| 1 | //! RLM turn loop — paper Algorithm 1 driven over a long-lived Python |
| 2 | //! subprocess + stdin/stdout RPC bridge (no HTTP sidecar). |
| 3 | |
| 4 | use std::path::PathBuf; |
| 5 | use std::sync::Arc; |
| 6 | use std::time::{Duration, Instant}; |
| 7 | |
| 8 | use tokio::sync::mpsc; |
| 9 | use uuid::Uuid; |
| 10 | |
| 11 | use crate::client::DeepSeekClient; |
| 12 | use crate::core::events::Event; |
| 13 | use crate::models::{ContentBlock, Message, MessageRequest, SystemPrompt, Usage}; |
| 14 | use crate::repl::PythonRuntime; |
| 15 | |
| 16 | use super::bridge::{RlmBridge, RlmLlmClient}; |
| 17 | use super::prompt::rlm_system_prompt; |
| 18 | |
| 19 | // --------------------------------------------------------------------------- |
| 20 | // Constants |
| 21 | // --------------------------------------------------------------------------- |
| 22 | |
| 23 | /// Maximum number of RLM iterations before the loop gives up. |
| 24 | const MAX_RLM_ITERATIONS: u32 = 25; |
| 25 | /// Max consecutive rounds where the model returns no `repl` fence before we |
| 26 | /// hard-fail. The paper requires `code → REPL → Final`; anything else is |
| 27 | /// not the RLM contract. |
| 28 | const MAX_CONSECUTIVE_NO_CODE: u32 = 3; |
| 29 | /// Max output tokens for the root LLM — it just needs to generate code. |
| 30 | const ROOT_MAX_TOKENS: u32 = 4096; |
| 31 | /// Max chars of stdout shown as metadata to the root LLM in next iteration. |
| 32 | const STDOUT_METADATA_PREVIEW_LEN: usize = 800; |
| 33 | /// Max chars of `context` shown as a preview in the metadata. |
| 34 | const PROMPT_PREVIEW_LEN: usize = 500; |
| 35 | /// Temperature for root LLM calls. |
| 36 | const ROOT_TEMPERATURE: f32 = 0.3; |
| 37 | /// Hard wall-clock cap on a whole RLM turn. |
| 38 | const TURN_TIMEOUT: Duration = Duration::from_secs(180); |
| 39 | /// Bound on conversation history we keep across iterations. |
| 40 | const MAX_HISTORY_MESSAGES: usize = 20; |
| 41 | |
| 42 | // --------------------------------------------------------------------------- |
| 43 | // Public API |
| 44 | // --------------------------------------------------------------------------- |
| 45 | |
| 46 | /// How an RLM turn ended. |
| 47 | #[derive(Debug, Clone, Copy, PartialEq, Eq)] |
| 48 | pub enum RlmTermination { |
| 49 | /// `FINAL(value)` was called inside the REPL or `FINAL(...)` appeared |
| 50 | /// at the top of the model's response on its own line. |
| 51 | Final, |
| 52 | /// The model failed to emit a `repl` block for too many rounds in a |
| 53 | /// row. The accumulated last response text is surfaced as the answer |
| 54 | /// rather than being thrown away. |
| 55 | NoCode, |
| 56 | /// Iteration cap reached without `FINAL`. |
| 57 | Exhausted, |
| 58 | /// Hard error — LLM call failed, REPL crashed, timeout. |
| 59 | Error, |
| 60 | } |
| 61 | |
| 62 | /// Per-round trace entry. Surfaced in the tool result so the user can see |
| 63 | /// exactly what the sub-agent did. |
| 64 | #[derive(Debug, Clone)] |
| 65 | pub struct RlmRoundTrace { |
| 66 | pub round: u32, |
| 67 | pub code_summary: String, |
| 68 | pub stdout_preview: String, |
| 69 | pub had_error: bool, |
| 70 | pub rpc_count: u32, |
| 71 | pub elapsed_ms: u64, |
| 72 | } |
| 73 | |
| 74 | /// Result of an RLM turn. |
| 75 | #[derive(Debug, Clone)] |
| 76 | pub struct RlmTurnResult { |
| 77 | pub answer: String, |
| 78 | pub iterations: u32, |
| 79 | pub duration: Duration, |
| 80 | pub error: Option<String>, |
| 81 | pub usage: Usage, |
| 82 | pub termination: RlmTermination, |
| 83 | /// Per-round trace. Empty when the loop never reached the REPL. |
| 84 | pub trace: Vec<RlmRoundTrace>, |
| 85 | /// Total sub-LLM RPCs made by the sub-agent (sum of `rpc_count` across |
| 86 | /// rounds). Useful for verifying that the model engaged with `context` |
| 87 | /// rather than answering directly. |
| 88 | pub total_rpcs: u32, |
| 89 | } |
| 90 | |
| 91 | /// Run a full RLM turn. `prompt` is loaded into the REPL as `context`; it |
| 92 | /// never enters the root LLM's window. |
| 93 | pub async fn run_rlm_turn( |
| 94 | client: &DeepSeekClient, |
| 95 | model: String, |
| 96 | prompt: String, |
| 97 | child_model: String, |
| 98 | tx_event: mpsc::Sender<Event>, |
| 99 | max_depth: u32, |
| 100 | ) -> RlmTurnResult { |
| 101 | run_rlm_turn_inner( |
| 102 | Arc::new(client.clone()), |
