Skill tools

When at least one skill is loaded, Anvia creates four tools that let the model progressively access skill content.

Generated tools

Tool	Purpose
get_skill_instructions	Load the complete SKILL.md instructions for one skill.
get_skill_reference	Read one listed file under references/.
get_skill_script	Read one listed file under scripts/.
run_skill_script	Execute one listed skill script.

These tools are included in the SkillSet and attached by .skills(skillSet).

What the agent sees first

The generated instruction block lists skill names, descriptions, reference paths, and script paths. The full bodies remain unloaded until the model calls the relevant tool.

User task
  → select skill by name and description
  → get_skill_instructions
  → get_skill_reference or get_skill_script when directed
  → run_skill_script only when execution is needed

This keeps unrelated skill content out of the model context.

Attach the complete skill set

const productSkills = await loadSkills(skill.local('skills'))

const agent = new AgentBuilder('release-assistant', model)
  .instructions('Use skills when they are relevant to the task.')
  .skills(productSkills)
  .defaultMaxTurns(4)
  .build()

const response = await agent
  .prompt(
    'Draft release notes for the streaming and retrieval improvements.',
  )
  .send()

The model may load the release-note instructions, then read a style guide or run a listed helper if the skill directs it to do so.

Keep the catalog selective

Descriptions should make skill choice clear. Too many overlapping skills make selection ambiguous and add model-facing tool context.

Load only skills appropriate for the current agent role and environment. A loaded skill exposes its catalog entry and generated access paths to that agent.

Treat script execution as a tool boundary

Containment prevents arbitrary paths, but run_skill_script still executes reviewed code. Do not load executable skill packages into a runtime that should be read-only. Keep secrets out of script output and apply normal tool-event filtering to user-facing streams.