跳至正文
OpenAgentSkill
浏览已开放 · 目录记录日期 2026-11-21 · 自动更新和账户服务暂时关闭。
浏览器与自动化

hive.browser-automation

Required before any hive-browser CLI command. The browser is driven from the terminal by running `hive-browser <command> ... --json` via terminal_exec — not via MCP tools. Teaches the browser lifecycle rules (the bridge attaches to the USER'S running Chrome — never kill or launch browser processes; timeouts are transport issues, not crashes), the screenshot + coordinate workflow (hive-browser interact with a fractional coordinate) that reaches shadow-DOM inputs selectors can't see, the viewport-fraction coordinate rule (not pixels), rich-text editor quirks ("send button stays disabled" failures), and CSP gotchas. Covers Chrome via CDP through the GCU Beeline extension.

aden-hive · 已有 AI 审核记录

安装与使用

运行命令前先阅读来源与依赖说明。复制命令不会自动安装或执行。

npx skills add aden-hive/hive --skill hive.browser-automation

技能说明


name: hive.browser-automation description: Required before any hive-browser CLI command. The browser is driven from the terminal by running hive-browser <command> ... --json via terminal_exec — not via MCP tools. Teaches the browser lifecycle rules (the bridge attaches to the USER'S running Chrome — never kill or launch browser processes; timeouts are transport issues, not crashes), the screenshot + coordinate workflow (hive-browser interact with a fractional coordinate) that reaches shadow-DOM inputs selectors can't see, the viewport-fraction coordinate rule (not pixels), rich-text editor quirks ("send button stays disabled" failures), and CSP gotchas. Covers Chrome via CDP through the GCU Beeline extension. metadata: author: hive type: default-skill version: "2.2"

GCU Browser Automation

All GCU browser automation drives a real Chrome instance through the Beeline extension and Chrome DevTools Protocol (CDP). You drive it from the terminal: every command is hive-browser <command> ... --json, run through terminal_exec. Always pass --json so the result is machine-readable. That means clicks, keystrokes, and screenshots are processed by the actual browser's native hit testing, focus, and layout engines — not a synthetic event layer. Understanding this unlocks strategies that make hard sites easy.

Browser lifecycle & recovery — read this before "fixing" anything

The bridge attaches to the user's already-running Chrome via the extension. The browser is not yours: it holds the user's logged-in sessions and other agents' work, and the runtime — not you — owns the connection to it. Consequences:

  • There is nothing to "start." If no browser is connected, hive-browser setup --json tells you and gives the user install steps. Launching Chrome yourself (any google-chrome/chromium command, --remote-debugging-port, --user-data-dir, headless flags) is forbidden and blocked — it opens a browser the bridge can't see, often under the wrong profile.
  • There is no situation where killing the browser helps. Never use terminal tools to kill/pkill/killall Chrome, the bridge, or gcu processes. These commands are blocked, and attempting them breaks every agent sharing the connection. The same applies to the Hive desktop app and bridge_host.
  • A timeout is NOT a stuck browser. All hive-browser commands share one transport across all agents; one slow call (a heavy hive-browser evaluate on a big page) can make YOUR calls time out while the browser is perfectly healthy. The timeout message tells you whether the server passed its liveness check and whether recovery is already running — believe it.
  • Allowed recovery, in order: (1) wait ~30s and retry ONCE, with a smaller/simpler request — e.g. split a page-wide hive-browser evaluate sweep into chunked queries; (2) close YOUR OWN tabs with hive-browser tab close <T> --json and reopen; (3) report the failure (report_to_parent or your reporting channel) and move on to work that doesn't need the browser. Escalate to the user; never to the process table.
  • Keep heavy hive-browser evaluate scripts cheap: avoid innerText over thousands of nodes (each read forces layout). Prefer textContent, scope the selector, and paginate the sweep.

Targeting a specific Chrome profile / account

If your machine has more than one Chrome profile connected (different logged-in accounts), say which one to act in — if you omit it, the bridge falls back to the first-connected profile, which may be the wrong account. Pass --browser-profile <label> to hive-browser open (and hive-browser navigate / hive-browser script):

  • See every connected profile in hive-browser status --json / hive-browser setup --json — both return a connected_profiles list of {label, is_default, starred} for ALL connected Chrome profiles (not just the one you're using). That's how you discover the labels. (hive-browser status reflects ALL connections; don't conclude "only one profile" from the single profile label of your own context.)
  • The label is a connected profile label — the name shown in that profile's Hive extension side panel (or its auto 3-word id). If your task assigns you a profile (e.g. "your profile is acct-jpn"), pass exactly that.
  • Every command's JSON echoes the profile it actually used. Check it: if it doesn't match what you intended, you opened the wrong account. Stop and fix the label (don't proceed).
  • A label that isn't connected fails fast with the list of connected labels — bind to one of those.
  • With one profile connected (or one starred default), you can omit it. With several connected and no star, omitting it uses the first-connected profile — fine for single-account work, risky for multi-account, so pass the label when the account matters.

Working on LinkedIn?

For ANY LinkedIn flow, load hive.linkedin-core first — it owns the auth check, rate limits, stop protocol, reply-text policy, and DOM gotchas every LinkedIn script depends on. Then load the capability skill for the task: hive.linkedin-discovery (scans / People search), hive.linkedin-messaging (lk_send_to_message_url, reply, inbox), hive.linkedin-connect (lk_send_invite, post comments), or hive.linkedin-sales-navigator (premium search + InMail).

Coordinates

Every hive-browser interact action that takes a --coordinate — and every command that returns one — operates in fractions of the viewport (0..1 for both axes). Read a target's proportional position off hive-browser screenshot — "this button is about 35% from the left and 20% from the top" → pass --coordinate 0.35,0.20. Rect-returning commands (hive-browser page shadow-query and the rect inside focused_element) also return fractions. The CLI converts to CSS pixels internally before dispatching to Chrome.

hive-browser screenshot --json                                      → image + cssWidth/cssHeight in
阅读完整来源 ↗