← 返回 Skills 市场
347
总下载
0
收藏
0
当前安装
1
版本数
在 OpenClaw 中安装
/install agile-workflow
功能描述
全自动敏捷协作工作流引擎 v7.18.1 - 细纲串行+正文并行、并发安全架构、零数据污染、智能合并
使用说明 (SKILL.md)
全自动敏捷协作工作流引擎 v7.0
重大升级: 并发安全架构,解决多 Agent 并发的数据污染问题
By OpenClaw Community — 智能敏捷协同系统
新一代工作流引擎:智能拆解、自动学习、持续优化
🎯 版本演进
v7.0 vs v6.1 (最新)
| 功能 | v6.1 | v7.0 | 提升 |
|---|---|---|---|
| 数据污染风险 | ⚠️ 中等 | 🟢 零 | ✅ 100% 消除 |
| 写入隔离 | ❌ 无 | ✅ 自动隔离 | ✅ 新增 |
| 依赖管理 | 基础 | DAG+ 环路检测 | ✅ 增强 |
| 合并策略 | 3 种 | 7 种 | ✅ 新增 4 种 |
| 冲突检测 | ❌ 无 | ✅ 自动检测 | ✅ 新增 |
| 并发安全 | 🔴 高风险 | 🟢 零风险 | ✅ 架构级 |
v6.1 vs v4.0
| 功能 | v4.0 | v6.1 | 提升 |
|---|---|---|---|
| 任务拆解 | 手动定义 | 智能识别依赖 | ⬆️ 自动化 |
| 依赖管理 | 静态配置 | 动态检测 | ⬆️ 灵活性 |
| Agent 协作 | 固定分配 | 负载均衡 | ⬆️ 效率 30% |
| 状态监控 | 分钟级 | 秒级 | ⬆️ 实时性 60x |
| 缓存机制 | ❌ 无 | ✅ LRU+TTL | ⬆️ 响应 -75% |
| 并发优化 | 串行 | 并行 | ⬆️ 10 倍 |
| 测试覆盖 | 基础 | >90% | ⬆️ 质量 |
🚀 核心功能
1. 智能任务拆解
输入:创作一部玄幻小说
↓ 智能拆解
输出:
- 世界观架构 (依赖:无)
- 人物体系 (依赖:世界观)
- 情节大纲 (依赖:人物体系)
- 章节细纲 (依赖:情节大纲)
- 正文创作 (依赖:章节细纲)
- 审查 (依赖:正文创作)
特点:
- ✅ 自动识别任务类型(小说/开发/文档)
- ✅ 自动分析依赖关系
- ✅ 应用历史经验优化拆解
2. 多 Agent 智能协作
任务 → 智能路由 → Agent 选择 → 负载均衡 → 执行
↓
实时负载检测
(避免过载)
特点:
- ✅ 根据任务类型自动选择最佳 Agent
- ✅ 实时检测 Agent 负载
- ✅ 动态调整任务分配
3. 实时状态追踪
状态机:
pending → ready → running → completed
↓
failed/timeout
特点:
- ✅ 秒级状态更新
- ✅ 自动检测超时任务
- ✅ 失败自动重试
4. 自动学习迭代
执行 → 记录结果 → 分析模式 → 生成优化 → 应用优化
↓ ↑
└──────────────────────────────────────┘
持续改进闭环
特点:
- ✅ 记录成功/失败模式
- ✅ 自动生成优化建议
- ✅ 持续改进工作流
📦 安装与配置
安装
# 方式 1: 从 ClawHub 安装
clawhub install agile-workflow
# 方式 2: 更新到 v4.0
cd ~/.openclaw/workspace/skills/agile-workflow
git pull origin main
npm install --omit=dev
配置
1. 启用 Skill
在 ~/.openclaw/openclaw.json 中:
{
"skills": {
"entries": {
"agile-workflow": {
"enabled": true,
"config": {
"autoTrigger": true,
"monitorInterval": 10, // v4.0: 10 秒 (原 60 秒)
"maxConcurrentTasks": 3,
"activeProjectThreshold": 24,
"autoLearn": true, // v4.0 新增:自动学习
"enableOptimizations": true // v4.0 新增:启用优化
}
}
}
}
}
2. 启动引擎
# 启动工作流引擎
node /home/ubutu/.openclaw/workspace/skills/agile-workflow/core/agile-workflow-engine.js start
# 后台运行(推荐)
nohup node /home/ubutu/.openclaw/workspace/skills/agile-workflow/core/agile-workflow-engine.js start > /workspace/logs/agile-workflow/engine.log 2>&1 &
3. 配置 Crontab
