<p>As computational thinking (CT) is increasingly recognized as a fundamental competency across disciplines, there is a growing emphasis on integrating it into educational settings for non-computer science (non-CS) majors, with collaborative block-based coding emerging as a promising approach. However, much of the existing research focuses on coding outcomes rather than the underlying cognitive and regulatory processes that unfold during collaboration, particularly across tasks of varying complexity. This study examines how CT behaviors and socially shared metacognitive regulation (SSMR) patterns differ by task difficulty and performance level in collaborative coding contexts. It employed a mixed-methods design, with 120 undergraduates completing basic and advanced coding tasks using App Inventor. Behavioral data on the use of computational concept (CC) and non-CC blocks were analyzed through machine learning-based clustering, while group discourse was examined using lag sequential analysis. Results revealed that high-performing groups, especially in advanced tasks, demonstrated concise block usage and engaged in coherent cycles of evaluation, adaptation, and monitoring. In contrast, low-performing groups demonstrated fragmented regulation and relied heavily on trial-and-error strategies. The findings suggest that analyzing both behavioral patterns and discourse context facilitates more efficient identification of underlying cognitive structures expressed through actions, whereas discourse analysis offers deeper insights into the reasoning processes and socially regulated strategies involved. The study carries practical implications for the design of scaffolding and adaptive feedback tools and contributes theoretically by deepening our understanding of how SSMR functions in collaborative coding education.</p>