Training Composer for longer horizons · Cursor

By making self-summarization part of Composer's training, we can get training signal from trajectories much longer than the model's max context window.