菜单
Drive the user's real browser (Chromium/Brave/Chrome) — open url, web search, read page text, click, type, screenshot, list/switch tabs. Real backend = Playwright connect_over_cdp (attaches to your existing browser); mock for the binding benchmark. Pure execution (slot-filled, no LLM).
用 Codex 或 Claude 帮你安装 复制这段 Prompt，粘贴到 Codex、Claude 或其他助手里，让它检查 Skill 页面并帮你完成安装。
基于 SOC 职业分类
Application control — open/launch, close/quit, and switch-to a running app by name. Mock backend. Pure execution.
Telephony actions — dial a contact/number, redial, hang up, answer. Mock backend (no real telephony). Pure execution.
Unified time control — countdown timers, absolute-time alarms (with optional repeat), and current/world time. The one skill that ACTUALLY fires (systemd-run --user) when ALFRED_CLOCK_REAL=1; mock for the binding benchmark. Pure execution.
Connectivity + UI toggles — wifi, bluetooth, airplane mode, do-not-disturb, dark mode, each on/off. Mock backend. Pure execution.
Manage Docker containers — start/stop/restart, remove, show logs, list running containers. Pure execution, no communication. Mock backend (binding-only) for the thesis benchmark.
Local media processing with ffmpeg — extract audio to MP3, transcode formats, trim segments, capture frames, resize, and concatenate. Real ffmpeg binary; deterministic, offline, no network. Binding-only in the thesis benchmark.
| name | browser |
| description | Drive the user's real browser (Chromium/Brave/Chrome) — open url, web search, read page text, click, type, screenshot, list/switch tabs. Real backend = Playwright connect_over_cdp (attaches to your existing browser); mock for the binding benchmark. Pure execution (slot-filled, no LLM). |
Deterministic control of the user's own browser. Canonical reflexer case — each request is one slot-filled command, no autonomous reasoning.
Backends. Default MOCK echoes JSON (binding-only). The REAL backend uses Playwright
connect_over_cdpto attach to a running Chromium/Brave/Chrome started with--remote-debugging-port=9222(preserves your logins/tabs). On Raspberry Pi / ARM, attach to the aptchromium-browser(never Playwright's bundled binary); raw CDP (pychrome) is the lightweight fallback. Seedocs/research/browser_skill_options.md.
CLI: python3 agents/skills/browser/browser.py <command> (aliased browser).
browser open --url "<url>"
browser search --query "<text>" [--engine google]
browser read-text [--selector "<css>"]
browser click --target "<css selector or visible text>"
browser type --selector "<css>" --text "<text>"
browser screenshot [--path <file>] [--full-page]
browser tabs
browser tab-switch --id <id>
One JSON line: {"ok":bool,"data":{...},"error":null|{...},"meta":{...}}. Exit 0/2.
On an ambiguous click target the real backend returns error.code = "ambiguous"
rather than guessing.
No smart-waits, no "do this whole task" verbs, no arbitrary JS-eval, no credential handling — those would require a thinker. The browser is just one more OS surface the reflexer fills slots against.
browser open --url vs app open --name (URL vs app); browser search vs maps search
(web vs places) vs spotify search (web vs music); browser tab-switch vs app switch.
