| 1 | //! Pure RLM result and feedback facts. The canonical Engine owns every model |
| 2 | //! request, code round, cancellation, session message and permission settlement. |
| 3 | use codewhale_models::{ContentBlock, Message, Role, Usage}; |
| 4 | use std::time::Duration; |
| 5 | |
| 6 | pub(crate) const MAX_RLM_ITERATIONS: u32 = 25; |
| 7 | pub(crate) const MAX_CONSECUTIVE_NO_CODE: u32 = 3; |
| 8 | pub(crate) const STDOUT_METADATA_PREVIEW_LEN: usize = 800; |
| 9 | const PROMPT_PREVIEW_LEN: usize = 500; |
| 10 | |
| 11 | /// How an RLM turn ended. |
| 12 | #[derive(Debug, Clone, Copy, PartialEq, Eq)] |
| 13 | pub enum RlmTermination { |
| 14 | /// `FINAL(value)` was called inside the REPL or `FINAL(...)` appeared |
| 15 | /// at the top of the model's response on its own line. |
| 16 | Final, |
| 17 | /// The model failed to emit a `repl` block for too many rounds in a |
| 18 | /// row. The accumulated last response text is surfaced as the answer |
| 19 | /// rather than being thrown away. |
| 20 | NoCode, |
| 21 | /// Iteration cap reached without `FINAL`. The last root response is |
| 22 | /// surfaced as the answer alongside the error. |
| 23 | Exhausted, |
| 24 | /// Hard error — LLM call failed, REPL crashed, timeout. |
| 25 | Error, |
| 26 | } |
| 27 | |
| 28 | /// Per-round trace entry. Surfaced in the tool result so the user can see |
| 29 | /// exactly what the sub-agent did. |
| 30 | #[derive(Debug, Clone)] |
| 31 | pub struct RlmRoundTrace { |
| 32 | pub round: u32, |
| 33 | pub code_summary: String, |
| 34 | pub stdout_preview: String, |
| 35 | pub had_error: bool, |
| 36 | pub rpc_count: u32, |
| 37 | pub elapsed_ms: u64, |
| 38 | } |
| 39 | |
| 40 | /// Result of an RLM turn. |
| 41 | #[derive(Debug, Clone)] |
| 42 | pub struct RlmTurnResult { |
| 43 | pub answer: String, |
| 44 | pub iterations: u32, |
| 45 | pub duration: Duration, |
| 46 | pub error: Option<String>, |
| 47 | pub usage: Usage, |
| 48 | /// One exact frozen route/quote receipt per admitted provider request. |
| 49 | /// Distinct calls are never coalesced, even when they share a route. |
| 50 | pub routed_usage: Vec<crate::cost_status::RuntimeUsageRecord>, |
| 51 | /// Exact routes for provider-success responses that omitted authoritative |
| 52 | /// usage metadata. |
| 53 | pub routed_usage_drop_records: Vec<crate::cost_status::RuntimeUsageDropRecord>, |
| 54 | pub routed_usage_dropped_records: u64, |
| 55 | pub termination: RlmTermination, |
| 56 | /// Per-round trace. Empty when the loop never reached the REPL. |
| 57 | pub trace: Vec<RlmRoundTrace>, |
| 58 | /// Total sub-LLM RPCs made by the sub-agent (sum of `rpc_count` across |
| 59 | /// rounds). Useful for verifying that the model engaged with `context` |
| 60 | /// rather than answering directly. |
| 61 | pub total_rpcs: u32, |
| 62 | } |
| 63 | |
| 64 | impl RlmTurnResult { |
| 65 | pub(crate) fn failed(error: String, duration: Duration) -> Self { |
| 66 | Self { |
| 67 | answer: String::new(), |
| 68 | iterations: 0, |
| 69 | duration, |
| 70 | error: Some(error), |
| 71 | usage: Usage::default(), |
| 72 | routed_usage: Vec::new(), |
| 73 | routed_usage_drop_records: Vec::new(), |
| 74 | routed_usage_dropped_records: 0, |
| 75 | termination: RlmTermination::Error, |
| 76 | trace: Vec::new(), |
| 77 | total_rpcs: 0, |
| 78 | } |
| 79 | } |
| 80 | pub(crate) fn require_answer_or_error(mut self) -> Self { |
| 81 | if self.termination != RlmTermination::Final |
| 82 | && self.answer.trim().is_empty() |
| 83 | && self.error.is_none() |
| 84 | { |
| 85 | self.error = Some(format!( |
| 86 | "RLM ended ({:?}) after {} iteration(s) with an empty answer", |
| 87 | self.termination, self.iterations |
| 88 | )); |
| 89 | } |
| 90 | self |
| 91 | } |
| 92 | } |
| 93 | |
| 94 | pub(crate) fn metadata_text( |
| 95 | prompt: &str, |
| 96 | iteration: u32, |
| 97 | code: Option<&str>, |
| 98 | stdout: Option<&str>, |
| 99 | ) -> String { |
| 100 | extract_text_blocks(&build_metadata_message(prompt, None, iteration, code, stdout).content) |
| 101 | } |
| 102 | |
| 103 | fn build_metadata_message( |
| 104 | prompt: &str, |
| 105 | root_prompt: Option<&str>, |
| 106 | iteration: u32, |
| 107 | previous_code: Option<&str>, |
| 108 | previous_stdout: Option<&str>, |
| 109 | ) -> Message { |
| 110 | let prompt_len = prompt.chars().count(); |
| 111 | let prompt_preview = truncate_text(prompt, PROMPT_PREVIEW_LEN); |
| 112 | |
| 113 | let mut parts = Vec::new(); |
| 114 | parts.push(format!("## REPL state (round {iteration})")); |
| 115 | parts.push(String::new()); |
| 116 | if let Some(rp) = root_prompt |
| 117 | && !rp.trim().is_empty() |
| 118 | { |
| 119 | parts.push("**Original task** (re-shown every round)".to_string()); |
| 120 | parts.push(format!("> {}", truncate_text(rp.trim(), 600))); |
| 121 | parts.push(String::new()); |
| 122 | } |
| 123 | parts.push("**`_context`** — the long input lives in the REPL only".to_string()); |
