Skip to Main Content

RepoFuse

Built at Meta Global AI Developer Hackathon · Oct 26, 2026 · Remote

Demo video · github.com/…

Useful open-source repositories are easy to find. Combining them into something that works means understanding APIs, checking licenses and dependencies, writing adapters, and testing the result. RepoFuse is a supervised prototype for that workflow. A builder supplies two public GitHub repository links. RepoFuse pins their commits and checks them, uses Meta models to propose integration ideas, then generates integration logic and tests for a selected idea. Independent source review gates contained verification and local serving. The current successful demonstration combines cachetools and boltons. The model-generated integration passed seven model-written tests, eight independent probes, and real Chrome checks of numeric input, cache configuration changes, repeat hits, invalid input and returning to defaults. That successful integration was not manually repaired. Earlier failed generations remain failures. RepoFuse supplies a fixed HTTP/UI workbench around generated compute logic. It does not generate an arbitrary bespoke frontend. Source review was supervised, not an available automatic judge service. One successful pair does not establish arbitrary-repository compatibility or unattended end-to-end generation. RepoFuse's new workflow was built during the event; upstream repositories predate it and retain their own licenses. The current boundary is selected compatible vendored Python packages and a Linux-contained Python WSGI runtime. This is a local prototype, not a production service or a hosted judge-ready product. The demo distinguishes the earlier notes-app replay, including assistant repairs and extensions, from the later unmodified cachetools-plus-boltons integration. It is not continuous live generation. I didn't have enough time to finish the full workflow during the hackathon. I plan to keep developing RepoFuse with my remaining Muse credits.

Team