Causal video question answering (QA) has garnered increasing interest, yet existing datasets often lack depth in causal reasoning analysis. To address this gap, we capitalize on the unique properties of cartoons and construct CausalChaos!, a novel, challenging causal Why-QA dataset built upon the iconic "Tom and Jerry" cartoon series. With thoughtful questions and multi-level answers, our dataset contains much longer causal chains embedded in dynamic interactions and visuals, at the same time principles of animation allows animators to create well-defined, unambiguous causal relationships. These factors allow models to solve more challenging, yet well-defined causal relationships. We also introduce hard negative mining, including CausalConfusion version. While models perform well, there is much room for improvement, especially, on open-ended answers. We identify more advanced/explicit causal relationship modeling and joint modeling of vision and language as the immediate areas for future efforts to focus upon. Along with the other complementary datasets, our new challenging dataset will pave the way for these developments in the field. We will release our dataset, codes, and models to help future efforts in this domain.