| 1 | #[test] |
| 2 | fn stream_read_scenario() { |
| 3 | // Scenario consolidation of: stream_read_error_message_explains_retry_before_output, stream_read_error_message_explains_no_replay_after_output |
| 4 | // from stream_read_error_message_explains_retry_before_output |
| 5 | { |
| 6 | let message = super::stream_read_error_user_message( |
| 7 | "Stream read error: error decoding response body", |
| 8 | false, |
| 9 | ); |
| 10 | |
| 11 | assert!(message.contains("Provider stream connection dropped")); |
| 12 | assert!(message.contains("No output had streamed yet")); |
| 13 | assert!(message.contains("retry automatically")); |
| 14 | assert!(message.contains("Stream read error: error decoding response body")); |
| 15 | } |
| 16 | // from stream_read_error_message_explains_no_replay_after_output |
| 17 | { |
| 18 | let message = super::stream_read_error_user_message( |
| 19 | "Stream read error: error decoding response body", |
| 20 | true, |
| 21 | ); |
| 22 | |
| 23 | assert!(message.contains("Provider stream connection dropped")); |
| 24 | assert!(message.contains("Some output had already streamed")); |
| 25 | assert!(message.contains("risking duplicated output")); |
| 26 | assert!(message.contains("Stream read error: error decoding response body")); |
| 27 | assert_eq!( |
| 28 | crate::error_taxonomy::classify_error_message(&message), |
| 29 | crate::error_taxonomy::ErrorCategory::Network |
| 30 | ); |
| 31 | } |
| 32 | } |
| 33 | |
| 34 | #[test] |
| 35 | fn stream_retry_budget_caps_transparent_retries_at_two() { |
| 36 | // Case 4 from issue #103: after MAX_TRANSPARENT_STREAM_RETRIES attempts |
| 37 | // we stop trying transparently and let the outer error path surface. |
| 38 | // (The outer per-turn `stream_retry_attempts` retry is a separate layer |
| 39 | // and is still in effect at the whole-turn level.) |
| 40 | assert!( |
| 41 | super::should_transparently_retry_stream( |
| 42 | false, |
| 43 | super::MAX_TRANSPARENT_STREAM_RETRIES - 1, |
| 44 | super::MAX_TRANSPARENT_STREAM_RETRIES, |
| 45 | false, |
| 46 | ), |
| 47 | "one short of the cap should still retry" |
| 48 | ); |
| 49 | assert!( |
| 50 | !super::should_transparently_retry_stream( |
| 51 | false, |
| 52 | super::MAX_TRANSPARENT_STREAM_RETRIES, |
| 53 | super::MAX_TRANSPARENT_STREAM_RETRIES, |
| 54 | false, |
| 55 | ), |
| 56 | "at the cap, no further transparent retries" |
| 57 | ); |
| 58 | assert!( |
| 59 | !super::should_transparently_retry_stream( |
| 60 | false, |
| 61 | super::MAX_TRANSPARENT_STREAM_RETRIES + 5, |
| 62 | super::MAX_TRANSPARENT_STREAM_RETRIES, |
| 63 | false, |
| 64 | ), |
| 65 | "well past the cap, definitely no transparent retries" |
| 66 | ); |
| 67 | } |
| 68 | |
| 69 | // === #2990 sleep-resume policy ================================================ |
| 70 | |
| 71 | #[test] |
| 72 | fn sleep_gap_requires_wallclock_to_outrun_monotonic_clock() { |
| 73 | use std::time::Duration; |
| 74 | // No divergence: ordinary network failure, clocks agree. |
| 75 | assert!( |
| 76 | !super::sleep_gap_detected(Duration::from_secs(30), Duration::from_secs(30)), |
| 77 | "equal elapsed times must not register as a sleep gap" |
| 78 | ); |
| 79 | // Divergence below the threshold: NTP slew / scheduling jitter. |
| 80 | assert!( |
| 81 | !super::sleep_gap_detected(Duration::from_secs(5), Duration::from_secs(14)), |
| 82 | "9s of divergence is below the 10s threshold" |
| 83 | ); |
| 84 | // Divergence above the threshold: the host was suspended. |
| 85 | assert!( |
| 86 | super::sleep_gap_detected(Duration::from_secs(5), Duration::from_secs(16)), |
| 87 | "11s of divergence must register as a sleep gap" |
| 88 | ); |
| 89 | // Wall clock went backwards (NTP step): saturating_sub → zero gap. |
| 90 | assert!( |
| 91 | !super::sleep_gap_detected(Duration::from_secs(60), Duration::from_secs(5)), |
| 92 | "wall clock behind monotonic must never register as a sleep gap" |
| 93 | ); |
| 94 | } |
| 95 | |
| 96 | #[test] |
| 97 | fn sleep_resume_scenario() { |
| 98 | // Scenario consolidation of: sleep_resume_retries_even_after_content_streamed, sleep_resume_requires_a_detected_gap, sleep_resume_respects_budget_and_cancellation |
| 99 | // from sleep_resume_retries_even_after_content_streamed |
| 100 | { |
| 101 | // The whole point of #2990: unlike the #103 transparent retry, a |
| 102 | // detected sleep gap retries regardless of streamed content — the |
| 103 | // partial output predates the sleep and the user was not watching. |
| 104 | assert!( |
| 105 | super::should_resume_after_sleep(true, 0, super::MAX_STREAM_RETRIES, false), |
| 106 | "detected sleep with full budget must resume" |
| 107 | ); |
| 108 | assert!( |
| 109 | super::should_resume_after_sleep( |
| 110 | true, |
| 111 | super::MAX_STREAM_RETRIES - 1, |
| 112 | super::MAX_STREAM_RETRIES, |
| 113 | false |
| 114 | ), |
| 115 | "detected sleep one short of the budget must still resume" |
| 116 | ); |
| 117 | } |
| 118 | // from sleep_resume_requires_a_detected_gap |
| 119 | { |
| 120 | // Without a sleep gap this layer stays out of the way entirely, so the |
| 121 | // deliberate no-retry-after-content policy for ordinary flakes (#103) |
| 122 | // is preserved. |
| 123 | assert!( |
| 124 | !super::should_resume_after_sleep(false, 0, super::MAX_STREAM_RETRIES, false), |
| 125 | "no sleep gap → never resume via this layer" |
| 126 | ); |
| 127 | } |
| 128 | // from sleep_resume_respects_budget_and_cancellation |
| 129 | { |
| 130 | assert!( |
| 131 | !super::should_resume_after_sleep( |
| 132 | true, |
| 133 | super::MAX_STREAM_RETRIES, |
| 134 | super::MAX_STREAM_RETRIES, |
| 135 | false |
| 136 | ), |
| 137 | "budget exhausted → surface the failure instead of looping" |
| 138 | ); |
| 139 | assert!( |
| 140 | !super::should_resume_after_sleep(true, 0, super::MAX_STREAM_RETRIES, true), |
| 141 | "cancelled turn must not be resumed behind the user's back" |
| 142 | ); |
| 143 | } |
| 144 | } |
| 145 | |
| 146 | // === headless mid-stream network-drop resume (v0.9.4 Terminal-Bench P0) ====== |
| 147 | // |
| 148 | // Terminal-Bench 2.1 on the 0.9.4 bundle forfeited tasks when the DeepSeek |
| 149 | // stream dropped mid-response ("error decoding response body" after partial |
| 150 | // content): the #103 policy surfaced the warning and failed the turn, and |
| 151 | // `codewhale exec` exited 1. In a headless host no operator watches the |
| 152 | // partial deltas and the fragment is never committed, so the turn loop now |
| 153 | // re-issues the request instead (bounded by MAX_STREAM_RETRIES), exactly |
| 154 | // like the #2990 sleep-resume. |
| 155 | |
| 156 | #[test] |
| 157 | fn network_drop_scenario() { |
| 158 | // Scenario consolidation of: network_drop_resume_only_fires_for_headless_hosts, network_drop_resume_requires_network_class_error, network_drop_resume_respects_budget_and_cancellation |
| 159 | // from network_drop_resume_only_fires_for_headless_hosts |
| 160 | { |
| 161 | assert!( |
| 162 | super::should_resume_after_network_drop( |
| 163 | true, |
| 164 | true, |
| 165 | 0, |
| 166 | super::MAX_STREAM_RETRIES, |
| 167 | false |
| 168 | ), |
| 169 | "headless host + network-class drop with budget must resume" |
| 170 | ); |
| 171 | assert!( |
| 172 | !super::should_resume_after_network_drop( |
| 173 | false, |
| 174 | true, |
| 175 | 0, |
| 176 | super::MAX_STREAM_RETRIES, |
| 177 | false |
| 178 | ), |
| 179 | "interactive sessions keep the #103 surface-the-warning policy: \ |
| 180 | the user saw the partial deltas and replay would render them twice" |
| 181 | ); |
| 182 | } |
| 183 | // from network_drop_resume_requires_network_class_error |
| 184 | { |
| 185 | assert!( |
| 186 | !super::should_resume_after_network_drop( |
| 187 | true, |
| 188 | false, |
| 189 | 0, |
| 190 | super::MAX_STREAM_RETRIES, |
| 191 | false |
| 192 | ), |
| 193 | "non-network failures (model/parse/auth) must never be replayed" |
| 194 | ); |
| 195 | } |
| 196 | // from network_drop_resume_respects_budget_and_cancellation |
| 197 | { |
| 198 | assert!( |
| 199 | super::should_resume_after_network_drop( |
| 200 | true, |
| 201 | true, |
| 202 | super::MAX_STREAM_RETRIES - 1, |
| 203 | super::MAX_STREAM_RETRIES, |
| 204 | false, |
| 205 | ), |
| 206 | "one short of the budget should still resume" |
| 207 | ); |
| 208 | assert!( |
| 209 | !super::should_resume_after_network_drop( |
| 210 | true, |
| 211 | true, |
| 212 | super::MAX_STREAM_RETRIES, |
| 213 | super::MAX_STREAM_RETRIES, |
| 214 | false |
| 215 | ), |
| 216 | "budget exhausted → surface the failure instead of looping" |
| 217 | ); |
| 218 | assert!( |
| 219 | !super::should_resume_after_network_drop( |
| 220 | true, |
| 221 | true, |
| 222 | 0, |
| 223 | super::MAX_STREAM_RETRIES, |
| 224 | true |
| 225 | ), |
| 226 | "cancelled turn must not be resumed behind the operator's back" |
| 227 | ); |
