WIGSS

WIGSS v0.1.4 went up on npm on April 3, 2026, the first release of it we have written up. It is a command-line tool for the part of frontend work that coding agents are worst at, which is not producing the page but nudging it afterwards.

What it does

You point it at a dev server you already have running, drag a component on the live page, and WIGSS writes the change back into the source file that produced it. It reads which of five CSS strategies your project uses and writes in that one, so the diff looks like something you would have typed.

Apache 2.0, Node 18 or newer, React and Next.js only. That last one is a real limit, and it is one of several the tech report sets out in full.

Get it

Nothing is added to your dependency list. Start your dev server the way you always do, then open a second terminal:

npx wigss@latest --port <your-port-number>

That starts the editor on port 4000 with your page loaded inside it. The --wigss-port flag moves the editor if 4000 is taken, and npx wigss@latest --demo runs the whole thing against a built-in demo page if you have no project handy. WIGSS reads OPENAI_API_KEY, takes --key, or asks for a key on first run. If you installed wigss globally at some point, uninstall it first, or npx will keep serving the old copy. The npm page shows its usual npm i wigss install box; ignore it, this is a CLI you run, not a dependency you add.

What shipped

• The wigss CLI at v0.1.4, published to npm on April 3, 2026. It is the fifth release of the package: 0.1.0 through 0.1.4.
• Five CSS strategies, detected from your project rather than configured: Tailwind utility classes, CSS Modules, plain CSS or SCSS, an HTML file with a linked stylesheet, and inline React styles as the universal fallback.
• Save-time verification. Every save produces a rollback token and a set of fidelity expectations, then compares the post-apply DOM against what you designed within a two-pixel tolerance. A mismatch is surfaced as a warning with a one-click rollback, not reported as a successful refactor.
• Apache 2.0, and the whole source on GitHub. 222 kB unpacked, 40 files, 17 dependencies.

What is not in it, because an omission found later reads as a claim withdrawn. There is no model and no hosted service: the component detection, the design suggestions, the drag feedback and the chat all call GPT-4o with your own OpenAI key, and without a key those four do not run. The refactoring is the half that never calls a model at all. Strategy detection, the Tailwind class mapping, the PostCSS work on CSS files and the Babel work on inline styles are ordinary code, which is the reason a save is reproducible and a suggestion is not.

Read the report

How a DOM node is mapped back to the source file that produced it, and how one apply path stays honest across five very different CSS strategies, is in the WIGSS tech report. The CLI flags, the architecture and the current limitations are in the README.