airflow.providers.common.ai.skills

Framework-agnostic Agent Skills sources for Airflow.

This module resolves skill sources – a local directory or a GitSkills descriptor – into local SKILL.md directories, cloning Git repositories with a token taken from an Airflow connection. The output is a list of directory paths, the interchange format every Agent Skills implementation consumes (pydantic-ai-skills, LangChain DeepAgents, Strands), so the same Airflow credential handling works across frameworks.

For pydantic-ai use the AgentSkillsToolset binding. For other frameworks, resolve the directories yourself:

from airflow.providers.common.ai.skills import GitSkills, resolve_skills

with resolve_skills(["./skills", GitSkills(repo_url="https://...", conn_id="github")]) as dirs:
    # LangChain DeepAgents
    agent = create_deep_agent(model=..., skills=dirs)
    # ...or Strands
    agent = Agent(plugins=[AgentSkills(skills=dirs)])

Resolution (connection lookup, clone) happens when resolve_skills is entered, so run it inside the task, not at module import / DAG-parse time. The context manager removes any cloned directories on exit.

Attributes

SkillSource

Classes

GitSkills

Agent Skills cloned from a Git repository when resolved.

Functions

resolve_skills(sources)

Resolve skill sources to local SKILL.md directories.

Module Contents

class airflow.providers.common.ai.skills.GitSkills[source]

Agent Skills cloned from a Git repository when resolved.

Parameters:

  • repo_url – HTTPS or SSH URL of the repository to clone.

  • conn_id – Airflow git connection used for credentials, resolved through the Git provider’s GitHook (HTTPS token in the connection password, or an SSH key in the connection’s extra). Set this for private repositories. Plain http:// is rejected when conn_id is set so a credential is never sent in cleartext, and a repo_url with embedded credentials is rejected (use conn_id instead). When conn_id is None the clone is unauthenticated; as with any git clone, the worker’s own git configuration (credential helpers, SSH agent) may still apply, so run workers without ambient git credentials if you need strict isolation.

  • path – Sub-path inside the repository that holds the skill directories (e.g. "skills"). Defaults to the repository root.

  • branch – Branch, tag, or ref to check out. Defaults to the repository’s default branch.

Warning

Skill bundles can contain scripts an agent may run on the worker. Because the repository is fetched at run time, anyone who can modify it can introduce code that runs in your environment, outside DAG review. Point repo_url at a trusted repository and pin branch to a trusted ref.

repo_url: str[source]
conn_id: str | None = None[source]
path: str = ''[source]
branch: str | None = None[source]
airflow.providers.common.ai.skills.SkillSource[source]
airflow.providers.common.ai.skills.resolve_skills(sources)[source]

Resolve skill sources to local SKILL.md directories.

Yields a list of directory paths suitable for any Agent Skills loader (pydantic-ai-skills, LangChain DeepAgents skills=, Strands AgentSkills(skills=...)). Cloned repositories are removed on exit, so use the returned directories inside the with block.

Was this entry helpful?