| 124 | parts.push(format!("- Length: {prompt_len} chars")); |
| 125 | parts.push(format!("- Preview: \"{prompt_preview}\"")); |
| 126 | parts.push(String::new()); |
| 127 | |
| 128 | parts.push("**REPL helpers** (use inside ```repl blocks)".to_string()); |
| 129 | parts.push( |
| 130 | "- `_context` / `_ctx` / `content` — the full input string" |
| 131 | .to_string(), |
| 132 | ); |
| 133 | parts.push( |
| 134 | "- `len(_context)` / `_context[a:b]` / `_context.splitlines()` — slice it".to_string(), |
| 135 | ); |
| 136 | parts.push( |
| 137 | "- `chunk(max_chars=20000, overlap=0)` — full-coverage chunks with index/start/end/text" |
| 138 | .to_string(), |
| 139 | ); |
| 140 | parts.push( |
| 141 | "- `chunk_coverage(chunks)` — coverage report for chunk output".to_string(), |
| 142 | ); |
| 143 | parts.push( |
| 144 | "- `llm_query(prompt, model=None)` — one-shot child LLM; `model` is ignored and child calls keep the captured Core route" |
| 145 | .to_string(), |
| 146 | ); |
| 147 | parts.push( |
| 148 | "- `llm_query_batched([p1, p2, ...], dependency_mode=\"independent\")` — concurrent fan-out for independent prompts only; `model` is ignored" |
| 149 | .to_string(), |
| 150 | ); |
| 151 | parts.push( |
| 152 | "- `rlm_query(prompt, model=None)` — recursive sub-RLM; `model` is ignored" |
| 153 | .to_string(), |
| 154 | ); |
| 155 | parts.push( |
| 156 | "- `rlm_query_batched([p1, p2, ...], dependency_mode=\"independent\")` — concurrent recursive sub-RLMs for independent prompts only; `model` is ignored" |
| 157 | .to_string(), |
| 158 | ); |
| 159 | parts.push( |
| 160 | "- `sub_query_sequence(prompt, slices)` — sequential child calls for A->B dependencies and rollback-sensitive work" |
| 161 | .to_string(), |
| 162 | ); |
| 163 | parts.push( |
| 164 | "- Batch safety: never batch dependent steps, global-state refactors, schema migrations, or rollback-sensitive tasks" |
| 165 | .to_string(), |
| 166 | ); |
| 167 | parts.push("- `SHOW_VARS()` — list user variables".to_string()); |
| 168 | parts.push("- `repl_set(name, value)` / `repl_get(name)` — explicit store".to_string()); |
| 169 | parts.push( |
| 170 | "- `FINAL(value)` — end the loop with this answer".to_string(), |
| 171 | ); |
| 172 | parts.push( |
| 173 | "- `FINAL_VAR(name)` — end the loop with a variable's value" |
| 174 | .to_string(), |
| 175 | ); |
| 176 | parts.push(String::new()); |
| 177 | |
| 178 | if iteration > 0 { |
| 179 | parts.push("**Previous round**".to_string()); |
| 180 | if let Some(code) = previous_code { |
| 181 | parts.push(format!("- Code: {}", summarize_code(code))); |
| 182 | } |
| 183 | if let Some(stdout) = previous_stdout { |
| 184 | let stdout_clean = stdout.trim(); |
| 185 | if !stdout_clean.is_empty() { |
| 186 | parts.push(format!("- Stdout preview: \"{stdout_clean}\"")); |
| 187 | } else { |
| 188 | parts.push("- Stdout: (empty)".to_string()); |
| 189 | } |
| 190 | } |
| 191 | } |
| 192 | |
| 193 | let text = parts.join("\n"); |
| 194 | |
| 195 | Message { |
| 196 | role: Role::User, |
| 197 | content: vec![ContentBlock::Text { |
| 198 | text, |
| 199 | cache_control: None, |
| 200 | }], |
| 201 | } |
| 202 | } |
| 203 | |
| 204 | pub(crate) fn summarize_code(code: &str) -> String { |
| 205 | let lines: Vec<&str> = code.lines().collect(); |
| 206 | if lines.len() <= 8 { |
| 207 | return code.to_string(); |
| 208 | } |
| 209 | let head = lines[..4].join("\n"); |
| 210 | let tail = lines[lines.len() - 4..].join("\n"); |
| 211 | format!("{} lines:\n{head}\n…\n{tail}", lines.len()) |
| 212 | } |
| 213 | |
| 214 | fn extract_text_blocks(blocks: &[ContentBlock]) -> String { |
| 215 | blocks |
| 216 | .iter() |
| 217 | .filter_map(|b| match b { |
| 218 | ContentBlock::Text { text, .. } => Some(text.as_str()), |
| 219 | _ => None, |
| 220 | }) |
| 221 | .collect::<Vec<_>>() |
| 222 | .join("\n") |
| 223 | } |
| 224 | |
| 225 | /// Extract the first ` ```repl ` block from `text`. Falls back to |
| 226 | /// ` ```python `/`` ```py `` for compatibility with prompts that learned |
| 227 | /// the older fence style. |
| 228 | /// |
| 229 | /// The opening fence must start its own line (up to three spaces of indent, |
| 230 | /// as in Markdown) with no other info string, so prose that mentions a fence |
| 231 | /// mid-line is never code to run. |
| 232 | pub(crate) fn extract_repl_code(text: &str) -> Option<String> { |
| 233 | let mut lines = text.split_inclusive('\n'); |
| 234 | let mut offset = 0; |
| 235 | let code_start = loop { |
| 236 | let line = lines.next()?; |
| 237 | offset += line.len(); |
| 238 | let indent = line.len() - line.trim_start_matches(' ').len(); |
| 239 | if indent <= 3 |
| 240 | && line[indent..].strip_prefix("```").is_some_and(|info| { |
| 241 | matches!(info.trim(), "repl" | "python" | "py") && line.ends_with('\n') |
| 242 | }) |
| 243 | { |
| 244 | break offset; |
| 245 | } |
| 246 | }; |
| 247 | let after_fence = &text[code_start..]; |
| 248 | let end_idx = if after_fence.starts_with("```") { |
