tamaratran/fast-jev-compaction ? reverse-engineered prompt

Build me a Claude Code plugin and small npm library that makes compaction smarter without rewriting my chat history.

When the conversation gets too long, I want it to keep the original user and assistant messages exactly as they are, and only decide which tool calls and tool results can be dropped, shortened, or kept. If a tool result is still important, keep it verbatim, and if only the call matters, keep the call and trim the result. If neither matters, remove both. It should work on a full transcript, preserve the newest few messages automatically, and fall back safely if the request fails or does not save enough space.

Please make it usable both as a library and as a Claude Code plugin, with a simple install and a clear way to pass in an API key from the environment. Include a small demo or example, tests, and enough types so it is pleasant to use from TypeScript. If you need to check current Claude Code hook docs online, go ahead.