agentsclimarketplace

Browser verify

Skill webkubor/agent-skills/skills/browser-verify

Production-tested, tool-agnostic skills for AI coding agents — Claude Code, Codex, Gemini CLI, Cursor and beyond. 生产验证的工具中立 Agent 技能库。

Install
npx -y skills add webkubor/agent-skills --skill browser-verify

Assembled from the repository path, not quoted from the project. Check it against their README if it does not work.

2 things to look at

  • 22 days oldThe repository was created 22 days ago. New is not bad, but a brand new repository carrying a familiar-sounding name is the shape a typosquat arrives in, and there has been no time for anyone else to find a problem with it.
  • 0 stars0 stars. Stars are a popularity signal and not a quality one, but at this level it is likely that nobody has read this closely except its author, and you would be relying on your own review.

What its author says it does

Copied from the file, not written here

Use when you need to control a real browser or visually verify rendered UI (screenshot, read on-screen values, click, check light/dark/mobile, diagnose blank/empty pages). Built on browser-harness (CDP), works cross-agent. Use for any UI visual DoD / "看渲染 / 截图验证 / 浏览器控制".

SKILL.md

3.3 KB, as published. Nobody here has run it

── 质量追踪 ──

{ "status": "production", "rating": "★★★★★", "last_used": "2026-07-16", "total_uses": 1, "trace": [ {"date": "2026-07-16", "usage": "收录 agent-skills", "result": "✅ 已投产使用"} ] }

浏览器控制 / UI 可视化验证(browser-harness)

跨 Agent 的浏览器控制能力,依赖 browser-harness(CDP 直控本机 Chrome)。它是命令行工具,Claude / Codex / Gemini 都能 shell 调用——不绑定某个 Agent 专属的浏览器 MCP,所以跨 Agent。

何时用

  • UI 改动的视觉验收(DoD):亲眼看渲染(明暗 / 移动宽度),不靠猜。线上改动要看部署后的真实环境(test/prod),不是只看 localhost。
  • 排查「空白 / 没数据」:区分是数据问题还是渲染问题(先看接口/日志,再看渲染)
  • 确认交互、布局、配色的实际效果

调用形态(heredoc,python 内核)

内网 / 本地一律带 NO_PROXY 防 Clash 劫持:

NO_PROXY=127.0.0.1,localhost browser-harness <<'PY'
new_tab("http://localhost:5173/dashboard")   # 首次导航用 new_tab;别用 goto(会抢用户当前标签页)
wait_for_load()
import time; time.sleep(2)                    # SPA 数据异步,留一点渲染时间
p = capture_screenshot()                      # 截图落 /tmp/shot.png,再用 Read 工具看图下结论
print("SHOT:", p)
PY
  • 读 DOM 文本/值js("(()=>{ ...return 字符串... })()")——比截图快,适合核对具体数值/文案
  • 点击:先截图读坐标 → click_at_xy(x, y) → 再截图验证(命中走 Chrome 合成层,穿透 iframe/shadow)
  • 滚动到元素js("document.querySelector('.sel')?.scrollIntoView({block:'center'})")
  • 输入type_text("...") / press_key("Enter")

项目级注意事项(接入你的项目时按此模板补充)

把下面几件事写进你项目自己的团队 skill,通用能力保持干净:

  • 端口与环境:dev 端口是多少;改动部署后要在哪个真实环境(test/prod)实测,别只看 localhost。看不到新内容先排除浏览器缓存(硬刷新)+ 确认 JS bundle hash 已更新
  • 登录墙:用测试账号或复用浏览器已登录会话;被重定向到登录页就停下来问人,永远不要从截图里读取/填写密码
  • 页面空白先看运行时日志(如接入了遥测:errors.log / api-calls.log):后端 503 会让整页空白,那是环境问题不是前端 bug
  • 截图后必须用读图工具看过再下结论——不允许裸报「渲染正常」(视觉 DoD 硬要求)

完整工具参考

npm i -g browser-harness 后运行 browser-harness --help;核心 API:new_tab / wait_for_load / page_info / capture_screenshot / js / click_at_xy / type_text / press_key 等;远程/并行(多 sub-agent 各自独立浏览器)见其 interaction-skills/ 与 remote daemon 章节。

Keep looking

Skills are one crate of 328,083. Ordering is by how many stacks a row turns up in, so the top of any crate is what has actually been picked rather than what has the most stars.