danshu3007-lang/jarvis-v2 — reverse-engineered prompt

Build me a Jarvis style AI assistant in Python that can chat in a simple web UI and handle three kinds of requests, research on the web, coding tasks in a sandbox, and very limited OS control with safe allowlisted commands only. I want the interesting part to be that when one of its handlers keeps failing, it reads its past failure logs, asks the model to rewrite that handler, tests the old and new versions on recent examples, and only swaps in the new code if it actually does better.

Please keep memory of each episode with the prompt, reply, reward, domain, and version so it can learn from feedback. Add thumbs up and thumbs down in the UI so those ratings feed the improvement loop. Include the quantum style reward component and the PPO training scaffold if they are practical, but make the whole thing runnable first, ideally on free Google Colab with a T4 and also locally if possible. Keep the tests CPU safe. If anything is unclear, check the current docs online and make sensible choices.