Local LLM engine where a tool call cut by the token limit still parses