Game Terminal-Bench judges fixes by running the game

Game Terminal-Bench grades coding agents by running their fix and prodding the live game with input, per the GamePhanes repository.

Most agent benchmarks stop at file diffs and a clean exit code. This one drives the running program instead, then checks that its state changed, a closer proxy for whether a fix shows up in play.

The evaluator's loop covers inspecting the code, editing it, running the game, and diagnosing what broke before it repairs and verifies the fix. Only after that does it send input probes into the running project to confirm the change took hold.

Anyone testing coding agents against interactive software hits the same gap. A suite that only checks compilation doesn't catch what breaks once a player starts pressing buttons.

So far one task is public. It asks an agent to fix a broken phase jump in a Godot platformer without breaking shard collection elsewhere in the level.