| 228 | } |
| 229 | } |
| 230 | |
| 231 | // === interactive mid-stream network-drop resume (0.9.4; reworked 0.9.10) ========= |
| 232 | // |
| 233 | // The interactive TUI used to fail the turn when a provider stream dropped |
| 234 | // after partial output because the #103 policy treated any post-content error |
| 235 | // as terminal. The model now preserves a visible partial reply as a committed |
| 236 | // assistant message and re-issues the request. Since 0.9.10 the recovery is |
| 237 | // typed engine-internal state (`StreamResume`): no synthetic `[runtime]` user |
| 238 | // continuation message is appended, and a thinking-only drop preserves |
| 239 | // nothing and never claims it did. |
| 240 | |
| 241 | #[test] |
| 242 | fn interactive_network_scenario() { |
| 243 | // Scenario consolidation of: interactive_network_drop_resume_only_fires_for_interactive_hosts, interactive_network_drop_resume_requires_partial_content_and_no_tools, interactive_network_drop_resume_requires_network_class_error |
| 244 | // from interactive_network_drop_resume_only_fires_for_interactive_hosts |
| 245 | { |
| 246 | assert!( |
| 247 | super::should_resume_interactive_after_network_drop( |
| 248 | true, |
| 249 | true, |
| 250 | true, |
| 251 | true, |
| 252 | 0, |
| 253 | super::MAX_STREAM_RETRIES, |
| 254 | false |
| 255 | ), |
| 256 | "interactive TUI + partial text + no tools + budget must resume" |
| 257 | ); |
| 258 | assert!( |
| 259 | !super::should_resume_interactive_after_network_drop( |
| 260 | false, |
| 261 | true, |
| 262 | true, |
| 263 | true, |
| 264 | 0, |
| 265 | super::MAX_STREAM_RETRIES, |
| 266 | false |
| 267 | ), |
| 268 | "headless hosts must use the headless resume path, not this one" |
| 269 | ); |
| 270 | } |
| 271 | // from interactive_network_drop_resume_requires_partial_content_and_no_tools |
| 272 | { |
| 273 | assert!( |
| 274 | !super::should_resume_interactive_after_network_drop( |
| 275 | true, |
| 276 | true, |
| 277 | false, |
| 278 | true, |
| 279 | 0, |
| 280 | super::MAX_STREAM_RETRIES, |
| 281 | false |
| 282 | ), |
| 283 | "no streamed content → transparent retry or nothing-streamed path" |
| 284 | ); |
| 285 | assert!( |
| 286 | !super::should_resume_interactive_after_network_drop( |
| 287 | true, |
| 288 | true, |
| 289 | true, |
| 290 | false, |
| 291 | 0, |
| 292 | super::MAX_STREAM_RETRIES, |
| 293 | false |
| 294 | ), |
| 295 | "in-flight tool calls must never be resumed (side-effect duplication)" |
| 296 | ); |
| 297 | } |
| 298 | // from interactive_network_drop_resume_requires_network_class_error |
| 299 | { |
| 300 | assert!( |
| 301 | !super::should_resume_interactive_after_network_drop( |
| 302 | true, |
| 303 | false, |
| 304 | true, |
| 305 | true, |
| 306 | 0, |
| 307 | super::MAX_STREAM_RETRIES, |
| 308 | false |
| 309 | ), |
| 310 | "non-network failures must surface normally" |
| 311 | ); |
| 312 | } |
| 313 | } |
| 314 | |
| 315 | #[test] |
| 316 | fn interactive_network_drop_resume_respects_budget_and_cancellation() { |
| 317 | assert!( |
| 318 | super::should_resume_interactive_after_network_drop( |
| 319 | true, |
| 320 | true, |
| 321 | true, |
| 322 | true, |
| 323 | super::MAX_STREAM_RETRIES - 1, |
| 324 | super::MAX_STREAM_RETRIES, |
| 325 | false, |
| 326 | ), |
| 327 | "one short of the budget should still resume" |
| 328 | ); |
| 329 | assert!( |
| 330 | !super::should_resume_interactive_after_network_drop( |
| 331 | true, |
| 332 | true, |
| 333 | true, |
| 334 | true, |
| 335 | super::MAX_STREAM_RETRIES, |
| 336 | super::MAX_STREAM_RETRIES, |
| 337 | false, |
| 338 | ), |
| 339 | "budget exhausted → surface the failure" |
| 340 | ); |
| 341 | assert!( |
| 342 | !super::should_resume_interactive_after_network_drop( |
| 343 | true, |
| 344 | true, |
| 345 | true, |
| 346 | true, |
| 347 | 0, |
| 348 | super::MAX_STREAM_RETRIES, |
| 349 | true |
| 350 | ), |
| 351 | "cancelled turn must not resume" |
| 352 | ); |
| 353 | } |
| 354 | |
| 355 | /// Model client whose first `failures` streams emit partial content and then |
| 356 | /// die with the network-class read error reqwest reports for a dropped |
| 357 | /// chunked-transfer body; later streams complete a normal text turn. |
| 358 | /// Exercises the headless mid-stream network-drop resume through the real |
| 359 | /// turn loop. |
| 360 | struct FlakyNetworkDropModelClient { |
| 361 | calls: std::sync::atomic::AtomicUsize, |
| 362 | failures: usize, |
| 363 | terminal_before_drop: bool, |
| 364 | content_before_drop: bool, |
| 365 | /// #6699: fail the request itself, before any stream exists, in the |
| 366 | /// given shape. |
| 367 | open_failure: Option<StreamOpenFailure>, |
| 368 | } |
| 369 | |
| 370 | /// #6699/#6711: how a request fails before any stream exists. |
| 371 | #[derive(Clone, Copy, Debug)] |
| 372 | enum StreamOpenFailure { |
| 373 | /// The typed transport error `open_sse_response` returns on a header |
| 374 | /// stall. |
| 375 | HeaderStall, |
| 376 | /// A reqwest connect error under the Anthropic adapter's outer context, |
| 377 | /// as an HTTP/1.1-pinned open returns it: the outer message names no |
| 378 | /// transport cause. |
| 379 | WrappedConnectError, |
| 380 | /// The Responses adapter shape: the typed retry-layer network error |
| 381 | /// under the adapter's outer context. |
| 382 | WrappedTypedNetworkError, |
| 383 | /// A provider answered with an HTTP rejection whose body mentions a |
| 384 | /// connection reset, as the Anthropic adapter reports it (untyped). |
| 385 | HttpRejectionMentioningConnection, |
| 386 | } |
| 387 | |
| 388 | impl StreamOpenFailure { |
| 389 | async fn error(self) -> anyhow::Error { |
| 390 | use anyhow::Context as _; |
| 391 | match self { |
| 392 | Self::HeaderStall => anyhow::Error::new(crate::llm_client::LlmError::NetworkError( |
| 393 | "SSE stream request did not receive response headers after 45s \ |
| 394 | (HTTP/2 and HTTP/1.1)." |
| 395 | .to_string(), |
| 396 | )), |
| 397 | Self::WrappedConnectError => { |
| 398 | // A port that was just released refuses the connection. |
| 399 | let listener = std::net::TcpListener::bind("127.0.0.1:0").expect("bind"); |
| 400 | let addr = listener.local_addr().expect("local addr"); |
| 401 | drop(listener); |
| 402 | let client = crate::tls::reqwest_client_builder() |
| 403 | .no_proxy() |
| 404 | .build() |
| 405 | .expect("client"); |
| 406 | let send = client |
| 407 | .post(format!("http://{addr}/v1/messages")) |
| 408 | .send() |
| 409 | .await; |
| 410 | let err = send |
| 411 | .context("Anthropic Messages API request failed") |
| 412 | .expect_err("closed port must refuse the connection"); |
| 413 | assert_eq!(err.to_string(), "Anthropic Messages API request failed"); |
| 414 | err |
| 415 | } |
| 416 | Self::WrappedTypedNetworkError => Err::<(), _>(anyhow::Error::new( |
| 417 | crate::llm_client::LlmError::NetworkError( |
| 418 | "Connection failed: error sending request".to_string(), |
| 419 | ), |
| 420 | )) |
| 421 | .context("Responses API request failed") |
| 422 | .expect_err("wrapped network error"), |
| 423 | Self::HttpRejectionMentioningConnection => anyhow::anyhow!( |
| 424 | "Anthropic API error (HTTP 500 Internal Server Error api_error): \ |
| 425 | upstream connection reset" |
| 426 | ), |
| 427 | } |
| 428 | } |
| 429 | } |
| 430 | |
| 431 | #[async_trait::async_trait] |
| 432 | impl crate::core::model_client::ModelClient for FlakyNetworkDropModelClient { |
| 433 | fn provider_name(&self) -> &str { |
| 434 | "flaky-network" |
| 435 | } |
| 436 | |
| 437 | fn model(&self) -> &str { |
| 438 | "local-model" |
| 439 | } |
| 440 | |
| 441 | async fn create_message( |
| 442 | &self, |
| 443 | _request: codewhale_models::MessageRequest, |
| 444 | ) -> anyhow::Result<codewhale_models::MessageResponse> { |
| 445 | anyhow::bail!("flaky-network regression uses the streaming model boundary") |
| 446 | } |
| 447 | |
| 448 | async fn create_message_stream( |
| 449 | &self, |
| 450 | _request: codewhale_models::MessageRequest, |
| 451 | ) -> anyhow::Result<crate::llm_client::StreamEventBox> { |
| 452 | use crate::llm_client::mock::canned; |
| 453 | let call = self |
| 454 | .calls |
| 455 | .fetch_add(1, std::sync::atomic::Ordering::SeqCst) |
| 456 | .saturating_add(1); |
| 457 | if call <= self.failures { |
| 458 | if let Some(open_failure) = self.open_failure { |
| 459 | return Err(open_failure.error().await); |
| 460 | } |
| 461 | if self.terminal_before_drop { |
| 462 | let start_usage = Usage { |
| 463 | input_tokens: 31, |
| 464 | ..Default::default() |
| 465 | }; |
| 466 | let delta_usage = Usage { |
| 467 | output_tokens: 8, |
| 468 | ..Default::default() |
| 469 | }; |
| 470 | let mut message_start = canned::message_start("terminal_then_drop"); |
| 471 | if let StreamEvent::MessageStart { message } = &mut message_start { |
| 472 | message.usage = start_usage; |
