Skills Loader Example

Map local skill modules to typed tools and execute them through the call loop.

Complete Example

TypeScript

import { TheRouter.ai } from "@therouter/sdk";

interface Skill {
  name: string;
  description: string;
  schema: Record<string, unknown>;
  run: (args: Record<string, unknown>) => Promise<unknown>;
}

const registry: Skill[] = [
  {
    name: "summarize_markdown",
    description: "Summarize markdown content",
    schema: {
      type: "object",
      properties: { markdown: { type: "string" } },
      required: ["markdown"],
    },
    run: async ({ markdown }) => String(markdown).slice(0, 120),
  },
];

const tools = registry.map((skill) => ({
  type: "function",
  function: {
    name: skill.name,
    description: skill.description,
    parameters: skill.schema,
  },
}));

const client = new TheRouter.ai({ apiKey: process.env.THEROUTER_API_KEY! });

const first = await client.callModel({
  model: "anthropic/claude-sonnet-4.5",
  input: [{ role: "user", content: "Summarize this markdown and highlight risks." }],
  tools,
});

for (const item of first.items) {
  if (item.type !== "tool_call") continue;
  const skill = registry.find((entry) => entry.name === item.name);
  if (!skill) continue;
  const output = await skill.run(item.arguments as Record<string, unknown>);
  await client.callModel({
    ...first.next_turn,
    input: [{ type: "tool_result", call_id: item.call_id, output }],
  });
}

Skill Metadata

Name	Type	Required	Description
name	string	Required	Unique skill key used as tool name.
description	string	Required	Natural language description for model planning.
schema	JSON Schema	Required	Input parameter schema.
run	function	Required	Skill runtime implementation.

Expected Output

skills-loader-output.json

{
  "skill": "summarize_markdown",
  "status": "executed",
  "output_preview": "Top risks: rollout timing, migration sequencing..."
}

Skill registry hygiene

Version skills explicitly and keep schemas backward compatible to avoid breaking long-running automations.