Real tool invocation and in-request runtime

Core registers and executes real tools and returns receipts carrying an invocation; business authorization is owned by the implementation registered by the App.

ToolRegistry input and output parsers run synchronously; invoke is a synchronous call and rejects when it encounters a Promise, while invokeAsync allows an asynchronous execute. Both results are always {value, invocation}, and tool failures throw instead of masquerading as successful values.

A bare ExecutableTool is an execution mechanism, and business authorization in parseInput/execute must be guaranteed by the App itself. The recommended approach is Support.defineToolCatalog, which generates the Schema, authorize, execute, parseOutput, and model tool from a single source, then binds the real App dependencies.

After binding the tool runtime once with bindToolRuntime, nodes supply only toolId/input; Core automatically wires in the node execution scope and cancellation. events is a summary of real calls, not a list of tools proposed by the model.

Progressive examples for the same business case

Minimal local read-only tool

Beginner · Node.js

TypeScript

import { ToolRegistry } from '@jangts/agent-core/tools';
const tools = new ToolRegistry().register({
  id: 'get_status', kind: 'read', description: '读取演示服务状态',
  parseInput(value: unknown) { if (!value || typeof value !== 'object' || Array.isArray(value)) throw new TypeError('需要对象'); return {}; },
  execute: () => '服务运行正常',
  parseOutput(value: unknown) { if (typeof value !== 'string') throw new TypeError('需要文本'); return value; },
}).seal();
const result = tools.invoke({ toolId: 'get_status', input: {} });
console.log(result.value, result.invocation.invocationId);

JavaScript

import { ToolRegistry } from '@jangts/agent-core/tools';
const tools = new ToolRegistry().register({
    id: 'get_status', kind: 'read', description: '读取演示服务状态',
    parseInput(value) { if (!value || typeof value !== 'object' || Array.isArray(value))
        throw new TypeError('需要对象'); return {}; },
    execute: () => '服务运行正常',
    parseOutput(value) { if (typeof value !== 'string')
        throw new TypeError('需要文本'); return value; },
}).seal();
const result = tools.invoke({ toolId: 'get_status', input: {} });
console.log(result.value, result.invocation.invocationId);

Bind in advance so nodes no longer inject tools repeatedly

Configuration · Node.js

The first example validates external input strictly; the second only shows the binding relationship, and real business code should use the appropriate Schema and authorization.

TypeScript

import { AgentCoreEngine, defineWorkflow, route } from '@jangts/agent-core/engine';
import { ToolRegistry } from '@jangts/agent-core/tools';
type State = { answer?: unknown };
const tools = new ToolRegistry().register({ id: 'get_status', kind: 'read',
  parseInput: () => ({}), execute: async () => '服务运行正常', parseOutput: value => value,
}).seal();
const workflow = defineWorkflow<State, unknown>({ id: 'status', nodes: [{
  id: 'read', description: '读取状态', startsRound: true,
  async run(_state, { tools }) { const { value } = await tools.invoke({ toolId: 'get_status', input: {} }); return { answer: value }; },
  routes: [route.deliver('answer')],
}] });
const agent = new AgentCoreEngine({ workflow }).bindToolRuntime(tools);
console.log(await agent.run({}));

JavaScript

import { AgentCoreEngine, defineWorkflow, route } from '@jangts/agent-core/engine';
import { ToolRegistry } from '@jangts/agent-core/tools';
const tools = new ToolRegistry().register({ id: 'get_status', kind: 'read',
    parseInput: () => ({}), execute: async () => '服务运行正常', parseOutput: value => value,
}).seal();
const workflow = defineWorkflow({ id: 'status', nodes: [{
            id: 'read', description: '读取状态', startsRound: true,
            async run(_state, { tools }) { const { value } = await tools.invoke({ toolId: 'get_status', input: {} }); return { answer: value }; },
            routes: [route.deliver('answer')],
        }] });
const agent = new AgentCoreEngine({ workflow }).bindToolRuntime(tools);
console.log(await agent.run({}));

Policies & implementations

• Check order for tool input, execution, output, and late results