers 5 -> 6, setuptools 83 ->
84, websockets 16 -> 17, xxhash 3 -> 4. Direct dependencies that moved
include fastapi 0.139.0 -> 0.141.1, instructor 1.15.4 -> 1.17.0, langchain-aws
1.6.2 -> 1.7.5, neo4j 6.2.0 -> 6.3.0, openai 2.45.0 -> 2.54.0, qdrant-client
1.18.0 -> 1.19.0, sqlalchemy 2.0.51 -> 2.0.52, ty 0.0.59 -> 0.0.79 and
uvicorn 0.51.0 -> 0.52.4.

Two specifiers had to change for the refresh to be honest:

- litellm was capped at <1.85 when the provider was added (#1386, which
  records no reason for the bound), and dependabot later widened it to
  <1.86. That range ends inside a stretch of releases -- 1.83.8 through
  1.92.x -- that declare Requires-Python >=3.10,<3.14. uv ignores that
  upper bound and installed 1.85.7 anyway, but pip does not: on Python
  3.14, `pip install memmachine-server[litellm]` resolved to 1.83.7, not
  the 1.85.7 in the lock. 1.93.0 is the first release that supports 3.14,
  so that is the floor now, and the lock moves to 1.100.0. litellm also
  requires openai<3, which is what holds the OpenAI SDK on the 2.x line;
  without the floor the resolver satisfies that by walking litellm
  backwards to 1.83.0 instead.
- boto3-stubs was pinned exactly at 1.43.13 while boto3 resolved to
  1.43.90, so the stubs no longer described the installed SDK.

Two source changes the new versions require, and one cleanup:

- ty 0.0.79 rejects assigning over a bound method, so the metrics-factory
  wiring test installs its fake engine with monkeypatch.setattr.
- `list(vector.flat)` is a list of numpy scalars, not `list[float]`, which
  the newer numpy typing now reports at the assignment. The replay path
  converts explicitly.
- The dev group listed complexipy twice.

Left behind deliberately: nebula5-python (see the cap and its reason),
openai 3.x (litellm requires openai<3, and instructor's jiter<0.15 pin
holds openai below 3.11 regardless; 3.0 also swaps the transport to
httpx2), and ruff, still pinned at 0.15.14 -- 0.16 formats Markdown code
blocks, which would reformat six docs, and adds LOG004, which flags four
`logger.exception` calls, two of them genuine misuses and two in a helper
that is only ever called from an except block.

All six ty jobs pass, as do `ruff check`, `ruff format --check`,
`uv lock --check` and the unit suite.

Co-Authored-By: Claude Opus 5 (1M context) <[email protected]>
Claude-Session: https://claude.ai/code/session_01BcPSoGsQjnVJS8A5NkoGZN

* Drop dependencies nothing declares them for

Four entries in the server's dependency list are not requirements of this
package:

- boto3-stubs is a type-stub distribution sitting in the runtime list.
  Nothing imports it, and `ty check --project packages/server` reports
  the same "All checks passed" with the package uninstalled, so it was
  buying neither runtime behaviour nor type coverage. (Base boto3-stubs
  only types the boto3 entry points; per-service clients need the
  mypy-boto3-* extras. If stronger boto3 typing is wanted later, that
  belongs in the dev group with those extras.)
- regex belongs to memmachine-common, which declares it and imports it in
  api/spec.py. Nothing in the server package uses it.
- The PyPI "dotenv" distribution ships no modules at all -- only
  dist-info -- and exists to depend on python-dotenv.
  `importlib.metadata.packages_distributions()["dotenv"]` is
  `["python-dotenv"]`, so `from dotenv import load_dotenv` has always
  been python-dotenv. Declare that instead.
- greenlet is what SQLAlchemy's asyncio extra installs, and this package
  uses create_async_engine, so ask for `sqlalchemy[asyncio]` and let it
  bring greenlet. That also covers the platforms SQLAlchemy's own
  platform_machine marker on greenlet leaves out.

