Game Terminal-Bench grades coding agents on whether the game actually runs, using a Harbor-compatible Godot task contract
GitHub·medium signal
GamePhanes reached 224 stars since 2026-08-21 with a benchmark that refuses to stop at files and exit codes. Its loop is inspect, edit, run, observe, diagnose, repair, verify, and the evaluator drives controlled input probes against the running project to confirm runtime state changed. Only one complete task is published so far, repair-neon-relay-jump, which asks an agent to fix a broken phase jump in a Godot platformer without breaking shard collection.