Skill flagged — suspicious patterns detected

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

long-running-agent

v1.0.0

长时间运行智能体编排框架。让AI智能体能够跨会话持续工作并完成复杂任务。 使用场景: - 用户说 "创建一个新项目"、"开始一个项目"、"新建任务"、"启动项目" - 用户说 "继续项目"、"继续工作"、"接着做"、"继续开发" - 用户说 "更新进度"、"记录进度"、"保存进度" - 用户说 "项目状态"、"查...

1· 224·0 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 sleepingzzzz/long-running-agent.

Previewing Install & Setup.
Prompt PreviewInstall & Setup
Install the skill "long-running-agent" (sleepingzzzz/long-running-agent) from ClawHub.
Skill page: https://clawhub.ai/sleepingzzzz/long-running-agent
Keep the work scoped to this skill only.
After install, inspect the skill metadata and help me finish setup.
Required binaries: git
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 long-running-agent

ClawHub CLI

Package manager switcher

npx clawhub@latest install long-running-agent
Security Scan
VirusTotalVirusTotal
Suspicious
View report →
OpenClawOpenClaw
Benign
high confidence
Purpose & Capability
Name/description describe a long-running agent that uses the filesystem for cross-session memory; the skill only requires git (reasonable given the documented Git coordination) and includes small helper scripts to initialize and list projects. The requested binaries, file layout, and scripts are proportionate to the described functionality.
Instruction Scope
SKILL.md instructs the agent to read/write workspace/tasks/{project} and memory/*.md files (expected). It also instructs running quick tests (e.g., npm test / test preflight) and using a 'Ralph Loop' that may iterate work until tests pass — running tests can execute arbitrary code present in the workspace. Recommend confirming or sandboxing any test/command execution. The instructions do not attempt to read unrelated system paths or external secrets.
Install Mechanism
There is no install spec (instruction-only) and included JS scripts are small and local. Nothing is downloaded from external URLs or written to unexpected system locations.
Credentials
The skill declares no required environment variables or credentials. It does require the git binary which matches its described Git coordination behavior. No unrelated secrets or config paths are requested.
Persistence & Privilege
always:false (default) and disable-model-invocation:false (normal). The skill writes/reads files under the workspace/tasks and memory directories (expected for persistent cross-session state) and does not request elevated system-wide privileges or modify other skills' configurations.
Assessment
This skill is internally consistent with its purpose: it creates and manages project folders and uses local files (PROJECT.md, CHANGELOG.md) as long-term memory. Things to consider before installing: 1) It will read and write files under the agent's current workspace (workspace/tasks/ by default) — ensure the agent's working directory is safe and you won't accidentally expose sensitive files. 2) SKILL.md encourages running quick tests and using an iteration loop (Ralph Loop); running tests may execute code in your workspace — review/sandbox any test commands before letting the agent run them autonomously. 3) The scripts assume reasonable file formats (parsing progress, etc.); they are simple but somewhat fragile (e.g., progress parsing). 4) git is required for the documented workflow (commits), so ensure git is present and configured as you expect. If you want stronger guarantees, run the skill in a dedicated workspace or container and review the two included scripts (scripts/init-project.js and scripts/list-projects.js) before enabling autonomous execution.

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

Runtime requirements

🔄 Clawdis
Binsgit
latestvk975qxb3403fn3h6zfxdqj45zh8362wj
224downloads
1stars
1versions
Updated 23h ago
v1.0.0
MIT-0

Long-Running Agent - 长时间运行智能体编排框架

基于 Anthropic 论文《Long-Running Claude for Scientific Research》实现,让 AI 智能体能够跨会话持续工作并完成复杂任务。

核心理念

问题:AI智能体每次会话都是"新生的",没有记忆延续。

解决方案:通过文件系统作为"外部大脑",实现跨会话的持续工作能力。


快速开始

创建新项目

当用户说 "创建一个[项目名]项目" 时:

