Anvia

Skill tools ​

When at least one skill is loaded, Anvia creates four tools for progressive disclosure.

get_skill_instructions loads the complete instructions for one skill.

get_skill_reference reads one discovered reference path.

get_skill_script reads one discovered script as text.

run_skill_script executes one discovered script with optional string arguments and timeout.

1. Understand the initial catalog ​

The agent initially sees each skill's name, description, reference paths, and script paths, plus descriptions of the four generated tools. Full file bodies remain unloaded.

text
Task
  -> choose a skill from name and description
  -> get_skill_instructions
  -> read a directed reference or script
  -> run a script only when execution is required

2. Attach the complete set ​

ts
import { Agent } from '@anvia/core'
import { loadSkills, skill } from '@anvia/core/skills'

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

const agent = new Agent({
  id: 'release-assistant',
  model,
  skills: productSkills,
  maxTurns: 4,
})

const response = await agent.generate({
    prompt: 'Draft release notes for the streaming improvements.'
})

skills: productSkills attaches catalog instructions and tools together. You may register skillSet.tools manually, but then you must also decide how the model learns the catalog.

3. Keep the catalog selective ​

Load only skills appropriate for the agent role and environment. Too many overlapping descriptions add tool context and make selection ambiguous.

4. Treat generated tools as a privileged boundary ​

Path containment prevents arbitrary requested paths, but listed scripts still execute with process permissions. Generated skill tools are specially marked and their outputs do not pass through ordinary onToolOutput middleware.

Review all loaded content, restrict executable skill sources, avoid sensitive output, and filter runtime events before sending them to users.

Next, verify package validation.

Built for Anvia.