Drop crewai from workspace install path; users install separately

crewai 1.14.x ships tight conservative pins (click~=8.1.7, tomli~=2.0.2,
plus a crewai-cli that pulls textual>=7.5 -> rich>=14.2) that conflict
with Airflow's resolved click 8.3+, tomli 2.4+, and FAB's transitive
rich<14. uv can satisfy the install via override-dependencies, but the
CI image build runs ``pip check`` (see
scripts/docker/install_airflow_when_building_images.sh) which treats
the metadata mismatch as fatal.

The cleanest fix is to keep crewai out of Airflow's resolution entirely:

- Remove the ``[project.optional-dependencies] crewai`` extra from the
  provider pyproject. There's now nothing for uv to resolve against
  crewai's pins -- the root override-dependencies and crewai-* cooldown
  bypass are no longer needed.
- Drop crewai from the provider's dev dependency group.
- Document in docs/hooks/crewai.rst that users ``pip install
  "crewai>=1.14.5"`` themselves after installing common-ai.
- Stub ``crewai`` in ``sys.modules`` for the hook tests (mirrors the
  langchain-stub pattern from #67237 -- justified for crewai because
  the pin conflict makes installation in CI infeasible).

The hook itself is unchanged; users who want it install crewai into
their own venv where they've accepted the pin cascade. The lock no
longer carries crewai / chromadb / onnxruntime entries.

Author: kaxil

providers/common/ai/docs/hooks/crewai.rst

@@ -51,7 +51,7 @@ Supported providers
 
 The hook forwards ``conn.password`` as ``api_key`` and ``conn.host`` as
 ``base_url`` to ``crewai.LLM``. The combinations that work out of the box
-with just the ``crewai`` extra installed:
+with just ``crewai`` installed:
 
 - ``openai/gpt-4o``, ``openai/gpt-4o-mini``
 - ``openai/...`` against an OpenAI-compatible endpoint, with ``host`` pointing
@@ -107,16 +107,27 @@ Parameters
 Dependencies
 ------------
 
-Install the ``crewai`` extra to use this hook::
+Install ``crewai`` separately into your venv after installing the
+``common-ai`` provider::
 
-    pip install apache-airflow-providers-common-ai[crewai]
+    pip install apache-airflow-providers-common-ai
+    pip install "crewai>=1.14.5"
 
-That extra installs ``crewai`` itself (which depends on ``openai`` and
-``anthropic`` at runtime). For other providers, install CrewAI's own
-``litellm`` extra plus the vendor SDK separately, e.g.::
+There is intentionally **no** ``[crewai]`` extra on this provider.
+``crewai`` 1.14.x ships tight conservative pins (``click~=8.1.7``,
+``tomli~=2.0.2``, plus a ``crewai-cli`` co-dependency that pulls
+``textual>=7.5`` → ``rich>=14.2``) that conflict with Airflow's own
+resolved ``click>=8.3``, ``tomli>=2.4``, and FAB's transitive
+``rich<14``. uv can resolve the mismatch via ``override-dependencies``,
+but the CI image build runs ``pip check`` afterwards and treats the
+metadata conflict as fatal. The cleanest answer is to keep ``crewai``
+out of Airflow's lock entirely; users who accept the pin cascade
+install ``crewai`` themselves.
 
-    pip install "crewai[litellm]" langchain-mistralai
+For non-OpenAI / non-Anthropic providers, install CrewAI's own
+``litellm`` extra plus the vendor SDK::
 
-The ``crewai`` extra is currently gated on ``python_version < "3.14"`` because
-CrewAI 1.14.x publishes only ``<3.14,>=3.10`` wheels. Newer Python envs will
-silently skip the extra.
+    pip install "crewai[litellm]>=1.14.5" langchain-mistralai
+
+CrewAI 1.14.x publishes wheels for Python 3.10-3.13; 3.14 isn't supported
+upstream yet.

providers/common/ai/pyproject.toml

@@ -103,13 +103,17 @@ dependencies = [
     "llama-index-embeddings-openai>=0.6.0",
     "llama-index-llms-openai>=0.6.0",
 ]
-# crewai pulls chromadb, which transitively requires onnxruntime.
-# onnxruntime >=1.24 dropped cp310 wheels, and crewai itself caps Python
-# to <3.14. Gate the extra to 3.11-3.13 so 3.10 / 3.14 envs don't fail
-# to resolve.
-"crewai" = [
-    'crewai>=1.14.5; python_version >= "3.11" and python_version < "3.14"',
-]
+# crewai is intentionally NOT declared as an optional extra here. crewai
+# 1.14.x ships tight conservative pins (click~=8.1.7, tomli~=2.0.2, and
+# a crewai-cli that pulls textual>=7.5 → rich>=14.2) that conflict with
+# Airflow's own resolved click 8.3+, tomli 2.4+, and FAB's transitive
+# rich<14. uv can satisfy the install via override-dependencies, but
+# Airflow's CI image build runs ``pip check`` afterwards (see
+# scripts/docker/install_airflow_when_building_images.sh) which treats
+# the metadata mismatch as fatal. Users who want the CrewAI hook should
+# ``pip install crewai>=1.14.5`` into their venv after installing
+# common-ai; see docs/hooks/crewai.rst. Hook tests stub ``crewai`` in
+# ``sys.modules``.
 "pdf" = ["pypdf>=4.0.0"]
 "docx" = ["python-docx>=1.0.0"]
 
@@ -129,7 +133,10 @@ dev = [
     "llama-index-core>=0.13.0",
     "llama-index-embeddings-openai>=0.6.0",
     "llama-index-llms-openai>=0.6.0",
-    'crewai>=1.14.5; python_version >= "3.11" and python_version < "3.14"',
+    # crewai deliberately NOT listed here, and also NOT declared as an
+    # optional extra above -- see the comment under
+    # ``[project.optional-dependencies]`` for the pip-check rationale.
+    # Tests stub ``crewai`` in ``sys.modules``.
 ]
 
 # To build docs:

providers/common/ai/tests/unit/common/ai/hooks/test_crewai.py

@@ -16,13 +16,29 @@
 # under the License.
 from __future__ import annotations
 
+import sys
 from unittest.mock import MagicMock, patch
 
 import pytest
 
 from airflow.providers.common.ai.hooks.crewai import CrewAIHook
 
 
+@pytest.fixture(autouse=True)
+def _stub_crewai_module():
+    """Stub ``crewai`` in sys.modules so @patch can resolve targets without
+    the real package installed.
+
+    crewai is intentionally NOT in the provider's dev group (see
+    ``providers/common/ai/pyproject.toml`` for the rationale -- crewai's
+    pins conflict with Airflow's resolved click/tomli/rich and break the
+    CI image build's ``pip check``). Tests have to stub it explicitly.
+    """
+    crewai_mod = MagicMock()
+    with patch.dict(sys.modules, {"crewai": crewai_mod}):
+        yield
+
+
 def _conn(password: str = "", host: str = "", extra: dict | None = None) -> MagicMock:
     mock_conn = MagicMock()
     mock_conn.password = password