Multi-behavior recommendations (MBRs) aim to enhance recommendation performance with multi-typed user-item interactions. This paper approaches MBR from a causal perspective, treating the predictions of MBR as outcomes, given various user behavioral data as treatments. However, with the incorporation of additional user behaviors, MBR becomes more vulnerable to including spurious correlations caused by unobserved confounders. Addressing such unobserved confounding effects with the current methods of frontdoor adjustment and proxy variables poses practical challenges in real-world MBRs. To solve these practical challenges, we debias the negative effects of unobserved confounders with stable counterfactual reasoning, which models the stable trend within the stratum of users and is enhanced with counterfactual examples. Specifically, we propose a counterfactual-enhanced multi-behavior recommender (C-MBR), which models user preferences from multi-behavior interactions and provides recommendations via stable counterfactual reasoning. Experiments on two real-world recommendation datasets demonstrate that our C-MBR outperforms baseline models in recommendation performance. The source code is available \(^1\) ( https://github.com/s1ruihuang/c-mbr ).