| 249 | 0 |
| 250 | } else { |
| 251 | after_fence.find("\n```")? |
| 252 | }; |
| 253 | |
| 254 | let code = after_fence[..end_idx].trim().to_string(); |
| 255 | if code.is_empty() { |
| 256 | return None; |
| 257 | } |
| 258 | Some(code) |
| 259 | } |
| 260 | |
| 261 | /// Parse a top-level `FINAL(...)` directive from the model's raw text. |
| 262 | /// Mirrors the reference RLM's `find_final_answer`: directive must appear |
| 263 | /// at the start of a line, *outside* any code fence. |
| 264 | pub(crate) fn parse_text_final(text: &str) -> Option<String> { |
| 265 | let outside_fence = strip_code_fences(text); |
| 266 | |
| 267 | for line in outside_fence.lines() { |
| 268 | let trimmed = line.trim_start(); |
| 269 | if trimmed.starts_with("FINAL_VAR(") { |
| 270 | // FINAL_VAR can't be resolved from text alone — defer to REPL. |
| 271 | continue; |
| 272 | } |
| 273 | if let Some(rest) = trimmed.strip_prefix("FINAL(") { |
| 274 | let inner = rest.trim_end(); |
| 275 | if let Some(end) = inner.rfind(')') { |
| 276 | let value = inner[..end].trim(); |
| 277 | if !value.is_empty() { |
| 278 | return Some(strip_quotes(value)); |
| 279 | } |
| 280 | } |
| 281 | } |
| 282 | } |
| 283 | None |
| 284 | } |
| 285 | |
| 286 | fn strip_code_fences(text: &str) -> String { |
| 287 | let mut out = String::with_capacity(text.len()); |
| 288 | let mut in_fence = false; |
| 289 | for line in text.lines() { |
| 290 | if line.trim_start().starts_with("```") { |
| 291 | in_fence = !in_fence; |
| 292 | continue; |
| 293 | } |
| 294 | if !in_fence { |
| 295 | out.push_str(line); |
| 296 | out.push('\n'); |
| 297 | } |
| 298 | } |
| 299 | out |
| 300 | } |
| 301 | |
| 302 | fn strip_quotes(s: &str) -> String { |
| 303 | let bytes = s.as_bytes(); |
| 304 | if bytes.len() >= 2 |
| 305 | && ((bytes[0] == b'"' && bytes[bytes.len() - 1] == b'"') |
| 306 | || (bytes[0] == b'\'' && bytes[bytes.len() - 1] == b'\'')) |
| 307 | { |
| 308 | return s[1..s.len() - 1].to_string(); |
| 309 | } |
| 310 | s.to_string() |
| 311 | } |
| 312 | |
| 313 | pub(crate) fn truncate_text(text: &str, max_chars: usize) -> String { |
| 314 | let count = text.chars().count(); |
| 315 | if count <= max_chars { |
| 316 | return text.to_string(); |
| 317 | } |
| 318 | let take = max_chars.saturating_sub(3); |
| 319 | let mut result: String = text.chars().take(take).collect(); |
| 320 | result.push_str("..."); |
| 321 | result |
| 322 | } |
| 323 | |
| 324 | #[cfg(test)] |
| 325 | mod tests { |
| 326 | use super::*; |
| 327 | use crate::core::engine::tests::rlm_host::{Replies, run_admitted_fixture, run_fixture}; |
| 328 | use crate::core::events::Event; |
| 329 | use crate::llm_client::mock::MockLlmClient; |
| 330 | use crate::rlm::bridge::RlmUsageAccumulator; |
| 331 | use codewhale_models::{MessageRequest, MessageResponse}; |
| 332 | use std::path::PathBuf; |
| 333 | use std::sync::Arc; |
| 334 | use tokio::sync::mpsc; |
| 335 | |
| 336 | /// One model round whose code writes a marker file, run under `gate`. |
| 337 | async fn marker_round( |
| 338 | gate: Option<crate::tools::codemode::NestedCallGate>, |
| 339 | ) -> (RlmTurnResult, bool) { |
| 340 | let workspace = tempfile::tempdir().expect("tempdir"); |
| 341 | let marker = workspace.path().join("rlm-round-executed.txt"); |
| 342 | let marker_literal = serde_json::to_string(&marker.to_string_lossy()).expect("literal"); |
| 343 | let mock = Arc::new(MockLlmClient::new(Vec::new())); |
| 344 | mock.push_message_response(MessageResponse { |
| 345 | id: "mock_gated_rlm".to_string(), |
| 346 | r#type: "message".to_string(), |
| 347 | role: "assistant".to_string(), |
| 348 | content: vec![ContentBlock::Text { |
| 349 | text: format!( |
| 350 | "```repl\nfrom pathlib import Path\nPath({marker_literal}).write_text('executed')\nFINAL('ran')\n```" |
| 351 | ), |
| 352 | cache_control: None, |
| 353 | }], |
| 354 | model: "mock-model".to_string(), |
| 355 | stop_reason: Some("end_turn".to_string()), |
| 356 | stop_sequence: None, |
| 357 | container: None, |
| 358 | usage: Usage::default(), |
| 359 | }); |
| 360 | let client: Arc<dyn Replies> = mock; |
| 361 | let (tx, _rx) = mpsc::channel(8); |
| 362 | let result = run_admitted_fixture( |
| 363 | client, |
| 364 | "root-model".to_string(), |
| 365 | "long context".to_string(), |
| 366 | None, |
| 367 | "child-model".to_string(), |
| 368 | tx, |
| 369 | 0, |
| 370 | RlmUsageAccumulator::new(), |
| 371 | tokio::time::Instant::now() + Duration::from_secs(60), |
| 372 | gate, |
| 373 | ) |
| 374 | .await; |
| 375 | (result, marker.exists()) |
| 376 | } |
| 377 | |
| 378 | #[tokio::test] |
| 379 | async fn code_round_runs_only_when_the_gate_admits_exactly_that_code() { |
| 380 | use crate::tools::codemode::{NestedCallGate, NestedCallVerdict, NestedDecision}; |
| 381 | |
| 382 | let (result, ran) = marker_round(Some(NestedCallGate::admitting_for_test())).await; |
| 383 | assert!(ran, "an admitted round runs: {:?}", result.error); |
