| 1 | use super::*; |
| 2 | use crate::core::engine::{MockApprovalEvent, mock_engine_handle}; |
| 3 | use crate::core::events::{Event as EngineEvent, TurnOutcomeStatus}; |
| 4 | use std::time::{Duration, Instant}; |
| 5 | use tokio::sync::{mpsc, oneshot}; |
| 6 | use tokio::time::sleep; |
| 7 | use uuid::Uuid; |
| 8 | |
| 9 | fn test_runtime_dir() -> PathBuf { |
| 10 | std::env::temp_dir().join(format!("deepseek-runtime-threads-{}", Uuid::new_v4())) |
| 11 | } |
| 12 | |
| 13 | const EVENT_PROCESS_ROLE_ENV: &str = "CODEWHALE_TEST_EVENT_PROCESS_ROLE"; |
| 14 | const EVENT_PROCESS_ROOT_ENV: &str = "CODEWHALE_TEST_EVENT_PROCESS_ROOT"; |
| 15 | const EVENT_PROCESS_THREAD_ENV: &str = "CODEWHALE_TEST_EVENT_PROCESS_THREAD"; |
| 16 | const EVENT_PROCESS_SIGNAL_ENV: &str = "CODEWHALE_TEST_EVENT_PROCESS_SIGNAL"; |
| 17 | const EVENT_PROCESS_START_ENV: &str = "CODEWHALE_TEST_EVENT_PROCESS_START"; |
| 18 | const EVENT_PROCESS_WORKER_ENV: &str = "CODEWHALE_TEST_EVENT_PROCESS_WORKER"; |
| 19 | const EVENT_PROCESS_COUNT_ENV: &str = "CODEWHALE_TEST_EVENT_PROCESS_COUNT"; |
| 20 | const EVENT_PROCESS_HELPER: &str = "runtime_threads::tests::runtime_event_process_child_helper"; |
| 21 | |
| 22 | #[test] |
| 23 | #[ignore = "spawned by real cross-process Runtime event tests"] |
| 24 | fn runtime_event_process_child_helper() { |
| 25 | let Ok(role) = std::env::var(EVENT_PROCESS_ROLE_ENV) else { |
| 26 | return; |
| 27 | }; |
| 28 | let root = PathBuf::from( |
| 29 | std::env::var_os(EVENT_PROCESS_ROOT_ENV).expect("event child needs a store root"), |
| 30 | ); |
| 31 | let thread_id = std::env::var(EVENT_PROCESS_THREAD_ENV).expect("event child needs a thread id"); |
| 32 | let signal = PathBuf::from( |
| 33 | std::env::var_os(EVENT_PROCESS_SIGNAL_ENV).expect("event child needs a signal path"), |
| 34 | ); |
| 35 | let store = RuntimeThreadStore::open(root).expect("event child opens Runtime store"); |
| 36 | |
| 37 | match role.as_str() { |
| 38 | "writer" => { |
| 39 | let start = PathBuf::from( |
| 40 | std::env::var_os(EVENT_PROCESS_START_ENV) |
| 41 | .expect("event writer needs a start barrier"), |
| 42 | ); |
| 43 | let worker = |
| 44 | std::env::var(EVENT_PROCESS_WORKER_ENV).expect("event writer needs a worker id"); |
| 45 | let count = std::env::var(EVENT_PROCESS_COUNT_ENV) |
| 46 | .expect("event writer needs a count") |
| 47 | .parse::<u64>() |
| 48 | .expect("event writer count must be numeric"); |
| 49 | std::fs::write(&signal, b"ready").expect("announce ready event writer"); |
| 50 | wait_for_runtime_event_test_file(&start, "writer start barrier"); |
| 51 | let runtime = tokio::runtime::Builder::new_current_thread() |
| 52 | .enable_all() |
| 53 | .build() |
| 54 | .expect("build event writer runtime"); |
| 55 | runtime.block_on(async { |
| 56 | for index in 0..count { |
| 57 | store |
| 58 | .append_event( |
| 59 | &thread_id, |
| 60 | None, |
| 61 | None, |
| 62 | "process.event", |
| 63 | json!({ "worker": worker, "index": index }), |
| 64 | ) |
| 65 | .await |
| 66 | .expect("append cross-process Runtime event"); |
| 67 | } |
| 68 | }); |
| 69 | } |
| 70 | "holder" => { |
| 71 | store |
| 72 | .with_event_transaction(Duration::from_secs(5), || { |
| 73 | std::fs::write(&signal, b"locked").expect("announce held event lock"); |
| 74 | std::thread::sleep(Duration::from_secs(60)); |
| 75 | Ok(()) |
| 76 | }) |
| 77 | .expect("hold Runtime event transaction"); |
| 78 | } |
| 79 | "torn" => { |
| 80 | store |
| 81 | .with_event_transaction(Duration::from_secs(5), || { |
| 82 | let mut state = load_runtime_store_state(&store.state_path)?; |
| 83 | let seq = state.next_seq; |
| 84 | state.next_seq = seq.checked_add(1).context("reserve torn event sequence")?; |
| 85 | write_json_atomic(&store.state_path, &state)?; |
| 86 | let path = store.events_path(&thread_id)?; |
| 87 | let mut file = |
| 88 | open_runtime_store_file(&path, "torn Runtime event fixture", |options| { |
| 89 | options.create(true).append(true); |
| 90 | })?; |
| 91 | let record = RuntimeEventRecord { |
| 92 | schema_version: CURRENT_RUNTIME_SCHEMA_VERSION, |
| 93 | seq, |
| 94 | timestamp: Utc::now(), |
| 95 | thread_id: thread_id.clone(), |
| 96 | turn_id: None, |
| 97 | item_id: None, |
| 98 | event: "torn.process.event".to_string(), |
| 99 | payload: json!({}), |
| 100 | }; |
| 101 | file.write_all(&serde_json::to_vec(&record)?)?; |
| 102 | file.flush()?; |
| 103 | file.sync_all()?; |
| 104 | std::fs::write(&signal, b"torn").expect("announce torn event append"); |
| 105 | std::thread::sleep(Duration::from_secs(60)); |
| 106 | Ok(()) |
| 107 | }) |
| 108 | .expect("hold torn Runtime event transaction"); |
| 109 | } |
| 110 | "phantom" => { |
| 111 | store |
| 112 | .with_event_transaction(Duration::from_secs(5), || { |
| 113 | let path = store.events_path(&thread_id)?; |
| 114 | let mut append = open_runtime_store_file( |
| 115 | &path, |
| 116 | "phantom Runtime event fixture", |
| 117 | |options| { |
| 118 | options.create(true).append(true); |
| 119 | }, |
| 120 | )?; |
| 121 | let rollback = open_runtime_store_file( |
| 122 | &path, |
| 123 | "phantom Runtime event rollback", |
| 124 | |options| { |
| 125 | options.write(true); |
| 126 | }, |
| 127 | )?; |
| 128 | validate_same_runtime_store_file_handles(&append, &rollback, &path)?; |
| 129 | let original_len = append.metadata()?.len(); |
| 130 | serde_json::to_writer( |
| 131 | &mut append, |
| 132 | &RuntimeEventRecord { |
| 133 | schema_version: CURRENT_RUNTIME_SCHEMA_VERSION, |
| 134 | seq: 999, |
| 135 | timestamp: Utc::now(), |
| 136 | thread_id: thread_id.clone(), |
| 137 | turn_id: None, |
| 138 | item_id: None, |
| 139 | event: "phantom.process.event".to_string(), |
| 140 | payload: json!({}), |
| 141 | }, |
| 142 | )?; |
| 143 | append.write_all(b"\n")?; |
| 144 | append.flush()?; |
| 145 | std::fs::write(&signal, b"visible").expect("announce rollback candidate"); |
| 146 | std::thread::sleep(Duration::from_millis(300)); |
| 147 | rollback_failed_event_append_handle(&rollback, original_len) |
| 148 | }) |
| 149 | .expect("roll back phantom Runtime event"); |
| 150 | } |
| 151 | other => panic!("unknown Runtime event child role {other}"), |
| 152 | } |
| 153 | } |
| 154 | |
| 155 | fn wait_for_runtime_event_test_file(path: &Path, label: &str) { |
| 156 | let deadline = Instant::now() + Duration::from_secs(30); |
| 157 | while !path.exists() { |
| 158 | assert!( |
| 159 | Instant::now() < deadline, |
| 160 | "timed out waiting for {label} at {}", |
| 161 | path.display() |
| 162 | ); |
| 163 | std::thread::sleep(Duration::from_millis(5)); |
| 164 | } |
| 165 | } |
| 166 | |
| 167 | fn spawn_runtime_event_child( |
| 168 | role: &str, |
| 169 | root: &Path, |
| 170 | thread_id: &str, |
| 171 | signal: &Path, |
| 172 | ) -> RuntimeEventChildGuard { |
| 173 | let mut command = std::process::Command::new( |
| 174 | std::env::current_exe().expect("current test executable for Runtime event child"), |
| 175 | ); |
| 176 | command |
| 177 | .arg(EVENT_PROCESS_HELPER) |
| 178 | .args(["--exact", "--ignored", "--test-threads", "1"]) |
| 179 | .env(EVENT_PROCESS_ROLE_ENV, role) |
| 180 | .env(EVENT_PROCESS_ROOT_ENV, root) |
| 181 | .env(EVENT_PROCESS_THREAD_ENV, thread_id) |
| 182 | .env(EVENT_PROCESS_SIGNAL_ENV, signal) |
| 183 | .env(EVENT_PROCESS_START_ENV, root.join("process-writers.start")) |
| 184 | .env(EVENT_PROCESS_WORKER_ENV, signal.as_os_str()) |
| 185 | .env(EVENT_PROCESS_COUNT_ENV, "8") |
| 186 | .stdin(std::process::Stdio::null()) |
| 187 | .stdout(std::process::Stdio::null()) |
| 188 | .stderr(std::process::Stdio::null()); |
| 189 | RuntimeEventChildGuard::new(command.spawn().expect("spawn Runtime event child")) |
| 190 | } |
| 191 | |
| 192 | struct RuntimeEventChildGuard(Option<std::process::Child>); |
| 193 | |
| 194 | impl RuntimeEventChildGuard { |
| 195 | const fn new(child: std::process::Child) -> Self { |
| 196 | Self(Some(child)) |
| 197 | } |
| 198 | |
| 199 | fn wait_success(mut self, label: &str) { |
| 200 | use wait_timeout::ChildExt as _; |
| 201 | |
| 202 | let outcome = self |
| 203 | .0 |
| 204 | .as_mut() |
| 205 | .expect("Runtime event child is present") |
| 206 | .wait_timeout(Duration::from_secs(30)); |
| 207 | match outcome { |
| 208 | Ok(Some(status)) => { |
| 209 | self.0.take(); |
| 210 | assert!(status.success(), "{label} failed with {status}"); |
| 211 | } |
| 212 | Ok(None) => panic!("timed out waiting for {label}"), |
| 213 | Err(error) => panic!("failed waiting for {label}: {error}"), |
| 214 | } |
| 215 | } |
| 216 | |
| 217 | fn kill_and_reap(mut self) -> std::io::Result<std::process::ExitStatus> { |
| 218 | let child = self.0.as_mut().expect("Runtime event child is present"); |
| 219 | let kill = child.kill(); |
| 220 | let status = child.wait(); |
| 221 | if status.is_ok() { |
| 222 | self.0.take(); |
| 223 | } |
| 224 | kill.and(status) |
| 225 | } |
| 226 | } |
| 227 | |
| 228 | impl Drop for RuntimeEventChildGuard { |
| 229 | fn drop(&mut self) { |
| 230 | if let Some(child) = self.0.as_mut() { |
| 231 | let _ = child.kill(); |
| 232 | let _ = child.wait(); |
| 233 | } |
| 234 | } |
| 235 | } |
| 236 | |
| 237 | fn test_manager_config(data_dir: PathBuf) -> RuntimeThreadManagerConfig { |
| 238 | RuntimeThreadManagerConfig { |
| 239 | task_data_dir: data_dir.clone(), |
| 240 | data_dir, |
| 241 | max_active_threads: 4, |
| 242 | } |
| 243 | } |
| 244 | |
| 245 | fn test_manager(data_dir: PathBuf) -> Result<RuntimeThreadManager> { |
| 246 | RuntimeThreadManager::open( |
| 247 | Config::default(), |
| 248 | PathBuf::from("."), |
| 249 | test_manager_config(data_dir), |
| 250 | ) |
| 251 | } |
| 252 | |
| 253 | struct ApprovalTimeoutGuard { |
| 254 | previous_ms: u64, |
| 255 | } |
| 256 | |
| 257 | impl Drop for ApprovalTimeoutGuard { |
| 258 | fn drop(&mut self) { |
| 259 | set_test_approval_decision_timeout_ms(self.previous_ms); |
| 260 | } |
| 261 | } |
| 262 | |
| 263 | fn test_approval_timeout_ms(ms: u64) -> ApprovalTimeoutGuard { |
| 264 | ApprovalTimeoutGuard { |
| 265 | previous_ms: set_test_approval_decision_timeout_ms(ms), |
| 266 | } |
| 267 | } |
| 268 | |
| 269 | struct DynamicToolTimeoutGuard { |
| 270 | previous_ms: u64, |
| 271 | } |
| 272 | |
| 273 | impl Drop for DynamicToolTimeoutGuard { |
| 274 | fn drop(&mut self) { |
| 275 | set_test_dynamic_tool_result_timeout_ms(self.previous_ms); |
| 276 | } |
| 277 | } |
| 278 | |
| 279 | fn test_dynamic_tool_timeout_ms(ms: u64) -> DynamicToolTimeoutGuard { |
| 280 | DynamicToolTimeoutGuard { |
| 281 | previous_ms: set_test_dynamic_tool_result_timeout_ms(ms), |
| 282 | } |
| 283 | } |
| 284 | |
| 285 | struct EventAppendFaultGuard { |
| 286 | restore: Option<EventAppendTestFaultRestore>, |
| 287 | } |
| 288 | |
| 289 | impl EventAppendFaultGuard { |
| 290 | fn arm(thread_id: &str, fault: EventAppendTestFault) -> Self { |
| 291 | Self::arm_repeated(thread_id, fault, 1) |
| 292 | } |
| 293 | |
| 294 | fn arm_repeated(thread_id: &str, fault: EventAppendTestFault, count: usize) -> Self { |
| 295 | Self { |
| 296 | restore: Some(set_test_event_append_fault(thread_id, fault, count)), |
| 297 | } |
| 298 | } |
| 299 | } |
| 300 | |
| 301 | impl Drop for EventAppendFaultGuard { |
| 302 | fn drop(&mut self) { |
| 303 | if let Some(restore) = self.restore.take() { |
| 304 | restore_test_event_append_fault(restore); |
| 305 | } |
| 306 | } |
| 307 | } |
| 308 | |
| 309 | fn sample_thread(thread_id: &str) -> ThreadRecord { |
| 310 | let now = Utc::now(); |
| 311 | ThreadRecord { |
| 312 | schema_version: CURRENT_RUNTIME_SCHEMA_VERSION, |
| 313 | id: thread_id.to_string(), |
| 314 | created_at: now, |
| 315 | updated_at: now, |
| 316 | model: DEFAULT_TEXT_MODEL.to_string(), |
| 317 | model_provider: None, |
| 318 | model_provider_id: None, |
| 319 | workspace: PathBuf::from("."), |
| 320 | mode: AppMode::Agent.as_setting().to_string(), |
| 321 | permission_posture: Some("ask".to_string()), |
| 322 | allow_shell: false, |
| 323 | trust_mode: false, |
| 324 | auto_approve: false, |
| 325 | latest_turn_id: None, |
| 326 | latest_response_bookmark: None, |
| 327 | archived: false, |
| 328 | system_prompt: None, |
| 329 | task_id: None, |
| 330 | title: None, |
| 331 | session_id: None, |
| 332 | } |
| 333 | } |
| 334 | |
| 335 | fn sample_turn(thread_id: &str, turn_id: &str, status: RuntimeTurnStatus) -> TurnRecord { |
| 336 | let now = Utc::now(); |
| 337 | TurnRecord { |
| 338 | schema_version: CURRENT_RUNTIME_SCHEMA_VERSION, |
| 339 | id: turn_id.to_string(), |
| 340 | thread_id: thread_id.to_string(), |
| 341 | status, |
| 342 | input_summary: "sample".to_string(), |
| 343 | created_at: now, |
| 344 | started_at: Some(now), |
| 345 | ended_at: None, |
| 346 | duration_ms: None, |
| 347 | usage: None, |
| 348 | permission_posture: None, |
| 349 | effective_provider: None, |
| 350 | effective_provider_id: None, |
| 351 | effective_billing_surface: None, |
| 352 | effective_endpoint_fingerprint: None, |
| 353 | effective_billing_mode: None, |
| 354 | effective_dispatched_at: None, |
| 355 | effective_model: None, |
| 356 | routed_usage: Vec::new(), |
| 357 | routed_usage_source_ids: Vec::new(), |
| 358 | routed_usage_dropped_records: 0, |
| 359 | error: None, |
| 360 | item_ids: Vec::new(), |
| 361 | steer_count: 0, |
| 362 | } |
| 363 | } |
| 364 | |
| 365 | fn set_test_turn_route( |
| 366 | turn: &mut TurnRecord, |
| 367 | provider: ApiProvider, |
| 368 | provider_identity: &str, |
| 369 | model: &str, |
| 370 | billing_surface: Option<&str>, |
| 371 | billing_mode: crate::cost_status::RouteBillingMode, |
| 372 | ) { |
| 373 | turn.persist_effective_route(&crate::cost_status::EffectiveRouteEnvelope { |
| 374 | provider, |
| 375 | provider_identity: provider_identity.to_string(), |
| 376 | model: model.to_string(), |
| 377 | billing_surface: billing_surface.map(str::to_string), |
| 378 | endpoint_fingerprint: None, |
| 379 | billing_mode, |
| 380 | dispatched_at: turn.created_at, |
| 381 | }); |
| 382 | } |
| 383 | |
| 384 | #[test] |
| 385 | fn runtime_compaction_uses_provider_route_context() { |
| 386 | let limits = codewhale_config::route::RouteLimits { |
| 387 | context_tokens: Some(272_000), |
| 388 | input_tokens: None, |
| 389 | output_tokens: None, |
| 390 | }; |
| 391 | let config = runtime_compaction_config( |
| 392 | ApiProvider::OpenaiCodex, |
| 393 | "gpt-5.5", |
| 394 | Some(limits), |
| 395 | false, |
| 396 | false, |
| 397 | 80.0, |
| 398 | ); |
| 399 | |
| 400 | assert!(config.enabled); |
| 401 | // The threshold is 80% of the route's spendable input budget after |
| 402 | // output reservation and headroom, not 80% of the raw context window. |
| 403 | assert_eq!(config.token_threshold, 213_504); |
| 404 | assert_eq!(config.effective_context_window, Some(272_000)); |
| 405 | } |
| 406 | |
| 407 | #[test] |
| 408 | fn legacy_turn_record_has_no_invented_route_provenance() { |
| 409 | let turn = sample_turn("thr_legacy", "turn_legacy", RuntimeTurnStatus::Completed); |
| 410 | let mut value = serde_json::to_value(turn).expect("serialize turn"); |
| 411 | let object = value.as_object_mut().expect("turn object"); |
| 412 | object.remove("effective_provider"); |
| 413 | object.remove("effective_provider_id"); |
| 414 | object.remove("effective_billing_surface"); |
| 415 | object.remove("effective_endpoint_fingerprint"); |
| 416 | object.remove("effective_billing_mode"); |
| 417 | object.remove("effective_dispatched_at"); |
| 418 | object.remove("effective_model"); |
| 419 | object.remove("routed_usage"); |
| 420 | |
| 421 | let restored: TurnRecord = serde_json::from_value(value).expect("deserialize legacy turn"); |
| 422 | assert_eq!(restored.effective_provider, None); |
| 423 | assert_eq!(restored.effective_billing_surface, None); |
| 424 | assert_eq!(restored.effective_billing_mode, None); |
| 425 | assert_eq!(restored.effective_dispatched_at, None); |
| 426 | assert_eq!(restored.effective_model, None); |
| 427 | assert!(restored.routed_usage.is_empty()); |
| 428 | } |
| 429 | |
| 430 | #[tokio::test] |
| 431 | async fn named_custom_thread_identity_round_trips_and_fails_closed_when_removed() -> Result<()> { |
| 432 | let mut custom = std::collections::HashMap::new(); |
| 433 | custom.insert( |
| 434 | "lm-studio".to_string(), |
| 435 | crate::config::ProviderConfig { |
| 436 | kind: Some("openai-compatible".to_string()), |
| 437 | base_url: Some("http://127.0.0.1:1234/v1".to_string()), |
| 438 | model: Some("local-default".to_string()), |
| 439 | ..crate::config::ProviderConfig::default() |
| 440 | }, |
| 441 | ); |
| 442 | let config = Config { |
| 443 | provider: Some("lm-studio".to_string()), |
| 444 | providers: Some(crate::config::ProvidersConfig { |
| 445 | custom, |
| 446 | ..crate::config::ProvidersConfig::default() |
| 447 | }), |
| 448 | ..Config::default() |
| 449 | }; |
| 450 | let manager = RuntimeThreadManager::open( |
| 451 | config.clone(), |
| 452 | PathBuf::from("."), |
| 453 | test_manager_config(test_runtime_dir()), |
| 454 | )?; |
| 455 | |
| 456 | let thread = manager |
| 457 | .create_thread(CreateThreadRequest { |
| 458 | model: Some("local-code-model".to_string()), |
| 459 | model_provider: Some("lm-studio".to_string()), |
| 460 | ..CreateThreadRequest::default() |
| 461 | }) |
| 462 | .await?; |
| 463 | let persisted = manager.get_thread(&thread.id).await?; |
| 464 | assert_eq!(persisted.model_provider.as_deref(), Some("custom")); |
| 465 | assert_eq!(persisted.model_provider_id.as_deref(), Some("lm-studio")); |
| 466 | let serialized = serde_json::to_string(&persisted)?; |
| 467 | assert!(serialized.contains("\"model_provider\":\"custom\"")); |
| 468 | assert!(serialized.contains("\"model_provider_id\":\"lm-studio\"")); |
| 469 | assert!(!serialized.contains("127.0.0.1:1234")); |
| 470 | |
| 471 | let route = manager.resolved_route_for_thread(&config, &persisted)?; |
| 472 | assert_eq!(route.identity.provider, ApiProvider::Custom); |
| 473 | assert_eq!(route.identity.key, "lm-studio"); |
| 474 | assert_eq!(route.model, "local-code-model"); |
| 475 | assert_eq!(route.config.deepseek_base_url(), "http://127.0.0.1:1234/v1"); |
| 476 | |
| 477 | let err = manager |
| 478 | .resolved_route_for_thread(&Config::default(), &persisted) |
| 479 | .expect_err("removed provider must fail closed"); |
| 480 | let message = err.to_string(); |
| 481 | assert!(message.contains("[providers.lm-studio]"), "{message}"); |
| 482 | assert!(message.contains("will not fall back"), "{message}"); |
| 483 | |
| 484 | let mut legacy_value = serde_json::to_value(&persisted)?; |
| 485 | legacy_value |
| 486 | .as_object_mut() |
| 487 | .expect("thread object") |
| 488 | .remove("model_provider"); |
| 489 | legacy_value |
| 490 | .as_object_mut() |
| 491 | .expect("thread object") |
| 492 | .remove("model_provider_id"); |
| 493 | let legacy: ThreadRecord = serde_json::from_value(legacy_value)?; |
| 494 | assert_eq!(legacy.model_provider, None); |
| 495 | Ok(()) |
| 496 | } |
| 497 | |
| 498 | #[test] |
| 499 | fn legacy_literal_custom_thread_resume_requires_and_keeps_root_route() -> Result<()> { |
| 500 | let config = Config { |
| 501 | provider: Some("custom".to_string()), |
| 502 | base_url: Some("http://127.0.0.1:18180/v1".to_string()), |
| 503 | default_text_model: Some("legacy-default-model".to_string()), |
| 504 | ..Config::default() |
| 505 | }; |
| 506 | let manager = RuntimeThreadManager::open( |
| 507 | config.clone(), |
| 508 | PathBuf::from("."), |
| 509 | test_manager_config(test_runtime_dir()), |
| 510 | )?; |
| 511 | let mut persisted = sample_thread("thr_legacy_custom"); |
| 512 | persisted.model = "legacy-saved-model".to_string(); |
| 513 | persisted.model_provider = Some("custom".to_string()); |
| 514 | let restored: ThreadRecord = serde_json::from_str(&serde_json::to_string(&persisted)?)?; |
| 515 | |
| 516 | let route = manager.resolved_route_for_thread(&config, &restored)?; |
| 517 | assert_eq!(route.identity.provider, ApiProvider::Custom); |
| 518 | assert_eq!(route.identity.key, "custom"); |
| 519 | assert_eq!(route.model, "legacy-saved-model"); |
| 520 | assert_eq!( |
| 521 | route.config.deepseek_base_url(), |
| 522 | "http://127.0.0.1:18180/v1" |
| 523 | ); |
| 524 | assert!( |
| 525 | route |
| 526 | .config |
| 527 | .providers |
| 528 | .as_ref() |
| 529 | .is_none_or(|providers| !providers.custom.contains_key("custom")), |
| 530 | "route resolution must not synthesize an ambiguous [providers.custom] table" |
| 531 | ); |
| 532 | assert_eq!( |
| 533 | route |
| 534 | .config |
| 535 | .resolve_provider_identity("custom") |
| 536 | .map_err(anyhow::Error::msg)?, |
| 537 | crate::config::ProviderIdentity { |
| 538 | provider: ApiProvider::Custom, |
| 539 | key: "custom".to_string(), |
| 540 | exact_id: None, |
| 541 | } |
| 542 | ); |
| 543 | let repeated = manager.resolved_route_for_thread(&route.config, &restored)?; |
| 544 | assert_eq!(repeated.identity.key, "custom"); |
| 545 | assert_eq!(repeated.model, "legacy-saved-model"); |
| 546 | assert_eq!( |
| 547 | repeated.config.deepseek_base_url(), |
| 548 | "http://127.0.0.1:18180/v1" |
| 549 | ); |
| 550 | |
| 551 | let named_config = { |
| 552 | let mut custom = std::collections::HashMap::new(); |
| 553 | custom.insert( |
| 554 | "lm-studio".to_string(), |
| 555 | crate::config::ProviderConfig { |
| 556 | kind: Some("openai-compatible".to_string()), |
| 557 | base_url: Some("http://127.0.0.1:18181/v1".to_string()), |
| 558 | model: Some("named-model".to_string()), |
| 559 | ..crate::config::ProviderConfig::default() |
| 560 | }, |
| 561 | ); |
| 562 | Config { |
| 563 | provider: Some("lm-studio".to_string()), |
| 564 | providers: Some(crate::config::ProvidersConfig { |
| 565 | custom, |
| 566 | ..crate::config::ProvidersConfig::default() |
| 567 | }), |
| 568 | ..Config::default() |
| 569 | } |
| 570 | }; |
| 571 | let error = manager |
| 572 | .resolved_route_for_thread(&named_config, &restored) |
| 573 | .expect_err("id-less root record must not migrate to a named table") |
| 574 | .to_string(); |
| 575 | assert!(error.contains("root-level"), "{error}"); |
| 576 | assert!(error.contains("will not guess or fall back"), "{error}"); |
| 577 | |
| 578 | Ok(()) |
| 579 | } |
| 580 | |
| 581 | #[tokio::test] |
| 582 | async fn root_custom_thread_and_turn_writers_omit_exact_id() -> Result<()> { |
| 583 | let config = Config { |
| 584 | provider: Some("custom".to_string()), |
| 585 | base_url: Some("http://127.0.0.1:18180/v1".to_string()), |
| 586 | default_text_model: Some("legacy-root-model".to_string()), |
| 587 | ..Config::default() |
| 588 | }; |
| 589 | let manager = RuntimeThreadManager::open( |
| 590 | config, |
| 591 | PathBuf::from("."), |
| 592 | test_manager_config(test_runtime_dir()), |
| 593 | )?; |
| 594 | let thread = manager |
| 595 | .create_thread(CreateThreadRequest { |
| 596 | model: Some("legacy-root-model".to_string()), |
| 597 | ..CreateThreadRequest::default() |
| 598 | }) |
| 599 | .await?; |
| 600 | assert_eq!(thread.model_provider.as_deref(), Some("custom")); |
| 601 | assert_eq!(thread.model_provider_id, None); |
| 602 | assert!(!serde_json::to_string(&thread)?.contains("model_provider_id")); |
| 603 | |
| 604 | let mut harness = install_mock_engine(&manager, &thread.id).await; |
| 605 | let turn = manager |
| 606 | .start_turn( |
| 607 | &thread.id, |
| 608 | StartTurnRequest { |
| 609 | prompt: "keep the root route".to_string(), |
| 610 | ..StartTurnRequest::default() |
| 611 | }, |
| 612 | ) |
| 613 | .await?; |
| 614 | assert_eq!(turn.effective_provider.as_deref(), Some("custom")); |
| 615 | assert_eq!(turn.effective_provider_id, None); |
| 616 | assert!(!serde_json::to_string(&turn)?.contains("effective_provider_id")); |
| 617 | match harness.rx_op.recv().await { |
| 618 | Some(Op::SendMessage { route, .. }) => { |
| 619 | assert_eq!(route.identity.key, "custom"); |
| 620 | assert_eq!(route.identity.exact_id, None); |
| 621 | assert_eq!( |
| 622 | route.config.deepseek_base_url(), |
| 623 | "http://127.0.0.1:18180/v1" |
| 624 | ); |
| 625 | } |
| 626 | other => panic!("expected root custom send, got {other:?}"), |
| 627 | } |
| 628 | Ok(()) |
| 629 | } |
| 630 | |
| 631 | #[tokio::test] |
| 632 | async fn real_turn_client_preflight_failure_writes_no_in_progress_record() -> Result<()> { |
| 633 | let mut custom = std::collections::HashMap::new(); |
| 634 | custom.insert( |
| 635 | "preflight-failure".to_string(), |
| 636 | crate::config::ProviderConfig { |
| 637 | kind: Some("openai-compatible".to_string()), |
| 638 | base_url: Some("https://preflight.invalid/v1".to_string()), |
| 639 | model: Some("preflight-model".to_string()), |
| 640 | api_key: Some("test-key".to_string()), |
| 641 | // Client construction rejects this independently of ambient auth, |
| 642 | // keeping the async regression hermetic without a global env lock. |
| 643 | insecure_skip_tls_verify: Some(true), |
| 644 | ..crate::config::ProviderConfig::default() |
| 645 | }, |
| 646 | ); |
| 647 | let manager = RuntimeThreadManager::open( |
| 648 | Config { |
| 649 | provider: Some("preflight-failure".to_string()), |
| 650 | providers: Some(crate::config::ProvidersConfig { |
| 651 | custom, |
| 652 | ..crate::config::ProvidersConfig::default() |
| 653 | }), |
| 654 | ..Config::default() |
| 655 | }, |
| 656 | PathBuf::from("."), |
| 657 | test_manager_config(test_runtime_dir()), |
| 658 | )?; |
| 659 | let thread = manager |
| 660 | .create_thread(CreateThreadRequest::default()) |
| 661 | .await?; |
| 662 | |
| 663 | let error = manager |
| 664 | .start_turn( |
| 665 | &thread.id, |
| 666 | StartTurnRequest { |
| 667 | prompt: "must not become a zombie turn".to_string(), |
| 668 | ..StartTurnRequest::default() |
| 669 | }, |
| 670 | ) |
| 671 | .await |
| 672 | .expect_err("missing credentials must fail before turn persistence") |
| 673 | .to_string(); |
| 674 | |
| 675 | assert!( |
| 676 | error.contains("TLS certificate verification cannot be disabled"), |
| 677 | "{error}" |
| 678 | ); |
| 679 | assert!(manager.store.list_turns_for_thread(&thread.id)?.is_empty()); |
| 680 | assert_eq!(manager.get_thread(&thread.id).await?.latest_turn_id, None); |
| 681 | Ok(()) |
| 682 | } |
| 683 | |
| 684 | #[tokio::test] |
| 685 | async fn closed_turn_mailbox_rolls_back_durable_records_and_active_claim() -> Result<()> { |
| 686 | let manager = test_manager(test_runtime_dir())?; |
| 687 | let thread = manager |
| 688 | .create_thread(CreateThreadRequest::default()) |
| 689 | .await?; |
| 690 | let harness = install_mock_engine(&manager, &thread.id).await; |
| 691 | let before_active = { |
| 692 | let active = manager.active.lock().await; |
| 693 | let state = active.engines.get(&thread.id).expect("installed engine"); |
| 694 | ( |
| 695 | state.active_turn.as_ref().map(|turn| turn.turn_id.clone()), |
| 696 | state.route_identity.clone(), |
| 697 | state.route_model.clone(), |
| 698 | active.lru.clone(), |
| 699 | ) |
| 700 | }; |
| 701 | let before_thread = serde_json::to_value(manager.get_thread(&thread.id).await?)?; |
| 702 | let before_events = serde_json::to_value(manager.events_since(&thread.id, None)?)?; |
| 703 | drop(harness.rx_op); |
| 704 | |
| 705 | let error = manager |
| 706 | .start_turn( |
| 707 | &thread.id, |
| 708 | StartTurnRequest { |
| 709 | prompt: "mailbox is already closed".to_string(), |
| 710 | ..StartTurnRequest::default() |
| 711 | }, |
| 712 | ) |
| 713 | .await |
| 714 | .expect_err("closed mailbox must reject the turn") |
| 715 | .to_string(); |
| 716 | assert!(error.contains("Failed to start turn"), "{error}"); |
| 717 | |
| 718 | assert!(manager.store.list_turns_for_thread(&thread.id)?.is_empty()); |
| 719 | assert_eq!( |
| 720 | serde_json::to_value(manager.get_thread(&thread.id).await?)?, |
| 721 | before_thread |
| 722 | ); |
| 723 | assert_eq!( |
| 724 | serde_json::to_value(manager.events_since(&thread.id, None)?)?, |
| 725 | before_events |
| 726 | ); |
| 727 | assert_eq!( |
| 728 | std::fs::read_dir(&manager.store.items_dir)?.count(), |
| 729 | 0, |
| 730 | "failed send must remove the optimistic user item" |
| 731 | ); |
| 732 | let after_active = { |
| 733 | let active = manager.active.lock().await; |
| 734 | let state = active.engines.get(&thread.id).expect("installed engine"); |
| 735 | ( |
| 736 | state.active_turn.as_ref().map(|turn| turn.turn_id.clone()), |
| 737 | state.route_identity.clone(), |
| 738 | state.route_model.clone(), |
| 739 | active.lru.clone(), |
| 740 | ) |
| 741 | }; |
| 742 | assert_eq!(after_active, before_active); |
| 743 | Ok(()) |
| 744 | } |
| 745 | |
| 746 | #[tokio::test] |
| 747 | async fn cancellation_while_waiting_for_mailbox_capacity_claims_nothing() -> Result<()> { |
| 748 | let manager = test_manager(test_runtime_dir())?; |
| 749 | let thread = manager |
| 750 | .create_thread(CreateThreadRequest::default()) |
| 751 | .await?; |
| 752 | let mut harness = install_mock_engine(&manager, &thread.id).await; |
| 753 | |
| 754 | for _ in 0..32 { |
| 755 | harness.handle.try_send(Op::ListSubAgents)?; |
| 756 | } |
| 757 | let start_sender_count = harness.handle.tx_op.strong_count(); |
| 758 | let start_manager = manager.clone(); |
| 759 | let start_thread_id = thread.id.clone(); |
| 760 | let start_task = tokio::spawn(async move { |
| 761 | start_manager |
| 762 | .start_turn( |
| 763 | &start_thread_id, |
| 764 | StartTurnRequest { |
| 765 | prompt: "cancel before mailbox capacity".to_string(), |
| 766 | ..StartTurnRequest::default() |
| 767 | }, |
| 768 | ) |
| 769 | .await |
| 770 | }); |
| 771 | wait_for_sender_strong_count(&harness.handle.tx_op, start_sender_count + 2).await?; |
| 772 | assert!( |
| 773 | !start_task.is_finished(), |
| 774 | "start should be waiting for capacity" |
| 775 | ); |
| 776 | assert!(manager.store.list_turns_for_thread(&thread.id)?.is_empty()); |
| 777 | assert_eq!(manager.get_thread(&thread.id).await?.latest_turn_id, None); |
| 778 | assert_eq!(manager.active_turn_flags(&thread.id, "missing").await, None); |
| 779 | start_task.abort(); |
| 780 | let _ = start_task.await; |
| 781 | for _ in 0..32 { |
| 782 | assert!(matches!( |
| 783 | harness.rx_op.recv().await, |
| 784 | Some(Op::ListSubAgents) |
| 785 | )); |
| 786 | } |
| 787 | |
| 788 | for _ in 0..32 { |
| 789 | harness.handle.try_send(Op::ListSubAgents)?; |
| 790 | } |
| 791 | let compact_sender_count = harness.handle.tx_op.strong_count(); |
| 792 | let compact_manager = manager.clone(); |
| 793 | let compact_thread_id = thread.id.clone(); |
| 794 | let compact_task = tokio::spawn(async move { |
| 795 | compact_manager |
| 796 | .compact_thread(&compact_thread_id, CompactThreadRequest::default()) |
| 797 | .await |
| 798 | }); |
| 799 | wait_for_sender_strong_count(&harness.handle.tx_op, compact_sender_count + 2).await?; |
| 800 | assert!( |
| 801 | !compact_task.is_finished(), |
| 802 | "compaction should be waiting for capacity" |
| 803 | ); |
| 804 | assert!(manager.store.list_turns_for_thread(&thread.id)?.is_empty()); |
| 805 | assert_eq!(manager.get_thread(&thread.id).await?.latest_turn_id, None); |
| 806 | compact_task.abort(); |
| 807 | let _ = compact_task.await; |
| 808 | for _ in 0..32 { |
| 809 | assert!(matches!( |
| 810 | harness.rx_op.recv().await, |
| 811 | Some(Op::ListSubAgents) |
| 812 | )); |
| 813 | } |
| 814 | Ok(()) |
| 815 | } |
| 816 | |
| 817 | #[tokio::test] |
| 818 | async fn caller_cancellation_after_engine_acceptance_keeps_owned_turn_lifecycle() -> Result<()> { |
| 819 | let manager = test_manager(test_runtime_dir())?; |
| 820 | let thread = manager |
| 821 | .create_thread(CreateThreadRequest::default()) |
| 822 | .await?; |
| 823 | let mut harness = install_mock_engine(&manager, &thread.id).await; |
| 824 | |
| 825 | // Block the first start-event append so the public future remains |
| 826 | // cancellable after the operation has entered the engine mailbox. |
| 827 | let mut event_lock = fd_lock::RwLock::new(manager.store.open_event_lock()?); |
| 828 | let event_state_guard = event_lock.try_write()?; |
| 829 | let start_manager = manager.clone(); |
| 830 | let thread_id = thread.id.clone(); |
| 831 | let start_task = tokio::spawn(async move { |
| 832 | start_manager |
| 833 | .start_turn( |
| 834 | &thread_id, |
| 835 | StartTurnRequest { |
| 836 | prompt: "the lifecycle outlives its caller".to_string(), |
| 837 | ..StartTurnRequest::default() |
| 838 | }, |
| 839 | ) |
| 840 | .await |
| 841 | }); |
| 842 | assert!(matches!( |
| 843 | tokio::time::timeout(Duration::from_secs(2), harness.rx_op.recv()).await?, |
| 844 | Some(Op::SendMessage { .. }) |
| 845 | )); |
| 846 | let turns = manager.store.list_turns_for_thread(&thread.id)?; |
| 847 | assert_eq!(turns.len(), 1); |
| 848 | let turn_id = turns[0].id.clone(); |
| 849 | assert_eq!(turns[0].status, RuntimeTurnStatus::InProgress); |
| 850 | assert_eq!(turns[0].item_ids.len(), 1); |
| 851 | assert_eq!( |
| 852 | manager.store.load_item(&turns[0].item_ids[0])?.turn_id, |
| 853 | turn_id |
| 854 | ); |
| 855 | assert!( |
| 856 | manager |
| 857 | .active_turn_flags(&thread.id, &turn_id) |
| 858 | .await |
| 859 | .is_some() |
| 860 | ); |
| 861 | |
| 862 | start_task.abort(); |
| 863 | let _ = start_task.await; |
| 864 | drop(event_state_guard); |
| 865 | |
| 866 | harness |
| 867 | .tx_event |
| 868 | .send(EngineEvent::MessageStarted { index: 0 }) |
| 869 | .await?; |
| 870 | harness |
| 871 | .tx_event |
| 872 | .send(EngineEvent::MessageDelta { |
| 873 | index: 0, |
| 874 | content: "owned monitor is live".to_string(), |
| 875 | }) |
| 876 | .await?; |
| 877 | harness |
| 878 | .tx_event |
| 879 | .send(EngineEvent::MessageComplete { index: 0 }) |
| 880 | .await?; |
| 881 | harness |
| 882 | .tx_event |
| 883 | .send(EngineEvent::TurnComplete { |
| 884 | usage: Usage::default(), |
| 885 | status: TurnOutcomeStatus::Completed, |
| 886 | error: None, |
| 887 | tool_catalog: None, |
| 888 | base_url: None, |
| 889 | }) |
| 890 | .await?; |
| 891 | let terminal = wait_for_terminal_turn(&manager, &turn_id, Duration::from_secs(2)).await?; |
| 892 | assert_eq!(terminal.status, RuntimeTurnStatus::Completed); |
| 893 | assert_eq!(manager.active_turn_flags(&thread.id, &turn_id).await, None); |
| 894 | |
| 895 | let lifecycle: Vec<String> = manager |
| 896 | .events_since(&thread.id, None)? |
| 897 | .iter() |
| 898 | .filter(|event| event.turn_id.as_deref() == Some(turn_id.as_str())) |
| 899 | .map(|event| event.event.clone()) |
| 900 | .collect(); |
| 901 | assert_eq!( |
| 902 | &lifecycle[..3], |
| 903 | &["turn.started", "item.started", "item.completed"] |
| 904 | ); |
| 905 | assert_eq!(lifecycle.last().map(String::as_str), Some("turn.completed")); |
| 906 | Ok(()) |
| 907 | } |
| 908 | |
| 909 | #[tokio::test] |
| 910 | async fn thread_updates_while_start_waits_for_capacity_survive_latest_turn_write() -> Result<()> { |
| 911 | let manager = test_manager(test_runtime_dir())?; |
| 912 | let thread = manager |
| 913 | .create_thread(CreateThreadRequest::default()) |
| 914 | .await?; |
| 915 | let mut harness = install_mock_engine(&manager, &thread.id).await; |
| 916 | for _ in 0..32 { |
| 917 | harness.handle.try_send(Op::ListSubAgents)?; |
| 918 | } |
| 919 | let sender_count = harness.handle.tx_op.strong_count(); |
| 920 | |
| 921 | let start_manager = manager.clone(); |
| 922 | let thread_id = thread.id.clone(); |
| 923 | let start_task = tokio::spawn(async move { |
| 924 | start_manager |
| 925 | .start_turn( |
| 926 | &thread_id, |
| 927 | StartTurnRequest { |
| 928 | prompt: "preserve concurrent metadata".to_string(), |
| 929 | ..StartTurnRequest::default() |
| 930 | }, |
| 931 | ) |
| 932 | .await |
| 933 | }); |
| 934 | wait_for_sender_strong_count(&harness.handle.tx_op, sender_count + 2).await?; |
| 935 | assert!(!start_task.is_finished()); |
| 936 | |
| 937 | manager |
| 938 | .update_thread( |
| 939 | &thread.id, |
| 940 | UpdateThreadRequest { |
| 941 | title: Some("new title while queued".to_string()), |
| 942 | ..UpdateThreadRequest::default() |
| 943 | }, |
| 944 | ) |
| 945 | .await?; |
| 946 | assert!(matches!( |
| 947 | harness.rx_op.recv().await, |
| 948 | Some(Op::ListSubAgents) |
| 949 | )); |
| 950 | let turn = tokio::time::timeout(Duration::from_secs(2), start_task).await???; |
| 951 | let mut saw_send = false; |
| 952 | for _ in 0..32 { |
| 953 | if matches!(harness.rx_op.recv().await, Some(Op::SendMessage { .. })) { |
| 954 | saw_send = true; |
| 955 | break; |
| 956 | } |
| 957 | } |
| 958 | assert!( |
| 959 | saw_send, |
| 960 | "accepted send must remain behind refresh operations" |
| 961 | ); |
| 962 | |
| 963 | harness |
| 964 | .tx_event |
| 965 | .send(EngineEvent::MessageStarted { index: 0 }) |
| 966 | .await?; |
| 967 | harness |
| 968 | .tx_event |
| 969 | .send(EngineEvent::MessageDelta { |
| 970 | index: 0, |
| 971 | content: "metadata retained".to_string(), |
| 972 | }) |
| 973 | .await?; |
| 974 | harness |
| 975 | .tx_event |
| 976 | .send(EngineEvent::MessageComplete { index: 0 }) |
| 977 | .await?; |
| 978 | harness |
| 979 | .tx_event |
| 980 | .send(EngineEvent::TurnComplete { |
| 981 | usage: Usage::default(), |
| 982 | status: TurnOutcomeStatus::Completed, |
| 983 | error: None, |
| 984 | tool_catalog: None, |
| 985 | base_url: None, |
| 986 | }) |
| 987 | .await?; |
| 988 | let terminal = wait_for_terminal_turn(&manager, &turn.id, Duration::from_secs(2)).await?; |
| 989 | assert_eq!(terminal.status, RuntimeTurnStatus::Completed); |
| 990 | assert_eq!(turn.item_ids.len(), 1); |
| 991 | assert!( |
| 992 | terminal.item_ids.contains(&turn.item_ids[0]), |
| 993 | "the accepted user item must survive later assistant-item writes" |
| 994 | ); |
| 995 | assert_eq!( |
| 996 | manager.store.load_turn(&turn.id)?.item_ids, |
| 997 | terminal.item_ids |
| 998 | ); |
| 999 | let updated = manager.get_thread(&thread.id).await?; |
| 1000 | assert_eq!(updated.title.as_deref(), Some("new title while queued")); |
| 1001 | assert_eq!(updated.latest_turn_id.as_deref(), Some(turn.id.as_str())); |
| 1002 | Ok(()) |
| 1003 | } |
| 1004 | |
| 1005 | #[tokio::test] |
| 1006 | async fn execution_update_while_start_waits_rejects_stale_operation() -> Result<()> { |
| 1007 | let manager = test_manager(test_runtime_dir())?; |
| 1008 | let thread = manager |
| 1009 | .create_thread(CreateThreadRequest::default()) |
| 1010 | .await?; |
| 1011 | let mut harness = install_mock_engine(&manager, &thread.id).await; |
| 1012 | for _ in 0..32 { |
| 1013 | harness.handle.try_send(Op::ListSubAgents)?; |
| 1014 | } |
| 1015 | let sender_count = harness.handle.tx_op.strong_count(); |
| 1016 | let start_manager = manager.clone(); |
| 1017 | let thread_id = thread.id.clone(); |