crontab -e
# v4.0 配置
*/1 * * * * node /home/ubutu/.openclaw/workspace/skills/agile-workflow/core/agile-workflow-engine.js monitor --quiet
0 */6 * * * node /home/ubutu/.openclaw/workspace/skills/agile-workflow/core/agile-workflow-engine.js learn --quiet
0 3 * * * /home/ubutu/.openclaw/workspace/skills/agile-workflow/scripts/auto-spec-discovery.sh --quiet
🛠️ 使用方式
命令列表
1. 任务管理
# 智能拆解任务
node agile-workflow-engine.js decompose novel_creation
node agile-workflow-engine.js decompose software_dev
# 查看任务状态
node agile-workflow-engine.js status
# 监控所有任务
node agile-workflow-engine.js monitor
2. 执行控制
# 启动引擎
node agile-workflow-engine.js start
# 清理僵尸任务
node agile-workflow-engine.js cleanup
# 生成优化建议
node agile-workflow-engine.js learn
3. 日志查看
# 查看引擎日志
tail -f /home/ubutu/.openclaw/workspace/logs/agile-workflow/engine.log
# 查看任务日志
tail -f /home/ubutu/.openclaw/workspace/logs/agile-workflow/task-monitor.log
# 查看学习日志
tail -f /home/ubutu/.openclaw/workspace/logs/agile-workflow/learning.log
📊 工作流程
完整流程图
用户输入任务
↓
[智能拆解引擎]
↓
生成子任务 + 依赖关系
↓
[任务分配器] → 选择最佳 Agent → 检查负载
↓
任务队列 (按依赖排序)
↓
[执行引擎] → 执行任务 → 监控状态
↓
任务完成 → 记录结果
↓
[学习系统] → 分析模式 → 生成优化
↓
触发下游任务
↓
循环直到所有任务完成
小说创作示例
# 1. 拆解任务
node agile-workflow-engine.js decompose novel_creation
# 输出:
{
"subtasks": [
{ "name": "世界观架构", "type": "world_building", "dependsOn": [] },
{ "name": "人物体系", "type": "character_design", "dependsOn": ["world_building"] },
{ "name": "情节大纲", "type": "plot_outline", "dependsOn": ["character_design"] },
{ "name": "章节细纲", "type": "chapter_outline", "dependsOn": ["plot_outline"] },
{ "name": "正文创作", "type": "chapter_write", "dependsOn": ["chapter_outline"] },
{ "name": "审查", "type": "review", "dependsOn": ["chapter_write"] }
]
}
# 2. 启动引擎
node agile-workflow-engine.js start
# 3. 监控进度
node agile-workflow-engine.js monitor
# 输出:
总计:6 | 待执行:0 | 进行中:1 | 完成:4 | 失败:0
🧠 学习系统
经验记录
成功经验:
{
"taskType": "chapter_write",
"agent": "chapter_writer",
"duration": 180000,
"timestamp": 1710288000000
}
失败经验:
{
"taskType": "world_building",
"agent": "world_builder",
"error": "context-length-exceeded",
"timestamp": 1710288000000
}
优化建议生成
# 生成优化建议
node agile-workflow-engine.js learn
# 输出:
📊 分析历史经验,生成优化建议...
