How many tokens was that before it diverged from your original request? If it filled the context window, the original prompt would no longer be available to the model and it would start predicting just based on the repeated word. That certainly looks like what's happened here, although I haven't checked what model context size you used or how long the response is.
The question is not how many prior messages there were in the conversation. That's irrelevant for context window purposes. The question is how many tokens there are between the prompt and the point where the reply breaks down.
