@easynet/agent-skill

Introduction

@easynet/agent-skill discovers and loads skills, builds a SkillSet, and registers it into the default AgentContext so runtime packages can inject or route skills.

API Reference

| API | What it does | Minimal usage | | --- | --- | --- | | createAgentSkills | Load skills from config or a directory and register the resulting SkillSet. | await createAgentSkills("./skill.yaml") |

Usage

Load skills from the default embedded config:

import { createAgentSkills } from "@easynet/agent-skill";

const skillSet = await createAgentSkills();
console.log(skillSet.list());

Load skills from a custom config file:

import { createAgentSkills } from "@easynet/agent-skill";

await createAgentSkills("./skill.yaml");

YAML config

Example skill.yaml:

apiVersion: easynet.world/v1
kind: SkillConfig
metadata:
  name: agent-skill-default
spec:
  path: ./skills
  mode: prompt
  inject_metadata: true
  embedding_threshold: 0.85

Explanation:

• spec.path: directory that contains skill folders with SKILL.md.
• spec.mode: how matched skills are applied, such as prompt or subagent.
• spec.inject_metadata: whether skill metadata is injected into the prompt.
• spec.embedding_threshold: minimum semantic match score.

SKILL Structure

agent-skill now treats a skill as three progressive layers:

1. Level 1: metadata in SKILL.md frontmatter (name, description).
2. Level 2: the SKILL.md Markdown body, used as detailed instructions.
3. Level 3: bundled resources loaded on demand, such as scripts/, references/, and assets/.

Example layout:

skills/
  code-review/
    SKILL.md
    references/
      CHECKLIST.md
    scripts/
      collect-context.sh
    assets/
      review-template.md

Level 3 is convention-based, not folder-name-locked:

• scripts/: executable helper code or command wrappers.
• references/: longer background docs, checklists, forms, protocol notes.
• assets/: templates, sample data, images, or other payload files.

Any file under the skill directory except SKILL.md is treated as a resource. Resource type is inferred by extension:

• Markdown/text/YAML/JSON -> instructions
• Shell/JS/TS/Python/etc. -> code
• everything else -> data

Level 3 Resource API

SkillSet exposes lazy resource access:

const skillSet = await createAgentSkills({ path: "./skills" });

const resources = await skillSet.listResources?.("code-review");
const docs = await skillSet.getResourcesByType?.("code-review", "instructions");
const script = await skillSet.readResource?.("code-review", "scripts/collect-context.sh");

Notes:

• listResources(name) returns resource metadata only.
• getResourcesByType(name, type) filters by inferred type.
• readResource(name, relativePath) reads the file as a Buffer.
• relative paths are constrained to stay inside the skill directory.