| 103 | model, |
| 104 | prompt, |
| 105 | None, |
| 106 | child_model, |
| 107 | tx_event, |
| 108 | max_depth, |
| 109 | ) |
| 110 | .await |
| 111 | } |
| 112 | |
| 113 | /// Variant that also passes a small `root_prompt` (the user-facing task) |
| 114 | /// shown to the root LLM each iteration so it remembers its objective. |
| 115 | pub async fn run_rlm_turn_with_root( |
| 116 | client: &DeepSeekClient, |
| 117 | model: String, |
| 118 | prompt: String, |
| 119 | root_prompt: Option<String>, |
| 120 | child_model: String, |
| 121 | tx_event: mpsc::Sender<Event>, |
| 122 | max_depth: u32, |
| 123 | ) -> RlmTurnResult { |
| 124 | run_rlm_turn_inner( |
| 125 | Arc::new(client.clone()), |
| 126 | model, |
| 127 | prompt, |
| 128 | root_prompt, |
| 129 | child_model, |
| 130 | tx_event, |
| 131 | max_depth, |
| 132 | ) |
| 133 | .await |
| 134 | } |
| 135 | |
| 136 | /// Inner entry point — also used by the bridge when it recurses. Returns |
| 137 | /// a boxed future to break the recursive opaque-future-type cycle: |
| 138 | /// `run_rlm_turn_inner` → `RlmBridge::dispatch` → `run_rlm_turn_inner`. |
| 139 | pub(crate) fn run_rlm_turn_inner( |
| 140 | client: Arc<dyn RlmLlmClient>, |
| 141 | model: String, |
| 142 | prompt: String, |
| 143 | root_prompt: Option<String>, |
| 144 | child_model: String, |
| 145 | tx_event: mpsc::Sender<Event>, |
| 146 | max_depth: u32, |
| 147 | ) -> std::pin::Pin<Box<dyn std::future::Future<Output = RlmTurnResult> + Send>> { |
| 148 | Box::pin(run_rlm_turn_impl( |
| 149 | client, |
| 150 | model, |
| 151 | prompt, |
| 152 | root_prompt, |
| 153 | child_model, |
| 154 | tx_event, |
| 155 | max_depth, |
| 156 | )) |
| 157 | } |
| 158 | |
| 159 | // --------------------------------------------------------------------------- |
| 160 | // Implementation |
| 161 | // --------------------------------------------------------------------------- |
| 162 | |
| 163 | async fn run_rlm_turn_impl( |
| 164 | client: Arc<dyn RlmLlmClient>, |
| 165 | model: String, |
| 166 | prompt: String, |
| 167 | root_prompt: Option<String>, |
| 168 | child_model: String, |
| 169 | tx_event: mpsc::Sender<Event>, |
| 170 | max_depth: u32, |
| 171 | ) -> RlmTurnResult { |
| 172 | let start = Instant::now(); |
| 173 | let mut total_usage = Usage::default(); |
| 174 | let mut trace: Vec<RlmRoundTrace> = Vec::new(); |
| 175 | let mut total_rpcs: u32 = 0; |
| 176 | |
| 177 | // 1. Stage `context` to a temp file. The REPL reads it on bootstrap so |
| 178 | // the big string never enters the process command line and doesn't |
| 179 | // show up in `ps`. |
| 180 | let ctx_path = match write_context_file(&prompt) { |
| 181 | Ok(p) => p, |
| 182 | Err(e) => { |
| 183 | return RlmTurnResult { |
| 184 | answer: String::new(), |
| 185 | iterations: 0, |
| 186 | duration: start.elapsed(), |
| 187 | error: Some(format!("rlm: failed to stage context: {e}")), |
| 188 | usage: total_usage, |
| 189 | termination: RlmTermination::Error, |
| 190 | trace, |
| 191 | total_rpcs, |
| 192 | }; |
| 193 | } |
| 194 | }; |
| 195 | |
| 196 | // 2. Spawn the long-lived REPL. |
| 197 | let mut repl = match PythonRuntime::spawn_with_context(&ctx_path).await { |
| 198 | Ok(rt) => rt, |
| 199 | Err(e) => { |
| 200 | let _ = tokio::fs::remove_file(&ctx_path).await; |
| 201 | return RlmTurnResult { |
| 202 | answer: String::new(), |
| 203 | iterations: 0, |
| 204 | duration: start.elapsed(), |
| 205 | error: Some(format!("rlm: failed to spawn REPL: {e}")), |
| 206 | usage: total_usage, |
| 207 | termination: RlmTermination::Error, |
| 208 | trace, |
| 209 | total_rpcs, |
| 210 | }; |
| 211 | } |
| 212 | }; |
| 213 | |
| 214 | // 3. Build the bridge that services llm_query / rlm_query RPCs. |
| 215 | let bridge = RlmBridge::new(Arc::clone(&client), child_model.clone(), max_depth); |
| 216 | let usage_handle = bridge.usage_handle(); |
| 217 | |
| 218 | let _ = tx_event |
| 219 | .send(Event::status(format!( |
| 220 | "RLM: spawned Python REPL (root={model}, child={child_model}, max_depth={max_depth}, ctx={} chars)", |
| 221 | prompt.chars().count() |
| 222 | ))) |
| 223 | .await; |
| 224 | |
| 225 | // 4. Build initial metadata-only history. |
| 226 | let system = rlm_system_prompt(); |
| 227 | let mut messages: Vec<Message> = vec![build_metadata_message( |
| 228 | &prompt, |
| 229 | root_prompt.as_deref(), |
| 230 | 0, |
| 231 | None, |
