nopus - deterministically detect and automatically rewrite complex responses
热度趋势
趋势数据积累中
百分比基于当前可用热度信号，而非评论数或独立用户人数。
一款名为 nopus 的新工具已被开发出来，旨在确定性地检测并自动重写编码代理响应中不必要的复杂散文。该工具在默认中等灵敏度下，对 5,337 个已完成的代理响应进行了测试，其中约 10% 的响应触发了该工具。nopus 旨在提供更清晰的重写，目前可用于 Pi、Codex 和 Claude Code，用户也可以选择自行派生版本。
TL;DR: I made nopus , a tool that deterministically detects unnecessarily complex prose in coding-agent responses and automatically asks the same agent for one clearer rewrite.
With recent LLMs (you can probably guess from the name which one caused me to write this plugin), I've had more and more trouble understanding the prose the agents are responding with - long "load-bearing" paragraphs, abstract language, overloaded phrases and all that fun stuff.
My logic is the cognitive capacity should be spent on the underlying problem not trying to decipher the message.
I found a few solutions that pass the response through a smaller model, but I wasn't a fan of the tradeoffs:
- The smaller model may receive only the isolated response which can cause it to simplify or reinterpret the wrong thing
The other option is to have a user-invocable skill you invoke when a response is hard to understand, that works but again, the onus is on you to first read the response and determine whether you even understand it, I also added a skill for that but it's a fallback not the main approach.
Measure those signals, identify the biggest offenders, and give that evidence back to the original agent.
nopus uses lifecycle hooks, checks the prose after the agent finishes responding. If enough signals cross the selected threshold, it asks for one clearer rewrite.
The rewrite happens in the same session, so the original agent keeps the full conversation context.
I ran it against 5,337 completed agent responses. At the default medium sensitivity, it triggered on around 10% of them, which I've evaluated myself and tuned so it doesn't trigger unnecessarily.
The repo has a few more details but I'd love for you to check it out.
Currently available for Pi, Codex and Claude Code (or just slop fork your own lol)