pierrejeambrun opened a new pull request, #76309: URL: https://github.com/apache/airflow/pull/76309
When an agent behaves differently from one run to the next, the first thing to check is which skills it had and which ones it actually loaded. Until now the AI Model tab only showed a total tool-call count, so that meant reading the task log. A skill pulled from Git can also change between two runs without the Dag changing. Recording the commit that was cloned next to each skill makes the difference traceable: "it behaved differently yesterday" becomes "the skill changed between these two commits". `AgentSkillsToolset` now publishes, under the namespaced XCom key `__AIRFLOW__COMMON_AI_SKILLS__`, every skill it offered the agent, where it came from (a local path, or the repository, path inside it, branch and cloned commit), and how many times the agent called `load_skill`, `read_skill_resource` and `run_skill_script` for it. The toolset pushes it itself rather than the operator, so it also works when the toolset is used from a plain `@task`. It respects `do_xcom_push`, and a failure to publish never fails the run. The AI Model tab lists those skills, loaded ones first. Hovering a skill shows its description and the calls made for it. A skill from a GitHub or GitLab repository links to its folder at the commit that ran. Recording a commit for local skills that live in a Git checkout (e.g. the Dag bundle) is left for a follow-up. SCREENSHOTS --- ##### Was generative AI tooling used to co-author this PR? - [X] Yes — Claude Code (Opus 5.5) Generated-by: Claude Code (Opus 5.5) following [the guidelines](https://github.com/apache/airflow/blob/main/contributing-docs/05_pull_requests.rst#gen-ai-assisted-contributions) -- 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]
