React Orchestrator

基于 ReAct 框架的双系统 AI 协调器，自动评估任务复杂度，智能切换快速执行和深度推理模式，支持多工具协作。

This skill appears to implement the advertised ReAct orchestration and includes helpful features (HITL, Code Mode, tool registry). However: - The package metadata lists no required env vars, but the code and examples reference TAVILY_API_KEY and call external APIs; treat these as optional integrations but be explicit about what you set. Do not expose sensitive env vars unless you trust the skill. - The Code Mode will write temp files and spawn child processes (node and PowerShell). That means it can read/write filesystem paths and execute arbitrary code — run in an isolated environment (container/VM) if you are unsure. - Built-in templates include file-read and file-write operations. Enable and configure HITL (requireApproval for file-write / execute-command) before letting the orchestrator act on your behalf. - package.json is minimal (only zod). The code references modules (e.g., 'tavily-search') not declared as dependencies; inspect and install required third‑party libs yourself from trusted sources. - If you plan to use networked features or the A2A functionality later, audit any networking endpoints and consider firewalling the runtime or limiting outbound access. If you want to proceed, run it in a sandbox, enable HITL approvals for dangerous operations, and avoid supplying real secrets (API keys, cloud credentials) until you have audited templates and tool implementations. If anything is unclear, ask the author to add explicit metadata listing required env vars, external endpoints, and a dependency list.

Type: OpenClaw Skill Name: yuyonghao-react-orchestrator Version: 0.1.0 The bundle implements a 'Code Mode' feature in `src/code-mode.js` that converts tool calls into executable JavaScript or PowerShell code, which is then run locally using `child_process.spawn`. This creates a significant Remote Code Execution (RCE) surface, especially as it includes built-in templates for file system operations and executes PowerShell with `-ExecutionPolicy Bypass`. While this is presented as a token-optimization feature and lacks explicit evidence of malicious intent (such as hardcoded exfiltration or backdoors), the capability to generate and execute arbitrary code based on LLM-provided parameters is inherently high-risk.