| 384 | assert_eq!(result.termination, RlmTermination::Final); |
| 385 | |
| 386 | let (result, ran) = marker_round(None).await; |
| 387 | assert!(!ran, "no gate, no code"); |
| 388 | assert_eq!(result.termination, RlmTermination::Error); |
| 389 | let error = result.error.expect("refusal is reported"); |
| 390 | assert!(error.contains("no permission gate"), "{error}"); |
| 391 | |
| 392 | let refusing = NestedCallGate::answering_for_test(|_, _| NestedCallVerdict::Refused { |
| 393 | error: crate::tools::spec::ToolError::permission_denied("not approved"), |
| 394 | decision: NestedDecision::Denied, |
| 395 | }); |
| 396 | let (result, ran) = marker_round(Some(refusing)).await; |
| 397 | assert!(!ran, "a refused round does not run"); |
| 398 | let error = result.error.expect("refusal is reported"); |
| 399 | assert!(error.contains("not approved"), "{error}"); |
| 400 | |
| 401 | // A hook-rewritten or otherwise different admitted input is not |
| 402 | // permission for the code the model wrote. |
| 403 | let rewriting = NestedCallGate::answering_for_test(|name, _| NestedCallVerdict::Run { |
| 404 | name: name.to_string(), |
| 405 | input: serde_json::json!({ "code": "print('something else')" }), |
| 406 | supports_parallel: false, |
| 407 | decision: NestedDecision::Auto, |
| 408 | hook_context: None, |
| 409 | }); |
| 410 | let (result, ran) = marker_round(Some(rewriting)).await; |
| 411 | assert!(!ran, "a different admitted input does not run this code"); |
| 412 | assert!( |
| 413 | result |
| 414 | .error |
| 415 | .is_some_and(|error| error.contains("was not this code")) |
| 416 | ); |
| 417 | } |
| 418 | |
| 419 | #[tokio::test] |
| 420 | async fn max_tokens_complete_repl_is_not_executed_or_accepted() { |
| 421 | let workspace = tempfile::tempdir().expect("tempdir"); |
| 422 | let marker = workspace.path().join("truncated-repl-executed.txt"); |
| 423 | let marker_literal = serde_json::to_string(&marker.to_string_lossy()) |
| 424 | .expect("marker path should serialize as a Python string literal"); |
| 425 | let partial = format!( |
| 426 | "```repl\nfrom pathlib import Path\nPath({marker_literal}).write_text('executed')\nFINAL('partial answer')\n```" |
| 427 | ); |
| 428 | let usage = Usage { |
| 429 | input_tokens: 17, |
| 430 | output_tokens: 4096, |
| 431 | ..Usage::default() |
| 432 | }; |
| 433 | let mock = Arc::new(MockLlmClient::new(Vec::new())); |
| 434 | mock.push_message_response(MessageResponse { |
| 435 | id: "mock_truncated_rlm".to_string(), |
| 436 | r#type: "message".to_string(), |
| 437 | role: "assistant".to_string(), |
| 438 | content: vec![ContentBlock::Text { |
| 439 | text: partial, |
| 440 | cache_control: None, |
| 441 | }], |
| 442 | model: "mock-model".to_string(), |
| 443 | stop_reason: Some("max_tokens".to_string()), |
| 444 | stop_sequence: None, |
| 445 | container: None, |
| 446 | usage: usage.clone(), |
| 447 | }); |
| 448 | let client: Arc<dyn Replies> = mock.clone(); |
| 449 | let (tx, _rx) = mpsc::channel(8); |
| 450 | |
| 451 | let result = run_fixture( |
| 452 | client, |
| 453 | "root-model".to_string(), |
| 454 | "long context".to_string(), |
| 455 | None, |
| 456 | "child-model".to_string(), |
| 457 | tx, |
| 458 | 0, |
| 459 | ) |
| 460 | .await; |
| 461 | |
| 462 | assert_eq!(result.termination, RlmTermination::Error); |
| 463 | assert!( |
| 464 | result.answer.is_empty(), |
| 465 | "partial FINAL must not be accepted" |
| 466 | ); |
| 467 | let error = result.error.expect("truncation must fail the RLM turn"); |
| 468 | assert!(error.contains("incomplete"), "{error}"); |
| 469 | assert!(error.contains("max_tokens"), "{error}"); |
| 470 | assert_eq!(result.usage, usage, "billed usage must still be charged"); |
| 471 | assert_eq!(result.routed_usage.len(), 1); |
| 472 | assert_eq!(result.routed_usage[0].usage.usage, usage); |
| 473 | assert_eq!(result.routed_usage_dropped_records, 0); |
| 474 | assert_eq!(mock.call_count(), 1, "truncation must not retry"); |
| 475 | assert!( |
| 476 | !marker.exists(), |
| 477 | "complete-looking code from a truncated response must not execute" |
| 478 | ); |
| 479 | } |
| 480 | |
| 481 | #[tokio::test] |
| 482 | async fn root_provider_success_without_usage_retains_exact_missing_receipt() { |
| 483 | let mock = Arc::new(MockLlmClient::new(Vec::new())); |
| 484 | mock.push_message_response(MessageResponse { |
| 485 | id: "mock_missing_rlm_usage".to_string(), |
| 486 | r#type: "message".to_string(), |
| 487 | role: "assistant".to_string(), |
| 488 | content: vec![ContentBlock::Text { |
| 489 | text: "partial".to_string(), |
| 490 | cache_control: None, |
| 491 | }], |
| 492 | model: "mock-model".to_string(), |
| 493 | stop_reason: Some("max_tokens".to_string()), |
| 494 | stop_sequence: None, |
| 495 | container: None, |
| 496 | usage: Usage::default(), |
| 497 | }); |
| 498 | let client: Arc<dyn Replies> = mock; |
| 499 | let (tx, _rx) = mpsc::channel(8); |
| 500 | |
| 501 | let result = run_fixture( |