The lock loses boto3-stubs, botocore-stubs, types-s3transfer and dotenv.
greenlet and python-dotenv stay, now for stated reasons.

reranker_manager took `runtime_checkable` from typing_extensions, an
undeclared dependency it reached through pydantic. It has been in
`typing` since 3.8 and this package requires 3.12, so it comes from
`typing` now, next to the `Protocol` the file already imports there.
memmachine-common and memmachine-client keep their typing_extensions
dependency: they support 3.10, and `Self` and `Unpack` are 3.11.

Deliberately kept, though nothing imports them, because each is named in
the source or by the tests:

- asyncpg and aiosqlite are the SQLAlchemy drivers this package builds
  URLs for. DatabaseBackend spells them out -- POSTGRES is
  ("postgres", SqlAlchemyConf, "postgresql", "asyncpg") -- SqlAlchemyConf
  renders `{dialect}+{driver}://`, and database_manager hands that to
  create_async_engine (and gates statement/connect timeouts on
  `conf.driver == "asyncpg"`).
- tzdata on Windows. prompt_utilities builds `zoneinfo.ZoneInfo(tz)`, and
  Windows ships no system tz database, so without it every lookup --
  including the UTC fallback in the except branch -- raises
  ZoneInfoNotFoundError.
- milvus-lite, which pymilvus does not declare. It is what serves a
  file-backed Milvus URI, both for the local `path` deployment mode and
  for test_milvus_vector_store, which opens
  `MilvusClient(uri=tmp_path / "test_milvus.db")` behind
  `pytest.importorskip("milvus_lite")` -- 22 tests in the default run.

All six ty jobs pass, as do `ruff check`, `ruff format --check`,
`uv lock --check` and the unit suite.

Co-Authored-By: Claude Opus 5 (1M context) <[email protected]>
Claude-Session: https://claude.ai/code/session_01BcPSoGsQjnVJS8A5NkoGZN

---------

Co-authored-by: Claude Opus 5 (1M context) <[email protected]>

The lock had drifted far behind: 30 of the 52 direct dependencies were
behind their latest release, some by a major version. Refreshing it moves
112 locked packages, adds 6 and removes 2, with no downgrades. Nine are
major bumps: complexipy 6 -> 8, cryptography 49 -> 50, fastmcp 3 -> 4 (and
fastmcp-slim), mcp 1 -> 2, sentence-transformers 5 -> 6, setuptools 83 ->
84, websockets 16 -> 17, xxhash 3 -> 4. Direct dependencies that moved
include fastapi 0.139.0 -> 0.141.1, instructor 1.15.4 -> 1.17.0, langchain-aws
1.6.2 -> 1.7.6, neo4j 6.2.0 -> 6.3.1, openai 2.45.0 -> 2.54.0, qdrant-client
1.18.0 -> 1.19.0, sqlalchemy 2.0.51 -> 2.0.53, ty 0.0.59 -> 0.0.81 and
uvicorn 0.51.0 -> 0.53.0.

Two specifiers had to change for the refresh to be honest:

- litellm was capped at <1.85 when the provider was added (MemMachine#1386, which
  records no reason for the bound), and dependabot later widened it to
  <1.86. That range ends inside a stretch of releases -- 1.83.8 through
  1.92.x -- that declare Requires-Python <3.14. uv ignores that upper
  bound and installed 1.85.7 anyway, but pip does not: on Python 3.14,
  `pip install memmachine-server[litellm]` resolves to 1.83.7, not the
  1.85.7 in the lock. 1.93.0 is the first release that supports 3.14,
  so that is the floor now, and the lock moves to 1.101.0. litellm also
  requires openai<3, which is what holds the OpenAI SDK on the 2.x line;
  without the floor the resolver satisfies that by walking litellm
  backwards to 1.83.0 instead.
- boto3-stubs was pinned exactly at 1.43.13 while boto3 resolved to
  1.43.47, so the stubs no longer described the installed SDK.

One source change the new versions require, and one cleanup:

- `list(vector.flat)` is a list of numpy scalars, not `list[float]`, which
  the newer numpy typing now reports at the assignment. The replay path
  converts explicitly.
- The dev group listed complexipy twice.

