Coding agents now write a real share of the code in a lot of teams. Three parts of that still go wrong often enough that I built an open-source tool for each, and today all three are public: The agen
Markdown supported · ⌘↵ to post
Thanks, that's the reason I added it. To your question: mostly rules onsystem doesn't have, not files it missed. @shadcn/lint scored only the one file each run created, which the hook also checks. Of the 2.64 left per run:
no-restyle: classes that change a design-system component's color, typography or spacing. onsystem doesn't report that, and whether an override is wrong depends on the design system.no-raw-colors, which flags the same kind of error onsystem shows the agent, so on colors it agrees with onsystem more than it confirms it. Most of that 0.60 comes from one model, gemini-3.5-flash-lite.One more caveat: I added this scoring after the standard run, so it wasn't one of the outcomes fixed before any run.
The per-rule and per-model numbers, and the command to rescore the runs yourself (it calls no model), are in the "Independent scorer" section of the benchmark page linked in the post.
Kartik N V J K
AI Developer | Making AI reliable, trustworthy & accessible to everyone | Active community contributor
I trust the 1,200-run result more because you also scored it with @shadcn/lint, which onsystem doesn't control. Going from 4.97 to 2.64 on an outside linter tells me more than the 0.00 from onsystem's own check. What's in the 2.64 that's left: rules onsystem doesn't cover yet, or code the agent wrote outside the files the hook checked?