Frankly a fun model.
I really hope you continue working on this idea.
I get that CoT models are generally very bad at long term roleplays and plotlines due to the way they are trained. They tend to over focus on parts of the prompt that no longer matter or are way too focused on the "last input".
But this one is kinda endearing. I like the fact that the CoT is talking from the character's perspective, given a non-shit character card and system prompt, which actually really helps keeping it grounded in the plot. The CoT part is actually fun to read, contrary to pretty much all other models where half of it is spent wasting tokens going "so i'm pretending to be X and the user is asking Y, so i probably should do Z + actual CoT chain".
Does the CoT bit actually helps with the end result's quality? Probably not in proper RP. Over dozens of comparisons between R1 and normal Cydonia the difference is marginal at best. It's more expressive (bolding and italicizing words) which is a double edged sword. But, it's less likely to just copy paste the same response structure over and over again. It still has that issue of, unchecked, writing longer and longer responses, though.
Overall, it's a lot better at 1:1 complex conversations than the non R1 version without a doubt (like most CoT models would be). But for RP it's more of a hit and miss. Sometimes surprisingly good if guided well, but also really off when the user's output is low quality.
Still, I like the approach. Good job :)