| 473 | } |
| 474 | let events: Vec<anyhow::Result<codewhale_models::StreamEvent>> = vec![ |
| 475 | Ok(message_start), |
| 476 | Ok(canned::text_block_start(0)), |
| 477 | Ok(canned::text_delta(0, "billed truncated fragment")), |
| 478 | Ok(canned::block_stop(0)), |
| 479 | Ok(canned::message_delta("max_tokens", Some(delta_usage))), |
| 480 | Err(anyhow::anyhow!( |
| 481 | "Stream read error: error decoding response body" |
| 482 | )), |
| 483 | ]; |
| 484 | return Ok(Box::pin(futures_util::stream::iter(events))); |
| 485 | } |
| 486 | if !self.content_before_drop { |
| 487 | return Ok(Box::pin(futures_util::stream::iter(vec![ |
| 488 | Ok(canned::message_start("empty_then_drop")), |
| 489 | Err(anyhow::anyhow!( |
| 490 | "Stream read error: error decoding response body" |
| 491 | )), |
| 492 | ]))); |
| 493 | } |
| 494 | // Partial content first — this flips `any_content_received` so |
| 495 | // the #103 transparent retry cannot fire — then the transport |
| 496 | // dies the way the 0.9.4 Terminal-Bench crashes did. |
| 497 | let events: Vec<anyhow::Result<codewhale_models::StreamEvent>> = vec![ |
| 498 | Ok(canned::message_start("flaky_msg")), |
| 499 | Ok(canned::text_block_start(0)), |
| 500 | Ok(canned::text_delta( |
| 501 | 0, |
| 502 | "partial answer that must be discarded", |
| 503 | )), |
| 504 | Err(anyhow::anyhow!( |
| 505 | "Stream read error: error decoding response body" |
| 506 | )), |
| 507 | ]; |
| 508 | return Ok(Box::pin(futures_util::stream::iter(events))); |
| 509 | } |
| 510 | let events = canned::simple_text_turn("recovered after retry") |
| 511 | .into_iter() |
| 512 | .map(Ok); |
| 513 | Ok(Box::pin(futures_util::stream::iter(events))) |
| 514 | } |
| 515 | |
| 516 | async fn health_check(&self) -> anyhow::Result<bool> { |
| 517 | Ok(true) |
| 518 | } |
| 519 | } |
| 520 | |
| 521 | /// Drive one headless (`terminal_chrome_enabled = false`, the exec / |
| 522 | /// stream-json posture) turn against the flaky client and collect every |
| 523 | /// event through the terminal TurnComplete. |
| 524 | async fn run_headless_turn_with_flaky_network( |
| 525 | failures: usize, |
| 526 | ) -> (std::sync::Arc<FlakyNetworkDropModelClient>, Vec<Event>) { |
| 527 | let model = std::sync::Arc::new(FlakyNetworkDropModelClient { |
| 528 | calls: std::sync::atomic::AtomicUsize::new(0), |
| 529 | failures, |
| 530 | terminal_before_drop: false, |
| 531 | content_before_drop: true, |
| 532 | open_failure: None, |
| 533 | }); |
| 534 | let client: crate::core::model_client::SharedModelClient = model.clone(); |
| 535 | let config = Config::default(); |
| 536 | let engine_config = EngineConfig { |
| 537 | max_steps: 1, |
| 538 | snapshots_enabled: false, |
| 539 | subagents_enabled: false, |
| 540 | terminal_chrome_enabled: false, |
| 541 | ..EngineConfig::default() |
| 542 | }; |
| 543 | let (engine, handle) = Engine::new_with_model_client(engine_config, &config, client); |
| 544 | let run_task = tokio::spawn(engine.run()); |
| 545 | |
| 546 | handle |
| 547 | .send(Op::SendMessage(TurnSpec { |
| 548 | max_output_tokens: None, |
| 549 | content: "solve the task".to_string(), |
| 550 | images: Vec::new(), |
| 551 | mode: AppMode::Agent, |
| 552 | route: resolved_route_for_test(&config, crate::config::DEFAULT_TEXT_MODEL), |
| 553 | compaction: Box::new(CompactionConfig::default()), |
| 554 | initial_routed_usage: Box::default(), |
| 555 | goal_objective: None, |
| 556 | goal_token_budget: None, |
| 557 | goal_status: crate::tools::goal::GoalStatus::Active, |
| 558 | reasoning_effort: None, |
| 559 | reasoning_effort_auto: false, |
| 560 | auto_model: false, |
| 561 | allow_shell: false, |
| 562 | trust_mode: false, |
| 563 | auto_approve: false, |
| 564 | approval_mode: ApprovalMode::Suggest, |
| 565 | translation_enabled: false, |
| 566 | allowed_tools: None, |
| 567 | dynamic_tools: Vec::new(), |
| 568 | hook_executor: None, |
| 569 | verbosity: None, |
| 570 | provenance: UserInputProvenance::ExternalUser, |
| 571 | submission_id: None, |
| 572 | })) |
| 573 | .await |
| 574 | .expect("send flaky-network turn"); |
| 575 | |
| 576 | let mut events = Vec::new(); |
| 577 | loop { |
| 578 | let event = tokio::time::timeout(model_turn_event_timeout(), async { |
| 579 | handle.rx_event.write().await.recv().await |
| 580 | }) |
| 581 | .await |
| 582 | .expect("flaky-network event timeout") |
| 583 | .expect("flaky-network event"); |
| 584 | let terminal = matches!(event, Event::TurnComplete { .. }); |
| 585 | events.push(event); |
| 586 | if terminal { |
| 587 | break; |
| 588 | } |
| 589 | } |
| 590 | handle.send(Op::Shutdown).await.expect("shutdown engine"); |
| 591 | run_task.await.expect("engine task"); |
| 592 | (model, events) |
| 593 | } |
| 594 | |
| 595 | #[tokio::test] |
| 596 | async fn headless_turn_retries_mid_stream_network_drop_and_recovers() { |
| 597 | let (model, events) = run_headless_turn_with_flaky_network(1).await; |
| 598 | |
| 599 | let terminal = events |
| 600 | .iter() |
| 601 | .find_map(|event| match event { |
| 602 | Event::ToolRequestSnapshot { snapshot } => snapshot.terminal.as_ref(), |
| 603 | _ => None, |
| 604 | }) |
| 605 | .expect("terminal diagnostics through existing event authority"); |
| 606 | assert_eq!(terminal.model_requests_started, 2); |
| 607 | assert_eq!(terminal.stream_resumes, 1); |
| 608 | assert_eq!(terminal.transparent_stream_retries, 0); |
| 609 | |
| 610 | assert_eq!( |
| 611 | model.calls.load(std::sync::atomic::Ordering::SeqCst), |
| 612 | 2, |
| 613 | "the dropped stream must be re-issued exactly once" |
| 614 | ); |
| 615 | let (status, error) = events |
| 616 | .iter() |
| 617 | .find_map(|event| match event { |
| 618 | Event::TurnComplete { status, error, .. } => Some((status, error)), |
| 619 | _ => None, |
| 620 | }) |
| 621 | .expect("terminal TurnComplete"); |
| 622 | assert_eq!( |
| 623 | *status, |
| 624 | TurnOutcomeStatus::Completed, |
| 625 | "a recovered retry must complete the turn: {error:?}" |
| 626 | ); |
| 627 | assert!(error.is_none(), "recovered turn must not report an error"); |
| 628 | assert!( |
| 629 | events.iter().any(|event| matches!( |
| 630 | event, |
| 631 | Event::Status { message } if message.starts_with("Retry attempt: stream-resume 1/") |
| 632 | )), |
| 633 | "the first actual resume must remain visible: {events:?}" |
| 634 | ); |
| 635 | assert!( |
| 636 | !events |
| 637 | .iter() |
| 638 | .any(|event| matches!(event, Event::Error { .. })), |
| 639 | "a transient drop that the retry recovers must not surface an error event: {events:?}" |
| 640 | ); |
| 641 | // The discarded fragment from the dropped attempt must never reach the |
| 642 | // transcript — only the retried turn's content is committed. |
| 643 | let transcript_text = events |
| 644 | .iter() |
| 645 | .filter_map(|event| match event { |
| 646 | Event::SessionUpdated { messages, .. } => Some(messages), |
| 647 | _ => None, |
| 648 | }) |
| 649 | .flat_map(|messages| messages.iter()) |
| 650 | .flat_map(|message| message.content.iter()) |
| 651 | .filter_map(|block| match block { |
| 652 | ContentBlock::Text { text, .. } => Some(text.as_str()), |
| 653 | _ => None, |
| 654 | }) |
| 655 | .collect::<Vec<_>>() |
| 656 | .join("\n"); |
| 657 | assert!( |
| 658 | transcript_text.contains("recovered after retry"), |
| 659 | "retried content must be committed: {transcript_text}" |
| 660 | ); |
| 661 | assert!( |
| 662 | !transcript_text.contains("partial answer that must be discarded"), |
| 663 | "the dropped attempt's fragment must be discarded, not committed: {transcript_text}" |
| 664 | ); |
| 665 | } |
| 666 | |
| 667 | #[tokio::test] |
| 668 | async fn terminal_diagnostics_count_transparent_stream_requests_without_extra_snapshots() { |
| 669 | let workspace = tempdir().expect("tempdir"); |
| 670 | let model = std::sync::Arc::new(FlakyNetworkDropModelClient { |
| 671 | calls: std::sync::atomic::AtomicUsize::new(0), |
| 672 | failures: 1, |
| 673 | terminal_before_drop: false, |
| 674 | content_before_drop: false, |
| 675 | open_failure: None, |
| 676 | }); |
| 677 | let client: crate::core::model_client::SharedModelClient = model.clone(); |
| 678 | let (mut engine, handle) = Engine::new_with_model_client( |
| 679 | deterministic_engine_config(workspace.path()), |
| 680 | &Config::default(), |
| 681 | client, |
| 682 | ); |
| 683 | let registry = crate::tools::ToolRegistry::new(crate::tools::ToolContext::new( |
| 684 | workspace.path().to_path_buf(), |
| 685 | )); |
| 686 | let surface = test_tool_surface(&engine, registry, None, AppMode::Agent); |
| 687 | let mut turn = crate::core::turn::TurnContext::new(4); |
| 688 | let (status, error) = engine.run_turn(&mut turn, surface, None, None).await; |
| 689 | assert_eq!(status, TurnOutcomeStatus::Completed, "{error:?}"); |
| 690 | assert_eq!(model.calls.load(std::sync::atomic::Ordering::SeqCst), 2); |