| 502 | client, |
| 503 | "root-model".to_string(), |
| 504 | "long context".to_string(), |
| 505 | None, |
| 506 | "child-model".to_string(), |
| 507 | tx, |
| 508 | 0, |
| 509 | ) |
| 510 | .await; |
| 511 | |
| 512 | assert_eq!(result.termination, RlmTermination::Error); |
| 513 | assert_eq!(result.usage, Usage::default()); |
| 514 | assert!(result.routed_usage.is_empty()); |
| 515 | assert_eq!(result.routed_usage_drop_records.len(), 1); |
| 516 | assert_eq!(result.routed_usage_dropped_records, 1); |
| 517 | assert_eq!( |
| 518 | result.routed_usage_drop_records[0].route.model, |
| 519 | "root-model" |
| 520 | ); |
| 521 | } |
| 522 | |
| 523 | fn text_response(text: &str) -> MessageResponse { |
| 524 | MessageResponse { |
| 525 | id: "mock_rlm_round".to_string(), |
| 526 | r#type: "message".to_string(), |
| 527 | role: "assistant".to_string(), |
| 528 | content: vec![ContentBlock::Text { |
| 529 | text: text.to_string(), |
| 530 | cache_control: None, |
| 531 | }], |
| 532 | model: "mock-model".to_string(), |
| 533 | stop_reason: Some("end_turn".to_string()), |
| 534 | stop_sequence: None, |
| 535 | container: None, |
| 536 | usage: Usage { |
| 537 | input_tokens: 5, |
| 538 | output_tokens: 5, |
| 539 | ..Usage::default() |
| 540 | }, |
| 541 | } |
| 542 | } |
| 543 | |
| 544 | /// #6511: an exhausted loop used to return `answer: String::new()` and |
| 545 | /// drop the middle of its history after 20 messages. |
| 546 | #[tokio::test] |
| 547 | async fn exhausted_loop_returns_last_response_and_keeps_whole_history() { |
| 548 | let mock = Arc::new(MockLlmClient::new(Vec::new())); |
| 549 | for round in 0..MAX_RLM_ITERATIONS { |
| 550 | mock.push_message_response(text_response(&format!( |
| 551 | "```repl\nprint('round {round}')\n```" |
| 552 | ))); |
| 553 | } |
| 554 | let client: Arc<dyn Replies> = mock.clone(); |
| 555 | let (tx, mut rx) = mpsc::channel(1024); |
| 556 | |
| 557 | let result = run_fixture( |
| 558 | client, |
| 559 | "root-model".to_string(), |
| 560 | "long context".to_string(), |
| 561 | None, |
| 562 | "child-model".to_string(), |
| 563 | tx, |
| 564 | 0, |
| 565 | ) |
| 566 | .await; |
| 567 | |
| 568 | assert_eq!(result.termination, RlmTermination::Exhausted); |
| 569 | assert_eq!(result.iterations, MAX_RLM_ITERATIONS); |
| 570 | let last_round = MAX_RLM_ITERATIONS - 1; |
| 571 | assert!( |
| 572 | result.answer.contains(&format!("round {last_round}")), |
| 573 | "exhaustion must surface the last root response, got {:?}", |
| 574 | result.answer |
| 575 | ); |
| 576 | assert!( |
| 577 | result |
| 578 | .error |
| 579 | .as_deref() |
| 580 | .is_some_and(|e| e.contains("exhausted")), |
| 581 | "{:?}", |
| 582 | result.error |
| 583 | ); |
| 584 | |
| 585 | let requests = mock.captured_requests(); |
| 586 | assert_eq!(requests.len(), MAX_RLM_ITERATIONS as usize); |
| 587 | // Initial metadata plus (code, result) per completed round: nothing |
| 588 | // from the middle is dropped. |
| 589 | let last = requests.last().expect("last root request"); |
| 590 | assert_eq!( |
| 591 | last.messages.len(), |
| 592 | 1 + 2 * (MAX_RLM_ITERATIONS as usize - 1) |
| 593 | ); |
| 594 | |
| 595 | let mut statuses = Vec::new(); |
| 596 | while let Ok(event) = rx.try_recv() { |
| 597 | if let Event::Status { message } = event { |
| 598 | statuses.push(message); |
| 599 | } |
| 600 | } |
| 601 | assert!( |
| 602 | statuses |
| 603 | .iter() |
| 604 | .any(|line| line.contains("RLM finished: Exhausted")), |
| 605 | "the loop must log how it ended: {statuses:#?}" |
| 606 | ); |
| 607 | } |
| 608 | |
| 609 | struct PendingAfterResponses(MockLlmClient, usize); |
| 610 | |
| 611 | impl Replies for PendingAfterResponses { |
| 612 | fn effective_route_envelope( |
| 613 | &self, |
| 614 | model: &str, |
| 615 | dispatched_at: chrono::DateTime<chrono::Utc>, |
| 616 | ) -> crate::cost_status::EffectiveRouteEnvelope { |
| 617 | self.0.effective_route_envelope(model, dispatched_at) |
| 618 | } |
| 619 | |
| 620 | fn effective_max_output_tokens(&self, model: &str) -> u32 { |
| 621 | self.0.effective_max_output_tokens(model) |
| 622 | } |
| 623 | |
| 624 | fn create_message_boxed( |
| 625 | &self, |
| 626 | request: MessageRequest, |
| 627 | ) -> std::pin::Pin< |
| 628 | Box<dyn std::future::Future<Output = anyhow::Result<MessageResponse>> + Send + '_>, |
| 629 | > { |
| 630 | if self.0.call_count() < self.1 { |
| 631 | self.0.create_message_boxed(request) |
| 632 | } else { |
| 633 | Box::pin(std::future::pending()) |
| 634 | } |
| 635 | } |
| 636 | } |
| 637 | |
| 638 | /// Checking time only between rounds cannot stop a pending root |
| 639 | /// request, a Python block, or an event send within the current round. |
| 640 | #[tokio::test] |
| 641 | async fn wall_clock_deadline_interrupts_pending_work_and_keeps_partial_result() { |
| 642 | for pending_model in [true, false] { |
| 643 | let partial = if pending_model { |
| 644 | "```repl\nprint(_os.environ['RLM_CONTEXT_FILE'])\n```" |
