Clothes-Changing person re-identification (CC Re-ID) attempts to cross-link the query of interest across diverse physical locations over an extended time span, e.g., covering days, weeks, and months, and thus unavoidably includes clothes-changing situations. For example, criminals routinely try to change their attire randomly at different times and places to alter their appearance between camera captures to evade being identified and tracked. Besides, clothing inconsistency unavoidably arises in other sectors such as healthcare, retail, and smart city crowd management, where reliable identity tracking is crucial. Therefore, existing state-of-the-art Re-ID systems that only leverage the unchanged clothing information are not adequate enough, disrupting the re-id acts and dropping great challenges for person Re-ID under clothing-changing scenarios. In this work, we introduce an effective feature-driven disentanglement approach for dissecting human component regions, which can eliminate the negative impact of clothing from identity-inherent clues. To improve the discriminativeness of the mined features, we explore a dynamic training process to seamlessly integrate multi-task losses, allowing for smooth extraction of relevant shared information among them. The experimental outcomes on the PRCC dataset highlight the supremacy of our proposed approach, outperforming other benchmark Re-ID methods by over 1.1% in mAP under the clothes-consistent setting, and by 0.2% in Top-1 and 4% in mAP under the clothes-changing setting. The source code is available at https://github.com/raisa1921/UISAD-CCReID .