zagens_runtime/rlm/
mod.rs

1//! Recursive Language Model (RLM) loop — paper-spec Algorithm 1.
2//!
3//! Implements Zhang, Kraska & Khattab (arXiv:2512.24601, §2 Algorithm 1):
4//!
5//! ```text
6//! state ← InitREPL(prompt=P)
7//! state ← AddFunction(state, sub_RLM)
8//! hist ← [Metadata(state)]
9//! while True:
10//!     code ← LLM(hist)
11//!     (state, stdout) ← REPL(state, code)
12//!     hist ← hist ∥ code ∥ Metadata(stdout)
13//!     if state[Final] is set:
14//!         return state[Final]
15//! ```
16//!
17//! Invariants:
18//! - `P` is held only as a REPL variable (`context` / `ctx`); never
19//!   appears in the root LLM's window.
20//! - The root LLM receives small metadata messages — length, preview,
21//!   helper list, prior-round summary.
22//! - Code rounds and sub-LLM calls travel over a single stdin/stdout
23//!   pipe to a long-lived `python3 -u` subprocess. No HTTP sidecar.
24
25pub mod bridge;
26pub mod prompt;
27pub mod turn;
28
29pub use bridge::RlmBridge;
30pub use prompt::rlm_system_prompt;
31pub use turn::{RlmTermination, RlmTurnResult, run_rlm_turn, run_rlm_turn_with_root};
zagens_runtime/rlm/mod.rs

zagens_runtime/rlm/
mod.rs
