agentscope-ai/Trinity-RFT ? reverse-engineered prompt

Build me a Python framework for reinforcement fine tuning of large language models.

I want it to feel like a practical toolkit, where I can plug in an agent, collect experience from an environment, store and process that data, then train the model from it. It should support a few different ways to do training, like online, offline, and asynchronous workflows, and it should be easy to swap in new algorithms without rewriting everything.

Please include a clean example or two so someone can see how to run a basic training loop on a simple reasoning task. Make it usable for both people who want to improve an agent for a real task and people who want to experiment with new RL ideas. Add helpful docs, tests, and a simple way to run it locally, and if you need current package or model details, look them up online first.