✅ 生成 3 条优化建议
建议列表:
1. chapter_write → chapter_writer 平均耗时 180000ms,建议优先使用 (置信度:高)
2. world_building → world_builder 失败率 30%,建议增加上下文限制 (置信度:中)
3. review 任务建议在凌晨执行,避免资源竞争 (置信度:中)
📈 监控指标
任务指标
| 指标 | 说明 | 获取方式 |
|---|---|---|
| 总任务数 | 所有任务总数 | status 命令 |
| 待执行 | 等待依赖完成 | status 命令 |
| 进行中 | 正在执行 | status 命令 |
| 已完成 | 成功完成 | status 命令 |
| 失败 | 执行失败 | status 命令 |
| 超时 | 超过 1 小时未完成 | 自动检测 |
Agent 指标
| 指标 | 说明 | 告警阈值 |
|---|---|---|
| 并发数 | 同时执行任务数 | > 3 告警 |
| 成功率 | 成功/总任务 | \x3C 80% 告警 |
| 平均耗时 | 任务平均执行时间 | > 30 分钟告警 |
| 失败率 | 失败/总任务 | > 20% 告警 |
🔧 故障排查
问题 1: 任务未自动触发
检查:
# 查看引擎状态
node agile-workflow-engine.js status
# 查看依赖关系
node agile-workflow-engine.js decompose novel_creation
# 查看日志
tail -100 /workspace/logs/agile-workflow/engine.log
解决:
- 确认引擎已启动
- 检查上游任务是否完成
- 验证依赖关系配置
问题 2: Agent 负载过高
检查:
# 查看 Agent 负载
ps aux | grep chapter_writer | wc -l
# 查看任务队列
node agile-workflow-engine.js status | grep running
解决:
# 增加最大并发数(配置文件)
"maxConcurrentTasks": 5
# 或等待任务完成
node agile-workflow-engine.js monitor
问题 3: 学习系统未生效
检查:
# 查看经验库
cat /workspace/logs/agile-workflow/experience-base.json
# 查看学习日志
tail -100 /workspace/logs/agile-workflow/learning.log
解决:
# 确认 autoLearn 配置
# 在 openclaw.json 中设置 "autoLearn": true
# 手动触发学习
node agile-workflow-engine.js learn
📊 性能对比
v3.6 vs v4.0
| 指标 | v3.6 | v4.0 | 提升 |
|---|---|---|---|
| 任务触发延迟 | 60 秒 | 10 秒 | 6x |
| 任务完成率 | 75% | 92% | 23% |
| 平均执行时间 | 25 分钟 | 18 分钟 | 28% |
| 失败恢复时间 | 手动 | 自动 | 100% |
| 优化建议 | 无 | 自动生成 | 新增 |
🎯 最佳实践
1. 任务拆解
# 大任务拆解为小任务
node agile-workflow-engine.js decompose novel_creation
# 审查拆解结果,确保依赖正确
# 手动调整不合理依赖(如需要)
2. 监控配置
# 启动引擎(后台运行)
nohup node agile-workflow-engine.js start > engine.log 2>&1 &
# 添加监控告警
# 编辑 crontab,每 5 分钟检查一次
*/5 * * * * node agile-workflow-engine.js monitor --quiet
3. 学习优化
# 每周生成优化建议
0 9 * * 1 node agile-workflow-engine.js learn
# 应用优化建议
# 根据建议调整配置或任务分配
📚 核心文件
| 文件 | 路径 | 用途 |
|---|---|---|
| 引擎核心 | /skills/agile-workflow/core/agile-workflow-engine.js |
工作流引擎 |
| 状态文件 | /workspace/logs/agile-workflow/workflow-state.json |
任务状态 |
| 经验库 | /workspace/logs/agile-workflow/experience-base.json |
学习经验 |
| 引擎日志 | /workspace/logs/agile-workflow/engine.log |
运行日志 |
| 学习日志 | /workspace/logs/agile-workflow/learning.log |
学习记录 |
🔄 任务依赖模型 v1.0(细纲串行 + 正文并行)
核心原则
- 细纲任务串行执行:
outline_N依赖outline_{N-1} - 正文任务只依赖细纲:
write_N依赖outline_N(不依赖write_{N-1}) - 流水线并行:细纲完成后正文可立即开始
依赖关系图
细纲任务链(串行):
outline_01 → outline_02 → outline_03 → outline_04 → ...
