// OVERVIEW
A platform for reproducing bugs across any language and any environment, from a single browser tab.
// 01 · WHAT IT GIVES YOU
Vivarium is a platform for reproducing bugs. Open a page, click once, and find out whether a previously-reported bug still reproduces — or whether a candidate fix actually fixes it.
For each upstream bug, the catalogue holds one recipe.
Every recipe satisfies the same verdict.json v1 contract,
so each page returns either reproduced (the bug
reproduces) or unreproduced (it does not). Results show
inline on the page and are also exposed as machine-readable JSON.
Vivarium ships bugs as runnable artefacts, not write-ups. Instead of someone telling you "this is a bug," you open it in your browser and watch.
// 02 · THREE EXECUTION SURFACES
Bugs differ wildly in what environment they need. Vivarium organises reproduction techniques into three layers, and you don't pick the layer — the recipe declares it for you.
The reproduction runs directly in your browser via WebAssembly. Startup is milliseconds to a few seconds. Right for algorithms, parsers, data processing, and in-memory database operations — bugs that don't need a full OS.
The reproduction runs inside Docker or a microVM. Startup is seconds to a minute. Right for bugs that depend on a real filesystem, real processes, or real networking.
Record-replay, deterministic simulation, WASI Preview 3+, snapshot-based restore — and techniques that don't exist yet. For bugs that L1 and L2 can't reach.
// 03 · HOW WE GOT HERE
The trigger was the surge of AI-generated bug reports and pull requests. Maintainers were drowning in plausible-but-unverified claims arriving at machine speed; contributors had no way to check whether their own AI-drafted fixes actually worked.
The shared root cause is that there is no cheap, universal way to verify whether a bug is real or whether a fix actually fixes it. Vivarium provides one — "does this reproduce?" reduced to a single click.