| 232 | None, |
| 233 | )]; |
| 234 | |
| 235 | let mut consecutive_no_code: u32 = 0; |
| 236 | let mut last_response_text = String::new(); |
| 237 | |
| 238 | let result = 'turn: { |
| 239 | for iteration in 0..MAX_RLM_ITERATIONS { |
| 240 | if start.elapsed() > TURN_TIMEOUT { |
| 241 | break 'turn RlmTurnResult { |
| 242 | answer: String::new(), |
| 243 | iterations: iteration, |
| 244 | duration: start.elapsed(), |
| 245 | error: Some(format!( |
| 246 | "RLM turn timed out after {}s", |
| 247 | TURN_TIMEOUT.as_secs() |
| 248 | )), |
| 249 | usage: total_usage, |
| 250 | termination: RlmTermination::Error, |
| 251 | trace: trace.clone(), |
| 252 | total_rpcs, |
| 253 | }; |
| 254 | } |
| 255 | |
| 256 | let _ = tx_event |
| 257 | .send(Event::status(format!( |
| 258 | "RLM iteration {}/{}", |
| 259 | iteration + 1, |
| 260 | MAX_RLM_ITERATIONS |
| 261 | ))) |
| 262 | .await; |
| 263 | |
| 264 | // 4a. Root LLM generates code from metadata-only context. |
| 265 | let request = build_root_request(&model, &messages, &system); |
| 266 | |
| 267 | let response = match client.create_message_boxed(request).await { |
| 268 | Ok(r) => r, |
| 269 | Err(e) => { |
| 270 | break 'turn RlmTurnResult { |
| 271 | answer: String::new(), |
| 272 | iterations: iteration + 1, |
| 273 | duration: start.elapsed(), |
| 274 | error: Some(format!("Root LLM call failed: {e}")), |
| 275 | usage: total_usage, |
| 276 | termination: RlmTermination::Error, |
| 277 | trace: trace.clone(), |
| 278 | total_rpcs, |
| 279 | }; |
| 280 | } |
| 281 | }; |
| 282 | |
| 283 | total_usage.input_tokens = total_usage |
| 284 | .input_tokens |
| 285 | .saturating_add(response.usage.input_tokens); |
| 286 | total_usage.output_tokens = total_usage |
| 287 | .output_tokens |
| 288 | .saturating_add(response.usage.output_tokens); |
| 289 | |
| 290 | let response_text = extract_text_blocks(&response.content); |
| 291 | last_response_text = response_text.clone(); |
| 292 | |
| 293 | // 4b. Top-level FINAL(...) lets the model close out without |
| 294 | // touching the REPL — but only if it has done some work |
| 295 | // (non-zero rpc_count) on a prior round. Otherwise it's a |
| 296 | // shortcut and we reject it. |
| 297 | if let Some(final_val) = parse_text_final(&response_text) { |
| 298 | if total_rpcs == 0 { |
| 299 | // Discard the top-level FINAL — the model is bypassing |
| 300 | // the loop. Force it to use the REPL by appending a |
| 301 | // strict reminder. |
| 302 | consecutive_no_code = consecutive_no_code.saturating_add(1); |
| 303 | if consecutive_no_code >= MAX_CONSECUTIVE_NO_CODE { |
| 304 | break 'turn RlmTurnResult { |
| 305 | answer: final_val, |
| 306 | iterations: iteration + 1, |
| 307 | duration: start.elapsed(), |
| 308 | error: None, |
| 309 | usage: total_usage, |
| 310 | termination: RlmTermination::NoCode, |
| 311 | trace: trace.clone(), |
| 312 | total_rpcs, |
| 313 | }; |
| 314 | } |
| 315 | messages.push(Message { |
| 316 | role: "assistant".to_string(), |
| 317 | content: vec![ContentBlock::Text { |
| 318 | text: response_text.clone(), |
| 319 | cache_control: None, |
| 320 | }], |
| 321 | }); |
| 322 | messages.push(Message { |
| 323 | role: "user".to_string(), |
| 324 | content: vec![ContentBlock::Text { |
| 325 | text: "You called FINAL(...) without ever running a ```repl block. \ |
| 326 | That defeats the recursive language model — you're guessing \ |
| 327 | from the preview alone. Emit a ```repl block now that uses \ |
| 328 | `llm_query`, `llm_query_batched`, or `rlm_query` against \ |
| 329 | `context` to actually compute the answer." |
| 330 | .to_string(), |
| 331 | cache_control: None, |
| 332 | }], |
| 333 | }); |
| 334 | continue; |
| 335 | } |
| 336 | let _ = tx_event |
| 337 | .send(Event::status( |
| 338 | "RLM: FINAL detected in response text".to_string(), |
| 339 | )) |
| 340 | .await; |
| 341 | break 'turn RlmTurnResult { |
| 342 | answer: final_val, |
| 343 | iterations: iteration + 1, |
| 344 | duration: start.elapsed(), |
| 345 | error: None, |
| 346 | usage: total_usage, |
| 347 | termination: RlmTermination::Final, |
| 348 | trace: trace.clone(), |
| 349 | total_rpcs, |
| 350 | }; |
| 351 | } |
| 352 | |
| 353 | // 4c. Extract a ```repl block. |
| 354 | let code = extract_repl_code(&response_text); |
| 355 | let code_to_run = match code { |
| 356 | Some(c) => { |
