As someone who is using agents to code for me a lot lately, I was starting to feel more and more like I'm losing touch with the code I write--it's starting to feel like Claude's code rather than mine. I'm pretty sure I'm not alone in feeling this way, so I made Rudder as a first earnest shot at trying to fix this problem.

Nowadays with AI coding agents, unit tests are pretty much dead code--your agent may write them because it has to hit coverage, but it is basically meaningless because it's just checking against its own implementation. So, I figured I would repurpose unit tests back to their actual core purpose: making sure that your code does what you want it to do.

Rudder is a local Codex/Claude Code plugin that captures your coding session history and forces your agent to rewrite unit tests exclusively from the intent expressed in your prompts, turning your unit test coverage into a proxy for what percent of the code actually comes from decisions you've made talking to your agent. After showing you an initial coverage percentage, Rudder prompts you with targeted questions to increase coverage to your goal target, and runs a red-green test-driven-development flow to get you up to your goal. Hitting coverage is the equivalent of saying "My decisions in chat conversation cover X% of the code here."

There's a few other ideas out there for getting agents to code closer to the user's intent including some spec-driven-development tools, etc. But I was pretty unsatisfied with the idea that I'd have to completely change my workflow to be able to make decisions when coding with AI. So, I tried to make Rudder as unintrusive as possible by embedding directly into your existing coding sessions.

Install the plugin into every coding agent you want to track and it will work cross-platform (i.e. if you code in the same worktree with both Codex and Claude Code, Rudder will be able to consolidate the prompts from both by tracking the working branch). When you want to use it, just tell your agent "Run Rudder" and it will kick off the flow.

It's Apache-2.0 licensed, because that seems popular nowadays. I'm actively improving this day-by-day for my own use, so let me know what you think! Especially curious how you guys go about making your agent's code your own and verifying that your actual intent makes it into the code your agents generate.

Check it out here: https://github.com/RudderCode/Rudder

Show HN: I Repurposed Unit Tests to Show How Much Coding Agents "Improvise"