| 691 | let snapshot = turn |
| 692 | .terminal_request_snapshot(status) |
| 693 | .expect("terminal snapshot"); |
| 694 | let terminal = snapshot.terminal.expect("terminal facts"); |
| 695 | assert_eq!(terminal.model_requests_started, 2); |
| 696 | assert_eq!(terminal.transparent_stream_retries, 1); |
| 697 | assert_eq!(terminal.stream_resumes, 0); |
| 698 | assert_eq!(terminal.model_step_index, 0); |
| 699 | let mut receiver = handle.rx_event.write().await; |
| 700 | let events = std::iter::from_fn(|| receiver.try_recv().ok()).collect::<Vec<_>>(); |
| 701 | assert!(events.iter().any(|event| matches!(event, |
| 702 | Event::Status { message } if message.starts_with("Retry attempt: transparent-stream 1/") |
| 703 | ))); |
| 704 | assert!(events.iter().any(|event| matches!(event, |
| 705 | Event::Status { message } if message == "Retry recovery: transparent stream recovered after 1 retries" |
| 706 | ))); |
| 707 | let snapshots = events |
| 708 | .iter() |
| 709 | .filter(|event| matches!(event, Event::ToolRequestSnapshot { .. })) |
| 710 | .count(); |
| 711 | assert_eq!( |
| 712 | snapshots, 1, |
| 713 | "request construction is distinct from stream retries" |
| 714 | ); |
| 715 | } |
| 716 | |
| 717 | /// #6699: drive one interactive turn whose first `failures` requests fail |
| 718 | /// before a stream opens, with `max_resumes` as the configured budget. |
| 719 | async fn run_turn_with_stream_open_failures( |
| 720 | failures: usize, |
| 721 | max_resumes: Option<u32>, |
| 722 | open_failure: StreamOpenFailure, |
| 723 | ) -> ( |
| 724 | std::sync::Arc<FlakyNetworkDropModelClient>, |
| 725 | TurnOutcomeStatus, |
| 726 | Option<String>, |
| 727 | crate::tool_inspection::TurnStopDiagnostics, |
| 728 | Vec<Event>, |
| 729 | ) { |
| 730 | let workspace = tempdir().expect("tempdir"); |
| 731 | let model = std::sync::Arc::new(FlakyNetworkDropModelClient { |
| 732 | calls: std::sync::atomic::AtomicUsize::new(0), |
| 733 | failures, |
| 734 | terminal_before_drop: false, |
| 735 | content_before_drop: false, |
| 736 | open_failure: Some(open_failure), |
| 737 | }); |
| 738 | let client: crate::core::model_client::SharedModelClient = model.clone(); |
| 739 | let engine_config = EngineConfig { |
| 740 | terminal_chrome_enabled: true, |
| 741 | stream_retry_limits: turn_budget::resolve_stream_retry_limits(max_resumes, None, None), |
| 742 | ..deterministic_engine_config(workspace.path()) |
| 743 | }; |
| 744 | let (mut engine, handle) = |
| 745 | Engine::new_with_model_client(engine_config, &Config::default(), client); |
| 746 | let registry = crate::tools::ToolRegistry::new(crate::tools::ToolContext::new( |
| 747 | workspace.path().to_path_buf(), |
| 748 | )); |
| 749 | let surface = test_tool_surface(&engine, registry, None, AppMode::Agent); |
| 750 | let mut turn = crate::core::turn::TurnContext::new(4); |
| 751 | let (status, error) = engine.run_turn(&mut turn, surface, None, None).await; |
| 752 | let terminal = turn |
| 753 | .terminal_request_snapshot(status) |
| 754 | .expect("terminal snapshot") |
| 755 | .terminal |
| 756 | .expect("terminal facts"); |
| 757 | let mut rx = handle.rx_event.write().await; |
| 758 | let events = std::iter::from_fn(|| rx.try_recv().ok()).collect(); |
| 759 | (model, status, error, terminal, events) |
| 760 | } |
| 761 | |
| 762 | #[tokio::test] |
| 763 | async fn stream_open_failure_is_retried_through_the_resume_budget() { |
| 764 | let (model, status, error, terminal, events) = |
| 765 | run_turn_with_stream_open_failures(1, None, StreamOpenFailure::HeaderStall).await; |
| 766 | assert_eq!(status, TurnOutcomeStatus::Completed, "{error:?}"); |
| 767 | assert_eq!( |
| 768 | model.calls.load(std::sync::atomic::Ordering::SeqCst), |
| 769 | 2, |
| 770 | "a request that never opened must be re-issued once" |
| 771 | ); |
| 772 | assert_eq!(terminal.model_requests_started, 2); |
| 773 | assert_eq!(terminal.stream_resumes, 1); |
| 774 | assert!( |
| 775 | !events |
| 776 | .iter() |
| 777 | .any(|event| matches!(event, Event::Error { .. })), |
| 778 | "a recovered open failure must not surface an error event: {events:?}" |
| 779 | ); |
| 780 | } |
| 781 | |
| 782 | #[tokio::test] |
| 783 | async fn stream_open_failure_fails_the_turn_once_the_budget_is_spent() { |
| 784 | let (model, status, error, terminal, events) = |
| 785 | run_turn_with_stream_open_failures(usize::MAX, None, StreamOpenFailure::HeaderStall).await; |
| 786 | assert_eq!(status, TurnOutcomeStatus::Failed); |
| 787 | assert_eq!( |
| 788 | model.calls.load(std::sync::atomic::Ordering::SeqCst), |
| 789 | 1 + super::MAX_STREAM_RETRIES as usize, |
| 790 | "initial attempt plus the default resume budget, then the turn fails" |
| 791 | ); |
| 792 | assert_eq!(terminal.stream_resumes, super::MAX_STREAM_RETRIES); |
| 793 | assert!( |
| 794 | error |
| 795 | .as_deref() |
| 796 | .is_some_and(|error| error.contains("did not receive response headers")), |
| 797 | "the real transport error must surface: {error:?}" |
| 798 | ); |
| 799 | assert_eq!( |
| 800 | events |
| 801 | .iter() |
| 802 | .filter(|event| matches!(event, Event::Error { .. })) |
| 803 | .count(), |
| 804 | 1, |
| 805 | "only the exhausted attempt emits an error event: {events:?}" |
| 806 | ); |
| 807 | } |
| 808 | |
| 809 | #[tokio::test] |
| 810 | async fn stream_open_failure_honors_a_configured_resume_budget() { |
| 811 | let (model, status, _error, terminal, _events) = |
| 812 | run_turn_with_stream_open_failures(usize::MAX, Some(0), StreamOpenFailure::HeaderStall) |
| 813 | .await; |
| 814 | assert_eq!(status, TurnOutcomeStatus::Failed); |
| 815 | assert_eq!( |
| 816 | model.calls.load(std::sync::atomic::Ordering::SeqCst), |
| 817 | 1, |
| 818 | "`stream_max_resumes = 0` disables the turn-level retry" |
| 819 | ); |
| 820 | assert_eq!(terminal.stream_resumes, 0); |
| 821 | } |
| 822 | |
| 823 | /// #6711: an HTTP/1.1-pinned Anthropic open returns the reqwest connect |
| 824 | /// error under an outer context that names no transport cause. The retry gate |
| 825 | /// must read the error chain, not the outer message. |
| 826 | #[tokio::test] |
| 827 | async fn stream_open_failure_behind_adapter_context_is_retried() { |
| 828 | for open_failure in [ |
| 829 | StreamOpenFailure::WrappedConnectError, |
| 830 | StreamOpenFailure::WrappedTypedNetworkError, |
| 831 | ] { |
| 832 | let (model, status, error, terminal, _events) = |
| 833 | run_turn_with_stream_open_failures(1, None, open_failure).await; |
| 834 | assert_eq!( |
| 835 | status, |
| 836 | TurnOutcomeStatus::Completed, |
| 837 | "{open_failure:?}: {error:?}" |
| 838 | ); |
| 839 | assert_eq!( |
| 840 | model.calls.load(std::sync::atomic::Ordering::SeqCst), |
| 841 | 2, |
| 842 | "{open_failure:?}: a wrapped transport failure must be re-issued" |
| 843 | ); |
| 844 | assert_eq!(terminal.stream_resumes, 1, "{open_failure:?}"); |
| 845 | } |
| 846 | } |
| 847 | |
| 848 | /// #6711: a provider that answered with an HTTP rejection is not an open |
| 849 | /// failure, even when its body mentions a connection or a timeout. The request |
| 850 | /// must not be re-issued through the resume budget. |
| 851 | #[tokio::test] |
| 852 | async fn stream_open_http_rejection_mentioning_connection_is_not_retried() { |
| 853 | let (model, status, _error, terminal, events) = run_turn_with_stream_open_failures( |
| 854 | usize::MAX, |
| 855 | None, |
| 856 | StreamOpenFailure::HttpRejectionMentioningConnection, |
| 857 | ) |
| 858 | .await; |
| 859 | assert_eq!(status, TurnOutcomeStatus::Failed); |
| 860 | assert_eq!( |
| 861 | model.calls.load(std::sync::atomic::Ordering::SeqCst), |
| 862 | 1, |
| 863 | "a provider rejection must fail the turn on the first answer" |
| 864 | ); |
| 865 | assert_eq!(terminal.stream_resumes, 0); |
| 866 | assert_eq!( |
| 867 | events |
| 868 | .iter() |
| 869 | .filter(|event| matches!(event, Event::Error { .. })) |
| 870 | .count(), |
| 871 | 1, |
| 872 | "{events:?}" |
| 873 | ); |
| 874 | } |
| 875 | |
| 876 | #[tokio::test] |
| 877 | async fn terminal_output_limit_followed_by_stream_error_is_charged_and_not_retried() { |
| 878 | let model = std::sync::Arc::new(FlakyNetworkDropModelClient { |
| 879 | calls: std::sync::atomic::AtomicUsize::new(0), |
| 880 | failures: 1, |
| 881 | terminal_before_drop: true, |
| 882 | content_before_drop: true, |
| 883 | open_failure: None, |
| 884 | }); |
| 885 | let client: crate::core::model_client::SharedModelClient = model.clone(); |
| 886 | let config = Config::default(); |
| 887 | let engine_config = EngineConfig { |
| 888 | max_steps: 1, |
| 889 | snapshots_enabled: false, |
| 890 | subagents_enabled: false, |
| 891 | terminal_chrome_enabled: false, |
| 892 | ..EngineConfig::default() |
| 893 | }; |
| 894 | let (engine, handle) = Engine::new_with_model_client(engine_config, &config, client); |
