← 返回 Skills 市场
yuyonghao-123

React Orchestrator

作者 yuyonghao-123 · GitHub ↗ · v0.1.0 · MIT-0
cross-platform ⚠ suspicious
112
总下载
0
收藏
0
当前安装
1
版本数
在 OpenClaw 中安装
/install react-orchestrator
功能描述
基于 ReAct 框架的双系统 AI 代理协调器,自动评估任务复杂度并在快速执行与深度推理间智能切换,实现高效分层推理。
安全使用建议
High-level things to check before installing or running: - Missing declarations: the package metadata lists no required env vars or binaries, but the code and examples use process.env.TAVILY_API_KEY and expect 'tavily-search'/'rag-retriever' style modules. Confirm what environment variables and extra packages you must provide. - Code execution risk: CodeMode writes temporary files and spawns child processes (node and PowerShell with ExecutionPolicy Bypass). If you run this on a host with sensitive files or with network access, generated code could read/write files or execute commands. Prefer running in an isolated container/VM and review/disable CodeMode if you don't trust generated code. - File I/O and tools: Built-in templates include file-read and file-write that perform arbitrary path access. Review which tools are registered by default and enable HITL for write/execute tools (HITLManager default requireApproval includes file-write/execute-command in the code; keep those enabled). - External endpoints and secrets: The tavily-search template uses an API key and calls api.tavily.com. If you supply such keys, verify the endpoint is expected and trustworthy. Avoid putting cloud credentials or high-privilege secrets into the environment unless necessary. - Dependencies: package.json only lists 'zod'. Examples/templates reference other packages and clients not in dependencies. Before npm install/run, inspect code for undeclared requires and add/lock dependencies from trusted registries. - Platform assumptions: PowerShell templates and invocation of 'powershell.exe' are Windows-specific; the code expects Node >=18. If running on non-Windows, PowerShell paths and behavior differ. If you want to proceed safely: - Run the skill inside an isolated container with no sensitive mounts and minimal network access. - Enable HITL for anything that writes files or executes commands; set autoApprove to an empty list. - Audit all tool templates and remove or replace any that call external services or perform filesystem access you don't want. - Add explicit required env var declarations (TAVILY_API_KEY etc.) to your deployment manifest so you know what secrets will be exposed. Given the mismatches between declared metadata and actual code behavior, treat this package as potentially dangerous until you review and constrain its runtime capabilities.
功能分析
Type: OpenClaw Skill Name: react-orchestrator Version: 0.1.0 The bundle implements a 'ReAct Orchestrator' framework that includes a high-risk 'Code Mode' feature in src/code-mode.js. This module dynamically generates and executes JavaScript and PowerShell scripts using child_process.spawn based on tool parameters, which creates a significant Remote Code Execution (RCE) surface if the driving LLM is manipulated via prompt injection. While the bundle includes a Human-in-the-loop (HITL) safety mechanism in src/hitl.js to mitigate unauthorized actions, the inherent capability to execute unvalidated code generated from AI prompts is a major security risk. No evidence of intentional malice or data exfiltration was found, but the architectural design is highly vulnerable.
能力评估
Purpose & Capability
The implementation (orchestrator, LLM integration, HITL, Code Mode, tool registry) matches the described ReAct orchestrator purpose. However the code references capabilities and external services not declared in the registry metadata: templates and examples reference TAVILY_API_KEY, tavily-search and rag-retriever integrations, and the CodeMode expects node/powershell execution support. The skill metadata declares no required environment variables or binaries despite clearly depending on them, which is an incoherence (likely sloppy packaging).
Instruction Scope
SKILL.md and other docs instruct the agent/user to register tools that may perform network calls and file I/O. The code includes templates that read/write arbitrary filesystem paths and spawn child processes to execute generated code (Node and PowerShell). SKILL.md does not declare that the orchestrator will execute generated code or require access to environment secrets like TAVILY_API_KEY, but the runtime templates and usage examples do. The agent instructions therefore have broader scope than the declared metadata and permit operations (arbitrary code execution, file writes, PowerShell ExecutionPolicy bypass, outbound network calls) that the registry did not surface.
Install Mechanism
There is no install spec in the registry (instruction-only), and the repository is a normal Node project (package.json). No remote download/install URLs or extract steps are present in registry metadata. That said, the package references external packages (e.g., 'tavily-search', 'rag-retriever') in examples and templates that are not listed in package.json, so installing via npm install may not provide all referenced code. The installation mechanism itself is low-risk, but missing dependency declarations are an operational coherence issue.
Credentials
The registry claims no required env vars, yet code templates and examples expect process.env.TAVILY_API_KEY and use Authorization headers for external APIs. CodeMode templates also call external endpoints (https://api.tavily.com) and spawn processes with full process.env forwarded. The skill therefore requires secrets / env access not declared in metadata. In addition, CodeMode uses filesystem access and spawns child processes (node, powershell.exe) which are powerful privileges relative to a library that could have been designed as purely in-process orchestration.
Persistence & Privilege
The skill does not request always:true and does not declare system-wide configuration changes. However it provides a Code Mode that writes temporary files and executes them (Node/Powershell), and a HITL flow that may require integrating UI/webhooks. Autonomous invocation is allowed by default (disable-model-invocation=false). Combined with the ability to execute arbitrary code and call external services, autonomous invocation increases blast radius — this is not flagged alone by policy but should inform user caution and configuration choices (e.g., enable HITL, restrict autoApprove).
如何使用
  1. 确保已安装 OpenClaw(本地或 Docker 部署)
  2. 在对话框中输入安装命令:/install react-orchestrator
  3. 安装完成后,直接呼叫该 Skill 的名称或使用 /react-orchestrator 触发
  4. 根据 Skill 的参数说明提供必要输入,即可获得结构化输出
版本历史
v0.1.0
react-orchestrator v0.1.0 - 首个发布,提供基于 ReAct+Reflexion 的双系统 AI 代理协调器 - 支持 System 1 (快速直觉) 和 System 2 (深度反思) 自动切换 - 内置工具注册中心、自动复杂度评估、超时保护、执行历史记录 - 实现 Reason → Act → Observe → Repeat + 定期自我反思机制 - 提供示例代码、基础 API、配置项说明及测试覆盖 - 已知限制包括 LLM 集成和参数提取需后续完善
元数据
Slug react-orchestrator
版本 0.1.0
许可证 MIT-0
累计安装 0
当前安装数 0
历史版本数 1
常见问题

React Orchestrator 是什么?

基于 ReAct 框架的双系统 AI 代理协调器,自动评估任务复杂度并在快速执行与深度推理间智能切换,实现高效分层推理。 它是一个面向 Claude Code / OpenClaw 的 AI Agent Skill 插件,目前累计下载 112 次。

如何安装 React Orchestrator?

在 OpenClaw 或 Claude Code 对话框中运行命令「/install react-orchestrator」即可一键安装,无需额外配置。

React Orchestrator 是免费的吗?

是的,React Orchestrator 完全免费,采用 MIT-0 许可证,可自由下载、安装和使用。

React Orchestrator 支持哪些平台?

React Orchestrator 跨平台运行,可在任意部署了 OpenClaw / Claude Code 的环境中使用(cross-platform)。

谁开发了 React Orchestrator?

由 yuyonghao-123(@yuyonghao-123)开发并维护,当前版本 v0.1.0。

💬 留言讨论