JiuTian-VL/Optimus-3 ? reverse-engineered prompt

Build me a Minecraft assistant that can run on a server and be controlled from a simple local GUI. I want it to understand what is on screen, answer questions about the world, find objects in view, plan multi step tasks, and then carry out actions in the game after planning first. The main goal is to let me type things like get a diamond sword or describe this view and have the agent respond in real time inside Minecraft.

Please make the server side handle the model and game logic, and make the client side a clean desktop app I can connect to remotely. I also want a reset button, pause control, and separate modes for planning, captioning, grounded object finding, embodied QA, and action execution. If you need to check current setup details or docs online, go ahead. Keep it practical so I can run it with a downloaded checkpoint and use it for live demos and evaluation.