| 895 | let run_task = tokio::spawn(engine.run()); |
| 896 | |
| 897 | handle |
| 898 | .send(Op::SendMessage(TurnSpec { |
| 899 | max_output_tokens: None, |
| 900 | content: "solve the task".to_string(), |
| 901 | images: Vec::new(), |
| 902 | mode: AppMode::Agent, |
| 903 | route: resolved_route_for_test(&config, crate::config::DEFAULT_TEXT_MODEL), |
| 904 | compaction: Box::new(CompactionConfig::default()), |
| 905 | initial_routed_usage: Box::default(), |
| 906 | goal_objective: None, |
| 907 | goal_token_budget: None, |
| 908 | goal_status: crate::tools::goal::GoalStatus::Active, |
| 909 | reasoning_effort: None, |
| 910 | reasoning_effort_auto: false, |
| 911 | auto_model: false, |
| 912 | allow_shell: false, |
| 913 | trust_mode: false, |
| 914 | auto_approve: false, |
| 915 | approval_mode: ApprovalMode::Suggest, |
| 916 | translation_enabled: false, |
| 917 | allowed_tools: None, |
| 918 | dynamic_tools: Vec::new(), |
| 919 | hook_executor: None, |
| 920 | verbosity: None, |
| 921 | provenance: UserInputProvenance::ExternalUser, |
| 922 | submission_id: None, |
| 923 | })) |
| 924 | .await |
| 925 | .expect("send terminal-then-drop turn"); |
| 926 | |
| 927 | let mut events = Vec::new(); |
| 928 | loop { |
| 929 | let event = tokio::time::timeout(model_turn_event_timeout(), async { |
| 930 | handle.rx_event.write().await.recv().await |
| 931 | }) |
| 932 | .await |
| 933 | .expect("terminal-then-drop event timeout") |
| 934 | .expect("terminal-then-drop event"); |
| 935 | let terminal = matches!(event, Event::TurnComplete { .. }); |
| 936 | events.push(event); |
| 937 | if terminal { |
| 938 | break; |
| 939 | } |
| 940 | } |
| 941 | |
| 942 | assert_eq!( |
| 943 | model.calls.load(std::sync::atomic::Ordering::SeqCst), |
| 944 | 1, |
| 945 | "a provider-declared terminal response must never be re-issued" |
| 946 | ); |
| 947 | assert!(events.iter().any(|event| matches!( |
| 948 | event, |
| 949 | Event::TurnUsage { usage, .. } |
| 950 | if usage.input_tokens == 31 && usage.output_tokens == 8 |
| 951 | ))); |
| 952 | let (status, error) = events |
| 953 | .iter() |
| 954 | .find_map(|event| match event { |
| 955 | Event::TurnComplete { status, error, .. } => Some((status, error)), |
| 956 | _ => None, |
| 957 | }) |
| 958 | .expect("terminal TurnComplete"); |
| 959 | assert_eq!(*status, TurnOutcomeStatus::Failed); |
| 960 | assert!( |
| 961 | error |
| 962 | .as_deref() |
| 963 | .is_some_and(|error| error.contains("max_tokens")), |
| 964 | "{error:?}" |
| 965 | ); |
| 966 | assert!(!events.iter().any(|event| matches!( |
| 967 | event, |
| 968 | Event::Status { message } if crate::core::events::is_retry_status_receipt(message) |
| 969 | ))); |
| 970 | |
| 971 | handle.send(Op::Shutdown).await.expect("shutdown engine"); |
| 972 | run_task.await.expect("engine task"); |
| 973 | } |
| 974 | |
| 975 | /// Run one user turn against scripted streams; returns its events and how |
| 976 | /// many model requests were made. |
| 977 | async fn error_frame_turn_events(turns: Vec<Vec<StreamEvent>>) -> (Vec<Event>, usize) { |
| 978 | let model = std::sync::Arc::new(crate::llm_client::mock::MockLlmClient::new(turns)); |
| 979 | let client: crate::core::model_client::SharedModelClient = model.clone(); |
| 980 | let config = Config::default(); |
| 981 | let engine_config = EngineConfig { |
| 982 | max_steps: 1, |
| 983 | snapshots_enabled: false, |
| 984 | subagents_enabled: false, |
| 985 | terminal_chrome_enabled: false, |
| 986 | ..EngineConfig::default() |
| 987 | }; |
| 988 | let (engine, handle) = Engine::new_with_model_client(engine_config, &config, client); |
| 989 | let run_task = tokio::spawn(engine.run()); |
| 990 | handle |
| 991 | .send(Op::SendMessage(TurnSpec { |
| 992 | max_output_tokens: None, |
| 993 | content: "solve the task".to_string(), |
| 994 | images: Vec::new(), |
| 995 | mode: AppMode::Agent, |
| 996 | route: resolved_route_for_test(&config, crate::config::DEFAULT_TEXT_MODEL), |
| 997 | compaction: Box::new(CompactionConfig::default()), |
| 998 | initial_routed_usage: Box::default(), |
| 999 | goal_objective: None, |
| 1000 | goal_token_budget: None, |
| 1001 | goal_status: crate::tools::goal::GoalStatus::Active, |
| 1002 | reasoning_effort: None, |
| 1003 | reasoning_effort_auto: false, |
| 1004 | auto_model: false, |
| 1005 | allow_shell: false, |
| 1006 | trust_mode: false, |
| 1007 | auto_approve: false, |
| 1008 | approval_mode: ApprovalMode::Suggest, |
| 1009 | translation_enabled: false, |
| 1010 | allowed_tools: None, |
| 1011 | dynamic_tools: Vec::new(), |
| 1012 | hook_executor: None, |
| 1013 | verbosity: None, |
| 1014 | provenance: UserInputProvenance::ExternalUser, |
| 1015 | submission_id: None, |
| 1016 | })) |
| 1017 | .await |
| 1018 | .expect("send turn"); |
| 1019 | let mut events = Vec::new(); |
| 1020 | loop { |
| 1021 | let event = tokio::time::timeout(model_turn_event_timeout(), async { |
| 1022 | handle.rx_event.write().await.recv().await |
| 1023 | }) |
| 1024 | .await |
| 1025 | .expect("turn event timeout") |
| 1026 | .expect("turn event"); |
| 1027 | let terminal = matches!(event, Event::TurnComplete { .. }); |
| 1028 | events.push(event); |
| 1029 | if terminal { |
| 1030 | break; |
| 1031 | } |
| 1032 | } |
| 1033 | handle.send(Op::Shutdown).await.expect("shutdown engine"); |
| 1034 | run_task.await.expect("engine task"); |
| 1035 | (events, model.call_count()) |
| 1036 | } |
| 1037 | |
| 1038 | /// #6795: a transient upstream failure delivered as an error frame inside a |
| 1039 | /// successful response, before anything streamed, is retried like every other |
| 1040 | /// no-content stream death. A terminal-class frame still fails on the first |
| 1041 | /// request. |
| 1042 | #[tokio::test] |
| 1043 | async fn transient_error_frame_with_no_content_is_retried_and_terminal_frame_is_not() { |
| 1044 | use crate::llm_client::mock::canned; |
| 1045 | let frame = |message: &str| { |
| 1046 | vec![StreamEvent::Error { |
| 1047 | error: serde_json::json!({ "message": message }), |
| 1048 | }] |
| 1049 | }; |
| 1050 | |
| 1051 | let (events, requests) = error_frame_turn_events(vec![ |
| 1052 | frame("Provider returned an empty response"), |
| 1053 | canned::simple_text_turn("recovered answer"), |
| 1054 | ]) |
| 1055 | .await; |
| 1056 | assert_eq!(requests, 2, "the empty-upstream frame is re-issued once"); |
| 1057 | assert!(events.iter().any(|event| matches!(event, |
| 1058 | Event::Status { message } if message.starts_with("Retry attempt: stream-resume 1/") |
| 1059 | ))); |
| 1060 | assert!(events.iter().any(|event| matches!(event, |
| 1061 | Event::Status { message } if message == "Retry recovery: stream recovered after 1 retries" |
| 1062 | ))); |
| 1063 | assert!( |
| 1064 | events.iter().any(|event| matches!( |
| 1065 | event, |
| 1066 | Event::TurnComplete { |
| 1067 | status: TurnOutcomeStatus::Completed, |
| 1068 | error: None, |
| 1069 | .. |
| 1070 | } |
| 1071 | )), |
| 1072 | "a successful retry completes the turn" |
| 1073 | ); |
| 1074 | assert!( |
| 1075 | !events |
| 1076 | .iter() |
| 1077 | .any(|event| matches!(event, Event::Error { .. })), |
| 1078 | "a retried frame leaves no error card behind" |
| 1079 | ); |
| 1080 | |
| 1081 | let (events, requests) = error_frame_turn_events(vec![ |
| 1082 | frame("Model not exist."), |
| 1083 | canned::simple_text_turn("must never be requested"), |
| 1084 | ]) |
| 1085 | .await; |
| 1086 | assert_eq!(requests, 1, "a terminal-class frame is never retried"); |
| 1087 | assert!(events.iter().any(|event| matches!( |
| 1088 | event, |
| 1089 | Event::TurnComplete { |
| 1090 | status: TurnOutcomeStatus::Failed, |
| 1091 | .. |
| 1092 | } |
| 1093 | ))); |
| 1094 | assert!(events.iter().any(|event| matches!( |
| 1095 | event, |
| 1096 | Event::Error { envelope, .. } |
| 1097 | if envelope.message.contains("Model not exist.") && !envelope.recoverable |
| 1098 | ))); |
| 1099 | } |
| 1100 | |
| 1101 | #[tokio::test] |
| 1102 | async fn midstream_error_frame_stops_the_stream_and_drops_trailing_deltas() { |
| 1103 | // The reported incident: a provider delivered a chunk-level error |
| 1104 | // object ("Model not exist.") mid-stream, and the stream kept parsing |
| 1105 | // later frames as deltas — so reasoning/text rendered *after* the |
| 1106 | // failure. The `StreamEvent::Error` arm must surface a terminal error |
| 1107 | // event, stop consuming the stream, and never forward deltas that |
| 1108 | // arrive after the failure frame. |
| 1109 | let model = std::sync::Arc::new(crate::llm_client::mock::MockLlmClient::new(vec![vec![ |
| 1110 | crate::llm_client::mock::canned::message_start("midstream-error"), |
| 1111 | crate::llm_client::mock::canned::text_block_start(0), |
| 1112 | StreamEvent::Error { |
| 1113 | error: serde_json::json!({ "message": "Model not exist." }), |
| 1114 | }, |
