Skip to content

rlm_repl

diapason.agents.rlm_repl

Sandboxed REPL environment for the RLM agent.

Provides a persistent Python namespace with injected helper functions (llm_query, llm_batch, FINAL, FINAL_VAR) that the RLM agent's generated code uses to decompose context and make recursive sub-LM calls.

RLMRepl

Sandboxed Python REPL with persistent namespace for the RLM agent.

Parameters

llm_query_fn: Callback invoked when REPL code calls llm_query(prompt). llm_batch_fn: Callback invoked when REPL code calls llm_batch(prompts). max_output_chars: Maximum characters captured from stdout per execution.

final_answer property

Return the termination value.

is_terminated property

Check if FINAL/FINAL_VAR was called or answer["ready"] is True.

execute(code)

Execute code in the persistent namespace and return captured stdout.

Raises are caught and returned as error strings.

get_variable(name)

Get a variable from the REPL namespace.

security_check(code)

Check code for dangerous patterns. Returns error message or None.

set_variable(name, value)

Set a variable in the REPL namespace.