To elevate AI tutors from simple chatbots to genuine pedagogical partners, this study introduces a novel diagnosis-driven methodology for adaptive Socratic teaching. Current LLM-based tutors often lack the deep diagnostic capabilities required for authentic Socratic dialogue, resulting in superficial interactions. Our approach overcomes this by tightly coupling a real-time, multi-dimensional student cognitive model with the dialogue generation process. This methodology is implemented within a multi-agent system where: a Planner Agent first dynamically structures learning paths from raw documents; a Diagnoser Agent continuously assesses student mastery to create precise cognitive reports; and a Tutor Agent leverages these reports to conduct targeted, high-quality Socratic questioning. Experimental results demonstrate the superiority of our diagnosis-driven method. In blind comparisons, Our system achieved an average win rate of 82% against baseline systems, showing decisive advantages in "Socratic style" and teaching effectiveness. This study confirms that a deep diagnostic mechanism is the key to unlocking personalized and effective Socratic teaching in AI tutors.