Skip to content

Skills System

Skills are defined as SKILL.md files following the Agent Skills Spec. They are discovered, loaded, and injected into agent prompts.

Skill Locations (Three Tiers)

  • ~/.claude/skills/*/SKILL.md — User-wide (default)
  • .claude/skills/*/SKILL.md — Project-wide (e.g., openspec-*)
  • .agents/skills/*/SKILL.md — Agent/workflow skills
  • MCP servers via skill:// resource URIs

Key Files

  • src/agentpool/skills/skill.py — Skill model: YAML frontmatter parsing, lazy instruction loading
  • src/agentpool/skills/registry.py — SkillsRegistry auto-discovers SKILL.md files from configured paths
  • src/agentpool/skills/manager.py — SkillsManager pool-level lifecycle
  • src/agentpool/skills/uri_resolver.py — skill:// URI scheme resolver
  • src/agentpool/skills/command.py — SkillCommand wraps skills as protocol-agnostic slash commands
  • src/agentpool/skills/instruction_provider.py — SkillsInstructionProvider injects skills as XML into prompts (metadata/full modes) — migrated from resource_providers/

Injection Modes

Via YAML skills.instruction:

  • off — No injection
  • metadata — <available-skills> XML block (names + descriptions)
  • full — <skill_content> XML block with complete instructions + parameters