The GitHub Actions job "Tests (AMD)" on airflow.git/aip99-crewai has failed. Run started by GitHub user kaxil (triggered by kaxil).
Head commit for run: 4e4de2b8f41b69bdefe2bf6b25b7c6ee3d7e5b26 / Kaxil Naik <[email protected]> Use dedicated crewai conn_type and reject non-OpenAI-compatible providers CrewAI's ``LLM`` is a thin LiteLLM wrapper that takes model identifiers in LiteLLM's ``provider/name`` slash format (e.g. ``openai/gpt-4o``). This is distinct from pydantic-ai's ``provider:name`` colon format, so reusing the ``pydanticai`` conn_type was wrong on two counts: the connection's ``model`` UI field is documented as colon-format, and the four pydanticai-* conn shapes don't map uniformly to CrewAI either. - Own ``crewai`` connection type with a ``model`` conn-field documenting the LiteLLM slash convention explicitly. ``default_conn_name`` resolves at runtime via ``llm_conn_id: str | None = None``. - ``_resolve_model_id`` honours ``extra["model"]`` as a fallback via ``conn.extra_dejson`` (parity with PydanticAIHook / LangChainHook). - ``_check_provider_supported`` raises ``NotImplementedError`` for ``bedrock/``, ``vertex_ai/``, ``azure/`` model prefixes. Their auth surface needs region/IAM, project/location, or endpoint+api_version kwargs that this hook does not extract; better to fail loudly than to construct an unusable LLM that explodes mid-agent loop. Per-vendor subclasses are a follow-up. - ``get_ui_field_behaviour`` and ``@staticmethod`` on helpers. - ``[crewai]`` extra pins ``crewai>=1.14.5; python_version < "3.14"``: 1.14.0 was yanked (anthropic dep conflict); 1.14.1-1.14.4 require ``click<8.2`` which conflicts with apache-airflow-breeze's ``click>=8.3``. crewai 1.14.x ships only ``<3.14`` wheels. - Honest "Supported providers" docs: only ``openai/`` and ``anthropic/`` work with the base ``crewai`` install (anthropic SDK is a hard runtime dep of crewai). Other LiteLLM-routed providers need the ``crewai[litellm]`` extra plus the vendor SDK installed separately; the hook does not pull them in. Drops the false "LiteLLM is a transitive dependency" claim. - ``example_crewai_stock_analysis.py``: separate ``crewai_default`` and ``pydanticai_default`` connections (mixed model formats in one conn were mutually exclusive, so the marquee example couldn't run as-is). Pass ``model_id="openai:gpt-4o"`` explicitly to ``LLMOperator``. Add ``execution_timeout=10m`` to the mapped CrewAI task so a stuck ticker doesn't hold a worker slot indefinitely. ``Param`` import goes through ``common.compat.sdk`` per the provider convention. - New ``example_crewai_hook.py``: minimal single-task pattern with ``[START howto_hook_crewai_basic]`` markers for the docs to ``exampleinclude`` (mirrors the LangChainHook docs pattern). Local note: the ``Validate provider.yaml files`` prek hook will fail until ~2026-05-22 because of the 4-day ``exclude-newer`` cutoff in Airflow's uv config (crewai 1.14.5 was published 2026-05-18). Upstream CI will resolve once the 4-day window catches up. Report URL: https://github.com/apache/airflow/actions/runs/26134695786 With regards, GitHub Actions via GitBox --------------------------------------------------------------------- To unsubscribe, e-mail: [email protected] For additional commands, e-mail: [email protected]