正文任务链(只依赖细纲):
write_01 (依赖 outline_01)
write_02 (依赖 outline_02) ← 不等待 write_01
write_03 (依赖 outline_03) ← 不等待 write_02
write_04 (依赖 outline_04) ← 不等待 write_03
并行执行效果
| 时间 | 完成任务 | 可开始任务 | 并行情况 |
|---|---|---|---|
| T1 | outline_01 | outline_02 + write_01 | 并行 |
| T2 | outline_02 | outline_03 + write_02 | 并行 |
| T3 | outline_03 | outline_04 + write_03 | 并行 |
任务命名规范
| 任务类型 | ID 格式 | 示例 |
|---|---|---|
| 细纲任务 | outline_NN |
outline_01, outline_02, ... |
| 正文任务 | write_NN |
write_01, write_02, ... |
| 审查任务 | review_NN |
review_01, review_02, ... |
任务属性
{
"id": "outline_01",
"name": "第1章_细纲_半块玉牌",
"chapter": 1,
"type": "outline", // outline | writing | review
"status": "pending",
"agent": "novel_architect",
"output": "/path/to/04_章节细纲/第01章_半块玉牌.md"
}
相关脚本
- 依赖生成器:
scripts/task-dependency-generator.js - 任务修复:
scripts/repair-task-states.js - 依赖检查:
core/dependency-manager.js
🔄 版本历史
v7.18.1 (2026-03-15)
- ✅ 清理老版本文件(v5/v7 引擎、v2 执行器、v2 健康检查)
- ✅ 删除冗余测试框架(stress-test、test-framework)
- ✅ 释放空间 120 KB,文件数从 53 减少到 47
- ✅ 备份位置:
backups/20260315_215924/
v7.18.0 (2026-03-15)
- ✅ 新增任务依赖模型(细纲串行 + 正文并行)
- ✅ 新增任务依赖生成器
- ✅ 新增任务状态修复脚本
- ✅ 优化流水线并行效率
v4.0.0 (2026-03-12)
- ✅ 新增智能任务拆解引擎
- ✅ 新增多 Agent 智能协作
- ✅ 新增实时状态追踪(秒级)
- ✅ 新增自动学习迭代系统
- ✅ 监控间隔从 60 秒降至 10 秒
- ✅ 性能提升 30%
v3.6.0 (2026-03-08)
- ✅ 修复文件名匹配问题
- ✅ 实现按顺序触发
- ✅ 新增活跃项目检测
v3.5.0 (2026-03-08)
- ✅ 新增通用任务依赖自动触发
- ✅ 支持所有项目类型
📖 参考文档
让敏捷协作全自动、智能化、持续优化! 🚀
安全使用建议
This package appears to be a full workflow engine and includes many executable JavaScript files plus instructions to enable background services and edit your agent config. Before installing or enabling: 1) Do not add the crontab entries or enable autoTrigger until you audit the code. 2) Inspect key files locally (llm-gateway.js, agent-manager.js, any files that use network, child_process, fs, or eval) for outbound endpoints, hard-coded URLs, or code that reads unrelated host files. 3) Search the code for child_process.exec/spawn, net/http/axios/fetch, websocket/socket usage, and any calls that post data externally. 4) Pay special attention to SKILL.md content flagged as 'system-prompt-override' — remove or question any lines that attempt to change model/system prompts. 5) Check package.json and package-lock.json for unexpected dependencies; run static scans (SAST) and dynamic tests in an isolated VM or container. 6) If you must test on your machine, run the service in a sandboxed account with no access to sensitive files and avoid enabling persistent crontab/nohup until you are confident. If you want, I can list high-priority files/strings to inspect (network calls, credential patterns, child_process) or help generate a review checklist.