Step 1: 确认项目名称和目标
Step 2: 在 workspace/tasks/{项目名}/ 创建目录
Step 3: 生成 PROJECT.md(项目简报)
Step 4: 生成 CHANGELOG.md(进度追踪)
Step 5: 开始第一项任务

示例对话

用户: 创建一个博客系统项目,用 Node.js 实现

QClaw: ✅ 项目已创建: 博客系统

📁 位置: tasks/博客系统/
📄 PROJECT.md - 项目简报
📄 CHANGELOG.md - 进度追踪

🎯 目标: 用 Node.js 实现博客系统

下一步: 定义详细需求

请告诉我:
1. 博客系统需要哪些功能?
2. 使用什么数据库?

继续项目

当用户说 "继续[项目名]项目" 时:

Step 1: 读取 tasks/{项目名}/PROJECT.md(了解上下文)
Step 2: 读取 tasks/{项目名}/CHANGELOG.md(了解进度)
Step 3: 执行定向协议(PROJECT.md 中定义)
Step 4: 从"下一步"继续工作

示例对话

用户: 继续博客系统项目

QClaw: 📋 继续项目: 博客系统

📊 当前状态: Phase 2 进行中
📈 进度: 5/12 任务,42%

🔄 Ralph Loop: 迭代 4/20

📍 下一步: 实现文章 API

[开始工作...]

查看项目状态

用户: 我有哪些项目?

QClaw: 📋 项目列表 (共 3 个)

名称                    | 状态           | 进度    
------------------------|----------------|---------
博客系统                | Phase 2        | 42%     
用户管理API             | 已完成         | 100%    
数据分析工具            | Phase 1        | 15%     

文件结构

每个项目都有标准结构:

workspace/
├── tasks/
│   └── {项目名}/
│       ├── PROJECT.md      # 项目简报(目标、成功标准、定向协议)
│       ├── CHANGELOG.md    # 进度追踪(跨会话记忆)
│       ├── tests/          # 测试预言机
│       └── src/            # 项目源代码
└── memory/
    └── YYYY-MM-DD.md       # 每日记录

核心组件

1. PROJECT.md - 项目简报

每个项目的"说明书":

# PROJECT.md - {项目名称}

## 项目概述
一句话描述项目做什么

## 交付物
- [ ] 交付物1
- [ ] 交付物2

## 成功标准
可量化的完成定义:
- 所有测试通过
- 测试覆盖率 > 80%
- 功能完整可用

## 定向协议
每个会话开始时执行:
1. 读取 CHANGELOG.md 的"当前状态"和"下一步"
2. 运行快速测试确认无回归
3. 从优先列表选择任务
4. 开始工作

## 不要做的事
- ❌ 跳过定向协议
- ❌ 重新尝试已记录为"失败"的方法
- ❌ 在测试失败时提交代码

2. CHANGELOG.md - 进度追踪

最重要的文件! 智能体的"长期记忆":

# CHANGELOG.md

## 当前状态: Phase 2 进行中
**进度**: 5/12 任务完成,42%
**下一步行动**: 
1. 实现文章 API
2. 编写测试

**Ralph Loop 状态**:
迭代: 4/20
成功标准: 所有测试通过,覆盖率>80%

## 失败的方法
> ⚠️ 最重要!防止重复尝试

### 方法:使用 Python 实现 - 2026-03-19
**尝试**: 使用 Python 创建项目
**结果**: 系统未安装 Python
**教训**: 当前环境优先使用 JavaScript/TypeScript

## 完成的任务
### 2026-03-19: 用户模型实现
- 实现 User 数据类
- 10 个测试用例通过
- 提交: abc123

## 会话日志
### Session 4 - 2026-03-19
**完成**: 用户 API 实现
**下次继续**: 文章 API

3. 测试预言机

让智能体能自主判断是否在进步:

