Inspiration
A code repair can look convincing while still breaking existing behavior. We built ReproLab to make a proposed fix easier to inspect: connect the model’s work to the tests it passed, the environment where it ran, and the exact patch a developer can review.
What it does
ReproLab runs a scoped repair workflow for a Preact ref-cleanup feature. NVIDIA Nemotron generates candidates through Nebius Token Factory, and candidate code runs in an isolated Nebius Sandbox. The workflow records rejected candidates, test results, model and execution provenance, and the tested patch. Reviewers can inspect and download that patch from the app.
How we built it
We published a source-built judge client that connects to our existing authenticated backend. The demonstrated run used three Token Factory calls and 14 Sandbox operations. It passed 29/29 native Mocha checks, 3/3 public checks, and 10/10 independent checks. The browser suite recorded 1,173 passes with 11 original skips; two type checks remained pending. These are results for this run, not a measure of general repository accuracy.
Challenges and what we learned
Reliable repair requires more than a plausible diff. We had to preserve failed attempts, separate candidate execution from the controller, handle uncertain execution without blindly retrying, and make the downloaded patch match the saved result. We also learned to state the project’s limits clearly: this demonstration covers one exposed feature with supplied scaffolding and a cached image. The published client depends on our hosted backend and does not recreate its private state.
What’s next
We want to test the workflow across more repositories and feature types, improve independent evaluation, and make setup less dependent on the hosted backend.
Built With
- nebius-sandbox
- nebius-token-factory
- node.js
- nvidia-nemotron
- preact
- typescript
Log in or sign up for Devpost to join the conversation.