A player that won’t jump might have perfectly reasonable GDScript. The input action could be missing, the script could be attached to the wrong node, or the collision shape could be catching on the floor.
Reading source files gets an agent part of the way. A Godot harness also needs to inspect scenes, node paths, signal connections and resources. I want it to answer “where is this script used?” before it rewrites the script.
Then it needs to run the scene. Capture the errors, inspect the live tree and take a frame. For movement bugs, a short recording may be more useful than a screenshot. The next edit should follow from that evidence rather than another guess about the code.
I’m experimenting with this kind of harness because game engines make the limits of text-only tooling obvious. A change can compile and still feel wrong. Animation timing, camera movement and controls need to be judged in motion.
I also want ordinary Godot files at the end. I should be able to open the project without the agent, inspect its changes in the editor, and carry on working.