Kimi K3.1 Pro added

Kimi K3.1 Pro used fewer calls to the browser tool than MiniMax M3.1 Flash on 33 of the same 35 tasks: a median of eight versus 23. That count leaves out an interesting part of Kimi’s process: six sessions also attempted custom Playwright or Node VM checks through the shell.

Platformer Dave is the clearest example. Its browser-tool trace contains just an inspection and one arrow-key input, but separate Playwright scripts drove timed movement and jumps, collected the trophy and reached the exit. Another route reached the exit without the trophy and received the locked-exit message. The scripts also exercised losing and restarting. The saved commands and returned game states make those checks traceable.

Space Invaders shows a concrete revision: after a scripted check returned a zero score, Kimi changed the Space-key handler to fire immediately during play. The next script returned a score of 25, then zero after Restart. Snake took a different route, extracting its JavaScript into a Node VM with mocked browser objects and assertions for turns, growth, collisions and a full-board win. Those checks cover logic rather than rendering.

Maze Generator is useful for following a longer debugging sequence. Temporary logs exposed a negative starting y-coordinate; Kimi changed the vertical offset calculation from the column count to the row count, removed the logs and reloaded. Together, these runs show why browser-tool counts alone miss some of the verification work. They document specific checks and revisions, rather than establish an overall quality ranking.