Job interview simulation with a virtual agents aims at improving people's social skills and supporting professional inclusion. In such simulators, the virtual agent must be capable of representing and reasoning about the user's mental state based on social cues that inform the system about his/her affects and social attitude. In this paper, we propose a formal model of Theory of Mind (ToM) for virtual agent in the context of human-agent interaction that focuses on the affective dimension. It relies on a hybrid ToM that combines the two major paradigms of the domain. Our framework is based on modal logic and inference rules about the mental states, emotions and social relations of both actors. Finally, we present preliminary results regarding the impact of such a model on natural interaction in the context of job interviews simulation.