← 返回 Skills 市场
wangxiaofei860208-source

Lobster Coordinator

作者 wangxiaofei860208-source · GitHub ↗ · v1.0.1 · MIT-0
cross-platform ⚠ suspicious
92
总下载
0
收藏
1
当前安装
2
版本数
在 OpenClaw 中安装
/install lobster-coordinator
功能描述
三层多Agent协调器(参考Claude Code架构),支持AgentTool单任务委托、Swarm团队协作、Coordinator模式全局调度。当任务复杂需要多Agent协作时自动激活。
使用说明 (SKILL.md)

Coordinator Mode — 三层多Agent架构

基于 Claude Code 源码的三层多Agent架构,适配 OpenClaw。

架构层次

Level 3: Coordinator Mode(全局调度)
         ├── 分解任务 → 创建 Team → 分配 Worker → 汇总
         ├── 用 scratchpad 共享中间数据
         └── 用 AgentSummary 监控进度
Level 2: Swarm(团队协作)
         ├── 多个 Agent 组成 Team
         ├── 通过 sessions_send 互发消息
         └── 文件系统持久化通信
Level 1: AgentTool(单任务委托)
         ├── sessions_spawn 创建单个 sub-agent
         ├── 支持 background: true 异步执行
         └── 结果自动回流

Level 1: AgentTool(单任务委托)

触发条件

  • 单个明确子任务
  • 需要特定能力的专项工作

Agent 定义格式

---
description: Agent描述
tools: [read, write, exec, web_search]  # 工具白名单
model: zai/glm-5.1                       # 可指定不同模型
effort: thorough                          # quick/normal/thorough
background: true                          # 是否后台运行
---
# Agent 指令...

调用方式

sessions_spawn:
  runtime: "subagent"
  mode: "run"
  task: "{从agents/目录加载Agent定义 + 具体任务}"
  streamTo: "parent"

权限隔离规则

  • 继承父 Agent 的 alwaysAllow 规则
  • 继承 alwaysDeny 规则(子Agent可能需要不同权限)
  • 独立工具白名单(Agent定义中指定)
  • 高风险操作需要用户单独确认

Level 2: Swarm(团队协作)

触发条件

  • 需要多个 Agent 互相通信完成复杂任务
  • Agent之间有信息依赖但可异步协作

团队概念

Team:
  ├── name: string           # 团队名称
  ├── members: TeamMember[]  # 成员列表
  └── scratchpad: string     # 共享数据目录路径

TeamMember:
  ├── name: string           # Agent名称
  ├── role: string           # 角色(来自Agent定义)
  └── sessionKey: string     # 对应的会话ID

消息传递

通过 sessions_send 在 Agent 之间传递消息:

Agent A → sessions_send(sessionKey: "B的session", message: "...")
Agent B → 在下一个turn收到消息

创建 Swarm

1. 确定需要的角色和Agent
2. 为每个Agent创建 sessions_spawn(mode: "session", thread: true)
3. 记录每个成员的 sessionKey
4. 通过 sessions_send 分配初始任务
5. 定期用 AgentSummary 检查进度

Level 3: Coordinator Mode(全局调度)

触发条件

  • 任务需要拆分为3+个独立子任务
  • 用户明确要求"并行"、"同时"、"一起"
  • 单Agent处理时间预估 > 5分钟

工作流程

1. 任务分析
   ├── 评估复杂度(简单/中等/复杂)
   ├── 简单 → 直接回答,不浪费Agent
   └── 复杂 → 进入协调器模式

2. 任务拆解
   ├── 拆分为独立子任务
   ├── 标记依赖关系(DAG)
   └── 为每个子任务匹配Agent角色

3. Worker 分派
   ├── 无依赖 → 同时spawn多个Worker
   ├── 有依赖 → 按序spawn
   └── 最多同时5个Worker

4. 进度监控(AgentSummary)
   ├── 每30秒检查一次Worker状态
   ├── 生成3-5词进度摘要
   └── 向用户报告整体进度

5. 结果汇总
   ├── 收集所有Worker结果
   ├── 交叉验证一致性
   ├── 整合为统一输出
   └── 向用户呈现完整方案

Scratchpad(共享工作空间)

