问HN：你是如何使用个人AI助手与本地编码代理的？
嗨，HN，
我最近尝试了一种工作流程，个人AI助手负责任务接收，而编码代理负责实现。
在我的设置中，我使用OpenClaw作为个人助手层：我可以发送语音或文本备注，它会将其转化为我代码库中的结构化Markdown任务卡。然后，我构建的开源本地CLI工具VibeDeck会获取这些任务卡，将编码代理分派到独立的Git工作树中，并让我以拉取请求的形式审查结果。
主要思想是将任务捕获、执行和审查分开：
- OpenClaw负责想法捕获/任务草拟
- VibeDeck负责本地协调
- 编码代理在独立的Git工作树——孤立的沙箱中工作
- 主分支从不直接修改
代码库链接：[https://github.com/quentinzhang/vibedeck](https://github.com/quentinzhang/vibedeck)
我很好奇这里其他人是如何处理任务隔离、上下文交接和本地AI编码代理的分支安全的。
欢迎提问技术相关的问题。
查看原文
Hi HN,<p>I’ve been trying a workflow where a personal AI assistant handles task intake, and coding agents handle implementation.<p>In my setup, I use OpenClaw as the personal assistant layer: I can send it a voice or text note, and it turns that into a structured Markdown task card in my repo. Then VibeDeck, an open source local CLI I built, picks up those task cards, dispatches coding agents into isolated Git worktrees, and lets me review the results as PRs.<p>The main idea is to keep task capture, execution, and review separate:<p>- OpenClaw handles idea capture / task drafting
- VibeDeck handles local orchestration
- coding agents work inside isolated git worktree --orphan sandboxes
- the main branch is never modified directly<p>Repo: https://github.com/quentinzhang/vibedeck<p>Curious how others here are handling task isolation, context handoff, and branch safety for local AI coding agents.<p>Happy to answer technical questions.