The convergence of networking and computing has become a critical architectural paradigm due to the increasing demand for services offering high availability and ultra-low latency, including immersive video and wide-area high-performance computing. However, increasing heterogeneity in computing resources and tighter coupling with algorithms have exacerbated the mismatch between resource supply and service demands. The imperative necessity for service-aware and resource-adaptive routing solutions is underscored by the fact that traditional location-based routing mechanisms neglect service semantics. Against this backdrop, the Computing Power Network (CPN) has emerged, integrating network and computing to enable dynamic, SLA-driven task orchestration. Although substantial progress has been achieved in task scheduling and orchestration in the existing work, the critical issue of service-level routing has not been completely addressed. To bridge this gap, we propose OSAF, a lightweight and backward-compatible service routing mechanism for CPN. OSAF introduced Service Tag (STag) as location-independent semantic identifier, which decouple service requests from host addresses. This enables users to access integrated services without being aware of resource locations. Our design of Service-Aware Routing (SAR) is based on OSAF. This routing algorithm utilizes reinforcement learning and multi-layer graph attention to model fine-grained resource dependencies and adapt to real-time service dynamics. Extensive evaluations of large-scale, dynamic service scenarios under CPN have shown that OSAF outperforms the current routing mechanisms in terms of SLA fulfillment, scalability, and load balancing.