Large Language Models (LLMs) have demonstrated great potential in robotic task and motion planning (TAMP). However, existing works rarely address tasks that demand both goal achievement and strict compliance with task rules, posing greater challenges for robots. Moreover, LLM-based motion failure correction methods relying solely on textual feedback struggle in complex environments. To address these issues, we propose a novel multi-agent TAMP framework that integrates consensus planning with a cross-layer motion failure correction mechanism. The framework enables hierarchical collaboration of LLM-driven multi-role agents. Specifically, at the task planning layer, multiple agents reach consensus on a semantic task plan through multi-round deliberation, ensuring its accuracy and rule compliance. At the action grounding layer, action sequences with continuous parameters are generated by the specialized agent under the guidance of the semantic plan, effectively bridging the task planning and motion planning layers. Furthermore, a cross-layer correction mechanism based on a visual reasoning agent enhances the ability to handle motion failures. Extensive experiments show that our framework significantly outperforms baselines in task success rate, rule compliance, and motion planning efficiency and robustness.