| 357 | consecutive_no_code = 0; |
| 358 | c |
| 359 | } |
| 360 | None => { |
| 361 | consecutive_no_code = consecutive_no_code.saturating_add(1); |
| 362 | if consecutive_no_code >= MAX_CONSECUTIVE_NO_CODE { |
| 363 | break 'turn RlmTurnResult { |
| 364 | answer: response_text, |
| 365 | iterations: iteration + 1, |
| 366 | duration: start.elapsed(), |
| 367 | error: Some(format!( |
| 368 | "RLM: model failed to emit ```repl after {MAX_CONSECUTIVE_NO_CODE} consecutive rounds" |
| 369 | )), |
| 370 | usage: total_usage, |
| 371 | termination: RlmTermination::NoCode, |
| 372 | trace: trace.clone(), |
| 373 | total_rpcs, |
| 374 | }; |
| 375 | } |
| 376 | messages.push(Message { |
| 377 | role: "assistant".to_string(), |
| 378 | content: vec![ContentBlock::Text { |
| 379 | text: response_text.clone(), |
| 380 | cache_control: None, |
| 381 | }], |
| 382 | }); |
| 383 | messages.push(Message { |
| 384 | role: "user".to_string(), |
| 385 | content: vec![ContentBlock::Text { |
| 386 | text: "Reminder: emit Python inside a ```repl … ``` fence. \ |
| 387 | Use `llm_query` / `llm_query_batched` / `rlm_query` to \ |
| 388 | process `context` and call `FINAL(value)` when done." |
| 389 | .to_string(), |
| 390 | cache_control: None, |
| 391 | }], |
| 392 | }); |
| 393 | continue; |
| 394 | } |
| 395 | }; |
| 396 | |
| 397 | let _ = tx_event |
| 398 | .send(Event::MessageDelta { |
| 399 | index: iteration as usize, |
| 400 | content: format!( |
| 401 | "\n[RLM round {} — code]\n```repl\n{code_to_run}\n```\n", |
| 402 | iteration + 1 |
| 403 | ), |
| 404 | }) |
| 405 | .await; |
| 406 | |
| 407 | // 4d. Execute the code in the REPL with the bridge servicing |
| 408 | // llm_query / rlm_query callbacks. |
| 409 | let round = match repl.run(&code_to_run, Some(&bridge)).await { |
| 410 | Ok(r) => r, |
| 411 | Err(e) => { |
| 412 | break 'turn RlmTurnResult { |
| 413 | answer: String::new(), |
| 414 | iterations: iteration + 1, |
| 415 | duration: start.elapsed(), |
| 416 | error: Some(format!("REPL execution failed: {e}")), |
| 417 | usage: total_usage, |
| 418 | termination: RlmTermination::Error, |
| 419 | trace: trace.clone(), |
| 420 | total_rpcs, |
| 421 | }; |
| 422 | } |
| 423 | }; |
| 424 | |
| 425 | total_rpcs = total_rpcs.saturating_add(round.rpc_count); |
| 426 | |
| 427 | // Trace this round. |
| 428 | let stdout_preview = truncate_text(round.stdout.trim(), STDOUT_METADATA_PREVIEW_LEN); |
| 429 | trace.push(RlmRoundTrace { |
| 430 | round: iteration + 1, |
| 431 | code_summary: summarize_code(&code_to_run), |
| 432 | stdout_preview: stdout_preview.clone(), |
| 433 | had_error: round.has_error, |
| 434 | rpc_count: round.rpc_count, |
| 435 | elapsed_ms: round.elapsed.as_millis() as u64, |
| 436 | }); |
| 437 | |
| 438 | let _ = tx_event |
| 439 | .send(Event::status(format!( |
| 440 | "RLM round {}: {} bytes stdout, {} sub-LLM call(s){}", |
| 441 | iteration + 1, |
| 442 | round.full_stdout.len(), |
| 443 | round.rpc_count, |
| 444 | if round.has_error { " (error)" } else { "" }, |
| 445 | ))) |
| 446 | .await; |
| 447 | |
| 448 | // 4e. FINAL detection. |
| 449 | if let Some(final_val) = round.final_value.clone() { |
| 450 | let _ = tx_event |
| 451 | .send(Event::status( |
| 452 | "RLM: FINAL detected in REPL, ending loop".to_string(), |
| 453 | )) |
| 454 | .await; |
| 455 | break 'turn RlmTurnResult { |
| 456 | answer: final_val, |
| 457 | iterations: iteration + 1, |
| 458 | duration: start.elapsed(), |
| 459 | error: None, |
| 460 | usage: total_usage, |
| 461 | termination: RlmTermination::Final, |
| 462 | trace: trace.clone(), |
| 463 | total_rpcs, |
| 464 | }; |
| 465 | } |
| 466 | |
| 467 | // 4f. Build metadata for next iteration. |
| 468 | messages.push(Message { |
| 469 | role: "assistant".to_string(), |
| 470 | content: vec![ContentBlock::Text { |
| 471 | text: format!("```repl\n{code_to_run}\n```"), |
| 472 | cache_control: None, |
| 473 | }], |
| 474 | }); |
| 475 | messages.push(build_metadata_message( |
| 476 | &prompt, |
| 477 | root_prompt.as_deref(), |
| 478 | iteration + 1, |
| 479 | Some(&code_to_run), |
| 480 | Some(&stdout_preview), |
| 481 | )); |
| 482 | |
| 483 | if messages.len() > MAX_HISTORY_MESSAGES { |
| 484 | let drop_from = messages.len() - MAX_HISTORY_MESSAGES + 1; |
| 485 | let mut kept = vec![messages[0].clone()]; |
| 486 | kept.extend(messages.drain(drop_from..)); |
