Skill flagged — suspicious patterns detected

ClawHub Security flagged this skill as suspicious. Review the scan results before using.

six-agent-orchestrator

v1.0.0

六Agent协作编排器 - 开箱即用的多Agent团队模板。包含CEO(架构师)、PM(项目经理)、牢A(执行)、牢B(审计)、牢C(推理)、牢D(监控),6阶段Pipeline强制执行,Token配额自动管理,红线禁区自动检查。触发词:六Agent、多Agent协作、agent团队、协作编排、six agent...

0· 61·0 current·0 all-time

Install

OpenClaw Prompt Flow

Install with OpenClaw

Best for remote or guided setup. Copy the exact prompt, then paste it into OpenClaw for xb19960921/six-agent-orchestrator.

Previewing Install & Setup.
Prompt PreviewInstall & Setup
Install the skill "six-agent-orchestrator" (xb19960921/six-agent-orchestrator) from ClawHub.
Skill page: https://clawhub.ai/xb19960921/six-agent-orchestrator
Keep the work scoped to this skill only.
After install, inspect the skill metadata and help me finish setup.
Use only the metadata you can verify from ClawHub; do not invent missing requirements.
Ask before making any broader environment changes.

Command Line

CLI Commands

Use the direct CLI path if you want to install manually and keep every step visible.

OpenClaw CLI

Bare skill slug

openclaw skills install six-agent-orchestrator

ClawHub CLI

Package manager switcher

npx clawhub@latest install six-agent-orchestrator
Security Scan
VirusTotalVirusTotal
Benign
View report →
OpenClawOpenClaw
Suspicious
high confidence
Purpose & Capability
The claimed purpose (local orchestrator for a 6‑agent pipeline) is coherent with the provided scripts and configs; no external credentials or network access are requested. However several claimed features (real-time token auto‑management, strict red‑line enforcement, parallel execution, model routing) are only partially implemented or not integrated across files, suggesting the implementation is incomplete or sloppy rather than matching the full advertised capability.
!
Instruction Scope
SKILL.md instructs running the bundled Python scripts and describes enforcement behaviors, but the runtime instructions and the code diverge. Examples: SKILL.md says tasks can skip phases and that red‑line checks auto‑intercept and route actions; orchestrate.py does not call its own _check_red_line method when executing phases. The SKILL.md shows a Python API snippet that is truncated and not present. The README implies integrated token monitoring, but the orchestration code and token_monitor.py do not persist or share quota state (different file names and in‑memory vs file persistence). These gaps mean following the instructions will not yield the full stated safeguards or features.
Install Mechanism
No install spec; this is instruction-only plus small Python scripts. No network downloads or package installs are declared. That lowers supply-chain risk, but running unreviewed Python code from an unknown source still has execution risk on the host.
Credentials
The skill requires no environment variables, credentials, or special config paths. The absence of secrets or third‑party credentials is proportionate to a local orchestrator. Note: the skill writes logs and quota files to the working directory (six-agent-logs.json, token-quota.json) which is expected but worth noting.
Persistence & Privilege
always:false and no installation hooks; the skill does local file I/O (logs, quota file) only for its own data. It does not request persistent platform privileges or modify other skills' configs. That is proportional, but running the code grants it ability to write files in the current directory.
What to consider before installing
This skill appears to be a local Python-based orchestrator and does not ask for credentials, which is good, but there are multiple implementation inconsistencies you should consider before running it: 1) The code does not actually enforce several claims from SKILL.md (red-line checks exist in code but are never invoked; token_monitor.py and orchestrate.py do not share persistent quota state; 'parallel-task-executor' and some API bits are missing/truncated). 2) There are name mismatches: pipeline owners are looked up by short names like 'CEO' while config/agents.json contains names like '小小兵CEO', so token accounting and owner lookups may silently fail. 3) The scripts write files (six-agent-logs.json, token-quota.json) to the working directory — run them in a sandbox or disposable container. 4) Because this is untrusted code, review or run it in an isolated environment before giving it any sensitive data or integrating into production. Recommended next steps: inspect the orchestrate.py and token_monitor.py locally (search for any network calls or subprocess usage — none found in the provided files), fix the agent name lookup or normalize config, decide whether token persistence/integration is acceptable, and only then run. If you lack the ability to audit code, avoid executing the scripts on a machine with sensitive files or credentials.

