Skills and repository instructions

A DeepSkill has a name, description, instructions, and optional resource files. At startup the agent sees only each skill’s name and description. It loads the detailed instructions through the skill tool when relevant, which keeps static context focused.

discoverSkills(root) adopts the Agent Skills layout. It scans for SKILL.md files two directories deep, expecting each at <root>/<skill>/SKILL.md, and includes the remaining files in that skill directory as relative resources. The Markdown file must begin with YAML frontmatter containing name and description.

---
name: release-check
description: Prepare and verify a release candidate.
---

Run the verification suite before preparing the release notes.

The same discovery call recursively loads repository AGENTS.md files as agent instructions. Use those files for repository-wide and directory-specific operating rules; use skills for focused procedures the model should load only when needed.

You may register skills directly with skills(...) when discovery from the filesystem is unsuitable. Since skill instructions can guide tool use, treat their source directory as trusted application configuration.