rlm-reasoning-skill

Python REPL enables multi-step reasoning by executing code in repl blocks, inspecting stdout, and refining outputs across iterations. You write Python to process context, run recursive sub-queries, and signal completion via FINAL or FINAL_VAR.

Recursive decomposition breaks complex problems into smaller sub-tasks by using llm_query for simple operations and rlm_query to spawn deeper recursive reasoning when code-backed iteration is needed.

You execute Python code iteratively by writing repl fenced blocks that run in a sandboxed exec environment. This allows you to read and transform input from the persistent context variable namespace across multiple iterations.

Persistent variable state in a Python REPL reasoning loop retains variable values across multiple iterations within the sandboxed exec environment. This allows subsequent repl blocks to access and modify data from previous execution steps.

Yes, you can use recursive model calls for algorithmic reasoning and planning by spawning deeper rlm_query sub-tasks. This handles decomposition-heavy tasks by combining executable Python code with nested language model calls.

You need FINAL or FINAL_VAR to signal completion because it reliably returns the final answer and ends the reasoning loop. Using exactly one of these structured completion signals prevents unbounded or ambiguous code execution.