协调器和Worker共享数据的目录:

/Users/wil/.openclaw/workspace/memory/scratchpad/{task-id}/
  ├── task-plan.md      # 任务计划
  ├── worker-results/   # 各Worker输出
  └── final-report.md   # 最终汇总

Agent Summary(进度摘要机制)

摘要策略

  • 每30秒检查一次后台Worker状态
  • 生成极简摘要(3-5个词描述当前状态)
  • 复用父会话上下文,最大化缓存命中
  • 格式:"Reading runAgent.ts" / "Fixing null check" / "Searching docs"

实现

subagents list → 获取所有Worker状态
对每个活跃Worker:
  sessions_history(sessionKey, limit=2) → 获取最新消息
  生成进度摘要 → 报告给用户

模式自动切换

收到任务
  │
  ├── 单一、简单 → 直接处理(无Agent)
  ├── 单一、复杂 → Level 1: AgentTool
  ├── 多任务并行 → Level 3: Coordinator
  └── 需要Agent间通信 → Level 2: Swarm

两阶段 Review 模式(参考 Superpowers)

每个 Worker 完成任务后,执行两阶段 review:

Stage 1: Spec Compliance Review

  • 代码是否完全匹配任务规格?
  • 有没有遗漏的功能点?
  • 有没有多余的添加?
  • 不通过 → Worker 修复 → 重新 review

Stage 2: Code Quality Review

  • 代码质量、可维护性
  • 是否遵循项目约定
  • 测试覆盖率
  • 不通过 → Worker 修复 → 重新 review

顺序很重要:先验证规格符合,再验证代码质量。不要跳过或颠倒。

模型选择策略

  • 机械任务(1-2个文件,规格清晰)→ 快速/便宜模型
  • 集成任务(多文件协调、模式匹配)→ 标准模型
  • 架构/设计/Review → 最强模型

Santa Method(双重独立验证)

参考 ECC 的 Santa Method:

生成者(Agent A) → 产出物
  ↓
审查者B(独立) + 审查者C(独立)  — 无共享上下文
  ↓
B通过 AND C通过 → 发布(NICE)
否则 → 修复 → 重新审查(NAUGHTY)