| 1115 | // Everything after the error frame must never reach the UI. |
| 1116 | crate::llm_client::mock::canned::text_delta(0, "TRAILING-DELTA-AFTER-FAILURE"), |
| 1117 | crate::llm_client::mock::canned::block_stop(0), |
| 1118 | ]])); |
| 1119 | let client: crate::core::model_client::SharedModelClient = model.clone(); |
| 1120 | let config = Config::default(); |
| 1121 | let engine_config = EngineConfig { |
| 1122 | max_steps: 1, |
| 1123 | snapshots_enabled: false, |
| 1124 | subagents_enabled: false, |
| 1125 | terminal_chrome_enabled: false, |
| 1126 | ..EngineConfig::default() |
| 1127 | }; |
| 1128 | let (engine, handle) = Engine::new_with_model_client(engine_config, &config, client); |
| 1129 | let run_task = tokio::spawn(engine.run()); |
| 1130 | |
| 1131 | handle |
| 1132 | .send(Op::SendMessage(TurnSpec { |
| 1133 | max_output_tokens: None, |
| 1134 | content: "solve the task".to_string(), |
| 1135 | images: Vec::new(), |
| 1136 | mode: AppMode::Agent, |
| 1137 | route: resolved_route_for_test(&config, crate::config::DEFAULT_TEXT_MODEL), |
| 1138 | compaction: Box::new(CompactionConfig::default()), |
| 1139 | initial_routed_usage: Box::default(), |
| 1140 | goal_objective: None, |
| 1141 | goal_token_budget: None, |
| 1142 | goal_status: crate::tools::goal::GoalStatus::Active, |
| 1143 | reasoning_effort: None, |
| 1144 | reasoning_effort_auto: false, |
| 1145 | auto_model: false, |
| 1146 | allow_shell: false, |
| 1147 | trust_mode: false, |
| 1148 | auto_approve: false, |
| 1149 | approval_mode: ApprovalMode::Suggest, |
| 1150 | translation_enabled: false, |
| 1151 | allowed_tools: None, |
| 1152 | dynamic_tools: Vec::new(), |
| 1153 | hook_executor: None, |
| 1154 | verbosity: None, |
| 1155 | provenance: UserInputProvenance::ExternalUser, |
| 1156 | submission_id: None, |
| 1157 | })) |
| 1158 | .await |
| 1159 | .expect("send midstream-error turn"); |
| 1160 | |
| 1161 | let mut events = Vec::new(); |
| 1162 | loop { |
| 1163 | let event = tokio::time::timeout(model_turn_event_timeout(), async { |
| 1164 | handle.rx_event.write().await.recv().await |
| 1165 | }) |
| 1166 | .await |
| 1167 | .expect("midstream-error event timeout") |
| 1168 | .expect("midstream-error event"); |
| 1169 | let terminal = matches!(event, Event::TurnComplete { .. }); |
| 1170 | events.push(event); |
| 1171 | if terminal { |
| 1172 | break; |
| 1173 | } |
| 1174 | } |
| 1175 | |
| 1176 | // The failure is surfaced as a typed error event at Error severity. |
| 1177 | assert!(events.iter().any(|event| matches!( |
| 1178 | event, |
| 1179 | Event::Error { envelope, .. } |
| 1180 | if envelope.message.contains("Model not exist.") |
| 1181 | && envelope.category == crate::error_taxonomy::ErrorCategory::InvalidInput |
| 1182 | && envelope.severity == crate::error_taxonomy::ErrorSeverity::Error |
| 1183 | ))); |
| 1184 | // Deltas after the failure frame never reach the UI. |
| 1185 | assert!( |
| 1186 | !events.iter().any(|event| matches!( |
| 1187 | event, |
| 1188 | Event::MessageDelta { content, .. } if content.contains("TRAILING-DELTA") |
| 1189 | )), |
| 1190 | "deltas after a mid-stream failure must be dropped: {events:?}" |
| 1191 | ); |
| 1192 | // The turn fails with the provider's message, not a generic truncation. |
| 1193 | let (status, error) = events |
| 1194 | .iter() |
| 1195 | .find_map(|event| match event { |
| 1196 | Event::TurnComplete { status, error, .. } => Some((status, error)), |
| 1197 | _ => None, |
| 1198 | }) |
| 1199 | .expect("midstream-error TurnComplete"); |
| 1200 | assert_eq!(*status, TurnOutcomeStatus::Failed); |
| 1201 | assert!( |
| 1202 | error |
| 1203 | .as_deref() |
| 1204 | .is_some_and(|e| e.contains("Model not exist.")), |
| 1205 | "{error:?}" |
| 1206 | ); |
| 1207 | // The stream is consumed exactly once — no re-issue of a terminal |
| 1208 | // rejection. |
| 1209 | assert_eq!(model.call_count(), 1); |
| 1210 | |
| 1211 | handle.send(Op::Shutdown).await.expect("shutdown engine"); |
| 1212 | run_task.await.expect("engine task"); |
| 1213 | } |
| 1214 | |
| 1215 | /// Emits a full billed response, then cancels the engine's own token while |
| 1216 | /// yielding the final stream event — modeling Esc arriving right after the |
| 1217 | /// provider finished charging for the response. |
| 1218 | struct CancelAfterTerminalUsageModelClient { |
| 1219 | calls: std::sync::atomic::AtomicUsize, |
| 1220 | // The engine mints a fresh token per turn; read the live one through the |
| 1221 | // engine's shared cell at stream time. |
| 1222 | token: std::sync::Mutex<Option<Arc<StdMutex<tokio_util::sync::CancellationToken>>>>, |
| 1223 | } |
| 1224 | |
| 1225 | #[async_trait::async_trait] |
| 1226 | impl crate::core::model_client::ModelClient for CancelAfterTerminalUsageModelClient { |
| 1227 | fn provider_name(&self) -> &str { |
| 1228 | "mock" |
| 1229 | } |
| 1230 | |
| 1231 | fn model(&self) -> &str { |
| 1232 | "mock-model" |
| 1233 | } |
| 1234 | |
| 1235 | async fn create_message( |
| 1236 | &self, |
| 1237 | _request: codewhale_models::MessageRequest, |
| 1238 | ) -> anyhow::Result<codewhale_models::MessageResponse> { |
| 1239 | anyhow::bail!("unused") |
| 1240 | } |
| 1241 | |
| 1242 | async fn create_message_stream( |
| 1243 | &self, |
| 1244 | _request: codewhale_models::MessageRequest, |
| 1245 | ) -> anyhow::Result<crate::llm_client::StreamEventBox> { |
| 1246 | use crate::llm_client::mock::canned; |
| 1247 | |
| 1248 | self.calls.fetch_add(1, std::sync::atomic::Ordering::SeqCst); |
| 1249 | let shared = self |
| 1250 | .token |
| 1251 | .lock() |
| 1252 | .expect("token cell") |
| 1253 | .clone() |
| 1254 | .expect("token installed before turn"); |
| 1255 | let token = shared |
| 1256 | .lock() |
| 1257 | .unwrap_or_else(std::sync::PoisonError::into_inner) |
| 1258 | .clone(); |
| 1259 | let mut message_start = canned::message_start("cancel_after_usage"); |
| 1260 | if let StreamEvent::MessageStart { message } = &mut message_start { |
| 1261 | message.usage = Usage { |
| 1262 | input_tokens: 47, |
| 1263 | ..Default::default() |
| 1264 | }; |
| 1265 | } |
| 1266 | let events = vec![ |
| 1267 | message_start, |
| 1268 | canned::text_block_start(0), |
| 1269 | canned::text_delta(0, "answer the user was billed for"), |
| 1270 | canned::block_stop(0), |
| 1271 | canned::message_delta( |
| 1272 | "end_turn", |
| 1273 | Some(Usage { |
| 1274 | output_tokens: 9, |
| 1275 | ..Default::default() |
| 1276 | }), |
| 1277 | ), |
| 1278 | canned::message_stop(), |
| 1279 | ]; |
| 1280 | let last = events.len() - 1; |
| 1281 | let stream = futures_util::stream::iter(events.into_iter().enumerate().map( |
| 1282 | move |(index, event)| { |
| 1283 | if index == last { |
| 1284 | token.cancel(); |
| 1285 | } |
| 1286 | Ok(event) |
| 1287 | }, |
| 1288 | )); |
| 1289 | Ok(Box::pin(stream)) |
| 1290 | } |
| 1291 | |
| 1292 | async fn health_check(&self) -> anyhow::Result<bool> { |
| 1293 | Ok(true) |
| 1294 | } |
| 1295 | } |
| 1296 | |
| 1297 | #[tokio::test] |
| 1298 | async fn cancellation_after_terminal_usage_still_charges_the_turn() { |
| 1299 | let model = std::sync::Arc::new(CancelAfterTerminalUsageModelClient { |
| 1300 | calls: std::sync::atomic::AtomicUsize::new(0), |
| 1301 | token: std::sync::Mutex::new(None), |
| 1302 | }); |
| 1303 | let client: crate::core::model_client::SharedModelClient = model.clone(); |
| 1304 | let config = Config::default(); |
| 1305 | let workspace = tempdir().expect("tempdir"); |
| 1306 | let (engine, handle) = Engine::new_with_model_client( |
| 1307 | deterministic_engine_config(workspace.path()), |
| 1308 | &config, |
| 1309 | client, |
| 1310 | ); |
| 1311 | *model.token.lock().expect("token cell") = Some(engine.shared_cancel_token.clone()); |
| 1312 | let run_task = tokio::spawn(engine.run()); |
| 1313 | |
| 1314 | handle |
| 1315 | .send(external_user_message_op( |
| 1316 | "solve the task", |
| 1317 | AppMode::Agent, |
| 1318 | &config, |
| 1319 | )) |
| 1320 | .await |
| 1321 | .expect("send cancel-after-usage turn"); |
| 1322 | |
| 1323 | let mut events = Vec::new(); |
| 1324 | loop { |
| 1325 | let event = tokio::time::timeout(model_turn_event_timeout(), async { |
| 1326 | handle.rx_event.write().await.recv().await |
| 1327 | }) |
| 1328 | .await |
| 1329 | .expect("cancel-after-usage event timeout") |
| 1330 | .expect("cancel-after-usage event"); |
| 1331 | let terminal = matches!(event, Event::TurnComplete { .. }); |
| 1332 | events.push(event); |
| 1333 | if terminal { |
| 1334 | break; |
| 1335 | } |
| 1336 | } |
| 1337 | |
| 1338 | assert_eq!(model.calls.load(std::sync::atomic::Ordering::SeqCst), 1); |
| 1339 | assert!( |
| 1340 | events.iter().any(|event| matches!( |
| 1341 | event, |
| 1342 | Event::TurnUsage { usage, .. } |
| 1343 | if usage.input_tokens == 47 && usage.output_tokens == 9 |
| 1344 | )), |
| 1345 | "billed usage must be accounted even though the turn was cancelled" |
