| 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 Python subprocess. No HTTP sidecar. |
| 24 | |
| 25 | use crate::models::Usage; |
| 26 | |
| 27 | pub mod bridge; |
| 28 | pub mod prompt; |
| 29 | pub mod session; |
| 30 | pub mod turn; |
| 31 | |
| 32 | pub(crate) use bridge::ModelClientRlmAdapter; |
| 33 | pub use bridge::RlmBridge; |
| 34 | pub use prompt::rlm_system_prompt; |
| 35 | pub use turn::{RlmTermination, RlmTurnResult, run_rlm_turn, run_rlm_turn_with_root}; |
| 36 | |
| 37 | fn add_usage_with_prompt_cache(total: &mut Usage, delta: &Usage) { |
| 38 | total.input_tokens = total.input_tokens.saturating_add(delta.input_tokens); |
| 39 | total.output_tokens = total.output_tokens.saturating_add(delta.output_tokens); |
| 40 | total.prompt_cache_hit_tokens = |
| 41 | add_optional_usage(total.prompt_cache_hit_tokens, delta.prompt_cache_hit_tokens); |
| 42 | total.prompt_cache_miss_tokens = add_optional_usage( |
| 43 | total.prompt_cache_miss_tokens, |
| 44 | delta.prompt_cache_miss_tokens, |
| 45 | ); |
| 46 | total.prompt_cache_write_tokens = add_optional_usage( |
| 47 | total.prompt_cache_write_tokens, |
| 48 | delta.prompt_cache_write_tokens, |
| 49 | ); |
| 50 | total.reasoning_tokens = add_optional_usage(total.reasoning_tokens, delta.reasoning_tokens); |
| 51 | total.reasoning_replay_tokens = |
| 52 | add_optional_usage(total.reasoning_replay_tokens, delta.reasoning_replay_tokens); |
| 53 | if let Some(delta_server) = delta.server_tool_use.as_ref() { |
| 54 | let total_server = total.server_tool_use.get_or_insert_default(); |
| 55 | total_server.code_execution_requests = add_optional_usage( |
| 56 | total_server.code_execution_requests, |
| 57 | delta_server.code_execution_requests, |
| 58 | ); |
| 59 | total_server.tool_search_requests = add_optional_usage( |
| 60 | total_server.tool_search_requests, |
| 61 | delta_server.tool_search_requests, |
| 62 | ); |
| 63 | } |
| 64 | } |
| 65 | |
| 66 | fn add_optional_usage(total: Option<u32>, delta: Option<u32>) -> Option<u32> { |
| 67 | match (total, delta) { |
| 68 | (Some(total), Some(delta)) => Some(total.saturating_add(delta)), |
| 69 | (None, Some(delta)) => Some(delta), |
| 70 | (Some(total), None) => Some(total), |
| 71 | (None, None) => None, |
| 72 | } |
| 73 | } |
| 74 | |
| 75 | #[cfg(test)] |
| 76 | mod tests { |
| 77 | use super::*; |
| 78 | |
| 79 | #[test] |
| 80 | fn add_usage_with_prompt_cache_preserves_cache_counts() { |
| 81 | let mut total = Usage { |
| 82 | input_tokens: 100, |
| 83 | output_tokens: 10, |
| 84 | prompt_cache_hit_tokens: Some(80), |
| 85 | prompt_cache_miss_tokens: Some(20), |
| 86 | ..Usage::default() |
| 87 | }; |
| 88 | let delta = Usage { |
| 89 | input_tokens: 50, |
| 90 | output_tokens: 5, |
| 91 | prompt_cache_hit_tokens: Some(30), |
| 92 | prompt_cache_miss_tokens: Some(20), |
| 93 | reasoning_tokens: Some(4), |
| 94 | reasoning_replay_tokens: Some(3), |
| 95 | server_tool_use: Some(crate::models::ServerToolUsage { |
| 96 | code_execution_requests: Some(2), |
| 97 | tool_search_requests: Some(1), |
| 98 | }), |
| 99 | ..Usage::default() |
| 100 | }; |
| 101 | |
| 102 | add_usage_with_prompt_cache(&mut total, &delta); |
| 103 | |
| 104 | assert_eq!(total.input_tokens, 150); |
| 105 | assert_eq!(total.output_tokens, 15); |
| 106 | assert_eq!(total.prompt_cache_hit_tokens, Some(110)); |
| 107 | assert_eq!(total.prompt_cache_miss_tokens, Some(40)); |
| 108 | assert_eq!(total.reasoning_tokens, Some(4)); |
| 109 | assert_eq!(total.reasoning_replay_tokens, Some(3)); |
| 110 | assert_eq!( |
| 111 | total.server_tool_use, |
| 112 | Some(crate::models::ServerToolUsage { |
| 113 | code_execution_requests: Some(2), |
| 114 | tool_search_requests: Some(1), |
| 115 | }) |
| 116 | ); |
| 117 | } |
| 118 | } |
| 119 |