| 1018 | let start_task = tokio::spawn(async move { |
| 1019 | start_manager |
| 1020 | .start_turn( |
| 1021 | &thread_id, |
| 1022 | StartTurnRequest { |
| 1023 | prompt: "must not use stale mode".to_string(), |
| 1024 | ..StartTurnRequest::default() |
| 1025 | }, |
| 1026 | ) |
| 1027 | .await |
| 1028 | }); |
| 1029 | wait_for_sender_strong_count(&harness.handle.tx_op, sender_count + 2).await?; |
| 1030 | |
| 1031 | manager |
| 1032 | .update_thread( |
| 1033 | &thread.id, |
| 1034 | UpdateThreadRequest { |
| 1035 | mode: Some(AppMode::Plan.as_setting().to_string()), |
| 1036 | ..UpdateThreadRequest::default() |
| 1037 | }, |
| 1038 | ) |
| 1039 | .await?; |
| 1040 | assert!(matches!( |
| 1041 | harness.rx_op.recv().await, |
| 1042 | Some(Op::ListSubAgents) |
| 1043 | )); |
| 1044 | let error = tokio::time::timeout(Duration::from_secs(2), start_task) |
| 1045 | .await?? |
| 1046 | .expect_err("stale operation must fail") |
| 1047 | .to_string(); |
| 1048 | assert!(error.contains("execution settings changed"), "{error}"); |
| 1049 | for _ in 0..31 { |
| 1050 | assert!(matches!( |
| 1051 | harness.rx_op.recv().await, |
| 1052 | Some(Op::ListSubAgents) |
| 1053 | )); |
| 1054 | } |
| 1055 | assert!(harness.rx_op.try_recv().is_err()); |
| 1056 | assert!(manager.store.list_turns_for_thread(&thread.id)?.is_empty()); |
| 1057 | let updated = manager.get_thread(&thread.id).await?; |
| 1058 | assert_eq!(updated.mode, AppMode::Plan.as_setting()); |
| 1059 | assert_eq!(updated.latest_turn_id, None); |
| 1060 | assert_eq!(manager.active_turn_flags(&thread.id, "missing").await, None); |
| 1061 | Ok(()) |
| 1062 | } |
| 1063 | |
| 1064 | #[tokio::test] |
| 1065 | async fn compact_lifecycle_outlives_caller_and_preserves_concurrent_thread_updates() -> Result<()> { |
| 1066 | let manager = test_manager(test_runtime_dir())?; |
| 1067 | let thread = manager |
| 1068 | .create_thread(CreateThreadRequest::default()) |
| 1069 | .await?; |
| 1070 | let mut harness = install_mock_engine(&manager, &thread.id).await; |
| 1071 | for _ in 0..32 { |
| 1072 | harness.handle.try_send(Op::ListSubAgents)?; |
| 1073 | } |
| 1074 | let sender_count = harness.handle.tx_op.strong_count(); |
| 1075 | |
| 1076 | let compact_manager = manager.clone(); |
| 1077 | let thread_id = thread.id.clone(); |
| 1078 | let compact_task = tokio::spawn(async move { |
| 1079 | compact_manager |
| 1080 | .compact_thread(&thread_id, CompactThreadRequest::default()) |
| 1081 | .await |
| 1082 | }); |
| 1083 | wait_for_sender_strong_count(&harness.handle.tx_op, sender_count + 2).await?; |
| 1084 | assert!(!compact_task.is_finished()); |
| 1085 | manager |
| 1086 | .update_thread( |
| 1087 | &thread.id, |
| 1088 | UpdateThreadRequest { |
| 1089 | title: Some("title before compact claim".to_string()), |
| 1090 | ..UpdateThreadRequest::default() |
| 1091 | }, |
| 1092 | ) |
| 1093 | .await?; |
| 1094 | // Once capacity is released, block the acknowledgement events so the |
| 1095 | // API future can be dropped after the engine accepted the operation. |
| 1096 | let mut event_lock = fd_lock::RwLock::new(manager.store.open_event_lock()?); |
| 1097 | let event_state_guard = event_lock.try_write()?; |
| 1098 | let mut saw_compact = false; |
| 1099 | for _ in 0..33 { |
| 1100 | if matches!( |
| 1101 | tokio::time::timeout(Duration::from_secs(2), harness.rx_op.recv()).await?, |
| 1102 | Some(Op::CompactContext { .. }) |
| 1103 | ) { |
| 1104 | saw_compact = true; |
| 1105 | break; |
| 1106 | } |
| 1107 | } |
| 1108 | assert!( |
| 1109 | saw_compact, |
| 1110 | "manual compaction must enter the engine mailbox" |
| 1111 | ); |
| 1112 | let turns = manager.store.list_turns_for_thread(&thread.id)?; |
| 1113 | assert_eq!(turns.len(), 1); |
| 1114 | let turn_id = turns[0].id.clone(); |
| 1115 | compact_task.abort(); |
| 1116 | let _ = compact_task.await; |
| 1117 | drop(event_state_guard); |
| 1118 | |
| 1119 | harness |
| 1120 | .tx_event |
| 1121 | .send(EngineEvent::CompactionStarted { |
| 1122 | id: "manual_owned".to_string(), |
| 1123 | auto: false, |
| 1124 | message: "compaction started".to_string(), |
| 1125 | }) |
| 1126 | .await?; |
| 1127 | harness |
| 1128 | .tx_event |
| 1129 | .send(EngineEvent::CompactionCompleted { |
| 1130 | id: "manual_owned".to_string(), |
| 1131 | auto: false, |
| 1132 | message: "compaction completed".to_string(), |
| 1133 | messages_before: Some(4), |
| 1134 | messages_after: Some(2), |
| 1135 | summary_prompt: None, |
| 1136 | }) |
| 1137 | .await?; |
| 1138 | harness |
| 1139 | .tx_event |
| 1140 | .send(EngineEvent::TurnComplete { |
| 1141 | usage: Usage::default(), |
| 1142 | status: TurnOutcomeStatus::Completed, |
| 1143 | error: None, |
| 1144 | tool_catalog: None, |
| 1145 | base_url: None, |
| 1146 | }) |
| 1147 | .await?; |
| 1148 | let terminal = wait_for_terminal_turn(&manager, &turn_id, Duration::from_secs(2)).await?; |
| 1149 | assert_eq!(terminal.status, RuntimeTurnStatus::Completed); |
| 1150 | assert_eq!(manager.active_turn_flags(&thread.id, &turn_id).await, None); |
| 1151 | let updated = manager.get_thread(&thread.id).await?; |
| 1152 | assert_eq!(updated.title.as_deref(), Some("title before compact claim")); |
| 1153 | assert_eq!(updated.latest_turn_id.as_deref(), Some(turn_id.as_str())); |
| 1154 | Ok(()) |
| 1155 | } |
| 1156 | |
| 1157 | #[tokio::test] |
| 1158 | async fn concurrent_turn_starts_leave_one_claim_and_one_consistent_durable_turn() -> Result<()> { |
| 1159 | let manager = test_manager(test_runtime_dir())?; |
| 1160 | let thread = manager |
| 1161 | .create_thread(CreateThreadRequest::default()) |
| 1162 | .await?; |
| 1163 | let mut harness = install_mock_engine(&manager, &thread.id).await; |
| 1164 | let first = manager.start_turn( |
| 1165 | &thread.id, |
| 1166 | StartTurnRequest { |
| 1167 | prompt: "first concurrent turn".to_string(), |
| 1168 | ..StartTurnRequest::default() |
| 1169 | }, |
| 1170 | ); |
| 1171 | let second = manager.start_turn( |
| 1172 | &thread.id, |
| 1173 | StartTurnRequest { |
| 1174 | prompt: "second concurrent turn".to_string(), |
| 1175 | ..StartTurnRequest::default() |
| 1176 | }, |
| 1177 | ); |
| 1178 | |
| 1179 | let (first, second) = tokio::join!(first, second); |
| 1180 | let (turn, rejection) = match (first, second) { |
| 1181 | (Ok(turn), Err(error)) | (Err(error), Ok(turn)) => (turn, error), |
| 1182 | (first, second) => { |
| 1183 | panic!("expected one accepted turn and one rejection: {first:?} {second:?}") |
| 1184 | } |
| 1185 | }; |
| 1186 | assert!( |
| 1187 | rejection.to_string().contains("already has an active turn"), |
| 1188 | "{rejection}" |
| 1189 | ); |
| 1190 | let turns = manager.store.list_turns_for_thread(&thread.id)?; |
| 1191 | assert_eq!(turns.len(), 1); |
| 1192 | assert_eq!(turns[0].id, turn.id); |
| 1193 | assert_eq!( |
| 1194 | manager.get_thread(&thread.id).await?.latest_turn_id, |
| 1195 | Some(turn.id.clone()) |
| 1196 | ); |
| 1197 | assert_eq!( |
| 1198 | manager.active_turn_flags(&thread.id, &turn.id).await, |
| 1199 | Some((false, false)) |
| 1200 | ); |
| 1201 | match harness.rx_op.recv().await { |
| 1202 | Some(Op::SendMessage { compaction, .. }) => { |
| 1203 | assert_eq!( |
| 1204 | compaction.runtime_cost_owner.as_deref(), |
| 1205 | Some(turn.id.as_str()) |
| 1206 | ); |
| 1207 | } |
| 1208 | other => panic!("expected SendMessage op, got {other:?}"), |
| 1209 | } |
| 1210 | assert!(harness.rx_op.try_recv().is_err()); |
| 1211 | Ok(()) |
| 1212 | } |
| 1213 | |
| 1214 | #[test] |
| 1215 | fn legacy_custom_thread_stays_on_root_when_literal_table_coexists() -> Result<()> { |
| 1216 | let mut custom = std::collections::HashMap::new(); |
| 1217 | custom.insert( |
| 1218 | "custom".to_string(), |
| 1219 | crate::config::ProviderConfig { |
| 1220 | kind: Some("openai-compatible".to_string()), |
| 1221 | base_url: Some("http://127.0.0.1:18182/v1".to_string()), |
| 1222 | model: Some("table-model".to_string()), |
| 1223 | ..crate::config::ProviderConfig::default() |
| 1224 | }, |
| 1225 | ); |
| 1226 | let config = Config { |
| 1227 | provider: Some("custom".to_string()), |
| 1228 | base_url: Some("http://127.0.0.1:18181/v1".to_string()), |
| 1229 | default_text_model: Some("legacy-root-model".to_string()), |
| 1230 | providers: Some(crate::config::ProvidersConfig { |
| 1231 | custom, |
| 1232 | ..crate::config::ProvidersConfig::default() |
| 1233 | }), |
| 1234 | ..Config::default() |
| 1235 | }; |
| 1236 | let manager = RuntimeThreadManager::open( |
| 1237 | config.clone(), |
| 1238 | PathBuf::from("."), |
| 1239 | test_manager_config(test_runtime_dir()), |
| 1240 | )?; |
| 1241 | let mut legacy = sample_thread("thr_ambiguous_legacy_custom"); |
| 1242 | legacy.model = "legacy-saved-model".to_string(); |
| 1243 | legacy.model_provider = Some("custom".to_string()); |
| 1244 | legacy.model_provider_id = None; |
| 1245 | |
| 1246 | let root = manager.resolved_route_for_thread(&config, &legacy)?; |
| 1247 | assert_eq!(root.identity.provider, ApiProvider::Custom); |
| 1248 | assert_eq!(root.identity.key, "custom"); |
| 1249 | assert_eq!(root.identity.exact_id, None); |
| 1250 | assert_eq!(root.config.deepseek_base_url(), "http://127.0.0.1:18181/v1"); |
| 1251 | |
| 1252 | legacy.model_provider_id = Some("custom".to_string()); |
| 1253 | let exact = manager.resolved_route_for_thread(&config, &legacy)?; |
| 1254 | assert_eq!(exact.identity.provider, ApiProvider::Custom); |
| 1255 | assert_eq!(exact.identity.key, "custom"); |
| 1256 | assert_eq!(exact.identity.exact_id.as_deref(), Some("custom")); |
| 1257 | assert_eq!( |
| 1258 | exact.config.deepseek_base_url(), |
| 1259 | "http://127.0.0.1:18182/v1" |
| 1260 | ); |
| 1261 | let root_only = Config { |
| 1262 | provider: Some("custom".to_string()), |
| 1263 | base_url: Some("http://127.0.0.1:18181/v1".to_string()), |
| 1264 | default_text_model: Some("legacy-root-model".to_string()), |
| 1265 | ..Config::default() |
| 1266 | }; |
| 1267 | let error = manager |
| 1268 | .resolved_route_for_thread(&root_only, &legacy) |
| 1269 | .expect_err("exact literal table thread must not fall back to root") |
| 1270 | .to_string(); |
| 1271 | assert!(error.contains("[providers.custom]"), "{error}"); |
| 1272 | assert!(error.contains("will not fall back"), "{error}"); |
| 1273 | Ok(()) |
| 1274 | } |
| 1275 | |
| 1276 | #[tokio::test] |
| 1277 | async fn empty_imported_custom_id_fails_closed_when_root_and_table_coexist() -> Result<()> { |
| 1278 | let mut custom = std::collections::HashMap::new(); |
| 1279 | custom.insert( |
| 1280 | "custom".to_string(), |
| 1281 | crate::config::ProviderConfig { |
| 1282 | kind: Some("openai-compatible".to_string()), |
| 1283 | base_url: Some("http://127.0.0.1:18182/v1".to_string()), |
| 1284 | model: Some("table-model".to_string()), |
| 1285 | ..crate::config::ProviderConfig::default() |
| 1286 | }, |
| 1287 | ); |
| 1288 | let config = Config { |
| 1289 | provider: Some("custom".to_string()), |
| 1290 | base_url: Some("http://127.0.0.1:18181/v1".to_string()), |
| 1291 | default_text_model: Some("legacy-root-model".to_string()), |
| 1292 | providers: Some(crate::config::ProvidersConfig { |
| 1293 | custom, |
| 1294 | ..crate::config::ProvidersConfig::default() |
| 1295 | }), |
| 1296 | ..Config::default() |
| 1297 | }; |
| 1298 | let manager = RuntimeThreadManager::open( |
| 1299 | config.clone(), |
| 1300 | PathBuf::from("."), |
| 1301 | test_manager_config(test_runtime_dir()), |
| 1302 | )?; |
| 1303 | |
| 1304 | let mut imported = sample_thread("thr_empty_custom_id"); |
| 1305 | imported.model_provider = Some("custom".to_string()); |
| 1306 | imported.model_provider_id = Some(" ".to_string()); |
| 1307 | let error = manager |
| 1308 | .resolved_route_for_thread(&config, &imported) |
| 1309 | .expect_err("malformed imported identity must not acquire the root route") |
| 1310 | .to_string(); |
| 1311 | assert!(error.contains("empty exact provider id"), "{error}"); |
| 1312 | |
| 1313 | let before = manager.store.list_threads()?.len(); |
| 1314 | let request_error = manager |
| 1315 | .create_thread(CreateThreadRequest { |
| 1316 | model_provider: Some("custom".to_string()), |
| 1317 | model_provider_id: Some(String::new()), |
| 1318 | ..CreateThreadRequest::default() |
| 1319 | }) |
| 1320 | .await |
| 1321 | .expect_err("malformed create request must fail before persistence") |
| 1322 | .to_string(); |
| 1323 | assert!( |
| 1324 | request_error.contains("empty exact provider id"), |
| 1325 | "{request_error}" |
| 1326 | ); |
| 1327 | assert_eq!(manager.store.list_threads()?.len(), before); |
| 1328 | Ok(()) |
| 1329 | } |
| 1330 | |
| 1331 | #[tokio::test] |
| 1332 | async fn thread_records_and_create_requests_preserve_provider_kind_id_pairing() -> Result<()> { |
| 1333 | let mut custom = std::collections::HashMap::new(); |
| 1334 | custom.insert( |
| 1335 | "openai".to_string(), |
| 1336 | crate::config::ProviderConfig { |
| 1337 | kind: Some("openai-compatible".to_string()), |
| 1338 | base_url: Some("http://127.0.0.1:18183/v1".to_string()), |
| 1339 | model: Some("custom-openai-model".to_string()), |
| 1340 | ..crate::config::ProviderConfig::default() |
| 1341 | }, |
| 1342 | ); |
| 1343 | let config = Config { |
| 1344 | provider: Some("openai".to_string()), |
| 1345 | providers: Some(crate::config::ProvidersConfig { |
| 1346 | custom, |
| 1347 | ..crate::config::ProvidersConfig::default() |
| 1348 | }), |
| 1349 | ..Config::default() |
| 1350 | }; |
| 1351 | let manager = RuntimeThreadManager::open( |
| 1352 | config.clone(), |
| 1353 | PathBuf::from("."), |
| 1354 | test_manager_config(test_runtime_dir()), |
| 1355 | )?; |
| 1356 | |
| 1357 | for provider_id in [None, Some("openai".to_string())] { |
| 1358 | let mut built_in = sample_thread("thr_builtin_openai_collision"); |
| 1359 | built_in.model_provider = Some("openai".to_string()); |
| 1360 | built_in.model_provider_id = provider_id; |
| 1361 | let error = manager |
| 1362 | .resolved_route_for_thread(&config, &built_in) |
| 1363 | .expect_err("built-in thread must not route through same-key custom endpoint") |
| 1364 | .to_string(); |
| 1365 | assert!(error.contains("requires built-in 'openai'"), "{error}"); |
| 1366 | assert!(error.contains("shadows"), "{error}"); |
| 1367 | } |
| 1368 | |
| 1369 | let mut exact_custom = sample_thread("thr_custom_openai_collision"); |
| 1370 | exact_custom.model = "custom-openai-model".to_string(); |
| 1371 | exact_custom.model_provider = Some("custom".to_string()); |
| 1372 | exact_custom.model_provider_id = Some("openai".to_string()); |
| 1373 | let route = manager.resolved_route_for_thread(&config, &exact_custom)?; |
| 1374 | assert_eq!(route.identity.provider, ApiProvider::Custom); |
| 1375 | assert_eq!(route.identity.key, "openai"); |
| 1376 | assert_eq!( |
| 1377 | route.config.deepseek_base_url(), |
| 1378 | "http://127.0.0.1:18183/v1" |
| 1379 | ); |
| 1380 | |
| 1381 | let mut auto_thread = exact_custom.clone(); |
| 1382 | auto_thread.id = "thr_auto_openai_collision".to_string(); |
| 1383 | auto_thread.model = "auto".to_string(); |
| 1384 | manager.store.save_thread(&auto_thread)?; |
| 1385 | let mut restored_turn = sample_turn( |
| 1386 | &auto_thread.id, |
| 1387 | "turn_openai_collision", |
| 1388 | RuntimeTurnStatus::Completed, |
| 1389 | ); |
| 1390 | restored_turn.effective_provider = Some("openai".to_string()); |
| 1391 | restored_turn.effective_provider_id = None; |
| 1392 | restored_turn.effective_model = Some("custom-openai-model".to_string()); |
| 1393 | manager.store.save_turn(&restored_turn)?; |
| 1394 | let turn_error = manager |
| 1395 | .resolved_route_for_thread(&config, &auto_thread) |
| 1396 | .expect_err("restored built-in turn must not be captured by custom endpoint") |
| 1397 | .to_string(); |
| 1398 | assert!( |
| 1399 | turn_error.contains("requires built-in 'openai'"), |
| 1400 | "{turn_error}" |
| 1401 | ); |
| 1402 | |
| 1403 | restored_turn.effective_provider = Some("custom".to_string()); |
| 1404 | restored_turn.effective_provider_id = Some("openai".to_string()); |
| 1405 | manager.store.save_turn(&restored_turn)?; |
| 1406 | let restored_custom = manager.resolved_route_for_thread(&config, &auto_thread)?; |
| 1407 | assert_eq!(restored_custom.identity.provider, ApiProvider::Custom); |
| 1408 | assert_eq!(restored_custom.identity.key, "openai"); |
| 1409 | assert_eq!(restored_custom.model, "custom-openai-model"); |
| 1410 | |
| 1411 | let request_error = manager |
| 1412 | .create_thread(CreateThreadRequest { |
| 1413 | model_provider: Some("openai".to_string()), |
| 1414 | model_provider_id: Some("openai".to_string()), |
| 1415 | ..CreateThreadRequest::default() |
| 1416 | }) |
| 1417 | .await |
| 1418 | .expect_err("built-in request must fail closed under exact custom shadow") |
| 1419 | .to_string(); |
| 1420 | assert!( |
| 1421 | request_error.contains("requires built-in 'openai'"), |
| 1422 | "{request_error}" |
| 1423 | ); |
| 1424 | |
| 1425 | let created = manager |
| 1426 | .create_thread(CreateThreadRequest { |
| 1427 | model_provider: Some("custom".to_string()), |
| 1428 | model_provider_id: Some("openai".to_string()), |
| 1429 | ..CreateThreadRequest::default() |
| 1430 | }) |
| 1431 | .await?; |
| 1432 | assert_eq!(created.model_provider.as_deref(), Some("custom")); |
| 1433 | assert_eq!(created.model_provider_id.as_deref(), Some("openai")); |
| 1434 | assert_eq!(created.model, "custom-openai-model"); |
| 1435 | Ok(()) |
| 1436 | } |
| 1437 | |
| 1438 | #[tokio::test] |
| 1439 | async fn config_reload_updates_next_turn_route_without_mutating_engine_route() -> Result<()> { |
| 1440 | let mut custom = std::collections::HashMap::new(); |
| 1441 | custom.insert( |
| 1442 | "lm-studio".to_string(), |
| 1443 | crate::config::ProviderConfig { |
| 1444 | kind: Some("openai-compatible".to_string()), |
| 1445 | base_url: Some("http://127.0.0.1:18181/v1".to_string()), |
| 1446 | model: Some("local-model".to_string()), |
| 1447 | api_key: Some("old-local-test-key".to_string()), |
| 1448 | ..crate::config::ProviderConfig::default() |
| 1449 | }, |
| 1450 | ); |
| 1451 | let config = Config { |
| 1452 | provider: Some("lm-studio".to_string()), |
| 1453 | providers: Some(crate::config::ProvidersConfig { |
| 1454 | custom, |
| 1455 | ..crate::config::ProvidersConfig::default() |
| 1456 | }), |
| 1457 | ..Config::default() |
| 1458 | }; |
| 1459 | let manager = RuntimeThreadManager::open( |
| 1460 | config.clone(), |
| 1461 | PathBuf::from("."), |
| 1462 | test_manager_config(test_runtime_dir()), |
| 1463 | )?; |
| 1464 | let thread = manager |
| 1465 | .create_thread(CreateThreadRequest { |
| 1466 | model: Some("local-model".to_string()), |
| 1467 | model_provider: Some("lm-studio".to_string()), |
| 1468 | ..CreateThreadRequest::default() |
| 1469 | }) |
| 1470 | .await?; |
| 1471 | let mut harness = install_mock_engine(&manager, &thread.id).await; |
| 1472 | |
| 1473 | let mut reloaded = config; |
| 1474 | let provider = reloaded |
| 1475 | .providers |
| 1476 | .as_mut() |
| 1477 | .and_then(|providers| providers.custom.get_mut("lm-studio")) |
| 1478 | .expect("named custom provider"); |
| 1479 | provider.base_url = Some("http://127.0.0.1:18182/v1".to_string()); |
| 1480 | provider.api_key = Some("new-local-test-key".to_string()); |
| 1481 | manager.reload_config(reloaded).await?; |
| 1482 | |
| 1483 | let refreshed = manager.resolved_route_for_thread(&manager.read_config(), &thread)?; |
| 1484 | assert_eq!(refreshed.identity.key, "lm-studio"); |
| 1485 | assert_eq!( |
| 1486 | refreshed.config.deepseek_base_url(), |
| 1487 | "http://127.0.0.1:18182/v1" |
| 1488 | ); |
| 1489 | for _ in 0..3 { |
| 1490 | let op = harness.rx_op.recv().await.expect("runtime control op"); |
| 1491 | assert!( |
| 1492 | matches!( |
| 1493 | op, |
| 1494 | Op::SetCompaction { .. } |
| 1495 | | Op::SetStreamChunkTimeout { .. } |
| 1496 | | Op::SetSubagentRuntimeConfig { .. } |
| 1497 | ), |
| 1498 | "reload must not mutate an engine provider route: {op:?}" |
| 1499 | ); |
| 1500 | } |
| 1501 | let compact_turn = manager |
| 1502 | .compact_thread( |
| 1503 | &thread.id, |
| 1504 | CompactThreadRequest { |
| 1505 | reason: Some("verify refreshed route".to_string()), |
| 1506 | }, |
| 1507 | ) |
| 1508 | .await?; |
| 1509 | assert_eq!(compact_turn.effective_provider.as_deref(), Some("custom")); |
| 1510 | assert_eq!( |
| 1511 | compact_turn.effective_provider_id.as_deref(), |
| 1512 | Some("lm-studio") |
| 1513 | ); |
| 1514 | assert_eq!(compact_turn.effective_model.as_deref(), Some("local-model")); |
| 1515 | match harness.rx_op.recv().await { |
| 1516 | Some(Op::CompactContext { route, compaction }) => { |
| 1517 | assert_eq!(route.identity.key, "lm-studio"); |
| 1518 | assert_eq!( |
| 1519 | route.config.deepseek_base_url(), |
| 1520 | "http://127.0.0.1:18182/v1" |
| 1521 | ); |
| 1522 | assert_eq!(compaction.model, "local-model"); |
| 1523 | assert_eq!( |
| 1524 | compaction.effective_context_window, |
| 1525 | Some(crate::route_budget::route_context_window_tokens( |
| 1526 | ApiProvider::Custom, |
| 1527 | "local-model", |
| 1528 | crate::route_budget::known_route_limits(route.candidate.limits()), |
| 1529 | )) |
| 1530 | ); |
| 1531 | } |
| 1532 | other => panic!("expected typed compact route, got {other:?}"), |
| 1533 | } |
| 1534 | |
| 1535 | Ok(()) |
| 1536 | } |
| 1537 | |
| 1538 | #[tokio::test] |
| 1539 | async fn config_sync_reports_removed_named_custom_route_and_keeps_mailbox_clean() -> Result<()> { |
| 1540 | let mut custom = std::collections::HashMap::new(); |
| 1541 | custom.insert( |
| 1542 | "lm-studio".to_string(), |
| 1543 | crate::config::ProviderConfig { |
| 1544 | kind: Some("openai-compatible".to_string()), |
| 1545 | base_url: Some("http://127.0.0.1:18181/v1".to_string()), |
| 1546 | model: Some("local-model".to_string()), |
| 1547 | api_key: Some("local-test-key".to_string()), |
| 1548 | ..crate::config::ProviderConfig::default() |
| 1549 | }, |
| 1550 | ); |
| 1551 | let config = Config { |
| 1552 | provider: Some("lm-studio".to_string()), |
| 1553 | providers: Some(crate::config::ProvidersConfig { |
| 1554 | custom, |
| 1555 | ..crate::config::ProvidersConfig::default() |
| 1556 | }), |
| 1557 | ..Config::default() |
| 1558 | }; |
| 1559 | let manager = RuntimeThreadManager::open( |
| 1560 | config, |
| 1561 | PathBuf::from("."), |
| 1562 | test_manager_config(test_runtime_dir()), |
| 1563 | )?; |
| 1564 | let thread = manager |
| 1565 | .create_thread(CreateThreadRequest { |
| 1566 | model: Some("local-model".to_string()), |
| 1567 | model_provider: Some("lm-studio".to_string()), |
| 1568 | ..CreateThreadRequest::default() |
| 1569 | }) |
| 1570 | .await?; |
| 1571 | let mut harness = install_mock_engine(&manager, &thread.id).await; |
| 1572 | |
| 1573 | let err = manager |
| 1574 | .reload_config(Config::default()) |
| 1575 | .await |
| 1576 | .expect_err("removed named custom route must fail config reload"); |
| 1577 | |
| 1578 | let message = err.to_string(); |
| 1579 | assert!(message.contains(&thread.id), "{message}"); |
| 1580 | assert!(message.contains("lm-studio"), "{message}"); |
| 1581 | assert!(harness.rx_op.try_recv().is_err()); |
| 1582 | Ok(()) |
| 1583 | } |
| 1584 | |
| 1585 | #[tokio::test] |
| 1586 | async fn create_thread_uses_requested_named_custom_provider_default_model() -> Result<()> { |
| 1587 | let mut custom = std::collections::HashMap::new(); |
| 1588 | for (name, base_url, model) in [ |
| 1589 | ("custom-a", "http://127.0.0.1:18181/v1", "model-a"), |
| 1590 | ("custom-b", "http://127.0.0.1:18182/v1", "model-b"), |
| 1591 | ] { |
| 1592 | custom.insert( |
| 1593 | name.to_string(), |
| 1594 | crate::config::ProviderConfig { |
| 1595 | kind: Some("openai-compatible".to_string()), |
| 1596 | base_url: Some(base_url.to_string()), |
| 1597 | model: Some(model.to_string()), |
| 1598 | ..Default::default() |
| 1599 | }, |
| 1600 | ); |
| 1601 | } |
| 1602 | let config = Config { |
| 1603 | provider: Some("custom-b".to_string()), |
| 1604 | providers: Some(crate::config::ProvidersConfig { |
| 1605 | custom, |
| 1606 | ..Default::default() |
| 1607 | }), |
| 1608 | ..Default::default() |
| 1609 | }; |
| 1610 | let manager = RuntimeThreadManager::open( |
| 1611 | config.clone(), |
| 1612 | PathBuf::from("."), |
| 1613 | test_manager_config(test_runtime_dir()), |
| 1614 | )?; |
| 1615 | |
| 1616 | let thread = manager |
| 1617 | .create_thread(CreateThreadRequest { |
| 1618 | model_provider: Some("custom-a".to_string()), |
| 1619 | ..Default::default() |
| 1620 | }) |
| 1621 | .await?; |
| 1622 | |
| 1623 | assert_eq!(thread.model_provider.as_deref(), Some("custom")); |
| 1624 | assert_eq!(thread.model_provider_id.as_deref(), Some("custom-a")); |
| 1625 | assert_eq!(thread.model, "model-a"); |
| 1626 | let route = manager.resolved_route_for_thread(&config, &thread)?; |
| 1627 | assert_eq!(route.identity.key, "custom-a"); |
| 1628 | assert_eq!( |
| 1629 | route.config.deepseek_base_url(), |
| 1630 | "http://127.0.0.1:18181/v1" |
| 1631 | ); |
| 1632 | Ok(()) |
| 1633 | } |
| 1634 | |
| 1635 | #[tokio::test] |
| 1636 | async fn create_thread_uses_requested_non_current_builtin_default_model() -> Result<()> { |
| 1637 | let config = Config { |
| 1638 | provider: Some("openrouter".to_string()), |
| 1639 | default_text_model: Some(DEFAULT_TEXT_MODEL.to_string()), |
| 1640 | ..Default::default() |
| 1641 | }; |
| 1642 | let manager = RuntimeThreadManager::open( |
| 1643 | config, |
| 1644 | PathBuf::from("."), |
| 1645 | test_manager_config(test_runtime_dir()), |
| 1646 | )?; |
| 1647 | |
| 1648 | let thread = manager |
| 1649 | .create_thread(CreateThreadRequest { |
| 1650 | model_provider: Some("zai".to_string()), |
| 1651 | ..Default::default() |
| 1652 | }) |
| 1653 | .await?; |
| 1654 | |
| 1655 | assert_eq!(thread.model_provider.as_deref(), Some("zai")); |
| 1656 | assert_eq!(thread.model, crate::config::DEFAULT_ZAI_MODEL); |
| 1657 | Ok(()) |
| 1658 | } |
| 1659 | |
| 1660 | #[tokio::test] |
| 1661 | async fn simultaneous_named_custom_auto_threads_keep_exact_routes() -> Result<()> { |
| 1662 | let mut custom = std::collections::HashMap::new(); |
| 1663 | for (name, base_url, model) in [ |
| 1664 | ("custom-a", "http://127.0.0.1:18181/v1", "model-a"), |
| 1665 | ("custom-b", "http://127.0.0.1:18182/v1", "model-b"), |
| 1666 | ] { |
| 1667 | custom.insert( |
| 1668 | name.to_string(), |
| 1669 | crate::config::ProviderConfig { |
| 1670 | kind: Some("openai-compatible".to_string()), |
| 1671 | base_url: Some(base_url.to_string()), |
| 1672 | model: Some(model.to_string()), |
| 1673 | ..Default::default() |
| 1674 | }, |
| 1675 | ); |
| 1676 | } |
| 1677 | let manager = RuntimeThreadManager::open( |
| 1678 | Config { |
| 1679 | provider: Some("custom-b".to_string()), |
| 1680 | providers: Some(crate::config::ProvidersConfig { |
| 1681 | custom, |
| 1682 | ..Default::default() |
| 1683 | }), |
| 1684 | ..Default::default() |
| 1685 | }, |
| 1686 | PathBuf::from("."), |
| 1687 | test_manager_config(test_runtime_dir()), |
| 1688 | )?; |
| 1689 | let thread_a = manager |
| 1690 | .create_thread(CreateThreadRequest { |
| 1691 | model: Some("auto".to_string()), |
| 1692 | model_provider: Some("custom-a".to_string()), |
| 1693 | ..Default::default() |
| 1694 | }) |
| 1695 | .await?; |
| 1696 | let thread_b = manager |
| 1697 | .create_thread(CreateThreadRequest { |
| 1698 | model: Some("auto".to_string()), |
| 1699 | model_provider: Some("custom-b".to_string()), |
| 1700 | ..Default::default() |
| 1701 | }) |
| 1702 | .await?; |
| 1703 | let mut harness_a = install_mock_engine(&manager, &thread_a.id).await; |
| 1704 | let mut harness_b = install_mock_engine(&manager, &thread_b.id).await; |
| 1705 | |
| 1706 | let request_a = manager.start_turn( |
| 1707 | &thread_a.id, |
| 1708 | StartTurnRequest { |
| 1709 | prompt: "route A".to_string(), |
| 1710 | ..Default::default() |
| 1711 | }, |
| 1712 | ); |
| 1713 | let request_b = manager.start_turn( |
| 1714 | &thread_b.id, |
| 1715 | StartTurnRequest { |
| 1716 | prompt: "route B".to_string(), |
| 1717 | ..Default::default() |
| 1718 | }, |
| 1719 | ); |
| 1720 | let (turn_a, turn_b) = tokio::join!(request_a, request_b); |
| 1721 | let turn_a = turn_a?; |
| 1722 | let turn_b = turn_b?; |
| 1723 | |
| 1724 | assert_eq!(turn_a.effective_provider.as_deref(), Some("custom")); |
| 1725 | assert_eq!(turn_a.effective_provider_id.as_deref(), Some("custom-a")); |
| 1726 | assert_eq!(turn_a.effective_model.as_deref(), Some("model-a")); |
| 1727 | assert_eq!(turn_b.effective_provider.as_deref(), Some("custom")); |
| 1728 | assert_eq!(turn_b.effective_provider_id.as_deref(), Some("custom-b")); |
| 1729 | assert_eq!(turn_b.effective_model.as_deref(), Some("model-b")); |
| 1730 | match harness_a.rx_op.recv().await { |
| 1731 | Some(Op::SendMessage { route, .. }) => { |
| 1732 | assert_eq!(route.identity.provider, ApiProvider::Custom); |
| 1733 | assert_eq!(route.identity.key, "custom-a"); |
| 1734 | assert_eq!(route.model, "model-a"); |
| 1735 | } |
| 1736 | other => panic!("expected custom A send, got {other:?}"), |
| 1737 | } |
| 1738 | match harness_b.rx_op.recv().await { |
| 1739 | Some(Op::SendMessage { route, .. }) => { |
| 1740 | assert_eq!(route.identity.provider, ApiProvider::Custom); |
| 1741 | assert_eq!(route.identity.key, "custom-b"); |
| 1742 | assert_eq!(route.model, "model-b"); |
| 1743 | } |
| 1744 | other => panic!("expected custom B send, got {other:?}"), |
| 1745 | } |
| 1746 | Ok(()) |
| 1747 | } |
| 1748 | |
| 1749 | #[test] |
| 1750 | fn turn_record_persists_billing_surface_without_raw_endpoint() { |
| 1751 | let mut turn = sample_turn("thr_surface", "turn_surface", RuntimeTurnStatus::Completed); |
| 1752 | let fingerprint = "a".repeat(64); |
| 1753 | turn.persist_effective_route(&crate::cost_status::EffectiveRouteEnvelope { |
| 1754 | provider: ApiProvider::Stepfun, |
| 1755 | provider_identity: "stepfun-primary".to_string(), |
| 1756 | model: "step-3.7-flash".to_string(), |
| 1757 | billing_surface: Some(crate::pricing::STEPFUN_PAYG_BILLING_SURFACE.to_string()), |
| 1758 | endpoint_fingerprint: Some(fingerprint.clone()), |
| 1759 | billing_mode: crate::cost_status::RouteBillingMode::Metered, |
| 1760 | dispatched_at: turn.created_at, |
| 1761 | }); |
| 1762 | |
| 1763 | let value = serde_json::to_value(turn).expect("serialize turn"); |
| 1764 | assert_eq!( |
| 1765 | value["effective_billing_surface"], |
| 1766 | crate::pricing::STEPFUN_PAYG_BILLING_SURFACE |
| 1767 | ); |
| 1768 | assert_eq!(value["effective_billing_mode"], "metered"); |
| 1769 | assert_eq!(value["effective_endpoint_fingerprint"], fingerprint); |
| 1770 | assert!(value["effective_dispatched_at"].is_string()); |
| 1771 | assert!(value.get("base_url").is_none()); |
| 1772 | assert!(value.get("effective_base_url").is_none()); |
| 1773 | } |
| 1774 | |
| 1775 | #[test] |
| 1776 | fn serialized_turn_record_redacts_all_route_and_source_fields() { |
| 1777 | let mut value = serde_json::to_value(sample_turn( |
| 1778 | "thr_secret_route", |
| 1779 | "turn_secret_route", |
| 1780 | RuntimeTurnStatus::Completed, |
| 1781 | )) |
| 1782 | .expect("serialize clean fixture"); |
| 1783 | value["effective_provider"] = serde_json::json!("Authorization: Bearer provider-secret"); |
| 1784 | value["effective_provider_id"] = serde_json::json!("CUSTOM_API_KEY=sk-provider-secret"); |
| 1785 | value["effective_model"] = serde_json::json!("../.ssh/model-secret"); |
| 1786 | value["effective_billing_surface"] = |
| 1787 | serde_json::json!("https://alice:password@example.test/v1?token=secret#fragment"); |
| 1788 | value["effective_endpoint_fingerprint"] = serde_json::json!("secret-endpoint"); |
| 1789 | value["routed_usage_source_ids"] = serde_json::json!(["raw-provider-response-secret"]); |
| 1790 | value["routed_usage"] = serde_json::json!([{ |
| 1791 | "route": { |
| 1792 | "provider": "custom", |
| 1793 | "provider_identity": "TOKEN=ghp_child_secret", |
| 1794 | "model": r"relative\secret\model", |
| 1795 | "billing_surface": "https://host.test/v1?key=child-secret", |
| 1796 | "endpoint_fingerprint": "child-endpoint-secret", |
| 1797 | "billing_mode": "metered", |
| 1798 | "dispatched_at": Utc::now(), |
| 1799 | }, |
| 1800 | "usage": Usage::default(), |
| 1801 | }]); |
| 1802 | let turn: TurnRecord = serde_json::from_value(value).expect("deserialize hostile fixture"); |
| 1803 | |
| 1804 | let serialized = serde_json::to_string(&turn).expect("serialize turn record"); |
| 1805 | for secret in [ |
| 1806 | "provider-secret", |
| 1807 | "sk-provider-secret", |
| 1808 | ".ssh", |
| 1809 | "alice", |
| 1810 | "password", |
| 1811 | "token=secret", |
| 1812 | "raw-provider-response-secret", |
| 1813 | "ghp_child_secret", |
| 1814 | "child-secret", |
| 1815 | "child-endpoint-secret", |
| 1816 | ] { |
| 1817 | assert!( |
| 1818 | !serialized.contains(secret), |
| 1819 | "serialized turn leaked {secret:?}: {serialized}" |
| 1820 | ); |
| 1821 | } |
| 1822 | } |
| 1823 | |
| 1824 | #[tokio::test] |
| 1825 | async fn aggregate_usage_keeps_codex_tokens_without_api_dollar_pricing() -> Result<()> { |
| 1826 | let manager = test_manager(test_runtime_dir())?; |
| 1827 | let mut thread = sample_thread("thr_mixed_routes"); |
| 1828 | thread.model = "auto".to_string(); |
| 1829 | manager.store.save_thread(&thread)?; |
| 1830 | |
| 1831 | let usage = Usage { |
| 1832 | input_tokens: 10_000, |
| 1833 | output_tokens: 1_000, |
| 1834 | ..Usage::default() |
| 1835 | }; |
| 1836 | let mut deepseek = sample_turn(&thread.id, "turn_deepseek", RuntimeTurnStatus::Completed); |
| 1837 | deepseek.usage = Some(usage.clone()); |
| 1838 | set_test_turn_route( |
| 1839 | &mut deepseek, |
| 1840 | ApiProvider::Deepseek, |
| 1841 | ApiProvider::Deepseek.as_str(), |
| 1842 | "deepseek-v4-flash", |
| 1843 | // A real dispatch persists the classified surface. Leaving it absent |
| 1844 | // here would price the row at official DeepSeek rates on nothing but |
| 1845 | // the provider name, which is what the route audit now refuses. |
| 1846 | Some(crate::pricing::FIRST_PARTY_PAYG_BILLING_SURFACE), |
| 1847 | crate::cost_status::RouteBillingMode::Metered, |
| 1848 | ); |
| 1849 | manager.store.save_turn(&deepseek)?; |
| 1850 | |
| 1851 | let mut codex = sample_turn(&thread.id, "turn_codex", RuntimeTurnStatus::Completed); |
| 1852 | codex.usage = Some(usage); |
| 1853 | set_test_turn_route( |
| 1854 | &mut codex, |
| 1855 | ApiProvider::OpenaiCodex, |
| 1856 | ApiProvider::OpenaiCodex.as_str(), |
| 1857 | "gpt-5.5", |
| 1858 | Some(crate::pricing::OAUTH_SUBSCRIPTION_BILLING_SURFACE), |
| 1859 | crate::cost_status::RouteBillingMode::Subscription, |
| 1860 | ); |
| 1861 | manager.store.save_turn(&codex)?; |
| 1862 | |
| 1863 | let report = manager |
| 1864 | .aggregate_usage(None, None, UsageGroupBy::Provider) |
| 1865 | .await?; |
| 1866 | assert_eq!(report.totals.turns, 2); |
| 1867 | assert_eq!(report.totals.input_tokens, 20_000); |
| 1868 | assert!(report.totals.cost_usd > 0.0); |
| 1869 | |
| 1870 | let deepseek_bucket = report |
| 1871 | .buckets |
| 1872 | .iter() |
| 1873 | .find(|bucket| bucket.key == ApiProvider::Deepseek.as_str()) |
| 1874 | .expect("DeepSeek bucket"); |
| 1875 | let codex_bucket = report |
| 1876 | .buckets |
| 1877 | .iter() |
| 1878 | .find(|bucket| bucket.key == ApiProvider::OpenaiCodex.as_str()) |
| 1879 | .expect("Codex bucket"); |
| 1880 | assert!(deepseek_bucket.cost_usd > 0.0); |
| 1881 | assert_eq!(codex_bucket.cost_usd, 0.0); |
| 1882 | assert_eq!(deepseek_bucket.priced_turns, 1); |
| 1883 | assert_eq!(deepseek_bucket.unpriced_turns, 0); |
| 1884 | assert_eq!(codex_bucket.priced_turns, 0); |
| 1885 | assert_eq!(codex_bucket.nonmetered_turns, 1); |
| 1886 | assert_eq!(codex_bucket.unpriced_turns, 0); |
| 1887 | assert!(codex_bucket.unpriced_reasons.is_empty()); |
| 1888 | assert_eq!(report.totals.priced_turns, 1); |
| 1889 | assert_eq!(report.totals.nonmetered_turns, 1); |
| 1890 | assert_eq!(report.totals.unpriced_turns, 0); |
| 1891 | assert!(report.totals.cost_complete); |
| 1892 | assert_eq!(codex_bucket.input_tokens, 10_000); |
| 1893 | assert_eq!(report.totals.cost_usd, deepseek_bucket.cost_usd); |
| 1894 | Ok(()) |
| 1895 | } |
| 1896 | |
| 1897 | /// An aggregate in which nothing could be priced is unavailable, not zero. |
| 1898 | /// |
| 1899 | /// `cost_usd` is a `f64` and an all-unknown run leaves it at `0.0`, so the |
| 1900 | /// only thing standing between a reader and a fabricated "$0.00 spent" is that |
| 1901 | /// the aggregate also reports zero priced turns and incomplete coverage. Those |
| 1902 | /// qualifiers are the contract; this pins them at the aggregate level, which is |
| 1903 | /// where per-turn honesty is most easily lost by summing. |
| 1904 | #[tokio::test] |
| 1905 | async fn aggregate_usage_reports_an_all_unknown_run_as_unavailable_not_zero() -> Result<()> { |
| 1906 | let manager = test_manager(test_runtime_dir())?; |
| 1907 | let thread = sample_thread("thr_all_unknown"); |
| 1908 | manager.store.save_thread(&thread)?; |
| 1909 | let usage = Usage { |
| 1910 | input_tokens: 10_000, |
| 1911 | output_tokens: 1_000, |
| 1912 | ..Usage::default() |
| 1913 | }; |
| 1914 | |
| 1915 | for (index, surface) in [ |
| 1916 | // Endpoint classified but unplaceable. |
| 1917 | Some(crate::pricing::UNCLASSIFIED_BILLING_SURFACE), |
| 1918 | // No endpoint classification at all: a provider name is not evidence |
| 1919 | // that the official endpoint served the turn. |
| 1920 | None, |
| 1921 | ] |
| 1922 | .into_iter() |
| 1923 | .enumerate() |
| 1924 | { |
| 1925 | let mut turn = sample_turn( |
| 1926 | &thread.id, |
| 1927 | &format!("turn_unknown_{index}"), |
| 1928 | RuntimeTurnStatus::Completed, |
| 1929 | ); |
| 1930 | turn.usage = Some(usage.clone()); |
| 1931 | set_test_turn_route( |
| 1932 | &mut turn, |
| 1933 | ApiProvider::Openai, |
| 1934 | ApiProvider::Openai.as_str(), |
| 1935 | "gpt-5.5", |
| 1936 | surface, |
| 1937 | crate::cost_status::RouteBillingMode::Metered, |
| 1938 | ); |
| 1939 | manager.store.save_turn(&turn)?; |
| 1940 | } |
| 1941 | |
| 1942 | let report = manager |
| 1943 | .aggregate_usage(None, None, UsageGroupBy::Thread) |
| 1944 | .await?; |
| 1945 | |
| 1946 | assert_eq!(report.totals.turns, 2); |
| 1947 | // Tokens are known and are reported. Only the money is unknown; an unknown |
| 1948 | // price must not suppress usage that the provider did report. |
| 1949 | assert_eq!(report.totals.input_tokens, 20_000); |
| 1950 | assert_eq!(report.totals.priced_turns, 0); |
| 1951 | assert_eq!(report.totals.unpriced_turns, 2); |
