<p>Wi-Fi-based gesture recognition has received considerable attention due to its non-contact nature and privacy-friendly characteristics. However, most existing methods perform well only under specific environments and predefined gestures, and struggle to cope with the complexity and variability of real-world scenarios, particularly when confronted with gestures beyond the predefined. To address these challenges, we propose a meta-learning-based gesture recognition framework Wi-CO. Wi-CO achieves cross-domain open-set gesture recognition by collaboratively optimizing from three aspects: task construction, feature extraction, and classification discrimination. Specifically, we design a dynamic meta-task generation mechanism that constructs diverse tasks during training to simulate environmental variations, thereby enabling improved generalization to unseen domains. By randomly regarding some categories as pseudo-unknown categories during the training process, the model is guided to distinguish between known and unknown gesture categories. To enable domain-invariant feature extraction, we propose a Channel Suppression Enhancement (CSE) mechanism. The CSE mechanism enhances the robustness of feature representation by modeling inter-channel feature covariance, probabilistically suppressing environment-dominated channels while highlighting discriminative features highly correlated with target gestures. Finally, we incorporate an energy-based open-set detection mechanism, coupled with Gradient Consistency Regularization (GCR), to effectively balance the optimization objectives of known gesture classification and unknown gesture rejection. Extensive experiments conducted on three Wi-Fi gesture recognition datasets demonstrate the superior performance of the proposed method in cross-domain open-set scenarios, achieving recognition accuracies of 96.18%, 92.91%, and 90.94%, respectively.</p>