类型说明示例
参考实现已知的正确输出预期输出文件
量化目标明确的数值指标覆盖率 > 80%
测试套件可执行的测试npm test

4. Ralph Loop

解决"智能体惰性"问题——智能体有时会在完成部分任务后找借口停下来。

实现:循环检查,确保真正完成

for i in 1..MAX_ITERATIONS:
    result = agent.work(task)
    if result.includes("DONE") and verify_completion():
        break
    else:
        agent.continue("你确定完成了吗?请验证:{success_criteria}")

在 CHANGELOG.md 中声明

**Ralph Loop 状态**:
迭代: 1/20
成功标准: 所有测试通过,覆盖率>80%

每次会话检查是否达到标准,未达到则继续推进。

5. Git 协调

每个"有意义的工作单元"后提交代码:

一个功能实现 → 一次提交
一个 bug 修复 → 一次提交
一个重构 → 一次提交

命令参考

用户指令触发条件QClaw 行为
创建项目"创建/开始/新建/启动" + 项目名生成 PROJECT.md, CHANGELOG.md,开始第一项任务
继续项目"继续/接着做/继续开发" + 项目名读取进度,从断点继续
更新进度"更新进度/记录进度/保存进度"更新 CHANGELOG.md
项目状态"项目状态/查看进度/做到哪了"显示当前进度和下一步
列出项目"列出项目/所有项目/我有哪些项目"显示所有项目及其状态
暂停项目"暂停项目/存档项目"标记项目为暂停状态

工作规则

必须做

  • ✅ 每个子任务完成后更新 CHANGELOG.md
  • ✅ 先写测试再写代码(TDD)
  • ✅ 记录所有失败的方法(最重要!)
  • ✅ 使用定向协议确保从正确位置开始
  • ✅ 简洁输出(成功5-10行,失败20行)

禁止做

  • ❌ 跳过定向协议
  • ❌ 重新尝试已记录为"失败"的方法
  • ❌ 在测试失败时提交代码
  • ❌ 打印完整数组或日志
  • ❌ 冗长的描述

输出规范

创建项目时

✅ 项目已创建: {项目名}

📁 位置: tasks/{项目名}/
📄 PROJECT.md - 项目简报
📄 CHANGELOG.md - 进度追踪

🎯 目标: {项目目标}
📊 成功标准: {成功标准}

下一步: {第一个任务}

继续项目时

📋 继续项目: {项目名}

📊 当前状态: {状态}
📈 进度: {X/Y 任务,Z%}

🔄 Ralph Loop: 迭代 {N}/20

📍 下一步: {下一个任务}

[开始工作...]

更新进度时

📝 进度已更新

状态: {新状态}
进度: {X/Y → X+1/Y}
完成: {刚完成的任务}
下一步: {新任务}

与 HEARTBEAT 集成

可以将项目检查加入 HEARTBEAT.md,实现后台持续推进:

## 当前 Ralph Loop 任务

项目: 博客系统 迭代: 4/20 成功标准: 所有测试通过


检查:是否需要继续推进?

模板文件

详见 references/ 目录:


脚本工具

init-project.js

初始化新项目:

node scripts/init-project.js <项目名> --desc "项目描述"

list-projects.js

列出所有项目:

node scripts/list-projects.js

错误处理

错误处理方式
项目不存在提示用户创建或列出可用项目
CHANGELOG.md 损坏尝试恢复,无法恢复则提示重新创建
测试失败记录失败原因到 CHANGELOG,继续调试
Git 冲突提示用户手动解决

示例项目

查看 examples/task-planner-demo/ 目录,这是一个完整的示例项目,展示了:

  • ✅ 完整的 PROJECT.md 和 CHANGELOG.md
  • ✅ 失败方法记录(Python → TypeScript 迁移)
  • ✅ Ralph Loop 迭代过程(3次迭代完成任务)
  • ✅ 测试驱动开发流程

参考资料

Comments

Loading comments...