| 1952 | assert_eq!(report.totals.nonmetered_turns, 0); |
| 1953 | assert!( |
| 1954 | !report.totals.cost_complete, |
| 1955 | "an all-unknown run must never be a complete total" |
| 1956 | ); |
| 1957 | assert_eq!( |
| 1958 | report.totals.cost_usd, 0.0, |
| 1959 | "no spend may be invented for an unpriced run" |
| 1960 | ); |
| 1961 | assert!( |
| 1962 | !report.totals.unpriced_reasons.is_empty(), |
| 1963 | "an unpriced aggregate must say why" |
| 1964 | ); |
| 1965 | for bucket in &report.buckets { |
| 1966 | assert_eq!(bucket.priced_turns, 0); |
| 1967 | assert!( |
| 1968 | bucket.unpriced_turns > 0, |
| 1969 | "a bucket cannot report a complete zero either" |
| 1970 | ); |
| 1971 | } |
| 1972 | Ok(()) |
| 1973 | } |
| 1974 | |
| 1975 | #[tokio::test] |
| 1976 | async fn aggregate_usage_marks_unknown_cost_as_subtotal_and_keeps_cache_writes() -> Result<()> { |
| 1977 | let manager = test_manager(test_runtime_dir())?; |
| 1978 | let thread = sample_thread("thr_cost_coverage"); |
| 1979 | manager.store.save_thread(&thread)?; |
| 1980 | let usage = Usage { |
| 1981 | input_tokens: 10_000, |
| 1982 | output_tokens: 1_000, |
| 1983 | prompt_cache_write_tokens: Some(321), |
| 1984 | ..Usage::default() |
| 1985 | }; |
| 1986 | |
| 1987 | let mut priced = sample_turn(&thread.id, "turn_priced", RuntimeTurnStatus::Completed); |
| 1988 | priced.usage = Some(usage.clone()); |
| 1989 | set_test_turn_route( |
| 1990 | &mut priced, |
| 1991 | ApiProvider::Deepseek, |
| 1992 | ApiProvider::Deepseek.as_str(), |
| 1993 | "deepseek-v4-flash", |
| 1994 | Some(crate::pricing::FIRST_PARTY_PAYG_BILLING_SURFACE), |
| 1995 | crate::cost_status::RouteBillingMode::Metered, |
| 1996 | ); |
| 1997 | manager.store.save_turn(&priced)?; |
| 1998 | |
| 1999 | let mut unknown = sample_turn(&thread.id, "turn_unknown", RuntimeTurnStatus::Completed); |
| 2000 | unknown.usage = Some(usage); |
| 2001 | set_test_turn_route( |
| 2002 | &mut unknown, |
| 2003 | ApiProvider::Openai, |
| 2004 | ApiProvider::Openai.as_str(), |
| 2005 | "gpt-5.5", |
| 2006 | Some(crate::pricing::UNCLASSIFIED_BILLING_SURFACE), |
| 2007 | crate::cost_status::RouteBillingMode::Unknown, |
| 2008 | ); |
| 2009 | manager.store.save_turn(&unknown)?; |
| 2010 | |
| 2011 | let report = manager |
| 2012 | .aggregate_usage(None, None, UsageGroupBy::Thread) |
| 2013 | .await?; |
| 2014 | assert!(report.totals.cost_usd > 0.0, "priced subtotal is retained"); |
| 2015 | assert_eq!(report.totals.priced_turns, 1); |
| 2016 | assert_eq!(report.totals.unpriced_turns, 1); |
| 2017 | assert!(!report.totals.cost_complete); |
| 2018 | assert_eq!(report.totals.nonmetered_turns, 0); |
| 2019 | assert_eq!(report.totals.cache_write_tokens, 642); |
| 2020 | assert!( |
| 2021 | report |
| 2022 | .totals |
| 2023 | .unpriced_reasons |
| 2024 | .contains("unknown_billing_basis") |
| 2025 | ); |
| 2026 | Ok(()) |
| 2027 | } |
| 2028 | |
| 2029 | #[tokio::test] |
| 2030 | async fn aggregate_usage_prices_slow_predispatch_turn_at_dispatch_boundary() -> Result<()> { |
| 2031 | let manager = test_manager(test_runtime_dir())?; |
| 2032 | let mut thread = sample_thread("thr_historical_pricing"); |
| 2033 | thread.model = "claude-sonnet-5".to_string(); |
| 2034 | manager.store.save_thread(&thread)?; |
| 2035 | |
| 2036 | let usage = Usage { |
| 2037 | input_tokens: 1_000_000, |
| 2038 | output_tokens: 0, |
| 2039 | ..Usage::default() |
| 2040 | }; |
| 2041 | for (turn_id, created_at, dispatched_at) in [ |
| 2042 | ("turn_intro", "2026-08-31T23:59:58Z", "2026-08-31T23:59:59Z"), |
| 2043 | ( |
| 2044 | "turn_slow_snapshot", |
| 2045 | "2026-08-31T23:59:59Z", |
| 2046 | "2026-09-01T00:00:00Z", |
| 2047 | ), |
| 2048 | ] { |
| 2049 | let mut turn = sample_turn(&thread.id, turn_id, RuntimeTurnStatus::Completed); |
| 2050 | turn.created_at = created_at.parse().expect("recorded turn time"); |
| 2051 | turn.started_at = Some(turn.created_at); |
| 2052 | turn.usage = Some(usage.clone()); |
| 2053 | set_test_turn_route( |
| 2054 | &mut turn, |
| 2055 | ApiProvider::Anthropic, |
| 2056 | ApiProvider::Anthropic.as_str(), |
| 2057 | "claude-sonnet-5", |
| 2058 | Some(crate::pricing::FIRST_PARTY_PAYG_BILLING_SURFACE), |
| 2059 | crate::cost_status::RouteBillingMode::Metered, |
| 2060 | ); |
| 2061 | turn.effective_dispatched_at = Some(dispatched_at.parse().expect("dispatch time")); |
| 2062 | manager.store.save_turn(&turn)?; |
| 2063 | } |
| 2064 | |
| 2065 | let report = manager |
| 2066 | .aggregate_usage(None, None, UsageGroupBy::Model) |
| 2067 | .await?; |
| 2068 | |
| 2069 | assert_eq!(report.totals.turns, 2); |
| 2070 | assert!((report.totals.cost_usd - 5.0).abs() < f64::EPSILON); |
| 2071 | assert_eq!(report.buckets.len(), 1); |
| 2072 | assert!((report.buckets[0].cost_usd - 5.0).abs() < f64::EPSILON); |
| 2073 | Ok(()) |
| 2074 | } |
| 2075 | |
| 2076 | #[tokio::test] |
| 2077 | async fn aggregate_usage_prices_only_stepfun_payg_surface() -> Result<()> { |
| 2078 | let manager = test_manager(test_runtime_dir())?; |
| 2079 | let mut thread = sample_thread("thr_stepfun_surfaces"); |
| 2080 | thread.model = "step-3.7-flash".to_string(); |
| 2081 | manager.store.save_thread(&thread)?; |
| 2082 | |
| 2083 | let usage = Usage { |
| 2084 | input_tokens: 1_000_000, |
| 2085 | output_tokens: 500_000, |
| 2086 | prompt_cache_hit_tokens: Some(250_000), |
| 2087 | ..Usage::default() |
| 2088 | }; |
| 2089 | for (turn_id, surface) in [ |
| 2090 | ( |
| 2091 | "turn_stepfun_payg", |
| 2092 | crate::pricing::STEPFUN_PAYG_BILLING_SURFACE, |
| 2093 | ), |
| 2094 | ( |
| 2095 | "turn_stepfun_plan", |
| 2096 | crate::pricing::STEPFUN_PLAN_BILLING_SURFACE, |
| 2097 | ), |
| 2098 | ] { |
| 2099 | let mut turn = sample_turn(&thread.id, turn_id, RuntimeTurnStatus::Completed); |
| 2100 | turn.usage = Some(usage.clone()); |
| 2101 | set_test_turn_route( |
| 2102 | &mut turn, |
| 2103 | ApiProvider::Stepfun, |
| 2104 | ApiProvider::Stepfun.as_str(), |
| 2105 | "step-3.7-flash", |
| 2106 | Some(surface), |
| 2107 | if surface == crate::pricing::STEPFUN_PLAN_BILLING_SURFACE { |
| 2108 | crate::cost_status::RouteBillingMode::Subscription |
| 2109 | } else { |
| 2110 | crate::cost_status::RouteBillingMode::Metered |
| 2111 | }, |
| 2112 | ); |
| 2113 | manager.store.save_turn(&turn)?; |
| 2114 | } |
| 2115 | |
| 2116 | let report = manager |
| 2117 | .aggregate_usage(None, None, UsageGroupBy::Provider) |
| 2118 | .await?; |
| 2119 | assert_eq!(report.totals.turns, 2); |
| 2120 | assert!((report.totals.cost_usd - 0.735).abs() < 1e-12); |
| 2121 | Ok(()) |
| 2122 | } |
| 2123 | |
| 2124 | #[tokio::test] |
| 2125 | async fn aggregate_usage_includes_exclusive_child_calls_and_zero_usage_receipts() -> Result<()> { |
| 2126 | let manager = test_manager(test_runtime_dir())?; |
| 2127 | let thread = sample_thread("thr_child_usage"); |
| 2128 | manager.store.save_thread(&thread)?; |
| 2129 | |
| 2130 | let mut turn = sample_turn(&thread.id, "turn_parent", RuntimeTurnStatus::Completed); |
| 2131 | turn.usage = Some(Usage { |
| 2132 | input_tokens: 100, |
| 2133 | output_tokens: 20, |
| 2134 | prompt_cache_hit_tokens: Some(40), |
| 2135 | prompt_cache_miss_tokens: Some(50), |
| 2136 | prompt_cache_write_tokens: Some(10), |
| 2137 | reasoning_tokens: Some(5), |
| 2138 | ..Usage::default() |
| 2139 | }); |
| 2140 | set_test_turn_route( |
| 2141 | &mut turn, |
| 2142 | ApiProvider::Deepseek, |
| 2143 | "deepseek-parent", |
| 2144 | "deepseek-v4-flash", |
| 2145 | Some(crate::pricing::FIRST_PARTY_PAYG_BILLING_SURFACE), |
| 2146 | crate::cost_status::RouteBillingMode::Metered, |
| 2147 | ); |
| 2148 | turn.routed_usage |
| 2149 | .push(crate::cost_status::EffectiveRouteUsage { |
| 2150 | route: crate::cost_status::EffectiveRouteEnvelope { |
| 2151 | provider: ApiProvider::Deepseek, |
| 2152 | provider_identity: "deepseek-child".to_string(), |
| 2153 | model: "deepseek-v4-flash".to_string(), |
| 2154 | // A child call carries the surface its own dispatch classified; |
| 2155 | // the route audit will not price a metered route without one. |
| 2156 | billing_surface: Some(crate::pricing::FIRST_PARTY_PAYG_BILLING_SURFACE.to_string()), |
| 2157 | endpoint_fingerprint: None, |
| 2158 | billing_mode: crate::cost_status::RouteBillingMode::Metered, |
| 2159 | dispatched_at: turn.created_at, |
| 2160 | }, |
| 2161 | usage: Usage { |
| 2162 | input_tokens: 30, |
| 2163 | output_tokens: 7, |
| 2164 | prompt_cache_hit_tokens: Some(10), |
| 2165 | prompt_cache_miss_tokens: Some(20), |
| 2166 | reasoning_tokens: Some(6), |
| 2167 | ..Usage::default() |
| 2168 | }, |
| 2169 | }); |
| 2170 | turn.routed_usage |
| 2171 | .push(crate::cost_status::EffectiveRouteUsage { |
| 2172 | route: crate::cost_status::EffectiveRouteEnvelope { |
| 2173 | provider: ApiProvider::OpenaiCodex, |
| 2174 | provider_identity: "codex-oauth".to_string(), |
| 2175 | model: "gpt-5.5".to_string(), |
| 2176 | billing_surface: Some( |
| 2177 | crate::pricing::OAUTH_SUBSCRIPTION_BILLING_SURFACE.to_string(), |
| 2178 | ), |
| 2179 | endpoint_fingerprint: None, |
| 2180 | billing_mode: crate::cost_status::RouteBillingMode::Subscription, |
| 2181 | dispatched_at: turn.created_at, |
| 2182 | }, |
| 2183 | usage: Usage::default(), |
| 2184 | }); |
| 2185 | manager.store.save_turn(&turn)?; |
| 2186 | |
| 2187 | let report = manager |
| 2188 | .aggregate_usage(None, None, UsageGroupBy::Thread) |
| 2189 | .await?; |
| 2190 | assert_eq!(report.totals.turns, 3, "parent plus two child calls"); |
| 2191 | assert_eq!(report.totals.input_tokens, 70); |
| 2192 | assert_eq!(report.totals.cached_tokens, 50); |
| 2193 | assert_eq!(report.totals.cache_write_tokens, 10); |
| 2194 | assert_eq!(report.totals.output_tokens, 27); |
| 2195 | assert_eq!(report.totals.reasoning_tokens, 11); |
| 2196 | assert_eq!(report.totals.priced_turns, 2); |
| 2197 | assert_eq!(report.totals.nonmetered_turns, 1); |
| 2198 | assert_eq!(report.totals.unpriced_turns, 0); |
| 2199 | assert!(report.totals.cost_complete); |
| 2200 | assert_eq!(report.totals.route_receipts.len(), 3); |
| 2201 | Ok(()) |
| 2202 | } |
| 2203 | |
| 2204 | #[tokio::test] |
| 2205 | async fn aggregate_usage_filters_each_call_by_its_dispatch_timestamp() -> Result<()> { |
| 2206 | let manager = test_manager(test_runtime_dir())?; |
| 2207 | let thread = sample_thread("thr_dispatch_window"); |
| 2208 | manager.store.save_thread(&thread)?; |
| 2209 | let window = Utc::now(); |
| 2210 | let mut turn = sample_turn( |
| 2211 | &thread.id, |
| 2212 | "turn_dispatch_window", |
| 2213 | RuntimeTurnStatus::Completed, |
| 2214 | ); |
| 2215 | turn.created_at = window - chrono::Duration::days(2); |
| 2216 | turn.usage = Some(Usage { |
| 2217 | input_tokens: 100, |
| 2218 | ..Usage::default() |
| 2219 | }); |
| 2220 | set_test_turn_route( |
| 2221 | &mut turn, |
| 2222 | ApiProvider::Deepseek, |
| 2223 | "deepseek-parent", |
| 2224 | "deepseek-v4-flash", |
| 2225 | Some(crate::pricing::FIRST_PARTY_PAYG_BILLING_SURFACE), |
| 2226 | crate::cost_status::RouteBillingMode::Metered, |
| 2227 | ); |
| 2228 | turn.routed_usage |
| 2229 | .push(crate::cost_status::EffectiveRouteUsage { |
| 2230 | route: crate::cost_status::EffectiveRouteEnvelope { |
| 2231 | provider: ApiProvider::Deepseek, |
| 2232 | provider_identity: "deepseek-child".to_string(), |
| 2233 | model: "deepseek-v4-flash".to_string(), |
| 2234 | billing_surface: None, |
| 2235 | endpoint_fingerprint: None, |
| 2236 | billing_mode: crate::cost_status::RouteBillingMode::Metered, |
| 2237 | dispatched_at: window, |
| 2238 | }, |
| 2239 | usage: Usage { |
| 2240 | input_tokens: 33, |
| 2241 | ..Usage::default() |
| 2242 | }, |
| 2243 | }); |
| 2244 | manager.store.save_turn(&turn)?; |
| 2245 | |
| 2246 | let report = manager |
| 2247 | .aggregate_usage( |
| 2248 | Some(window - chrono::Duration::minutes(1)), |
| 2249 | Some(window + chrono::Duration::minutes(1)), |
| 2250 | UsageGroupBy::Thread, |
| 2251 | ) |
| 2252 | .await?; |
| 2253 | assert_eq!(report.totals.turns, 1); |
| 2254 | assert_eq!(report.totals.input_tokens, 33); |
| 2255 | Ok(()) |
| 2256 | } |
| 2257 | |
| 2258 | #[tokio::test] |
| 2259 | async fn aggregate_usage_marks_bounded_journal_truncation_incomplete() -> Result<()> { |
| 2260 | let manager = test_manager(test_runtime_dir())?; |
| 2261 | let thread = sample_thread("thr_truncated_usage"); |
| 2262 | manager.store.save_thread(&thread)?; |
| 2263 | let mut turn = sample_turn( |
| 2264 | &thread.id, |
| 2265 | "turn_truncated_usage", |
| 2266 | RuntimeTurnStatus::Completed, |
| 2267 | ); |
| 2268 | turn.routed_usage_dropped_records = 2; |
| 2269 | manager.store.save_turn(&turn)?; |
| 2270 | |
| 2271 | let report = manager |
| 2272 | .aggregate_usage(None, None, UsageGroupBy::Thread) |
| 2273 | .await?; |
| 2274 | assert_eq!(report.totals.dropped_usage_records, 2); |
| 2275 | assert_eq!(report.totals.unpriced_turns, 2); |
| 2276 | assert_eq!(report.totals.turns, 2); |
| 2277 | assert!(!report.totals.cost_complete); |
| 2278 | assert!( |
| 2279 | report |
| 2280 | .totals |
| 2281 | .unpriced_reasons |
| 2282 | .contains("runtime_usage_journal_truncated") |
| 2283 | ); |
| 2284 | Ok(()) |
| 2285 | } |
| 2286 | |
| 2287 | #[test] |
| 2288 | fn runtime_usage_sink_survives_parent_terminal_and_restart_exactly_once() -> Result<()> { |
| 2289 | let _cost_scope = crate::cost_status::test_scope(); |
| 2290 | let data_dir = test_runtime_dir(); |
| 2291 | let manager = test_manager(data_dir.clone())?; |
| 2292 | let thread = sample_thread("thr_compaction_sink"); |
| 2293 | manager.store.save_thread(&thread)?; |
| 2294 | let turn = sample_turn( |
| 2295 | &thread.id, |
| 2296 | "turn_compaction_sink", |
| 2297 | RuntimeTurnStatus::InProgress, |
| 2298 | ); |
| 2299 | manager.store.save_turn(&turn)?; |
| 2300 | manager.register_runtime_usage_sink(&turn.id); |
| 2301 | let lease = crate::cost_status::acquire_runtime_usage_lease(&turn.id) |
| 2302 | .expect("active runtime owner lease"); |
| 2303 | crate::cost_status::finish_runtime_usage_owner(&turn.id); |
| 2304 | |
| 2305 | let route = crate::cost_status::EffectiveRouteEnvelope::capture( |
| 2306 | None, |
| 2307 | ApiProvider::Deepseek, |
| 2308 | "deepseek-compaction", |
| 2309 | "deepseek-v4-flash", |
| 2310 | Some(ApiProvider::Deepseek.default_base_url()), |
| 2311 | Utc::now(), |
| 2312 | ); |
| 2313 | crate::cost_status::report_effective_route_for_runtime( |
| 2314 | crate::cost_status::scope_token(), |
| 2315 | Some(&turn.id), |
| 2316 | "compaction:response:test", |
| 2317 | &route, |
| 2318 | &Usage { |
| 2319 | input_tokens: 17, |
| 2320 | output_tokens: 3, |
| 2321 | ..Usage::default() |
| 2322 | }, |
| 2323 | ); |
| 2324 | crate::cost_status::report_effective_route_for_runtime( |
| 2325 | crate::cost_status::scope_token(), |
| 2326 | Some(&turn.id), |
| 2327 | "compaction:response:test", |
| 2328 | &route, |
| 2329 | &Usage { |
| 2330 | input_tokens: 17, |
| 2331 | output_tokens: 3, |
| 2332 | ..Usage::default() |
| 2333 | }, |
| 2334 | ); |
| 2335 | drop(lease); |
| 2336 | |
| 2337 | let reopened = test_manager(data_dir.clone())?; |
| 2338 | reopened.register_runtime_usage_sink(&turn.id); |
| 2339 | crate::cost_status::report_effective_route_for_runtime( |
| 2340 | crate::cost_status::scope_token(), |
| 2341 | Some(&turn.id), |
| 2342 | "compaction:response:test", |
| 2343 | &route, |
| 2344 | &Usage { |
| 2345 | input_tokens: 17, |
| 2346 | output_tokens: 3, |
| 2347 | ..Usage::default() |
| 2348 | }, |
| 2349 | ); |
| 2350 | crate::cost_status::finish_runtime_usage_owner(&turn.id); |
| 2351 | let reopened = test_manager(data_dir)?; |
| 2352 | let persisted = reopened.store.load_turn(&turn.id)?; |
| 2353 | assert_eq!(persisted.routed_usage.len(), 1); |
| 2354 | assert_eq!(persisted.routed_usage[0].usage.input_tokens, 17); |
| 2355 | assert!( |
| 2356 | crate::cost_status::take_runtime_usage(&turn.id) |
| 2357 | .records |
| 2358 | .is_empty(), |
| 2359 | "successful synchronous persistence must not duplicate into fallback" |
| 2360 | ); |
| 2361 | assert!( |
| 2362 | crate::cost_status::drain().is_empty(), |
| 2363 | "runtime-owned usage must not duplicate into the TUI pool" |
| 2364 | ); |
| 2365 | Ok(()) |
| 2366 | } |
| 2367 | |
| 2368 | #[test] |
| 2369 | fn routed_usage_append_is_bounded_and_idempotent_for_every_delivery_path() { |
| 2370 | let mut turn = sample_turn( |
| 2371 | "thr_all_paths", |
| 2372 | "turn_all_paths", |
| 2373 | RuntimeTurnStatus::InProgress, |
| 2374 | ); |
| 2375 | let route = crate::cost_status::EffectiveRouteEnvelope::capture( |
| 2376 | None, |
| 2377 | ApiProvider::Deepseek, |
| 2378 | "deepseek-primary", |
| 2379 | "deepseek-v4-flash", |
| 2380 | Some(ApiProvider::Deepseek.default_base_url()), |
| 2381 | Utc::now(), |
| 2382 | ); |
| 2383 | for index in 0..90 { |
| 2384 | let path = match index % 3 { |
| 2385 | 0 => "sink", |
| 2386 | 1 => "mailbox", |
| 2387 | _ => "fallback", |
| 2388 | }; |
| 2389 | assert!(append_routed_usage_record( |
| 2390 | &mut turn, |
| 2391 | &format!("{path}:response:{index}"), |
| 2392 | crate::cost_status::EffectiveRouteUsage { |
| 2393 | route: route.clone(), |
| 2394 | usage: Usage { |
| 2395 | input_tokens: index, |
| 2396 | ..Usage::default() |
| 2397 | }, |
| 2398 | }, |
| 2399 | )); |
| 2400 | } |
| 2401 | |
| 2402 | assert_eq!(turn.routed_usage.len(), MAX_ROUTED_USAGE_RECORDS_PER_TURN); |
| 2403 | assert_eq!(turn.routed_usage_dropped_records, 26); |
| 2404 | assert_eq!(turn.routed_usage_source_ids.len(), 90); |
| 2405 | let before = turn.clone(); |
| 2406 | assert!(!append_routed_usage_record( |
| 2407 | &mut turn, |
| 2408 | "sink:response:0", |
| 2409 | crate::cost_status::EffectiveRouteUsage { |
| 2410 | route, |
| 2411 | usage: Usage { |
| 2412 | input_tokens: 999, |
| 2413 | ..Usage::default() |
| 2414 | }, |
| 2415 | }, |
| 2416 | )); |
| 2417 | assert_eq!(turn.routed_usage, before.routed_usage); |
| 2418 | assert_eq!( |
| 2419 | turn.routed_usage_dropped_records, |
| 2420 | before.routed_usage_dropped_records |
| 2421 | ); |
| 2422 | } |
| 2423 | |
| 2424 | #[tokio::test] |
| 2425 | async fn aggregate_usage_fails_closed_for_legacy_reconstructed_route() -> Result<()> { |
| 2426 | let manager = test_manager(test_runtime_dir())?; |
| 2427 | let thread = sample_thread("thr_legacy_cost"); |
| 2428 | manager.store.save_thread(&thread)?; |
| 2429 | let mut turn = sample_turn(&thread.id, "turn_legacy_cost", RuntimeTurnStatus::Completed); |
| 2430 | turn.usage = Some(Usage { |
| 2431 | input_tokens: 1_000, |
| 2432 | output_tokens: 10, |
| 2433 | ..Usage::default() |
| 2434 | }); |
| 2435 | // These mutable-era fields are intentionally insufficient: no exact |
| 2436 | // identity, billing mode, or dispatch timestamp was persisted. |
| 2437 | turn.effective_provider = Some(ApiProvider::Deepseek.as_str().to_string()); |
| 2438 | turn.effective_model = Some("deepseek-v4-flash".to_string()); |
| 2439 | manager.store.save_turn(&turn)?; |
| 2440 | |
| 2441 | let report = manager |
| 2442 | .aggregate_usage(None, None, UsageGroupBy::Thread) |
| 2443 | .await?; |
| 2444 | assert_eq!(report.totals.cost_usd, 0.0); |
| 2445 | assert_eq!(report.totals.unpriced_turns, 1); |
| 2446 | assert!( |
| 2447 | report |
| 2448 | .totals |
| 2449 | .unpriced_reasons |
| 2450 | .contains("unknown_provider_route") |
| 2451 | ); |
| 2452 | assert!(report.totals.route_receipts.is_empty()); |
| 2453 | Ok(()) |
| 2454 | } |
| 2455 | |
| 2456 | #[test] |
| 2457 | fn runtime_usage_accumulators_saturate_tokens_and_currency() { |
| 2458 | let mut total = f64::MAX; |
| 2459 | saturating_add_usd(&mut total, f64::MAX); |
| 2460 | assert_eq!(total, f64::MAX); |
| 2461 | |
| 2462 | let thread = sample_thread("thr_saturating"); |
| 2463 | let turn = sample_turn(&thread.id, "turn_saturating", RuntimeTurnStatus::Completed); |
| 2464 | let mut totals = UsageTotals { |
| 2465 | input_tokens: u64::MAX, |
| 2466 | output_tokens: u64::MAX, |
| 2467 | cached_tokens: u64::MAX, |
| 2468 | reasoning_tokens: u64::MAX, |
| 2469 | reasoning_replay_tokens: u64::MAX, |
| 2470 | cache_write_tokens: u64::MAX, |
| 2471 | ..UsageTotals::default() |
| 2472 | }; |
| 2473 | let mut buckets = std::collections::BTreeMap::new(); |
| 2474 | accumulate_runtime_usage_record( |
| 2475 | &mut totals, |
| 2476 | &mut buckets, |
| 2477 | UsageGroupBy::Thread, |
| 2478 | None, |
| 2479 | &Usage { |
| 2480 | input_tokens: 1, |
| 2481 | output_tokens: 1, |
| 2482 | prompt_cache_hit_tokens: Some(1), |
| 2483 | prompt_cache_write_tokens: Some(1), |
| 2484 | reasoning_tokens: Some(1), |
| 2485 | reasoning_replay_tokens: Some(1), |
| 2486 | ..Usage::default() |
| 2487 | }, |
| 2488 | &turn, |
| 2489 | &thread, |
| 2490 | ); |
| 2491 | assert_eq!(totals.input_tokens, u64::MAX); |
| 2492 | assert_eq!(totals.output_tokens, u64::MAX); |
| 2493 | assert_eq!(totals.cached_tokens, u64::MAX); |
| 2494 | assert_eq!(totals.reasoning_tokens, u64::MAX); |
| 2495 | assert_eq!(totals.reasoning_replay_tokens, u64::MAX); |
| 2496 | assert_eq!(totals.cache_write_tokens, u64::MAX); |
| 2497 | } |
| 2498 | |
| 2499 | fn sample_item(turn_id: &str, item_id: &str, status: TurnItemLifecycleStatus) -> TurnItemRecord { |
| 2500 | TurnItemRecord { |
| 2501 | schema_version: CURRENT_RUNTIME_SCHEMA_VERSION, |
| 2502 | id: item_id.to_string(), |
| 2503 | turn_id: turn_id.to_string(), |
| 2504 | kind: TurnItemKind::Status, |
| 2505 | status, |
| 2506 | summary: "sample item".to_string(), |
| 2507 | detail: None, |
| 2508 | metadata: None, |
| 2509 | artifact_refs: Vec::new(), |
| 2510 | started_at: Some(Utc::now()), |
| 2511 | ended_at: None, |
| 2512 | } |
| 2513 | } |
| 2514 | |
| 2515 | async fn install_mock_engine( |
| 2516 | manager: &RuntimeThreadManager, |
| 2517 | thread_id: &str, |
| 2518 | ) -> crate::core::engine::MockEngineHandle { |
| 2519 | let harness = mock_engine_handle(); |
| 2520 | manager |
| 2521 | .install_test_engine(thread_id, harness.handle.clone()) |
| 2522 | .await |
| 2523 | .expect("install mock engine"); |
| 2524 | harness |
| 2525 | } |
| 2526 | |
| 2527 | async fn wait_for_sender_strong_count<T>( |
| 2528 | sender: &tokio::sync::mpsc::Sender<T>, |
| 2529 | minimum: usize, |
| 2530 | ) -> Result<()> { |
| 2531 | tokio::time::timeout(Duration::from_secs(2), async { |
| 2532 | while sender.strong_count() < minimum { |
| 2533 | tokio::task::yield_now().await; |
| 2534 | } |
| 2535 | }) |
| 2536 | .await |
| 2537 | .map_err(|_| anyhow!("Timed out waiting for mailbox reservation"))?; |
| 2538 | Ok(()) |
| 2539 | } |
| 2540 | |
| 2541 | async fn wait_for_terminal_turn( |
| 2542 | manager: &RuntimeThreadManager, |
| 2543 | turn_id: &str, |
| 2544 | timeout: Duration, |
| 2545 | ) -> Result<TurnRecord> { |
| 2546 | let deadline = Instant::now() + timeout; |
| 2547 | loop { |
| 2548 | let turn = manager.store.load_turn(turn_id)?; |
| 2549 | let terminal = matches!( |
| 2550 | turn.status, |
| 2551 | RuntimeTurnStatus::Completed |
| 2552 | | RuntimeTurnStatus::Failed |
| 2553 | | RuntimeTurnStatus::Interrupted |
| 2554 | | RuntimeTurnStatus::Canceled |
| 2555 | ); |
| 2556 | if terminal { |
| 2557 | let receipt_is_durable = |
| 2558 | manager |
| 2559 | .events_since(&turn.thread_id, None)? |
| 2560 | .iter() |
| 2561 | .any(|event| { |
| 2562 | event.event == "turn.completed" && event.turn_id.as_deref() == Some(turn_id) |
| 2563 | }); |
| 2564 | let claim_is_clear = manager |
| 2565 | .active_turn_flags(&turn.thread_id, turn_id) |
| 2566 | .await |
| 2567 | .is_none(); |
| 2568 | if receipt_is_durable && claim_is_clear { |
| 2569 | return Ok(turn); |
| 2570 | } |
| 2571 | } |
| 2572 | if Instant::now() >= deadline { |
| 2573 | bail!("Timed out waiting for turn {turn_id}"); |
| 2574 | } |
| 2575 | sleep(Duration::from_millis(20)).await; |
| 2576 | } |
| 2577 | } |
| 2578 | |
| 2579 | #[test] |
| 2580 | fn store_load_thread_rejects_newer_schema_version() { |
| 2581 | let dir = test_runtime_dir(); |
| 2582 | let store = RuntimeThreadStore::open(dir.clone()).expect("open store"); |
| 2583 | |
| 2584 | // Construct a thread record persisted with a future schema version. |
| 2585 | let mut thread = sample_thread("thr_future"); |
| 2586 | thread.schema_version = CURRENT_RUNTIME_SCHEMA_VERSION + 1; |
| 2587 | |
| 2588 | // Bypass save_thread (which would respect our local schema_version) |
| 2589 | // by writing the JSON directly so we can simulate a future writer. |
| 2590 | let path = store.threads_dir.join(format!("{}.json", thread.id)); |
| 2591 | std::fs::create_dir_all(path.parent().unwrap()).expect("mkdirs"); |
| 2592 | let payload = serde_json::to_string(&thread).expect("serialize thread"); |
| 2593 | std::fs::write(&path, payload).expect("write thread"); |
| 2594 | |
| 2595 | let err = store |
| 2596 | .load_thread(&thread.id) |
| 2597 | .expect_err("load_thread must reject newer schema"); |
| 2598 | let msg = format!("{err:#}"); |
| 2599 | assert!(msg.contains("newer than supported"), "got: {msg}"); |
| 2600 | |
| 2601 | // Cleanup so we don't leak across tests. |
| 2602 | let _ = std::fs::remove_dir_all(dir); |
| 2603 | } |
| 2604 | |
| 2605 | #[test] |
| 2606 | fn runtime_event_sequences_serialize_across_real_processes() -> Result<()> { |
| 2607 | let dir = test_runtime_dir(); |
| 2608 | let store = RuntimeThreadStore::open(dir.clone())?; |
| 2609 | let thread_id = "thr_process_writers"; |
| 2610 | let start = dir.join("process-writers.start"); |
| 2611 | let writer_count = 4_u64; |
| 2612 | let events_per_writer = 8_u64; |
| 2613 | let mut children = Vec::new(); |
| 2614 | |
| 2615 | for worker in 0..writer_count { |
| 2616 | let signal = dir.join(format!("process-writer-{worker}.ready")); |
| 2617 | children.push(( |
| 2618 | spawn_runtime_event_child("writer", &dir, thread_id, &signal), |
| 2619 | signal, |
| 2620 | )); |
| 2621 | } |
| 2622 | for (_, signal) in &children { |
| 2623 | wait_for_runtime_event_test_file(signal, "event writer readiness"); |
| 2624 | } |
| 2625 | // The children already opened their stores before this barrier. The old |
| 2626 | // store-local sequence cache therefore allocated duplicates deterministically. |
| 2627 | std::fs::write(&start, b"go")?; |
| 2628 | for (child, _) in children { |
| 2629 | child.wait_success("cross-process event writer"); |
| 2630 | } |
| 2631 | |
| 2632 | let events = store.events_since(thread_id, None)?; |
| 2633 | let expected = writer_count * events_per_writer; |
| 2634 | assert_eq!(u64::try_from(events.len())?, expected); |
| 2635 | assert_eq!( |
| 2636 | events.iter().map(|event| event.seq).collect::<Vec<_>>(), |
| 2637 | (1..=expected).collect::<Vec<_>>() |
| 2638 | ); |
| 2639 | let identities = events |
| 2640 | .iter() |
| 2641 | .map(|event| { |
| 2642 | ( |
| 2643 | event.payload["worker"].as_str().unwrap().to_string(), |
| 2644 | event.payload["index"].as_u64().unwrap(), |
| 2645 | ) |
| 2646 | }) |
| 2647 | .collect::<HashSet<_>>(); |
| 2648 | assert_eq!(u64::try_from(identities.len())?, expected); |
| 2649 | let runtime = tokio::runtime::Builder::new_current_thread() |
| 2650 | .enable_all() |
| 2651 | .build()?; |
| 2652 | assert_eq!(runtime.block_on(store.current_seq())?, expected); |
| 2653 | assert_eq!( |
| 2654 | load_runtime_store_state(&store.state_path)?.next_seq, |
| 2655 | expected + 1 |
| 2656 | ); |
| 2657 | |
| 2658 | std::fs::remove_dir_all(dir)?; |
| 2659 | Ok(()) |
| 2660 | } |
| 2661 | |
| 2662 | #[test] |
| 2663 | fn runtime_event_lock_timeout_is_non_mutating_and_holder_death_releases_lock() -> Result<()> { |
| 2664 | let dir = test_runtime_dir(); |
| 2665 | let store = RuntimeThreadStore::open(dir.clone())?; |
| 2666 | let thread_id = "thr_process_holder"; |
| 2667 | let signal = dir.join("process-holder.ready"); |
| 2668 | let child = spawn_runtime_event_child("holder", &dir, thread_id, &signal); |
| 2669 | wait_for_runtime_event_test_file(&signal, "held Runtime event lock"); |
| 2670 | |
| 2671 | let state_before = std::fs::read(&store.state_path)?; |
| 2672 | let event_path = store.events_path(thread_id)?; |
| 2673 | let started = Instant::now(); |
| 2674 | let error = store |
| 2675 | .append_event_transaction( |
| 2676 | thread_id.to_string(), |
| 2677 | None, |
| 2678 | None, |
| 2679 | "must.timeout".to_string(), |
| 2680 | json!({}), |
| 2681 | Duration::from_millis(75), |
| 2682 | ) |
| 2683 | .expect_err("contended Runtime event transaction must time out"); |
| 2684 | let elapsed = started.elapsed(); |
| 2685 | assert!(error.downcast_ref::<RuntimeEventLockTimeout>().is_some()); |
| 2686 | assert!( |
| 2687 | elapsed >= Duration::from_millis(75), |
| 2688 | "returned too early: {elapsed:?}" |
| 2689 | ); |
| 2690 | assert!( |
| 2691 | elapsed < Duration::from_secs(2), |
| 2692 | "timeout was unbounded: {elapsed:?}" |
| 2693 | ); |
| 2694 | assert_eq!(std::fs::read(&store.state_path)?, state_before); |
| 2695 | assert!(!event_path.exists()); |
| 2696 | |
| 2697 | let status = child.kill_and_reap()?; |
| 2698 | assert!( |
| 2699 | !status.success(), |
| 2700 | "killed lock holder unexpectedly succeeded" |
| 2701 | ); |
| 2702 | let runtime = tokio::runtime::Builder::new_current_thread() |
| 2703 | .enable_all() |
| 2704 | .build()?; |
| 2705 | let committed = runtime.block_on(store.append_event( |
| 2706 | thread_id, |
| 2707 | None, |
| 2708 | None, |
| 2709 | "after.holder.death", |
| 2710 | json!({}), |
| 2711 | ))?; |
| 2712 | assert_eq!(committed.seq, 1); |
| 2713 | |
| 2714 | std::fs::remove_dir_all(dir)?; |
| 2715 | Ok(()) |
| 2716 | } |
| 2717 | |
| 2718 | #[test] |
| 2719 | fn killed_torn_append_is_repaired_and_preserves_reserved_gap() -> Result<()> { |
| 2720 | let dir = test_runtime_dir(); |
| 2721 | let store = RuntimeThreadStore::open(dir.clone())?; |
| 2722 | let thread_id = "thr_process_torn"; |
| 2723 | let signal = dir.join("process-torn.ready"); |
| 2724 | let child = spawn_runtime_event_child("torn", &dir, thread_id, &signal); |
| 2725 | wait_for_runtime_event_test_file(&signal, "durable torn Runtime event tail"); |
| 2726 | let status = child.kill_and_reap()?; |
| 2727 | assert!( |
| 2728 | !status.success(), |
| 2729 | "killed torn writer unexpectedly succeeded" |
| 2730 | ); |
| 2731 | |
| 2732 | let runtime = tokio::runtime::Builder::new_current_thread() |
| 2733 | .enable_all() |
| 2734 | .build()?; |
| 2735 | let committed = runtime.block_on(store.append_event( |
| 2736 | thread_id, |
| 2737 | None, |
| 2738 | None, |
| 2739 | "after.torn.death", |
| 2740 | json!({}), |
| 2741 | ))?; |
| 2742 | assert_eq!(committed.seq, 2); |
| 2743 | assert_eq!(runtime.block_on(store.current_seq())?, 2); |
| 2744 | let events = store.events_since(thread_id, None)?; |
| 2745 | assert_eq!(events.len(), 1); |
| 2746 | assert_eq!(events[0].seq, 2); |
| 2747 | let raw = std::fs::read(store.events_path(thread_id)?)?; |
| 2748 | assert_eq!(raw.last(), Some(&b'\n')); |
| 2749 | for line in raw |
| 2750 | .split(|byte| *byte == b'\n') |
| 2751 | .filter(|line| !line.is_empty()) |
| 2752 | { |
| 2753 | serde_json::from_slice::<RuntimeEventRecord>(line)?; |
| 2754 | } |
| 2755 | |
| 2756 | std::fs::remove_dir_all(dir)?; |
| 2757 | Ok(()) |
| 2758 | } |
| 2759 | |
| 2760 | #[test] |
| 2761 | fn event_reader_never_observes_cross_process_rollback_candidate() -> Result<()> { |
| 2762 | let dir = test_runtime_dir(); |
| 2763 | let store = RuntimeThreadStore::open(dir.clone())?; |
| 2764 | let thread_id = "thr_process_phantom"; |
| 2765 | let signal = dir.join("process-phantom.ready"); |
| 2766 | let child = spawn_runtime_event_child("phantom", &dir, thread_id, &signal); |
| 2767 | wait_for_runtime_event_test_file(&signal, "visible rollback candidate"); |
| 2768 | |
| 2769 | let started = Instant::now(); |
| 2770 | assert!(store.events_since(thread_id, None)?.is_empty()); |
| 2771 | assert!( |
| 2772 | started.elapsed() >= Duration::from_millis(200), |
| 2773 | "reader did not wait for rollback disposition" |
| 2774 | ); |
| 2775 | child.wait_success("phantom rollback child"); |
| 2776 | assert!(store.events_since(thread_id, None)?.is_empty()); |
| 2777 | |
| 2778 | std::fs::remove_dir_all(dir)?; |
| 2779 | Ok(()) |
| 2780 | } |
| 2781 | |
| 2782 | #[tokio::test(flavor = "current_thread")] |
| 2783 | async fn event_history_lock_wait_does_not_block_tokio_worker() -> Result<()> { |
| 2784 | let dir = test_runtime_dir(); |
| 2785 | let manager = test_manager(dir.clone())?; |
| 2786 | let thread_id = "thr_async_history_wait"; |
| 2787 | let signal = dir.join("async-history-holder.ready"); |
| 2788 | let child = spawn_runtime_event_child("holder", &dir, thread_id, &signal); |
| 2789 | wait_for_runtime_event_test_file(&signal, "async history lock holder"); |
| 2790 | |
| 2791 | let history = manager.events_since_async(thread_id, None); |
| 2792 | tokio::pin!(history); |
| 2793 | tokio::select! { |
| 2794 | result = &mut history => panic!("history lock wait returned early: {result:?}"), |
| 2795 | () = sleep(Duration::from_millis(50)) => {} |
| 2796 | } |
| 2797 | let status = child.kill_and_reap()?; |
| 2798 | assert!(!status.success(), "killed history lock holder succeeded"); |
| 2799 | assert!(history.await?.is_empty()); |
| 2800 | |
| 2801 | std::fs::remove_dir_all(dir)?; |
| 2802 | Ok(()) |
| 2803 | } |
| 2804 | |
| 2805 | #[test] |
| 2806 | fn runtime_event_sequence_exhaustion_fails_before_mutation() -> Result<()> { |
| 2807 | let dir = test_runtime_dir(); |
| 2808 | let store = RuntimeThreadStore::open(dir.clone())?; |
| 2809 | store.with_event_transaction(Duration::from_secs(1), || { |
| 2810 | write_json_atomic( |
| 2811 | &store.state_path, |
| 2812 | &RuntimeStoreState { |
| 2813 | schema_version: CURRENT_RUNTIME_SCHEMA_VERSION, |
| 2814 | next_seq: u64::MAX, |
| 2815 | }, |
| 2816 | ) |
| 2817 | })?; |
| 2818 | let state_before = std::fs::read(&store.state_path)?; |
| 2819 | let thread_id = "thr_sequence_exhausted"; |
| 2820 | let error = store |
| 2821 | .append_event_transaction( |
| 2822 | thread_id.to_string(), |
| 2823 | None, |
| 2824 | None, |
| 2825 | "must.not.append".to_string(), |
| 2826 | json!({}), |
| 2827 | Duration::from_secs(1), |
| 2828 | ) |
| 2829 | .expect_err("u64 sequence exhaustion must fail closed"); |
| 2830 | assert!( |
| 2831 | format!("{error:#}").contains("Runtime event sequence exhausted"), |
| 2832 | "unexpected exhaustion error: {error:#}" |
| 2833 | ); |
| 2834 | assert_eq!(std::fs::read(&store.state_path)?, state_before); |
| 2835 | assert!(!store.events_path(thread_id)?.exists()); |
| 2836 | |
| 2837 | std::fs::remove_dir_all(dir)?; |
| 2838 | Ok(()) |
| 2839 | } |
| 2840 | |
| 2841 | #[tokio::test] |
| 2842 | async fn store_open_truncates_only_torn_final_event_record_and_preserves_sequence_gap() { |
| 2843 | let dir = test_runtime_dir(); |
| 2844 | let store = RuntimeThreadStore::open(dir.clone()).expect("open store"); |
| 2845 | let first = store |
| 2846 | .append_event("thr_torn_tail", None, None, "first", json!({ "value": 1 })) |
| 2847 | .await |
| 2848 | .expect("append first event"); |
| 2849 | let torn = store |
| 2850 | .append_event("thr_torn_tail", None, None, "torn", json!({ "value": 2 })) |
| 2851 | .await |
| 2852 | .expect("append event to tear"); |
| 2853 | let path = store.events_path("thr_torn_tail").expect("event path"); |
| 2854 | let original_len = std::fs::metadata(&path).expect("event metadata").len(); |
| 2855 | assert!(original_len > 16); |
| 2856 | std::fs::OpenOptions::new() |
| 2857 | .write(true) |
| 2858 | .open(&path) |
| 2859 | .expect("open event log for simulated crash") |
| 2860 | .set_len(original_len - 16) |
| 2861 | .expect("tear final event record"); |
| 2862 | drop(store); |
| 2863 | |
| 2864 | let reopened = RuntimeThreadStore::open(dir.clone()).expect("repair torn event tail"); |
| 2865 | let replay = reopened |
| 2866 | .events_since("thr_torn_tail", None) |
| 2867 | .expect("replay repaired event log"); |
| 2868 | assert_eq!(replay.len(), 1); |
| 2869 | assert_eq!(replay[0].seq, first.seq); |
| 2870 | |
| 2871 | let after_repair = reopened |
| 2872 | .append_event( |
| 2873 | "thr_torn_tail", |
| 2874 | None, |
| 2875 | None, |
| 2876 | "after_repair", |
| 2877 | json!({ "value": 3 }), |
| 2878 | ) |
| 2879 | .await |
| 2880 | .expect("append after repair"); |
| 2881 | assert_eq!(after_repair.seq, torn.seq.saturating_add(1)); |
| 2882 | assert_eq!( |
| 2883 | reopened |
| 2884 | .events_since("thr_torn_tail", None) |
| 2885 | .expect("replay repaired and appended events") |
| 2886 | .iter() |
| 2887 | .map(|event| event.seq) |
| 2888 | .collect::<Vec<_>>(), |
| 2889 | vec![first.seq, after_repair.seq] |
| 2890 | ); |
| 2891 | |
| 2892 | let _ = std::fs::remove_dir_all(dir); |
| 2893 | } |
| 2894 | |
| 2895 | #[tokio::test] |
| 2896 | async fn store_open_treats_valid_json_without_newline_as_uncommitted_append() { |
| 2897 | let dir = test_runtime_dir(); |
| 2898 | let store = RuntimeThreadStore::open(dir.clone()).expect("open store"); |
| 2899 | let committed = store |
| 2900 | .append_event( |
| 2901 | "thr_missing_commit_marker", |
| 2902 | None, |
| 2903 | None, |
| 2904 | "committed", |
| 2905 | json!({ "value": 1 }), |
| 2906 | ) |
| 2907 | .await |
| 2908 | .expect("append committed event"); |
| 2909 | let uncommitted = store |
| 2910 | .append_event( |
| 2911 | "thr_missing_commit_marker", |
| 2912 | None, |
| 2913 | None, |
| 2914 | "missing_marker", |
| 2915 | json!({ "value": 2 }), |
| 2916 | ) |
| 2917 | .await |
| 2918 | .expect("append event whose commit marker will be removed"); |
| 2919 | let path = store |
| 2920 | .events_path("thr_missing_commit_marker") |
| 2921 | .expect("event path"); |
| 2922 | let encoded = std::fs::read(&path).expect("read event log"); |
| 2923 | assert_eq!(encoded.last(), Some(&b'\n')); |
| 2924 | let without_marker = &encoded[..encoded.len() - 1]; |
| 2925 | let final_json = without_marker |
| 2926 | .rsplit(|byte| *byte == b'\n') |
| 2927 | .next() |
| 2928 | .expect("final JSON record"); |
| 2929 | serde_json::from_slice::<RuntimeEventRecord>(final_json) |
| 2930 | .expect("the unterminated tail must otherwise be valid JSON"); |
| 2931 | std::fs::OpenOptions::new() |
| 2932 | .write(true) |
| 2933 | .open(&path) |
| 2934 | .expect("open event log for simulated crash") |
| 2935 | .set_len(u64::try_from(without_marker.len()).expect("event log length fits u64")) |
| 2936 | .expect("remove only the newline commit marker"); |
| 2937 | drop(store); |
| 2938 | |
| 2939 | let reopened = RuntimeThreadStore::open(dir.clone()).expect("repair uncommitted event tail"); |
| 2940 | let replay = reopened |
| 2941 | .events_since("thr_missing_commit_marker", None) |
| 2942 | .expect("replay repaired event log"); |
