About

freecode-rlm

A production-grade Recursive Language Model (RLM) runtime in TypeScript. The LLM writes JavaScript inside a sandboxed REPL, inspects large contexts programmatically, and calls sub-LM or sub-RLM functions when it needs help.

“If I cannot make it then I do not understand it.” Reimplement, never import from upstream RLM repos.

What this project is

• A learning project. We study three reference implementations and re-build the architecture from scratch in TypeScript with strict typing.
• A faithful, smaller runtime. The Python reference is the source of truth; we keep the same shape (RLM loop, sandbox REPL, sub-LM calls, budget, depth limits, compaction) but with TS idioms where they help.
• Open development. Every public function has a test. Every commit is small. Every decision has a reason.

Where to look

• docs/superpowers/specs/2026-08-10-freecode-rlm-design.md — the design spec.
• docs/superpowers/plans/2026-08-10-freecode-rlm.md — the milestone plan (M0 bootstrap → M5 polish).
• packages/rlm-core — the RLM loop, budget, compaction, final-answer detection.
• packages/rlm-client — the LM provider adapters (Vercel AI SDK).
• packages/rlm-repl — the sandboxed REPL (isolated-vm) and host bridge.
• apps/cli — the command-line entry point.

References we learn from

• Recursive Language Models — Zhang, Kraska, Khattab 
• Author write-up 
• Official Python implementation: alexzhang13/rlm

License

MIT.