| 645 | } else { |
| 646 | "```repl\nimport time\ntime.sleep(60)\nFINAL('too late')\n```" |
| 647 | }; |
| 648 | let mock = MockLlmClient::new(Vec::new()); |
| 649 | mock.push_message_response(text_response(partial)); |
| 650 | let client = Arc::new(PendingAfterResponses(mock, 1)); |
| 651 | let (tx, mut rx) = mpsc::channel(32); |
| 652 | let usage = RlmUsageAccumulator::new(); |
| 653 | |
| 654 | let result = tokio::time::timeout( |
| 655 | Duration::from_secs(5), |
| 656 | run_admitted_fixture( |
| 657 | client.clone(), |
| 658 | "root-model".to_string(), |
| 659 | "long context".to_string(), |
| 660 | None, |
| 661 | "child-model".to_string(), |
| 662 | tx, |
| 663 | 0, |
| 664 | usage.clone(), |
| 665 | tokio::time::Instant::now() + Duration::from_secs(1), |
| 666 | Some(crate::tools::codemode::NestedCallGate::admitting_for_test()), |
| 667 | ), |
| 668 | ) |
| 669 | .await |
| 670 | .expect("the turn deadline must interrupt in-flight work"); |
| 671 | |
| 672 | assert_eq!(result.termination, RlmTermination::Error); |
| 673 | assert_eq!(result.answer, partial); |
| 674 | assert_eq!(result.iterations, if pending_model { 2 } else { 1 }); |
| 675 | assert!( |
| 676 | result |
| 677 | .error |
| 678 | .as_deref() |
| 679 | .unwrap() |
| 680 | .contains("wall-clock deadline") |
| 681 | ); |
| 682 | assert_eq!(result.usage, text_response(partial).usage); |
| 683 | assert_eq!(client.0.call_count(), 1); |
| 684 | let snapshot = usage.snapshot().await; |
| 685 | assert_eq!(snapshot.records.len(), 1, "keep completed usage"); |
| 686 | assert_eq!(snapshot.dropped_records, u64::from(pending_model)); |
| 687 | if pending_model { |
| 688 | assert_eq!(result.trace.len(), 1, "keep completed rounds"); |
| 689 | assert!(!result.trace[0].had_error); |
| 690 | let context_path = PathBuf::from(result.trace[0].stdout_preview.trim()); |
| 691 | assert!( |
| 692 | context_path.is_absolute(), |
| 693 | "REPL must report its context path" |
| 694 | ); |
| 695 | assert!(!context_path.exists(), "timeout must clean up context"); |
| 696 | } |
| 697 | let mut saw_timeout = false; |
| 698 | while let Ok(event) = rx.try_recv() { |
| 699 | if let Event::Status { message } = event { |
| 700 | saw_timeout |= message.contains("RLM finished: Error") |
| 701 | && message.contains("wall-clock deadline"); |
| 702 | } |
| 703 | } |
| 704 | assert!(saw_timeout, "timeout must record how the turn ended"); |
| 705 | let terminal_receipts: Vec<_> = snapshot |
| 706 | .nested_events |
| 707 | .iter() |
| 708 | .filter(|event| { |
| 709 | event["kind"] == "status" |
| 710 | && event["content"].as_str().is_some_and(|content| { |
| 711 | content.contains("RLM finished: Error") |
| 712 | && content.contains("wall-clock deadline") |
| 713 | }) |
| 714 | }) |
| 715 | .collect(); |
| 716 | assert_eq!(terminal_receipts.len(), 1, "one canonical terminal receipt"); |
| 717 | } |
| 718 | } |
| 719 | |
| 720 | #[tokio::test] |
| 721 | async fn wall_clock_deadline_returns_when_event_stream_is_full() { |
| 722 | // An empty response queue fails immediately and cannot test the deadline. |
| 723 | // Keep the admitted provider future pending while the host channel is full. |
| 724 | let client = Arc::new(PendingAfterResponses(MockLlmClient::new(Vec::new()), 0)); |
| 725 | let usage = RlmUsageAccumulator::new(); |
| 726 | let (tx, _rx) = mpsc::channel(1); |
| 727 | tx.try_send(Event::status("fixture occupies the caller event channel")) |
| 728 | .unwrap(); |
| 729 | let result = tokio::time::timeout( |
| 730 | Duration::from_secs(5), |
| 731 | run_admitted_fixture( |
| 732 | client.clone(), |
| 733 | "root-model".to_string(), |
| 734 | "long context".to_string(), |
| 735 | None, |
| 736 | "child-model".to_string(), |
| 737 | tx, |
| 738 | 0, |
| 739 | usage.clone(), |
| 740 | tokio::time::Instant::now() + Duration::from_secs(1), |
| 741 | Some(crate::tools::codemode::NestedCallGate::admitting_for_test()), |
| 742 | ), |
| 743 | ) |
| 744 | .await |
| 745 | .expect("deadline hand-back must not wait on a full event stream"); |
| 746 | |
| 747 | assert_eq!(result.termination, RlmTermination::Error); |
| 748 | assert!( |
| 749 | result |
| 750 | .error |
| 751 | .as_deref() |
| 752 | .unwrap() |
| 753 | .contains("wall-clock deadline") |
| 754 | ); |
| 755 | assert_eq!(client.0.call_count(), 0, "no scripted response completed"); |
| 756 | let snapshot = usage.snapshot().await; |
| 757 | assert_eq!(snapshot.dropped_records, 1, "one pending admitted request"); |
| 758 | assert_eq!( |
| 759 | snapshot |
| 760 | .nested_events |
| 761 | .iter() |
| 762 | .filter(|event| { |
| 763 | event["kind"] == "status" |
| 764 | && event["content"].as_str().is_some_and(|content| { |
| 765 | content.contains("RLM finished: Error") |
| 766 | && content.contains("wall-clock deadline") |
| 767 | }) |
| 768 | }) |
| 769 | .count(), |
| 770 | 1, |
| 771 | "full host channel must retain one canonical terminal receipt" |
| 772 | ); |
| 773 | } |
| 774 | |