| 2943 | assert_eq!(replay.len(), 1); |
| 2944 | assert_eq!(replay[0].seq, committed.seq); |
| 2945 | |
| 2946 | let after_repair = reopened |
| 2947 | .append_event( |
| 2948 | "thr_missing_commit_marker", |
| 2949 | None, |
| 2950 | None, |
| 2951 | "after_repair", |
| 2952 | json!({ "value": 3 }), |
| 2953 | ) |
| 2954 | .await |
| 2955 | .expect("append after repair"); |
| 2956 | assert_eq!(after_repair.seq, uncommitted.seq.saturating_add(1)); |
| 2957 | |
| 2958 | let _ = std::fs::remove_dir_all(dir); |
| 2959 | } |
| 2960 | |
| 2961 | #[tokio::test] |
| 2962 | async fn store_open_does_not_discard_newline_terminated_malformed_event() { |
| 2963 | let dir = test_runtime_dir(); |
| 2964 | let store = RuntimeThreadStore::open(dir.clone()).expect("open store"); |
| 2965 | store |
| 2966 | .append_event("thr_bad_tail", None, None, "valid", json!({})) |
| 2967 | .await |
| 2968 | .expect("append valid event"); |
| 2969 | let path = store.events_path("thr_bad_tail").expect("event path"); |
| 2970 | let mut file = std::fs::OpenOptions::new() |
| 2971 | .append(true) |
| 2972 | .open(&path) |
| 2973 | .expect("open event log"); |
| 2974 | std::io::Write::write_all(&mut file, b"{malformed-but-terminated}\n") |
| 2975 | .expect("append malformed event"); |
| 2976 | std::io::Write::flush(&mut file).expect("flush malformed event"); |
| 2977 | drop(file); |
| 2978 | drop(store); |
| 2979 | |
| 2980 | let reopened = RuntimeThreadStore::open(dir.clone()).expect("open terminated event log"); |
| 2981 | let error = reopened |
| 2982 | .events_since("thr_bad_tail", None) |
| 2983 | .expect_err("terminated malformed event must fail closed"); |
| 2984 | assert!( |
| 2985 | format!("{error:#}").contains("Failed to parse event line"), |
| 2986 | "unexpected replay error: {error:#}" |
| 2987 | ); |
| 2988 | |
| 2989 | let _ = std::fs::remove_dir_all(dir); |
| 2990 | } |
| 2991 | |
| 2992 | #[tokio::test] |
| 2993 | async fn event_replay_is_bounded_and_tail_cursor_skips_only_omitted_history() -> Result<()> { |
| 2994 | let manager = test_manager(test_runtime_dir())?; |
| 2995 | let thread_id = "thr_bounded_replay"; |
| 2996 | let path = manager.store.events_path(thread_id)?; |
| 2997 | let mut encoded = Vec::new(); |
| 2998 | for seq in 1_u64..=600 { |
| 2999 | let event = RuntimeEventRecord { |
| 3000 | schema_version: CURRENT_RUNTIME_SCHEMA_VERSION, |
| 3001 | seq, |
| 3002 | timestamp: Utc::now(), |
| 3003 | thread_id: thread_id.to_string(), |
| 3004 | turn_id: None, |
| 3005 | item_id: None, |
| 3006 | event: "test.event".to_string(), |
| 3007 | payload: json!({ "seq": seq }), |
| 3008 | }; |
| 3009 | serde_json::to_writer(&mut encoded, &event)?; |
| 3010 | encoded.push(b'\n'); |
| 3011 | } |
| 3012 | std::fs::write(path, encoded)?; |
| 3013 | |
| 3014 | let mut full = manager.replay_events(thread_id, None, None).await?; |
| 3015 | assert_eq!(full.base_seq, 0); |
| 3016 | let mut full_sequences = Vec::new(); |
| 3017 | while let Some(batch) = full.batches.recv().await { |
| 3018 | let batch = batch.map_err(anyhow::Error::msg)?; |
| 3019 | assert!( |
| 3020 | batch.len() <= RUNTIME_EVENT_REPLAY_BATCH_SIZE, |
| 3021 | "replay batch exceeded its memory bound" |
| 3022 | ); |
| 3023 | full_sequences.extend(batch.into_iter().map(|event| event.seq)); |
| 3024 | } |
| 3025 | assert_eq!(full_sequences, (1_u64..=600).collect::<Vec<_>>()); |
| 3026 | |
| 3027 | let mut tail = manager.replay_events(thread_id, None, Some(10)).await?; |
| 3028 | assert_eq!(tail.base_seq, 590); |
| 3029 | let mut tail_sequences = Vec::new(); |
| 3030 | while let Some(batch) = tail.batches.recv().await { |
| 3031 | tail_sequences.extend( |
| 3032 | batch |
| 3033 | .map_err(anyhow::Error::msg)? |
| 3034 | .into_iter() |
| 3035 | .map(|event| event.seq), |
| 3036 | ); |
| 3037 | } |
| 3038 | assert_eq!(tail_sequences, (591_u64..=600).collect::<Vec<_>>()); |
| 3039 | |
| 3040 | let mut empty_tail = manager.replay_events(thread_id, None, Some(0)).await?; |
| 3041 | assert_eq!(empty_tail.base_seq, 600); |
| 3042 | assert!(empty_tail.batches.recv().await.is_none()); |
| 3043 | assert!( |
| 3044 | manager |
| 3045 | .replay_events( |
| 3046 | thread_id, |
| 3047 | None, |
| 3048 | Some(MAX_RUNTIME_EVENT_REPLAY_TAIL.saturating_add(1)), |
| 3049 | ) |
| 3050 | .await |
| 3051 | .is_err() |
| 3052 | ); |
| 3053 | Ok(()) |
| 3054 | } |
| 3055 | |
| 3056 | #[tokio::test] |
| 3057 | async fn event_reader_ignores_an_unterminated_live_append_tail() -> Result<()> { |
| 3058 | let dir = test_runtime_dir(); |
| 3059 | let store = RuntimeThreadStore::open(dir.clone())?; |
| 3060 | let committed = store |
| 3061 | .append_event("thr_live_tail", None, None, "committed", json!({})) |
| 3062 | .await?; |
| 3063 | let path = store.events_path("thr_live_tail")?; |
| 3064 | let mut file = std::fs::OpenOptions::new().append(true).open(path)?; |
| 3065 | std::io::Write::write_all(&mut file, br#"{"schema_version":2,"seq":999"#)?; |
| 3066 | std::io::Write::flush(&mut file)?; |
| 3067 | |
| 3068 | let replay = store.events_since("thr_live_tail", None)?; |
| 3069 | assert_eq!(replay.len(), 1); |
| 3070 | assert_eq!(replay[0].seq, committed.seq); |
| 3071 | let _ = std::fs::remove_dir_all(dir); |
| 3072 | Ok(()) |
| 3073 | } |
| 3074 | |
| 3075 | #[cfg(unix)] |
| 3076 | #[test] |
| 3077 | fn store_open_rejects_symlinked_state_file() { |
| 3078 | let dir = test_runtime_dir(); |
| 3079 | std::fs::create_dir_all(&dir).expect("mkdir runtime dir"); |
| 3080 | let target = dir.join("outside-state.json"); |
| 3081 | let link = dir.join("state.json"); |
| 3082 | std::fs::write( |
| 3083 | &target, |
| 3084 | serde_json::to_string(&RuntimeStoreState::default()).unwrap(), |
| 3085 | ) |
| 3086 | .expect("write target"); |
| 3087 | std::os::unix::fs::symlink(&target, &link).expect("symlink state"); |
| 3088 | |
| 3089 | let err = RuntimeThreadStore::open(dir.clone()).expect_err("symlink state should fail"); |
| 3090 | assert!(format!("{err:#}").contains("must not be a symlink")); |
| 3091 | |
| 3092 | let _ = std::fs::remove_dir_all(dir); |
| 3093 | } |
| 3094 | |
| 3095 | #[cfg(unix)] |
| 3096 | #[test] |
| 3097 | fn event_append_rollback_handle_rejects_a_swapped_symlink_target() -> Result<()> { |
| 3098 | let dir = test_runtime_dir(); |
| 3099 | std::fs::create_dir_all(&dir)?; |
| 3100 | let outside = dir.join("outside.jsonl"); |
| 3101 | let events_path = dir.join("events.jsonl"); |
| 3102 | std::fs::write(&outside, b"must-remain-intact\n")?; |
| 3103 | std::os::unix::fs::symlink(&outside, &events_path)?; |
| 3104 | |
| 3105 | let error = open_runtime_store_file(&events_path, "Runtime event rollback", |options| { |
| 3106 | options.write(true); |
| 3107 | }) |
| 3108 | .expect_err("rollback followed a swapped symlink target"); |
| 3109 | assert!(format!("{error:#}").contains("must not be a symlink")); |
| 3110 | assert_eq!(std::fs::read(&outside)?, b"must-remain-intact\n"); |
| 3111 | std::fs::remove_dir_all(&dir)?; |
| 3112 | Ok(()) |
| 3113 | } |
| 3114 | |
| 3115 | #[test] |
| 3116 | fn store_open_rejects_root_traversal() { |
| 3117 | let dir = test_runtime_dir(); |
| 3118 | let bad_root = dir.join("runtime").join("..").join("outside"); |
| 3119 | |
| 3120 | let err = RuntimeThreadStore::open(bad_root).expect_err("traversal root should fail"); |
| 3121 | assert!(format!("{err:#}").contains("cannot contain '..'")); |
| 3122 | |
| 3123 | let _ = std::fs::remove_dir_all(dir); |
| 3124 | } |
| 3125 | |
| 3126 | #[cfg(unix)] |
| 3127 | #[test] |
| 3128 | fn store_open_rejects_symlinked_store_directory() { |
| 3129 | let dir = test_runtime_dir(); |
| 3130 | std::fs::create_dir_all(&dir).expect("mkdir runtime dir"); |
| 3131 | let outside = dir.join("outside-items"); |
| 3132 | let link = dir.join("items"); |
| 3133 | std::fs::create_dir_all(&outside).expect("mkdir outside"); |
| 3134 | std::os::unix::fs::symlink(&outside, &link).expect("symlink items dir"); |
| 3135 | |
| 3136 | let err = RuntimeThreadStore::open(dir.clone()).expect_err("symlink items dir should fail"); |
| 3137 | assert!( |
| 3138 | format!("{err:#}").contains("directory must not be a symlink"), |
| 3139 | "got: {err:#}" |
| 3140 | ); |
| 3141 | |
| 3142 | let _ = std::fs::remove_dir_all(dir); |
| 3143 | } |
| 3144 | |
| 3145 | #[cfg(unix)] |
| 3146 | #[test] |
| 3147 | fn store_list_items_rejects_symlinked_item_file() { |
| 3148 | let dir = test_runtime_dir(); |
| 3149 | let store = RuntimeThreadStore::open(dir.clone()).expect("open store"); |
| 3150 | let item = sample_item("turn_link", "item_link", TurnItemLifecycleStatus::Completed); |
| 3151 | let target = dir.join("outside-item.json"); |
| 3152 | let link = store.items_dir.join(format!("{}.json", item.id)); |
| 3153 | std::fs::write(&target, serde_json::to_string(&item).unwrap()).expect("write target"); |
| 3154 | std::os::unix::fs::symlink(&target, &link).expect("symlink item"); |
| 3155 | |
| 3156 | let err = store |
| 3157 | .list_items_for_turn(&item.turn_id) |
| 3158 | .expect_err("symlink item should fail"); |
| 3159 | assert!(format!("{err:#}").contains("must not be a symlink")); |
| 3160 | |
| 3161 | let _ = std::fs::remove_dir_all(dir); |
| 3162 | } |
| 3163 | |
| 3164 | #[cfg(unix)] |
| 3165 | #[test] |
| 3166 | fn store_list_items_rejects_swapped_symlinked_store_directory() { |
| 3167 | let dir = test_runtime_dir(); |
| 3168 | let store = RuntimeThreadStore::open(dir.clone()).expect("open store"); |
| 3169 | let outside = dir.join("outside-items"); |
| 3170 | std::fs::create_dir_all(&outside).expect("mkdir outside"); |
| 3171 | std::fs::remove_dir_all(&store.items_dir).expect("remove items dir"); |
| 3172 | std::os::unix::fs::symlink(&outside, &store.items_dir).expect("symlink items dir"); |
| 3173 | |
| 3174 | let err = store |
| 3175 | .list_items_for_turn("turn_link") |
| 3176 | .expect_err("swapped symlink items dir should fail"); |
| 3177 | assert!( |
| 3178 | format!("{err:#}").contains("directory must not be a symlink"), |
| 3179 | "got: {err:#}" |
| 3180 | ); |
| 3181 | |
| 3182 | let _ = std::fs::remove_dir_all(dir); |
| 3183 | } |
| 3184 | |
| 3185 | #[test] |
| 3186 | fn store_load_thread_defaults_missing_session_id() { |
| 3187 | let dir = test_runtime_dir(); |
| 3188 | let store = RuntimeThreadStore::open(dir.clone()).expect("open store"); |
| 3189 | let thread = sample_thread("thr_legacy_session"); |
| 3190 | let path = store.threads_dir.join(format!("{}.json", thread.id)); |
| 3191 | std::fs::create_dir_all(path.parent().unwrap()).expect("mkdirs"); |
| 3192 | let mut payload = serde_json::to_value(&thread).expect("serialize thread"); |
| 3193 | payload |
| 3194 | .as_object_mut() |
| 3195 | .expect("thread object") |
| 3196 | .remove("session_id"); |
| 3197 | std::fs::write( |
| 3198 | &path, |
| 3199 | serde_json::to_string(&payload).expect("encode thread"), |
| 3200 | ) |
| 3201 | .expect("write thread"); |
| 3202 | |
| 3203 | let loaded = store |
| 3204 | .load_thread(&thread.id) |
| 3205 | .expect("legacy thread should load"); |
| 3206 | assert_eq!(loaded.session_id, None); |
| 3207 | |
| 3208 | let _ = std::fs::remove_dir_all(dir); |
| 3209 | } |
| 3210 | |
| 3211 | #[tokio::test] |
| 3212 | async fn seed_thread_keeps_tool_results_on_preceding_turn() -> Result<()> { |
| 3213 | let dir = test_runtime_dir(); |
| 3214 | let manager = test_manager(dir.clone())?; |
| 3215 | let thread = sample_thread("thr_seed_blocks"); |
| 3216 | manager.store.save_thread(&thread)?; |
| 3217 | let messages = vec![ |
| 3218 | Message { |
| 3219 | role: "user".to_string(), |
| 3220 | content: vec![ContentBlock::Text { |
| 3221 | text: "check the files".to_string(), |
| 3222 | cache_control: None, |
| 3223 | }], |
| 3224 | }, |
| 3225 | Message { |
| 3226 | role: "assistant".to_string(), |
| 3227 | content: vec![ |
| 3228 | ContentBlock::Thinking { |
| 3229 | thinking: "need a tool".to_string(), |
| 3230 | signature: Some("sig-1".to_string()), |
| 3231 | }, |
| 3232 | ContentBlock::ToolUse { |
| 3233 | id: "tool-1".to_string(), |
| 3234 | name: "shell".to_string(), |
| 3235 | input: json!({ "cmd": "one" }), |
| 3236 | caller: None, |
| 3237 | }, |
| 3238 | ContentBlock::ToolUse { |
| 3239 | id: "tool-2".to_string(), |
| 3240 | name: "shell".to_string(), |
| 3241 | input: json!({ "cmd": "two" }), |
| 3242 | caller: None, |
| 3243 | }, |
| 3244 | ], |
| 3245 | }, |
| 3246 | Message { |
| 3247 | role: "user".to_string(), |
| 3248 | content: vec![ContentBlock::ToolResult { |
| 3249 | tool_use_id: "tool-1".to_string(), |
| 3250 | content: "one".to_string(), |
| 3251 | is_error: None, |
| 3252 | content_blocks: Some(vec![json!({ |
| 3253 | "type": "text", |
| 3254 | "text": "structured one" |
| 3255 | })]), |
| 3256 | }], |
| 3257 | }, |
| 3258 | Message { |
| 3259 | role: "user".to_string(), |
| 3260 | content: vec![ContentBlock::ToolResult { |
| 3261 | tool_use_id: "tool-2".to_string(), |
| 3262 | content: "two".to_string(), |
| 3263 | is_error: Some(true), |
| 3264 | content_blocks: None, |
| 3265 | }], |
| 3266 | }, |
| 3267 | Message { |
| 3268 | role: "assistant".to_string(), |
| 3269 | content: vec![ContentBlock::Text { |
| 3270 | text: "done".to_string(), |
| 3271 | cache_control: None, |
| 3272 | }], |
| 3273 | }, |
| 3274 | ]; |
| 3275 | |
| 3276 | manager |
| 3277 | .seed_thread_from_messages(&thread.id, &messages) |
| 3278 | .await?; |
| 3279 | let turns = manager.store.list_turns_for_thread(&thread.id)?; |
| 3280 | assert_eq!(turns.len(), 1); |
| 3281 | |
| 3282 | let restored = manager.reconstruct_messages_from_turns(&turns)?; |
| 3283 | let roles = restored |
| 3284 | .iter() |
| 3285 | .map(|message| message.role.as_str()) |
| 3286 | .collect::<Vec<_>>(); |
| 3287 | assert_eq!(roles, vec!["user", "assistant", "user", "assistant"]); |
| 3288 | assert_eq!(restored[2].content.len(), 2); |
| 3289 | |
| 3290 | match &restored[2].content[0] { |
| 3291 | ContentBlock::ToolResult { |
| 3292 | tool_use_id, |
| 3293 | content, |
| 3294 | is_error, |
| 3295 | content_blocks, |
| 3296 | } => { |
| 3297 | assert_eq!(tool_use_id, "tool-1"); |
| 3298 | assert_eq!(content, "one"); |
| 3299 | assert_eq!(*is_error, None); |
| 3300 | assert_eq!( |
| 3301 | content_blocks |
| 3302 | .as_ref() |
| 3303 | .and_then(|blocks| blocks[0].get("text")), |
| 3304 | Some(&json!("structured one")) |
| 3305 | ); |
| 3306 | } |
| 3307 | other => panic!("expected first tool result, got {other:?}"), |
| 3308 | } |
| 3309 | match &restored[2].content[1] { |
| 3310 | ContentBlock::ToolResult { |
| 3311 | tool_use_id, |
| 3312 | content, |
| 3313 | is_error, |
| 3314 | content_blocks, |
| 3315 | } => { |
| 3316 | assert_eq!(tool_use_id, "tool-2"); |
| 3317 | assert_eq!(content, "two"); |
| 3318 | assert_eq!(*is_error, Some(true)); |
| 3319 | assert!(content_blocks.is_none()); |
| 3320 | } |
| 3321 | other => panic!("expected second tool result, got {other:?}"), |
| 3322 | } |
| 3323 | |
| 3324 | let _ = std::fs::remove_dir_all(dir); |
| 3325 | Ok(()) |
| 3326 | } |
| 3327 | |
| 3328 | #[test] |
| 3329 | fn current_runtime_schema_version_is_two_on_v066() { |
| 3330 | // Locks the bump in (issue #124). Bump deliberately when persisted |
| 3331 | // shape changes. |
| 3332 | assert_eq!(CURRENT_RUNTIME_SCHEMA_VERSION, 2); |
| 3333 | } |
| 3334 | |
| 3335 | #[test] |
| 3336 | fn store_rejects_path_like_record_ids() { |
| 3337 | let dir = test_runtime_dir(); |
| 3338 | let store = RuntimeThreadStore::open(dir.clone()).expect("open store"); |
| 3339 | |
| 3340 | let err = store |
| 3341 | .load_thread("../outside") |
| 3342 | .expect_err("path traversal id should fail"); |
| 3343 | assert!( |
| 3344 | format!("{err:#}").contains("unsupported characters"), |
| 3345 | "got: {err:#}" |
| 3346 | ); |
| 3347 | |
| 3348 | let mut thread = sample_thread("thr_bad/id"); |
| 3349 | let err = store |
| 3350 | .save_thread(&thread) |
| 3351 | .expect_err("path separator id should fail"); |
| 3352 | assert!( |
| 3353 | format!("{err:#}").contains("unsupported characters"), |
| 3354 | "got: {err:#}" |
| 3355 | ); |
| 3356 | |
| 3357 | thread.id = " thr_bad".to_string(); |
| 3358 | let err = store |
| 3359 | .save_thread(&thread) |
| 3360 | .expect_err("whitespace id should fail"); |
| 3361 | assert!(format!("{err:#}").contains("whitespace"), "got: {err:#}"); |
| 3362 | |
| 3363 | let _ = std::fs::remove_dir_all(dir); |
| 3364 | } |
| 3365 | |
| 3366 | #[test] |
| 3367 | fn store_load_turn_rejects_newer_schema_version() { |
| 3368 | let dir = test_runtime_dir(); |
| 3369 | let store = RuntimeThreadStore::open(dir.clone()).expect("open store"); |
| 3370 | |
| 3371 | let mut turn = sample_turn("thr_t", "trn_future", RuntimeTurnStatus::InProgress); |
| 3372 | turn.schema_version = CURRENT_RUNTIME_SCHEMA_VERSION + 1; |
| 3373 | |
| 3374 | let path = store.turns_dir.join(format!("{}.json", turn.id)); |
| 3375 | std::fs::create_dir_all(path.parent().unwrap()).expect("mkdirs"); |
| 3376 | std::fs::write(&path, serde_json::to_string(&turn).expect("serialize turn")) |
| 3377 | .expect("write turn"); |
| 3378 | |
| 3379 | let err = store |
| 3380 | .load_turn(&turn.id) |
| 3381 | .expect_err("load_turn must reject newer schema"); |
| 3382 | assert!( |
| 3383 | format!("{err:#}").contains("newer than supported"), |
| 3384 | "got: {err:#}" |
| 3385 | ); |
| 3386 | |
| 3387 | let _ = std::fs::remove_dir_all(dir); |
| 3388 | } |
| 3389 | |
| 3390 | #[test] |
| 3391 | fn store_load_item_rejects_newer_schema_version() { |
| 3392 | let dir = test_runtime_dir(); |
| 3393 | let store = RuntimeThreadStore::open(dir.clone()).expect("open store"); |
| 3394 | |
| 3395 | let mut item = sample_item("trn_t", "itm_future", TurnItemLifecycleStatus::InProgress); |
| 3396 | item.schema_version = CURRENT_RUNTIME_SCHEMA_VERSION + 1; |
| 3397 | |
| 3398 | let path = store.items_dir.join(format!("{}.json", item.id)); |
| 3399 | std::fs::create_dir_all(path.parent().unwrap()).expect("mkdirs"); |
| 3400 | std::fs::write(&path, serde_json::to_string(&item).expect("serialize item")) |
| 3401 | .expect("write item"); |
| 3402 | |
| 3403 | let err = store |
| 3404 | .load_item(&item.id) |
| 3405 | .expect_err("load_item must reject newer schema"); |
| 3406 | assert!( |
| 3407 | format!("{err:#}").contains("newer than supported"), |
| 3408 | "got: {err:#}" |
| 3409 | ); |
| 3410 | |
| 3411 | let _ = std::fs::remove_dir_all(dir); |
| 3412 | } |
| 3413 | |
| 3414 | #[test] |
| 3415 | fn enforce_lru_capacity_does_not_loop_when_all_threads_are_active() { |
| 3416 | let mut active = ActiveThreads::default(); |
| 3417 | let harness_a = mock_engine_handle(); |
| 3418 | let harness_b = mock_engine_handle(); |
| 3419 | |
| 3420 | active.engines.insert( |
| 3421 | "thr_a".to_string(), |
| 3422 | ActiveThreadState { |
| 3423 | engine: harness_a.handle, |
| 3424 | active_turn: Some(ActiveTurnState { |
| 3425 | turn_id: "turn_a".to_string(), |
| 3426 | interrupt_requested: false, |
| 3427 | }), |
| 3428 | route_identity: crate::config::ProviderIdentity { |
| 3429 | provider: ApiProvider::Deepseek, |
| 3430 | key: "deepseek".to_string(), |
| 3431 | exact_id: Some("deepseek".to_string()), |
| 3432 | }, |
| 3433 | route_model: DEFAULT_TEXT_MODEL.to_string(), |
| 3434 | client_preflight_required: false, |
| 3435 | }, |
| 3436 | ); |
| 3437 | active.engines.insert( |
| 3438 | "thr_b".to_string(), |
| 3439 | ActiveThreadState { |
| 3440 | engine: harness_b.handle, |
| 3441 | active_turn: Some(ActiveTurnState { |
| 3442 | turn_id: "turn_b".to_string(), |
| 3443 | interrupt_requested: false, |
| 3444 | }), |
| 3445 | route_identity: crate::config::ProviderIdentity { |
| 3446 | provider: ApiProvider::Deepseek, |
| 3447 | key: "deepseek".to_string(), |
| 3448 | exact_id: Some("deepseek".to_string()), |
| 3449 | }, |
| 3450 | route_model: DEFAULT_TEXT_MODEL.to_string(), |
| 3451 | client_preflight_required: false, |
| 3452 | }, |
| 3453 | ); |
| 3454 | active.lru.push_back("thr_a".to_string()); |
| 3455 | active.lru.push_back("thr_b".to_string()); |
| 3456 | |
| 3457 | let evicted = enforce_lru_capacity(&mut active, 2); |
| 3458 | assert!(evicted.is_empty(), "no idle threads should be evicted"); |
| 3459 | assert_eq!(active.engines.len(), 2); |
| 3460 | assert_eq!(active.lru.len(), 2); |
| 3461 | } |
| 3462 | |
| 3463 | #[test] |
| 3464 | fn approval_decision_keeps_trust_mode_out_of_tool_approval() { |
| 3465 | assert!(matches!( |
| 3466 | RuntimeThreadManager::approval_decision(false, false, false), |
| 3467 | RuntimeApprovalDecision::DenyTool |
| 3468 | )); |
| 3469 | assert!(matches!( |
| 3470 | RuntimeThreadManager::approval_decision(false, true, false), |
| 3471 | RuntimeApprovalDecision::DenyTool |
| 3472 | )); |
| 3473 | assert!(matches!( |
| 3474 | RuntimeThreadManager::approval_decision(true, false, false), |
| 3475 | RuntimeApprovalDecision::ApproveTool |
| 3476 | )); |
| 3477 | assert!(matches!( |
| 3478 | RuntimeThreadManager::approval_decision(true, false, true), |
| 3479 | RuntimeApprovalDecision::DenyTool |
| 3480 | )); |
| 3481 | assert!(matches!( |
| 3482 | RuntimeThreadManager::approval_decision(true, true, true), |
| 3483 | RuntimeApprovalDecision::RetryWithFullAccess |
| 3484 | )); |
| 3485 | } |
| 3486 | |
| 3487 | #[test] |
| 3488 | fn open_recovers_queued_and_in_progress_turns() -> Result<()> { |
| 3489 | let runtime_dir = test_runtime_dir(); |
| 3490 | let store = RuntimeThreadStore::open(runtime_dir.clone())?; |
| 3491 | let thread = sample_thread("thr_recover"); |
| 3492 | store.save_thread(&thread)?; |
| 3493 | |
| 3494 | let mut queued_turn = sample_turn(&thread.id, "turn_queued", RuntimeTurnStatus::Queued); |
| 3495 | let mut in_progress_turn = |
| 3496 | sample_turn(&thread.id, "turn_running", RuntimeTurnStatus::InProgress); |
| 3497 | let completed_turn = sample_turn(&thread.id, "turn_done", RuntimeTurnStatus::Completed); |
| 3498 | |
| 3499 | let queued_item = sample_item( |
| 3500 | &queued_turn.id, |
| 3501 | "item_queued", |
| 3502 | TurnItemLifecycleStatus::Queued, |
| 3503 | ); |
| 3504 | let in_progress_item = sample_item( |
| 3505 | &in_progress_turn.id, |
| 3506 | "item_running", |
| 3507 | TurnItemLifecycleStatus::InProgress, |
| 3508 | ); |
| 3509 | let completed_item = sample_item( |
| 3510 | &completed_turn.id, |
| 3511 | "item_done", |
| 3512 | TurnItemLifecycleStatus::Completed, |
| 3513 | ); |
| 3514 | |
| 3515 | queued_turn.item_ids = vec![queued_item.id.clone()]; |
| 3516 | in_progress_turn.item_ids = vec![in_progress_item.id.clone()]; |
| 3517 | |
| 3518 | store.save_item(&queued_item)?; |
| 3519 | store.save_item(&in_progress_item)?; |
| 3520 | store.save_item(&completed_item)?; |
| 3521 | store.save_turn(&queued_turn)?; |
| 3522 | store.save_turn(&in_progress_turn)?; |
| 3523 | store.save_turn(&completed_turn)?; |
| 3524 | |
| 3525 | let manager = test_manager(runtime_dir)?; |
| 3526 | |
| 3527 | let queued_turn = manager.store.load_turn(&queued_turn.id)?; |
| 3528 | assert_eq!(queued_turn.status, RuntimeTurnStatus::Interrupted); |
| 3529 | assert_eq!(queued_turn.error.as_deref(), Some(RUNTIME_RESTART_REASON)); |
| 3530 | assert!(queued_turn.ended_at.is_some()); |
| 3531 | assert!(queued_turn.duration_ms.is_some()); |
| 3532 | |
| 3533 | let in_progress_turn = manager.store.load_turn(&in_progress_turn.id)?; |
| 3534 | assert_eq!(in_progress_turn.status, RuntimeTurnStatus::Interrupted); |
| 3535 | assert_eq!( |
| 3536 | in_progress_turn.error.as_deref(), |
| 3537 | Some(RUNTIME_RESTART_REASON) |
| 3538 | ); |
| 3539 | assert!(in_progress_turn.ended_at.is_some()); |
| 3540 | assert!(in_progress_turn.duration_ms.is_some()); |
| 3541 | |
| 3542 | let completed_turn = manager.store.load_turn(&completed_turn.id)?; |
| 3543 | assert_eq!(completed_turn.status, RuntimeTurnStatus::Completed); |
| 3544 | assert!(completed_turn.error.is_none()); |
| 3545 | |
| 3546 | let queued_item = manager.store.load_item("item_queued")?; |
| 3547 | assert_eq!(queued_item.status, TurnItemLifecycleStatus::Interrupted); |
| 3548 | assert!(queued_item.ended_at.is_some()); |
| 3549 | |
| 3550 | let in_progress_item = manager.store.load_item("item_running")?; |
| 3551 | assert_eq!( |
| 3552 | in_progress_item.status, |
| 3553 | TurnItemLifecycleStatus::Interrupted |
| 3554 | ); |
| 3555 | assert!(in_progress_item.ended_at.is_some()); |
| 3556 | |
| 3557 | let completed_item = manager.store.load_item("item_done")?; |
| 3558 | assert_eq!(completed_item.status, TurnItemLifecycleStatus::Completed); |
| 3559 | |
| 3560 | Ok(()) |
| 3561 | } |
| 3562 | |
| 3563 | #[tokio::test] |
| 3564 | async fn thread_lifecycle_persists_across_restart() -> Result<()> { |
| 3565 | let runtime_dir = test_runtime_dir(); |
| 3566 | let manager = test_manager(runtime_dir.clone())?; |
| 3567 | let thread = manager |
| 3568 | .create_thread(CreateThreadRequest { |
| 3569 | model: None, |
| 3570 | workspace: None, |
| 3571 | mode: None, |
| 3572 | allow_shell: None, |
| 3573 | trust_mode: None, |
| 3574 | auto_approve: None, |
| 3575 | archived: false, |
| 3576 | system_prompt: None, |
| 3577 | task_id: None, |
| 3578 | ..Default::default() |
| 3579 | }) |
| 3580 | .await?; |
| 3581 | |
| 3582 | let harness = install_mock_engine(&manager, &thread.id).await; |
| 3583 | let mut rx_op = harness.rx_op; |
| 3584 | let tx_event = harness.tx_event; |
| 3585 | tokio::spawn(async move { |
| 3586 | if matches!(rx_op.recv().await, Some(Op::SendMessage { .. })) { |
| 3587 | let _ = tx_event |
| 3588 | .send(EngineEvent::TurnStarted { |
| 3589 | turn_id: "engine_turn_1".to_string(), |
| 3590 | created_at: chrono::Utc::now(), |
| 3591 | route: None, |
| 3592 | }) |
| 3593 | .await; |
| 3594 | let _ = tx_event |
| 3595 | .send(EngineEvent::MessageStarted { index: 0 }) |
| 3596 | .await; |
| 3597 | let _ = tx_event |
| 3598 | .send(EngineEvent::MessageDelta { |
| 3599 | index: 0, |
| 3600 | content: "mock response".to_string(), |
| 3601 | }) |
| 3602 | .await; |
| 3603 | let _ = tx_event |
| 3604 | .send(EngineEvent::MessageComplete { index: 0 }) |
| 3605 | .await; |
| 3606 | let _ = tx_event |
| 3607 | .send(EngineEvent::TurnComplete { |
| 3608 | usage: Usage { |
| 3609 | input_tokens: 10, |
| 3610 | output_tokens: 12, |
| 3611 | ..Usage::default() |
| 3612 | }, |
| 3613 | status: TurnOutcomeStatus::Completed, |
| 3614 | error: None, |
| 3615 | tool_catalog: None, |
| 3616 | base_url: None, |
| 3617 | }) |
| 3618 | .await; |
| 3619 | } |
| 3620 | }); |
| 3621 | |
| 3622 | let turn = manager |
| 3623 | .start_turn( |
| 3624 | &thread.id, |
| 3625 | StartTurnRequest { |
| 3626 | prompt: "first prompt".to_string(), |
| 3627 | input_summary: None, |
| 3628 | model: None, |
| 3629 | mode: None, |
| 3630 | allow_shell: None, |
| 3631 | trust_mode: None, |
| 3632 | auto_approve: None, |
| 3633 | ..Default::default() |
| 3634 | }, |
| 3635 | ) |
| 3636 | .await?; |
| 3637 | let completed = wait_for_terminal_turn(&manager, &turn.id, Duration::from_secs(2)).await?; |
| 3638 | assert_eq!(completed.status, RuntimeTurnStatus::Completed); |
| 3639 | |
| 3640 | drop(manager); |
| 3641 | |
| 3642 | let reopened = test_manager(runtime_dir)?; |
| 3643 | let detail = reopened.get_thread_detail(&thread.id).await?; |
| 3644 | assert_eq!(detail.thread.id, thread.id); |
| 3645 | assert_eq!(detail.turns.len(), 1); |
| 3646 | assert!(detail.latest_seq >= 1); |
| 3647 | assert!(!detail.items.is_empty()); |
| 3648 | let events = reopened.events_since(&thread.id, None)?; |
| 3649 | assert!( |
| 3650 | events.iter().any(|ev| ev.event == "turn.completed"), |
| 3651 | "expected turn.completed event after restart" |
| 3652 | ); |
| 3653 | Ok(()) |
| 3654 | } |
| 3655 | |
| 3656 | #[tokio::test] |
| 3657 | async fn monitor_separates_lifecycle_start_from_billing_dispatch_and_child_usage() -> Result<()> { |
| 3658 | let manager = test_manager(test_runtime_dir())?; |
| 3659 | let thread = manager |
| 3660 | .create_thread(CreateThreadRequest::default()) |
| 3661 | .await?; |
| 3662 | let mut harness = install_mock_engine(&manager, &thread.id).await; |
| 3663 | let turn = manager |
| 3664 | .start_turn( |
| 3665 | &thread.id, |
| 3666 | StartTurnRequest { |
| 3667 | prompt: "capture immutable route".to_string(), |
| 3668 | ..StartTurnRequest::default() |
| 3669 | }, |
| 3670 | ) |
| 3671 | .await?; |
| 3672 | assert!(matches!( |
| 3673 | harness.rx_op.recv().await, |
| 3674 | Some(Op::SendMessage { .. }) |
| 3675 | )); |
| 3676 | |
| 3677 | let started_at = Utc::now() - chrono::Duration::minutes(5); |
| 3678 | let dispatched_at = Utc::now() - chrono::Duration::seconds(2); |
| 3679 | let endpoint_fingerprint = "d".repeat(64); |
| 3680 | harness |
| 3681 | .tx_event |
| 3682 | .send(EngineEvent::TurnStarted { |
| 3683 | turn_id: "engine_route_receipt".to_string(), |
| 3684 | created_at: started_at, |
| 3685 | route: None, |
| 3686 | }) |
| 3687 | .await?; |
| 3688 | harness |
| 3689 | .tx_event |
| 3690 | .send(EngineEvent::RouteDispatched { |
| 3691 | turn_id: "engine_route_receipt".to_string(), |
| 3692 | route: crate::core::events::TurnRoute { |
| 3693 | provider: ApiProvider::Stepfun, |
| 3694 | provider_identity: "stepfun-payg-primary".to_string(), |
| 3695 | model: "step-3.7-flash".to_string(), |
| 3696 | auto_model: false, |
| 3697 | receipt: None, |
| 3698 | billing: Some(crate::core::events::RouteBillingEnvelope { |
| 3699 | billing_surface: Some(crate::pricing::STEPFUN_PAYG_BILLING_SURFACE.to_string()), |
| 3700 | endpoint_fingerprint: Some(endpoint_fingerprint.clone()), |
| 3701 | billing_mode: crate::cost_status::RouteBillingMode::Metered, |
| 3702 | dispatched_at, |
| 3703 | }), |
| 3704 | base_url: ApiProvider::Stepfun.default_base_url().to_string(), |
| 3705 | billing_product: crate::route_billing::RouteProduct::Unproven, |
| 3706 | }, |
| 3707 | }) |
| 3708 | .await?; |
| 3709 | harness |
| 3710 | .tx_event |
| 3711 | .send(EngineEvent::SubAgentMailbox { |
| 3712 | turn_id: "engine_route_receipt".to_string(), |
| 3713 | seq: 77, |
| 3714 | message: crate::tools::subagent::MailboxMessage::TokenUsage { |
| 3715 | agent_id: "agent_child".to_string(), |
| 3716 | source_id: "response-child".to_string(), |
| 3717 | route: crate::cost_status::EffectiveRouteEnvelope { |
| 3718 | provider: ApiProvider::OpenaiCodex, |
| 3719 | provider_identity: "codex-child".to_string(), |
| 3720 | model: "gpt-5.5".to_string(), |
| 3721 | billing_surface: Some( |
| 3722 | crate::pricing::OAUTH_SUBSCRIPTION_BILLING_SURFACE.to_string(), |
| 3723 | ), |
| 3724 | endpoint_fingerprint: None, |
| 3725 | billing_mode: crate::cost_status::RouteBillingMode::Subscription, |
| 3726 | dispatched_at, |
| 3727 | }, |
| 3728 | usage: Usage { |
| 3729 | input_tokens: 3, |
| 3730 | output_tokens: 2, |
| 3731 | reasoning_tokens: Some(2), |
| 3732 | ..Usage::default() |
| 3733 | }, |
| 3734 | }, |
| 3735 | }) |
| 3736 | .await?; |
| 3737 | harness |
| 3738 | .tx_event |
| 3739 | .send(EngineEvent::TurnComplete { |
| 3740 | usage: Usage { |
| 3741 | input_tokens: 10, |
| 3742 | output_tokens: 4, |
| 3743 | reasoning_replay_tokens: Some(6), |
| 3744 | server_tool_use: Some(crate::models::ServerToolUsage { |
| 3745 | code_execution_requests: Some(2), |
| 3746 | tool_search_requests: Some(3), |
| 3747 | }), |
| 3748 | ..Usage::default() |
| 3749 | }, |
| 3750 | status: TurnOutcomeStatus::Completed, |
| 3751 | error: None, |
| 3752 | tool_catalog: None, |
| 3753 | // Completion-time endpoint evidence must be ignored. |
| 3754 | base_url: Some("https://api.stepfun.com/v1/coding".to_string()), |
| 3755 | }) |
| 3756 | .await?; |
| 3757 | |
| 3758 | let completed = wait_for_terminal_turn(&manager, &turn.id, Duration::from_secs(2)).await?; |
| 3759 | assert_eq!(completed.effective_provider.as_deref(), Some("stepfun")); |
| 3760 | assert_eq!( |
| 3761 | completed.effective_provider_id.as_deref(), |
| 3762 | Some("stepfun-payg-primary") |
| 3763 | ); |
| 3764 | assert_eq!( |
| 3765 | completed.effective_billing_surface.as_deref(), |
| 3766 | Some(crate::pricing::STEPFUN_PAYG_BILLING_SURFACE) |
| 3767 | ); |
| 3768 | assert_eq!( |
| 3769 | completed.effective_endpoint_fingerprint.as_deref(), |
| 3770 | Some(endpoint_fingerprint.as_str()) |
| 3771 | ); |
| 3772 | assert_eq!( |
| 3773 | completed.effective_billing_mode, |
| 3774 | Some(crate::cost_status::RouteBillingMode::Metered) |
| 3775 | ); |
| 3776 | assert_eq!(completed.effective_dispatched_at, Some(dispatched_at)); |
| 3777 | assert_eq!(completed.started_at, Some(started_at)); |
| 3778 | assert_eq!(completed.routed_usage.len(), 1); |
| 3779 | assert_eq!(completed.routed_usage[0].usage.reasoning_tokens, Some(2)); |
| 3780 | let persisted_usage = completed.usage.expect("parent usage"); |
| 3781 | assert_eq!(persisted_usage.reasoning_replay_tokens, Some(6)); |
| 3782 | assert_eq!( |
| 3783 | persisted_usage.server_tool_use, |
| 3784 | Some(crate::models::ServerToolUsage { |
| 3785 | code_execution_requests: Some(2), |
| 3786 | tool_search_requests: Some(3), |
| 3787 | }) |
| 3788 | ); |
| 3789 | |
| 3790 | // A new engine mailbox starts at the same sequence number. Runtime |
| 3791 | // dedupe is scoped by the engine turn identity, so the second receipt is |
| 3792 | // persisted exactly once instead of colliding with the prior turn. |
| 3793 | let second = manager |
| 3794 | .start_turn( |
| 3795 | &thread.id, |
| 3796 | StartTurnRequest { |
| 3797 | prompt: "capture second mailbox".to_string(), |
| 3798 | ..StartTurnRequest::default() |
| 3799 | }, |
| 3800 | ) |
| 3801 | .await?; |
| 3802 | assert!(matches!( |
| 3803 | harness.rx_op.recv().await, |
| 3804 | Some(Op::SendMessage { .. }) |
| 3805 | )); |
| 3806 | let second_engine_turn = "engine_route_receipt_second"; |
| 3807 | harness |
| 3808 | .tx_event |
| 3809 | .send(EngineEvent::TurnStarted { |
| 3810 | turn_id: second_engine_turn.to_string(), |
| 3811 | created_at: Utc::now(), |
| 3812 | route: None, |
| 3813 | }) |
| 3814 | .await?; |
| 3815 | harness |
| 3816 | .tx_event |
| 3817 | .send(EngineEvent::SubAgentMailbox { |
| 3818 | turn_id: second_engine_turn.to_string(), |
| 3819 | seq: 77, |
| 3820 | message: crate::tools::subagent::MailboxMessage::TokenUsage { |
| 3821 | agent_id: "agent-child-second".to_string(), |
| 3822 | source_id: "response-child-second".to_string(), |
| 3823 | route: crate::cost_status::EffectiveRouteEnvelope { |
| 3824 | provider: ApiProvider::OpenaiCodex, |
| 3825 | provider_identity: "codex-child".to_string(), |
| 3826 | model: "gpt-5.5".to_string(), |
| 3827 | billing_surface: Some( |
| 3828 | crate::pricing::OAUTH_SUBSCRIPTION_BILLING_SURFACE.to_string(), |
| 3829 | ), |
| 3830 | endpoint_fingerprint: None, |
| 3831 | billing_mode: crate::cost_status::RouteBillingMode::Subscription, |
| 3832 | dispatched_at: Utc::now(), |
| 3833 | }, |
| 3834 | usage: Usage { |
| 3835 | input_tokens: 5, |
| 3836 | output_tokens: 1, |
| 3837 | ..Usage::default() |
| 3838 | }, |
| 3839 | }, |
| 3840 | }) |
| 3841 | .await?; |
| 3842 | harness |
| 3843 | .tx_event |
| 3844 | .send(EngineEvent::TurnComplete { |
| 3845 | usage: Usage::default(), |
| 3846 | status: TurnOutcomeStatus::Completed, |
| 3847 | error: None, |
| 3848 | tool_catalog: None, |
| 3849 | base_url: None, |
| 3850 | }) |
| 3851 | .await?; |
| 3852 | let second = wait_for_terminal_turn(&manager, &second.id, Duration::from_secs(2)).await?; |
| 3853 | assert_eq!(second.routed_usage.len(), 1); |
| 3854 | assert_eq!(second.routed_usage[0].usage.input_tokens, 5); |
| 3855 | Ok(()) |
| 3856 | } |
| 3857 | |
| 3858 | #[tokio::test] |
| 3859 | async fn completed_turn_without_engine_output_fails() -> Result<()> { |
| 3860 | let manager = test_manager(test_runtime_dir())?; |
| 3861 | let thread = manager |
| 3862 | .create_thread(CreateThreadRequest { |
| 3863 | model: None, |
| 3864 | workspace: None, |
| 3865 | mode: None, |
| 3866 | allow_shell: None, |
| 3867 | trust_mode: None, |
| 3868 | auto_approve: None, |
| 3869 | archived: false, |
| 3870 | system_prompt: None, |
| 3871 | task_id: None, |
| 3872 | ..Default::default() |
| 3873 | }) |
| 3874 | .await?; |
| 3875 | |
| 3876 | let harness = install_mock_engine(&manager, &thread.id).await; |
| 3877 | let mut rx_op = harness.rx_op; |
| 3878 | let tx_event = harness.tx_event; |
| 3879 | tokio::spawn(async move { |
| 3880 | if matches!(rx_op.recv().await, Some(Op::SendMessage { .. })) { |
| 3881 | let _ = tx_event |
| 3882 | .send(EngineEvent::TurnStarted { |
| 3883 | turn_id: "engine_empty_turn".to_string(), |
| 3884 | created_at: chrono::Utc::now(), |
| 3885 | route: None, |
| 3886 | }) |
| 3887 | .await; |
| 3888 | let _ = tx_event |
| 3889 | .send(EngineEvent::TurnComplete { |
| 3890 | usage: Usage { |
| 3891 | input_tokens: 10, |
| 3892 | output_tokens: 0, |
| 3893 | ..Usage::default() |
| 3894 | }, |
| 3895 | status: TurnOutcomeStatus::Completed, |
| 3896 | error: None, |
| 3897 | tool_catalog: None, |
| 3898 | base_url: None, |
| 3899 | }) |
| 3900 | .await; |
| 3901 | } |
| 3902 | }); |
| 3903 | |
| 3904 | let turn = manager |
| 3905 | .start_turn( |
| 3906 | &thread.id, |
| 3907 | StartTurnRequest { |
| 3908 | prompt: "empty turn".to_string(), |
| 3909 | input_summary: None, |
| 3910 | model: None, |
| 3911 | mode: None, |
| 3912 | allow_shell: None, |
| 3913 | trust_mode: None, |
| 3914 | auto_approve: None, |
| 3915 | ..Default::default() |
| 3916 | }, |
| 3917 | ) |
| 3918 | .await?; |
| 3919 | |
| 3920 | let failed = wait_for_terminal_turn(&manager, &turn.id, Duration::from_secs(2)).await?; |
| 3921 | assert_eq!(failed.status, RuntimeTurnStatus::Failed); |
| 3922 | assert_eq!(failed.error.as_deref(), Some(EMPTY_TURN_REASON)); |
| 3923 | |
| 3924 | let events = manager.events_since(&thread.id, None)?; |
| 3925 | assert!(events.iter().any(|ev| { |
| 3926 | ev.event == "item.failed" |
| 3927 | && ev |
| 3928 | .payload |
| 3929 | .get("item") |
| 3930 | .and_then(|item| item.get("kind")) |
| 3931 | .and_then(Value::as_str) |
| 3932 | == Some("error") |
| 3933 | })); |
| 3934 | assert!(events.iter().any(|ev| { |
| 3935 | ev.event == "turn.completed" |
| 3936 | && ev |
| 3937 | .payload |
| 3938 | .get("turn") |
| 3939 | .and_then(|turn| turn.get("status")) |
| 3940 | .and_then(Value::as_str) |
| 3941 | == Some("failed") |
| 3942 | })); |
| 3943 | Ok(()) |
| 3944 | } |
| 3945 | |
| 3946 | #[tokio::test] |
| 3947 | async fn preturn_control_status_does_not_make_empty_turn_succeed() -> Result<()> { |
| 3948 | let manager = test_manager(test_runtime_dir())?; |
| 3949 | let thread = manager |
| 3950 | .create_thread(CreateThreadRequest::default()) |
| 3951 | .await?; |
| 3952 | let harness = install_mock_engine(&manager, &thread.id).await; |
| 3953 | let mut rx_op = harness.rx_op; |