Left behind deliberately: nebula5-python (see the cap and its reason),
openai 3.x (litellm requires openai<3, and instructor's jiter<0.15 pin
holds openai below 3.11 regardless; 3.0 also swaps the transport to
httpx2), and ruff, still pinned at 0.15.14 -- 0.16 formats Markdown code
blocks, which would reformat six docs, and adds LOG004, which flags four
`logger.exception` calls, two of them genuine misuses and two in a helper
that is only ever called from an except block.

Same change as the first half of MemMachine#1596 on speedkick, applied to main.

Co-Authored-By: Claude Opus 5 (1M context) <[email protected]>

…e ruff to 0.16 (#1646)

* Bring the dependency lock up to date

The lock had drifted far behind: 30 of the 52 direct dependencies were
behind their latest release, some by a major version. Refreshing it moves
112 locked packages, adds 6 and removes 2, with no downgrades. Nine are
major bumps: complexipy 6 -> 8, cryptography 49 -> 50, fastmcp 3 -> 4 (and
fastmcp-slim), mcp 1 -> 2, sentence-transformers 5 -> 6, setuptools 83 ->
84, websockets 16 -> 17, xxhash 3 -> 4. Direct dependencies that moved
include fastapi 0.139.0 -> 0.141.1, instructor 1.15.4 -> 1.17.0, langchain-aws
1.6.2 -> 1.7.6, neo4j 6.2.0 -> 6.3.1, openai 2.45.0 -> 2.54.0, qdrant-client
1.18.0 -> 1.19.0, sqlalchemy 2.0.51 -> 2.0.53, ty 0.0.59 -> 0.0.81 and
uvicorn 0.51.0 -> 0.53.0.

Two specifiers had to change for the refresh to be honest:

- litellm was capped at <1.85 when the provider was added (#1386, which
  records no reason for the bound), and dependabot later widened it to
  <1.86. That range ends inside a stretch of releases -- 1.83.8 through
  1.92.x -- that declare Requires-Python <3.14. uv ignores that upper
  bound and installed 1.85.7 anyway, but pip does not: on Python 3.14,
  `pip install memmachine-server[litellm]` resolves to 1.83.7, not the
  1.85.7 in the lock. 1.93.0 is the first release that supports 3.14,
  so that is the floor now, and the lock moves to 1.101.0. litellm also
  requires openai<3, which is what holds the OpenAI SDK on the 2.x line;
  without the floor the resolver satisfies that by walking litellm
  backwards to 1.83.0 instead.
- boto3-stubs was pinned exactly at 1.43.13 while boto3 resolved to
  1.43.47, so the stubs no longer described the installed SDK.

One source change the new versions require, and one cleanup:

- `list(vector.flat)` is a list of numpy scalars, not `list[float]`, which
  the newer numpy typing now reports at the assignment. The replay path
  converts explicitly.
- The dev group listed complexipy twice.

Left behind deliberately: nebula5-python (see the cap and its reason),
openai 3.x (litellm requires openai<3, and instructor's jiter<0.15 pin
holds openai below 3.11 regardless; 3.0 also swaps the transport to
httpx2), and ruff, still pinned at 0.15.14 -- 0.16 formats Markdown code
blocks, which would reformat six docs, and adds LOG004, which flags four
`logger.exception` calls, two of them genuine misuses and two in a helper
that is only ever called from an except block.

Same change as the first half of #1596 on speedkick, applied to main.

Co-Authored-By: Claude Opus 5 (1M context) <[email protected]>

* Drop dependencies nothing declares them for

Four entries in the server's dependency list are not requirements of this
package:

- boto3-stubs is a type-stub distribution sitting in the runtime list.
  Nothing imports it, and `ty check --project packages/server` reports
  the same "All checks passed" with the package uninstalled, so it was
  buying neither runtime behavior nor type coverage. (Base boto3-stubs
  only types the boto3 entry points; per-service clients need the
  mypy-boto3-* extras. If stronger boto3 typing is wanted later, that
  belongs in the dev group with those extras.)
- regex belongs to memmachine-common, which declares it and imports it in
  api/spec.py. Nothing in the server package uses it.
- The PyPI "dotenv" distribution ships no modules at all -- only
  dist-info -- and exists to depend on python-dotenv.
  `importlib.metadata.packages_distributions()["dotenv"]` is
  `["python-dotenv"]`, so `from dotenv import load_dotenv` has always
  been python-dotenv. Declare that instead.
- greenlet is what SQLAlchemy's asyncio extra installs, and this package
  uses create_async_engine, so ask for `sqlalchemy[asyncio]` and let it
  bring greenlet. That also covers the platforms SQLAlchemy's own
  platform_machine marker on greenlet leaves out.

