Cursor released Composer 2.5 on its platform, built on Moonshot Kimi K2.5 with targeted RL using textual feedback and 25x more synthetic tasks than version 2 to improve long-horizon performance and instruction following. https://cursor.com/blog/composer-2-5. The update addresses credit assignment in multi-turn rollouts by inserting localized hints during training, a method that refines specific behaviors such as tool selection without altering global rewards. https://cursor.com/blog/composer-2-5. Composer 2.5 training also incorporates dynamic task generation grounded in real codebases, including feature deletion exercises that maintain test coverage. https://cursor.com/blog/composer-2-5.