<p>Urban traffic anomalies, including accidents and adverse weather, can cascade into widespread congestion, imposing significant economic and safety costs. This paper presents HTGAT, a schema-guided heterogeneous temporal graph attention network, integrated with proximal policy optimization (PPO) reinforcement learning. This framework optimizes response pathways from pre-identified anomaly seeds to recovery actions. Grounded in a three-layer event schema (anomaly-response-recovery), HTGAT leverages role similarities and location-aware edge weighting to enhance decision-making in dynamic road graphs. Assuming the availability of operational seeds from dispatch logs, our approach focuses on downstream causal optimization rather than raw detection, addressing critical gaps in causal linkage and localization. Evaluations on Beijing traffic logs (real-world end-to-end annotations) and SUMO simulations (controlled experiments) demonstrate a 28–32% reduction in response latency and an 18–22% decrease in recovery MSE—providing operational validation of our approach. On the sensor-only PeMS-Bay benchmark, where labels are artificially constructed proxies under transparent assumptions, HTGAT shows consistent relative improvements in a controlled proxy setting; these results serve as a proof-of-concept for adapting our framework to widely used traffic forecasting datasets, but do not represent operational performance. The proposed framework outperforms baseline models such as T-GCN while maintaining a propagation accuracy above 0.90. By unifying semantic reasoning with spatiotemporal optimization, this work advances the development of resilient traffic management systems.</p>