Browser Ops Routing

A bilingual routing skill for browser-heavy agent workflows.
Use the safest workable execution layer first: API/CLI -> Structured Browser Automation -> Visual Browser Control -> Human Confirmation

中文介绍 · English Overview · Examples · What's Included · Release Notes

Why This Exists

Agent browser work often fails for the same reason: teams jump straight into browser control before deciding whether the task should have been done through an API, a CLI, deterministic browser automation, or a human confirmation boundary.

This repository packages a reusable routing policy that helps agents:

• avoid unnecessary browser work
• prefer deterministic paths over fragile ones
• escalate to visual control only when it is justified
• stop before high-risk actions such as login approval, posting, deletion, payment, or security changes

中文介绍

Browser Ops Routing 是一个面向智能体网页任务的双语 skill。它不强调“让智能体接管浏览器”，而是强调先做正确的分流。

核心原则是：

• 能走 API/CLI 就不要先上浏览器
• 能走结构化浏览器自动化，就不要直接上视觉模式
• 视觉浏览器只作为兜底
• 登录授权、验证码、发布、删除、付款等高风险动作必须停下来人工确认

这套方法适合：

• Codex
• OpenClaw
• 其他带浏览器能力的 agent 系统

它的目标不是绑定某个产品，而是提供一套稳定、可解释、可复用的网页任务路由规则。

English Overview

Browser Ops Routing is a lightweight policy skill for agent-driven browser work.

It gives agents a simple but practical execution order:

1. API/CLI
2. Structured browser automation
3. Visual browser control
4. Human confirmation

This keeps browser-heavy workflows more reliable, easier to reason about, and safer to operate across different agent stacks.

Examples

Example prompts

• Use $browser-ops-routing to handle this login-gated dashboard task and stop before any irreversible action.
• 使用 $browser-ops-routing 判断这一步应该走 API、结构化浏览器还是视觉模式。
• Use $browser-ops-routing to review this browser workflow and identify where human confirmation is required.

For longer examples, see examples/.

What's Included

• SKILL.md Shared routing policy for browser work
• agents/openai.yaml UI metadata for skill pickers and skill libraries
• references/openclaw.md OpenClaw-specific implementation guidance kept outside the core shared policy
• examples/ Short bilingual usage examples for common browser-heavy agent tasks

Design Principles

• Keep human and AI browser state separated
• Use the least fragile layer that can complete the task
• Treat login, publishing, deletion, payments, and verification as confirmation boundaries
• Keep machine-specific setup outside the core skill

Suggested Uses

• browser-heavy agent workflows
• mixed API plus browser operations
• login-gated dashboards
• anti-bot-fragile web tasks
• OpenClaw or Codex browser policy standardization

Scope

This repository intentionally ships a generic policy layer.

It does not include:

• machine-specific ports or paths
• browser-brand-specific personal preferences
• private environment fixes
• CAPTCHA bypass logic

Release Notes

v0.1.0

Initial public release:

• bilingual README with Chinese and English framing
• reusable SKILL.md for browser task routing
• OpenClaw-specific reference guidance
• UI metadata for skill pickers

License

MIT