Description

亚马逊外贸B2B多CMS Agent协作系统。支持Shopify/WooCommerce/Magento三大平台，A2A架构协调库存/采购/财务/物流四大专家Agent，三层安全网保障审批合规。即装即用，零配置开箱。

README (SKILL.md)

产业互联网硅基军团 v2.0

Name: Agent Cluster
Author: wangm-a3

企业级Multi-Agent智能体集群系统，基于1+N架构（1个幕僚长+20个专业Agent），参考OpenClaw Main Agent、腾讯ADP Router设计。v2.0全面升级：真实ERP API接入、协作流程细化、错误处理体系。

核心能力

v2.0三大升级

1. 真实API接入层（api_integration/）

多ERP适配器：SAP S/4HANA、用友U8/NC/YonBIP、金蝶K3 Cloud/EAS、通用REST
断路器模式（Circuit Breaker）+ 故障自动降级
健康检查轮询（10s间隔，自动摘除异常节点）
演示模式保留（MockDataGenerator，variance=0.1随机波动）

2. 跨Agent协作细化（collaboration/）

细粒度任务协议（TaskMessage）：依赖声明、优先级、TTL
状态同步（SharedStateManager）：TTL+pub/sub通知机制
全链路追踪（CollaborationTracker）：trace_id/span_id + Mermaid时序图可视化

3. 错误处理与状态管理（error_handling/）

7状态任务状态机：pending→running→success/failed/retry/timeout/cancelled
10类异常自动分类：VALIDATION/NETWORK/TIMEOUT/AUTH/RESOURCE/NOT_FOUND等
5种重试策略：FIXED/EXPONENTIAL/FIBONACCI/JITTER/ADAPTIVE
敏感信息脱敏 + SOC2合规审计日志

系统架构

用户请求 → Orchestrator（意图识别→任务拆解→智能体调度）
    ↓
20个专业Agent：采购/生产/销售/财务/运营/战略/研发/人力/合规
    ↓
API Integration Layer（v2.0新增）
  ├─ SAP/用友/金蝶适配器（真实ERP）
  └─ 断路器+健康检查+Mock降级

目录结构

agent-cluster/
├── orchestrator.py              # 指挥智能体（核心调度器）
├── api_integration/              # v2.0新增：真实API接入层
│   ├── api_adapter.py           # 多ERP适配器（SAP/用友/金蝶）
│   ├── api_config.py            # 配置化管理
│   ├── api_health.py            # 健康检查+断路器
│   └── mock_data.py             # 模拟数据（开发/演示）
├── collaboration/                # v2.0新增：跨Agent协作
│   ├── task_protocol.py         # 细粒度任务协议
│   ├── state_sync.py            # 状态同步+TTL+pub/sub
│   ├── trace_tracker.py         # 全链路追踪+Mermaid
│   └── workflow_engine.py       # 混合执行引擎
├── error_handling/              # v2.0新增：错误处理
│   ├── task_state_machine.py    # 7状态任务状态机
│   ├── exception_middleware.py  # 统一异常处理
│   ├── retry_policy.py          # 5种重试策略
│   └── operation_log.py         # 操作日志+脱敏+合规
├── specialists/                 # 专业智能体
│   ├── inventory_agent.py      # 库存智能体
│   ├── logistics_agent.py       # 物流智能体
│   ├── procurement_agent.py     # 采购智能体
│   ├── finance_agent.py         # 财务智能体
│   └── doc_agent.py            # 工艺文档智能体
├── mcp_servers/                 # MCP协议封装
│   ├── erp_server.py           # ERP系统接口
│   ├── wms_server.py           # WMS仓库管理接口
│   └── srm_server.py           # SRM供应商管理接口
├── safety/                      # 安全围栏
│   ├── permission_manager.py    # RBAC权限管理
│   ├── audit_logger.py          # 全链路审计日志
│   └── human_loop.py            # 人机回环审批
└── config/                      # 配置文件
    ├── agents.yaml             # 智能体定义
    ├── workflows.yaml          # 工作流配置
    └── permissions.yaml        # 权限矩阵

快速开始

环境要求

Python 3.10+
依赖：pip install pyyaml fastapi uvicorn httpx aiofiles

运行

cd agent-cluster
python orchestrator.py

配置（生产模式）

export SYSTEM_MODE=production
export SAP_BASE_URL=https://sap.example.com
export SAP_API_KEY=sk-xxx
export YONYOU_BASE_URL=https://yonyou.example.com

触发词

分类

效率工具

版本

v2.0.0 - 三大升级：真实API接入层、协作流程细化、错误处理与状态管理

Usage Guidance

What to consider before installing/running this skill: - Provenance: the repository homepage is provided but the registry metadata showed 'source: unknown'. Confirm you trust the author and that the GitHub repo at the homepage is genuine and unchanged. - Undeclared secrets: SKILL.md/README instruct you to set ERP API keys and other env vars (SAP_*, YONYOU_*, KINGDEE_*, SYSTEM_MODE) and the code references model engines (claude-ma, deepseek). However the skill metadata did not declare these as required. Expect to supply sensitive credentials if you run it; do not put production credentials into an untrusted environment. - Install/run in isolation: because the package includes many Python modules and persistent stores, run it in a throwaway VM/container or CI runner with least privilege. Avoid running on a machine that contains other credentials you care about. - Review network endpoints: inspect connector code (api_integration/, mcp_servers/, connectors/) for all outbound endpoints and where API keys are sent. If you need ERP integration, consider using test/demo ERP endpoints first. - Prompt-injection pattern: the SKILL.md contains structured/system-like content that the scanner flagged. Before granting the agent autonomous invocation or broad network access, review those instruction sections and remove/neutralize anything that tries to override system prompts or inject untrusted instructions. - Credential handling: prefer short‑lived/test credentials or proxy endpoints. If you must provide real ERP credentials, create a dedicated, scoped account with minimal privileges and rotate keys after testing. - Ask the maintainer for clarity: request an explicit manifest listing required environment variables and any external services/credentials (including LLM engines), and ask for an install script or container image. A clear SECURITY.md and reproducible build/install steps would reduce risk. - If you lack ability to audit: consider not installing on production systems. If you want to try it, use 'SYSTEM_MODE=demo' and dummy/test ERP endpoints to validate behavior first.