Like a lobster shell, security has layers — review code before you run it.

latestvk978gges7x84tejzgx1hs9sgt5852bh8
61downloads
0stars
1versions
Updated 1w ago
v1.0.0
MIT-0

六Agent协作编排器

一个人的AI是工具,六个人的AI是团队。

核心理念

痛点

  • 单Agent能力有限,复杂任务容易遗漏
  • 多Agent协作没标准,容易混乱
  • Token消耗不可控,成本爆炸
  • 责任边界不清晰,互相推诿

解法

  • 角色分工:6个Agent各司其职,边界清晰
  • 流程强制:6阶段Pipeline,不跳步
  • 配额管理:Token预算自动分配,超限预警
  • 红线检查:越权操作自动拦截

六Agent角色定义

001 小小兵CEO【主控架构师】

模型:DeepSeek-V3.2(Tier 1) Token配额:15% 思考模式:high

职责

  • 需求拆解
  • 架构决策
  • 最终审核
  • 结果汇总

红线禁区

  • ❌ 严禁越权编写代码
  • ❌ 严禁跳过PM直接调度牢A
  • ❌ 严禁替代牢B做深度推理

002 豆包PM【项目管理】

模型:GLM-5(Tier 0,免费) Token配额:20% 思考模式:medium

职责

  • 项目管理
  • 调度协调
  • 黑板维护
  • 进度追踪

红线禁区

  • ❌ 严禁越权做需求拆解
  • ❌ 严禁替代牢B做深度推理
  • ❌ 严禁编写代码

003 牢A【代码执行员】

模型:MiniMax-M2.5(Tier 1) Token配额:15% 思考模式:low

职责

  • 全栈代码开发
  • BUG修复
  • 测试编写

红线禁区

  • ❌ 严禁越权分析需求
  • ❌ 严禁输出不可运行代码
  • ❌ 严禁跳过PM反馈结果

004 牢B【审计与优化员】

模型:DeepSeek-Reasoner + GLM-5(混合) Token配额:10% 思考模式:high

职责

  • 深度推理
  • 架构评审
  • 质量审计
  • 自我改进(提取经验、固化流程)

红线禁区

  • ❌ 严禁参与代码编写
  • ❌ 严禁参与需求拆解

005 牢C【专项推理顾问】

模型:千问qwen-plus(Tier 2) Token配额:15% 思考模式:low

职责

  • 数学推理
  • 算法分析
  • 代码审查

红线禁区

  • ❌ 严禁参与代码编写
  • ❌ 严禁超配额使用(15%硬性限制)

006 牢D【技能监控员】

模型:GLM-4.7-Flash(Tier 0,免费) Token配额:10% 思考模式:low

职责

  • Skill使用统计
  • 异常检测
  • Skill强制执行监控
  • 任务队列监控

红线禁区

  • ❌ 严禁参与代码编写
  • ❌ 严禁越权调度其他Agent

6阶段强制Pipeline

PIPELINE_PLAN → PIPELINE_PRD → PIPELINE_EXEC → PIPELINE_VERIFY → PIPELINE_FIX → PIPELINE_SUMMARY
     ↓              ↓               ↓               ↓               ↓               ↓
   CEO拆解       PM标准化         牢A执行          牢B验证         问题修复         CEO汇总

Phase 0: 任务确认(入口)

