Service knowledge graphs (SKGs) have emerged as a crucial framework for capturing the temporal evolution of entities and their dynamic interactions within service-oriented environments. Although current SKG reasoning approaches excel at modeling structural and temporal dependencies from historical observations, they often struggle in cases involving sparse or previously unobserved interactions, mainly due to their limited capacity to handle uncertainty in temporal patterns. To tackle this challenge, recent efforts have explored diffusion-based generative models, showing notable progress. However, these models usually apply uniform denoising strategies across all timesteps, which can result in semantic misalignment during the generation process. Additionally, their unsupervised design often fails to preserve critical temporal and relational semantics inherent to service ecosystems. To overcome these limitations, we introduce GCLP, a novel generative framework that combines adaptive prompt-driven diffusion with historical contrastive learning. In GCLP, adaptive prompts are injected at each denoising step, enabling dynamic guidance of the generation process according to the shifting semantics of services. Simultaneously, contrastive supervision grounded in historical service interaction traces is employed to strengthen semantic consistency throughout the diffusion sequence. Comprehensive evaluations on four real-world knowledge graph datasets show that GCLP surpasses state-of-the-art baselines across multiple metrics, demonstrating its superior ability to model intricate temporal dynamics and enhance temporal reasoning in intelligent service systems. The anonymized code repository is accessible at https://anonymous.4open.science/status/GCLP .