Support excluding files from common.ai Agent Skills discovery

[kaxil]

AgentSkillsToolset gains an optional exclude_resources parameter — a list of glob patterns forwarded to pydantic-ai-skills' SkillsToolset that keeps matched files out of a skill's discovered resources. It mirrors the existing exclude_tools escape hatch and is resolved per-run on the worker (so nothing is baked into the serialized DAG).

Why now

pydantic-ai-skills 1.2.0 broadened resource discovery from a fixed seven-extension allowlist (.md/.json/.yaml/.yml/.csv/.xml/.txt) to every readable text file in a skill bundle, so .sql, .toml, .jinja, and similar assets are exposed with no extra config (DougTrajano/pydantic-ai-skills#54, #55).

That broader discovery also means more files reach the model, so exclude_resources is the trim knob for secrets or scratch files that happen to live in a skill directory. The two ship together: broaden discovery, and give users a way to narrow it back.

Usage

from airflow.providers.common.ai.toolsets.skills import AgentSkillsToolset

toolset = AgentSkillsToolset(
    sources=["./skills"],
    exclude_resources=["*.env", "secrets/*"],  # added on top of the built-in defaults
)

Patterns extend the built-in defaults (__pycache__, *.pyc, *.pyo, .DS_Store, .git) and match either the full skill-relative path or any single path component.

Scope of the guarantee

exclude_resources hides files from resource discovery only — the resource list and the read_skill_resource tool. It does not stop a skill's run_skill_script from reading an excluded file off disk (a skill script runs as a subprocess with the worker's filesystem permissions). When files are genuinely sensitive, pair it with exclude_tools={"run_skill_script"}. This is called out in both the docstring and the toolset guide.

Packaging notes

• The [skills] extra floor moves to pydantic-ai-skills>=1.2.0 (the version that adds the parameter).
• 1.2.0 was published inside the repo's exclude-newer = "4 days" window, so a temporary per-package exclude-newer-package cutoff (2026-07-16T00:00:00Z) is added in both [tool.uv.exclude-newer-package] and [tool.uv.pip.exclude-newer-package] so it resolves. The comment marks it for removal once the rolling window advances past 2026-07-15.
• Regenerating uv.lock for the floor bump also picked up pre-existing drift on main unrelated to this change: the openai provider had already dropped its [datalib] extra in pyproject.toml but the lock was never refreshed, and common-dataquality is re-sorted. uv lock always fully syncs, so this churn is unavoidable when touching the lock.

[gopidesupavan]

Nice :)