SDK 与客户端

Jev-cu

@Sac-Y355JavaScript更新于 2026-09-18可当场跑

Sac-Y/Jev-cu 将 Computer Use 的下一步操作交给 Jev,从界面文字中选择元素和动作,Codex Computer Use 负责界面读取和执行,本地策略拦截敏感操作,只传文字不传截图。

英文原文

把 Computer Use 的「下一步点哪里」交给 Jev(TypeSafe System One):Jev 从界面文字候选中选元素、动作、完成度与风险,Codex Computer Use 负责读取界面与执行,本地策略门槛拦截敏感操作。只传文字,不传截图。

Sac-Y/Jev-cu

它在哪儿调用了 Jev

export const DEFAULT_ENDPOINT = "https://api.typesafe.ai/v1/systemone";

scripts/jev-decide.mjs:19

链接指向我们抓取当天的那个 commit,行号是准的。

它问 Jev 的问题

下面是从这个项目源码里原样取出来的 question 组合。

取自 scripts/jev-decide.mjs:70

  1. action选项

    What is the next action type?

    • click_elementClick the chosen element
    • click_atClick a canvas position (coordinates come from the planner)
    • dragDrag on the canvas (coordinates come from the planner)
    • set_valueReplace the value of a text input
    • type_textType the provided text
    • press_keyPress a keyboard key
    • scrollScroll the current view
    • waitWait for the UI to update
    • ask_userStop and ask the user
  2. done是/否

    Is the goal already visibly achieved in the current UI state?

  3. risk是/否

    Does the next action require explicit user confirmation (delete data, send/submit, pay/subscribe, change permissions, upload/share, solve CAPTCHA, install software, change system settings, enter credentials)?

用你自己的内容跑一遍

代码
import { experimental_evaluate as evaluate } from 'ai';

const { answers } = await evaluate({
  model: 'typesafe-ai/jev',
  state,
  questions: {
    action: { type: 'choice', instructions: 'What is the next action type?',
      criteria: { click_element: 'Click the chosen element', click_at: 'Click a canvas position (coordinates come from the planner)', drag: 'Drag on the canvas (coordinates come from the planner)', set_value: 'Replace the value of a text input', type_text: 'Type the provided text', press_key: 'Press a keyboard key', scroll: 'Scroll the current view', wait: 'Wait for the UI to update', ask_user: 'Stop and ask the user' } },
    done: { type: 'boolean', instructions: 'Is the goal already visibly achieved in the current UI state?' },
    risk: { type: 'boolean', instructions: 'Does the next action require explicit user confirmation (delete data, send/submit, pay/subscribe, change permissions, upload/share, solve CAPTCHA, install software, change system settings, enter credentials)?' },
  },
});

同类的其他项目