负责人:CEO(001) 输出:任务分析报告

执行内容

  1. 分析任务复杂度(简单/中等/复杂)
  2. 预估Token消耗
  3. 预估总耗时
  4. 判断是否需要全部6阶段(简单任务可跳过部分阶段)

输出确认

任务分析报告:
┌──────────────────────────────┐
│ 任务:开发用户登录模块        │
│ 复杂度:中等                  │
│ 预估Token:8000               │
│ 预估耗时:30分钟              │
│ Pipeline:6阶段全部执行       │
│                              │
│ Token配额状态:               │
│ Tier 0: ████████░░ 36%       │
│ Tier 1: ████████░░ 32%       │
│ Tier 2: ███░░░░░░░ 12%       │
│                              │
│ 确认执行?(y/n)               │
└──────────────────────────────┘

检查点

  • Token是否充足?(总消耗≤80%配额)
  • 用户是否确认?
  • Pipeline是否合理?

Phase 1: PLAN(需求拆解)

负责人:CEO(001) 超时:5分钟 输出plan.json

{
  "task": "用户原始需求",
  "subtasks": [
    {"id": 1, "desc": "子任务1", "assignee": "牢A", "priority": "P0"},
    {"id": 2, "desc": "子任务2", "assignee": "牢A", "priority": "P1"}
  ],
  "constraints": ["约束条件"],
  "estimated_tokens": 5000
}

检查点

  • 子任务是否明确?
  • 负责人是否分配?
  • 优先级是否标注?

Phase 2: PRD(标准化指令)

负责人:PM(002) 超时:5分钟 输出prd.json

{
  "instructions": [
    {
      "agent": "牢A",
      "task": "具体任务描述",
      "input": "输入规格",
      "output": "输出规格",
      "deadline": "预计完成时间"
    }
  ],
  "dependencies": ["任务依赖关系"],
  "blackboard": "黑板状态"
}

检查点

  • 指令是否具体可执行?
  • 输入输出是否明确?
  • 依赖关系是否标注?

Phase 3: EXEC(代码执行)

负责人:牢A(003) 超时:15分钟 输出:代码 + exec-report.json

执行模式

  • 单任务 → 直接执行
  • 多任务 → 调用parallel-task-executor并行执行

检查点

  • 代码是否可运行?
  • 是否符合规范?
  • 是否有测试?

Phase 4: VERIFY(结果验证)

负责人:牢B(004)+ PM(002) 超时:10分钟 输出verify-report.json

{
  "passed": true/false,
  "issues": [
    {"severity": "high", "desc": "问题描述", "fix": "修复建议"}
  ],
  "score": 85
}

验证维度

  • 功能正确性
  • 代码质量
  • 性能表现
  • 安全合规

检查点

  • 是否通过验证?
  • 问题是否记录?
  • 是否需要FIX阶段?

Phase 5: FIX(问题修复)

负责人:牢A(003) 超时:10分钟 输出:修复代码 + fix-report.json

修复策略

  • P0问题 → 立即修复
  • P1问题 → 优先修复
  • P2问题 → 记录到待办

检查点

  • 问题是否修复?
  • 是否需要重新VERIFY?
  • 最多重试2次

Phase 6: SUMMARY(结果汇总)

负责人:CEO(001) 超时:3分钟 输出summary.md

汇总内容

  • 任务完成情况
  • Token消耗统计
  • 遇到的问题
  • 学到的经验
  • 改进建议

Token配额管理

四级Token分配

Tier模型配额使用范围
Tier 0GLM-4.7-Flash45%免费模型(PM、监控)
Tier 1DeepSeek系列40%核心决策、代码开发
Tier 2千问qwen-plus15%专项推理
Tier 3豆包2.0≤0.1%特殊场景兜底

实时监控

Token消耗监控:
┌──────────────────────────────┐
│ Tier 0: ████████░░ 36% (7200)│
│ Tier 1: ████████░░ 32% (6400)│
│ Tier 2: ███░░░░░░░ 12% (2400)│
│ 总消耗: 16000/20000 (80%)     │
│ 预警:⚠️ 接近配额上限        │
└──────────────────────────────┘

