cdp-browser（CDP 浏览器）

cdp-browser skill

Install this file in your repo as .agents/skills/cdp-browser/SKILL.md.

Use this skill when a task needs real-browser execution rather than static HTTP fetching.

Quick workflow

bunx cdp-browser start
bunx cdp-browser nav https://example.com --new
bunx cdp-browser wait-network-idle
bunx cdp-browser dismiss-cookies
bunx cdp-browser eval "document.title"
bunx cdp-browser screenshot
bunx cdp-browser watch
bunx cdp-browser logs-tail --follow
bunx cdp-browser net-summary

Optional version pinning

bunx cdp-browser@0.1.3 nav https://example.com

Default operating pattern

1. Start a managed browser session.
2. Navigate to the target page and wait for network idle.
3. Remove overlays when needed (dismiss-cookies).
4. Extract data with eval; use pick when selector discovery is needed.
5. Capture artifacts (screenshot, logs-tail, net-summary) when requested.
6. Return results with all generated file paths.

Commands

• start [--fresh] [--copy-profile [name]] [--browser <path-or-name>]
• nav <url> [--new]
• eval '<expression>'
• screenshot
• pick '<message>'
• dismiss-cookies [--reject]
• watch
• logs-tail [--file <path>] [--follow]
• net-summary [--file <path>]
• wait-network-idle [--timeout <ms>] [--idle-time <ms>] [--max-inflight <count>]

Response expectations

• Include exact paths for any generated artifacts.
• Include the failing command and concise stderr when a command errors.
• Keep extraction steps deterministic and reproducible.

Troubleshooting

• If connection fails, verify http://localhost:9222/json/version.
• If browser detection fails, run start --browser <path-or-name>.
• Use --fresh for a clean session.
• Use --copy-profile [name] for logged-in state.
• Use CDP_BROWSER_BASE_DIR to control managed data paths.