long-running-harness

What This Skill Does

The long-running-harness skill implements a robust framework for managing complex, multi-session software projects within the OpenClaw environment. Inspired by Anthropic's 'Effective Harnesses for Long-Running Agents' methodology, this skill enforces a structured, file-based approach to state management. Instead of relying on volatile AI memory, it mandates a directory-based architecture including project definitions (PROJECT.md), progress logs (progress.md), and granular task tracking (features.json). By enforcing specific startup and teardown routines for every session, it ensures that project state remains consistent, rollbacks are possible, and complex goals are broken down into manageable, verifiable increments.

Installation

To install the skill in your OpenClaw environment, execute the following command in your terminal: clawhub install openclaw/skills/skills/aowind/long-running-harness

Use Cases

• Software Development: Managing the lifecycle of a new application from scaffolding to deployment.
• Research Projects: Tracking long-term data gathering or documentation tasks that span multiple sessions.
• Refactoring: Executing systematic code improvements where state verification is critical after every change.
• Team Handoffs: Ensuring that a new agent session can immediately pick up where the previous one left off without needing a full summary of history.

Example Prompts

1. "启动项目：创建一个名为 'web-scraper' 的新项目，目标是抓取特定网站的数据并保存为 CSV，使用 Python 技术栈。"
2. "继续开发：我已经完成了基础环境搭建，请查看 features.json，并执行下一个高优先级任务。"
3. "项目进度：请生成当前项目 'web-scraper' 的进度报告，并告诉我目前还有哪些高优先级的开发任务待处理。"

Tips & Limitations

• Incremental Progress: The golden rule is to complete one small feature per session. Do not attempt large-scale refactors in a single pass.
• Verification is Non-negotiable: Never mark a feature as passes: true in features.json without running the associated tests or verifying the output.
• Persistence: Always check progress.md before starting a session to understand the context of previous failures or partial successes.
• Limitations: This skill requires a disciplined file system approach. It is not suitable for ephemeral tasks or quick, one-off queries. If you fail to maintain the file structure, the framework's ability to 'resume' state will be compromised.