Recommender systems are extensively used across various real-world applications, but they often struggle with the persistent challenge of the user cold-start problem. Cross-domain recommendation (CDR), which leverages user interactions observed in one domain and transfers them to improve prediction performance in another domain, has emerged as a promising solution. However, we argue that users who share similar preferences in the source domain may have different interests in the target domain. Therefore, directly transferring embedding will introduce irrelevant source-domain collaborative information. In this paper, we introduce a novel graph-based disentangled contrastive learning framework to capture fine-grained user intent and filter out irrelevant collaborative information to avoid negative transfer. Specifically, for each domain, we first utilize a multi-channel graph encoder to learn diverse user intent. Then, we construct the affinity graph in the embedding space and perform multi-step random walks to learn the high-order user similarity relationships. Based on that, we treat one of the domains as the target domain and propose a disentangled intent-wise contrastive learning approach guided by the user similarity to retrain the user rationale intent bridging two domains. Extensive experiments on four commonly used CDR datasets show that DisCo consistently outperforms existing state-of-the-art baselines and verifies the effectiveness of each proposed module.