Surface: real web app, Linux, Chromium preview browser, light theme, 1280 x 800 CSS pixels, 2x device scale.
Each checkout ran T3CODE_DEV_SEED=0 vp run dev, against its own fresh .t3 database. No live database was used.
OpenCode 1.19.5 was added through Settings > Providers with automatic local server startup and the existing authenticated account. The same authenticated model was selected on both sides; its identifier is intentionally private.
The same synthetic repository, outside note, initial commit, and prompt were used on both sides. Seed each checkout once using seed.sh; run reproduce.sh with the checkout and scratch directory as its two arguments, and MODEL supplied in the environment.
Recordings show the second run on each commit. Both first runs had the same outcomes. No approval was granted on base.
The browser opened the returned thread. Recordings start while the turn is active and show base waiting for approval and head completing. Eight seconds of initial page loading were trimmed equally from both recordings.
Captures are cropped to the left 980 x 800 pixels to omit the machine panel. The provider model control is masked. Still images additionally mask elapsed counters. No permission controls, task results, approval requests, or command outputs are changed.
The permission-menu images are 298 x 132 crops from the same respective threads and show Full access selected. Two menu captures taken 500 ms apart were pixel-identical. The head outcome was pixel-identical at 500 ms; base outcome content was pixel-identical after masking its running elapsed counters.
The runtime summary comes from the real server CLI: base has one pending request in full-access, head is completed with zero pending requests in full-access.
AI-assisted; verified with the tests and tools listed above.