- SignalDesk3小时前
Original Summary
Agents ignore AGENTS.md mid-session.<p>Then you tell them: hey, this is written in AGENTS.md.<p>- "You're absolutely right!" (proceeds with fixing)<p>But not anymore! This is a tool that would check the code immediately with a classifier (Jev etc.) and give the agent advice in the same turn.<p>It catches some mistakes that are difficult to put into lint rules and excels at the cases when semantic understanding is required - e.g. field names/types in a nested structure that contradict each other (more commonly known as "invalid states") (that we strive to make, e.g. "unrepresentable", some would say)<p>Even reviewer agent loops won't do much (as Primeagen's X post a month ago shows in all the colours [which I don't know if I can or can not link, and will stay on the safe side of not doing so]). This may or may not have an influence in nerd sniping me into writing that tool.<p>In cases like these, Hapsland (or its approach, at least) comes to help.<p>I know, I know, Abide (look it up!) exists, but I wanted (a) semantic expansion of code (b) full control over what's sent to Jev (hey, their ToS aren't that privacy-friendly) (c) less footprint on the file system. So I made it compatible. In case you want to use both, I mean.<p>TLDR: catch a change with a coding agent hook => expand it into semantic code structure, respecting privacy settings => call Jev with your own code semantic rules => deliver feedback to the agent in the same turn, in 0.5s.<p>P.S. This is a repost from last week when I was being a genius; I forgot to add "Show HN:" to the title, and then 30m passed and I couldn't delete that post anymore; renaming didn't put it into Show HN either.
- 情报分类：技术学习与提效
- 分类依据：内容涉及技术、AI、软件工具或工程实践
- 信息来源：Hacker News 新项目
- 发布时间：2026/10/13 23:19:32
- 暂无回复