Installation
Option 1: ClawhHub CLI (recommended)
CODEBLOCK0
Option 2: From GitHub
CODEBLOCK1
Project Coordinator Skill
A skill for structuring multi-agent project execution with isolated session architecture.
Manifest
CODEBLOCK2
Architecture
CODEBLOCK3
Token efficiency: Each project Coordinator is isolated. Old project sessions accumulate zero main-session tokens after completion.
Least-privilege design: The Coordinator does NOT read session transcript files directly. All transcript reading, sanitization, and archiving is delegated to a dedicated archive-subagent using the archive-project skill.
Trigger Conditions
Activated when:
- - User says "start a new project" or "let's work on [project]"
- A task is complex and needs multiple subagents
- A task will take more than a few minutes
- When user says "archive this" after a project is done → spawn archive-subagent
Do NOT activate for: quick questions, simple lookups, one-liner tasks.
Trigger 3: Slash command
Type
//start followed by your project description to activate the Project Coordinator.
Example: "//start build a Chrome extension"
Project Coordinator Pattern
Step 1: Define the Project
Collect from the user:
- - Project name (e.g., "cureforge-hr-assessment")
- Project description (1-2 sentences)
- Key deliverables
- Any known constraints
Step 2: Spawn Project Coordinator
CODEBLOCK4
Step 3: Monitor Progress
- - Use
subagents(action="list") to track subagent status - Receive completion announcements via inter-session messages
- If a subagent fails, decide whether to retry or adapt
Step 4: Compile and Report
When all subagents complete:
- 1. Review all outputs
- Write a final summary
- Save any deliverables before the Coordinator session ends
Step 5: Archiving (Delegated)
When a project is complete and archiving is requested:
- 1. The Coordinator spawns an archive-subagent
- The archive-subagent uses the archive-project skill to handle:
- Locating the correct session transcripts
- Sanitizing credentials
- Writing ARCHIVE.md
- Committing to workspace
- 3. The Coordinator monitors the archive-subagent and reports completion
The Coordinator itself does NOT read session transcript files directly.
Coordinator's Tool Usage
Note: The Coordinator should NOT directly read, copy, move, or delete session transcript files. All such operations must be performed by a dedicated archive-subagent using the archive-project skill.
Subagent sandboxing: When spawning subagents, each subagent runs in an isolated sandbox with workspace-only filesystem access. Subagents cannot access credentials, environment variables, or session transcripts outside their scope. Network access is restricted per platform policy.
The Coordinator SHOULD directly call tools:
- -
exec — run commands, check files - INLINECODE3 — create output files
- INLINECODE4 — examine code or documents
- INLINECODE5 — spawn subagents for parallel work
- INLINECODE6 — monitor subagent status
Subagent Naming Convention