| 1346 | ); |
| 1347 | let status = events |
| 1348 | .iter() |
| 1349 | .find_map(|event| match event { |
| 1350 | Event::TurnComplete { status, .. } => Some(*status), |
| 1351 | _ => None, |
| 1352 | }) |
| 1353 | .expect("terminal TurnComplete"); |
| 1354 | assert_eq!(status, TurnOutcomeStatus::Interrupted); |
| 1355 | |
| 1356 | handle.send(Op::Shutdown).await.expect("shutdown engine"); |
| 1357 | run_task.await.expect("engine task"); |
| 1358 | } |
| 1359 | |
| 1360 | /// Drive one interactive (`terminal_chrome_enabled = true`) turn against the |
| 1361 | /// flaky client and collect every event through the terminal TurnComplete. |
| 1362 | async fn run_interactive_turn_with_flaky_network( |
| 1363 | failures: usize, |
| 1364 | ) -> (std::sync::Arc<FlakyNetworkDropModelClient>, Vec<Event>) { |
| 1365 | let model = std::sync::Arc::new(FlakyNetworkDropModelClient { |
| 1366 | calls: std::sync::atomic::AtomicUsize::new(0), |
| 1367 | failures, |
| 1368 | terminal_before_drop: false, |
| 1369 | content_before_drop: true, |
| 1370 | open_failure: None, |
| 1371 | }); |
| 1372 | let client: crate::core::model_client::SharedModelClient = model.clone(); |
| 1373 | let config = Config::default(); |
| 1374 | let engine_config = EngineConfig { |
| 1375 | max_steps: 1, |
| 1376 | snapshots_enabled: false, |
| 1377 | subagents_enabled: false, |
| 1378 | terminal_chrome_enabled: true, |
| 1379 | ..EngineConfig::default() |
| 1380 | }; |
| 1381 | let (engine, handle) = Engine::new_with_model_client(engine_config, &config, client); |
| 1382 | let run_task = tokio::spawn(engine.run()); |
| 1383 | |
| 1384 | handle |
| 1385 | .send(Op::SendMessage(TurnSpec { |
| 1386 | max_output_tokens: None, |
| 1387 | content: "solve the task".to_string(), |
| 1388 | images: Vec::new(), |
| 1389 | mode: AppMode::Agent, |
| 1390 | route: resolved_route_for_test(&config, crate::config::DEFAULT_TEXT_MODEL), |
| 1391 | compaction: Box::new(CompactionConfig::default()), |
| 1392 | initial_routed_usage: Box::default(), |
| 1393 | goal_objective: None, |
| 1394 | goal_token_budget: None, |
| 1395 | goal_status: crate::tools::goal::GoalStatus::Active, |
| 1396 | reasoning_effort: None, |
| 1397 | reasoning_effort_auto: false, |
| 1398 | auto_model: false, |
| 1399 | allow_shell: false, |
| 1400 | trust_mode: false, |
| 1401 | auto_approve: false, |
| 1402 | approval_mode: ApprovalMode::Suggest, |
| 1403 | translation_enabled: false, |
| 1404 | allowed_tools: None, |
| 1405 | dynamic_tools: Vec::new(), |
| 1406 | hook_executor: None, |
| 1407 | verbosity: None, |
| 1408 | provenance: UserInputProvenance::ExternalUser, |
| 1409 | submission_id: None, |
| 1410 | })) |
| 1411 | .await |
| 1412 | .expect("send interactive flaky-network turn"); |
| 1413 | |
| 1414 | let mut events = Vec::new(); |
| 1415 | loop { |
| 1416 | let event = tokio::time::timeout(model_turn_event_timeout(), async { |
| 1417 | handle.rx_event.write().await.recv().await |
| 1418 | }) |
| 1419 | .await |
| 1420 | .expect("interactive flaky-network event timeout") |
| 1421 | .expect("interactive flaky-network event"); |
| 1422 | let terminal = matches!(event, Event::TurnComplete { .. }); |
| 1423 | events.push(event); |
| 1424 | if terminal { |
| 1425 | break; |
| 1426 | } |
| 1427 | } |
| 1428 | handle.send(Op::Shutdown).await.expect("shutdown engine"); |
| 1429 | run_task.await.expect("engine task"); |
| 1430 | (model, events) |
| 1431 | } |
| 1432 | |
| 1433 | #[tokio::test] |
| 1434 | async fn interactive_turn_preserves_partial_reply_and_recovers_after_network_drop() { |
| 1435 | let (model, events) = run_interactive_turn_with_flaky_network(1).await; |
| 1436 | |
| 1437 | assert_eq!( |
| 1438 | model.calls.load(std::sync::atomic::Ordering::SeqCst), |
| 1439 | 2, |
| 1440 | "the dropped stream must be re-issued exactly once" |
| 1441 | ); |
| 1442 | let (status, error) = events |
| 1443 | .iter() |
| 1444 | .find_map(|event| match event { |
| 1445 | Event::TurnComplete { status, error, .. } => Some((status, error)), |
| 1446 | _ => None, |
| 1447 | }) |
| 1448 | .expect("terminal TurnComplete"); |
| 1449 | assert_eq!( |
| 1450 | *status, |
| 1451 | TurnOutcomeStatus::Completed, |
| 1452 | "a recovered retry must complete the turn: {error:?}" |
| 1453 | ); |
| 1454 | assert!(error.is_none(), "recovered turn must not report an error"); |
| 1455 | assert!( |
| 1456 | events.iter().any(|event| matches!( |
| 1457 | event, |
| 1458 | Event::Status { message } if message.starts_with("Retry attempt: stream-resume 1/") |
| 1459 | )), |
| 1460 | "the first actual resume must remain visible: {events:?}" |
| 1461 | ); |
| 1462 | assert!( |
| 1463 | !events |
| 1464 | .iter() |
| 1465 | .any(|event| matches!(event, Event::Error { .. })), |
| 1466 | "a transient drop that the retry recovers must not surface an error event: {events:?}" |
| 1467 | ); |
| 1468 | |
| 1469 | // The visible fragment must survive as an assistant message, followed by |
| 1470 | // the retried assistant content — and nothing else. The recovery is |
| 1471 | // typed internal state, so no synthetic `[runtime]` user turn may appear. |
| 1472 | let transcript = events |
| 1473 | .iter() |
| 1474 | .rev() |
| 1475 | .find_map(|event| match event { |
| 1476 | Event::SessionUpdated { messages, .. } => Some(messages.clone()), |
| 1477 | _ => None, |
| 1478 | }) |
| 1479 | .expect("final SessionUpdated"); |
| 1480 | assert!( |
| 1481 | transcript |
| 1482 | .iter() |
| 1483 | .flat_map(|message| message.content.iter()) |
| 1484 | .all(|block| match block { |
| 1485 | ContentBlock::Text { text, .. } => !text.contains("[runtime]"), |
| 1486 | _ => true, |
| 1487 | }), |
| 1488 | "a retried turn must not insert a synthetic [runtime] user message: {transcript:?}" |
| 1489 | ); |
| 1490 | assert_eq!( |
| 1491 | transcript |
| 1492 | .iter() |
| 1493 | .filter(|message| message.role == "user") |
| 1494 | .count(), |
| 1495 | 1, |
| 1496 | "the operator's own turn must be the only user message: {transcript:?}" |
| 1497 | ); |
| 1498 | let assistant_cells = transcript |
| 1499 | .iter() |
| 1500 | .filter(|message| message.role == "assistant") |
| 1501 | .count(); |
| 1502 | assert_eq!( |
| 1503 | assistant_cells, 2, |
| 1504 | "preserved fragment + one authoritative continuation: {transcript:?}" |
| 1505 | ); |
| 1506 | let transcript_text = transcript |
| 1507 | .iter() |
| 1508 | .flat_map(|message| message.content.iter()) |
| 1509 | .filter_map(|block| match block { |
| 1510 | ContentBlock::Text { text, .. } => Some(text.as_str()), |
| 1511 | _ => None, |
| 1512 | }) |
| 1513 | .collect::<Vec<_>>() |
| 1514 | .join("\n"); |
| 1515 | assert!( |
| 1516 | transcript_text.contains("partial answer that must be discarded"), |
| 1517 | "the visible partial reply must be preserved in the session: {transcript_text}" |
| 1518 | ); |
| 1519 | assert_eq!( |
| 1520 | transcript_text.matches("recovered after retry").count(), |
| 1521 | 1, |
| 1522 | "exactly one authoritative final answer, not a duplicate: {transcript_text}" |
| 1523 | ); |
| 1524 | } |
| 1525 | |
| 1526 | /// Streams only hidden reasoning and then dies with the network-class read |
| 1527 | /// error; later streams complete a normal text turn. This is the shape of the |
| 1528 | /// 0.9.10 regression: a thinking-only drop used to persist a synthetic |
| 1529 | /// `[runtime]` user message claiming a partial answer had been preserved. |
| 1530 | struct ThinkingOnlyDropModelClient { |
| 1531 | calls: std::sync::atomic::AtomicUsize, |
| 1532 | failures: usize, |
| 1533 | } |
| 1534 | |
| 1535 | #[async_trait::async_trait] |
| 1536 | impl crate::core::model_client::ModelClient for ThinkingOnlyDropModelClient { |
| 1537 | fn provider_name(&self) -> &str { |
| 1538 | "flaky-network" |
| 1539 | } |
| 1540 | |
| 1541 | fn model(&self) -> &str { |
| 1542 | "local-model" |
| 1543 | } |
| 1544 | |
| 1545 | async fn create_message( |
| 1546 | &self, |
| 1547 | _request: codewhale_models::MessageRequest, |
| 1548 | ) -> anyhow::Result<codewhale_models::MessageResponse> { |
| 1549 | anyhow::bail!("thinking-only drop regression uses the streaming model boundary") |
| 1550 | } |
| 1551 | |
| 1552 | async fn create_message_stream( |
| 1553 | &self, |
| 1554 | _request: codewhale_models::MessageRequest, |
| 1555 | ) -> anyhow::Result<crate::llm_client::StreamEventBox> { |
| 1556 | use crate::llm_client::mock::canned; |
| 1557 | let call = self |
| 1558 | .calls |
| 1559 | .fetch_add(1, std::sync::atomic::Ordering::SeqCst) |
| 1560 | .saturating_add(1); |
| 1561 | if call <= self.failures { |
| 1562 | // Hidden reasoning only — no text block is ever opened, so nothing |
| 1563 | // visible streams before the transport dies. This still flips |
| 1564 | // `any_content_received`, which is what routes the drop to the |
