TL;DR: KeyFlow adapts a robot policy without retraining: a coding agent picks a few semantic 3D keypoints, writes short programs that turn an episodic memory of ten demonstrations into a flow field the robot follows, and revises both from its own play rollouts. On 13 RoboCasa tasks it beats π0.5 fine-tuned on ten times as many demonstrations (62.6% vs. 47.8%), and in IsaacSim its flow fields raise a base π0.5 policy's success by 37.2 points.