Format: INLINECODE7
Examples:
- - INLINECODE8
- INLINECODE9
- INLINECODE10
Session Management Rules
- - One Coordinator per project (isolated token accounting)
- Do NOT run multiple unrelated projects in one Coordinator
- When a project is complete, the Coordinator stops — no lingering sessions
- For long-running projects: Coordinator can spawn child Coordinators for sub-phases
After Project Completion
- 1. Report completion to Main session (brief summary)
- Notify the user
- User decides: archive / continue / hold
安装
选项一:ClawhHub CLI(推荐)
bash
openclaw skills install project-coordinator
或
clawhub install project-coordinator
选项二:从GitHub安装
bash
克隆仓库
git clone https://github.com/KaigeGao1110/Project-Coordinator.git ~/.openclaw/skills/project-coordinator
或直接下载
curl -L https://github.com/KaigeGao1110/Project-Coordinator/archive/refs/heads/main.zip -o /tmp/project-coordinator.zip
unzip /tmp/project-coordinator.zip -d ~/.openclaw/skills/
mv ~/.openclaw/skills/Project-Coordinator-main ~/.openclaw/skills/project-coordinator
项目协调器技能
一种用于构建多智能体项目执行的技能,采用隔离会话架构。
清单
yaml
name: project-coordinator
version: 1.0.11
description: |
生成一个隔离的项目协调器会话,该会话拥有项目上下文,
将工作分解为任务,并生成子智能体进行并行执行。
协调器在完成后向主会话报告。
permissions:
- spawn: 子智能体会话(通过 sessions_spawn)
- read: 工作区文件
- exec: 通过子智能体执行 shell 命令
dataPolicy:
archivedData: 仅限内部工作区
neverExternal: true
架构
主会话(从不直接运行代码)
└── 项目协调器(按项目生成,模式=run)
└── 子智能体(由协调器生成,并行执行任务)
Token 效率:每个项目协调器都是隔离的。旧项目会话在完成后不会累积主会话 token。
最小权限设计:协调器不直接读取会话记录文件。所有会话记录的读取、清理和归档工作都委托给专门的归档子智能体,使用 archive-project 技能完成。
触发条件
在以下情况下激活:
- - 用户说开始一个新项目或我们来处理 [项目]
- 任务复杂且需要多个子智能体
- 任务需要几分钟以上时间
- 项目完成后用户说归档这个 → 生成归档子智能体
不要为以下情况激活:快速提问、简单查询、一句话任务。
触发条件 3:斜杠命令
输入 //start 后跟项目描述即可激活项目协调器。
示例://start 构建一个 Chrome 扩展
项目协调器模式
步骤 1:定义项目
从用户处收集:
- - 项目名称(例如:cureforge-hr-assessment)
- 项目描述(1-2 句话)
- 关键交付物
- 任何已知约束
步骤 2:生成项目协调器
python
sessions_spawn(
task=f你是 {project_name} 的项目协调器。
项目:{project_description}
你的职责:
- 1. 理解项目的完整范围
- 将其分解为独立任务
- 生成子智能体并行执行任务
- 监控子智能体进度
- 编译结果并向主会话报告
完成后,编写摘要并等待进一步指示。
,
label=fcoordinator-{project_name},
runtime=subagent,
model=minimax-m2.7-highspeed,
mode=run
)
步骤 3:监控进度
- - 使用 subagents(action=list) 跟踪子智能体状态
- 通过会话间消息接收完成通知
- 如果子智能体失败,决定是重试还是调整
步骤 4:编译并报告
当所有子智能体完成后:
- 1. 审查所有输出
- 编写最终摘要
- 在协调器会话结束前保存所有交付物
步骤 5:归档(委托)
当项目完成并请求归档时:
- 1. 协调器生成一个归档子智能体
- 归档子智能体使用 archive-project 技能处理:
- 定位正确的会话记录
- 清理凭证
- 编写 ARCHIVE.md
- 提交到工作区
- 3. 协调器监控归档子智能体并报告完成
协调器本身不直接读取会话记录文件。
协调器的工具使用
注意: 协调器不应直接读取、复制、移动或删除会话记录文件。所有此类操作必须由专门的归档子智能体使用 archive-project 技能执行。
子智能体沙箱: 生成子智能体时,每个子智能体在隔离的沙箱中运行,仅可访问工作区文件系统。子智能体无法访问其范围之外的凭证、环境变量或会话记录。网络访问根据平台策略受限。
协调器应直接调用以下工具:
- - exec — 运行命令、检查文件
- write — 创建输出文件
- read — 检查代码或文档
- sessions_spawn — 生成子智能体进行并行工作
- subagents — 监控子智能体状态
子智能体命名约定
格式:{项目}-{角色}
示例:
- - cureforge-gke-migration
- cureforge-rabbitmq-queue
- cureforge-review
会话管理规则
- - 每个项目一个协调器(隔离的 token 核算)
- 不要在一个协调器中运行多个不相关的项目
- 项目完成后,协调器停止运行——不遗留会话
- 对于长期运行的项目:协调器可以为子阶段生成子协调器
项目完成后
- 1. 向主会话报告完成(简要摘要)
- 通知用户
- 用户决定:归档 / 继续 / 保留