| 3954 | let tx_event = harness.tx_event; |
| 3955 | tokio::spawn(async move { |
| 3956 | if matches!(rx_op.recv().await, Some(Op::SendMessage { .. })) { |
| 3957 | let _ = tx_event |
| 3958 | .send(EngineEvent::AgentComplete { |
| 3959 | id: "stale_agent".to_string(), |
| 3960 | result: "stale completion".to_string(), |
| 3961 | }) |
| 3962 | .await; |
| 3963 | let _ = tx_event |
| 3964 | .send(EngineEvent::status("Compaction settings updated")) |
| 3965 | .await; |
| 3966 | let _ = tx_event |
| 3967 | .send(EngineEvent::TurnStarted { |
| 3968 | turn_id: "engine_empty_after_control_status".to_string(), |
| 3969 | created_at: chrono::Utc::now(), |
| 3970 | route: None, |
| 3971 | }) |
| 3972 | .await; |
| 3973 | let _ = tx_event |
| 3974 | .send(EngineEvent::TurnComplete { |
| 3975 | usage: Usage::default(), |
| 3976 | status: TurnOutcomeStatus::Completed, |
| 3977 | error: None, |
| 3978 | tool_catalog: None, |
| 3979 | base_url: None, |
| 3980 | }) |
| 3981 | .await; |
| 3982 | } |
| 3983 | }); |
| 3984 | |
| 3985 | let turn = manager |
| 3986 | .start_turn( |
| 3987 | &thread.id, |
| 3988 | StartTurnRequest { |
| 3989 | prompt: "empty after setup".to_string(), |
| 3990 | ..Default::default() |
| 3991 | }, |
| 3992 | ) |
| 3993 | .await?; |
| 3994 | let terminal = wait_for_terminal_turn(&manager, &turn.id, Duration::from_secs(2)).await?; |
| 3995 | assert_eq!(terminal.status, RuntimeTurnStatus::Failed); |
| 3996 | assert_eq!(terminal.error.as_deref(), Some(EMPTY_TURN_REASON)); |
| 3997 | assert!( |
| 3998 | manager |
| 3999 | .store |
| 4000 | .list_items_for_turn(&turn.id)? |
| 4001 | .iter() |
| 4002 | .all(|item| { |
| 4003 | item.summary != "Compaction settings updated" |
| 4004 | && !item.summary.contains("stale_agent") |
| 4005 | }) |
| 4006 | ); |
| 4007 | Ok(()) |
| 4008 | } |
| 4009 | |
| 4010 | #[tokio::test] |
| 4011 | async fn engine_error_remains_failed_after_nominal_turn_complete() -> Result<()> { |
| 4012 | let manager = test_manager(test_runtime_dir())?; |
| 4013 | let thread = manager |
| 4014 | .create_thread(CreateThreadRequest::default()) |
| 4015 | .await?; |
| 4016 | let harness = install_mock_engine(&manager, &thread.id).await; |
| 4017 | let mut rx_op = harness.rx_op; |
| 4018 | let tx_event = harness.tx_event; |
| 4019 | tokio::spawn(async move { |
| 4020 | if matches!(rx_op.recv().await, Some(Op::SendMessage { .. })) { |
| 4021 | let _ = tx_event |
| 4022 | .send(EngineEvent::TurnStarted { |
| 4023 | turn_id: "engine_error_then_complete".to_string(), |
| 4024 | created_at: chrono::Utc::now(), |
| 4025 | route: None, |
| 4026 | }) |
| 4027 | .await; |
| 4028 | let _ = tx_event |
| 4029 | .send(EngineEvent::error( |
| 4030 | crate::error_taxonomy::ErrorEnvelope::fatal("provider exploded"), |
| 4031 | )) |
| 4032 | .await; |
| 4033 | let _ = tx_event |
| 4034 | .send(EngineEvent::TurnComplete { |
| 4035 | usage: Usage::default(), |
| 4036 | status: TurnOutcomeStatus::Completed, |
| 4037 | error: None, |
| 4038 | tool_catalog: None, |
| 4039 | base_url: None, |
| 4040 | }) |
| 4041 | .await; |
| 4042 | } |
| 4043 | }); |
| 4044 | |
| 4045 | let turn = manager |
| 4046 | .start_turn( |
| 4047 | &thread.id, |
| 4048 | StartTurnRequest { |
| 4049 | prompt: "surface the failure".to_string(), |
| 4050 | ..Default::default() |
| 4051 | }, |
| 4052 | ) |
| 4053 | .await?; |
| 4054 | let terminal = wait_for_terminal_turn(&manager, &turn.id, Duration::from_secs(2)).await?; |
| 4055 | assert_eq!(terminal.status, RuntimeTurnStatus::Failed); |
| 4056 | assert_eq!(terminal.error.as_deref(), Some("provider exploded")); |
| 4057 | Ok(()) |
| 4058 | } |
| 4059 | |
| 4060 | #[tokio::test] |
| 4061 | async fn create_thread_defaults_auto_approve_to_false() -> Result<()> { |
| 4062 | let manager = test_manager(test_runtime_dir())?; |
| 4063 | let thread = manager |
| 4064 | .create_thread(CreateThreadRequest { |
| 4065 | model: None, |
| 4066 | workspace: None, |
| 4067 | mode: None, |
| 4068 | allow_shell: None, |
| 4069 | trust_mode: None, |
| 4070 | auto_approve: None, |
| 4071 | archived: false, |
| 4072 | system_prompt: None, |
| 4073 | task_id: None, |
| 4074 | ..Default::default() |
| 4075 | }) |
| 4076 | .await?; |
| 4077 | |
| 4078 | assert!(!thread.auto_approve); |
| 4079 | Ok(()) |
| 4080 | } |
| 4081 | |
| 4082 | #[tokio::test] |
| 4083 | async fn update_thread_workspace_persists_event_and_evicts_idle_engine() -> Result<()> { |
| 4084 | let manager = test_manager(test_runtime_dir())?; |
| 4085 | let old_workspace = std::env::temp_dir().join("codewhale-runtime-old-workspace"); |
| 4086 | let new_workspace = std::env::temp_dir().join("codewhale-runtime-new-workspace"); |
| 4087 | let thread = manager |
| 4088 | .create_thread(CreateThreadRequest { |
| 4089 | model: None, |
| 4090 | workspace: Some(old_workspace.clone()), |
| 4091 | mode: None, |
| 4092 | allow_shell: None, |
| 4093 | trust_mode: None, |
| 4094 | auto_approve: None, |
| 4095 | archived: false, |
| 4096 | system_prompt: None, |
| 4097 | task_id: None, |
| 4098 | ..Default::default() |
| 4099 | }) |
| 4100 | .await?; |
| 4101 | |
| 4102 | let harness = install_mock_engine(&manager, &thread.id).await; |
| 4103 | let mut rx_op = harness.rx_op; |
| 4104 | |
| 4105 | let updated = manager |
| 4106 | .update_thread( |
| 4107 | &thread.id, |
| 4108 | UpdateThreadRequest { |
| 4109 | workspace: Some(new_workspace.clone()), |
| 4110 | ..UpdateThreadRequest::default() |
| 4111 | }, |
| 4112 | ) |
| 4113 | .await?; |
| 4114 | |
| 4115 | assert_eq!(updated.workspace, new_workspace); |
| 4116 | assert_eq!( |
| 4117 | manager.store.load_thread(&thread.id)?.workspace, |
| 4118 | new_workspace |
| 4119 | ); |
| 4120 | { |
| 4121 | let active = manager.active.lock().await; |
| 4122 | assert!( |
| 4123 | !active.engines.contains_key(&thread.id), |
| 4124 | "workspace changes must evict the stale cached engine" |
| 4125 | ); |
| 4126 | assert!(!active.lru.iter().any(|id| id == &thread.id)); |
| 4127 | } |
| 4128 | |
| 4129 | match tokio::time::timeout(Duration::from_secs(1), rx_op.recv()).await { |
| 4130 | Ok(Some(Op::Shutdown)) => {} |
| 4131 | other => panic!("expected cached engine shutdown, got {other:?}"), |
| 4132 | } |
| 4133 | |
| 4134 | let events = manager.events_since(&thread.id, None)?; |
| 4135 | let event = events |
| 4136 | .iter() |
| 4137 | .rev() |
| 4138 | .find(|event| event.event == "thread.updated") |
| 4139 | .expect("thread.updated event"); |
| 4140 | let workspace_value = serde_json::to_value(&updated.workspace)?; |
| 4141 | assert_eq!( |
| 4142 | event |
| 4143 | .payload |
| 4144 | .get("changes") |
| 4145 | .and_then(|changes| changes.get("workspace")), |
| 4146 | Some(&workspace_value) |
| 4147 | ); |
| 4148 | Ok(()) |
| 4149 | } |
| 4150 | |
| 4151 | #[tokio::test] |
| 4152 | async fn update_thread_workspace_rejects_empty_path() -> Result<()> { |
| 4153 | let manager = test_manager(test_runtime_dir())?; |
| 4154 | let thread = manager |
| 4155 | .create_thread(CreateThreadRequest { |
| 4156 | model: None, |
| 4157 | workspace: None, |
| 4158 | mode: None, |
| 4159 | allow_shell: None, |
| 4160 | trust_mode: None, |
| 4161 | auto_approve: None, |
| 4162 | archived: false, |
| 4163 | system_prompt: None, |
| 4164 | task_id: None, |
| 4165 | ..Default::default() |
| 4166 | }) |
| 4167 | .await?; |
| 4168 | |
| 4169 | let err = manager |
| 4170 | .update_thread( |
| 4171 | &thread.id, |
| 4172 | UpdateThreadRequest { |
| 4173 | workspace: Some(PathBuf::new()), |
| 4174 | ..UpdateThreadRequest::default() |
| 4175 | }, |
| 4176 | ) |
| 4177 | .await |
| 4178 | .expect_err("empty workspace must be rejected"); |
| 4179 | assert!(format!("{err:#}").contains("workspace must not be empty")); |
| 4180 | Ok(()) |
| 4181 | } |
| 4182 | |
| 4183 | #[tokio::test] |
| 4184 | async fn update_thread_workspace_rejects_active_turn() -> Result<()> { |
| 4185 | let manager = test_manager(test_runtime_dir())?; |
| 4186 | let old_workspace = std::env::temp_dir().join("codewhale-runtime-active-old"); |
| 4187 | let new_workspace = std::env::temp_dir().join("codewhale-runtime-active-new"); |
| 4188 | let thread = manager |
| 4189 | .create_thread(CreateThreadRequest { |
| 4190 | model: None, |
| 4191 | workspace: Some(old_workspace.clone()), |
| 4192 | mode: None, |
| 4193 | allow_shell: None, |
| 4194 | trust_mode: None, |
| 4195 | auto_approve: None, |
| 4196 | archived: false, |
| 4197 | system_prompt: None, |
| 4198 | task_id: None, |
| 4199 | ..Default::default() |
| 4200 | }) |
| 4201 | .await?; |
| 4202 | |
| 4203 | let harness = install_mock_engine(&manager, &thread.id).await; |
| 4204 | let mut rx_op = harness.rx_op; |
| 4205 | { |
| 4206 | let mut active = manager.active.lock().await; |
| 4207 | let state = active.engines.get_mut(&thread.id).expect("mock engine"); |
| 4208 | state.active_turn = Some(ActiveTurnState { |
| 4209 | turn_id: "turn_live".to_string(), |
| 4210 | interrupt_requested: false, |
| 4211 | }); |
| 4212 | } |
| 4213 | |
| 4214 | let err = manager |
| 4215 | .update_thread( |
| 4216 | &thread.id, |
| 4217 | UpdateThreadRequest { |
| 4218 | workspace: Some(new_workspace), |
| 4219 | ..UpdateThreadRequest::default() |
| 4220 | }, |
| 4221 | ) |
| 4222 | .await |
| 4223 | .expect_err("workspace update during active turn must fail"); |
| 4224 | |
| 4225 | assert!(format!("{err:#}").contains("active turn")); |
| 4226 | assert_eq!( |
| 4227 | manager.store.load_thread(&thread.id)?.workspace, |
| 4228 | old_workspace |
| 4229 | ); |
| 4230 | { |
| 4231 | let active = manager.active.lock().await; |
| 4232 | assert!( |
| 4233 | active.engines.contains_key(&thread.id), |
| 4234 | "active engine should stay cached after rejected update" |
| 4235 | ); |
| 4236 | } |
| 4237 | assert!( |
| 4238 | tokio::time::timeout(Duration::from_millis(100), rx_op.recv()) |
| 4239 | .await |
| 4240 | .is_err(), |
| 4241 | "rejected workspace update must not shut down the active engine" |
| 4242 | ); |
| 4243 | Ok(()) |
| 4244 | } |
| 4245 | |
| 4246 | #[tokio::test] |
| 4247 | async fn start_turn_passes_effective_auto_approve_to_engine() -> Result<()> { |
| 4248 | let manager = test_manager(test_runtime_dir())?; |
| 4249 | let thread = manager |
| 4250 | .create_thread(CreateThreadRequest { |
| 4251 | model: None, |
| 4252 | workspace: None, |
| 4253 | mode: None, |
| 4254 | allow_shell: None, |
| 4255 | trust_mode: None, |
| 4256 | auto_approve: Some(false), |
| 4257 | archived: false, |
| 4258 | system_prompt: None, |
| 4259 | task_id: None, |
| 4260 | ..Default::default() |
| 4261 | }) |
| 4262 | .await?; |
| 4263 | |
| 4264 | let harness = install_mock_engine(&manager, &thread.id).await; |
| 4265 | let mut rx_op = harness.rx_op; |
| 4266 | |
| 4267 | let turn = manager |
| 4268 | .start_turn( |
| 4269 | &thread.id, |
| 4270 | StartTurnRequest { |
| 4271 | prompt: "override approval".to_string(), |
| 4272 | input_summary: None, |
| 4273 | model: None, |
| 4274 | mode: None, |
| 4275 | allow_shell: None, |
| 4276 | trust_mode: None, |
| 4277 | auto_approve: Some(true), |
| 4278 | ..Default::default() |
| 4279 | }, |
| 4280 | ) |
| 4281 | .await?; |
| 4282 | assert_eq!(turn.permission_posture.as_deref(), Some("full_access")); |
| 4283 | |
| 4284 | match rx_op.recv().await { |
| 4285 | Some(Op::SendMessage { |
| 4286 | auto_approve, |
| 4287 | approval_mode, |
| 4288 | .. |
| 4289 | }) => { |
| 4290 | assert!(auto_approve); |
| 4291 | assert_eq!(approval_mode, crate::tui::approval::ApprovalMode::Bypass); |
| 4292 | } |
| 4293 | other => panic!("expected SendMessage op, got {other:?}"), |
| 4294 | } |
| 4295 | |
| 4296 | Ok(()) |
| 4297 | } |
| 4298 | |
| 4299 | #[tokio::test] |
| 4300 | async fn start_turn_can_override_thread_auto_approve_to_false() -> Result<()> { |
| 4301 | let manager = test_manager(test_runtime_dir())?; |
| 4302 | let thread = manager |
| 4303 | .create_thread(CreateThreadRequest { |
| 4304 | model: None, |
| 4305 | workspace: None, |
| 4306 | mode: None, |
| 4307 | allow_shell: None, |
| 4308 | trust_mode: None, |
| 4309 | auto_approve: Some(true), |
| 4310 | archived: false, |
| 4311 | system_prompt: None, |
| 4312 | task_id: None, |
| 4313 | ..Default::default() |
| 4314 | }) |
| 4315 | .await?; |
| 4316 | |
| 4317 | let harness = install_mock_engine(&manager, &thread.id).await; |
| 4318 | let mut rx_op = harness.rx_op; |
| 4319 | |
| 4320 | let turn = manager |
| 4321 | .start_turn( |
| 4322 | &thread.id, |
| 4323 | StartTurnRequest { |
| 4324 | prompt: "disable approval".to_string(), |
| 4325 | input_summary: None, |
| 4326 | model: None, |
| 4327 | mode: None, |
| 4328 | allow_shell: None, |
| 4329 | trust_mode: None, |
| 4330 | auto_approve: Some(false), |
| 4331 | ..Default::default() |
| 4332 | }, |
| 4333 | ) |
| 4334 | .await?; |
| 4335 | assert_eq!(turn.permission_posture.as_deref(), Some("ask")); |
| 4336 | |
| 4337 | match rx_op.recv().await { |
| 4338 | Some(Op::SendMessage { |
| 4339 | auto_approve, |
| 4340 | approval_mode, |
| 4341 | .. |
| 4342 | }) => { |
| 4343 | assert!(!auto_approve); |
| 4344 | assert_eq!(approval_mode, crate::tui::approval::ApprovalMode::Suggest); |
| 4345 | } |
| 4346 | other => panic!("expected SendMessage op, got {other:?}"), |
| 4347 | } |
| 4348 | |
| 4349 | Ok(()) |
| 4350 | } |
| 4351 | |
| 4352 | #[tokio::test] |
| 4353 | async fn start_turn_enforces_and_records_auto_review_without_legacy_bypass() -> Result<()> { |
| 4354 | let manager = test_manager(test_runtime_dir())?; |
| 4355 | let thread = manager |
| 4356 | .create_thread(CreateThreadRequest { |
| 4357 | permission_posture: Some("ask".to_string()), |
| 4358 | ..Default::default() |
| 4359 | }) |
| 4360 | .await?; |
| 4361 | let harness = install_mock_engine(&manager, &thread.id).await; |
| 4362 | let mut rx_op = harness.rx_op; |
| 4363 | |
| 4364 | let turn = manager |
| 4365 | .start_turn( |
| 4366 | &thread.id, |
| 4367 | StartTurnRequest { |
| 4368 | prompt: "review autonomously".to_string(), |
| 4369 | permission_posture: Some("auto-review".to_string()), |
| 4370 | ..Default::default() |
| 4371 | }, |
| 4372 | ) |
| 4373 | .await?; |
| 4374 | |
| 4375 | assert_eq!(turn.permission_posture.as_deref(), Some("auto_review")); |
| 4376 | assert_eq!( |
| 4377 | manager |
| 4378 | .store |
| 4379 | .load_turn(&turn.id)? |
| 4380 | .permission_posture |
| 4381 | .as_deref(), |
| 4382 | Some("auto_review") |
| 4383 | ); |
| 4384 | match rx_op.recv().await { |
| 4385 | Some(Op::SendMessage { |
| 4386 | auto_approve, |
| 4387 | approval_mode, |
| 4388 | .. |
| 4389 | }) => { |
| 4390 | assert!(!auto_approve); |
| 4391 | assert_eq!(approval_mode, crate::tui::approval::ApprovalMode::Auto); |
| 4392 | } |
| 4393 | other => panic!("expected SendMessage op, got {other:?}"), |
| 4394 | } |
| 4395 | Ok(()) |
| 4396 | } |
| 4397 | |
| 4398 | #[tokio::test] |
| 4399 | async fn active_turn_permission_posture_switches_use_the_engine_live_authority() -> Result<()> { |
| 4400 | let manager = test_manager(test_runtime_dir())?; |
| 4401 | let thread = manager |
| 4402 | .create_thread(CreateThreadRequest { |
| 4403 | permission_posture: Some("ask".to_string()), |
| 4404 | ..Default::default() |
| 4405 | }) |
| 4406 | .await?; |
| 4407 | let mut harness = install_mock_engine(&manager, &thread.id).await; |
| 4408 | let turn = manager |
| 4409 | .start_turn( |
| 4410 | &thread.id, |
| 4411 | StartTurnRequest { |
| 4412 | prompt: "keep working while permissions change".to_string(), |
| 4413 | ..Default::default() |
| 4414 | }, |
| 4415 | ) |
| 4416 | .await?; |
| 4417 | assert!(matches!( |
| 4418 | harness.rx_op.recv().await, |
| 4419 | Some(Op::SendMessage { |
| 4420 | approval_mode: crate::tui::approval::ApprovalMode::Suggest, |
| 4421 | .. |
| 4422 | }) |
| 4423 | )); |
| 4424 | |
| 4425 | for (requested, canonical, expected_auto, expected_approval) in [ |
| 4426 | ( |
| 4427 | "auto-review", |
| 4428 | "auto_review", |
| 4429 | false, |
| 4430 | crate::tui::approval::ApprovalMode::Auto, |
| 4431 | ), |
| 4432 | ( |
| 4433 | "full-access", |
| 4434 | "full_access", |
| 4435 | true, |
| 4436 | crate::tui::approval::ApprovalMode::Bypass, |
| 4437 | ), |
| 4438 | ( |
| 4439 | "ask", |
| 4440 | "ask", |
| 4441 | false, |
| 4442 | crate::tui::approval::ApprovalMode::Suggest, |
| 4443 | ), |
| 4444 | ] { |
| 4445 | let updated = manager |
| 4446 | .update_thread( |
| 4447 | &thread.id, |
| 4448 | UpdateThreadRequest { |
| 4449 | permission_posture: Some(requested.to_string()), |
| 4450 | ..Default::default() |
| 4451 | }, |
| 4452 | ) |
| 4453 | .await?; |
| 4454 | assert_eq!(updated.permission_posture.as_deref(), Some(canonical)); |
| 4455 | assert_eq!(updated.auto_approve, expected_auto); |
| 4456 | |
| 4457 | match harness.rx_op.recv().await { |
| 4458 | Some(Op::ChangeMode { |
| 4459 | auto_approve, |
| 4460 | approval_mode, |
| 4461 | .. |
| 4462 | }) => { |
| 4463 | assert_eq!(auto_approve, expected_auto, "{requested}"); |
| 4464 | assert_eq!(approval_mode, expected_approval, "{requested}"); |
| 4465 | } |
| 4466 | other => panic!("expected ChangeMode for {requested}, got {other:?}"), |
| 4467 | } |
| 4468 | |
| 4469 | let authority = { |
| 4470 | let active = manager.active.lock().await; |
| 4471 | active |
| 4472 | .engines |
| 4473 | .get(&thread.id) |
| 4474 | .expect("active engine") |
| 4475 | .engine |
| 4476 | .runtime_permission_authority() |
| 4477 | }; |
| 4478 | assert_eq!(authority.auto_approve, expected_auto, "{requested}"); |
| 4479 | assert_eq!(authority.approval_mode, expected_approval, "{requested}"); |
| 4480 | assert_eq!( |
| 4481 | manager.active_turn_flags(&thread.id, &turn.id).await, |
| 4482 | Some((expected_auto, false)), |
| 4483 | "{requested} must replace the running turn's authority immediately" |
| 4484 | ); |
| 4485 | } |
| 4486 | |
| 4487 | Ok(()) |
| 4488 | } |
| 4489 | |
| 4490 | #[tokio::test] |
| 4491 | async fn compact_thread_preserves_thread_auto_approve_policy() -> Result<()> { |
| 4492 | let manager = test_manager(test_runtime_dir())?; |
| 4493 | let thread = manager |
| 4494 | .create_thread(CreateThreadRequest { |
| 4495 | model: None, |
| 4496 | workspace: None, |
| 4497 | mode: None, |
| 4498 | allow_shell: None, |
| 4499 | trust_mode: None, |
| 4500 | auto_approve: Some(false), |
| 4501 | archived: false, |
| 4502 | system_prompt: None, |
| 4503 | task_id: None, |
| 4504 | ..Default::default() |
| 4505 | }) |
| 4506 | .await?; |
| 4507 | |
| 4508 | let harness = install_mock_engine(&manager, &thread.id).await; |
| 4509 | let mut rx_op = harness.rx_op; |
| 4510 | |
| 4511 | let turn = manager |
| 4512 | .compact_thread(&thread.id, CompactThreadRequest::default()) |
| 4513 | .await?; |
| 4514 | |
| 4515 | match rx_op.recv().await { |
| 4516 | Some(Op::CompactContext { compaction, .. }) => { |
| 4517 | assert_eq!( |
| 4518 | compaction.runtime_cost_owner.as_deref(), |
| 4519 | Some(turn.id.as_str()) |
| 4520 | ); |
| 4521 | } |
| 4522 | other => panic!("expected CompactContext op, got {other:?}"), |
| 4523 | } |
| 4524 | assert_eq!( |
| 4525 | manager.active_turn_flags(&thread.id, &turn.id).await, |
| 4526 | Some((false, false)) |
| 4527 | ); |
| 4528 | |
| 4529 | Ok(()) |
| 4530 | } |
| 4531 | |
| 4532 | #[tokio::test] |
| 4533 | async fn closed_compaction_mailbox_rolls_back_durable_records_and_active_claim() -> Result<()> { |
| 4534 | let manager = test_manager(test_runtime_dir())?; |
| 4535 | let thread = manager |
| 4536 | .create_thread(CreateThreadRequest::default()) |
| 4537 | .await?; |
| 4538 | let harness = install_mock_engine(&manager, &thread.id).await; |
| 4539 | let before_active = { |
| 4540 | let active = manager.active.lock().await; |
| 4541 | let state = active.engines.get(&thread.id).expect("installed engine"); |
| 4542 | ( |
| 4543 | state.active_turn.as_ref().map(|turn| turn.turn_id.clone()), |
| 4544 | state.route_identity.clone(), |
| 4545 | state.route_model.clone(), |
| 4546 | active.lru.clone(), |
| 4547 | ) |
| 4548 | }; |
| 4549 | let before_thread = serde_json::to_value(manager.get_thread(&thread.id).await?)?; |
| 4550 | let before_events = serde_json::to_value(manager.events_since(&thread.id, None)?)?; |
| 4551 | drop(harness.rx_op); |
| 4552 | |
| 4553 | let error = manager |
| 4554 | .compact_thread(&thread.id, CompactThreadRequest::default()) |
| 4555 | .await |
| 4556 | .expect_err("closed mailbox must reject compaction") |
| 4557 | .to_string(); |
| 4558 | assert!(error.contains("Failed to trigger compaction"), "{error}"); |
| 4559 | |
| 4560 | assert!(manager.store.list_turns_for_thread(&thread.id)?.is_empty()); |
| 4561 | assert_eq!( |
| 4562 | serde_json::to_value(manager.get_thread(&thread.id).await?)?, |
| 4563 | before_thread |
| 4564 | ); |
| 4565 | assert_eq!( |
| 4566 | serde_json::to_value(manager.events_since(&thread.id, None)?)?, |
| 4567 | before_events |
| 4568 | ); |
| 4569 | let after_active = { |
| 4570 | let active = manager.active.lock().await; |
| 4571 | let state = active.engines.get(&thread.id).expect("installed engine"); |
| 4572 | ( |
| 4573 | state.active_turn.as_ref().map(|turn| turn.turn_id.clone()), |
| 4574 | state.route_identity.clone(), |
| 4575 | state.route_model.clone(), |
| 4576 | active.lru.clone(), |
| 4577 | ) |
| 4578 | }; |
| 4579 | assert_eq!(after_active, before_active); |
| 4580 | Ok(()) |
| 4581 | } |
| 4582 | |
| 4583 | #[tokio::test] |
| 4584 | async fn compact_thread_receipt_keeps_exact_named_custom_identity() -> Result<()> { |
| 4585 | let mut custom = std::collections::HashMap::new(); |
| 4586 | custom.insert( |
| 4587 | "lm-studio".to_string(), |
| 4588 | crate::config::ProviderConfig { |
| 4589 | kind: Some("openai-compatible".to_string()), |
| 4590 | base_url: Some("http://127.0.0.1:1234/v1".to_string()), |
| 4591 | model: Some("local-code-model".to_string()), |
| 4592 | ..Default::default() |
| 4593 | }, |
| 4594 | ); |
| 4595 | let manager = RuntimeThreadManager::open( |
| 4596 | Config { |
| 4597 | provider: Some("lm-studio".to_string()), |
| 4598 | providers: Some(crate::config::ProvidersConfig { |
| 4599 | custom, |
| 4600 | ..Default::default() |
| 4601 | }), |
| 4602 | ..Default::default() |
| 4603 | }, |
| 4604 | PathBuf::from("."), |
| 4605 | test_manager_config(test_runtime_dir()), |
| 4606 | )?; |
| 4607 | let thread = manager |
| 4608 | .create_thread(CreateThreadRequest::default()) |
| 4609 | .await?; |
| 4610 | let harness = install_mock_engine(&manager, &thread.id).await; |
| 4611 | let mut rx_op = harness.rx_op; |
| 4612 | |
| 4613 | let turn = manager |
| 4614 | .compact_thread(&thread.id, CompactThreadRequest::default()) |
| 4615 | .await?; |
| 4616 | |
| 4617 | assert!(matches!( |
| 4618 | rx_op.recv().await, |
| 4619 | Some(Op::CompactContext { .. }) |
| 4620 | )); |
| 4621 | assert_eq!(turn.effective_provider.as_deref(), Some("custom")); |
| 4622 | assert_eq!(turn.effective_provider_id.as_deref(), Some("lm-studio")); |
| 4623 | Ok(()) |
| 4624 | } |
| 4625 | |
| 4626 | #[tokio::test] |
| 4627 | async fn compact_thread_with_real_engine_reaches_terminal_status() -> Result<()> { |
| 4628 | let manager = RuntimeThreadManager::open( |
| 4629 | Config { |
| 4630 | // This test intentionally crosses the real-engine boundary. Give |
| 4631 | // client preflight a hermetic credential and closed-loopback URL; |
| 4632 | // the assertion permits the resulting terminal failure. |
| 4633 | api_key: Some("runtime-thread-test-key".to_string()), |
| 4634 | base_url: Some("http://127.0.0.1:1/v1".to_string()), |
| 4635 | ..Config::default() |
| 4636 | }, |
| 4637 | PathBuf::from("."), |
| 4638 | test_manager_config(test_runtime_dir()), |
| 4639 | )?; |
| 4640 | let thread = manager |
| 4641 | .create_thread(CreateThreadRequest { |
| 4642 | model: None, |
| 4643 | workspace: None, |
| 4644 | mode: None, |
| 4645 | allow_shell: None, |
| 4646 | trust_mode: None, |
| 4647 | auto_approve: None, |
| 4648 | archived: false, |
| 4649 | system_prompt: None, |
| 4650 | task_id: None, |
| 4651 | ..Default::default() |
| 4652 | }) |
| 4653 | .await?; |
| 4654 | |
| 4655 | let turn = manager |
| 4656 | .compact_thread(&thread.id, CompactThreadRequest::default()) |
| 4657 | .await?; |
| 4658 | let terminal = wait_for_terminal_turn(&manager, &turn.id, Duration::from_secs(2)).await?; |
| 4659 | |
| 4660 | assert!(matches!( |
| 4661 | terminal.status, |
| 4662 | RuntimeTurnStatus::Completed | RuntimeTurnStatus::Failed |
| 4663 | )); |
| 4664 | assert!( |
| 4665 | terminal.ended_at.is_some(), |
| 4666 | "manual compaction should reach a terminal turn state" |
| 4667 | ); |
| 4668 | assert_eq!(manager.active_turn_flags(&thread.id, &turn.id).await, None); |
| 4669 | |
| 4670 | let expected_status = match terminal.status { |
| 4671 | RuntimeTurnStatus::Completed => "completed", |
| 4672 | RuntimeTurnStatus::Failed => "failed", |
| 4673 | other => panic!("unexpected non-terminal compaction status: {other:?}"), |
| 4674 | }; |
| 4675 | let events = manager.events_since(&thread.id, None)?; |
| 4676 | assert!(events.iter().any(|ev| { |
| 4677 | ev.event == "turn.completed" |
| 4678 | && ev |
| 4679 | .payload |
| 4680 | .get("turn") |
| 4681 | .and_then(|turn| turn.get("status")) |
| 4682 | .and_then(Value::as_str) |
| 4683 | == Some(expected_status) |
| 4684 | })); |
| 4685 | Ok(()) |
| 4686 | } |
| 4687 | |
| 4688 | #[tokio::test] |
| 4689 | async fn multi_turn_continuity_same_thread() -> Result<()> { |
| 4690 | let manager = test_manager(test_runtime_dir())?; |
| 4691 | let thread = manager |
| 4692 | .create_thread(CreateThreadRequest { |
| 4693 | model: None, |
| 4694 | workspace: None, |
| 4695 | mode: None, |
| 4696 | allow_shell: None, |
| 4697 | trust_mode: None, |
| 4698 | auto_approve: None, |
| 4699 | archived: false, |
| 4700 | system_prompt: None, |
| 4701 | task_id: None, |
| 4702 | ..Default::default() |
| 4703 | }) |
| 4704 | .await?; |
| 4705 | |
| 4706 | let harness = install_mock_engine(&manager, &thread.id).await; |
| 4707 | let mut rx_op = harness.rx_op; |
| 4708 | let tx_event = harness.tx_event; |
| 4709 | tokio::spawn(async move { |
| 4710 | let mut turn_index = 0u8; |
| 4711 | while let Some(op) = rx_op.recv().await { |
| 4712 | if !matches!(op, Op::SendMessage { .. }) { |
| 4713 | continue; |
| 4714 | } |
| 4715 | turn_index = turn_index.saturating_add(1); |
| 4716 | let _ = tx_event |
| 4717 | .send(EngineEvent::TurnStarted { |
| 4718 | turn_id: format!("engine_turn_{turn_index}"), |
| 4719 | created_at: chrono::Utc::now(), |
| 4720 | route: None, |
| 4721 | }) |
| 4722 | .await; |
| 4723 | let _ = tx_event |
| 4724 | .send(EngineEvent::MessageStarted { index: 0 }) |
| 4725 | .await; |
| 4726 | let _ = tx_event |
| 4727 | .send(EngineEvent::MessageDelta { |
| 4728 | index: 0, |
| 4729 | content: format!("reply {turn_index}"), |
| 4730 | }) |
| 4731 | .await; |
| 4732 | let _ = tx_event |
| 4733 | .send(EngineEvent::MessageComplete { index: 0 }) |
| 4734 | .await; |
| 4735 | let _ = tx_event |
| 4736 | .send(EngineEvent::TurnComplete { |
| 4737 | usage: Usage { |
| 4738 | input_tokens: 5, |
| 4739 | output_tokens: 5, |
| 4740 | ..Usage::default() |
| 4741 | }, |
| 4742 | status: TurnOutcomeStatus::Completed, |
| 4743 | error: None, |
| 4744 | tool_catalog: None, |
| 4745 | base_url: None, |
| 4746 | }) |
| 4747 | .await; |
| 4748 | if turn_index >= 2 { |
| 4749 | break; |
| 4750 | } |
| 4751 | } |
| 4752 | }); |
| 4753 | |
| 4754 | let turn_1 = manager |
| 4755 | .start_turn( |
| 4756 | &thread.id, |
| 4757 | StartTurnRequest { |
| 4758 | prompt: "first".to_string(), |
| 4759 | input_summary: None, |
| 4760 | model: None, |
| 4761 | mode: None, |
| 4762 | allow_shell: None, |
| 4763 | trust_mode: None, |
| 4764 | auto_approve: None, |
| 4765 | ..Default::default() |
| 4766 | }, |
| 4767 | ) |
| 4768 | .await?; |
| 4769 | let turn_1 = wait_for_terminal_turn(&manager, &turn_1.id, Duration::from_secs(2)).await?; |
| 4770 | assert_eq!(turn_1.status, RuntimeTurnStatus::Completed); |
| 4771 | |
| 4772 | let turn_2 = manager |
| 4773 | .start_turn( |
| 4774 | &thread.id, |
| 4775 | StartTurnRequest { |
| 4776 | prompt: "second".to_string(), |
| 4777 | input_summary: None, |
| 4778 | model: None, |
| 4779 | mode: None, |
| 4780 | allow_shell: None, |
| 4781 | trust_mode: None, |
| 4782 | auto_approve: None, |
| 4783 | ..Default::default() |
| 4784 | }, |
| 4785 | ) |
| 4786 | .await?; |
| 4787 | let turn_2 = wait_for_terminal_turn(&manager, &turn_2.id, Duration::from_secs(2)).await?; |
| 4788 | assert_eq!(turn_2.status, RuntimeTurnStatus::Completed); |
| 4789 | |
| 4790 | let detail = manager.get_thread_detail(&thread.id).await?; |
| 4791 | assert_eq!( |
| 4792 | detail.thread.latest_turn_id.as_deref(), |
| 4793 | Some(turn_2.id.as_str()) |
| 4794 | ); |
| 4795 | assert_eq!(detail.turns.len(), 2); |
| 4796 | assert!(detail.items.iter().any(|item| { |
| 4797 | item.kind == TurnItemKind::UserMessage && item.detail.as_deref() == Some("first") |
| 4798 | })); |
| 4799 | assert!(detail.items.iter().any(|item| { |
| 4800 | item.kind == TurnItemKind::UserMessage && item.detail.as_deref() == Some("second") |
| 4801 | })); |
| 4802 | |
| 4803 | let events = manager.events_since(&thread.id, None)?; |
| 4804 | let started = events |
| 4805 | .iter() |
| 4806 | .filter(|ev| ev.event == "turn.started") |
| 4807 | .count(); |
| 4808 | let completed = events |
| 4809 | .iter() |
| 4810 | .filter(|ev| ev.event == "turn.completed") |
| 4811 | .count(); |
| 4812 | assert_eq!(started, 2); |
| 4813 | assert_eq!(completed, 2); |
| 4814 | Ok(()) |
| 4815 | } |
| 4816 | |
| 4817 | #[tokio::test] |
| 4818 | async fn get_thread_detail_batches_items_by_turn_without_losing_order() -> Result<()> { |
| 4819 | let manager = test_manager(test_runtime_dir())?; |
| 4820 | let thread = manager |
| 4821 | .create_thread(CreateThreadRequest { |
| 4822 | model: None, |
| 4823 | workspace: None, |
| 4824 | mode: None, |
| 4825 | allow_shell: None, |
| 4826 | trust_mode: None, |
| 4827 | auto_approve: None, |
| 4828 | archived: false, |
| 4829 | system_prompt: None, |
| 4830 | task_id: None, |
| 4831 | ..Default::default() |
| 4832 | }) |
| 4833 | .await?; |
| 4834 | |
| 4835 | let base = Utc::now(); |
| 4836 | let mut first_turn = sample_turn( |
| 4837 | &thread.id, |
| 4838 | "turn_detail_batch_first", |
| 4839 | RuntimeTurnStatus::Completed, |
| 4840 | ); |
| 4841 | first_turn.created_at = base; |
| 4842 | let mut second_turn = sample_turn( |
| 4843 | &thread.id, |
| 4844 | "turn_detail_batch_second", |
| 4845 | RuntimeTurnStatus::Completed, |
| 4846 | ); |
| 4847 | second_turn.created_at = base + chrono::Duration::seconds(1); |
| 4848 | manager.store.save_turn(&first_turn)?; |
| 4849 | manager.store.save_turn(&second_turn)?; |
| 4850 | |
| 4851 | let mut first_late = sample_item( |
| 4852 | &first_turn.id, |
| 4853 | "item_detail_first_late", |
| 4854 | TurnItemLifecycleStatus::Completed, |
| 4855 | ); |
| 4856 | first_late.started_at = Some(base + chrono::Duration::seconds(5)); |
| 4857 | let mut first_early = sample_item( |
| 4858 | &first_turn.id, |
| 4859 | "item_detail_first_early", |
| 4860 | TurnItemLifecycleStatus::Completed, |
| 4861 | ); |
| 4862 | first_early.started_at = Some(base + chrono::Duration::seconds(1)); |
| 4863 | let mut second_item = sample_item( |
| 4864 | &second_turn.id, |
| 4865 | "item_detail_second", |
| 4866 | TurnItemLifecycleStatus::Completed, |
| 4867 | ); |
| 4868 | second_item.started_at = Some(base + chrono::Duration::seconds(2)); |
| 4869 | let unrelated = sample_item( |
| 4870 | "turn_detail_batch_unrelated", |
| 4871 | "item_detail_unrelated", |
| 4872 | TurnItemLifecycleStatus::Completed, |
| 4873 | ); |
| 4874 | |
| 4875 | manager.store.save_item(&first_late)?; |
| 4876 | manager.store.save_item(&second_item)?; |
| 4877 | manager.store.save_item(&unrelated)?; |
| 4878 | manager.store.save_item(&first_early)?; |
| 4879 | |
| 4880 | let detail = manager.get_thread_detail(&thread.id).await?; |
| 4881 | let item_ids: Vec<&str> = detail.items.iter().map(|item| item.id.as_str()).collect(); |
| 4882 | assert_eq!( |
| 4883 | item_ids, |
| 4884 | vec![ |
| 4885 | "item_detail_first_early", |
| 4886 | "item_detail_first_late", |
| 4887 | "item_detail_second" |
| 4888 | ] |
| 4889 | ); |
| 4890 | Ok(()) |
| 4891 | } |
| 4892 | |
| 4893 | #[tokio::test] |
| 4894 | async fn interrupt_turn_marks_interrupted_after_cleanup() -> Result<()> { |
| 4895 | let manager = test_manager(test_runtime_dir())?; |
| 4896 | let thread = manager |
| 4897 | .create_thread(CreateThreadRequest { |
| 4898 | model: None, |
| 4899 | workspace: None, |
| 4900 | mode: None, |
| 4901 | allow_shell: None, |
| 4902 | trust_mode: None, |
| 4903 | auto_approve: None, |
| 4904 | archived: false, |
| 4905 | system_prompt: None, |
| 4906 | task_id: None, |
| 4907 | ..Default::default() |
| 4908 | }) |
| 4909 | .await?; |
| 4910 | |
| 4911 | let harness = install_mock_engine(&manager, &thread.id).await; |
| 4912 | let mut rx_op = harness.rx_op; |
| 4913 | let tx_event = harness.tx_event; |
| 4914 | let cancel_token = harness.cancel_token; |
| 4915 | let cleanup_delay = Duration::from_millis(140); |
| 4916 | tokio::spawn(async move { |
| 4917 | if matches!(rx_op.recv().await, Some(Op::SendMessage { .. })) { |
| 4918 | let _ = tx_event |
| 4919 | .send(EngineEvent::TurnStarted { |
| 4920 | turn_id: "engine_turn_interrupt".to_string(), |
| 4921 | created_at: chrono::Utc::now(), |
| 4922 | route: None, |
| 4923 | }) |
| 4924 | .await; |
| 4925 | let _ = tx_event |
| 4926 | .send(EngineEvent::MessageStarted { index: 0 }) |
| 4927 | .await; |
| 4928 | let _ = tx_event |
| 4929 | .send(EngineEvent::MessageDelta { |
| 4930 | index: 0, |
| 4931 | content: "partial".to_string(), |
| 4932 | }) |
| 4933 | .await; |
| 4934 | cancel_token.cancelled().await; |
| 4935 | sleep(cleanup_delay).await; |
| 4936 | } |
| 4937 | }); |
| 4938 | |
| 4939 | let turn = manager |
| 4940 | .start_turn( |
| 4941 | &thread.id, |
| 4942 | StartTurnRequest { |
| 4943 | prompt: "interrupt me".to_string(), |
| 4944 | input_summary: None, |
| 4945 | model: None, |
| 4946 | mode: None, |
| 4947 | allow_shell: None, |
| 4948 | trust_mode: None, |
| 4949 | auto_approve: None, |
| 4950 | ..Default::default() |
| 4951 | }, |
| 4952 | ) |
| 4953 | .await?; |
| 4954 | |
| 4955 | sleep(Duration::from_millis(20)).await; |
| 4956 | let interrupted_at = Instant::now(); |
| 4957 | let interrupt_result = manager.interrupt_turn(&thread.id, &turn.id).await?; |
| 4958 | assert_eq!(interrupt_result.status, RuntimeTurnStatus::InProgress); |
| 4959 | |
| 4960 | let final_turn = wait_for_terminal_turn(&manager, &turn.id, Duration::from_secs(3)).await?; |
| 4961 | assert_eq!(final_turn.status, RuntimeTurnStatus::Interrupted); |
| 4962 | assert!( |
| 4963 | interrupted_at.elapsed() >= cleanup_delay, |
| 4964 | "turn transitioned before cleanup finished" |
| 4965 | ); |
| 4966 | |
| 4967 | let events = manager.events_since(&thread.id, None)?; |
| 4968 | let interrupt_seq = events |
| 4969 | .iter() |
| 4970 | .find(|ev| ev.event == "turn.interrupt_requested") |
| 4971 | .map(|ev| ev.seq) |
| 4972 | .context("missing turn.interrupt_requested event")?; |
| 4973 | let completed = events |
| 4974 | .iter() |
| 4975 | .find(|ev| ev.event == "turn.completed") |
| 4976 | .context("missing turn.completed event")?; |
| 4977 | assert!(completed.seq > interrupt_seq); |
| 4978 | assert_eq!( |
| 4979 | completed |
| 4980 | .payload |
| 4981 | .get("turn") |
| 4982 | .and_then(|turn| turn.get("status")) |
| 4983 | .and_then(Value::as_str), |
| 4984 | Some("interrupted") |
| 4985 | ); |
| 4986 | Ok(()) |
| 4987 | } |
| 4988 | |
| 4989 | #[tokio::test] |
| 4990 | async fn approval_required_with_stale_active_turn_is_denied() -> Result<()> { |
| 4991 | let manager = test_manager(test_runtime_dir())?; |
| 4992 | let thread = manager |
| 4993 | .create_thread(CreateThreadRequest { |
| 4994 | model: None, |
| 4995 | workspace: None, |
| 4996 | mode: None, |
| 4997 | allow_shell: None, |
| 4998 | trust_mode: None, |
| 4999 | auto_approve: Some(true), |
| 5000 | archived: false, |
| 5001 | system_prompt: None, |
| 5002 | task_id: None, |
| 5003 | ..Default::default() |
| 5004 | }) |
| 5005 | .await?; |
| 5006 | |
| 5007 | let mut harness = install_mock_engine(&manager, &thread.id).await; |
| 5008 | let turn = manager |
| 5009 | .start_turn( |
| 5010 | &thread.id, |
| 5011 | StartTurnRequest { |
| 5012 | prompt: "needs approval".to_string(), |
| 5013 | input_summary: None, |
| 5014 | model: None, |
| 5015 | mode: None, |
| 5016 | allow_shell: None, |
| 5017 | trust_mode: None, |
| 5018 | auto_approve: Some(true), |
| 5019 | ..Default::default() |
| 5020 | }, |
| 5021 | ) |
| 5022 | .await?; |
| 5023 | |
| 5024 | assert!(matches!( |
| 5025 | harness.rx_op.recv().await, |
| 5026 | Some(Op::SendMessage { .. }) |
| 5027 | )); |
| 5028 | { |
| 5029 | let mut active = manager.active.lock().await; |
| 5030 | let state = active |
| 5031 | .engines |
| 5032 | .get_mut(&thread.id) |
| 5033 | .context("missing active thread state")?; |
| 5034 | state.active_turn = None; |
| 5035 | } |
| 5036 | |
| 5037 | harness |
| 5038 | .tx_event |
| 5039 | .send(EngineEvent::ApprovalRequired { |
| 5040 | approval_key: "test_key".to_string(), |
| 5041 | approval_grouping_key: "test_key".to_string(), |
| 5042 | id: "tool_stale".to_string(), |
| 5043 | tool_name: "exec_command".to_string(), |
| 5044 | description: "stale approval".to_string(), |
| 5045 | input: serde_json::json!({}), |
| 5046 | intent_summary: None, |
| 5047 | approval_force_prompt: false, |
| 5048 | }) |
| 5049 | .await?; |
| 5050 | |
| 5051 | assert_eq!( |
| 5052 | harness.recv_approval_event().await, |
| 5053 | Some(MockApprovalEvent::Denied { |
| 5054 | id: "tool_stale".to_string(), |
| 5055 | }) |
| 5056 | ); |
| 5057 | |
| 5058 | harness |
| 5059 | .tx_event |
| 5060 | .send(EngineEvent::TurnComplete { |
| 5061 | usage: Usage { |
| 5062 | input_tokens: 0, |
| 5063 | output_tokens: 0, |
| 5064 | ..Usage::default() |
| 5065 | }, |
| 5066 | status: TurnOutcomeStatus::Completed, |
| 5067 | error: None, |
| 5068 | tool_catalog: None, |
| 5069 | base_url: None, |
| 5070 | }) |
| 5071 | .await?; |
| 5072 | |
| 5073 | let terminal = wait_for_terminal_turn(&manager, &turn.id, Duration::from_secs(2)).await?; |
| 5074 | assert_eq!(terminal.status, RuntimeTurnStatus::Completed); |
| 5075 | Ok(()) |
| 5076 | } |
| 5077 | |
| 5078 | #[tokio::test] |
| 5079 | async fn approval_required_awaits_external_decision_allow() -> Result<()> { |
| 5080 | let manager = test_manager(test_runtime_dir())?; |
| 5081 | let thread = manager |
| 5082 | .create_thread(CreateThreadRequest { |
