‹ 首页

browser-automation-router

@dirtytrii · 收录于 1 周前 · 上游提交 1 周前

Choose the narrowest reliable browser surface across Codex Browser, signed-in Chrome, Playwright CLI, and platform scripts. Use for browser automation, logged-in workflows, UI inspection, screenshots, form filling, publish previews, or frontend QA.

适合你,如果需要在多种浏览器环境中自动执行操作

/ 通过 npx 安装 校验哈希
npx oh-my-skill add dirtytrii/codex-skills/browser-automation-router
/ 通过 bash 安装
curl -fsSL https://oh-my-skill.com/install.sh | bash -s -- dirtytrii/codex-skills/browser-automation-router
/ 已经装过?验证本机副本,不用重装
npx oh-my-skill verify dirtytrii/codex-skills/browser-automation-router
安装目标可用 --agent / --scope 或 --to 明确指定;省略时只会在唯一已存在的 agent 目录上自动选择,零命中或多命中会停止并提示。content_hash 缺失或不一致均拒装。
9GitHub stars
~711最小装载
~711含声明引用
~758文本包总量
索引托管

怎么用

商店整理自技能原文 · 版本 f142db6 · 表述以原文为准
它做什么

Claude 会自动选择最合适的浏览器工具(如内置浏览器、Chrome 扩展、Playwright 等)来执行自动化任务,例如截图、填写表单、检查页面等。

什么时候触发

当你要求 Claude 进行浏览器自动化操作,如打开网页、截图、填写表单、检查 UI 或运行测试时触发。

装好后可以这样说
Claude 会使用内置浏览器打开百度并截图。
Claude 会通过 Chrome 扩展使用你的登录状态。
Claude 会使用 Playwright 执行自动化测试。
技能原文 SKILL.md作者撰写 · MIT · f142db6

Browser Automation Router

Choose the narrowest reliable browser surface before acting.

Route Order
  1. Prefer a purpose-built connector, API, or CLI when it fully supports the requested read or write action.
  2. Use Codex's in-app Browser for public pages, localhost, visual review, annotations, downloads, and isolated sign-in state.
  3. Use the Codex Chrome extension when the task requires the user's existing Chrome profile, signed-in session, open tabs, or extensions.
  4. Use $playwright for deterministic terminal/CI runs, repeatable regression suites, tracing, or unattended batch execution.
  5. Use a platform-specific automation script only when it exposes a tested batch/export contract that the native browser surface does not replace.
  6. Use Computer Use only for browser-external desktop UI or a browser flow unsupported by the Browser/Chrome surfaces.
Native Browser Rule
  • Do not maintain custom JavaScript snippets, selectors, browser launchers, cookie importers, or profile-path logic for ordinary interactive browser work.
  • Follow the installed Browser or Chrome plugin skill and its live documentation. Do not copy its internal bootstrap code into this repository.
  • Prefer visible page state, snapshots, semantic locators, normal clicks/fills, and screenshots.
  • Use page evaluation only when a high-level read-only operation cannot obtain the evidence. Never use it to inspect cookies, local storage, passwords, tokens, or profile files.
  • Reuse the current browser binding or claimed tab during the task; do not create duplicate sessions without a reason.
Fail-Closed Capability Gate

The native route requires Codex Desktop with the Browser or Chrome plugin available in the current task.

  • Capability floor: Codex Desktop release dated 2026-06-11 or later for the Browser/Chrome Developer mode used by this routing policy.
  • OpenAI does not publish a stable numeric desktop build floor for this feature. Treat the release date plus live plugin availability as the minimum contract.
  • The in-app Browser has its own browser state. It does not inherit the user's Chrome profile.
  • Chrome profile reuse requires the Codex Chrome extension and user-approved access to the intended tab/account.
  • Plan, workspace policy, region, and rollout state may still disable a surface on a nominally new enough app.
  • If the required plugin is absent or disabled, report the missing capability and select an explicit Playwright/platform-script fallback. Do not pretend the native route ran.
Write Gates
  • Keep navigation and read-only extraction separate from publish, comment, follow, like, bookmark, delete, account-switch, or settings changes.
  • Confirm the active account and exact write action immediately before execution.
  • Stop on captcha, unexpected login, wrong account, permission prompts, ambiguous targets, or changed page state.
  • Never commit credentials, cookies, login state, QR payloads, or local Chrome profile paths.
Evidence

Return the selected surface, capability check, target account boundary, actions performed, screenshots or output artifacts, blocked steps, and fallback used. Do not claim success from planned actions alone.

按 MIT 许可原样转载,未经改动 · 在 GitHub 查看 →

评论

登录即可评论;带「已验证安装」的,是发布者名下有本店的安装或持有记录。