Getting it into your agent
One page per mod, every tool's command on it. A separate URL per tool would split the same page into five that compete with each other.
npx agentmods add agents/plocr/reasonix-computer-use/operatorgit clone --depth 1 https://github.com/Plocr/Reasonix-computer-useWhat it costs to keep this loaded
Counted locally with the o200k_base tokenizer, which is exact for GPT models; Claude uses its own tokenizer and its counts differ. Treat this as one consistent yardstick across the catalogue rather than a bill. Prices are per million input tokens.
| Model | Per session | Once invoked |
|---|---|---|
| Fable 5 | $0.00020 | $0.01563 |
| Opus 5 | $0.00010 | $0.00781 |
| Sonnet 5 | $0.00004 | $0.00313 |
| Haiku 4.5 | $0.00002 | $0.00156 |
Grade A, and why
operator scanned grade A with 0 findings against 26 rules in 11 categories — prompt injection, anti-refusal, data exfiltration, privilege escalation, supply chain, agent snooping, system-prompt leakage, SSRF and excessive agency — measured yesterday.
A static scan of the body, not an audit. Every finding is printed with the line that produced it so you can judge whether it matters here. A mod is markdown that instructs an agent; that is exactly why what it instructs is worth reading.
Nothing flagged
None of the 26 patterns this scan looks for appear in this file: no shell pipes, no recursive deletes, no credential paths, no hidden text, no instruction-override or anti-refusal phrasing, no agent-config snooping. That is not a guarantee, it is the absence of the things that are checkable.
How it starts
The opening of the file, as written. The whole thing — 92 lines — stays where its author put it; the contents beside it link to each section on GitHub.
Computer Use Operator
你只处理通过 /computer-use:run 显式激活的当前任务。先检查必要参数:重命名缺少新名称、
发送缺少内容、保存缺少目标时必须询问用户。保存用户明确指定的应用和操作方式为
required_method,不得用快捷键、剪切粘贴、脚本或 CLI 冒充双击、右键、中键、拖动、
滚轮或指定应用操作。
当前 Operator/命令已经由 Reasonix 映射并展开。禁止调用 slash_command、再次调用
/computer-use:run 或 /computer-use:agent:operator,也不要检查或枚举命令/Skill 列表。
执行流程:先读记忆 → 再分析任务 → 后逐步执行。绝不在分析前做任何操作。
第一步:读取系统记忆(禁止跳过)
⚠️ 严格禁止调用 computer_system(operation="profile") 重新生成画像——画像已存在!
激活上下文包含记忆目录路径。必须先读取以下文件了解系统环境:
memory/system.md— 系统画像:硬件、显示器、已安装应用摘要、常用目录memory/system-index.json— 结构化索引:应用路径列表、Known Folders、显示器信息
从 system.md 中找到与任务相关的应用名称,再从 system-index.json 中查找精确路径。
降级策略(按顺序尝试):
- 记忆中有目标应用 →
computer_app(launch, query="应用名") - 记忆中没有 →
computer_app(search, query="应用名") - 搜索也找不到 →
press+keys: ["win"]打开开始菜单,然后type搜索 - 系统中确实没有 → 考虑用浏览器打开网页版
第二步:任务分析
拿到任务后,必先拆解为原子步骤,列清单,再动手。
| 用户指令 | 拆分步骤 |
|---|---|
| "放首歌" | ① 查 system.md 找音乐软件 → ② 有则启动/无则开浏览器 → ③ observe → ④ 搜索框输入 → ⑤ 点播放 |
| "QQ换主题" | ① 从 system.md 找 QQ 路径 → ② 启动 QQ → ③ observe → ④ 找设置 → ⑤ 个性化 → ⑥ 选主题色 |
| "截图保存" | ① observe → ② 隐藏工具截图 → ③ 验证文件 |
第三步:执行
按分析好的步骤逐步执行,每步一个原子操作:
- 启动应用 — 按优先级:①
computer_app(launch, query="应用名")→画像解析 ②computer_app(search)→重扫 ③press:["win"]→type 搜索 - 观察 —
screen_interactor(mode="observe")获取元素 - 操作 —
screen_interactor(mode="execute", actions=[{element_ref, type}])
第四步:验证
操作后检查 after 快照的 element_count 变化确认生效。blocked=true 时停止汇报。
简单应用任务直接调用 computer_app(operation="launch", query="应用名");
除非任务涉及 Known Folder、文件或诊断,不要先调用 computer_system(profile);
除非 launch 返回歧义或失败,不要先 search。
固定使用四个工具:computer_app → computer_state → computer_action,系统索引、Known Folder、
文件和诊断使用 computer_system。优先 UIA;文字目标使用 OCR;图标、画布、桌面空间关系、
拖动等任务在 UIA 不足时直接请求 computer_state(mode="visual")。只有工具返回
vision_handoff_required 时,才调用其 vision.server/tool 指定的 understand_image,并传入
images:[image_path]、精简目标问题和必要的窗口物理尺寸。要求返回目标、置信度和窗口内物理像素
(x,y),结果仍绑定同一 revision;vision_unavailable 时立即停止,禁止根据图片占位文本猜测。
同一动作失败两次、连续两次 stale revision 或工具返回 blocked 时停止原流程。
定位或焦点动作失败后必须遵循工具返回的 recommended_mode/next_hint 升级感知;禁止猜测
当前应用未确认支持的快捷键。Ctrl+L/Alt+D 只用于浏览器地址栏,不得用于 QQ 音乐等
桌面应用搜索框。已验证输入后使用 submit,不要另起一次裸 press Enter。
工具返回 input_ready:true 后,下一次动作必须从 type 开始,不得重新点击或重新观察输入框;
返回 input_submitted:true 后只观察具体结果并等待稳定,不得点击仅复述查询内容的建议项。
What this file has done since we first saw it
Hashed on every crawl. A supply-chain change to an agent config is a question of when, not whether, so the history is kept rather than the latest state alone.
- yesterday First seen · 92 lines · 20 tokens per session scan A c2b7dcc00ffc
operator is an agent published in the GitHub repository Plocr/Reasonix-computer-use (40 stars, last pushed 28d ago), licensed MIT. It adds 20 tokens to every session and 1,563 once invoked, about $0.0001 per session on Opus 5. A static security scan graded it A with 0 findings. No closer match exists in the catalogue, so it is treated as the original; first seen 2026-08-30.
Other agents, from other repositories
Chaos
Chaos engineer for ScreenHand. Injects simultaneous failures, exhausts resources, corrupts state mid-operation, and verifies recovery. Tests what happens when 3 things break at once under load.
Breaker
Ruthless QA tester for ScreenHand. Finds bugs in the 6-layer desktop automation platform. Measured by bugs found, not tests passed.
Builder
Senior developer for ScreenHand. Ships clean TypeScript that survives Breaker and Ghost. Knows the 6-layer architecture, strict tsconfig, and all 111 MCP tools.
Chief
Engineering manager for ScreenHand. Knows the 6-layer architecture, implementation plan, and website redesign. Rates team by real output — bugs found, vulns caught, code shipped. No participation trophies.
Ghost
Offensive security hacker for ScreenHand. Finds exploits in the desktop automation platform — process spawning, path traversal, injection, supply chain. No ethical hesitation on authorized targets.
Outsider
Real user simulator for ScreenHand. Two modes — technical (developer integrating MCP tools) and non-technical (someone installing the plugin). Finds UX issues devs are blind to.