Live view recording (#139, shown for #140): a real agent on a Mobilebox phone, watched in the live viewer Captured: 2026-11-05 by verifier/capture-live.ts on a Mobilebox cloud Android 14 phone (arm64, 1080x2400), with the phone look from verifier/phone-look.ts. The agent drives the phone over MCP (the run from verifier/capture-run.ts: open Wikipedia, search "Matterhorn", open the article, scroll to Quick facts) while a browser shows /watch for that phone. One step taps an element that is not on screen ("No such tab"), so the failed state is shown as it happens. Original: live-view-run.mp4 on branch assets/issue-139 (43.2 s, 1180x860, 25 fps, sha256 52c232f549e8a93802aa779116bf20631f28254980682f203b9077a0c3545fc3). live-view.mp4: 2.5 s to 32.0 s of the original, real time, not sped up; re-encoded with ffmpeg -ss 2.5 -to 32 -i live-view-run.mp4 -an -c:v libx264 -preset slow -crf 25 -pix_fmt yuv420p -movflags +faststart sha256 ed94db93913a84760c7fff20676a4aff4653e5f42ac09c9a0a5ee416ae674079 live-view.webm: the same clip in VP9 for browsers without H.264 (ffmpeg -c:v libvpx-vp9 -b:v 0 -crf 38) sha256 2d2aa45acd02f126bb34175e9104ab2e5035e49332fe79090c746009982ad8b9 poster.webp: the clip's last frame (a swipe trail on the article, the failed step in the log). sha256 0c17970c3d619e445a52ab68176b32e825f94c54e8158c46c84ad88a3b859797 Nothing is drawn after capture: the cursor, swipe trail, typed-text bubble, captions and times are the viewer's own.