| 775 | #[test] |
| 776 | fn empty_answer_is_never_returned_without_an_error() { |
| 777 | let empty = |termination| RlmTurnResult { |
| 778 | answer: " ".to_string(), |
| 779 | iterations: 2, |
| 780 | duration: Duration::ZERO, |
| 781 | error: None, |
| 782 | usage: Usage::default(), |
| 783 | routed_usage: Vec::new(), |
| 784 | routed_usage_drop_records: Vec::new(), |
| 785 | routed_usage_dropped_records: 0, |
| 786 | termination, |
| 787 | trace: Vec::new(), |
| 788 | total_rpcs: 0, |
| 789 | }; |
| 790 | let error = empty(RlmTermination::NoCode) |
| 791 | .require_answer_or_error() |
| 792 | .error |
| 793 | .expect("empty answer needs a reason"); |
| 794 | assert!(error.contains("empty answer"), "{error}"); |
| 795 | assert!(error.contains("NoCode"), "{error}"); |
| 796 | |
| 797 | // `FINAL("")` is an answer the model chose; callers keep getting "". |
| 798 | let deliberate = empty(RlmTermination::Final).require_answer_or_error(); |
| 799 | assert!(deliberate.error.is_none(), "{:?}", deliberate.error); |
| 800 | } |
| 801 | |
| 802 | #[test] |
| 803 | fn extract_repl_code_finds_simple_block() { |
| 804 | let text = "Here:\n```repl\nprint('hi')\n```\nEnd."; |
| 805 | let code = extract_repl_code(text).unwrap(); |
| 806 | assert_eq!(code, "print('hi')"); |
| 807 | } |
| 808 | |
| 809 | #[test] |
| 810 | fn extract_repl_code_falls_back_to_python_marker() { |
| 811 | let text = "Code:\n```python\nx = 1 + 2\n```"; |
| 812 | let code = extract_repl_code(text).unwrap(); |
| 813 | assert_eq!(code, "x = 1 + 2"); |
| 814 | } |
| 815 | |
| 816 | #[test] |
| 817 | fn extract_repl_code_returns_none_when_missing() { |
| 818 | assert!(extract_repl_code("Just text.").is_none()); |
| 819 | } |
| 820 | |
| 821 | #[test] |
| 822 | fn extract_repl_code_returns_none_on_empty_block() { |
| 823 | assert!(extract_repl_code("```repl\n\n```").is_none()); |
| 824 | } |
| 825 | |
| 826 | #[test] |
| 827 | fn extract_repl_code_handles_multiple_blocks() { |
| 828 | let text = "```repl\na=1\n```\n```repl\nb=2\n```"; |
| 829 | let code = extract_repl_code(text).unwrap(); |
| 830 | assert_eq!(code, "a=1"); |
| 831 | } |
| 832 | |
| 833 | #[test] |
| 834 | fn extract_repl_code_ignores_other_fences() { |
| 835 | let text = "```\nfoo\n```\n```repl\nreal_code()\n```"; |
| 836 | let code = extract_repl_code(text).unwrap(); |
| 837 | assert_eq!(code, "real_code()"); |
| 838 | } |
| 839 | |
| 840 | #[test] |
| 841 | fn extract_repl_code_requires_a_line_anchored_fence() { |
| 842 | assert!(extract_repl_code("see ```python\nx = 1\n```").is_none()); |
| 843 | assert!(extract_repl_code("a ```repl mention\nx = 1\n```").is_none()); |
| 844 | assert!(extract_repl_code("```python-output\nx = 1\n```").is_none()); |
| 845 | let text = "prose ```py inline\n```py\nreal()\n```"; |
| 846 | assert_eq!(extract_repl_code(text).as_deref(), Some("real()")); |
| 847 | } |
| 848 | |
| 849 | #[test] |
| 850 | fn parse_text_final_extracts_simple_value() { |
| 851 | let text = "OK.\nFINAL(42)\nThanks."; |
| 852 | assert_eq!(parse_text_final(text).as_deref(), Some("42")); |
| 853 | } |
| 854 | |
| 855 | #[test] |
| 856 | fn parse_text_final_strips_quotes() { |
| 857 | let text = "FINAL(\"the answer is yes\")"; |
| 858 | assert_eq!(parse_text_final(text).as_deref(), Some("the answer is yes")); |
| 859 | } |
| 860 | |
| 861 | #[test] |
| 862 | fn parse_text_final_ignores_inside_code_fence() { |
| 863 | let text = |
| 864 | "Some prose.\n```repl\n# Note: when ready, call FINAL(value)\nx = 1\n```\nMore prose."; |
| 865 | assert!(parse_text_final(text).is_none()); |
| 866 | } |
| 867 | |
| 868 | #[test] |
| 869 | fn parse_text_final_returns_none_when_absent() { |
| 870 | assert!(parse_text_final("just talking, no final.").is_none()); |
| 871 | } |
| 872 | |
| 873 | #[test] |
| 874 | fn build_metadata_contains_key_information() { |
| 875 | let msg = build_metadata_message("Hello, world!", None, 0, None, None); |
| 876 | let text = extract_text_blocks(&msg.content); |
| 877 | assert!(text.contains("context")); |
| 878 | assert!(text.contains("Hello, world!")); |
| 879 | assert!(text.contains("round 0")); |
| 880 | assert!(text.contains("llm_query")); |
| 881 | assert!(text.contains("rlm_query")); |
| 882 | assert!(text.contains("FINAL")); |
| 883 | } |
| 884 | |
| 885 | #[test] |
| 886 | fn build_metadata_truncates_long_context_without_leaking_tail() { |
| 887 | let secret_tail = "DO_NOT_LEAK_CONTEXT_TAIL"; |
| 888 | let prompt = format!("{}{}", "a".repeat(PROMPT_PREVIEW_LEN + 100), secret_tail); |
| 889 | let msg = build_metadata_message(&prompt, None, 0, None, None); |
| 890 | let text = extract_text_blocks(&msg.content); |
| 891 | |
| 892 | assert!(text.contains(&format!("- Length: {} chars", prompt.chars().count()))); |
| 893 | assert!(text.contains("- Preview: \"")); |
| 894 | assert!(text.contains("...")); |
| 895 | assert!( |
| 896 | !text.contains(secret_tail), |
| 897 | "metadata leaked the non-preview tail of context" |
| 898 | ); |
