Skill System

The Skill System allows the host application to expose business capabilities to the AI. You simply register functions, and the AI can call them via scripts — no extra adaptation needed, and documentation is auto-generated.

Two Registration Approaches

	Declarative (Recommended)	Imperative (Advanced)
Registration Method	Set agentConfig property	Call defineRegistry
Ease of Use	Zero config, works out of the box	Requires manual management
Use Case	Most scenarios	Dynamic or conditional registration needed

Declarative registration example:

agent.agentConfig = {
  name: 'MyApp',
  persona: 'You are a ... assistant',
  groups: [{
    name: 'editor',
    description: 'Editor operations',
    functions: [
      {
        name: 'getCode',
        description: 'Get the current code in the editor',
        handler: () => editor.getCode()
      }
    ]
  }]
};

💡 After setting agentConfig, the system automatically creates the Registry and completes function registration — the entire process requires no manual intervention.

Function Definition

Each function consists of the following fields:

Field	Type	Description
name	string	Function name
description	string	Function description, included in auto-generated documentation
parameters	ParameterDef[]	Parameter definition array (ParameterDef[] format)
returns	object	Return value definition
handler	function	Execution function (supports async)
hooks	object	UI hooks (see Hook System)

parameters uses ParameterDef[] array format to describe parameters, enabling the AI to generate correct call code:

{
  name: 'createOrder',
  description: 'Create a new order',
  parameters: [
    { name: 'productId', schema: { type: 'string', description: 'Product ID' }, required: true },
    { name: 'quantity', schema: { type: 'number', description: 'Quantity' }, required: false }
  ],
  handler: async ({ productId, quantity }) => {
    return await api.createOrder(productId, quantity ?? 1);
  }
}

💡 Each parameter consists of name (parameter name), schema (parameter definition in OpenAPI Schema format), and required (whether it is required).

Function Groups

Functions are organized through groups; group name + function name = full invocation path:

Groups keep function namespaces clear and organized, avoiding naming conflicts.

AI Invocation Method

The AI executes code in a sandbox via the script tool, using Proxy chain syntax to call functions:

// Proxy chain invocation — natural API style
await rtcAgent.order.create({ productId: '123', quantity: 2 });

// Equivalent direct call
await rtcAgent.execute('order.create', { productId: '123', quantity: 2 });

💡 After reading the auto-generated function documentation, the AI knows how to call these functions — no extra configuration needed.

Hook System

Hooks allow the host to inject custom logic at various stages of function execution:

Hook	Trigger	Description
onStart	Before execution	Can throw CancelledError to cancel execution
onSuccess	After success	Runs asynchronously, does not block return
onError	After failure	Runs asynchronously, does not block return
onProgress	Progress update	Triggered when handler calls onProgress(n)

onStart is the only hook that can intercept execution — ideal for confirmation dialogs, permission checks, etc.

Auto-generated Documentation

Each time a function is registered, the system generates comprehensive documentation:

Generated Content	Description
Function Docs	Includes description, parameter table, return values, and call examples
INDEX.md	Function index; the AI uses this to discover available functions
AGENT.md	Updates the Agent capability description

📌 The AI learns how to use functions by reading this documentation — documentation quality directly affects the AI’s calling accuracy.

Scenarios

Scenarios load business workflow documentation, telling the AI how to handle specific business processes:

Feature	Description
Scenario Docs	Business workflow descriptions in Markdown format
Loading Method	Specified via the <rtc-agent scenarios-url="..."> attribute
Storage Location	/scenarios/{slug}.md
AI Usage	AI reads scenario docs to understand business processes and operational guidelines

For example, you can write scenario docs like “How to handle refunds” or “How to create an order”, and the AI will execute operations according to your business specifications.

Next Steps

• RTC Protocol — Learn about the core mechanism for AI calling frontend tools
• Virtual File System — Learn where function docs and scenario docs are stored
• Command System — Learn about user-facing command interaction capabilities