Nopus Catches Your Coding Agent When It Starts Talking Like a Research Paper

Coding agents have a failure mode where a simple answer turns into dense academic prose about "verification-feasibility signals" and "proportionate reliable evidence." Nopus is an open source tool that runs deterministic readability checks on every coding agent response and automatically forces a rewrite in plain English when the text crosses your complexity threshold. The video is a 31-second walkthrough of the GitHub README for vistyy/nopus, and the project is worth a look if you spend real time decoding what your agent just told you.

The Problem It Targets

The narration describes a specific annoyance: coding agent responses that "slip into dense, abstract, academic language." You ask for a change, and the agent replies with walls of jargon you have to translate back into concrete actions. The usual fix is rephrasing your prompt or asking "can you say that simply," which burns a turn every time. Nopus removes that manual step by intercepting the response before you read it.

How It Works

The README shown in the video states the mechanism plainly: "nopus catches coding-agent answers that disappear into abstract LLM babble. Using deterministic prose checks, it flags responses that cross your chosen complexity threshold and sends them back for one clearer rewrite." The key word is deterministic. It is not another LLM judging an LLM. It measures the prose itself, flags failures, and requests exactly one rewrite, so it does not spiral into repeated regeneration.

What the Rewrite Looks Like

The "See it in action" screenshot in the README shows a before-and-after example. The original response talks about treating "unreliable or disproportionate Evidence as a verification-feasibility signal" that "requires reconsidering the Evidence, Claim, and Required Seam before adding machinery." Nopus tags three issue types in that text: heavy abstract wording (dense language that makes meaning hard to grasp), overloaded phrases (a term like "verification-feasibility signal" compressing too many ideas into one label), and buried instructions (the key action hidden inside a long sentence). The rewrite says the same thing directly: when evidence is unreliable or too costly, reconsider before adding machinery, and escalate to the responsible authority if no practical evidence can prove the claim.

Tunable Sensitivity

The video notes you can adjust how aggressive the check is, "ranging from catching only the worst offenders to demanding simple answers every time." The repo documents three levels with measured rewrite rates: low triggers on about 5.3% of responses, medium (the default) on about 9.9%, and high on about 18.6%. You set the level with a command like /nopus-configure medium. That gives you a concrete dial between occasional intervention and a standing plain-language requirement.

Setup and Support

The README badges in the frames show npm v1.1.0, Node 22 or newer required, and an MIT license. It currently supports three agents: Pi, Claude Code, and Codex. For Pi you install it with pi install npm:@syzom/nopus. For Claude Code and Codex you add the GitHub marketplace and run /plugin install nopus@nopus. The repo sits at 276 stars as of this writing.

Key Takeaways

• Nopus intercepts coding agent responses and forces a plain-English rewrite when readability checks fail
• The checks are deterministic prose measurements, not a second LLM, and each flagged response gets exactly one rewrite
• It flags three issue types: heavy abstract wording, overloaded phrases, and buried instructions
• Three sensitivity levels control rewrite rates, from about 5.3% (low) to 18.6% (high), with medium at 9.9% as the default
• Works with Pi, Claude Code, and Codex; requires Node 22+, MIT licensed, installed via npm or agent plugin marketplace

Resources

• vistyy/nopus -- open source readability guard for coding agent responses, MIT licensed, supports Pi, Claude Code, and Codex