| 899 | } |
| 900 | |
| 901 | #[tokio::test] |
| 902 | async fn build_root_request_keeps_context_tail_out_of_root_payload() { |
| 903 | let secret_tail = "DO_NOT_LEAK_ROOT_REQUEST"; |
| 904 | let prompt = format!("{}{}", "a".repeat(PROMPT_PREVIEW_LEN + 100), secret_tail); |
| 905 | let mock = Arc::new(MockLlmClient::new(Vec::new())); |
| 906 | mock.push_message_response(MessageResponse { |
| 907 | id: "context-metadata".into(), |
| 908 | r#type: "message".into(), |
| 909 | role: "assistant".into(), |
| 910 | content: vec![ContentBlock::Text { |
| 911 | text: "```repl\nFINAL('metadata only')\n```".into(), |
| 912 | cache_control: None, |
| 913 | }], |
| 914 | model: "root-model".into(), |
| 915 | stop_reason: Some("end_turn".into()), |
| 916 | stop_sequence: None, |
| 917 | container: None, |
| 918 | usage: Usage::default(), |
| 919 | }); |
| 920 | let (tx, _rx) = mpsc::channel(64); |
| 921 | let result = run_fixture( |
| 922 | mock.clone(), |
| 923 | "root-model".into(), |
| 924 | prompt.clone(), |
| 925 | Some("answer from the long context".into()), |
| 926 | "child-model".into(), |
| 927 | tx, |
| 928 | 0, |
| 929 | ) |
| 930 | .await; |
| 931 | assert_eq!( |
| 932 | result.termination, |
| 933 | RlmTermination::Final, |
| 934 | "{:?}", |
| 935 | result.error |
| 936 | ); |
| 937 | let requests = mock.captured_requests(); |
| 938 | assert_eq!(requests.len(), 1); |
| 939 | let request = &requests[0]; |
| 940 | let payload = serde_json::to_string(request).expect("actual Core request serializes"); |
| 941 | assert_eq!(request.temperature, None); |
| 942 | assert_eq!(request.top_p, None); |
| 943 | assert!(payload.contains(&format!("- Length: {} chars", prompt.chars().count()))); |
| 944 | assert!( |
| 945 | !payload.contains(secret_tail), |
| 946 | "Core request leaked the non-preview context tail" |
| 947 | ); |
| 948 | assert!(payload.contains("Captured operator Core policy")); |
| 949 | assert!(payload.contains("answer from the long context")); |
| 950 | } |
| 951 | |
| 952 | #[test] |
| 953 | fn build_metadata_with_iteration_shows_previous_code() { |
| 954 | let msg = build_metadata_message("Test prompt", None, 3, Some("print('hi')"), Some("hi")); |
| 955 | let text = extract_text_blocks(&msg.content); |
| 956 | assert!(text.contains("round 3")); |
| 957 | assert!(text.contains("print('hi')")); |
| 958 | assert!(text.contains("hi")); |
| 959 | } |
| 960 | |
| 961 | #[test] |
| 962 | fn build_metadata_includes_root_prompt() { |
| 963 | let msg = build_metadata_message( |
| 964 | "long context", |
| 965 | Some("Summarize the security model"), |
| 966 | 1, |
| 967 | Some("# noop"), |
| 968 | Some("ok"), |
| 969 | ); |
| 970 | let text = extract_text_blocks(&msg.content); |
| 971 | assert!(text.contains("Original task")); |
| 972 | assert!(text.contains("Summarize the security model")); |
| 973 | } |
| 974 | |
| 975 | #[test] |
| 976 | fn truncate_text_leaves_short_alone() { |
| 977 | assert_eq!(truncate_text("hello", 100), "hello"); |
| 978 | } |
| 979 | |
| 980 | #[test] |
| 981 | fn truncate_text_shortens_long_text() { |
| 982 | let long = "a".repeat(1000); |
| 983 | let truncated = truncate_text(&long, 10); |
| 984 | assert_eq!(truncated.chars().count(), 10); |
| 985 | assert!(truncated.ends_with("...")); |
| 986 | } |
| 987 | |
| 988 | #[test] |
| 989 | fn truncate_text_is_unicode_safe() { |
| 990 | let s = "日本語テスト"; |
| 991 | let out = truncate_text(s, 4); |
| 992 | assert_eq!(out.chars().count(), 4); |
| 993 | assert!(out.ends_with("...")); |
| 994 | assert!(std::str::from_utf8(out.as_bytes()).is_ok()); |
| 995 | } |
| 996 | |
| 997 | #[test] |
| 998 | fn extract_text_blocks_joins_text() { |
| 999 | let blocks = vec![ |
| 1000 | ContentBlock::Text { |
| 1001 | text: "first".to_string(), |
| 1002 | cache_control: None, |
| 1003 | }, |
| 1004 | ContentBlock::Thinking { |
| 1005 | signature: None, |
| 1006 | state: None, |
| 1007 | thinking: "skip".to_string(), |
| 1008 | }, |
| 1009 | ContentBlock::Text { |
| 1010 | text: "second".to_string(), |
| 1011 | cache_control: None, |
| 1012 | }, |
| 1013 | ]; |
| 1014 | assert_eq!(extract_text_blocks(&blocks), "first\nsecond"); |
| 1015 | } |
| 1016 | |
| 1017 | #[test] |
| 1018 | fn metadata_msg_role_is_user() { |
| 1019 | let msg = build_metadata_message("test", None, 0, None, None); |
| 1020 | assert_eq!(msg.role, "user"); |
| 1021 | } |
| 1022 | |
| 1023 | #[test] |
| 1024 | fn summarize_code_keeps_short() { |
| 1025 | assert_eq!(summarize_code("a\nb\nc"), "a\nb\nc"); |
| 1026 | } |
| 1027 | |
| 1028 | #[test] |
| 1029 | fn summarize_code_compresses_long() { |
| 1030 | let lines: Vec<String> = (0..20).map(|i| format!("line{i}")).collect(); |
| 1031 | let code = lines.join("\n"); |
| 1032 | let s = summarize_code(&code); |
| 1033 | assert!(s.starts_with("20 lines:")); |
| 1034 | assert!(s.contains("line0")); |
| 1035 | assert!(s.contains("line19")); |
| 1036 | assert!(s.contains("…")); |
| 1037 | } |
| 1038 | } |
| 1039 |