| 1565 | // interactive resume path instead of the transparent retry. |
| 1566 | let events: Vec<anyhow::Result<codewhale_models::StreamEvent>> = vec![ |
| 1567 | Ok(canned::message_start("thinking_only_msg")), |
| 1568 | Ok(StreamEvent::ContentBlockStart { |
| 1569 | index: 0, |
| 1570 | content_block: codewhale_models::ContentBlockStart::Thinking { |
| 1571 | thinking: String::new(), |
| 1572 | }, |
| 1573 | }), |
| 1574 | Ok(canned::thinking_delta( |
| 1575 | 0, |
| 1576 | "hidden reasoning that no operator ever saw", |
| 1577 | )), |
| 1578 | Err(anyhow::anyhow!( |
| 1579 | "Stream read error: error decoding response body" |
| 1580 | )), |
| 1581 | ]; |
| 1582 | return Ok(Box::pin(futures_util::stream::iter(events))); |
| 1583 | } |
| 1584 | let events = canned::simple_text_turn("the one authoritative answer") |
| 1585 | .into_iter() |
| 1586 | .map(Ok); |
| 1587 | Ok(Box::pin(futures_util::stream::iter(events))) |
| 1588 | } |
| 1589 | |
| 1590 | async fn health_check(&self) -> anyhow::Result<bool> { |
| 1591 | Ok(true) |
| 1592 | } |
| 1593 | } |
| 1594 | |
| 1595 | struct TransportRetryModelClient { |
| 1596 | attempts: std::sync::atomic::AtomicUsize, |
| 1597 | failures: usize, |
| 1598 | } |
| 1599 | |
| 1600 | #[async_trait::async_trait] |
| 1601 | impl crate::core::model_client::ModelClient for TransportRetryModelClient { |
| 1602 | fn provider_name(&self) -> &str { |
| 1603 | "custom" |
| 1604 | } |
| 1605 | fn model(&self) -> &str { |
| 1606 | crate::config::DEFAULT_TEXT_MODEL |
| 1607 | } |
| 1608 | async fn create_message( |
| 1609 | &self, |
| 1610 | _: codewhale_models::MessageRequest, |
| 1611 | ) -> anyhow::Result<codewhale_models::MessageResponse> { |
| 1612 | anyhow::bail!("streaming fixture only") |
| 1613 | } |
| 1614 | async fn create_message_stream( |
| 1615 | &self, |
| 1616 | _: codewhale_models::MessageRequest, |
| 1617 | ) -> anyhow::Result<crate::llm_client::StreamEventBox> { |
| 1618 | crate::llm_client::with_retry( |
| 1619 | &crate::llm_client::RetryConfig { |
| 1620 | max_retries: 2, |
| 1621 | initial_delay: 0.0, |
| 1622 | jitter: false, |
| 1623 | ..Default::default() |
| 1624 | }, |
| 1625 | || async { |
| 1626 | let attempt = self |
| 1627 | .attempts |
| 1628 | .fetch_add(1, std::sync::atomic::Ordering::SeqCst); |
| 1629 | if attempt < self.failures { |
| 1630 | Err(crate::llm_client::LlmError::ServerError { |
| 1631 | status: 503, |
| 1632 | message: "RAW-ENGINE-RETRY-FAILURE".into(), |
| 1633 | }) |
| 1634 | } else { |
| 1635 | Ok(()) |
| 1636 | } |
| 1637 | }, |
| 1638 | None, |
| 1639 | ) |
| 1640 | .await |
| 1641 | .map_err(|error| anyhow::Error::new(error.last_error))?; |
| 1642 | Ok(Box::pin(futures_util::stream::iter( |
| 1643 | crate::llm_client::mock::canned::simple_text_turn("transport recovered answer") |
| 1644 | .into_iter() |
| 1645 | .map(Ok), |
| 1646 | ))) |
| 1647 | } |
| 1648 | async fn health_check(&self) -> anyhow::Result<bool> { |
| 1649 | Ok(true) |
| 1650 | } |
| 1651 | } |
| 1652 | |
| 1653 | #[tokio::test] |
| 1654 | async fn engine_transport_retry_receipts_and_counts_match_actual_scripted_attempts() { |
| 1655 | for (failures, quiet) in [(1, false), (3, false), (1, true)] { |
| 1656 | let workspace = tempdir().unwrap(); |
| 1657 | let model = std::sync::Arc::new(TransportRetryModelClient { |
| 1658 | attempts: std::sync::atomic::AtomicUsize::new(0), |
| 1659 | failures, |
| 1660 | }); |
| 1661 | let config = Config { |
| 1662 | notifications: Some(crate::config::NotificationsConfig { |
| 1663 | quiet, |
| 1664 | ..Default::default() |
| 1665 | }), |
| 1666 | ..Config::default() |
| 1667 | }; |
| 1668 | let (mut engine, handle) = Engine::new_with_model_client( |
| 1669 | deterministic_engine_config(workspace.path()), |
| 1670 | &config, |
| 1671 | model.clone(), |
| 1672 | ); |
| 1673 | let surface = test_tool_surface( |
| 1674 | &engine, |
| 1675 | crate::tools::ToolRegistry::new(crate::tools::ToolContext::new( |
| 1676 | workspace.path().to_path_buf(), |
| 1677 | )), |
| 1678 | None, |
| 1679 | AppMode::Agent, |
| 1680 | ); |
| 1681 | let mut turn = crate::core::turn::TurnContext::new(1); |
| 1682 | let (status, error) = engine.run_turn(&mut turn, surface, None, None).await; |
| 1683 | let expected_retries = failures.min(2) as u32; |
| 1684 | assert_eq!( |
| 1685 | model.attempts.load(std::sync::atomic::Ordering::SeqCst), |
| 1686 | (expected_retries + 1) as usize |
| 1687 | ); |
| 1688 | assert_eq!(turn.stop_diagnostics.model_requests_started, 1); |
| 1689 | assert_eq!(turn.stop_diagnostics.transport_retries, expected_retries); |
| 1690 | assert_eq!(turn.stop_diagnostics.stream_resumes, 0); |
| 1691 | assert_eq!(turn.stop_diagnostics.transparent_stream_retries, 0); |
| 1692 | let mut receiver = handle.rx_event.write().await; |
| 1693 | let events = std::iter::from_fn(|| receiver.try_recv().ok()).collect::<Vec<_>>(); |
| 1694 | if quiet { |
| 1695 | assert!( |
| 1696 | !events.iter().any(|event| matches!(event, |
| 1697 | Event::Status { message } if message.starts_with("Retry attempt:") |
| 1698 | )), |
| 1699 | "quiet suppresses attempt lines, not actual counts or completed summaries" |
| 1700 | ); |
| 1701 | } else { |
| 1702 | for attempt in 1..=expected_retries { |
| 1703 | assert!(events.iter().any(|event| matches!(event, |
| 1704 | Event::Status { message } if message.starts_with(&format!("Retry attempt: transport {attempt}/2;")) |
| 1705 | ))); |
| 1706 | } |
| 1707 | } |
| 1708 | assert!( |
| 1709 | !events.iter().any(|event| matches!(event, |
| 1710 | Event::Status { message } if message.contains("RAW-ENGINE-RETRY-FAILURE") |
| 1711 | )), |
| 1712 | "remote payload belongs to the original error, never a retry receipt" |
| 1713 | ); |
| 1714 | if failures == 1 { |
| 1715 | assert_eq!(status, TurnOutcomeStatus::Completed, "{error:?}"); |
| 1716 | assert!(events.iter().any(|event| matches!(event, |
| 1717 | Event::Status { message } if message == "Retry recovery: transport request recovered after 1 retries" |
| 1718 | ))); |
| 1719 | } else { |
| 1720 | assert_eq!(status, TurnOutcomeStatus::Failed); |
| 1721 | assert!(error.unwrap().contains("RAW-ENGINE-RETRY-FAILURE")); |
| 1722 | assert!(events.iter().any(|event| matches!(event, |
| 1723 | Event::Status { message } if message.starts_with("Retry exhaustion: transport request stopped after 2 retries;") |
| 1724 | ))); |
| 1725 | } |
| 1726 | assert!(engine.session.messages.iter().all(|message| message.content.iter().all(|block| |
| 1727 | !matches!(block, ContentBlock::Text { text, .. } if crate::core::events::is_retry_status_receipt(text)) |
| 1728 | )), "receipts must never join the provider prompt or session message graph"); |
| 1729 | } |
| 1730 | } |
| 1731 | |
| 1732 | #[tokio::test] |
| 1733 | async fn full_retry_receipt_queue_cancellation_does_not_dispatch_an_extra_request() { |
| 1734 | let workspace = tempdir().unwrap(); |
| 1735 | let (engine, _handle) = Engine::new_with_model_client( |
| 1736 | deterministic_engine_config(workspace.path()), |
| 1737 | &Config::default(), |
| 1738 | std::sync::Arc::new(TransportRetryModelClient { |
| 1739 | attempts: std::sync::atomic::AtomicUsize::new(0), |
| 1740 | failures: 0, |
| 1741 | }), |
| 1742 | ); |
| 1743 | while engine.tx_event.try_send(Event::status("occupied")).is_ok() {} |
| 1744 | let observation = engine.request_retry_observation(); |
| 1745 | let count = observation.retries.clone(); |
| 1746 | let mut calls = 0; |
| 1747 | let policy = crate::llm_client::RetryConfig { |
| 1748 | max_retries: 2, |
| 1749 | initial_delay: 0.0, |
| 1750 | jitter: false, |
| 1751 | ..Default::default() |
| 1752 | }; |
| 1753 | let outcome = { |
| 1754 | let request = crate::llm_client::observe_request_retries( |
| 1755 | Some(observation), |
| 1756 | crate::llm_client::with_retry( |
| 1757 | &policy, |
| 1758 | || { |
| 1759 | calls += 1; |
| 1760 | async { |
| 1761 | Err::<(), _>(crate::llm_client::LlmError::ServerError { |
| 1762 | status: 503, |
| 1763 | message: "private".into(), |
| 1764 | }) |
| 1765 | } |
| 1766 | }, |
| 1767 | None, |
| 1768 | ), |
| 1769 | ); |
| 1770 | tokio::pin!(request); |
| 1771 | assert!( |
| 1772 | futures_util::poll!(request.as_mut()).is_pending(), |
| 1773 | "the receipt waits on the one full queue" |
| 1774 | ); |
| 1775 | engine.cancel_token.cancel(); |
| 1776 | tokio::time::timeout(Duration::from_secs(1), async { |
| 1777 | tokio::select! { |
| 1778 | biased; |
| 1779 | () = engine.cancel_token.cancelled() => None, |
| 1780 | result = request.as_mut() => Some(result), |
| 1781 | } |
| 1782 | }) |
| 1783 | .await |
| 1784 | .unwrap() |
| 1785 | }; |
| 1786 | assert!(outcome.is_none()); |
| 1787 | assert_eq!(calls, 1); |
| 1788 | assert_eq!( |
| 1789 | count.load(std::sync::atomic::Ordering::Relaxed), |
| 1790 | 0, |
| 1791 | "a scheduled retry cancelled before dispatch is not spent" |
| 1792 | ); |
| 1793 | } |
| 1794 |