| 487 | messages = kept; |
| 488 | } |
| 489 | } |
| 490 | |
| 491 | let _ = last_response_text; |
| 492 | RlmTurnResult { |
| 493 | answer: String::new(), |
| 494 | iterations: MAX_RLM_ITERATIONS, |
| 495 | duration: start.elapsed(), |
| 496 | error: Some(format!( |
| 497 | "RLM loop exhausted after {MAX_RLM_ITERATIONS} iterations without FINAL" |
| 498 | )), |
| 499 | usage: total_usage, |
| 500 | termination: RlmTermination::Exhausted, |
| 501 | trace: trace.clone(), |
| 502 | total_rpcs, |
| 503 | } |
| 504 | }; |
| 505 | |
| 506 | // Fold bridge usage (children + nested sub_rlm) into totals. |
| 507 | let bridge_usage = usage_handle.lock().await; |
| 508 | let mut final_usage = result.usage.clone(); |
| 509 | final_usage.input_tokens = final_usage |
| 510 | .input_tokens |
| 511 | .saturating_add(bridge_usage.input_tokens); |
| 512 | final_usage.output_tokens = final_usage |
| 513 | .output_tokens |
| 514 | .saturating_add(bridge_usage.output_tokens); |
| 515 | drop(bridge_usage); |
| 516 | |
| 517 | repl.shutdown().await; |
| 518 | |
| 519 | RlmTurnResult { |
| 520 | usage: final_usage, |
| 521 | ..result |
| 522 | } |
| 523 | } |
| 524 | |
| 525 | // --------------------------------------------------------------------------- |
| 526 | // Helpers |
| 527 | // --------------------------------------------------------------------------- |
| 528 | |
| 529 | fn write_context_file(prompt: &str) -> std::io::Result<PathBuf> { |
| 530 | let dir = std::env::temp_dir().join("deepseek_rlm_ctx"); |
| 531 | std::fs::create_dir_all(&dir)?; |
| 532 | let path = dir.join(format!( |
| 533 | "ctx_{}_{}.txt", |
| 534 | std::process::id(), |
| 535 | Uuid::new_v4().simple() |
| 536 | )); |
| 537 | std::fs::write(&path, prompt)?; |
| 538 | Ok(path) |
| 539 | } |
| 540 | |
| 541 | fn build_root_request(model: &str, messages: &[Message], system: &SystemPrompt) -> MessageRequest { |
| 542 | MessageRequest { |
| 543 | model: model.to_string(), |
| 544 | messages: messages.to_vec(), |
| 545 | max_tokens: ROOT_MAX_TOKENS, |
| 546 | system: Some(system.clone()), |
| 547 | tools: None, |
| 548 | tool_choice: None, |
| 549 | metadata: None, |
| 550 | thinking: None, |
| 551 | reasoning_effort: None, |
| 552 | stream: Some(false), |
| 553 | temperature: Some(ROOT_TEMPERATURE), |
| 554 | top_p: Some(0.9_f32), |
| 555 | } |
| 556 | } |
| 557 | |
| 558 | /// Build `Metadata(state)` from the paper. Surfaces: |
| 559 | /// - the small `root_prompt` (if any) — repeated each iteration |
| 560 | /// - `context` length + preview |
| 561 | /// - the REPL helpers |
| 562 | /// - the previous round's code summary + stdout preview |
| 563 | fn build_metadata_message( |
| 564 | prompt: &str, |
| 565 | root_prompt: Option<&str>, |
| 566 | iteration: u32, |
| 567 | previous_code: Option<&str>, |
| 568 | previous_stdout: Option<&str>, |
| 569 | ) -> Message { |
| 570 | let prompt_len = prompt.chars().count(); |
| 571 | let prompt_preview = truncate_text(prompt, PROMPT_PREVIEW_LEN); |
| 572 | |
| 573 | let mut parts = Vec::new(); |
| 574 | parts.push(format!("## REPL state (round {iteration})")); |
| 575 | parts.push(String::new()); |
| 576 | if let Some(rp) = root_prompt |
| 577 | && !rp.trim().is_empty() |
| 578 | { |
| 579 | parts.push("**Original task** (re-shown every round)".to_string()); |
| 580 | parts.push(format!("> {}", truncate_text(rp.trim(), 600))); |
| 581 | parts.push(String::new()); |
| 582 | } |
| 583 | parts.push("**`context`** — the long input lives in the REPL only".to_string()); |
| 584 | parts.push(format!("- Length: {prompt_len} chars")); |
| 585 | parts.push(format!("- Preview: \"{prompt_preview}\"")); |
| 586 | parts.push(String::new()); |
| 587 | |
| 588 | parts.push("**REPL helpers** (use inside ```repl blocks)".to_string()); |
| 589 | parts.push("- `context` / `ctx` — the full input string".to_string()); |
| 590 | parts.push("- `len(context)` / `context[a:b]` / `context.splitlines()` — slice it".to_string()); |
| 591 | parts.push( |
| 592 | "- `llm_query(prompt, model=None)` — one-shot child LLM; `model` is ignored and child calls stay pinned to Flash" |
| 593 | .to_string(), |
| 594 | ); |
| 595 | parts.push( |
| 596 | "- `llm_query_batched([p1, p2, ...])` — concurrent fan-out; `model` is ignored" |
| 597 | .to_string(), |
| 598 | ); |
| 599 | parts.push( |
| 600 | "- `rlm_query(prompt, model=None)` — recursive sub-RLM; `model` is ignored" |