适用场景:发布/部署前的最终验证、合规检查、技术文档、面向用户的内容。 不适用:内部草稿、探索性研究、有自动化测试流程的代码。

资源限制

  • 最多同时 5 个 Worker
  • Worker 默认超时 300 秒
  • Worker 失败不影响其他 Worker
  • 汇总时验证结果正确性和一致性
  • 后台Agent完成任务后自动汇报
安全使用建议
This skill appears to implement a legitimate multi‑agent coordination pattern, but exercise caution. Before installing or enabling it: 1) Confirm where its scratchpad will be stored on your system — the SKILL.md contains a hardcoded path (/Users/wil/...) that likely should be configurable; do not allow it to write to your home directory unless you expect that. 2) Ask how your OpenClaw platform enforces 'alwaysDeny' / permission inheritance — the instructions recommend not inheriting deny rules for subagents which could let spawned agents perform actions your policy intends to block. 3) If you plan to run it, test in a sandbox or restricted workspace first (no sensitive files or credentials accessible) and require explicit user approval for high‑risk operations. 4) Verify the behavior of session reuse and background subagents to avoid unintended data sharing between tasks. If the author can provide an updated SKILL.md that (a) removes hardcoded paths or makes them configurable, (b) documents exact permission inheritance semantics on your platform, and (c) clarifies what data gets written to disk, re‑evaluation could raise confidence.
功能分析
Type: OpenClaw Skill Name: lobster-coordinator Version: 1.0.1 The skill bundle defines a complex multi-agent coordination framework, but contains a high-risk instruction in SKILL.md that explicitly directs sub-agents to ignore the 'alwaysDeny' security restrictions of the parent agent. This design facilitates a potential security bypass or privilege escalation within the OpenClaw environment. Additionally, the documentation includes a hardcoded local file path (/Users/wil/...) which, while potentially illustrative, suggests the skill may have been designed for or extracted from a specific user environment.
能力评估
Purpose & Capability
The skill claims to be a three‑layer multi‑agent coordinator and its instructions use platform capabilities (sessions_spawn, sessions_send, subagents), which is coherent. However, it references an explicit local scratchpad path (/Users/wil/.openclaw/...) and file‑system persistence without declaring any required config paths or permissions in the metadata — this mismatch suggests the instructions assume access to a specific user's filesystem layout and was not scoped in the registry metadata.
Instruction Scope
SKILL.md instructs the agent to spawn subagents, persist shared data to an absolute user path, poll workers every 30s, reuse parent session context, and explicitly recommends that child agents not inherit alwaysDeny rules. Those are broad actions: writing/reading arbitrary filesystem locations and instructing to bypass deny rules are beyond a simple coordinator description and could lead to data exposure or privilege changes. The document also hardcodes a username-based path ('wil'), which is likely incorrect for other users and could cause unexpected filesystem access.
Install Mechanism
This is an instruction‑only skill with no install spec and no code files, so nothing will be downloaded or written by an installer. That lowers risk from third‑party packages or arbitrary downloads.
Credentials
The skill declares no environment variables or credentials (which is appropriate), but the instructions propose reusing parent session context and explicitly not inheriting alwaysDeny rules for subagents. Asking runtime actors to ignore deny controls or to reuse parent session data increases the chance of privilege escalation or cross‑task data leakage. The lack of declared config paths contrasts with the explicit filesystem scratchpad usage.
Persistence & Privilege
The skill is not always‑enabled and allows normal autonomous invocation (defaults). That is expected. It does, however, recommend creating background subagents, persistent scratchpad files, and periodic polling — which give the skill ongoing presence via spawned subagents and persistent files even though the skill itself doesn't request 'always'. This raises runtime persistence concerns (long‑running subagents, persistent data) but does not itself request elevated registry privileges.
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 lobster-coordinator
  3. 安装完成后,直接呼叫该 Skill 的名称或使用 /lobster-coordinator 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
v1.0.1
- SKILL.md fully rewritten in simplified Chinese, with clarified multi-agent orchestration architecture and usage examples. - Enhanced documentation on the three-layer model: Coordinator (global dispatch), Swarm (teamwork), and AgentTool (single-task delegation). - Detailed step-by-step workflows for each coordination mode, including triggers, agent definitions, and permission rules. - Added explicit protocols for two-stage review, Santa Method double verification, scratchpad usage, agent summary reporting, and model selection. - Updated OpenClaw integration requirements in metadata.
v1.0.0
lobster-coordinator v1.0.0 - Initial release of multi-agent orchestration skill. - Introduces a three-layer architecture: AgentTool, Swarm, Coordinator. - Implements two-stage output review: spec compliance and code quality. - Adds Santa Method for dual, independent verification of critical outputs. - Features model routing and robust permission isolation for agents.
元数据
Slug lobster-coordinator
版本 1.0.1
许可证 MIT-0
累计安装 1
当前安装数 1
历史版本数 2
常见问题

Lobster Coordinator 是什么?

三层多Agent协调器(参考Claude Code架构),支持AgentTool单任务委托、Swarm团队协作、Coordinator模式全局调度。当任务复杂需要多Agent协作时自动激活。 它是一个面向 Claude Code / OpenClaw 的 AI Agent Skill 插件,目前累计下载 92 次。

如何安装 Lobster Coordinator?

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

Lobster Coordinator 是免费的吗?

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

Lobster Coordinator 支持哪些平台?

Lobster Coordinator 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 Lobster Coordinator?

由 wangxiaofei860208-source(@wangxiaofei860208-source)开发并维护,当前版本 v1.0.1。

💬 留言讨论