| 5083 | model: None, |
| 5084 | workspace: None, |
| 5085 | mode: None, |
| 5086 | allow_shell: None, |
| 5087 | trust_mode: None, |
| 5088 | auto_approve: None, |
| 5089 | archived: false, |
| 5090 | system_prompt: None, |
| 5091 | task_id: None, |
| 5092 | ..Default::default() |
| 5093 | }) |
| 5094 | .await?; |
| 5095 | |
| 5096 | let mut harness = install_mock_engine(&manager, &thread.id).await; |
| 5097 | let _turn = manager |
| 5098 | .start_turn( |
| 5099 | &thread.id, |
| 5100 | StartTurnRequest { |
| 5101 | prompt: "needs approval".to_string(), |
| 5102 | input_summary: None, |
| 5103 | model: None, |
| 5104 | mode: None, |
| 5105 | allow_shell: None, |
| 5106 | trust_mode: None, |
| 5107 | auto_approve: None, |
| 5108 | ..Default::default() |
| 5109 | }, |
| 5110 | ) |
| 5111 | .await?; |
| 5112 | assert!(matches!( |
| 5113 | harness.rx_op.recv().await, |
| 5114 | Some(Op::SendMessage { .. }) |
| 5115 | )); |
| 5116 | |
| 5117 | harness |
| 5118 | .tx_event |
| 5119 | .send(EngineEvent::ApprovalRequired { |
| 5120 | approval_key: "key1".to_string(), |
| 5121 | approval_grouping_key: "key1".to_string(), |
| 5122 | id: "tool_external_allow".to_string(), |
| 5123 | tool_name: "exec_command".to_string(), |
| 5124 | description: "external allow".to_string(), |
| 5125 | input: serde_json::json!({}), |
| 5126 | intent_summary: Some("I will update the config file.".to_string()), |
| 5127 | approval_force_prompt: false, |
| 5128 | }) |
| 5129 | .await?; |
| 5130 | |
| 5131 | let deadline = Instant::now() + Duration::from_secs(2); |
| 5132 | while Instant::now() < deadline && manager.pending_approvals_count() == 0 { |
| 5133 | sleep(Duration::from_millis(20)).await; |
| 5134 | } |
| 5135 | assert_eq!(manager.pending_approvals_count(), 1); |
| 5136 | |
| 5137 | let detail = manager.get_thread_detail(&thread.id).await?; |
| 5138 | assert_eq!(detail.pending_approvals.len(), 1); |
| 5139 | assert_eq!(detail.pending_approvals[0].id, "tool_external_allow"); |
| 5140 | assert_eq!(detail.pending_approvals[0].turn_id, _turn.id); |
| 5141 | assert_eq!(detail.pending_approvals[0].tool_name, "exec_command"); |
| 5142 | assert_eq!(detail.pending_user_inputs.len(), 0); |
| 5143 | |
| 5144 | let events = manager.events_since(&thread.id, None)?; |
| 5145 | let approval_event = events |
| 5146 | .iter() |
| 5147 | .rev() |
| 5148 | .find(|event| event.event == "approval.required") |
| 5149 | .context("missing approval.required event")?; |
| 5150 | assert_eq!( |
| 5151 | approval_event |
| 5152 | .payload |
| 5153 | .get("intent_summary") |
| 5154 | .and_then(Value::as_str), |
| 5155 | Some("I will update the config file.") |
| 5156 | ); |
| 5157 | |
| 5158 | assert!(manager.deliver_external_approval( |
| 5159 | "tool_external_allow", |
| 5160 | ExternalApprovalDecision::Allow { remember: false }, |
| 5161 | )); |
| 5162 | assert_eq!( |
| 5163 | harness.recv_approval_event().await, |
| 5164 | Some(MockApprovalEvent::Approved { |
| 5165 | id: "tool_external_allow".to_string(), |
| 5166 | }) |
| 5167 | ); |
| 5168 | assert_eq!(manager.pending_approvals_count(), 0); |
| 5169 | assert!( |
| 5170 | manager |
| 5171 | .get_thread_detail(&thread.id) |
| 5172 | .await? |
| 5173 | .pending_approvals |
| 5174 | .is_empty() |
| 5175 | ); |
| 5176 | |
| 5177 | harness |
| 5178 | .tx_event |
| 5179 | .send(EngineEvent::TurnComplete { |
| 5180 | usage: Usage::default(), |
| 5181 | status: TurnOutcomeStatus::Completed, |
| 5182 | error: None, |
| 5183 | tool_catalog: None, |
| 5184 | base_url: None, |
| 5185 | }) |
| 5186 | .await?; |
| 5187 | Ok(()) |
| 5188 | } |
| 5189 | |
| 5190 | #[tokio::test] |
| 5191 | async fn user_input_snapshot_survives_reload_and_clears_after_submission() -> Result<()> { |
| 5192 | let manager = test_manager(test_runtime_dir())?; |
| 5193 | let thread = manager |
| 5194 | .create_thread(CreateThreadRequest::default()) |
| 5195 | .await?; |
| 5196 | let mut harness = install_mock_engine(&manager, &thread.id).await; |
| 5197 | let turn = manager |
| 5198 | .start_turn( |
| 5199 | &thread.id, |
| 5200 | StartTurnRequest { |
| 5201 | prompt: "needs a choice".to_string(), |
| 5202 | ..StartTurnRequest::default() |
| 5203 | }, |
| 5204 | ) |
| 5205 | .await?; |
| 5206 | assert!(matches!( |
| 5207 | harness.rx_op.recv().await, |
| 5208 | Some(Op::SendMessage { .. }) |
| 5209 | )); |
| 5210 | |
| 5211 | harness |
| 5212 | .tx_event |
| 5213 | .send(EngineEvent::UserInputRequired { |
| 5214 | id: "input_reload".to_string(), |
| 5215 | request: crate::tools::user_input::UserInputRequest { |
| 5216 | questions: vec![crate::tools::user_input::UserInputQuestion { |
| 5217 | header: "Continue".to_string(), |
| 5218 | id: "continue".to_string(), |
| 5219 | question: "Continue with the check?".to_string(), |
| 5220 | options: vec![ |
| 5221 | crate::tools::user_input::UserInputOption { |
| 5222 | label: "Yes".to_string(), |
| 5223 | description: "Continue now".to_string(), |
| 5224 | }, |
| 5225 | crate::tools::user_input::UserInputOption { |
| 5226 | label: "No".to_string(), |
| 5227 | description: "Stop here".to_string(), |
| 5228 | }, |
| 5229 | ], |
| 5230 | allow_free_text: false, |
| 5231 | multi_select: false, |
| 5232 | }], |
| 5233 | }, |
| 5234 | }) |
| 5235 | .await?; |
| 5236 | |
| 5237 | let deadline = Instant::now() + Duration::from_secs(2); |
| 5238 | let detail = loop { |
| 5239 | let detail = manager.get_thread_detail(&thread.id).await?; |
| 5240 | if !detail.pending_user_inputs.is_empty() { |
| 5241 | break detail; |
| 5242 | } |
| 5243 | if Instant::now() >= deadline { |
| 5244 | bail!("pending user input did not reach the canonical snapshot"); |
| 5245 | } |
| 5246 | sleep(Duration::from_millis(20)).await; |
| 5247 | }; |
| 5248 | assert_eq!(detail.pending_approvals.len(), 0); |
| 5249 | assert_eq!(detail.pending_user_inputs.len(), 1); |
| 5250 | assert_eq!(detail.pending_user_inputs[0].id, "input_reload"); |
| 5251 | assert_eq!(detail.pending_user_inputs[0].turn_id, turn.id); |
| 5252 | assert_eq!( |
| 5253 | detail.pending_user_inputs[0].request.questions[0].question, |
| 5254 | "Continue with the check?" |
| 5255 | ); |
| 5256 | |
| 5257 | manager |
| 5258 | .submit_user_input( |
| 5259 | &thread.id, |
| 5260 | "input_reload", |
| 5261 | crate::tools::user_input::UserInputResponse { |
| 5262 | answers: vec![crate::tools::user_input::UserInputAnswer { |
| 5263 | id: "continue".to_string(), |
| 5264 | label: "Yes".to_string(), |
| 5265 | value: "Yes".to_string(), |
| 5266 | }], |
| 5267 | }, |
| 5268 | ) |
| 5269 | .await?; |
| 5270 | match harness.recv_user_input_submission().await { |
| 5271 | Some((id, response)) => { |
| 5272 | assert_eq!(id, "input_reload"); |
| 5273 | assert_eq!(response.answers[0].id, "continue"); |
| 5274 | } |
| 5275 | other => panic!("expected submitted user input, got {other:?}"), |
| 5276 | } |
| 5277 | assert!( |
| 5278 | manager |
| 5279 | .get_thread_detail(&thread.id) |
| 5280 | .await? |
| 5281 | .pending_user_inputs |
| 5282 | .is_empty() |
| 5283 | ); |
| 5284 | assert!(manager.events_since(&thread.id, None)?.iter().any(|event| { |
| 5285 | event.event == "user_input.answered" |
| 5286 | && event.payload.get("input_id").and_then(Value::as_str) == Some("input_reload") |
| 5287 | })); |
| 5288 | |
| 5289 | harness |
| 5290 | .tx_event |
| 5291 | .send(EngineEvent::TurnComplete { |
| 5292 | usage: Usage::default(), |
| 5293 | status: TurnOutcomeStatus::Completed, |
| 5294 | error: None, |
| 5295 | tool_catalog: None, |
| 5296 | base_url: None, |
| 5297 | }) |
| 5298 | .await?; |
| 5299 | Ok(()) |
| 5300 | } |
| 5301 | |
| 5302 | #[tokio::test] |
| 5303 | async fn unknown_user_input_id_is_not_delivered_to_engine() -> Result<()> { |
| 5304 | let manager = test_manager(test_runtime_dir())?; |
| 5305 | let thread = manager |
| 5306 | .create_thread(CreateThreadRequest::default()) |
| 5307 | .await?; |
| 5308 | let mut harness = install_mock_engine(&manager, &thread.id).await; |
| 5309 | let delivered = manager |
| 5310 | .submit_user_input( |
| 5311 | &thread.id, |
| 5312 | "input_missing", |
| 5313 | crate::tools::user_input::UserInputResponse { |
| 5314 | answers: vec![crate::tools::user_input::UserInputAnswer { |
| 5315 | id: "choice".to_string(), |
| 5316 | label: "Missing".to_string(), |
| 5317 | value: "must-not-enter-engine-mailbox".to_string(), |
| 5318 | }], |
| 5319 | }, |
| 5320 | ) |
| 5321 | .await?; |
| 5322 | assert!(!delivered); |
| 5323 | assert!( |
| 5324 | tokio::time::timeout( |
| 5325 | Duration::from_millis(25), |
| 5326 | harness.recv_user_input_submission() |
| 5327 | ) |
| 5328 | .await |
| 5329 | .is_err(), |
| 5330 | "unknown request entered the engine mailbox" |
| 5331 | ); |
| 5332 | assert!(manager.events_since(&thread.id, None)?.iter().all(|event| { |
| 5333 | !matches!( |
| 5334 | event.event.as_str(), |
| 5335 | "user_input.answered" | "user_input.canceled" |
| 5336 | ) |
| 5337 | })); |
| 5338 | Ok(()) |
| 5339 | } |
| 5340 | |
| 5341 | #[tokio::test] |
| 5342 | async fn user_input_receipt_append_failure_restores_request_without_delivery() -> Result<()> { |
| 5343 | const SECRET: &str = "answer-only-for-engine-after-retry"; |
| 5344 | let manager = test_manager(test_runtime_dir())?; |
| 5345 | let thread = manager |
| 5346 | .create_thread(CreateThreadRequest::default()) |
| 5347 | .await?; |
| 5348 | let mut harness = install_mock_engine(&manager, &thread.id).await; |
| 5349 | manager.register_pending_user_input( |
| 5350 | &thread.id, |
| 5351 | PendingUserInputRequest { |
| 5352 | id: "input_retry".to_string(), |
| 5353 | turn_id: "turn_retry".to_string(), |
| 5354 | request: crate::tools::user_input::UserInputRequest { |
| 5355 | questions: Vec::new(), |
| 5356 | }, |
| 5357 | }, |
| 5358 | ); |
| 5359 | let response = || crate::tools::user_input::UserInputResponse { |
| 5360 | answers: vec![crate::tools::user_input::UserInputAnswer { |
| 5361 | id: "choice".to_string(), |
| 5362 | label: "Retry".to_string(), |
| 5363 | value: SECRET.to_string(), |
| 5364 | }], |
| 5365 | }; |
| 5366 | |
| 5367 | let fault_guard = EventAppendFaultGuard::arm(&thread.id, EventAppendTestFault::AfterSync); |
| 5368 | let error = manager |
| 5369 | .submit_user_input(&thread.id, "input_retry", response()) |
| 5370 | .await |
| 5371 | .expect_err("injected receipt append unexpectedly succeeded"); |
| 5372 | drop(fault_guard); |
| 5373 | assert!(format!("{error:#}").contains("rolled back")); |
| 5374 | assert_eq!( |
| 5375 | manager |
| 5376 | .get_thread_detail(&thread.id) |
| 5377 | .await? |
| 5378 | .pending_user_inputs |
| 5379 | .len(), |
| 5380 | 1, |
| 5381 | "retry-safe append failure removed the authoritative prompt" |
| 5382 | ); |
| 5383 | assert!( |
| 5384 | tokio::time::timeout( |
| 5385 | Duration::from_millis(25), |
| 5386 | harness.recv_user_input_submission() |
| 5387 | ) |
| 5388 | .await |
| 5389 | .is_err(), |
| 5390 | "answer reached the engine before its receipt was durable" |
| 5391 | ); |
| 5392 | |
| 5393 | assert!( |
| 5394 | manager |
| 5395 | .submit_user_input(&thread.id, "input_retry", response()) |
| 5396 | .await?, |
| 5397 | "restored request was not retryable" |
| 5398 | ); |
| 5399 | let (_, delivered) = |
| 5400 | tokio::time::timeout(Duration::from_secs(2), harness.recv_user_input_submission()) |
| 5401 | .await |
| 5402 | .context("retried answer did not reach the engine")? |
| 5403 | .context("retried answer was canceled")?; |
| 5404 | assert_eq!(delivered.answers[0].value, SECRET); |
| 5405 | let events = manager.events_since(&thread.id, None)?; |
| 5406 | assert_eq!( |
| 5407 | events |
| 5408 | .iter() |
| 5409 | .filter(|event| event.event == "user_input.answered") |
| 5410 | .count(), |
| 5411 | 1 |
| 5412 | ); |
| 5413 | assert!(!serde_json::to_string(&events)?.contains(SECRET)); |
| 5414 | Ok(()) |
| 5415 | } |
| 5416 | |
| 5417 | #[tokio::test] |
| 5418 | async fn user_input_settlement_outlives_canceled_api_future() -> Result<()> { |
| 5419 | const SECRET: &str = "answer-survives-request-disconnect"; |
| 5420 | let manager = test_manager(test_runtime_dir())?; |
| 5421 | let thread = manager |
| 5422 | .create_thread(CreateThreadRequest::default()) |
| 5423 | .await?; |
| 5424 | let mut harness = install_mock_engine(&manager, &thread.id).await; |
| 5425 | manager.register_pending_user_input( |
| 5426 | &thread.id, |
| 5427 | PendingUserInputRequest { |
| 5428 | id: "input_detached".to_string(), |
| 5429 | turn_id: "turn_detached".to_string(), |
| 5430 | request: crate::tools::user_input::UserInputRequest { |
| 5431 | questions: Vec::new(), |
| 5432 | }, |
| 5433 | }, |
| 5434 | ); |
| 5435 | |
| 5436 | let emit_guard = manager.event_emit.lock().await; |
| 5437 | let submit_manager = manager.clone(); |
| 5438 | let thread_id = thread.id.clone(); |
| 5439 | let submission = tokio::spawn(async move { |
| 5440 | submit_manager |
| 5441 | .submit_user_input( |
| 5442 | &thread_id, |
| 5443 | "input_detached", |
| 5444 | crate::tools::user_input::UserInputResponse { |
| 5445 | answers: vec![crate::tools::user_input::UserInputAnswer { |
| 5446 | id: "choice".to_string(), |
| 5447 | label: "Continue".to_string(), |
| 5448 | value: SECRET.to_string(), |
| 5449 | }], |
| 5450 | }, |
| 5451 | ) |
| 5452 | .await |
| 5453 | }); |
| 5454 | tokio::time::timeout(Duration::from_secs(2), async { |
| 5455 | loop { |
| 5456 | if manager |
| 5457 | .pending_user_inputs |
| 5458 | .lock() |
| 5459 | .get(&(thread.id.clone(), "input_detached".to_string())) |
| 5460 | .is_some_and(|entry| entry.settling) |
| 5461 | { |
| 5462 | break; |
| 5463 | } |
| 5464 | tokio::task::yield_now().await; |
| 5465 | } |
| 5466 | }) |
| 5467 | .await |
| 5468 | .context("submission did not claim the pending request")?; |
| 5469 | submission.abort(); |
| 5470 | assert!( |
| 5471 | tokio::time::timeout( |
| 5472 | Duration::from_millis(25), |
| 5473 | harness.recv_user_input_submission() |
| 5474 | ) |
| 5475 | .await |
| 5476 | .is_err(), |
| 5477 | "answer reached the engine before its receipt append was released" |
| 5478 | ); |
| 5479 | drop(emit_guard); |
| 5480 | |
| 5481 | let (_, delivered) = |
| 5482 | tokio::time::timeout(Duration::from_secs(2), harness.recv_user_input_submission()) |
| 5483 | .await |
| 5484 | .context("detached settlement did not reach the engine")? |
| 5485 | .context("detached settlement was canceled")?; |
| 5486 | assert_eq!(delivered.answers[0].value, SECRET); |
| 5487 | let detail = manager.get_thread_detail(&thread.id).await?; |
| 5488 | assert!(detail.pending_user_inputs.is_empty()); |
| 5489 | let events = manager.events_since(&thread.id, None)?; |
| 5490 | assert_eq!( |
| 5491 | events |
| 5492 | .iter() |
| 5493 | .filter(|event| event.event == "user_input.answered") |
| 5494 | .count(), |
| 5495 | 1 |
| 5496 | ); |
| 5497 | assert!(!serde_json::to_string(&events)?.contains(SECRET)); |
| 5498 | Ok(()) |
| 5499 | } |
| 5500 | |
| 5501 | #[tokio::test] |
| 5502 | async fn terminal_user_input_cancellation_is_durable_before_engine_delivery() -> Result<()> { |
| 5503 | let manager = test_manager(test_runtime_dir())?; |
| 5504 | let thread = manager |
| 5505 | .create_thread(CreateThreadRequest::default()) |
| 5506 | .await?; |
| 5507 | let mut harness = install_mock_engine(&manager, &thread.id).await; |
| 5508 | manager.register_pending_user_input( |
| 5509 | &thread.id, |
| 5510 | PendingUserInputRequest { |
| 5511 | id: "input_terminal_order".to_string(), |
| 5512 | turn_id: "turn_terminal_order".to_string(), |
| 5513 | request: crate::tools::user_input::UserInputRequest { |
| 5514 | questions: Vec::new(), |
| 5515 | }, |
| 5516 | }, |
| 5517 | ); |
| 5518 | |
| 5519 | let emit_guard = manager.event_emit.lock().await; |
| 5520 | let engine = harness.handle.clone(); |
| 5521 | let settle_manager = manager.clone(); |
| 5522 | let thread_id = thread.id.clone(); |
| 5523 | let settlement = tokio::spawn(async move { |
| 5524 | settle_manager |
| 5525 | .settle_user_inputs_for_terminal_turn(&thread_id, "turn_terminal_order", Some(engine)) |
| 5526 | .await |
| 5527 | }); |
| 5528 | tokio::time::timeout(Duration::from_secs(2), async { |
| 5529 | loop { |
| 5530 | if manager |
| 5531 | .pending_user_inputs |
| 5532 | .lock() |
| 5533 | .get(&(thread.id.clone(), "input_terminal_order".to_string())) |
| 5534 | .is_some_and(|entry| entry.settling) |
| 5535 | { |
| 5536 | break; |
| 5537 | } |
| 5538 | tokio::task::yield_now().await; |
| 5539 | } |
| 5540 | }) |
| 5541 | .await |
| 5542 | .context("terminal cancellation did not claim the request")?; |
| 5543 | assert!( |
| 5544 | tokio::time::timeout( |
| 5545 | Duration::from_millis(25), |
| 5546 | harness.recv_user_input_submission() |
| 5547 | ) |
| 5548 | .await |
| 5549 | .is_err(), |
| 5550 | "terminal cancellation reached the engine before durable append" |
| 5551 | ); |
| 5552 | drop(emit_guard); |
| 5553 | settlement |
| 5554 | .await |
| 5555 | .context("terminal settlement task panicked")??; |
| 5556 | assert!( |
| 5557 | tokio::time::timeout(Duration::from_secs(2), harness.recv_user_input_submission()) |
| 5558 | .await |
| 5559 | .context("engine did not receive terminal cancellation")? |
| 5560 | .is_none(), |
| 5561 | "terminal cancellation delivered a submitted response" |
| 5562 | ); |
| 5563 | let events = manager.events_since(&thread.id, None)?; |
| 5564 | let canceled = events |
| 5565 | .iter() |
| 5566 | .find(|event| { |
| 5567 | event.event == "user_input.canceled" |
| 5568 | && event.payload.get("input_id").and_then(Value::as_str) |
| 5569 | == Some("input_terminal_order") |
| 5570 | }) |
| 5571 | .context("missing terminal cancellation receipt")?; |
| 5572 | assert_eq!(canceled.payload["terminal"], true); |
| 5573 | assert!( |
| 5574 | manager |
| 5575 | .get_thread_detail(&thread.id) |
| 5576 | .await? |
| 5577 | .pending_user_inputs |
| 5578 | .is_empty() |
| 5579 | ); |
| 5580 | Ok(()) |
| 5581 | } |
| 5582 | |
| 5583 | #[tokio::test] |
| 5584 | async fn thread_detail_cursor_precedes_projection_reads_at_terminal_boundary() -> Result<()> { |
| 5585 | let manager = test_manager(test_runtime_dir())?; |
| 5586 | let thread = manager |
| 5587 | .create_thread(CreateThreadRequest::default()) |
| 5588 | .await?; |
| 5589 | let mut harness = install_mock_engine(&manager, &thread.id).await; |
| 5590 | let turn = manager |
| 5591 | .start_turn( |
| 5592 | &thread.id, |
| 5593 | StartTurnRequest { |
| 5594 | prompt: "complete while the snapshot is paused".to_string(), |
| 5595 | ..StartTurnRequest::default() |
| 5596 | }, |
| 5597 | ) |
| 5598 | .await?; |
| 5599 | assert!(matches!( |
| 5600 | harness.rx_op.recv().await, |
| 5601 | Some(Op::SendMessage { .. }) |
| 5602 | )); |
| 5603 | |
| 5604 | let (hook_tx, mut hook_rx) = mpsc::unbounded_channel(); |
| 5605 | manager.set_snapshot_test_hook(hook_tx); |
| 5606 | let snapshot_manager = manager.clone(); |
| 5607 | let snapshot_thread_id = thread.id.clone(); |
| 5608 | let snapshot_task = tokio::spawn(async move { |
| 5609 | snapshot_manager |
| 5610 | .get_thread_detail(&snapshot_thread_id) |
| 5611 | .await |
| 5612 | }); |
| 5613 | |
| 5614 | let point = tokio::time::timeout(Duration::from_secs(2), hook_rx.recv()) |
| 5615 | .await |
| 5616 | .context("snapshot did not capture its replay cursor")? |
| 5617 | .context("snapshot test hook closed")?; |
| 5618 | assert_eq!(point.thread_id, thread.id); |
| 5619 | |
| 5620 | harness |
| 5621 | .tx_event |
| 5622 | .send(EngineEvent::TurnStarted { |
| 5623 | turn_id: "snapshot_terminal".to_string(), |
| 5624 | created_at: Utc::now(), |
| 5625 | route: None, |
| 5626 | }) |
| 5627 | .await?; |
| 5628 | harness |
| 5629 | .tx_event |
| 5630 | .send(EngineEvent::MessageStarted { index: 0 }) |
| 5631 | .await?; |
| 5632 | harness |
| 5633 | .tx_event |
| 5634 | .send(EngineEvent::MessageComplete { index: 0 }) |
| 5635 | .await?; |
| 5636 | harness |
| 5637 | .tx_event |
| 5638 | .send(EngineEvent::TurnComplete { |
| 5639 | usage: Usage::default(), |
| 5640 | status: TurnOutcomeStatus::Completed, |
| 5641 | error: None, |
| 5642 | tool_catalog: None, |
| 5643 | base_url: None, |
| 5644 | }) |
| 5645 | .await?; |
| 5646 | |
| 5647 | assert!( |
| 5648 | tokio::time::timeout(Duration::from_millis(100), async { |
| 5649 | loop { |
| 5650 | if manager |
| 5651 | .events_since(&thread.id, Some(point.latest_seq))? |
| 5652 | .iter() |
| 5653 | .any(|event| { |
| 5654 | event.turn_id.as_deref() == Some(&turn.id) |
| 5655 | && event.event == "turn.completed" |
| 5656 | }) |
| 5657 | { |
| 5658 | break Ok::<_, anyhow::Error>(()); |
| 5659 | } |
| 5660 | sleep(Duration::from_millis(10)).await; |
| 5661 | } |
| 5662 | }) |
| 5663 | .await |
| 5664 | .is_err(), |
| 5665 | "terminal publication crossed a snapshot projection boundary" |
| 5666 | ); |
| 5667 | point |
| 5668 | .resume |
| 5669 | .send(()) |
| 5670 | .map_err(|_| anyhow!("snapshot dropped its resume barrier"))?; |
| 5671 | |
| 5672 | let detail = snapshot_task.await.context("snapshot task panicked")??; |
| 5673 | assert_eq!(detail.latest_seq, point.latest_seq); |
| 5674 | assert_eq!( |
| 5675 | detail |
| 5676 | .turns |
| 5677 | .iter() |
| 5678 | .find(|record| record.id == turn.id) |
| 5679 | .map(|record| record.status), |
| 5680 | Some(RuntimeTurnStatus::InProgress), |
| 5681 | "the paused snapshot must retain the pre-terminal projection" |
| 5682 | ); |
| 5683 | |
| 5684 | let completed_event = tokio::time::timeout(Duration::from_secs(2), async { |
| 5685 | loop { |
| 5686 | if let Some(event) = manager |
| 5687 | .events_since(&thread.id, Some(detail.latest_seq))? |
| 5688 | .into_iter() |
| 5689 | .find(|event| { |
| 5690 | event.turn_id.as_deref() == Some(&turn.id) && event.event == "turn.completed" |
| 5691 | }) |
| 5692 | { |
| 5693 | break Ok::<_, anyhow::Error>(event); |
| 5694 | } |
| 5695 | sleep(Duration::from_millis(10)).await; |
| 5696 | } |
| 5697 | }) |
| 5698 | .await |
| 5699 | .context("terminal event did not follow the released snapshot boundary")??; |
| 5700 | assert!(completed_event.seq > detail.latest_seq); |
| 5701 | assert!( |
| 5702 | manager |
| 5703 | .events_since(&thread.id, Some(detail.latest_seq))? |
| 5704 | .iter() |
| 5705 | .any(|event| event.seq == completed_event.seq), |
| 5706 | "the same terminal transition must remain replayable from the snapshot cursor" |
| 5707 | ); |
| 5708 | Ok(()) |
| 5709 | } |
| 5710 | |
| 5711 | #[tokio::test] |
| 5712 | async fn thread_detail_does_not_reenter_recovery_while_projection_is_locked() -> Result<()> { |
| 5713 | let manager = test_manager(test_runtime_dir())?; |
| 5714 | let thread = manager |
| 5715 | .create_thread(CreateThreadRequest::default()) |
| 5716 | .await?; |
| 5717 | let turn_id = "turn_recovery_queued_during_snapshot"; |
| 5718 | let mut turn = sample_turn(&thread.id, turn_id, RuntimeTurnStatus::Completed); |
| 5719 | turn.ended_at = Some(Utc::now()); |
| 5720 | turn.duration_ms = Some(1); |
| 5721 | manager.store.save_turn(&turn)?; |
| 5722 | { |
| 5723 | let _thread_mutation = manager.store.thread_mutation.lock(); |
| 5724 | let mut persisted_thread = manager.store.load_thread(&thread.id)?; |
| 5725 | persisted_thread.latest_turn_id = Some(turn_id.to_string()); |
| 5726 | manager.store.save_thread(&persisted_thread)?; |
| 5727 | } |
| 5728 | |
| 5729 | let (hook_tx, mut hook_rx) = mpsc::unbounded_channel(); |
| 5730 | manager.set_snapshot_test_hook(hook_tx); |
| 5731 | let snapshot_manager = manager.clone(); |
| 5732 | let snapshot_thread_id = thread.id.clone(); |
| 5733 | let snapshot_task = tokio::spawn(async move { |
| 5734 | snapshot_manager |
| 5735 | .get_thread_detail(&snapshot_thread_id) |
| 5736 | .await |
| 5737 | }); |
| 5738 | let point = tokio::time::timeout(Duration::from_secs(2), hook_rx.recv()) |
| 5739 | .await |
| 5740 | .context("snapshot did not acquire its projection boundary")? |
| 5741 | .context("snapshot test hook closed")?; |
| 5742 | |
| 5743 | // Queue after get_thread_detail's initial recovery flush, while its |
| 5744 | // projection lock is held. A nested get_thread call would see this receipt |
| 5745 | // and deadlock trying to reacquire that same lock. |
| 5746 | manager.queue_recovery_receipt(RecoveredTurnReceipt { |
| 5747 | turn: turn.clone(), |
| 5748 | unresolved_dynamic_tools: Vec::new(), |
| 5749 | }); |
| 5750 | point |
| 5751 | .resume |
| 5752 | .send(()) |
| 5753 | .map_err(|_| anyhow!("snapshot dropped its resume barrier"))?; |
| 5754 | let detail = tokio::time::timeout(Duration::from_secs(2), snapshot_task) |
| 5755 | .await |
| 5756 | .context("snapshot re-entered recovery while holding its projection lock")? |
| 5757 | .context("snapshot task panicked")??; |
| 5758 | assert_eq!(detail.thread.id, thread.id); |
| 5759 | assert_eq!(detail.latest_seq, point.latest_seq); |
| 5760 | assert!(manager.recovery_receipts.lock().contains_key(&thread.id)); |
| 5761 | |
| 5762 | // The next top-level observation flushes the receipt after the prior |
| 5763 | // snapshot has released its projection boundary. |
| 5764 | manager.get_thread(&thread.id).await?; |
| 5765 | let completed = manager |
| 5766 | .events_since(&thread.id, None)? |
| 5767 | .into_iter() |
| 5768 | .filter(|event| { |
| 5769 | event.event == "turn.completed" && event.turn_id.as_deref() == Some(turn_id) |
| 5770 | }) |
| 5771 | .collect::<Vec<_>>(); |
| 5772 | assert_eq!(completed.len(), 1); |
| 5773 | assert_eq!(completed[0].payload["recovered"], true); |
| 5774 | assert!(!manager.recovery_receipts.lock().contains_key(&thread.id)); |
| 5775 | Ok(()) |
| 5776 | } |
| 5777 | |
| 5778 | #[tokio::test] |
| 5779 | async fn thread_detail_materializes_stream_prefixes_before_their_delta_cursor() -> Result<()> { |
| 5780 | let manager = test_manager(test_runtime_dir())?; |
| 5781 | let thread = manager |
| 5782 | .create_thread(CreateThreadRequest::default()) |
| 5783 | .await?; |
| 5784 | let mut harness = install_mock_engine(&manager, &thread.id).await; |
| 5785 | let turn = manager |
| 5786 | .start_turn( |
| 5787 | &thread.id, |
| 5788 | StartTurnRequest { |
| 5789 | prompt: "snapshot both streamed prefixes".to_string(), |
| 5790 | ..StartTurnRequest::default() |
| 5791 | }, |
| 5792 | ) |
| 5793 | .await?; |
| 5794 | assert!(matches!( |
| 5795 | harness.rx_op.recv().await, |
| 5796 | Some(Op::SendMessage { .. }) |
| 5797 | )); |
| 5798 | |
| 5799 | harness |
| 5800 | .tx_event |
| 5801 | .send(EngineEvent::TurnStarted { |
| 5802 | turn_id: "delta_snapshot".to_string(), |
| 5803 | created_at: Utc::now(), |
| 5804 | route: None, |
| 5805 | }) |
| 5806 | .await?; |
| 5807 | harness |
| 5808 | .tx_event |
| 5809 | .send(EngineEvent::MessageStarted { index: 0 }) |
| 5810 | .await?; |
| 5811 | harness |
| 5812 | .tx_event |
| 5813 | .send(EngineEvent::MessageDelta { |
| 5814 | index: 0, |
| 5815 | content: "durable message prefix".to_string(), |
| 5816 | }) |
| 5817 | .await?; |
| 5818 | harness |
| 5819 | .tx_event |
| 5820 | .send(EngineEvent::ThinkingStarted { index: 1 }) |
| 5821 | .await?; |
| 5822 | harness |
| 5823 | .tx_event |
| 5824 | .send(EngineEvent::ThinkingDelta { |
| 5825 | index: 1, |
| 5826 | content: "durable reasoning prefix".to_string(), |
| 5827 | }) |
| 5828 | .await?; |
| 5829 | |
| 5830 | let deltas = tokio::time::timeout(Duration::from_secs(2), async { |
| 5831 | loop { |
| 5832 | let deltas = manager |
| 5833 | .events_since(&thread.id, None)? |
| 5834 | .into_iter() |
| 5835 | .filter(|event| { |
| 5836 | event.turn_id.as_deref() == Some(&turn.id) && event.event == "item.delta" |
| 5837 | }) |
| 5838 | .collect::<Vec<_>>(); |
| 5839 | if deltas.len() == 2 { |
| 5840 | break Ok::<_, anyhow::Error>(deltas); |
| 5841 | } |
| 5842 | sleep(Duration::from_millis(10)).await; |
| 5843 | } |
| 5844 | }) |
| 5845 | .await |
| 5846 | .context("stream deltas were not durably sequenced")??; |
| 5847 | |
| 5848 | let detail = manager.get_thread_detail(&thread.id).await?; |
| 5849 | let latest_delta_seq = deltas.iter().map(|event| event.seq).max().unwrap_or(0); |
| 5850 | assert!(detail.latest_seq >= latest_delta_seq); |
| 5851 | let message = detail |
| 5852 | .items |
| 5853 | .iter() |
| 5854 | .find(|item| item.kind == TurnItemKind::AgentMessage) |
| 5855 | .context("snapshot omitted the streaming message item")?; |
| 5856 | let reasoning = detail |
| 5857 | .items |
| 5858 | .iter() |
| 5859 | .find(|item| item.kind == TurnItemKind::AgentReasoning) |
| 5860 | .context("snapshot omitted the streaming reasoning item")?; |
| 5861 | assert_eq!(message.status, TurnItemLifecycleStatus::InProgress); |
| 5862 | assert_eq!(message.detail.as_deref(), Some("durable message prefix")); |
| 5863 | assert_eq!(reasoning.status, TurnItemLifecycleStatus::InProgress); |
| 5864 | assert_eq!( |
| 5865 | reasoning.detail.as_deref(), |
| 5866 | Some("durable reasoning prefix") |
| 5867 | ); |
| 5868 | assert_eq!( |
| 5869 | manager.store.load_item(&message.id)?.detail, |
| 5870 | message.detail, |
| 5871 | "message prefix must already be on disk before its delta cursor" |
| 5872 | ); |
| 5873 | assert_eq!( |
| 5874 | manager.store.load_item(&reasoning.id)?.detail, |
| 5875 | reasoning.detail, |
| 5876 | "reasoning prefix must already be on disk before its delta cursor" |
| 5877 | ); |
| 5878 | assert!( |
| 5879 | manager |
| 5880 | .events_since(&thread.id, Some(detail.latest_seq))? |
| 5881 | .iter() |
| 5882 | .all(|event| event.event != "item.delta"), |
| 5883 | "the snapshot itself must carry every delta at or before latest_seq" |
| 5884 | ); |
| 5885 | |
| 5886 | harness |
| 5887 | .tx_event |
| 5888 | .send(EngineEvent::TurnComplete { |
| 5889 | usage: Usage::default(), |
| 5890 | status: TurnOutcomeStatus::Interrupted, |
| 5891 | error: None, |
| 5892 | tool_catalog: None, |
| 5893 | base_url: None, |
| 5894 | }) |
| 5895 | .await?; |
| 5896 | Ok(()) |
| 5897 | } |
| 5898 | |
| 5899 | #[tokio::test] |
| 5900 | async fn thread_detail_delta_boundary_is_replay_idempotent() -> Result<()> { |
| 5901 | let manager = test_manager(test_runtime_dir())?; |
| 5902 | let thread = manager |
| 5903 | .create_thread(CreateThreadRequest::default()) |
| 5904 | .await?; |
| 5905 | let mut harness = install_mock_engine(&manager, &thread.id).await; |
| 5906 | let turn = manager |
| 5907 | .start_turn( |
| 5908 | &thread.id, |
| 5909 | StartTurnRequest { |
| 5910 | prompt: "pause a snapshot across streamed deltas".to_string(), |
| 5911 | ..StartTurnRequest::default() |
| 5912 | }, |
| 5913 | ) |
| 5914 | .await?; |
| 5915 | assert!(matches!( |
| 5916 | harness.rx_op.recv().await, |
| 5917 | Some(Op::SendMessage { .. }) |
| 5918 | )); |
| 5919 | |
| 5920 | harness |
| 5921 | .tx_event |
| 5922 | .send(EngineEvent::TurnStarted { |
| 5923 | turn_id: "delta_boundary".to_string(), |
| 5924 | created_at: Utc::now(), |
| 5925 | route: None, |
| 5926 | }) |
| 5927 | .await?; |
| 5928 | harness |
| 5929 | .tx_event |
| 5930 | .send(EngineEvent::MessageStarted { index: 0 }) |
| 5931 | .await?; |
| 5932 | |
| 5933 | let started = tokio::time::timeout(Duration::from_secs(2), async { |
| 5934 | loop { |
| 5935 | if let Some(event) = manager |
| 5936 | .events_since(&thread.id, None)? |
| 5937 | .into_iter() |
| 5938 | .find(|event| { |
| 5939 | event.turn_id.as_deref() == Some(&turn.id) |
| 5940 | && event.event == "item.started" |
| 5941 | && event.payload.pointer("/item/kind").and_then(Value::as_str) |
| 5942 | == Some("agent_message") |
| 5943 | }) |
| 5944 | { |
| 5945 | break Ok::<_, anyhow::Error>(event); |
| 5946 | } |
| 5947 | sleep(Duration::from_millis(10)).await; |
| 5948 | } |
| 5949 | }) |
| 5950 | .await |
| 5951 | .context("message item did not start")??; |
| 5952 | let item_id = started.item_id.context("started event omitted item id")?; |
| 5953 | |
| 5954 | let (hook_tx, mut hook_rx) = mpsc::unbounded_channel(); |
| 5955 | manager.set_snapshot_test_hook(hook_tx); |
| 5956 | let snapshot_manager = manager.clone(); |
| 5957 | let snapshot_thread_id = thread.id.clone(); |
| 5958 | let snapshot_task = tokio::spawn(async move { |
| 5959 | snapshot_manager |
| 5960 | .get_thread_detail(&snapshot_thread_id) |
| 5961 | .await |
| 5962 | }); |
| 5963 | let point = tokio::time::timeout(Duration::from_secs(2), hook_rx.recv()) |
| 5964 | .await |
| 5965 | .context("snapshot did not capture its replay cursor")? |
| 5966 | .context("snapshot test hook closed")?; |
| 5967 | assert!(point.latest_seq >= started.seq); |
| 5968 | |
| 5969 | for content in ["A", "B"] { |
| 5970 | harness |
| 5971 | .tx_event |
| 5972 | .send(EngineEvent::MessageDelta { |
| 5973 | index: 0, |
| 5974 | content: content.to_string(), |
| 5975 | }) |
| 5976 | .await?; |
| 5977 | } |
| 5978 | sleep(STREAM_DELTA_BATCH_MAX_LATENCY + Duration::from_millis(50)).await; |
| 5979 | assert!( |
| 5980 | manager |
| 5981 | .events_since(&thread.id, Some(point.latest_seq))? |
| 5982 | .iter() |
| 5983 | .all(|event| event.event != "item.delta"), |
| 5984 | "a delta must not publish while the snapshot holds its projection boundary" |
| 5985 | ); |
| 5986 | |
| 5987 | point |
| 5988 | .resume |
| 5989 | .send(()) |
| 5990 | .map_err(|_| anyhow!("snapshot dropped its resume barrier"))?; |
| 5991 | let detail = snapshot_task.await.context("snapshot task panicked")??; |
| 5992 | let snapshotted = detail |
| 5993 | .items |
| 5994 | .iter() |
| 5995 | .find(|item| item.id == item_id) |
| 5996 | .context("snapshot omitted the streaming item")?; |
| 5997 | assert_eq!(snapshotted.detail.as_deref(), Some("")); |
| 5998 | |
| 5999 | let delta = tokio::time::timeout(Duration::from_secs(2), async { |
| 6000 | loop { |
| 6001 | if let Some(event) = manager |
| 6002 | .events_since(&thread.id, Some(detail.latest_seq))? |
| 6003 | .into_iter() |
| 6004 | .find(|event| { |
| 6005 | event.item_id.as_deref() == Some(&item_id) && event.event == "item.delta" |
| 6006 | }) |
| 6007 | { |
| 6008 | break Ok::<_, anyhow::Error>(event); |
| 6009 | } |
| 6010 | sleep(Duration::from_millis(10)).await; |
| 6011 | } |
| 6012 | }) |
| 6013 | .await |
| 6014 | .context("batched delta did not publish after snapshot release")??; |
| 6015 | assert_eq!( |
| 6016 | delta.payload.get("delta").and_then(Value::as_str), |
| 6017 | Some("AB") |
| 6018 | ); |
| 6019 | assert_eq!( |
| 6020 | manager.store.load_item(&item_id)?.detail.as_deref(), |
| 6021 | Some("AB") |
| 6022 | ); |
| 6023 | |
| 6024 | harness |
| 6025 | .tx_event |
| 6026 | .send(EngineEvent::TurnComplete { |
| 6027 | usage: Usage::default(), |
| 6028 | status: TurnOutcomeStatus::Interrupted, |
| 6029 | error: None, |
| 6030 | tool_catalog: None, |
| 6031 | base_url: None, |
| 6032 | }) |
| 6033 | .await?; |
| 6034 | Ok(()) |
| 6035 | } |
| 6036 | |
| 6037 | #[tokio::test] |
| 6038 | async fn terminal_turn_cancels_pending_user_input_and_clears_snapshot() -> Result<()> { |
| 6039 | let manager = test_manager(test_runtime_dir())?; |
| 6040 | let thread = manager |
| 6041 | .create_thread(CreateThreadRequest::default()) |
| 6042 | .await?; |
| 6043 | let mut harness = install_mock_engine(&manager, &thread.id).await; |
| 6044 | let turn = manager |
| 6045 | .start_turn( |
| 6046 | &thread.id, |
| 6047 | StartTurnRequest { |
| 6048 | prompt: "needs input before completion".to_string(), |
| 6049 | ..StartTurnRequest::default() |
| 6050 | }, |
| 6051 | ) |
| 6052 | .await?; |
| 6053 | assert!(matches!( |
| 6054 | harness.rx_op.recv().await, |
| 6055 | Some(Op::SendMessage { .. }) |
| 6056 | )); |
| 6057 | harness |
| 6058 | .tx_event |
| 6059 | .send(EngineEvent::UserInputRequired { |
| 6060 | id: "input_terminal".to_string(), |
| 6061 | request: crate::tools::user_input::UserInputRequest { |
| 6062 | questions: vec![crate::tools::user_input::UserInputQuestion { |
| 6063 | header: "Continue".to_string(), |
| 6064 | id: "continue".to_string(), |
| 6065 | question: "Continue?".to_string(), |
| 6066 | options: vec![crate::tools::user_input::UserInputOption { |
| 6067 | label: "Yes".to_string(), |
| 6068 | description: "Continue now".to_string(), |
| 6069 | }], |
| 6070 | allow_free_text: false, |
| 6071 | multi_select: false, |
| 6072 | }], |
| 6073 | }, |
| 6074 | }) |
| 6075 | .await?; |
| 6076 | |
| 6077 | let deadline = Instant::now() + Duration::from_secs(2); |
| 6078 | loop { |
| 6079 | if !manager |
| 6080 | .get_thread_detail(&thread.id) |
| 6081 | .await? |
| 6082 | .pending_user_inputs |
| 6083 | .is_empty() |
| 6084 | { |
| 6085 | break; |
| 6086 | } |
| 6087 | if Instant::now() >= deadline { |
| 6088 | bail!("pending user input did not reach the canonical snapshot"); |
| 6089 | } |
| 6090 | sleep(Duration::from_millis(20)).await; |
| 6091 | } |
| 6092 | |
| 6093 | harness |
| 6094 | .tx_event |
| 6095 | .send(EngineEvent::TurnComplete { |
| 6096 | usage: Usage::default(), |
| 6097 | status: TurnOutcomeStatus::Completed, |
| 6098 | error: None, |
| 6099 | tool_catalog: None, |
| 6100 | base_url: None, |
| 6101 | }) |
| 6102 | .await?; |
| 6103 | let canceled = tokio::time::timeout( |
| 6104 | Duration::from_secs(2), |
| 6105 | harness.recv_user_input_cancellation(), |
| 6106 | ) |
| 6107 | .await |
| 6108 | .expect("terminal user-input cancellation timed out"); |
| 6109 | assert_eq!(canceled.as_deref(), Some("input_terminal")); |
| 6110 | |
| 6111 | let deadline = Instant::now() + Duration::from_secs(2); |
| 6112 | loop { |
| 6113 | let detail = manager.get_thread_detail(&thread.id).await?; |
| 6114 | if detail.pending_user_inputs.is_empty() |
| 6115 | && manager.events_since(&thread.id, None)?.iter().any(|event| { |
| 6116 | event.event == "user_input.canceled" |
| 6117 | && event.turn_id.as_deref() == Some(turn.id.as_str()) |
| 6118 | && event.payload.get("input_id").and_then(Value::as_str) |
| 6119 | == Some("input_terminal") |
| 6120 | && event.payload.get("terminal").and_then(Value::as_bool) == Some(true) |
| 6121 | }) |
| 6122 | { |
| 6123 | break; |
| 6124 | } |
| 6125 | if Instant::now() >= deadline { |
| 6126 | bail!( |
| 6127 | "terminal user input was not cleared from the snapshot with a cancellation event" |
| 6128 | ); |
| 6129 | } |
| 6130 | sleep(Duration::from_millis(20)).await; |
| 6131 | } |
| 6132 | Ok(()) |
| 6133 | } |
| 6134 | |
| 6135 | #[tokio::test] |
| 6136 | async fn dynamic_tool_result_settles_snapshot_and_emits_one_safe_resolution() -> Result<()> { |
| 6137 | use crate::tools::spec::DynamicToolExecutor; |
| 6138 | |
| 6139 | let manager = test_manager(test_runtime_dir())?; |
| 6140 | let thread = manager |
| 6141 | .create_thread(CreateThreadRequest::default()) |
| 6142 | .await?; |
| 6143 | let mut harness = install_mock_engine(&manager, &thread.id).await; |
| 6144 | let turn = manager |
| 6145 | .start_turn( |
