Traditional smart home systems usually rely on users to manually define numerous rules to enable devices to respond proactively. However, when facing new scenarios or devices without pre-set rules, these systems fall into a “response blind spot”. To address this pain point, this study proposes an end-to-end proactive responsive smart home system architecture based on large language models. Through an innovative structured event representation mechanism, this system transforms sensor-perceived data into a semantically understandable text event stream. By combining customized prompt template design with a commonsense reasoning framework, it constructs a full-closed-loop intelligent decision-making chain from environmental perception, semantic understanding to device linkage. Experimental data shows that in multi-scenario mixed tests, the device response accuracy of the system reaches 83%, and the average response delay is controlled within 7 s. This solution breaks through the scene adaptation bottleneck of traditional rule engines. By leveraging the generalization and reasoning capabilities of large language models, the system is empowered with the ability to actively understand the semantics of complex environments, providing a solution that combines engineering feasibility and technological foresight for the paradigm shift of smart homes from “passive execution” to “active decision-making”.