Cyber-physical systems (CPS) operate in dynamic and uncertain environments, where maintaining operational objectives without manual intervention is critical. Self-adaptive systems (SAS) have emerged as a promising solution, leveraging machine learning (ML) models within feedback control loops to make runtime adaptation decisions. However, the black-box nature of these models poses challenges related to transparency and efficiency, particularly in safety-critical domains. This paper introduces CSA- \(\varPhi \) , a counterfactual-based self-adaptation approach that integrates model-agnostic interpretable ML into the adaptation loop. CSA- \(\varPhi \) includes two main components: an offline pre-processor with pre-trained classifiers for requirement evaluation, and an online opportunistic adaptation engine that uses counterfactual explanations to guide adaptation decisions efficiently. By bridging the gap between explainability and actionable adaptation, CSA- \(\varPhi \) enhances decision-making while reducing adaptation cost. Experimental results show that CSA- \(\varPhi \) achieves significant improvements in execution efficiency and provides adaptation decisions that are both effective and interpretable, outperforming selected baseline approaches.