rand/rlm-claude-code — reverse-engineered prompt

Build me a Claude Code plugin that turns Claude Code into a smarter recursive assistant for huge coding tasks. I want it to notice when a request is too big, break it into smaller questions, search and peek through large context, run safe helper steps, combine the results, and give a better final answer without stuffing everything into one prompt.

It should remember useful facts across sessions, keep a trace of its reasoning steps and costs, and choose cheaper or stronger model routes depending on how complex the job looks. Please include a simple setup flow, a sample config, startup hook that says RLM initialized, and basic commands or scripts to verify it works.

Use Python for the plugin side and the bundled Rust core for the fast pattern and memory parts. Add tests for activation, recursion depth limits, memory, trajectory logging, and install behavior. Look up current Claude Code plugin docs online if you need to.