kaxil opened a new pull request, #74899: URL: https://github.com/apache/airflow/pull/74899
`AgentSkillsToolset` fails on every run in a fresh install of the Common AI provider with the `skills` extra. The extra asks for `pydantic-ai-skills>=1.2.0` with no upper bound, so today it resolves 2.1.0 (checked with `uv pip compile` against the released 0.10.0). pydantic-ai-skills 2.0 removed `SkillsToolset`, which the toolset wraps, so the import in `__aenter__` fails before the first model request. The error then tells the user to install the `skills` extra, which they already have. This caps the extra and the dev dependency group at `>=1.2.0,<2`. 1.2.0, 1.3.0 and 1.4.0 all export `SkillsToolset`; 2.0.0 and 2.1.0 do not. The error now separates the two cases: if `pydantic_ai_skills` is missing, it still points at the extra; if it is installed without `SkillsToolset`, it names the installed version and the supported range. **Why CI didn't catch it.** The root `pyproject.toml` had a temporary `exclude-newer-package` cutoff for pydantic-ai-skills (in both blocks), added so 1.2.0 could resolve inside the exclude-newer window. It kept `uv.lock` on 1.2.0, so no later release was ever tested. With the cap in place the cutoff is no longer needed, so this removes it, and the lock now moves to 1.4.0, the latest 1.x. That is the only package that changes in `uv.lock`. The existing `AgentSkillsToolset` tests pass on 1.4.0. **Why a cap rather than a port to 2.x.** 2.0 is a rewrite, not a rename: `SkillsToolset` becomes `SkillsCapability`, each skill loads as a deferred pydantic-ai capability, and the tool names the model sees change. Supporting it means moving `AgentSkillsToolset` onto `capabilities=` and remapping `exclude_tools` / `exclude_resources`. That belongs in its own PR, and the cap fixes installs now. Users on an affected release can pin `pydantic-ai-skills<2` next to `apache-airflow-providers-common-ai[skills]`. Related to #74741 -- This is an automated message from the Apache Git Service. To respond to the message, please log on to GitHub and use the URL above to go to the specific comment. To unsubscribe, e-mail: [email protected] For queries about this service, please contact Infrastructure at: [email protected]
