When I ran the first version of my static UI linter against OpenCode-App (536 .tsx files on my machine), two things broke immediately:
Code blocks and syntax themes triggered false positives. Scanning raw file strings flagged <pre><code> documentation blocks for sample href="#" tags and terminal ANSI color maps for indigo tokens.
Global weight subtraction punished large repos. Subtracting raw rule weights without normalizing for file count gave a 536-file desktop app a 68/100 (Grade C) when 99.1% of its files had zero findings (only 2 dialog components used indigo utilities and 3 routes had an em-dash).
After masking out comments, <pre> , <code> , and <svg> blocks while preserving exact \n line numbers, and normalizing cosmetic penalties by sqrt(4 / filesScanned) , I published uislop ( npx uislop ) to catch 16 deterministic "vibecoded UI" tells across React, Next.js, Vue, Svelte, Tailwind, and HTML.
npx uislop ./src
Enter fullscreen mode
Exit fullscreen mode
Real benchmark across 687 UI files
Target…