| 601 | .to_string(), |
| 602 | ); |
| 603 | parts.push( |
| 604 | "- `rlm_query_batched([p1, p2, ...])` — concurrent recursive sub-RLMs; `model` is ignored" |
| 605 | .to_string(), |
| 606 | ); |
| 607 | parts.push("- `SHOW_VARS()` — list user variables".to_string()); |
| 608 | parts.push("- `repl_set(name, value)` / `repl_get(name)` — explicit store".to_string()); |
| 609 | parts.push( |
| 610 | "- `FINAL(value)` — end the loop with this answer".to_string(), |
| 611 | ); |
| 612 | parts.push( |
| 613 | "- `FINAL_VAR(name)` — end the loop with a variable's value" |
| 614 | .to_string(), |
| 615 | ); |
| 616 | parts.push(String::new()); |
| 617 | |
| 618 | if iteration > 0 { |
| 619 | parts.push("**Previous round**".to_string()); |
| 620 | if let Some(code) = previous_code { |
| 621 | parts.push(format!("- Code: {}", summarize_code(code))); |
| 622 | } |
| 623 | if let Some(stdout) = previous_stdout { |
| 624 | let stdout_clean = stdout.trim(); |
| 625 | if !stdout_clean.is_empty() { |
| 626 | parts.push(format!("- Stdout preview: \"{stdout_clean}\"")); |
| 627 | } else { |
| 628 | parts.push("- Stdout: (empty)".to_string()); |
| 629 | } |
| 630 | } |
| 631 | } |
| 632 | |
| 633 | let text = parts.join("\n"); |
| 634 | |
| 635 | Message { |
| 636 | role: "user".to_string(), |
| 637 | content: vec![ContentBlock::Text { |
| 638 | text, |
| 639 | cache_control: None, |
| 640 | }], |
| 641 | } |
| 642 | } |
| 643 | |
| 644 | fn summarize_code(code: &str) -> String { |
| 645 | let lines: Vec<&str> = code.lines().collect(); |
| 646 | if lines.len() <= 8 { |
| 647 | return code.to_string(); |
| 648 | } |
| 649 | let head = lines[..4].join("\n"); |
| 650 | let tail = lines[lines.len() - 4..].join("\n"); |
| 651 | format!("{} lines:\n{head}\n…\n{tail}", lines.len()) |
| 652 | } |
| 653 | |
| 654 | fn extract_text_blocks(blocks: &[ContentBlock]) -> String { |
| 655 | blocks |
| 656 | .iter() |
| 657 | .filter_map(|b| match b { |
| 658 | ContentBlock::Text { text, .. } => Some(text.as_str()), |
| 659 | _ => None, |
| 660 | }) |
| 661 | .collect::<Vec<_>>() |
| 662 | .join("\n") |
| 663 | } |
| 664 | |
| 665 | /// Extract the first ` ```repl ` block from `text`. Falls back to |
| 666 | /// ` ```python `/`` ```py `` for compatibility with prompts that learned |
| 667 | /// the older fence style. |
| 668 | fn extract_repl_code(text: &str) -> Option<String> { |
| 669 | let start_markers = [ |
| 670 | "```repl\n", |
| 671 | "```repl\r\n", |
| 672 | "```python\n", |
| 673 | "```py\n", |
| 674 | "```python\r\n", |
| 675 | "```py\r\n", |
| 676 | ]; |
| 677 | let mut best_start: Option<(usize, &str)> = None; |
| 678 | |
| 679 | for marker in &start_markers { |
| 680 | if let Some(idx) = text.find(marker) { |
| 681 | let end_pos = idx + marker.len(); |
| 682 | match best_start { |
| 683 | Some((best_idx, _)) if idx < best_idx => { |
| 684 | best_start = Some((idx, &text[end_pos..])); |
| 685 | } |
| 686 | None => { |
| 687 | best_start = Some((idx, &text[end_pos..])); |
| 688 | } |
| 689 | _ => {} |
| 690 | } |
| 691 | } |
| 692 | } |
| 693 | |
| 694 | let after_fence = best_start.map(|(_, rest)| rest)?; |
| 695 | |
| 696 | let end_idx = after_fence |
| 697 | .find("\n```") |
| 698 | .or_else(|| after_fence.find("```"))?; |
| 699 | |
| 700 | let code = after_fence[..end_idx].trim().to_string(); |
| 701 | if code.is_empty() { |
| 702 | return None; |
| 703 | } |
| 704 | Some(code) |
| 705 | } |
| 706 | |
| 707 | /// Parse a top-level `FINAL(...)` directive from the model's raw text. |
| 708 | /// Mirrors the reference RLM's `find_final_answer`: directive must appear |
| 709 | /// at the start of a line, *outside* any code fence. |
| 710 | fn parse_text_final(text: &str) -> Option<String> { |
| 711 | let outside_fence = strip_code_fences(text); |
| 712 | |
| 713 | for line in outside_fence.lines() { |
| 714 | let trimmed = line.trim_start(); |
| 715 | if trimmed.starts_with("FINAL_VAR(") { |
| 716 | // FINAL_VAR can't be resolved from text alone — defer to REPL. |
| 717 | continue; |
| 718 | } |
| 719 | if let Some(rest) = trimmed.strip_prefix("FINAL(") { |
| 720 | let inner = rest.trim_end(); |
| 721 | if let Some(end) = inner.rfind(')') { |
| 722 | let value = inner[..end].trim(); |
| 723 | if !value.is_empty() { |
| 724 | return Some(strip_quotes(value)); |
| 725 | } |
| 726 | } |
| 727 | } |
