freecode-rlm

A Recursive Language Model (RLM) runtime in TypeScript.

The LLM writes JavaScript inside a sandboxed REPL; it can inspect a large context programmatically and call sub-LM or sub-RLM functions.

References

• Paper: Recursive Language Models (Zhang, Kraska, Khattab)
• Author write-up: https://alexzhang13.github.io/blog/2025/rlm/
• Official implementation: https://github.com/alexzhang13/rlm

Quickstart

pnpm install
pnpm build
export MINIMAX_API_KEY=sk-...
pnpm -F @freecode-rs/cli dev -- "Print the first 100 powers of two"

Default model is MiniMax-M3 against the OpenAI-compatible endpoint at https://api.minimax.io/v1. Pass --model, --base-url, and --api-key to override.

Layout

• packages/rlm-core — RLM class, types, loop, recursion depth + budget
• packages/rlm-client — LM provider adapters (Vercel AI SDK)
• packages/rlm-repl — sandbox REPL (isolated-vm) + host bridge
• apps/cli — command-line entry (freecode-rlm <prompt> [...])
• apps/web — Next.js landing page

See docs/superpowers/specs/2026-08-10-freecode-rlm-design.md for the design.

Tests

pnpm test                # unit tests (live API tests are skipped without MINIMAX_API_KEY)
pnpm -r build            # strict tsc across all packages and apps

Live coverage:

• packages/rlm-client — MiniMax-M3 chat round-trip
• packages/rlm-core — end-to-end: print the first 100 powers of two; needle-in-the-haystack (50k-line lorem, find a secret token)

Learning project

We study and re-implement (no copy/paste from, no imports of) the following references:

• /home/ayan-de/Projects/githubProjects/agents/rlm/rlm (full Python impl)
• /home/ayan-de/Projects/githubProjects/agents/rlm/rlm-minimal (minimal impl)
• /home/ayan-de/Projects/githubProjects/agents/rlm/prime-agent (TS agent runtime)
• /home/ayan-de/Projects/freecode (sibling harness project)