Capability Analysis

Type: OpenClaw Skill Name: agent-cluster Version: 1.0.0 The skill bundle implements a highly sophisticated and well-structured multi-agent orchestration framework for e-commerce (Shopify, Amazon SP-API) and ERP (SAP, Yonyou) integration. It features a robust safety architecture, including a dedicated permission manager for RBAC (permission_manager.py), an audit logger with PII masking (audit_logger.py), and a human-in-the-loop approval system for high-risk operations (human_loop.py). The execution layer supports intelligent routing between multiple LLM engines with built-in circuit breakers and health monitoring (circuit_breaker.py). No evidence of malicious intent, data exfiltration, or unauthorized execution was found; the system's broad capabilities are aligned with its stated purpose as an enterprise-grade 'Chief of Staff' agent.

Capability Tags

cryptorequires-walletcan-make-purchasescan-sign-transactionsrequires-oauth-tokenrequires-sensitive-credentials

Capability Assessment

✓ Purpose & Capability

Name, description, and the included code (connectors for SAP/Shopify/Magento, orchestrator, agents for inventory/finance/logistics, API integration layer) are coherent: the repository implements an enterprise multi‑agent e‑commerce/ERP integrator as claimed.

⚠ Instruction Scope

SKILL.md tells an agent to run orchestrator.py and to set environment variables for ERP endpoints/API keys (SAP_BASE_URL, SAP_API_KEY, YONYOU_BASE_URL, etc.). The SKILL.md also contains a YAML/system-like header and the static scan flagged a 'system-prompt-override' pattern — this suggests the skill's runtime instructions include embedded agent/system directives which could attempt to affect agent behavior. The instructions reference external APIs and credentials (reasonable for ERP integration) but give the agent broad autonomy to contact those endpoints and use environment variables that were not declared in the registry metadata.

⚠ Install Mechanism

Registry metadata states 'No install spec — instruction-only', yet the package contains >100 code files and a README that instructs pip installing dependencies and running python orchestrator.py. There is no formal install script in the registry metadata; the user/platform would need to install Python packages and run code manually. This mismatch (instruction-only claim vs. large codebase) is an incoherence and increases operational risk because files will be executed without an automated vetted install step.

⚠ Credentials

The SKILL.md and README expect multiple environment variables (ERP base URLs and API keys, SYSTEM_MODE, possibly model engine credentials for 'claude-ma'/'deepseek'), but the skill manifest metadata lists no required env vars or primary credential. Requesting ERP API keys is proportionate to ERP integration, however the skill also references external LLM engines (claude-ma, deepseek) that typically require credentials but those are not declared. Undeclared/unnamed secret usage is a notable mismatch and raises the risk of accidental credential exposure or misconfiguration.

ℹ Persistence & Privilege

The skill is not marked always:true and uses the normal model-invocation defaults. That is appropriate. However, the codebase includes memory/persistent_store and an audit/logger and describes cross‑agent shared memory and trace export; if you install/run it it will store persistent state locally (SQLite/JSONL) and may retain data across runs. Combined with the other concerns, consider this persistence when deciding where/how to run it.

Version History

v1.0.0

foreign-trade-silicon-army v3.0.0 introduces a major update with enhanced multi-agent B2B collaboration, ERP integration, and robust error handling. - Added real ERP API integration layer supporting SAP, 用友, 金蝶, with circuit breaker, health checks, and mock mode. - Enhanced cross-agent B2B collaboration with fine-grained task protocol, state sync, and end-to-end traceability. - Integrated advanced error handling: 7-state task state machine, automatic exception classification, and flexible retry strategies. - Expanded to support Shopify, WooCommerce, and Magento for multi-CMS coordination. - Reinforced security and compliance with multi-layer approvals and audit logging. - Streamlined setup: plug and play, zero configuration required.

Metadata

Slug agent-cluster

Version 1.0.0

License MIT-0

All-time Installs 0

Active Installs 0

Total Versions 1

Frequently Asked Questions

What is Agent Cluster?

亚马逊外贸B2B多CMS Agent协作系统。支持Shopify/WooCommerce/Magento三大平台，A2A架构协调库存/采购/财务/物流四大专家Agent，三层安全网保障审批合规。即装即用，零配置开箱。 It is an AI Agent Skill for Claude Code / OpenClaw, with 79 downloads so far.

How do I install Agent Cluster?

Run "/install agent-cluster" in the OpenClaw or Claude Code chat to install it in one step — no extra setup required.

Is Agent Cluster free?

Yes, Agent Cluster is completely free, licensed under MIT-0. You can download, install and use it at no cost.

Which platforms does Agent Cluster support?

Agent Cluster is cross-platform and runs anywhere OpenClaw / Claude Code is available (cross-platform).

Who created Agent Cluster?

It is built and maintained by WangM-A3 (@wangm-a3); the current version is v1.0.0.

More Skills

Agent Cluster