While Large Language Models (LLMs) offer scalable mental health support, they typically operate within the superficial text space, providing generic comfort that fails to resolve a patient’s deep-rooted cognitive distortions. To address this, we propose **MindUntangler** a novel multi-agent framework that models psychological counseling as a dynamic topological routing task. By conceptualizing cognitive distortions as "pathological topologies"—tangible knots in a patient’s Cognitive Topology Graph (CTG)—our system dynamically tracks and resolves these complex connections through pluralistic therapeutic routing. To ensure clinical safety, we integrate topological guardrails and align the model via a novel topology-driven Direct Preference Optimization. Experiments on two benchmarks demonstrate state-of-the-art performance in cognitive restructuring and clinical efficacy. Ultimately, this work shifts AI therapy from textual empathy to computable, structural interventions.