| 6146 | &thread.id, |
| 6147 | StartTurnRequest { |
| 6148 | prompt: "run an external lookup".to_string(), |
| 6149 | ..StartTurnRequest::default() |
| 6150 | }, |
| 6151 | ) |
| 6152 | .await?; |
| 6153 | assert!(matches!( |
| 6154 | harness.rx_op.recv().await, |
| 6155 | Some(Op::SendMessage { .. }) |
| 6156 | )); |
| 6157 | harness |
| 6158 | .tx_event |
| 6159 | .send(EngineEvent::TurnStarted { |
| 6160 | turn_id: "dynamic_result".to_string(), |
| 6161 | created_at: Utc::now(), |
| 6162 | route: None, |
| 6163 | }) |
| 6164 | .await?; |
| 6165 | |
| 6166 | const RESULT_SECRET: &str = "dynamic-result-secret"; |
| 6167 | let executor = manager.clone(); |
| 6168 | let executor_thread_id = thread.id.clone(); |
| 6169 | let execution = tokio::spawn(async move { |
| 6170 | DynamicToolExecutor::execute_dynamic_tool( |
| 6171 | &executor, |
| 6172 | Some(executor_thread_id), |
| 6173 | Some("bench".to_string()), |
| 6174 | "lookup".to_string(), |
| 6175 | json!({ "record_id": "record-7" }), |
| 6176 | ) |
| 6177 | .await |
| 6178 | }); |
| 6179 | |
| 6180 | let pending = tokio::time::timeout(Duration::from_secs(2), async { |
| 6181 | loop { |
| 6182 | let detail = manager.get_thread_detail(&thread.id).await?; |
| 6183 | if let Some(call) = detail.pending_dynamic_tool_calls.first() { |
| 6184 | break Ok::<_, anyhow::Error>((detail.latest_seq, call.clone())); |
| 6185 | } |
| 6186 | sleep(Duration::from_millis(10)).await; |
| 6187 | } |
| 6188 | }) |
| 6189 | .await |
| 6190 | .context("dynamic tool call did not reach the canonical snapshot")??; |
| 6191 | let (snapshot_seq, call) = pending; |
| 6192 | assert_eq!(call.thread_id, thread.id); |
| 6193 | assert_eq!(call.turn_id, turn.id); |
| 6194 | assert_eq!(call.namespace.as_deref(), Some("bench")); |
| 6195 | assert_eq!(call.tool, "lookup"); |
| 6196 | assert_eq!(call.arguments["record_id"], "record-7"); |
| 6197 | let requested = manager |
| 6198 | .events_since(&thread.id, None)? |
| 6199 | .into_iter() |
| 6200 | .find(|event| { |
| 6201 | event.event == "tool_call.requested" |
| 6202 | && event.payload.get("call_id").and_then(Value::as_str) |
| 6203 | == Some(call.call_id.as_str()) |
| 6204 | }) |
| 6205 | .context("dynamic tool request was not durable")?; |
| 6206 | assert!(requested.seq <= snapshot_seq); |
| 6207 | assert!( |
| 6208 | manager |
| 6209 | .events_since(&thread.id, Some(snapshot_seq))? |
| 6210 | .iter() |
| 6211 | .all(|event| event.event != "tool_call.requested"), |
| 6212 | "the pending call must be recoverable from the snapshot once replay starts at latest_seq" |
| 6213 | ); |
| 6214 | |
| 6215 | assert!( |
| 6216 | manager |
| 6217 | .deliver_dynamic_tool_result( |
| 6218 | &thread.id, |
| 6219 | &turn.id, |
| 6220 | &call.call_id, |
| 6221 | DynamicToolCallResult { |
| 6222 | success: true, |
| 6223 | content: vec![DynamicToolCallContent::InputText { |
| 6224 | text: RESULT_SECRET.to_string(), |
| 6225 | }], |
| 6226 | }, |
| 6227 | ) |
| 6228 | .await? |
| 6229 | ); |
| 6230 | let result = execution.await.context("dynamic tool task panicked")??; |
| 6231 | assert_eq!( |
| 6232 | result.content, RESULT_SECRET, |
| 6233 | "the model-facing result changed" |
| 6234 | ); |
| 6235 | assert!( |
| 6236 | manager |
| 6237 | .get_thread_detail(&thread.id) |
| 6238 | .await? |
| 6239 | .pending_dynamic_tool_calls |
| 6240 | .is_empty() |
| 6241 | ); |
| 6242 | let resolved = manager |
| 6243 | .events_since(&thread.id, None)? |
| 6244 | .into_iter() |
| 6245 | .filter(|event| { |
| 6246 | event.event == "tool_call.resolved" |
| 6247 | && event.payload.get("call_id").and_then(Value::as_str) |
| 6248 | == Some(call.call_id.as_str()) |
| 6249 | }) |
| 6250 | .collect::<Vec<_>>(); |
| 6251 | assert_eq!(resolved.len(), 1); |
| 6252 | assert_eq!(resolved[0].payload["status"], "resolved"); |
| 6253 | assert_eq!(resolved[0].payload["success"], true); |
| 6254 | assert!( |
| 6255 | !serde_json::to_string(&resolved)?.contains(RESULT_SECRET), |
| 6256 | "terminal dynamic-tool lifecycle must not echo result content" |
| 6257 | ); |
| 6258 | assert!( |
| 6259 | !manager |
| 6260 | .deliver_dynamic_tool_result( |
| 6261 | &thread.id, |
| 6262 | &turn.id, |
| 6263 | &call.call_id, |
| 6264 | DynamicToolCallResult { |
| 6265 | success: true, |
| 6266 | content: Vec::new(), |
| 6267 | }, |
| 6268 | ) |
| 6269 | .await?, |
| 6270 | "a duplicate result must not settle or emit twice" |
| 6271 | ); |
| 6272 | assert_eq!( |
| 6273 | manager |
| 6274 | .events_since(&thread.id, None)? |
| 6275 | .iter() |
| 6276 | .filter(|event| event.event == "tool_call.resolved") |
| 6277 | .count(), |
| 6278 | 1 |
| 6279 | ); |
| 6280 | |
| 6281 | harness |
| 6282 | .tx_event |
| 6283 | .send(EngineEvent::TurnComplete { |
| 6284 | usage: Usage::default(), |
| 6285 | status: TurnOutcomeStatus::Completed, |
| 6286 | error: None, |
| 6287 | tool_catalog: None, |
| 6288 | base_url: None, |
| 6289 | }) |
| 6290 | .await?; |
| 6291 | Ok(()) |
| 6292 | } |
| 6293 | |
| 6294 | #[tokio::test] |
| 6295 | async fn dynamic_tool_result_receipt_outlives_canceled_delivery_future() -> Result<()> { |
| 6296 | use crate::tools::spec::DynamicToolExecutor; |
| 6297 | |
| 6298 | let manager = test_manager(test_runtime_dir())?; |
| 6299 | let thread = manager |
| 6300 | .create_thread(CreateThreadRequest::default()) |
| 6301 | .await?; |
| 6302 | let mut harness = install_mock_engine(&manager, &thread.id).await; |
| 6303 | let turn = manager |
| 6304 | .start_turn( |
| 6305 | &thread.id, |
| 6306 | StartTurnRequest { |
| 6307 | prompt: "settle a result after its HTTP future disappears".to_string(), |
| 6308 | ..StartTurnRequest::default() |
| 6309 | }, |
| 6310 | ) |
| 6311 | .await?; |
| 6312 | assert!(matches!( |
| 6313 | harness.rx_op.recv().await, |
| 6314 | Some(Op::SendMessage { .. }) |
| 6315 | )); |
| 6316 | harness |
| 6317 | .tx_event |
| 6318 | .send(EngineEvent::TurnStarted { |
| 6319 | turn_id: "dynamic_detached_settlement".to_string(), |
| 6320 | created_at: Utc::now(), |
| 6321 | route: None, |
| 6322 | }) |
| 6323 | .await?; |
| 6324 | |
| 6325 | const RESULT_SECRET: &str = "detached-dynamic-result-secret"; |
| 6326 | let executor = manager.clone(); |
| 6327 | let executor_thread_id = thread.id.clone(); |
| 6328 | let execution = tokio::spawn(async move { |
| 6329 | DynamicToolExecutor::execute_dynamic_tool( |
| 6330 | &executor, |
| 6331 | Some(executor_thread_id), |
| 6332 | Some("bench".to_string()), |
| 6333 | "detached_lookup".to_string(), |
| 6334 | json!({ "record_id": "record-detached" }), |
| 6335 | ) |
| 6336 | .await |
| 6337 | }); |
| 6338 | |
| 6339 | let call = tokio::time::timeout(Duration::from_secs(2), async { |
| 6340 | loop { |
| 6341 | if let Some(call) = manager |
| 6342 | .get_thread_detail(&thread.id) |
| 6343 | .await? |
| 6344 | .pending_dynamic_tool_calls |
| 6345 | .first() |
| 6346 | .cloned() |
| 6347 | { |
| 6348 | break Ok::<_, anyhow::Error>(call); |
| 6349 | } |
| 6350 | tokio::task::yield_now().await; |
| 6351 | } |
| 6352 | }) |
| 6353 | .await |
| 6354 | .context("dynamic tool call did not become pending")??; |
| 6355 | |
| 6356 | // Stall terminal publication, submit the result, and wait until that path |
| 6357 | // owns the call. Canceling the API future from this point must not cancel |
| 6358 | // its detached receipt task or wake the model before the receipt is durable. |
| 6359 | let emit_guard = manager.event_emit.lock().await; |
| 6360 | let delivery_manager = manager.clone(); |
| 6361 | let delivery_thread_id = thread.id.clone(); |
| 6362 | let delivery_turn_id = turn.id.clone(); |
| 6363 | let delivery_call_id = call.call_id.clone(); |
| 6364 | let delivery = tokio::spawn(async move { |
| 6365 | delivery_manager |
| 6366 | .deliver_dynamic_tool_result( |
| 6367 | &delivery_thread_id, |
| 6368 | &delivery_turn_id, |
| 6369 | &delivery_call_id, |
| 6370 | DynamicToolCallResult { |
| 6371 | success: true, |
| 6372 | content: vec![DynamicToolCallContent::InputText { |
| 6373 | text: RESULT_SECRET.to_string(), |
| 6374 | }], |
| 6375 | }, |
| 6376 | ) |
| 6377 | .await |
| 6378 | }); |
| 6379 | tokio::time::timeout(Duration::from_secs(2), async { |
| 6380 | loop { |
| 6381 | let settling = manager |
| 6382 | .pending_dynamic_tools |
| 6383 | .lock() |
| 6384 | .get(&call.call_id) |
| 6385 | .is_some_and(|entry| entry.sender.is_none()); |
| 6386 | if settling { |
| 6387 | break; |
| 6388 | } |
| 6389 | tokio::task::yield_now().await; |
| 6390 | } |
| 6391 | }) |
| 6392 | .await |
| 6393 | .context("result delivery never claimed the pending call")?; |
| 6394 | assert!( |
| 6395 | !execution.is_finished(), |
| 6396 | "the model consumed the result before its terminal receipt could commit" |
| 6397 | ); |
| 6398 | assert!( |
| 6399 | !manager |
| 6400 | .deliver_dynamic_tool_result( |
| 6401 | &thread.id, |
| 6402 | &turn.id, |
| 6403 | &call.call_id, |
| 6404 | DynamicToolCallResult { |
| 6405 | success: false, |
| 6406 | content: Vec::new(), |
| 6407 | }, |
| 6408 | ) |
| 6409 | .await?, |
| 6410 | "a duplicate result stole a call whose terminal receipt was settling" |
| 6411 | ); |
| 6412 | delivery.abort(); |
| 6413 | assert!( |
| 6414 | delivery |
| 6415 | .await |
| 6416 | .expect_err("delivery API future must be canceled") |
| 6417 | .is_cancelled() |
| 6418 | ); |
| 6419 | assert!( |
| 6420 | !execution.is_finished(), |
| 6421 | "canceling the delivery future woke the model without a receipt" |
| 6422 | ); |
| 6423 | |
| 6424 | drop(emit_guard); |
| 6425 | let model_result = tokio::time::timeout(Duration::from_secs(2), execution) |
| 6426 | .await |
| 6427 | .context("model did not receive the result after terminal publication")? |
| 6428 | .context("dynamic tool task panicked")??; |
| 6429 | assert_eq!(model_result.content, RESULT_SECRET); |
| 6430 | |
| 6431 | let detail = manager.get_thread_detail(&thread.id).await?; |
| 6432 | assert!(detail.pending_dynamic_tool_calls.is_empty()); |
| 6433 | let terminal = manager |
| 6434 | .events_since(&thread.id, None)? |
| 6435 | .into_iter() |
| 6436 | .filter(|event| { |
| 6437 | matches!( |
| 6438 | event.event.as_str(), |
| 6439 | "tool_call.resolved" | "tool_call.timeout" | "tool_call.canceled" |
| 6440 | ) && event.payload.get("call_id").and_then(Value::as_str) == Some(call.call_id.as_str()) |
| 6441 | }) |
| 6442 | .collect::<Vec<_>>(); |
| 6443 | assert_eq!(terminal.len(), 1); |
| 6444 | assert_eq!(terminal[0].event, "tool_call.resolved"); |
| 6445 | assert_eq!(terminal[0].payload["success"], true); |
| 6446 | assert!( |
| 6447 | !serde_json::to_string(&terminal)?.contains(RESULT_SECRET), |
| 6448 | "the terminal receipt exposed result content" |
| 6449 | ); |
| 6450 | assert!( |
| 6451 | !manager |
| 6452 | .deliver_dynamic_tool_result( |
| 6453 | &thread.id, |
| 6454 | &turn.id, |
| 6455 | &call.call_id, |
| 6456 | DynamicToolCallResult { |
| 6457 | success: true, |
| 6458 | content: Vec::new(), |
| 6459 | }, |
| 6460 | ) |
| 6461 | .await?, |
| 6462 | "a duplicate retry settled an already terminal call" |
| 6463 | ); |
| 6464 | |
| 6465 | harness |
| 6466 | .tx_event |
| 6467 | .send(EngineEvent::TurnComplete { |
| 6468 | usage: Usage::default(), |
| 6469 | status: TurnOutcomeStatus::Completed, |
| 6470 | error: None, |
| 6471 | tool_catalog: None, |
| 6472 | base_url: None, |
| 6473 | }) |
| 6474 | .await?; |
| 6475 | Ok(()) |
| 6476 | } |
| 6477 | |
| 6478 | #[tokio::test] |
| 6479 | async fn dynamic_tool_result_acceptance_survives_receiver_close_before_append() -> Result<()> { |
| 6480 | let manager = test_manager(test_runtime_dir())?; |
| 6481 | let thread = manager |
| 6482 | .create_thread(CreateThreadRequest::default()) |
| 6483 | .await?; |
| 6484 | let receiver = manager.register_pending_dynamic_tool_for_test( |
| 6485 | &thread.id, |
| 6486 | "turn_closed_receiver", |
| 6487 | "call_closed_receiver", |
| 6488 | )?; |
| 6489 | let emit_guard = manager.event_emit.lock().await; |
| 6490 | let delivery_manager = manager.clone(); |
| 6491 | let delivery_thread_id = thread.id.clone(); |
| 6492 | let delivery = tokio::spawn(async move { |
| 6493 | delivery_manager |
| 6494 | .deliver_dynamic_tool_result( |
| 6495 | &delivery_thread_id, |
| 6496 | "turn_closed_receiver", |
| 6497 | "call_closed_receiver", |
| 6498 | DynamicToolCallResult { |
| 6499 | success: true, |
| 6500 | content: vec![DynamicToolCallContent::InputText { |
| 6501 | text: "closed-receiver-secret".to_string(), |
| 6502 | }], |
| 6503 | }, |
| 6504 | ) |
| 6505 | .await |
| 6506 | }); |
| 6507 | tokio::time::timeout(Duration::from_secs(2), async { |
| 6508 | loop { |
| 6509 | if manager |
| 6510 | .pending_dynamic_tools |
| 6511 | .lock() |
| 6512 | .get("call_closed_receiver") |
| 6513 | .is_some_and(|entry| entry.sender.is_none()) |
| 6514 | { |
| 6515 | break; |
| 6516 | } |
| 6517 | tokio::task::yield_now().await; |
| 6518 | } |
| 6519 | }) |
| 6520 | .await |
| 6521 | .context("result did not claim the call before append")?; |
| 6522 | drop(receiver); |
| 6523 | drop(emit_guard); |
| 6524 | |
| 6525 | assert!( |
| 6526 | delivery.await.context("delivery task panicked")??, |
| 6527 | "durably accepted result was reported as missing after receiver close" |
| 6528 | ); |
| 6529 | assert!( |
| 6530 | manager |
| 6531 | .get_thread_detail(&thread.id) |
| 6532 | .await? |
| 6533 | .pending_dynamic_tool_calls |
| 6534 | .is_empty() |
| 6535 | ); |
| 6536 | let terminal = manager |
| 6537 | .events_since(&thread.id, None)? |
| 6538 | .into_iter() |
| 6539 | .filter(|event| { |
| 6540 | event.payload.get("call_id").and_then(Value::as_str) == Some("call_closed_receiver") |
| 6541 | && matches!( |
| 6542 | event.event.as_str(), |
| 6543 | "tool_call.resolved" | "tool_call.timeout" | "tool_call.canceled" |
| 6544 | ) |
| 6545 | }) |
| 6546 | .collect::<Vec<_>>(); |
| 6547 | assert_eq!(terminal.len(), 1); |
| 6548 | assert_eq!(terminal[0].event, "tool_call.resolved"); |
| 6549 | assert_eq!(terminal[0].payload["result_accepted"], true); |
| 6550 | assert!( |
| 6551 | !serde_json::to_string(&terminal)?.contains("closed-receiver-secret"), |
| 6552 | "closed-receiver acceptance exposed result content" |
| 6553 | ); |
| 6554 | Ok(()) |
| 6555 | } |
| 6556 | |
| 6557 | #[cfg(unix)] |
| 6558 | #[tokio::test] |
| 6559 | async fn dynamic_tool_receipt_append_failure_rolls_back_for_retry() -> Result<()> { |
| 6560 | let manager = test_manager(test_runtime_dir())?; |
| 6561 | let thread = manager |
| 6562 | .create_thread(CreateThreadRequest::default()) |
| 6563 | .await?; |
| 6564 | let mut receiver = manager.register_pending_dynamic_tool_for_test( |
| 6565 | &thread.id, |
| 6566 | "turn_retry_after_append_failure", |
| 6567 | "call_retry_after_append_failure", |
| 6568 | )?; |
| 6569 | let claim = match manager.claim_pending_dynamic_tool( |
| 6570 | &thread.id, |
| 6571 | "turn_retry_after_append_failure", |
| 6572 | "call_retry_after_append_failure", |
| 6573 | ) { |
| 6574 | PendingDynamicToolClaim::Claimed(claim) => claim, |
| 6575 | PendingDynamicToolClaim::Settling(_) |
| 6576 | | PendingDynamicToolClaim::Indeterminate |
| 6577 | | PendingDynamicToolClaim::Missing => { |
| 6578 | bail!("failed to claim the append-failure fixture") |
| 6579 | } |
| 6580 | }; |
| 6581 | |
| 6582 | // Replace this throwaway thread's event log with a symlink. Runtime store |
| 6583 | // hardening rejects the append deterministically, exercising settlement |
| 6584 | // rollback without relying on platform permission behavior. |
| 6585 | let events_path = manager.store.events_path(&thread.id)?; |
| 6586 | let backup_path = events_path.with_extension("jsonl.append-failure-backup"); |
| 6587 | std::fs::rename(&events_path, &backup_path)?; |
| 6588 | std::os::unix::fs::symlink(&backup_path, &events_path)?; |
| 6589 | let ack = manager.spawn_dynamic_tool_settlement( |
| 6590 | claim, |
| 6591 | DynamicToolTerminalOutcome::Resolved(DynamicToolCallResult { |
| 6592 | success: true, |
| 6593 | content: vec![DynamicToolCallContent::InputText { |
| 6594 | text: "discarded-before-retry".to_string(), |
| 6595 | }], |
| 6596 | }), |
| 6597 | ); |
| 6598 | let failed = RuntimeThreadManager::await_dynamic_tool_settlement(ack).await; |
| 6599 | std::fs::remove_file(&events_path)?; |
| 6600 | std::fs::rename(&backup_path, &events_path)?; |
| 6601 | assert!( |
| 6602 | failed.is_err(), |
| 6603 | "symlinked event append unexpectedly succeeded" |
| 6604 | ); |
| 6605 | |
| 6606 | { |
| 6607 | let pending = manager.pending_dynamic_tools.lock(); |
| 6608 | let entry = pending |
| 6609 | .get("call_retry_after_append_failure") |
| 6610 | .context("failed receipt append stranded or removed the pending call")?; |
| 6611 | assert!( |
| 6612 | entry |
| 6613 | .sender |
| 6614 | .as_ref() |
| 6615 | .is_some_and(|sender| !sender.is_closed()), |
| 6616 | "failed receipt append left a Settling entry without its sender" |
| 6617 | ); |
| 6618 | } |
| 6619 | assert_eq!( |
| 6620 | manager |
| 6621 | .get_thread_detail(&thread.id) |
| 6622 | .await? |
| 6623 | .pending_dynamic_tool_calls |
| 6624 | .len(), |
| 6625 | 1, |
| 6626 | "rollback removed the snapshot-authoritative pending request" |
| 6627 | ); |
| 6628 | |
| 6629 | assert!( |
| 6630 | manager |
| 6631 | .deliver_dynamic_tool_result( |
| 6632 | &thread.id, |
| 6633 | "turn_retry_after_append_failure", |
| 6634 | "call_retry_after_append_failure", |
| 6635 | DynamicToolCallResult { |
| 6636 | success: true, |
| 6637 | content: vec![DynamicToolCallContent::InputText { |
| 6638 | text: "retry-result".to_string(), |
| 6639 | }], |
| 6640 | }, |
| 6641 | ) |
| 6642 | .await?, |
| 6643 | "retry did not settle the restored call" |
| 6644 | ); |
| 6645 | let delivered = tokio::time::timeout(Duration::from_secs(2), &mut receiver) |
| 6646 | .await |
| 6647 | .context("restored receiver was not woken by retry")??; |
| 6648 | assert_eq!( |
| 6649 | delivered.content, |
| 6650 | vec![DynamicToolCallContent::InputText { |
| 6651 | text: "retry-result".to_string(), |
| 6652 | }] |
| 6653 | ); |
| 6654 | let resolved = manager |
| 6655 | .events_since(&thread.id, None)? |
| 6656 | .into_iter() |
| 6657 | .filter(|event| { |
| 6658 | event.event == "tool_call.resolved" |
| 6659 | && event.payload.get("call_id").and_then(Value::as_str) |
| 6660 | == Some("call_retry_after_append_failure") |
| 6661 | }) |
| 6662 | .collect::<Vec<_>>(); |
| 6663 | assert_eq!(resolved.len(), 1); |
| 6664 | assert!( |
| 6665 | !serde_json::to_string(&resolved)?.contains("retry-result"), |
| 6666 | "retried terminal receipt exposed result content" |
| 6667 | ); |
| 6668 | Ok(()) |
| 6669 | } |
| 6670 | |
| 6671 | #[tokio::test] |
| 6672 | async fn dynamic_tool_post_write_failures_rollback_without_duplicate_receipts() -> Result<()> { |
| 6673 | let manager = test_manager(test_runtime_dir())?; |
| 6674 | let thread = manager |
| 6675 | .create_thread(CreateThreadRequest::default()) |
| 6676 | .await?; |
| 6677 | |
| 6678 | for (index, fault) in [ |
| 6679 | EventAppendTestFault::AfterFlush, |
| 6680 | EventAppendTestFault::AfterSync, |
| 6681 | ] |
| 6682 | .into_iter() |
| 6683 | .enumerate() |
| 6684 | { |
| 6685 | let turn_id = format!("turn_post_write_{index}"); |
| 6686 | let call_id = format!("call_post_write_{index}"); |
| 6687 | let result_text = format!("post-write-result-{index}"); |
| 6688 | let mut receiver = |
| 6689 | manager.register_pending_dynamic_tool_for_test(&thread.id, &turn_id, &call_id)?; |
| 6690 | let fault_guard = EventAppendFaultGuard::arm(&thread.id, fault); |
| 6691 | let error = manager |
| 6692 | .deliver_dynamic_tool_result( |
| 6693 | &thread.id, |
| 6694 | &turn_id, |
| 6695 | &call_id, |
| 6696 | DynamicToolCallResult { |
| 6697 | success: true, |
| 6698 | content: vec![DynamicToolCallContent::InputText { |
| 6699 | text: result_text.clone(), |
| 6700 | }], |
| 6701 | }, |
| 6702 | ) |
| 6703 | .await |
| 6704 | .expect_err("injected post-write failure unexpectedly settled"); |
| 6705 | drop(fault_guard); |
| 6706 | assert!( |
| 6707 | error.to_string().contains("rolled back"), |
| 6708 | "post-write failure was not classified retry-safe: {error}" |
| 6709 | ); |
| 6710 | |
| 6711 | let failed_snapshot = manager.get_thread_detail(&thread.id).await?; |
| 6712 | assert!( |
| 6713 | failed_snapshot |
| 6714 | .pending_dynamic_tool_calls |
| 6715 | .iter() |
| 6716 | .any(|call| call.call_id == call_id), |
| 6717 | "rolled-back call disappeared from the canonical snapshot" |
| 6718 | ); |
| 6719 | assert!( |
| 6720 | manager |
| 6721 | .events_since(&thread.id, Some(failed_snapshot.latest_seq))? |
| 6722 | .is_empty(), |
| 6723 | "failed append left a replay-visible terminal suffix" |
| 6724 | ); |
| 6725 | assert!( |
| 6726 | manager.events_since(&thread.id, None)?.iter().all(|event| { |
| 6727 | event.payload.get("call_id").and_then(Value::as_str) != Some(call_id.as_str()) |
| 6728 | }), |
| 6729 | "failed append left a visible terminal record before retry" |
| 6730 | ); |
| 6731 | |
| 6732 | assert!( |
| 6733 | manager |
| 6734 | .deliver_dynamic_tool_result( |
| 6735 | &thread.id, |
| 6736 | &turn_id, |
| 6737 | &call_id, |
| 6738 | DynamicToolCallResult { |
| 6739 | success: true, |
| 6740 | content: vec![DynamicToolCallContent::InputText { |
| 6741 | text: result_text.clone(), |
| 6742 | }], |
| 6743 | }, |
| 6744 | ) |
| 6745 | .await?, |
| 6746 | "retry did not durably accept the rolled-back result" |
| 6747 | ); |
| 6748 | let delivered = tokio::time::timeout(Duration::from_secs(2), &mut receiver) |
| 6749 | .await |
| 6750 | .context("retried result did not reach its model receiver")??; |
| 6751 | assert_eq!( |
| 6752 | delivered.content, |
| 6753 | vec![DynamicToolCallContent::InputText { |
| 6754 | text: result_text.clone(), |
| 6755 | }] |
| 6756 | ); |
| 6757 | |
| 6758 | let replay = manager.events_since(&thread.id, Some(failed_snapshot.latest_seq))?; |
| 6759 | let terminal = replay |
| 6760 | .iter() |
| 6761 | .filter(|event| { |
| 6762 | event.event == "tool_call.resolved" |
| 6763 | && event.payload.get("call_id").and_then(Value::as_str) |
| 6764 | == Some(call_id.as_str()) |
| 6765 | }) |
| 6766 | .collect::<Vec<_>>(); |
| 6767 | assert_eq!(terminal.len(), 1); |
| 6768 | assert!(terminal[0].seq > failed_snapshot.latest_seq); |
| 6769 | assert_eq!(terminal[0].payload["result_accepted"], true); |
| 6770 | assert!( |
| 6771 | !serde_json::to_string(&terminal)?.contains(&result_text), |
| 6772 | "retried terminal receipt exposed result content" |
| 6773 | ); |
| 6774 | let settled_snapshot = manager.get_thread_detail(&thread.id).await?; |
| 6775 | assert!( |
| 6776 | settled_snapshot |
| 6777 | .pending_dynamic_tool_calls |
| 6778 | .iter() |
| 6779 | .all(|call| call.call_id != call_id) |
| 6780 | ); |
| 6781 | assert!(settled_snapshot.latest_seq >= terminal[0].seq); |
| 6782 | } |
| 6783 | Ok(()) |
| 6784 | } |
| 6785 | |
| 6786 | #[tokio::test] |
| 6787 | async fn restart_recovers_terminal_turn_after_dynamic_receipt_append_failure() -> Result<()> { |
| 6788 | let data_dir = test_runtime_dir(); |
| 6789 | let manager = test_manager(data_dir.clone())?; |
| 6790 | let thread = manager |
| 6791 | .create_thread(CreateThreadRequest::default()) |
| 6792 | .await?; |
| 6793 | let turn_id = "turn_terminal_recovery"; |
| 6794 | let call_id = "call_terminal_recovery"; |
| 6795 | let mut turn = sample_turn(&thread.id, turn_id, RuntimeTurnStatus::Completed); |
| 6796 | turn.ended_at = Some(Utc::now()); |
| 6797 | turn.duration_ms = Some(1); |
| 6798 | manager.store.save_turn(&turn)?; |
| 6799 | { |
| 6800 | let _thread_mutation = manager.store.thread_mutation.lock(); |
| 6801 | let mut persisted_thread = manager.store.load_thread(&thread.id)?; |
| 6802 | persisted_thread.latest_turn_id = Some(turn_id.to_string()); |
| 6803 | manager.store.save_thread(&persisted_thread)?; |
| 6804 | } |
| 6805 | let params = DynamicToolCallParams { |
| 6806 | thread_id: thread.id.clone(), |
| 6807 | turn_id: turn_id.to_string(), |
| 6808 | call_id: call_id.to_string(), |
| 6809 | namespace: Some("recovery".to_string()), |
| 6810 | tool: "recover_lookup".to_string(), |
| 6811 | arguments: json!({ "record": "recovery-only" }), |
| 6812 | }; |
| 6813 | manager |
| 6814 | .emit_event_for_test( |
| 6815 | &thread.id, |
| 6816 | Some(turn_id), |
| 6817 | "tool_call.requested", |
| 6818 | json!(¶ms), |
| 6819 | ) |
| 6820 | .await?; |
| 6821 | let receiver = manager.register_pending_dynamic_tool(params)?; |
| 6822 | |
| 6823 | let fault_guard = EventAppendFaultGuard::arm(&thread.id, EventAppendTestFault::AfterSync); |
| 6824 | let error = manager |
| 6825 | .deliver_dynamic_tool_result( |
| 6826 | &thread.id, |
| 6827 | turn_id, |
| 6828 | call_id, |
| 6829 | DynamicToolCallResult { |
| 6830 | success: true, |
| 6831 | content: vec![DynamicToolCallContent::InputText { |
| 6832 | text: "never-committed-result".to_string(), |
| 6833 | }], |
| 6834 | }, |
| 6835 | ) |
| 6836 | .await |
| 6837 | .expect_err("injected terminal receipt append unexpectedly succeeded"); |
| 6838 | drop(fault_guard); |
| 6839 | assert!(error.to_string().contains("rolled back")); |
| 6840 | assert!(manager.events_since(&thread.id, None)?.iter().all(|event| { |
| 6841 | event.event != "turn.completed" |
| 6842 | && !matches!( |
| 6843 | event.event.as_str(), |
| 6844 | "tool_call.resolved" | "tool_call.canceled" | "tool_call.timeout" |
| 6845 | ) |
| 6846 | })); |
| 6847 | drop(receiver); |
| 6848 | drop(manager); |
| 6849 | |
| 6850 | let recovered = test_manager(data_dir.clone())?; |
| 6851 | // Opening is synchronous; the first async observation flushes queued |
| 6852 | // recovery receipts in terminal-call-before-turn order. |
| 6853 | let recovered_turn = recovered.get_thread(&thread.id).await?; |
| 6854 | assert_eq!(recovered_turn.latest_turn_id.as_deref(), Some(turn_id)); |
| 6855 | let events = recovered.events_since(&thread.id, None)?; |
| 6856 | let canceled = events |
| 6857 | .iter() |
| 6858 | .filter(|event| { |
| 6859 | event.event == "tool_call.canceled" |
| 6860 | && event.payload.get("call_id").and_then(Value::as_str) == Some(call_id) |
| 6861 | }) |
| 6862 | .collect::<Vec<_>>(); |
| 6863 | let completed = events |
| 6864 | .iter() |
| 6865 | .filter(|event| { |
| 6866 | event.event == "turn.completed" && event.turn_id.as_deref() == Some(turn_id) |
| 6867 | }) |
| 6868 | .collect::<Vec<_>>(); |
| 6869 | assert_eq!(canceled.len(), 1); |
| 6870 | assert_eq!(canceled[0].payload["reason"], "process_restart"); |
| 6871 | assert_eq!(canceled[0].payload["recovered"], true); |
| 6872 | assert_eq!(completed.len(), 1); |
| 6873 | assert_eq!(completed[0].payload["recovered"], true); |
| 6874 | assert!(canceled[0].seq < completed[0].seq); |
| 6875 | assert_eq!( |
| 6876 | completed[0] |
| 6877 | .payload |
| 6878 | .pointer("/turn/status") |
| 6879 | .and_then(Value::as_str), |
| 6880 | Some("completed") |
| 6881 | ); |
| 6882 | |
| 6883 | // Re-observation and a second manager restart both remain idempotent. |
| 6884 | recovered.get_thread(&thread.id).await?; |
| 6885 | assert_eq!( |
| 6886 | recovered |
| 6887 | .events_since(&thread.id, None)? |
| 6888 | .iter() |
| 6889 | .filter(|event| { |
| 6890 | event.event == "turn.completed" && event.turn_id.as_deref() == Some(turn_id) |
| 6891 | }) |
| 6892 | .count(), |
| 6893 | 1 |
| 6894 | ); |
| 6895 | drop(recovered); |
| 6896 | let reopened = test_manager(data_dir)?; |
| 6897 | reopened.get_thread(&thread.id).await?; |
| 6898 | let reopened_events = reopened.events_since(&thread.id, None)?; |
| 6899 | assert_eq!( |
| 6900 | reopened_events |
| 6901 | .iter() |
| 6902 | .filter(|event| { |
| 6903 | event.event == "tool_call.canceled" |
| 6904 | && event.payload.get("call_id").and_then(Value::as_str) == Some(call_id) |
| 6905 | }) |
| 6906 | .count(), |
| 6907 | 1 |
| 6908 | ); |
| 6909 | assert_eq!( |
| 6910 | reopened_events |
| 6911 | .iter() |
| 6912 | .filter(|event| { |
| 6913 | event.event == "turn.completed" && event.turn_id.as_deref() == Some(turn_id) |
| 6914 | }) |
| 6915 | .count(), |
| 6916 | 1 |
| 6917 | ); |
| 6918 | Ok(()) |
| 6919 | } |
| 6920 | |
| 6921 | #[tokio::test] |
| 6922 | async fn restart_reconciles_unresolved_dynamic_call_after_existing_turn_completion() -> Result<()> { |
| 6923 | let data_dir = test_runtime_dir(); |
| 6924 | let manager = test_manager(data_dir.clone())?; |
| 6925 | let thread = manager |
| 6926 | .create_thread(CreateThreadRequest::default()) |
| 6927 | .await?; |
| 6928 | let turn_id = "turn_legacy_completed_request"; |
| 6929 | let call_id = "call_legacy_completed_request"; |
| 6930 | let mut turn = sample_turn(&thread.id, turn_id, RuntimeTurnStatus::Completed); |
| 6931 | turn.ended_at = Some(Utc::now()); |
| 6932 | turn.duration_ms = Some(1); |
| 6933 | manager.store.save_turn(&turn)?; |
| 6934 | let params = DynamicToolCallParams { |
| 6935 | thread_id: thread.id.clone(), |
| 6936 | turn_id: turn_id.to_string(), |
| 6937 | call_id: call_id.to_string(), |
| 6938 | namespace: Some("legacy".to_string()), |
| 6939 | tool: "legacy_lookup".to_string(), |
| 6940 | arguments: json!({ "record": "persisted-before-terminal-receipts" }), |
| 6941 | }; |
| 6942 | manager |
| 6943 | .emit_event_for_test( |
| 6944 | &thread.id, |
| 6945 | Some(turn_id), |
| 6946 | "tool_call.requested", |
| 6947 | json!(¶ms), |
| 6948 | ) |
| 6949 | .await?; |
| 6950 | manager |
| 6951 | .emit_event_for_test( |
| 6952 | &thread.id, |
| 6953 | Some(turn_id), |
| 6954 | "turn.completed", |
| 6955 | json!({ "turn": &turn }), |
| 6956 | ) |
| 6957 | .await?; |
| 6958 | drop(manager); |
| 6959 | |
| 6960 | let recovered = test_manager(data_dir)?; |
| 6961 | recovered.get_thread(&thread.id).await?; |
| 6962 | let events = recovered.events_since(&thread.id, None)?; |
| 6963 | let terminal_calls = events |
| 6964 | .iter() |
| 6965 | .filter(|event| { |
| 6966 | event.turn_id.as_deref() == Some(turn_id) |
| 6967 | && event.payload.get("call_id").and_then(Value::as_str) == Some(call_id) |
| 6968 | && matches!( |
| 6969 | event.event.as_str(), |
| 6970 | "tool_call.resolved" | "tool_call.canceled" | "tool_call.timeout" |
| 6971 | ) |
| 6972 | }) |
| 6973 | .collect::<Vec<_>>(); |
| 6974 | assert_eq!(terminal_calls.len(), 1); |
| 6975 | assert_eq!(terminal_calls[0].event, "tool_call.canceled"); |
| 6976 | assert_eq!(terminal_calls[0].payload["reason"], "process_restart"); |
| 6977 | assert_eq!(terminal_calls[0].payload["recovered"], true); |
| 6978 | assert_eq!( |
| 6979 | events |
| 6980 | .iter() |
| 6981 | .filter(|event| { |
| 6982 | event.event == "turn.completed" && event.turn_id.as_deref() == Some(turn_id) |
| 6983 | }) |
| 6984 | .count(), |
| 6985 | 1, |
| 6986 | "recovery duplicated an already durable turn completion" |
| 6987 | ); |
| 6988 | Ok(()) |
| 6989 | } |
| 6990 | |
| 6991 | #[tokio::test] |
| 6992 | async fn concurrent_recovery_dedupe_scans_emit_each_terminal_receipt_once() -> Result<()> { |
| 6993 | let manager = test_manager(test_runtime_dir())?; |
| 6994 | let thread = manager |
| 6995 | .create_thread(CreateThreadRequest::default()) |
| 6996 | .await?; |
| 6997 | let turn_id = "turn_concurrent_recovery_dedupe"; |
| 6998 | let call_id = "call_concurrent_recovery_dedupe"; |
| 6999 | let mut turn = sample_turn(&thread.id, turn_id, RuntimeTurnStatus::Completed); |
| 7000 | turn.ended_at = Some(Utc::now()); |
| 7001 | let params = DynamicToolCallParams { |
| 7002 | thread_id: thread.id.clone(), |
| 7003 | turn_id: turn_id.to_string(), |
| 7004 | call_id: call_id.to_string(), |
| 7005 | namespace: Some("recovery".to_string()), |
| 7006 | tool: "dedupe_lookup".to_string(), |
| 7007 | arguments: json!({ "record": "same-terminal-receipt" }), |
| 7008 | }; |
| 7009 | |
| 7010 | let (first_call, second_call) = tokio::join!( |
| 7011 | manager.emit_recovered_dynamic_cancellation_if_missing(¶ms), |
| 7012 | manager.emit_recovered_dynamic_cancellation_if_missing(¶ms), |
| 7013 | ); |
| 7014 | assert_eq!( |
| 7015 | usize::from(first_call?) + usize::from(second_call?), |
| 7016 | 1, |
| 7017 | "the event_emit boundary must linearize dynamic-terminal dedupe" |
| 7018 | ); |
| 7019 | |
| 7020 | let (first_turn, second_turn) = tokio::join!( |
| 7021 | manager.emit_turn_completed_if_missing(&turn, true), |
| 7022 | manager.emit_turn_completed_if_missing(&turn, true), |
| 7023 | ); |
| 7024 | assert_eq!( |
| 7025 | usize::from(first_turn?) + usize::from(second_turn?), |
| 7026 | 1, |
| 7027 | "the event_emit boundary must linearize turn-completion dedupe" |
| 7028 | ); |
| 7029 | |
| 7030 | let events = manager.events_since(&thread.id, None)?; |
| 7031 | assert_eq!( |
| 7032 | events |
| 7033 | .iter() |
| 7034 | .filter(|event| { |
| 7035 | event.event == "tool_call.canceled" |
| 7036 | && event.turn_id.as_deref() == Some(turn_id) |
| 7037 | && event.payload.get("call_id").and_then(Value::as_str) == Some(call_id) |
| 7038 | }) |
| 7039 | .count(), |
| 7040 | 1 |
| 7041 | ); |
| 7042 | assert_eq!( |
| 7043 | events |
| 7044 | .iter() |
| 7045 | .filter(|event| { |
| 7046 | event.event == "turn.completed" && event.turn_id.as_deref() == Some(turn_id) |
| 7047 | }) |
| 7048 | .count(), |
| 7049 | 1 |
| 7050 | ); |
| 7051 | Ok(()) |
| 7052 | } |
| 7053 | |
| 7054 | #[tokio::test] |
| 7055 | async fn consecutive_dynamic_receipt_failures_queue_in_process_recovery() -> Result<()> { |
| 7056 | let manager = test_manager(test_runtime_dir())?; |
| 7057 | let thread = manager |
| 7058 | .create_thread(CreateThreadRequest::default()) |
| 7059 | .await?; |
| 7060 | let turn_id = "turn_consecutive_receipt_failures"; |
| 7061 | let call_id = "call_consecutive_receipt_failures"; |
| 7062 | let turn = sample_turn(&thread.id, turn_id, RuntimeTurnStatus::InProgress); |
| 7063 | manager.store.save_turn(&turn)?; |
| 7064 | { |
| 7065 | let _thread_mutation = manager.store.thread_mutation.lock(); |
| 7066 | let mut persisted_thread = manager.store.load_thread(&thread.id)?; |
| 7067 | persisted_thread.latest_turn_id = Some(turn_id.to_string()); |
| 7068 | manager.store.save_thread(&persisted_thread)?; |
| 7069 | } |
| 7070 | let params = DynamicToolCallParams { |
| 7071 | thread_id: thread.id.clone(), |
| 7072 | turn_id: turn_id.to_string(), |
| 7073 | call_id: call_id.to_string(), |
| 7074 | namespace: Some("recovery".to_string()), |
| 7075 | tool: "retry_lookup".to_string(), |
| 7076 | arguments: json!({ "record": "in-process" }), |
| 7077 | }; |
| 7078 | let requested = manager |
| 7079 | .emit_event_for_test( |
| 7080 | &thread.id, |
| 7081 | Some(turn_id), |
| 7082 | "tool_call.requested", |
| 7083 | json!(¶ms), |
| 7084 | ) |
| 7085 | .await?; |
| 7086 | let mut receiver = manager.register_pending_dynamic_tool(params)?; |
| 7087 | |
| 7088 | // The submitted result and the monitor's terminal cancellation both fail |
| 7089 | // after fsync, with each JSONL line transactionally removed. The monitor |
| 7090 | // must retain an in-process recovery path instead of evicting the engine |
| 7091 | // with an Awaiting call and no future owner. |
| 7092 | let fault_guard = |
| 7093 | EventAppendFaultGuard::arm_repeated(&thread.id, EventAppendTestFault::AfterSync, 2); |
| 7094 | let result_error = manager |
| 7095 | .deliver_dynamic_tool_result( |
| 7096 | &thread.id, |
| 7097 | turn_id, |
| 7098 | call_id, |
| 7099 | DynamicToolCallResult { |
| 7100 | success: true, |
| 7101 | content: vec![DynamicToolCallContent::InputText { |
| 7102 | text: "rolled-back-result".to_string(), |
| 7103 | }], |
| 7104 | }, |
| 7105 | ) |
| 7106 | .await |
| 7107 | .expect_err("first injected receipt failure unexpectedly succeeded"); |
| 7108 | assert!(result_error.to_string().contains("rolled back")); |
| 7109 | manager |
| 7110 | .settle_claimed_turn_failure(&thread.id, turn_id, "forced monitor failure") |
| 7111 | .await; |
| 7112 | drop(fault_guard); |
| 7113 | |
| 7114 | assert!( |
| 7115 | manager |
| 7116 | .recovery_receipts |
| 7117 | .lock() |
| 7118 | .get(&thread.id) |
| 7119 | .is_some_and(|receipts| receipts.iter().any(|receipt| receipt.turn.id == turn_id)), |
| 7120 | "second retry-safe failure did not queue in-process recovery" |
| 7121 | ); |
| 7122 | assert_eq!(manager.pending_dynamic_tools_count(), 1); |
| 7123 | assert!( |
| 7124 | tokio::time::timeout(Duration::from_millis(25), &mut receiver) |
| 7125 | .await |
| 7126 | .is_err(), |
| 7127 | "failed cancellation unexpectedly closed the model receiver" |
| 7128 | ); |
| 7129 | assert!(manager.events_since(&thread.id, None)?.iter().all(|event| { |
| 7130 | event.event != "turn.completed" |
| 7131 | && !matches!( |
| 7132 | event.event.as_str(), |
| 7133 | "tool_call.resolved" | "tool_call.canceled" | "tool_call.timeout" |
| 7134 | ) |
| 7135 | })); |
| 7136 | |
| 7137 | // The next async observation owns the queued retry. It durably cancels |
| 7138 | // the call before publishing exactly one recovered turn completion. |
| 7139 | manager.get_thread(&thread.id).await?; |
| 7140 | let closed = tokio::time::timeout(Duration::from_secs(2), &mut receiver) |
| 7141 | .await |
| 7142 | .context("recovery did not wake the model receiver")?; |
| 7143 | assert!(closed.is_err(), "terminal recovery delivered a tool result"); |
| 7144 | let events = manager.events_since(&thread.id, None)?; |
| 7145 | let canceled = events |
| 7146 | .iter() |
| 7147 | .filter(|event| { |
| 7148 | event.event == "tool_call.canceled" |
| 7149 | && event.payload.get("call_id").and_then(Value::as_str) == Some(call_id) |
| 7150 | }) |
| 7151 | .collect::<Vec<_>>(); |
| 7152 | let completed = events |
| 7153 | .iter() |
| 7154 | .filter(|event| { |
| 7155 | event.event == "turn.completed" && event.turn_id.as_deref() == Some(turn_id) |
| 7156 | }) |
| 7157 | .collect::<Vec<_>>(); |
| 7158 | assert_eq!(canceled.len(), 1); |
| 7159 | assert_eq!(canceled[0].payload["reason"], "turn_terminal"); |
| 7160 | assert_eq!(canceled[0].payload["terminal"], true); |
| 7161 | assert_eq!(completed.len(), 1); |
| 7162 | assert_eq!(completed[0].payload["recovered"], true); |
| 7163 | assert!(canceled[0].seq < completed[0].seq); |
| 7164 | assert!( |
| 7165 | canceled[0].seq > requested.seq.saturating_add(1), |
| 7166 | "rolled-back append sequence values were unexpectedly reused" |
| 7167 | ); |
| 7168 | assert_eq!(manager.pending_dynamic_tools_count(), 0); |
| 7169 | assert!(!manager.recovery_receipts.lock().contains_key(&thread.id)); |
| 7170 | |
| 7171 | manager.get_thread(&thread.id).await?; |
| 7172 | let replay = manager.events_since(&thread.id, None)?; |
| 7173 | assert_eq!( |
| 7174 | replay |
| 7175 | .iter() |
| 7176 | .filter(|event| { |
| 7177 | event.event == "tool_call.canceled" |