| 728 | } |
| 729 | None |
| 730 | } |
| 731 | |
| 732 | fn strip_code_fences(text: &str) -> String { |
| 733 | let mut out = String::with_capacity(text.len()); |
| 734 | let mut in_fence = false; |
| 735 | for line in text.lines() { |
| 736 | if line.trim_start().starts_with("```") { |
| 737 | in_fence = !in_fence; |
| 738 | continue; |
| 739 | } |
| 740 | if !in_fence { |
| 741 | out.push_str(line); |
| 742 | out.push('\n'); |
| 743 | } |
| 744 | } |
| 745 | out |
| 746 | } |
| 747 | |
| 748 | fn strip_quotes(s: &str) -> String { |
| 749 | let bytes = s.as_bytes(); |
| 750 | if bytes.len() >= 2 |
| 751 | && ((bytes[0] == b'"' && bytes[bytes.len() - 1] == b'"') |
| 752 | || (bytes[0] == b'\'' && bytes[bytes.len() - 1] == b'\'')) |
| 753 | { |
| 754 | return s[1..s.len() - 1].to_string(); |
| 755 | } |
| 756 | s.to_string() |
| 757 | } |
| 758 | |
| 759 | fn truncate_text(text: &str, max_chars: usize) -> String { |
| 760 | let count = text.chars().count(); |
| 761 | if count <= max_chars { |
| 762 | return text.to_string(); |
| 763 | } |
| 764 | let take = max_chars.saturating_sub(3); |
| 765 | let mut result: String = text.chars().take(take).collect(); |
| 766 | result.push_str("..."); |
| 767 | result |
| 768 | } |
| 769 | |
| 770 | // --------------------------------------------------------------------------- |
| 771 | // Tests |
| 772 | // --------------------------------------------------------------------------- |
| 773 | |
| 774 | #[cfg(test)] |
| 775 | mod tests { |
| 776 | use super::*; |
| 777 | |
| 778 | #[test] |
| 779 | fn extract_repl_code_finds_simple_block() { |
| 780 | let text = "Here:\n```repl\nprint('hi')\n```\nEnd."; |
| 781 | let code = extract_repl_code(text).unwrap(); |
| 782 | assert_eq!(code, "print('hi')"); |
| 783 | } |
| 784 | |
| 785 | #[test] |
| 786 | fn extract_repl_code_falls_back_to_python_marker() { |
| 787 | let text = "Code:\n```python\nx = 1 + 2\n```"; |
| 788 | let code = extract_repl_code(text).unwrap(); |
| 789 | assert_eq!(code, "x = 1 + 2"); |
| 790 | } |
| 791 | |
| 792 | #[test] |
| 793 | fn extract_repl_code_returns_none_when_missing() { |
| 794 | assert!(extract_repl_code("Just text.").is_none()); |
| 795 | } |
| 796 | |
| 797 | #[test] |
| 798 | fn extract_repl_code_returns_none_on_empty_block() { |
| 799 | assert!(extract_repl_code("```repl\n\n```").is_none()); |
| 800 | } |
| 801 | |
| 802 | #[test] |
| 803 | fn extract_repl_code_handles_multiple_blocks() { |
| 804 | let text = "```repl\na=1\n```\n```repl\nb=2\n```"; |
| 805 | let code = extract_repl_code(text).unwrap(); |
| 806 | assert_eq!(code, "a=1"); |
| 807 | } |
| 808 | |
| 809 | #[test] |
| 810 | fn extract_repl_code_ignores_other_fences() { |
| 811 | let text = "```\nfoo\n```\n```repl\nreal_code()\n```"; |
| 812 | let code = extract_repl_code(text).unwrap(); |
| 813 | assert_eq!(code, "real_code()"); |
| 814 | } |
| 815 | |
| 816 | #[test] |
| 817 | fn parse_text_final_extracts_simple_value() { |
| 818 | let text = "OK.\nFINAL(42)\nThanks."; |
| 819 | assert_eq!(parse_text_final(text).as_deref(), Some("42")); |
| 820 | } |
| 821 | |
| 822 | #[test] |
| 823 | fn parse_text_final_strips_quotes() { |
| 824 | let text = "FINAL(\"the answer is yes\")"; |
| 825 | assert_eq!(parse_text_final(text).as_deref(), Some("the answer is yes")); |
| 826 | } |
| 827 | |
| 828 | #[test] |
| 829 | fn parse_text_final_ignores_inside_code_fence() { |
| 830 | let text = |
| 831 | "Some prose.\n```repl\n# Note: when ready, call FINAL(value)\nx = 1\n```\nMore prose."; |
| 832 | assert!(parse_text_final(text).is_none()); |
| 833 | } |
| 834 | |
| 835 | #[test] |
| 836 | fn parse_text_final_returns_none_when_absent() { |
| 837 | assert!(parse_text_final("just talking, no final.").is_none()); |
| 838 | } |
| 839 | |
| 840 | #[test] |
| 841 | fn build_metadata_contains_key_information() { |
| 842 | let msg = build_metadata_message("Hello, world!", None, 0, None, None); |
| 843 | let text = extract_text_blocks(&msg.content); |
| 844 | assert!(text.contains("context")); |
| 845 | assert!(text.contains("Hello, world!")); |
| 846 | assert!(text.contains("round 0")); |
| 847 | assert!(text.contains("llm_query")); |
| 848 | assert!(text.contains("rlm_query")); |
| 849 | assert!(text.contains("FINAL")); |
| 850 | } |
| 851 | |
| 852 | #[test] |
| 853 | fn build_metadata_truncates_long_context_without_leaking_tail() { |