The lock loses boto3-stubs, botocore-stubs, types-s3transfer and dotenv.

reranker_manager also took runtime_checkable from typing_extensions, an
undeclared dependency it reached through pydantic. It has been in
typing since 3.8 and this package requires 3.12, so it now comes from
typing, next to the Protocol the file already imports there.
memmachine-common and memmachine-client keep their typing_extensions
dependency: they support 3.10, and Self and Unpack are 3.11.

Same change as the second half of #1596 on speedkick, applied to main.

Co-Authored-By: Claude Opus 5 (1M context) <[email protected]>

* Log without a traceback where no exception is in flight

openai_embedder's dimensionality check and amazon_bedrock_reranker's
score-count check each call logger.exception() and then raise their own
ExternalServiceAPIError. Nothing is being handled at that point:
logger.exception is error(..., exc_info=True), and with no active
exception it appends "NoneType: None" to the record. logger.error is
what both sites mean; the raise that follows carries the message.

ruff 0.16 reports both as LOG004 (`.exception()` call outside exception
handlers).

Co-Authored-By: Claude Opus 5 (1M context) <[email protected]>

* Re-raise the HTTPError the client's project-error helper is given

_handle_get_project_http_error takes the caught HTTPError as a
parameter, but its bare `raise` statements and logger.exception() calls
ignored it and used the exception the *caller's* except block had
active. That works only while the helper is called from inside that
block, and ruff 0.16's LOG004 cannot see the call boundary, so it
reports the two logging calls as .exception() outside a handler.

The helper now uses what it is handed: logger.error(..., exc_info=error)
attaches the same traceback logger.exception would have, and
`raise error` re-raises the same object. Callers observe no difference:
the exception identity, its original traceback and the 422 ValueError's
__cause__ are unchanged (the helper's own frame is appended to the
traceback, as it is for any re-raise from a callee).

Co-Authored-By: Claude Opus 5 (1M context) <[email protected]>

* Move ruff to 0.16.8

ruff was held at 0.15.14 by the dependency refresh because 0.16 brings
two things that touch committed content, and both now have their own
commits ahead of this one:

- LOG004 (`.exception()` outside an exception handler), selected via
  the LOG group. Its four findings are resolved by the two preceding
  commits, so `ruff check` is clean at this pin with no suppressions.
- The formatter now formats Python code blocks inside Markdown. The six
  docs it touches are reformatted here, by `ruff format` and nothing
  else: USAGE.md, examples/v1/README.md,
  integrations/aws_strands_agent_sdk/README.md,
  integrations/langgraph/README.md, maintainers/build-pip-packages.md,
  packages/client/README.md. Trailing commas, collapsed argument
  lists, and column-aligned comments brought to two spaces; no words
  change.

The lint workflow's ruff-action reads the pin from pyproject.toml, so
CI enforces 0.16.8 from this commit on. No .py file is reformatted by
the new version.

Co-Authored-By: Claude Opus 5 (1M context) <[email protected]>

* Explain only the litellm floor in the server's pyproject

The litellm extra's comment also said litellm's openai<3 is what keeps
the resolved OpenAI SDK on the 2.x line. That describes the universal
lock, which resolves every extra, and not a requirement of this package:
the server hands the SDK no httpx client, which is the only thing
openai 3.0 changed, and its suite passes on openai 3.3.0 with the extra
absent (1869 passed, 3 skipped, the same as on the lock). Installs
without the extra may resolve openai 3.x. The comment now states only
the floor, which is this package's own decision.

Co-Authored-By: Claude Opus 5 (1M context) <[email protected]>

---------

Co-authored-by: Claude Opus 5 (1M context) <[email protected]>