GOKU

Bug reports are often just a screen recording, and someone has to watch it, guess the steps and reproduce the bug before any fix can start. Bug2PR takes any .mp4 of a web bug and returns a verified pull request: 1. Understand: SAM 3.1 tracks the cursor in every frame; Muse Spark turns the video into repro steps, and every click must cite a SAM-tracked interaction, so it cannot invent steps. 2. Reproduce: Muse Spark computer use drives a real Chromium until it sees the bug itself; each action is saved as a stable locator. 3. Failing test: a Playwright test is generated from those actions and kept only if it fails on the bug at the expected assertion, 3/3 runs. 4. Fix: Muse Wonder edits the code; the fix is accepted only if the new test passes, the repo's own checks stay green and no test file is touched. 5. Pull request: branch, evidence and PR description are prepared locally; nothing is pushed until a human approves. Results on 3 seeded bugs + 1 real Actual Budget bug (#8475): 95.5% step recall with 0 invented steps, 4/4 reproduced live, 0 wrong tests accepted, 3/3 fixed on the first attempt with both Muse Wonder and open-weight Muse Spark Open, raw video → verified PR in 129–159 s.