| 854 | let secret_tail = "DO_NOT_LEAK_CONTEXT_TAIL"; |
| 855 | let prompt = format!("{}{}", "a".repeat(PROMPT_PREVIEW_LEN + 100), secret_tail); |
| 856 | let msg = build_metadata_message(&prompt, None, 0, None, None); |
| 857 | let text = extract_text_blocks(&msg.content); |
| 858 | |
| 859 | assert!(text.contains(&format!("- Length: {} chars", prompt.chars().count()))); |
| 860 | assert!(text.contains("- Preview: \"")); |
| 861 | assert!(text.contains("...")); |
| 862 | assert!( |
| 863 | !text.contains(secret_tail), |
| 864 | "metadata leaked the non-preview tail of context" |
| 865 | ); |
| 866 | } |
| 867 | |
| 868 | #[test] |
| 869 | fn build_root_request_keeps_context_tail_out_of_root_payload() { |
| 870 | let secret_tail = "DO_NOT_LEAK_ROOT_REQUEST"; |
| 871 | let prompt = format!("{}{}", "a".repeat(PROMPT_PREVIEW_LEN + 100), secret_tail); |
| 872 | let messages = vec![build_metadata_message( |
| 873 | &prompt, |
| 874 | Some("answer from the long context"), |
| 875 | 0, |
| 876 | None, |
| 877 | None, |
| 878 | )]; |
| 879 | |
| 880 | let request = build_root_request("root-model", &messages, &rlm_system_prompt()); |
| 881 | let payload = serde_json::to_string(&request).expect("request should serialize"); |
| 882 | |
| 883 | assert!(payload.contains(&format!("- Length: {} chars", prompt.chars().count()))); |
| 884 | assert!( |
| 885 | !payload.contains(secret_tail), |
| 886 | "root LLM request leaked the non-preview tail of context" |
| 887 | ); |
| 888 | } |
| 889 | |
| 890 | #[test] |
| 891 | fn build_metadata_with_iteration_shows_previous_code() { |
| 892 | let msg = build_metadata_message("Test prompt", None, 3, Some("print('hi')"), Some("hi")); |
| 893 | let text = extract_text_blocks(&msg.content); |
| 894 | assert!(text.contains("round 3")); |
| 895 | assert!(text.contains("print('hi')")); |
| 896 | assert!(text.contains("hi")); |
| 897 | } |
| 898 | |
| 899 | #[test] |
| 900 | fn build_metadata_includes_root_prompt() { |
| 901 | let msg = build_metadata_message( |
| 902 | "long context", |
| 903 | Some("Summarize the security model"), |
| 904 | 1, |
| 905 | Some("# noop"), |
| 906 | Some("ok"), |
| 907 | ); |
| 908 | let text = extract_text_blocks(&msg.content); |
| 909 | assert!(text.contains("Original task")); |
| 910 | assert!(text.contains("Summarize the security model")); |
| 911 | } |
| 912 | |
| 913 | #[test] |
| 914 | fn truncate_text_leaves_short_alone() { |
| 915 | assert_eq!(truncate_text("hello", 100), "hello"); |
| 916 | } |
| 917 | |
| 918 | #[test] |
| 919 | fn truncate_text_shortens_long_text() { |
| 920 | let long = "a".repeat(1000); |
| 921 | let truncated = truncate_text(&long, 10); |
| 922 | assert_eq!(truncated.chars().count(), 10); |
| 923 | assert!(truncated.ends_with("...")); |
| 924 | } |
| 925 | |
| 926 | #[test] |
| 927 | fn truncate_text_is_unicode_safe() { |
| 928 | let s = "日本語テスト"; |
| 929 | let out = truncate_text(s, 4); |
| 930 | assert_eq!(out.chars().count(), 4); |
| 931 | assert!(out.ends_with("...")); |
| 932 | assert!(std::str::from_utf8(out.as_bytes()).is_ok()); |
| 933 | } |
| 934 | |
| 935 | #[test] |
| 936 | fn extract_text_blocks_joins_text() { |
| 937 | let blocks = vec![ |
| 938 | ContentBlock::Text { |
| 939 | text: "first".to_string(), |
| 940 | cache_control: None, |
| 941 | }, |
| 942 | ContentBlock::Thinking { |
| 943 | thinking: "skip".to_string(), |
| 944 | }, |
| 945 | ContentBlock::Text { |
| 946 | text: "second".to_string(), |
| 947 | cache_control: None, |
| 948 | }, |
| 949 | ]; |
| 950 | assert_eq!(extract_text_blocks(&blocks), "first\nsecond"); |
| 951 | } |
| 952 | |
| 953 | #[test] |
| 954 | fn metadata_msg_role_is_user() { |
| 955 | let msg = build_metadata_message("test", None, 0, None, None); |
| 956 | assert_eq!(msg.role, "user"); |
| 957 | } |
| 958 | |
| 959 | #[test] |
| 960 | fn summarize_code_keeps_short() { |
| 961 | assert_eq!(summarize_code("a\nb\nc"), "a\nb\nc"); |
| 962 | } |
| 963 | |
| 964 | #[test] |
| 965 | fn summarize_code_compresses_long() { |
| 966 | let lines: Vec<String> = (0..20).map(|i| format!("line{i}")).collect(); |
| 967 | let code = lines.join("\n"); |
| 968 | let s = summarize_code(&code); |
| 969 | assert!(s.starts_with("20 lines:")); |
| 970 | assert!(s.contains("line0")); |
| 971 | assert!(s.contains("line19")); |
| 972 | assert!(s.contains("…")); |
| 973 | } |
| 974 | } |
| 975 |