@executor-js/execution

中文文档
1.6.11发布于 2026/10/23MITTypeScriptES Module构建来源已验证近 30 天下载 39.8万

Sandboxed JavaScript execution for an executor. Hands a tools.<namespace>.<name>(...) proxy into a code sandbox so an agent (or any caller) can run generated TypeScript/JavaScript that invokes the executor's registered tools.

原始使用文档
文档目录8 节
  1. Install
  2. Usage
  3. Custom tool discovery
  4. Pause/resume for elicitation
  5. Workflow description for LLMs
  6. Using with Effect
  7. Status
  8. License

@executor-js/execution

Sandboxed JavaScript execution for an executor. Hands a tools.<namespace>.<name>(...) proxy into a code sandbox so an agent (or any caller) can run generated TypeScript/JavaScript that invokes the executor's registered tools.

Supports pause/resume for elicitation-driven flows: tools that need user input (OAuth, form fill, approval) suspend the sandbox, surface a PausedExecution, and resume on a ResumeResponse.

Install

bun add @executor-js/sdk @executor-js/execution @executor-js/runtime-quickjs
# or
npm install @executor-js/sdk @executor-js/execution @executor-js/runtime-quickjs

@executor-js/runtime-quickjs is the sandbox runtime. It's not a dependency of @executor-js/execution — you bring your own so consumers with a different runtime don't ship ~13 MB of WASM they never use.

Usage

import { createExecutor } from "@executor-js/sdk";
import { createExecutionEngine } from "@executor-js/execution";
import { makeQuickJsExecutor } from "@executor-js/runtime-quickjs";

const executor = await createExecutor({
  onElicitation: "accept-all",
});

const engine = createExecutionEngine({
  executor,
  codeExecutor: makeQuickJsExecutor({
    timeoutMs: 2_000,
    memoryLimitBytes: 32 * 1024 * 1024,
  }),
});

const result = await engine.execute(
  `
    const pets = await tools.petstore.findPetsByStatus({ status: "available" });
    return pets.length;
  `,
  {
    onElicitation: async (ctx) => {
      // A tool asked for user input mid-execution. Your UI decides what to do.
      console.log("tool needs input:", ctx.request);
      return { action: "decline" };
    },
  },
);

console.log(result);
// { result: 12, logs: [...] }

Custom tool discovery

tools.search(...) uses Executor's built-in lexical tool discovery by default. Hosts can provide their own implementation, such as an indexed or semantic search provider, without replacing the sandbox runtime:

import { createExecutionEngine, type ToolDiscoveryProvider } from "@executor-js/execution";

const toolDiscoveryProvider: ToolDiscoveryProvider = {
  searchTools: ({ query, namespace, limit, offset }) =>
    mySearchIndex.searchTools({ query, namespace, limit, offset }),
};

const engine = createExecutionEngine({
  executor,
  codeExecutor: makeQuickJsExecutor(),
  toolDiscoveryProvider,
});

Pause/resume for elicitation

When the host doesn't support inline elicitation, use executeWithPause to intercept the first request as a pause point:

import type { ExecutionEngine } from "@executor-js/execution";

declare const engine: ExecutionEngine;
declare const code: string;

const started = await engine.executeWithPause(code);

if (started.status === "paused") {
  const { id, elicitationContext } = started.execution;
  // Render the elicitation request in your UI. Later:
  const resumed = await engine.resume(id, {
    action: "accept",
    content: { name: "Ada" },
  });
}

Workflow description for LLMs

import type { ExecutionEngine } from "@executor-js/execution";

declare const engine: ExecutionEngine;

const description = await engine.getDescription();
// Returns the short `execute` tool description: a one-line intro, a pointer to
// the `execute` skill, and the live connection-prefix inventory. Feed this to
// an LLM as the execute tool's description.

The full "use tools.search(), then tools.describe.tool(), then call ..." workflow prose lives in the execute skill rather than the always-loaded description, so a model that never runs code does not pay for it. MCP hosts expose it through the skills tool; to read it directly:

import { EXECUTE_SKILL } from "@executor-js/execution";

const howTo = EXECUTE_SKILL.body;

Using with Effect

If you're building on @executor-js/sdk/core (the raw Effect entry), import from the /core subpath. The returned engine is Effect-native: execute, executeWithPause, and resume all become Effect.Effect<...>, and onElicitation is an ElicitationHandler returning Effect.Effect<ElicitationResponse>.

import { createExecutionEngine } from "@executor-js/execution/core";

Status

Pre-1.0. APIs may still change between beta releases. Part of the executor monorepo.

License

MIT

版本与发布记录

@executor-js/execution 共发布 78 个版本,首次发布于 2026/04/18;过去 12 个月发布了 78 个版本。最新版本附带 npm provenance 构建来源证明,可追溯到公开的 CI 构建记录。

版本发布日期距今
1.6.11latest2026/10/2322 天前
1.6.102026/09/191 个月前
1.6.92026/09/181 个月前
1.6.82026/09/062 个月前
1.6.72026/08/302 个月前
1.6.62026/08/302 个月前

其他发布通道:beta → 0.0.1-beta.4。

关于 @executor-js/execution 的常见问题

@executor-js/execution 是什么?

Sandboxed JavaScript execution for an executor. Hands a tools.<namespace>.<name>(...) proxy into a code sandbox so an agent (or any caller) can run generated TypeScript/JavaScript that invokes the executor's registered tools.

如何安装 @executor-js/execution?

在项目目录运行 npm install @executor-js/execution,或使用 pnpm add / yarn add / bun add。它还需要宿主项目提供 peer 依赖:effect(4.0.0-beta.59)。npm 7 及以上与 pnpm 8 及以上默认会自动安装必需的 peer 依赖,Yarn 需要手动安装。

@executor-js/execution 支持 TypeScript 吗?

支持。包内自带 TypeScript 类型声明(package.json 声明了 types 入口),安装后即可获得类型提示,无需再安装 @types 包。

@executor-js/execution 支持 ESM 还是 CommonJS?

以 ES Module 形式发布,推荐使用 import 引入。在 CommonJS 项目中,Node.js 20.19+ / 22.12+ 可以直接 require() 不含顶层 await 的 ES Module,更早的版本需要改用动态 import()。

@executor-js/execution 可以商用吗?

@executor-js/execution 使用 MIT 许可。宽松许可:允许商业使用、修改和再分发,通常只需在分发时保留原版权与许可声明。以上为概括,不构成法律意见,请以许可原文为准。

@executor-js/execution 的下载量有多少?

根据 npm 官方下载统计,@executor-js/execution 在 2026/10/14 – 2026/11/12 这 30 天内被下载 397,794 次,日均约 13,260 次。下载量会把 CI 构建、镜像同步和作为其他包依赖的安装一并计入,适合横向比较热度,不能直接等同于使用人数。

@executor-js/execution 还在维护吗?

最新版本 1.6.11 发布于 2026/10/23(22 天前);至今共发布 78 个版本,其中过去 12 个月发布了 78 个;首次发布于 2026/04/18。发布频率只是参考信号:功能稳定的工具可能很少需要更新,建议结合源码仓库的 issue 和提交记录判断。