ranxianglei/billion-context ? reverse-engineered prompt

Build me a context compression proxy for coding agents that sits between my assistant and the model API, so long sessions can keep going without running out of context. I want it to work with tools like Claude Code and other similar agents by pointing their base URL at the proxy, then have it automatically compress older parts of the conversation into clear summaries while keeping the important details and cache friendliness intact.

Please make it feel reliable for real day to day coding, with support for streaming responses, a way to inspect whether compression and cache health are working, and sensible defaults that just work out of the box. If there are current integration details or API shapes that matter, look up the latest docs online if you need to.

I also want the project to be easy to install and run locally, with good docs that explain how to connect an agent, what the compression is doing, and how to troubleshoot if the session starts losing quality.