Luffy_akaza

BugArena turns AI-driven browser exploration into reproducible bug reports and executable regression tests. Developers need more than an AI saying something looks wrong. BugArena connects Meta Muse to Playwright so the model can explore a web interface, choose actions, and report suspected defects. Conventional verification code independently checks the observed behavior against explicit product rules. Our prototype demonstrates this workflow on a controlled shopping app with a deliberately seeded coupon bug. In a verified hosted run, Muse added a notebook, applied SAVE10, and added another notebook. The discount remained $2.50 instead of increasing to $5.00, producing a $47.50 total instead of $45.00. BugArena reproduced the discrepancy in a fresh browser and replayed the same actions against a prewritten corrected version, where the amount checks passed. It generated reproduction steps, expected-versus-actual values, screenshots, a browser trace, and a downloadable Playwright regression test. A separate scripted suite checks ten scenarios against both versions. All ten corrected-shop cases passed; five buggy-shop scenarios exposed the same underlying coupon defect. These scripted checks are clearly separated from autonomous AI exploration. The dashboard provides run history, progress logs, evidence inspection, and downloads. The public demo runs on Render, with open evidence viewing and protected access for launching tests. Built with Node.js, Playwright, the Meta Model API, and Docker. The current prototype supports this controlled shop rather than arbitrary websites. Development used AI coding assistance. Live demo: https://bugarena-96ki.onrender.com