What it is
Executable benchmark of web-app feature restoration tasks for coding agents, with masked apps, browser traces, verifiers, and reference solutions. 4,063 executable feature-restoration tasks across 26 web apps, plus a 300-task public subset and 12 full reconstructions; each task bundles masked source, reference app, replayable browser traces, verifier, and solution.
Key results
- 4,063 feature-restoration tasks across 26 apps
- 300-task public evaluation subset
- MIT benchmark artifacts; upstream app licenses also apply
Useful for benchmarking coding agents that edit web app code and validate fixes via browser-trace replay.