Multi Agent Collaboration

Claude Grade 多智能体协作技能。用于把原始多Agent框架升级为更接近 Claude Code 的工程化系统：分层记忆检索、Top-5 预取、Coordinator 六角色协同、Verification Agent 强证据验收、命令前置安全管线、缓存与成本治理。适用于需要“多Agent不空转、记忆不瞎...

Install only if you are comfortable with local persistent memory and personal/workflow profiling. Review and restrict the memory directory, avoid entering secrets or sensitive personal details, and prefer a sandboxed/test namespace until the publisher clearly documents retention, deletion, consent, and which modules are actually in scope.

Type: OpenClaw Skill Name: multi-agent-collaboration Version: 2.0.0 The skill bundle provides a comprehensive framework for multi-agent collaboration, featuring a structured five-level memory system (L0–L4), a command safety auditor, and a verification agent. The core logic, implemented in files like `dist/core/memory.js` and `dist/systems/safety.js`, focuses on categorizing information, tracking goals, and orchestrating agent roles (coordinator, explorer, verifier, etc.). The `SafetyGatePipeline` specifically includes defensive regex-based guards to detect and block dangerous operations such as recursive deletions (`rm -rf`), privilege escalation (`sudo`), and remote script execution (`curl | bash`). No evidence of data exfiltration, unauthorized persistence, or malicious intent was found; the code and documentation are entirely consistent with the stated purpose of an engineering-grade agent orchestration system.