VoidZero

Before agents, bundlers made sites faster mostly through static analysis: tree shaking and minification ship less code. That still matters. But bundle size is only a proxy. Users feel how soon the page shows up: FCP (first content on screen) and LCP (largest content on screen, a Core Web Vital). The usual fix is to load only the code the first paint needs, and lazy-load the rest. Which code that is depends on what runs in the browser, and static analysis can't see it. By hand, that means maintaining a measuring harness, reading browser coverage, and testing every change against noise. It takes hours or days per fix, so speed work became a separate phase after deploy.

In the next version of Rolldown The cycle check in chunk merging is now incremental: it builds the chunk graph once, not once per merge O(merges × graph) → O(graph + merges × reach) Apps with thousands of dynamic imports build up to 37× faster

In the next version of Rolldown Plugins reading `chunk.code` or `renderedLength` no longer copy the chunk code from Rust to JS on every read The bigger the bundle and the more plugins read `chunk.code`, the more you save