Pass initial message separately from prompt The llama.cpp logic is built around the prompt ending with the reverse-prompt and the actual user input being passed separately. Adjust Sherpa to do the same, rather than appending the first line of user input to the prompt.