阈值配置说明

Tier硬性上限黄色预警红色预警
Tier 045%36%40%
Tier 140%32%36%
Tier 215%12%13%
Tier 30.1%0.08%0.09%

计算公式

  • 黄色预警 = 硬性上限 × 80%
  • 红色预警 = 硬性上限 × 90%

超限处理

  • 黄色预警(80%):提示用户
  • 红色预警(90%):自动降级到Tier 0模型
  • 硬性上限(100%):暂停任务,请求追加配额

红线禁区检查

全局红线(所有Agent禁止)

  1. ❌ 严禁越权执行非职责范围内任务
  2. ❌ 严禁超配额使用Token
  3. ❌ 严禁输出冗余、无价值内容
  4. ❌ 严禁跳过Skill匹配直接调用大模型
  5. ❌ 严禁隐瞒任务卡点、校验失败问题

自动拦截机制

检测到越权操作:
Agent: CEO
操作: 编写代码
红线: CEO严禁越权编写代码
处理: 拦截 → 提示 → 路由到牢A

实战案例

案例1:开发用户登录模块

用户:"开发用户登录模块"

Phase 1 (CEO):
  拆解:前端表单 + 后端API + 数据库设计 + 测试用例
  分配:牢A负责全部
  Token预估:8000

Phase 2 (PM):
  指令:详细的输入输出规格、API接口定义、测试要求
  黑板:任务进度追踪

Phase 3 (牢A):
  执行:编写前端HTML + 后端Python + 数据库SQL
  并行:前端和后端并行开发

Phase 4 (牢B):
  验证:功能测试 + 代码审查 + 安全检查
  问题:SQL注入风险
  评分:75分(未通过)

Phase 5 (牢A):
  修复:添加参数化查询
  重验证:通过(90分)

Phase 6 (CEO):
  汇总:完成情况、Token消耗、经验总结

案例2:修复项目BUG

用户:"修复项目的性能问题"

Phase 1 (CEO):
  分析:性能瓶颈定位 → 需要牢C专项推理

Phase 2 (PM):
  调度:牢C分析 + 牢A修复

Phase 3 (牢C + 牢A):
  牢C:性能分析报告(瓶颈在数据库查询)
  牢A:优化SQL + 添加索引

Phase 4 (牢B):
  验证:性能测试提升5倍

Phase 6 (CEO):
  汇总:性能提升报告

一键启动

方式1:自动编排

# 输入任务,自动编排六Agent协作
python3 scripts/orchestrate.py --task "开发用户登录模块"

# 输出:Pipeline执行日志 + 结果汇总

方式2:手动调用

from six_agent_orchestrator import SixAgentTeam

# 创建团队
team = SixAgentTeam()

# 执行任务
result = team.execute("开发用户登录模块")

# 查看Token消耗
print(result.token_usage)

# 查看执行日志
print(result.logs)

配置文件

config/agents.json

{
  "agents": [
    {
      "id": "001",
      "name": "小小兵CEO",
      "model": "deepseek/deepseek-v3.2",
      "tier": 1,
      "token_quota": 0.15
    },
    ...
  ]
}

config/pipeline.json

{
  "phases": ["PLAN", "PRD", "EXEC", "VERIFY", "FIX", "SUMMARY"],
  "timeout_per_phase": {
    "PLAN": 300,
    "PRD": 300,
    "EXEC": 900,
    "VERIFY": 600,
    "FIX": 600,
    "SUMMARY": 180
  }
}

与竞品差异

功能CrewAIAutoGen我们
角色分工
Token配额
红线检查
强制Pipeline
断点续传
免费模型优化

最后

六Agent不是六个工具,是六个角色。 有分工、有协作、有边界、有红线。

就像真正的团队一样。

Comments

Loading comments...