360 Secure SkillHub
360 安全技能中心

fly-coworker

该 Skill 先收集 2–64 个候选任务,再调用 OpenClaw 配套插件工具创建一轮选择,并在 Chrome 中打开仪表盘的 fly-picker 组件;用户观看果蝇或射击后点击 Accept & start 确认,选定任务会被派发回同一会话开始执行,还可用状态工具查询是否已受理。

grahammcbain
grahammcbain
热度 18.48
最近更新 2026-09-17 05:35:54
分类:自动化依赖:联网浏览搜索 / MCP 服务应用场景:工作流 / 任务管理

安装技能

复制下面的话术发给你的 Agent,它会通过 secure-skills 检索并安装该技能。

执行命令安装此 skill: npx secure-skills-cli add fly-coworker
Metadata
name
fly-coworker
description
Let the fly decide between tasks or PRs in Chrome; watch, veto, then Accept & start the chosen work in OpenClaw.
metadata
{"openclaw":{"emoji":"🪰","requires":{"bins":["openclaw"]}}}

Fly Coworker

Browser-first workflow

  1. Collect 2–64 distinct, actionable tasks of 1–300 characters. Preserve PR URLs, repositories and constraints. Inspect the conversation for missing details; ask for the task list when absent. Identify the requesting session with dashboard read or session_status. Complete when every option identifies work the agent can begin and the owning session is known.

  2. Discover fly_coworker_create and fly_coworker_status. If unavailable, check openclaw plugins info fly-coworker and consult plugin-install.md for the companion. Verify the plugin is loaded and Custom plugin UI is enabled. Reuse the installed companion; changing from the Mac app to Chrome needs no Gateway restart or auth changes.

  3. Call fly_coworker_create with {tasks:[...]} in the requesting conversation. If the tool catalog is stale, call its registered session action with openclaw gateway call plugins.sessionAction --json --params '<JSON>', using {pluginId:"fly-coworker",actionId:"create",sessionKey:<current key>,agentId:<current agent>,payload:{tasks:[...]}}. Pass task text as structured arguments or JSON through a file/script, not shell interpolation. Verify the returned round ID and session key.

  4. Read the current dashboard, reuse or create tab fly, and widget_put with name fly-picker, kind fly-coworker:picker, props:{roundId}, size full. Open the Gateway's Control UI in Google Chrome, using the private browserUrl returned by openclaw dashboard --json as described in references/browser-launch.md. Navigate that browser to the owning session, focus the Fly tab, and expand the dashboard. This is the browser-hosted OpenClaw dashboard, not the native Mac window or the standalone viewer on port 4318. Verify the anatomical fly, brain panel, and Accept & start control render; report an actual error if they do not.

  5. Leave the browser open for the human to watch the fly and click Accept & start, or Salt blaster, aim, and click the fly. A hit excludes the choice and respawns; a miss leaves it alive. The fullscreen button gives an arena-only view for recording. Never accept real work on the human's behalf. Yield once the browser is ready; no result polling or background watcher is needed.

  6. Acceptance submits the exact chosen task into this same OpenClaw conversation using authenticated chat.send and a durable one-time dispatch key. When that message arrives, start the chosen task using the conversation's existing context and ordinary permissions. Do not start other options. A browser opened on the Gateway host still targets the requesting conversation, not a new chat or whichever chat happens to be visible elsewhere.

  7. For status, call fly_coworker_status with the round ID (or its status session action). chosen awaits acceptance; accepted with dispatch:started proves a run receipt, not completed work. Check the conversation before any manual retry after an uncertain handoff. Reloading an accepted round must not send again. Verify completion from the task's output, separately from the handoff.

Boundaries

The companion plugin is still required: Chrome loads its renderer and uses its existing secure handoff. The older standalone viewer's Accept pick only saves a local result and is not this workflow. Use Chrome for local HTTP on this host; the native Mac WebKit cookie issue is deferred, not fixed. Do not modify OpenClaw core or weaken authentication for this browser flow.

The fly uses FlyGym/NeuroMechFly anatomy and a 1,045-neuron MaleCNS circuit, not full Flyhard/Eon emulation or MuJoCo physics. Do not claim task comprehension or uniform randomness.