功能分析
Type: OpenClaw Skill
Name: agile-workflow
Version: 7.18.1
The bundle implements a highly complex and well-documented workflow engine for multi-agent collaboration. It is classified as suspicious primarily due to significant Command Injection vulnerabilities in core/task-scheduler.js, core/failure-handler.js, and core/agent-supervisor.js, where unsanitized task descriptions and project names are interpolated directly into shell command strings via child_process.exec. The skill also establishes persistence through crontab modifications and opens a network port (8080) for a monitoring dashboard (dashboard/backend/server.js). While these capabilities are consistent with the stated goal of an automated, self-healing workflow engine, the lack of input sanitization during shell execution creates a high risk of remote code execution if an agent processes a maliciously crafted task description.
能力评估
Purpose & Capability
The name, description and many included JS modules (agent manager, scheduler, circuit breaker, llm-gateway, token manager, etc.) are internally consistent with an on-host workflow engine. However the metadata claims 'instruction-only' / no install spec while the bundle contains dozens of executable code files — that mismatch is unexpected and worth noting. The engine's need to run background processes and modify its own agent config is reasonable for this purpose.
Instruction Scope
SKILL.md instructs editing ~/.openclaw/openclaw.json, creating crontab entries, running node on explicit absolute paths (e.g., /home/ubutu/.openclaw/...), and running background nohup processes. Those actions give the skill persistent runtime presence and cause it to execute bundled JS code on the host. Additionally, the pre-scan flagged a 'system-prompt-override' pattern in SKILL.md — instructions may attempt to influence model/system prompts. These instructions go beyond a simple helper and warrant manual code review before enabling.
Install Mechanism
No install specification is provided (the registry metadata says 'instruction-only'), yet the package includes many code files and scripts meant to be executed. The absence of an install step means the skill expects manual git/npm steps and direct node execution; that's not inherently malicious but increases the risk that bundled code will be executed with host privileges without an automated vetted install process. No remote download URLs were present in the SKILL.md, which reduces one class of risk.
Credentials
The skill does not request environment variables or external credentials, which aligns with its stated purpose. However the runtime instructions write and read host-specific paths (e.g., ~/.openclaw/openclaw.json, /home/ubutu/.openclaw/..., /workspace/logs/...) and add crontab entries. Those file and crontab modifications are proportionate for a service but still grant persistent access to logs/configs — review whether those paths overlap with sensitive data on your host.
Persistence & Privilege
The skill does not set always:true and does not declare extra credentials, but the SKILL.md explicitly directs enabling the skill in agent config and adding crontab/nohup entries to run the engine continuously. That grants long-term execution capability on the host (normal for a workflow engine), so only proceed after code review. This persistent presence combined with the prompt-injection signal increases risk.
如何使用
- 确保已安装 OpenClaw(本地或 Docker 部署)
- 在对话框中输入安装命令:
/install agile-workflow - 安装完成后,直接呼叫该 Skill 的名称或使用
/agile-workflow触发 - 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
v7.18.1
agile-workflow v7.18.1 更新日志
- 全面并发安全架构升级,彻底消除多 Agent 场景下的数据污染
- 新一代任务依赖管理系统,支持 DAG+环路检测和增强型冲突检测
- 智能任务拆解优化:细纲任务串行、正文任务并行,提升分工效率
- 自动学习与持续优化能力增强,根据任务执行经验生成配置建议
- 实时监控和报警机制升级,包括秒级状态更新和详细性能指标
- 丰富命令行管理与故障排查指引,提升易用性和自动化运维能力
元数据
常见问题
Agile Workflow 是什么?
全自动敏捷协作工作流引擎 v7.18.1 - 细纲串行+正文并行、并发安全架构、零数据污染、智能合并. 它是一个面向 Claude Code / OpenClaw 的 AI Agent Skill 插件,目前累计下载 347 次。
如何安装 Agile Workflow?
在 OpenClaw 或 Claude Code 对话框中运行命令「/install agile-workflow」即可一键安装,无需额外配置。
Agile Workflow 是免费的吗?
是的,Agile Workflow 完全免费,采用 MIT-0 许可证,可自由下载、安装和使用。
Agile Workflow 支持哪些平台?
Agile Workflow 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 Agile Workflow?
由 XBCS(@xbcs)开发并维护,当前版本 v7.18.1。
推荐 Skills