Skills

Read-only knowledge packages for agents

Skills are read-only knowledge packages mounted into agent sandboxes at /skills/<name>/. They provide context the agent can reference — coding guidelines, API documentation, workflow guides, etc.

Two Sources of Skills

1. Standalone Skills (via skills config)

{
  skills: {
    "code-review": { path: "./skills/code-review" },
  },
  agents: {
    assistant: {
      skills: ["code-review"],
    },
  },
}

2. Plugin-Provided Skills

Plugins can register skills during their register() call:

register(reg) {
  reg.skill({
    name: "telegram-guide",
    path: resolve(__dirname, "skill"),
    description: "Guide for using Telegram tools",
  });
}

Plugin-provided skills are referenced the same way in agent config:

agents: {
  assistant: {
    skills: ["code-review", "telegram-guide"],
  },
}

Skill Package Structure

skills/code-review/
├── skill.json     # Manifest: name, description, dependencies
├── README.md      # Context file included in agent's system prompt
└── SKILL.md       # Alternative context file name

skill.json

{
  "name": "code-review",
  "description": "Code review guidelines and best practices",
  "contextFile": "README.md",
  "requires": {
    "tools": ["git"],
    "skills": []
  }
}

The requires field declares dependencies:

• tools: The agent must have these tools to use this skill
• skills: Other skills that should also be loaded

How Skills Are Mounted

Host:     ./skills/code-review/
Sandbox:  /skills/code-review/ (read-only)

The skill’s context file (README.md by default) is included in the agent’s system prompt, giving the agent knowledge without consuming tool calls.