While multi-device cooperative task-oriented semantic communication (TOSC) enhances task performance through comprehensive information representation, it inevitably introduces redundancy, thereby increasing communication overhead. Existing redundancy elimination methods suffer from limitations in interpretability and coarse granularity, hindering the optimal utilization of communication resources. To this end, we propose a disentangled information bottleneck guided TOSC framework (DisenIB-TOSC). The framework first employs the basic IB for initial feature compression, then formulates a novel DisenIB specifically designed for multi-device cooperation inference, which enhances task performance and achieves interpretable feature disentanglement by separating features into common and private components, thereby establishing a theoretical foundation for redundancy identification. Subsequently, we derive differentiable, computationally tractable forms for both IB objectives by combining variational approximation, consistency constraints, and density ratio trick. Leveraging the disentangled features, we further design a feature importance-aware selective transmission strategy, DisenIB-TOSC-ST, which quantifies feature importance via mutual information estimation to dynamically discriminate and control redundant feature transmission. Experimental results on several tasks demonstrate that our method outperforms baselines in task performance while reducing communication costs, verifying the effectiveness and interpretability of feature disentanglement.