RLM REPL Environment

A persistent Python REPL maintains a global context variable containing the original input, enabling recursive reasoning and iterative exploration without losing context. It exposes an accessible API with context, llm_query, FINAL, and FINAL_VAR functions.

Batched sub-LM querying is handled via the llm_query(prompt) function, which enables recursive delegation to sub-LMs for chunked analysis. The workflow supports batching limits and error-handling guidance to manage large contexts safely.

You signal completion in a recursive REPL loop by calling FINAL() or FINAL_VAR() to emit explicit, verifiable outputs. This ensures safe completion signaling within the persistent Python environment.

Yes, print statements reveal intermediate results with truncation to support iterative reasoning during recursive exploration. This allows you to inspect available tools and use llm_query() to delegate tasks in batches.

The recursive REPL environment requires no external dependencies, implementing its persistent context and batched sub-LM querying interface entirely within a standalone Python environment.

Use a persistent REPL for recursive prompting when targeting scenarios requiring large-context reasoning, iterative exploration, and safe completion signaling. Standard LLM queries lose context, whereas this REPL loop preserves a global context variable.