| 7178 | && event.payload.get("call_id").and_then(Value::as_str) == Some(call_id) |
| 7179 | }) |
| 7180 | .count(), |
| 7181 | 1 |
| 7182 | ); |
| 7183 | assert_eq!( |
| 7184 | replay |
| 7185 | .iter() |
| 7186 | .filter(|event| { |
| 7187 | event.event == "turn.completed" && event.turn_id.as_deref() == Some(turn_id) |
| 7188 | }) |
| 7189 | .count(), |
| 7190 | 1 |
| 7191 | ); |
| 7192 | Ok(()) |
| 7193 | } |
| 7194 | |
| 7195 | #[test] |
| 7196 | fn pending_dynamic_tool_registry_rejects_duplicates_and_is_bounded() -> Result<()> { |
| 7197 | let manager = test_manager(test_runtime_dir())?; |
| 7198 | let mut receivers = Vec::with_capacity(MAX_PENDING_DYNAMIC_TOOL_CALLS); |
| 7199 | receivers.push(manager.register_pending_dynamic_tool_for_test( |
| 7200 | "thread-bound", |
| 7201 | "turn-bound", |
| 7202 | "call-0", |
| 7203 | )?); |
| 7204 | assert!( |
| 7205 | manager |
| 7206 | .register_pending_dynamic_tool_for_test("thread-bound", "turn-bound", "call-0",) |
| 7207 | .is_err(), |
| 7208 | "duplicate call IDs must not replace an existing result channel" |
| 7209 | ); |
| 7210 | |
| 7211 | for index in 1..MAX_PENDING_DYNAMIC_TOOL_CALLS { |
| 7212 | receivers.push(manager.register_pending_dynamic_tool_for_test( |
| 7213 | "thread-bound", |
| 7214 | "turn-bound", |
| 7215 | &format!("call-{index}"), |
| 7216 | )?); |
| 7217 | } |
| 7218 | assert_eq!( |
| 7219 | manager.pending_dynamic_tools_count(), |
| 7220 | MAX_PENDING_DYNAMIC_TOOL_CALLS |
| 7221 | ); |
| 7222 | let error = manager |
| 7223 | .register_pending_dynamic_tool_for_test("thread-bound", "turn-bound", "call-over-limit") |
| 7224 | .expect_err("pending dynamic tool registry exceeded its hard limit"); |
| 7225 | assert!( |
| 7226 | error |
| 7227 | .to_string() |
| 7228 | .contains("pending dynamic tool call limit") |
| 7229 | ); |
| 7230 | Ok(()) |
| 7231 | } |
| 7232 | |
| 7233 | #[tokio::test] |
| 7234 | async fn dynamic_tool_timeout_clears_snapshot_and_emits_once() -> Result<()> { |
| 7235 | use crate::tools::spec::{DynamicToolExecutor, ToolError}; |
| 7236 | |
| 7237 | let _timeout_guard = test_dynamic_tool_timeout_ms(25); |
| 7238 | let manager = test_manager(test_runtime_dir())?; |
| 7239 | let thread = manager |
| 7240 | .create_thread(CreateThreadRequest::default()) |
| 7241 | .await?; |
| 7242 | let mut harness = install_mock_engine(&manager, &thread.id).await; |
| 7243 | let turn = manager |
| 7244 | .start_turn( |
| 7245 | &thread.id, |
| 7246 | StartTurnRequest { |
| 7247 | prompt: "let an external lookup time out".to_string(), |
| 7248 | ..StartTurnRequest::default() |
| 7249 | }, |
| 7250 | ) |
| 7251 | .await?; |
| 7252 | assert!(matches!( |
| 7253 | harness.rx_op.recv().await, |
| 7254 | Some(Op::SendMessage { .. }) |
| 7255 | )); |
| 7256 | harness |
| 7257 | .tx_event |
| 7258 | .send(EngineEvent::TurnStarted { |
| 7259 | turn_id: "dynamic_timeout".to_string(), |
| 7260 | created_at: Utc::now(), |
| 7261 | route: None, |
| 7262 | }) |
| 7263 | .await?; |
| 7264 | |
| 7265 | let error = DynamicToolExecutor::execute_dynamic_tool( |
| 7266 | &manager, |
| 7267 | Some(thread.id.clone()), |
| 7268 | None, |
| 7269 | "slow_lookup".to_string(), |
| 7270 | json!({ "marker": "request-only" }), |
| 7271 | ) |
| 7272 | .await |
| 7273 | .expect_err("dynamic tool unexpectedly resolved"); |
| 7274 | assert!(matches!(error, ToolError::Timeout { .. })); |
| 7275 | assert!( |
| 7276 | manager |
| 7277 | .get_thread_detail(&thread.id) |
| 7278 | .await? |
| 7279 | .pending_dynamic_tool_calls |
| 7280 | .is_empty() |
| 7281 | ); |
| 7282 | let timeout_events = manager |
| 7283 | .events_since(&thread.id, None)? |
| 7284 | .into_iter() |
| 7285 | .filter(|event| event.event == "tool_call.timeout") |
| 7286 | .collect::<Vec<_>>(); |
| 7287 | assert_eq!(timeout_events.len(), 1); |
| 7288 | assert_eq!(timeout_events[0].turn_id.as_deref(), Some(turn.id.as_str())); |
| 7289 | assert_eq!(timeout_events[0].payload["status"], "timeout"); |
| 7290 | assert_eq!(timeout_events[0].payload["timeout_secs"], 0); |
| 7291 | assert!(timeout_events[0].payload.get("arguments").is_none()); |
| 7292 | |
| 7293 | harness |
| 7294 | .tx_event |
| 7295 | .send(EngineEvent::TurnComplete { |
| 7296 | usage: Usage::default(), |
| 7297 | status: TurnOutcomeStatus::Completed, |
| 7298 | error: None, |
| 7299 | tool_catalog: None, |
| 7300 | base_url: None, |
| 7301 | }) |
| 7302 | .await?; |
| 7303 | Ok(()) |
| 7304 | } |
| 7305 | |
| 7306 | #[tokio::test] |
| 7307 | async fn terminal_turn_cancels_pending_dynamic_tool_exactly_once() -> Result<()> { |
| 7308 | use crate::tools::spec::DynamicToolExecutor; |
| 7309 | |
| 7310 | let manager = test_manager(test_runtime_dir())?; |
| 7311 | let thread = manager |
| 7312 | .create_thread(CreateThreadRequest::default()) |
| 7313 | .await?; |
| 7314 | let mut harness = install_mock_engine(&manager, &thread.id).await; |
| 7315 | let turn = manager |
| 7316 | .start_turn( |
| 7317 | &thread.id, |
| 7318 | StartTurnRequest { |
| 7319 | prompt: "cancel an external lookup with the turn".to_string(), |
| 7320 | ..StartTurnRequest::default() |
| 7321 | }, |
| 7322 | ) |
| 7323 | .await?; |
| 7324 | assert!(matches!( |
| 7325 | harness.rx_op.recv().await, |
| 7326 | Some(Op::SendMessage { .. }) |
| 7327 | )); |
| 7328 | harness |
| 7329 | .tx_event |
| 7330 | .send(EngineEvent::TurnStarted { |
| 7331 | turn_id: "dynamic_cancel".to_string(), |
| 7332 | created_at: Utc::now(), |
| 7333 | route: None, |
| 7334 | }) |
| 7335 | .await?; |
| 7336 | |
| 7337 | let executor = manager.clone(); |
| 7338 | let executor_thread_id = thread.id.clone(); |
| 7339 | let execution = tokio::spawn(async move { |
| 7340 | DynamicToolExecutor::execute_dynamic_tool( |
| 7341 | &executor, |
| 7342 | Some(executor_thread_id), |
| 7343 | None, |
| 7344 | "cancel_lookup".to_string(), |
| 7345 | json!({ "id": "pending" }), |
| 7346 | ) |
| 7347 | .await |
| 7348 | }); |
| 7349 | tokio::time::timeout(Duration::from_secs(2), async { |
| 7350 | loop { |
| 7351 | if !manager |
| 7352 | .get_thread_detail(&thread.id) |
| 7353 | .await? |
| 7354 | .pending_dynamic_tool_calls |
| 7355 | .is_empty() |
| 7356 | { |
| 7357 | break Ok::<_, anyhow::Error>(()); |
| 7358 | } |
| 7359 | sleep(Duration::from_millis(10)).await; |
| 7360 | } |
| 7361 | }) |
| 7362 | .await |
| 7363 | .context("dynamic call did not become pending")??; |
| 7364 | |
| 7365 | harness |
| 7366 | .tx_event |
| 7367 | .send(EngineEvent::TurnComplete { |
| 7368 | usage: Usage::default(), |
| 7369 | status: TurnOutcomeStatus::Interrupted, |
| 7370 | error: None, |
| 7371 | tool_catalog: None, |
| 7372 | base_url: None, |
| 7373 | }) |
| 7374 | .await?; |
| 7375 | execution |
| 7376 | .await |
| 7377 | .context("dynamic tool task panicked")? |
| 7378 | .expect_err("terminal turn unexpectedly resolved the dynamic tool"); |
| 7379 | |
| 7380 | tokio::time::timeout(Duration::from_secs(2), async { |
| 7381 | loop { |
| 7382 | let detail = manager.get_thread_detail(&thread.id).await?; |
| 7383 | let canceled = manager |
| 7384 | .events_since(&thread.id, None)? |
| 7385 | .into_iter() |
| 7386 | .filter(|event| event.event == "tool_call.canceled") |
| 7387 | .collect::<Vec<_>>(); |
| 7388 | if detail.pending_dynamic_tool_calls.is_empty() && canceled.len() == 1 { |
| 7389 | assert_eq!(canceled[0].payload["status"], "canceled"); |
| 7390 | assert_eq!(canceled[0].payload["terminal"], true); |
| 7391 | break Ok::<_, anyhow::Error>(()); |
| 7392 | } |
| 7393 | sleep(Duration::from_millis(10)).await; |
| 7394 | } |
| 7395 | }) |
| 7396 | .await |
| 7397 | .context("terminal dynamic call did not disappear exactly once")??; |
| 7398 | assert_eq!( |
| 7399 | turn.id, |
| 7400 | manager |
| 7401 | .get_thread(&thread.id) |
| 7402 | .await? |
| 7403 | .latest_turn_id |
| 7404 | .unwrap() |
| 7405 | ); |
| 7406 | Ok(()) |
| 7407 | } |
| 7408 | |
| 7409 | #[tokio::test] |
| 7410 | async fn approval_required_external_deny_is_denied() -> Result<()> { |
| 7411 | let manager = test_manager(test_runtime_dir())?; |
| 7412 | let thread = manager |
| 7413 | .create_thread(CreateThreadRequest { |
| 7414 | model: None, |
| 7415 | workspace: None, |
| 7416 | mode: None, |
| 7417 | allow_shell: None, |
| 7418 | trust_mode: None, |
| 7419 | auto_approve: None, |
| 7420 | archived: false, |
| 7421 | system_prompt: None, |
| 7422 | task_id: None, |
| 7423 | ..Default::default() |
| 7424 | }) |
| 7425 | .await?; |
| 7426 | |
| 7427 | let mut harness = install_mock_engine(&manager, &thread.id).await; |
| 7428 | let _turn = manager |
| 7429 | .start_turn( |
| 7430 | &thread.id, |
| 7431 | StartTurnRequest { |
| 7432 | prompt: "needs approval".to_string(), |
| 7433 | input_summary: None, |
| 7434 | model: None, |
| 7435 | mode: None, |
| 7436 | allow_shell: None, |
| 7437 | trust_mode: None, |
| 7438 | auto_approve: None, |
| 7439 | ..Default::default() |
| 7440 | }, |
| 7441 | ) |
| 7442 | .await?; |
| 7443 | assert!(matches!( |
| 7444 | harness.rx_op.recv().await, |
| 7445 | Some(Op::SendMessage { .. }) |
| 7446 | )); |
| 7447 | |
| 7448 | harness |
| 7449 | .tx_event |
| 7450 | .send(EngineEvent::ApprovalRequired { |
| 7451 | approval_key: "key2".to_string(), |
| 7452 | approval_grouping_key: "key2".to_string(), |
| 7453 | id: "tool_external_deny".to_string(), |
| 7454 | tool_name: "exec_command".to_string(), |
| 7455 | description: "external deny".to_string(), |
| 7456 | input: serde_json::json!({}), |
| 7457 | intent_summary: None, |
| 7458 | approval_force_prompt: false, |
| 7459 | }) |
| 7460 | .await?; |
| 7461 | |
| 7462 | let deadline = Instant::now() + Duration::from_secs(2); |
| 7463 | while Instant::now() < deadline && manager.pending_approvals_count() == 0 { |
| 7464 | sleep(Duration::from_millis(20)).await; |
| 7465 | } |
| 7466 | assert_eq!(manager.pending_approvals_count(), 1); |
| 7467 | |
| 7468 | assert!(manager.deliver_external_approval( |
| 7469 | "tool_external_deny", |
| 7470 | ExternalApprovalDecision::Deny { remember: false }, |
| 7471 | )); |
| 7472 | assert_eq!( |
| 7473 | harness.recv_approval_event().await, |
| 7474 | Some(MockApprovalEvent::Denied { |
| 7475 | id: "tool_external_deny".to_string(), |
| 7476 | }) |
| 7477 | ); |
| 7478 | |
| 7479 | harness |
| 7480 | .tx_event |
| 7481 | .send(EngineEvent::TurnComplete { |
| 7482 | usage: Usage::default(), |
| 7483 | status: TurnOutcomeStatus::Completed, |
| 7484 | error: None, |
| 7485 | tool_catalog: None, |
| 7486 | base_url: None, |
| 7487 | }) |
| 7488 | .await?; |
| 7489 | Ok(()) |
| 7490 | } |
| 7491 | |
| 7492 | #[tokio::test] |
| 7493 | async fn auto_review_force_prompt_is_denied_without_opening_a_modal() -> Result<()> { |
| 7494 | let manager = test_manager(test_runtime_dir())?; |
| 7495 | let thread = manager |
| 7496 | .create_thread(CreateThreadRequest { |
| 7497 | permission_posture: Some("auto-review".to_string()), |
| 7498 | ..Default::default() |
| 7499 | }) |
| 7500 | .await?; |
| 7501 | let mut harness = install_mock_engine(&manager, &thread.id).await; |
| 7502 | let turn = manager |
| 7503 | .start_turn( |
| 7504 | &thread.id, |
| 7505 | StartTurnRequest { |
| 7506 | prompt: "review the gated action".to_string(), |
| 7507 | ..Default::default() |
| 7508 | }, |
| 7509 | ) |
| 7510 | .await?; |
| 7511 | assert!(matches!( |
| 7512 | harness.rx_op.recv().await, |
| 7513 | Some(Op::SendMessage { |
| 7514 | approval_mode: crate::tui::approval::ApprovalMode::Auto, |
| 7515 | .. |
| 7516 | }) |
| 7517 | )); |
| 7518 | |
| 7519 | harness |
| 7520 | .tx_event |
| 7521 | .send(EngineEvent::ApprovalRequired { |
| 7522 | approval_key: "auto_hold".to_string(), |
| 7523 | approval_grouping_key: "auto_hold".to_string(), |
| 7524 | id: "tool_auto_hold".to_string(), |
| 7525 | tool_name: "exec_command".to_string(), |
| 7526 | description: "policy hold under auto review".to_string(), |
| 7527 | input: serde_json::json!({}), |
| 7528 | intent_summary: None, |
| 7529 | approval_force_prompt: true, |
| 7530 | }) |
| 7531 | .await?; |
| 7532 | |
| 7533 | let decision = tokio::time::timeout(Duration::from_secs(2), harness.recv_approval_event()) |
| 7534 | .await |
| 7535 | .context("Auto-Review hold should resolve without a modal")?; |
| 7536 | assert_eq!( |
| 7537 | decision, |
| 7538 | Some(MockApprovalEvent::Denied { |
| 7539 | id: "tool_auto_hold".to_string(), |
| 7540 | }) |
| 7541 | ); |
| 7542 | assert_eq!(manager.pending_approvals_count(), 0); |
| 7543 | assert!(manager.events_since(&thread.id, None)?.iter().any(|event| { |
| 7544 | event.event == "approval.decided" |
| 7545 | && event.payload.get("approval_id").and_then(Value::as_str) == Some("tool_auto_hold") |
| 7546 | && event.payload.get("posture").and_then(Value::as_str) == Some("auto_review") |
| 7547 | })); |
| 7548 | |
| 7549 | harness |
| 7550 | .tx_event |
| 7551 | .send(EngineEvent::TurnComplete { |
| 7552 | usage: Usage::default(), |
| 7553 | status: TurnOutcomeStatus::Completed, |
| 7554 | error: None, |
| 7555 | tool_catalog: None, |
| 7556 | base_url: None, |
| 7557 | }) |
| 7558 | .await?; |
| 7559 | let terminal = wait_for_terminal_turn(&manager, &turn.id, Duration::from_secs(2)).await?; |
| 7560 | assert_eq!(terminal.status, RuntimeTurnStatus::Completed); |
| 7561 | Ok(()) |
| 7562 | } |
| 7563 | |
| 7564 | #[tokio::test] |
| 7565 | async fn approval_timeout_denies_clears_ui_and_next_turn_can_start() -> Result<()> { |
| 7566 | let _timeout_guard = test_approval_timeout_ms(25); |
| 7567 | let manager = test_manager(test_runtime_dir())?; |
| 7568 | let thread = manager |
| 7569 | .create_thread(CreateThreadRequest { |
| 7570 | model: None, |
| 7571 | workspace: None, |
| 7572 | mode: None, |
| 7573 | allow_shell: None, |
| 7574 | trust_mode: None, |
| 7575 | auto_approve: None, |
| 7576 | archived: false, |
| 7577 | system_prompt: None, |
| 7578 | task_id: None, |
| 7579 | ..Default::default() |
| 7580 | }) |
| 7581 | .await?; |
| 7582 | |
| 7583 | let mut harness = install_mock_engine(&manager, &thread.id).await; |
| 7584 | let turn = manager |
| 7585 | .start_turn( |
| 7586 | &thread.id, |
| 7587 | StartTurnRequest { |
| 7588 | prompt: "needs approval".to_string(), |
| 7589 | input_summary: None, |
| 7590 | model: None, |
| 7591 | mode: None, |
| 7592 | allow_shell: None, |
| 7593 | trust_mode: None, |
| 7594 | auto_approve: None, |
| 7595 | ..Default::default() |
| 7596 | }, |
| 7597 | ) |
| 7598 | .await?; |
| 7599 | assert!(matches!( |
| 7600 | harness.rx_op.recv().await, |
| 7601 | Some(Op::SendMessage { .. }) |
| 7602 | )); |
| 7603 | |
| 7604 | harness |
| 7605 | .tx_event |
| 7606 | .send(EngineEvent::ApprovalRequired { |
| 7607 | approval_key: "timeout_key".to_string(), |
| 7608 | approval_grouping_key: "timeout_key".to_string(), |
| 7609 | id: "tool_timeout".to_string(), |
| 7610 | tool_name: "exec_command".to_string(), |
| 7611 | description: "external timeout".to_string(), |
| 7612 | input: serde_json::json!({}), |
| 7613 | intent_summary: None, |
| 7614 | approval_force_prompt: false, |
| 7615 | }) |
| 7616 | .await?; |
| 7617 | |
| 7618 | let decision = tokio::time::timeout(Duration::from_secs(2), harness.recv_approval_event()) |
| 7619 | .await |
| 7620 | .context("approval timeout should deny the engine")?; |
| 7621 | assert_eq!( |
| 7622 | decision, |
| 7623 | Some(MockApprovalEvent::Denied { |
| 7624 | id: "tool_timeout".to_string(), |
| 7625 | }) |
| 7626 | ); |
| 7627 | assert_eq!(manager.pending_approvals_count(), 0); |
| 7628 | |
| 7629 | let events = manager.events_since(&thread.id, None)?; |
| 7630 | assert!( |
| 7631 | events.iter().any(|event| { |
| 7632 | event.event == "approval.timeout" |
| 7633 | && event.payload.get("approval_id").and_then(Value::as_str) == Some("tool_timeout") |
| 7634 | }), |
| 7635 | "timeout event should be persisted" |
| 7636 | ); |
| 7637 | assert!( |
| 7638 | events.iter().any(|event| { |
| 7639 | event.event == "approval.decided" |
| 7640 | && event.payload.get("approval_id").and_then(Value::as_str) == Some("tool_timeout") |
| 7641 | && event.payload.get("decision").and_then(Value::as_str) == Some("deny") |
| 7642 | && event.payload.get("timeout").and_then(Value::as_bool) == Some(true) |
| 7643 | }), |
| 7644 | "timeout should also emit approval.decided so clients can clear pending UI" |
| 7645 | ); |
| 7646 | |
| 7647 | harness |
| 7648 | .tx_event |
| 7649 | .send(EngineEvent::TurnComplete { |
| 7650 | usage: Usage::default(), |
| 7651 | status: TurnOutcomeStatus::Completed, |
| 7652 | error: None, |
| 7653 | tool_catalog: None, |
| 7654 | base_url: None, |
| 7655 | }) |
| 7656 | .await?; |
| 7657 | let terminal = wait_for_terminal_turn(&manager, &turn.id, Duration::from_secs(2)).await?; |
| 7658 | assert_eq!(terminal.status, RuntimeTurnStatus::Completed); |
| 7659 | |
| 7660 | let _next = manager |
| 7661 | .start_turn( |
| 7662 | &thread.id, |
| 7663 | StartTurnRequest { |
| 7664 | prompt: "after timeout".to_string(), |
| 7665 | input_summary: None, |
| 7666 | model: None, |
| 7667 | mode: None, |
| 7668 | allow_shell: None, |
| 7669 | trust_mode: None, |
| 7670 | auto_approve: None, |
| 7671 | ..Default::default() |
| 7672 | }, |
| 7673 | ) |
| 7674 | .await?; |
| 7675 | assert!( |
| 7676 | matches!(harness.rx_op.recv().await, Some(Op::SendMessage { .. })), |
| 7677 | "thread should accept a fresh turn after approval timeout cleanup" |
| 7678 | ); |
| 7679 | |
| 7680 | Ok(()) |
| 7681 | } |
| 7682 | |
| 7683 | #[tokio::test] |
| 7684 | async fn thinking_delta_emits_agent_reasoning_item() -> Result<()> { |
| 7685 | let manager = test_manager(test_runtime_dir())?; |
| 7686 | let thread = manager |
| 7687 | .create_thread(CreateThreadRequest { |
| 7688 | model: None, |
| 7689 | workspace: None, |
| 7690 | mode: None, |
| 7691 | allow_shell: None, |
| 7692 | trust_mode: None, |
| 7693 | auto_approve: Some(true), |
| 7694 | archived: false, |
| 7695 | system_prompt: None, |
| 7696 | task_id: None, |
| 7697 | ..Default::default() |
| 7698 | }) |
| 7699 | .await?; |
| 7700 | let mut harness = install_mock_engine(&manager, &thread.id).await; |
| 7701 | let mut event_rx = manager.subscribe_events(); |
| 7702 | let _turn = manager |
| 7703 | .start_turn( |
| 7704 | &thread.id, |
| 7705 | StartTurnRequest { |
| 7706 | prompt: "show your thinking".to_string(), |
| 7707 | input_summary: None, |
| 7708 | model: None, |
| 7709 | mode: None, |
| 7710 | allow_shell: None, |
| 7711 | trust_mode: None, |
| 7712 | auto_approve: Some(true), |
| 7713 | ..Default::default() |
| 7714 | }, |
| 7715 | ) |
| 7716 | .await?; |
| 7717 | assert!(matches!( |
| 7718 | harness.rx_op.recv().await, |
| 7719 | Some(Op::SendMessage { .. }) |
| 7720 | )); |
| 7721 | |
| 7722 | harness |
| 7723 | .tx_event |
| 7724 | .send(EngineEvent::ThinkingStarted { index: 0 }) |
| 7725 | .await?; |
| 7726 | harness |
| 7727 | .tx_event |
| 7728 | .send(EngineEvent::ThinkingDelta { |
| 7729 | index: 0, |
| 7730 | content: "Let me reason about this.".to_string(), |
| 7731 | }) |
| 7732 | .await?; |
| 7733 | harness |
| 7734 | .tx_event |
| 7735 | .send(EngineEvent::ThinkingComplete { index: 0 }) |
| 7736 | .await?; |
| 7737 | harness |
| 7738 | .tx_event |
| 7739 | .send(EngineEvent::TurnComplete { |
| 7740 | usage: Usage::default(), |
| 7741 | status: TurnOutcomeStatus::Completed, |
| 7742 | error: None, |
| 7743 | tool_catalog: None, |
| 7744 | base_url: None, |
| 7745 | }) |
| 7746 | .await?; |
| 7747 | |
| 7748 | // A busy or constrained runner can be quiet for more than one 200 ms poll |
| 7749 | // even though the engine is still making progress. Keep polling until the |
| 7750 | // actual deadline instead of treating the first quiet interval as failure. |
| 7751 | let deadline = Instant::now() + Duration::from_secs(5); |
| 7752 | let mut delta_seen = false; |
| 7753 | let mut completed_seen = false; |
| 7754 | while Instant::now() < deadline && (!delta_seen || !completed_seen) { |
| 7755 | match tokio::time::timeout(Duration::from_millis(200), event_rx.recv()).await { |
| 7756 | Ok(Ok(record)) => { |
| 7757 | if record.event == "item.delta" |
| 7758 | && record.payload.get("kind").and_then(|v| v.as_str()) |
| 7759 | == Some("agent_reasoning") |
| 7760 | { |
| 7761 | delta_seen = true; |
| 7762 | assert_eq!( |
| 7763 | record.payload.get("delta").and_then(|v| v.as_str()), |
| 7764 | Some("Let me reason about this.") |
| 7765 | ); |
| 7766 | } |
| 7767 | if record.event == "item.completed" |
| 7768 | && record |
| 7769 | .payload |
| 7770 | .get("item") |
| 7771 | .and_then(|v| v.get("kind")) |
| 7772 | .and_then(|v| v.as_str()) |
| 7773 | == Some("agent_reasoning") |
| 7774 | { |
| 7775 | completed_seen = true; |
| 7776 | } |
| 7777 | } |
| 7778 | Ok(Err(_)) => break, |
| 7779 | Err(_) => continue, |
| 7780 | } |
| 7781 | } |
| 7782 | assert!(delta_seen, "expected item.delta with kind=agent_reasoning"); |
| 7783 | assert!( |
| 7784 | completed_seen, |
| 7785 | "expected item.completed for the reasoning item" |
| 7786 | ); |
| 7787 | Ok(()) |
| 7788 | } |
| 7789 | |
| 7790 | #[tokio::test] |
| 7791 | async fn deliver_external_approval_for_unknown_id_returns_false() { |
| 7792 | let manager = test_manager(test_runtime_dir()).expect("manager"); |
| 7793 | assert!(!manager.deliver_external_approval( |
| 7794 | "no_such_approval", |
| 7795 | ExternalApprovalDecision::Allow { remember: false }, |
| 7796 | )); |
| 7797 | assert_eq!(manager.pending_approvals_count(), 0); |
| 7798 | } |
| 7799 | |
| 7800 | #[tokio::test] |
| 7801 | async fn approval_required_remember_flips_thread_auto_approve() -> Result<()> { |
| 7802 | let manager = test_manager(test_runtime_dir())?; |
| 7803 | let thread = manager |
| 7804 | .create_thread(CreateThreadRequest { |
| 7805 | model: None, |
| 7806 | workspace: None, |
| 7807 | mode: None, |
| 7808 | allow_shell: None, |
| 7809 | trust_mode: None, |
| 7810 | auto_approve: None, |
| 7811 | archived: false, |
| 7812 | system_prompt: None, |
| 7813 | task_id: None, |
| 7814 | ..Default::default() |
| 7815 | }) |
| 7816 | .await?; |
| 7817 | assert!(!manager.store.load_thread(&thread.id)?.auto_approve); |
| 7818 | |
| 7819 | let mut harness = install_mock_engine(&manager, &thread.id).await; |
| 7820 | let turn = manager |
| 7821 | .start_turn( |
| 7822 | &thread.id, |
| 7823 | StartTurnRequest { |
| 7824 | prompt: "needs approval".to_string(), |
| 7825 | input_summary: None, |
| 7826 | model: None, |
| 7827 | mode: None, |
| 7828 | allow_shell: None, |
| 7829 | trust_mode: None, |
| 7830 | auto_approve: None, |
| 7831 | ..Default::default() |
| 7832 | }, |
| 7833 | ) |
| 7834 | .await?; |
| 7835 | assert!(matches!( |
| 7836 | harness.rx_op.recv().await, |
| 7837 | Some(Op::SendMessage { .. }) |
| 7838 | )); |
| 7839 | |
| 7840 | harness |
| 7841 | .tx_event |
| 7842 | .send(EngineEvent::ApprovalRequired { |
| 7843 | approval_key: "key3".to_string(), |
| 7844 | approval_grouping_key: "key3".to_string(), |
| 7845 | id: "tool_remember".to_string(), |
| 7846 | tool_name: "exec_command".to_string(), |
| 7847 | description: "remember=true".to_string(), |
| 7848 | input: serde_json::json!({}), |
| 7849 | intent_summary: None, |
| 7850 | approval_force_prompt: false, |
| 7851 | }) |
| 7852 | .await?; |
| 7853 | |
| 7854 | let deadline = Instant::now() + Duration::from_secs(2); |
| 7855 | while Instant::now() < deadline && manager.pending_approvals_count() == 0 { |
| 7856 | sleep(Duration::from_millis(20)).await; |
| 7857 | } |
| 7858 | assert!(manager.deliver_external_approval( |
| 7859 | "tool_remember", |
| 7860 | ExternalApprovalDecision::Allow { remember: true }, |
| 7861 | )); |
| 7862 | let _ = harness.recv_approval_event().await; |
| 7863 | |
| 7864 | assert!( |
| 7865 | manager.store.load_thread(&thread.id)?.auto_approve, |
| 7866 | "remember=true should flip thread auto_approve" |
| 7867 | ); |
| 7868 | assert_eq!( |
| 7869 | manager.active_turn_flags(&thread.id, &turn.id).await, |
| 7870 | Some((true, false)), |
| 7871 | "remember=true should update the active turn used by subsequent approvals" |
| 7872 | ); |
| 7873 | |
| 7874 | harness |
| 7875 | .tx_event |
| 7876 | .send(EngineEvent::TurnComplete { |
| 7877 | usage: Usage::default(), |
| 7878 | status: TurnOutcomeStatus::Completed, |
| 7879 | error: None, |
| 7880 | tool_catalog: None, |
| 7881 | base_url: None, |
| 7882 | }) |
| 7883 | .await?; |
| 7884 | Ok(()) |
| 7885 | } |
| 7886 | |
| 7887 | #[tokio::test] |
| 7888 | async fn elevation_required_with_stale_active_turn_is_denied() -> Result<()> { |
| 7889 | let manager = test_manager(test_runtime_dir())?; |
| 7890 | let thread = manager |
| 7891 | .create_thread(CreateThreadRequest { |
| 7892 | model: None, |
| 7893 | workspace: None, |
| 7894 | mode: None, |
| 7895 | allow_shell: None, |
| 7896 | trust_mode: Some(true), |
| 7897 | auto_approve: Some(true), |
| 7898 | archived: false, |
| 7899 | system_prompt: None, |
| 7900 | task_id: None, |
| 7901 | ..Default::default() |
| 7902 | }) |
| 7903 | .await?; |
| 7904 | |
| 7905 | let mut harness = install_mock_engine(&manager, &thread.id).await; |
| 7906 | let turn = manager |
| 7907 | .start_turn( |
| 7908 | &thread.id, |
| 7909 | StartTurnRequest { |
| 7910 | prompt: "needs elevation".to_string(), |
| 7911 | input_summary: None, |
| 7912 | model: None, |
| 7913 | mode: None, |
| 7914 | allow_shell: None, |
| 7915 | trust_mode: Some(true), |
| 7916 | auto_approve: Some(true), |
| 7917 | ..Default::default() |
| 7918 | }, |
| 7919 | ) |
| 7920 | .await?; |
| 7921 | |
| 7922 | assert!(matches!( |
| 7923 | harness.rx_op.recv().await, |
| 7924 | Some(Op::SendMessage { .. }) |
| 7925 | )); |
| 7926 | { |
| 7927 | let mut active = manager.active.lock().await; |
| 7928 | let state = active |
| 7929 | .engines |
| 7930 | .get_mut(&thread.id) |
| 7931 | .context("missing active thread state")?; |
| 7932 | state.active_turn = None; |
| 7933 | } |
| 7934 | |
| 7935 | harness |
| 7936 | .tx_event |
| 7937 | .send(EngineEvent::ElevationRequired { |
| 7938 | tool_id: "tool_stale_elevated".to_string(), |
| 7939 | tool_name: "exec_command".to_string(), |
| 7940 | command: None, |
| 7941 | denial_reason: "sandbox denied".to_string(), |
| 7942 | blocked_network: false, |
| 7943 | blocked_write: false, |
| 7944 | }) |
| 7945 | .await?; |
| 7946 | |
| 7947 | assert_eq!( |
| 7948 | harness.recv_approval_event().await, |
| 7949 | Some(MockApprovalEvent::Denied { |
| 7950 | id: "tool_stale_elevated".to_string(), |
| 7951 | }) |
| 7952 | ); |
| 7953 | |
| 7954 | harness |
| 7955 | .tx_event |
| 7956 | .send(EngineEvent::TurnComplete { |
| 7957 | usage: Usage { |
| 7958 | input_tokens: 0, |
| 7959 | output_tokens: 0, |
| 7960 | ..Usage::default() |
| 7961 | }, |
| 7962 | status: TurnOutcomeStatus::Completed, |
| 7963 | error: None, |
| 7964 | tool_catalog: None, |
| 7965 | base_url: None, |
| 7966 | }) |
| 7967 | .await?; |
| 7968 | |
| 7969 | let terminal = wait_for_terminal_turn(&manager, &turn.id, Duration::from_secs(2)).await?; |
| 7970 | assert_eq!(terminal.status, RuntimeTurnStatus::Completed); |
| 7971 | Ok(()) |
| 7972 | } |
| 7973 | |
| 7974 | #[tokio::test] |
| 7975 | async fn steer_turn_on_active_turn_records_item_and_event() -> Result<()> { |
| 7976 | let manager = test_manager(test_runtime_dir())?; |
| 7977 | let thread = manager |
| 7978 | .create_thread(CreateThreadRequest { |
| 7979 | model: None, |
| 7980 | workspace: None, |
| 7981 | mode: None, |
| 7982 | allow_shell: None, |
| 7983 | trust_mode: None, |
| 7984 | auto_approve: None, |
| 7985 | archived: false, |
| 7986 | system_prompt: None, |
| 7987 | task_id: None, |
| 7988 | ..Default::default() |
| 7989 | }) |
| 7990 | .await?; |
| 7991 | |
| 7992 | let harness = install_mock_engine(&manager, &thread.id).await; |
| 7993 | let mut rx_op = harness.rx_op; |
| 7994 | let mut rx_steer = harness.rx_steer; |
| 7995 | let tx_event = harness.tx_event; |
| 7996 | let (steer_seen_tx, steer_seen_rx) = oneshot::channel::<String>(); |
| 7997 | tokio::spawn(async move { |
| 7998 | if matches!(rx_op.recv().await, Some(Op::SendMessage { .. })) { |
| 7999 | let _ = tx_event |
| 8000 | .send(EngineEvent::TurnStarted { |
| 8001 | turn_id: "engine_turn_steer".to_string(), |
| 8002 | created_at: chrono::Utc::now(), |
| 8003 | route: None, |
| 8004 | }) |
| 8005 | .await; |
| 8006 | if let Some(steer) = rx_steer.recv().await { |
| 8007 | let _ = steer_seen_tx.send(steer); |
| 8008 | } |
| 8009 | let _ = tx_event |
| 8010 | .send(EngineEvent::MessageStarted { index: 0 }) |
| 8011 | .await; |
| 8012 | let _ = tx_event |
| 8013 | .send(EngineEvent::MessageDelta { |
| 8014 | index: 0, |
| 8015 | content: "steered response".to_string(), |
| 8016 | }) |
| 8017 | .await; |
| 8018 | let _ = tx_event |
| 8019 | .send(EngineEvent::MessageComplete { index: 0 }) |
| 8020 | .await; |
| 8021 | let _ = tx_event |
| 8022 | .send(EngineEvent::TurnComplete { |
| 8023 | usage: Usage { |
| 8024 | input_tokens: 8, |
| 8025 | output_tokens: 9, |
| 8026 | ..Usage::default() |
| 8027 | }, |
| 8028 | status: TurnOutcomeStatus::Completed, |
| 8029 | error: None, |
| 8030 | tool_catalog: None, |
| 8031 | base_url: None, |
| 8032 | }) |
| 8033 | .await; |
| 8034 | } |
| 8035 | }); |
| 8036 | |
| 8037 | let turn = manager |
| 8038 | .start_turn( |
| 8039 | &thread.id, |
| 8040 | StartTurnRequest { |
| 8041 | prompt: "initial".to_string(), |
| 8042 | input_summary: None, |
| 8043 | model: None, |
| 8044 | mode: None, |
| 8045 | allow_shell: None, |
| 8046 | trust_mode: None, |
| 8047 | auto_approve: None, |
| 8048 | ..Default::default() |
| 8049 | }, |
| 8050 | ) |
| 8051 | .await?; |
| 8052 | |
| 8053 | let steer_text = "add bullet list".to_string(); |
| 8054 | let steered_turn = manager |
| 8055 | .steer_turn( |
| 8056 | &thread.id, |
| 8057 | &turn.id, |
| 8058 | SteerTurnRequest { |
| 8059 | prompt: steer_text.clone(), |
| 8060 | }, |
| 8061 | ) |
| 8062 | .await?; |
| 8063 | assert_eq!(steered_turn.steer_count, 1); |
| 8064 | let observed_steer = steer_seen_rx |
| 8065 | .await |
| 8066 | .context("driver did not receive steer")?; |
| 8067 | assert_eq!(observed_steer, steer_text); |
| 8068 | |
| 8069 | let final_turn = wait_for_terminal_turn(&manager, &turn.id, Duration::from_secs(2)).await?; |
| 8070 | assert_eq!(final_turn.status, RuntimeTurnStatus::Completed); |
| 8071 | assert_eq!(final_turn.steer_count, 1); |
| 8072 | |
| 8073 | let events = manager.events_since(&thread.id, None)?; |
| 8074 | assert!(events.iter().any(|ev| ev.event == "turn.steered")); |
| 8075 | assert!(events.iter().any(|ev| { |
| 8076 | ev.event == "item.completed" |
| 8077 | && ev |
| 8078 | .payload |
| 8079 | .get("item") |
| 8080 | .and_then(|item| item.get("detail")) |
| 8081 | .and_then(Value::as_str) |
| 8082 | == Some("add bullet list") |
| 8083 | })); |
| 8084 | Ok(()) |
| 8085 | } |
| 8086 | |
| 8087 | #[tokio::test] |
| 8088 | async fn steer_receipts_outlive_caller_cancellation_after_engine_acceptance() -> Result<()> { |
| 8089 | let manager = test_manager(test_runtime_dir())?; |
| 8090 | let thread = manager |
| 8091 | .create_thread(CreateThreadRequest::default()) |
| 8092 | .await?; |
| 8093 | let harness = install_mock_engine(&manager, &thread.id).await; |
| 8094 | let mut rx_op = harness.rx_op; |
| 8095 | let mut rx_steer = harness.rx_steer; |
| 8096 | let tx_event = harness.tx_event; |
| 8097 | |
| 8098 | let turn = manager |
| 8099 | .start_turn( |
| 8100 | &thread.id, |
| 8101 | StartTurnRequest { |
| 8102 | prompt: "initial".to_string(), |
| 8103 | ..Default::default() |
| 8104 | }, |
| 8105 | ) |
| 8106 | .await?; |
| 8107 | assert!(matches!(rx_op.recv().await, Some(Op::SendMessage { .. }))); |
| 8108 | |
| 8109 | // Hold publication after durable persistence and mailbox acceptance so the |
| 8110 | // API future can be cancelled while the detached receipt task is pending. |
| 8111 | let emit_guard = manager.event_emit.lock().await; |
| 8112 | let steer_manager = manager.clone(); |
| 8113 | let thread_id = thread.id.clone(); |
| 8114 | let turn_id = turn.id.clone(); |
| 8115 | let steer_task = tokio::spawn(async move { |
| 8116 | steer_manager |
| 8117 | .steer_turn( |
| 8118 | &thread_id, |
| 8119 | &turn_id, |
| 8120 | SteerTurnRequest { |
| 8121 | prompt: "keep the accepted steer".to_string(), |
| 8122 | }, |
| 8123 | ) |
| 8124 | .await |
| 8125 | }); |
| 8126 | assert_eq!( |
| 8127 | tokio::time::timeout(Duration::from_secs(2), rx_steer.recv()).await?, |
| 8128 | Some("keep the accepted steer".to_string()) |
| 8129 | ); |
| 8130 | steer_task.abort(); |
| 8131 | assert!( |
| 8132 | steer_task |
| 8133 | .await |
| 8134 | .expect_err("caller task must be cancelled") |
| 8135 | .is_cancelled() |
| 8136 | ); |
| 8137 | drop(emit_guard); |
| 8138 | |
| 8139 | let deadline = Instant::now() + Duration::from_secs(2); |
| 8140 | loop { |
| 8141 | let events = manager.events_since(&thread.id, None)?; |
| 8142 | let steered = events.iter().any(|event| event.event == "turn.steered"); |
| 8143 | let completed = events.iter().any(|event| { |
| 8144 | event.event == "item.completed" |
| 8145 | && event |
| 8146 | .payload |
| 8147 | .get("item") |
| 8148 | .and_then(|item| item.get("detail")) |
| 8149 | .and_then(Value::as_str) |
| 8150 | == Some("keep the accepted steer") |
| 8151 | }); |
| 8152 | if steered && completed { |
| 8153 | break; |
| 8154 | } |
| 8155 | if Instant::now() >= deadline { |
| 8156 | bail!("detached steer receipts were not persisted after caller cancellation"); |
| 8157 | } |
| 8158 | tokio::task::yield_now().await; |
| 8159 | } |
| 8160 | |
| 8161 | let persisted_turn = manager.store.load_turn(&turn.id)?; |
| 8162 | assert_eq!(persisted_turn.steer_count, 1); |
| 8163 | let items = manager.store.list_items_for_turn(&turn.id)?; |
| 8164 | let steer_item = items |
| 8165 | .iter() |
| 8166 | .find(|item| item.detail.as_deref() == Some("keep the accepted steer")) |
| 8167 | .context("accepted steer item must remain durable")?; |
| 8168 | assert!(persisted_turn.item_ids.contains(&steer_item.id)); |
| 8169 | |
| 8170 | tx_event |
| 8171 | .send(EngineEvent::MessageStarted { index: 0 }) |
| 8172 | .await?; |
| 8173 | tx_event |
| 8174 | .send(EngineEvent::MessageDelta { |
| 8175 | index: 0, |
| 8176 | content: "accepted steer completed".to_string(), |
| 8177 | }) |
| 8178 | .await?; |
| 8179 | tx_event |
| 8180 | .send(EngineEvent::MessageComplete { index: 0 }) |
| 8181 | .await?; |
| 8182 | tx_event |
| 8183 | .send(EngineEvent::TurnComplete { |
| 8184 | usage: Usage::default(), |
| 8185 | status: TurnOutcomeStatus::Completed, |
| 8186 | error: None, |
| 8187 | tool_catalog: None, |
| 8188 | base_url: None, |
| 8189 | }) |
| 8190 | .await?; |
| 8191 | let terminal = wait_for_terminal_turn(&manager, &turn.id, Duration::from_secs(2)).await?; |
| 8192 | assert_eq!(terminal.status, RuntimeTurnStatus::Completed); |
| 8193 | Ok(()) |
| 8194 | } |
| 8195 | |
| 8196 | #[tokio::test] |
| 8197 | async fn steer_rejects_a_terminal_durable_turn_without_dispatch_or_item() -> Result<()> { |
| 8198 | let manager = test_manager(test_runtime_dir())?; |
| 8199 | let thread = manager |
| 8200 | .create_thread(CreateThreadRequest::default()) |
| 8201 | .await?; |
| 8202 | let harness = install_mock_engine(&manager, &thread.id).await; |
| 8203 | let mut rx_op = harness.rx_op; |
| 8204 | let mut rx_steer = harness.rx_steer; |
| 8205 | let tx_event = harness.tx_event; |
| 8206 | |
| 8207 | let turn = manager |
| 8208 | .start_turn( |
| 8209 | &thread.id, |
| 8210 | StartTurnRequest { |
| 8211 | prompt: "initial".to_string(), |
| 8212 | ..Default::default() |
| 8213 | }, |
| 8214 | ) |
| 8215 | .await?; |
| 8216 | assert!(matches!(rx_op.recv().await, Some(Op::SendMessage { .. }))); |
| 8217 | let original_item_ids = turn.item_ids.clone(); |
| 8218 | { |
| 8219 | let _turn_mutation = manager.store.turn_mutation.lock(); |
| 8220 | let mut terminal = manager.store.load_turn(&turn.id)?; |
| 8221 | terminal.status = RuntimeTurnStatus::Completed; |
| 8222 | terminal.ended_at = Some(Utc::now()); |
| 8223 | manager.store.save_turn(&terminal)?; |
| 8224 | } |
| 8225 | |
| 8226 | let error = manager |
| 8227 | .steer_turn( |
| 8228 | &thread.id, |
| 8229 | &turn.id, |
| 8230 | SteerTurnRequest { |
| 8231 | prompt: "must be rejected".to_string(), |
| 8232 | }, |
| 8233 | ) |
| 8234 | .await |
| 8235 | .expect_err("terminal turn must reject steering"); |
| 8236 | assert!(error.to_string().contains("no longer in progress")); |
| 8237 | assert!( |
| 8238 | tokio::time::timeout(Duration::from_millis(100), rx_steer.recv()) |
| 8239 | .await |
| 8240 | .is_err(), |
| 8241 | "rejected terminal steer must not reach the engine" |
| 8242 | ); |
| 8243 | let persisted = manager.store.load_turn(&turn.id)?; |
| 8244 | assert_eq!(persisted.steer_count, 0); |
| 8245 | assert_eq!(persisted.item_ids, original_item_ids); |
| 8246 | assert_eq!(manager.store.list_items_for_turn(&turn.id)?.len(), 1); |
| 8247 | |
| 8248 | // Restore the synthetic record and let the real monitor settle normally. |
| 8249 | { |
| 8250 | let _turn_mutation = manager.store.turn_mutation.lock(); |
| 8251 | let mut active = manager.store.load_turn(&turn.id)?; |
| 8252 | active.status = RuntimeTurnStatus::InProgress; |
| 8253 | active.ended_at = None; |
| 8254 | manager.store.save_turn(&active)?; |
| 8255 | } |
| 8256 | tx_event |
| 8257 | .send(EngineEvent::MessageStarted { index: 0 }) |
| 8258 | .await?; |
| 8259 | tx_event |
| 8260 | .send(EngineEvent::MessageDelta { |
| 8261 | index: 0, |
| 8262 | content: "terminal rejection test completed".to_string(), |
| 8263 | }) |
| 8264 | .await?; |
| 8265 | tx_event |
| 8266 | .send(EngineEvent::MessageComplete { index: 0 }) |
| 8267 | .await?; |
| 8268 | tx_event |
| 8269 | .send(EngineEvent::TurnComplete { |
| 8270 | usage: Usage::default(), |
| 8271 | status: TurnOutcomeStatus::Completed, |
| 8272 | error: None, |
| 8273 | tool_catalog: None, |
| 8274 | base_url: None, |
| 8275 | }) |
| 8276 | .await?; |
| 8277 | let terminal = wait_for_terminal_turn(&manager, &turn.id, Duration::from_secs(2)).await?; |
| 8278 | assert_eq!(terminal.status, RuntimeTurnStatus::Completed); |
| 8279 | Ok(()) |
| 8280 | } |
| 8281 | |
| 8282 | #[tokio::test] |
| 8283 | async fn concurrent_event_publication_keeps_live_and_durable_sequence_order() -> Result<()> { |
| 8284 | let manager = test_manager(test_runtime_dir())?; |
| 8285 | let thread = manager |
| 8286 | .create_thread(CreateThreadRequest::default()) |
| 8287 | .await?; |
| 8288 | let mut live_rx = manager.subscribe_events(); |
| 8289 | |
| 8290 | let mut emitters = Vec::new(); |
| 8291 | for index in 0..24_u64 { |
| 8292 | let emitter = manager.clone(); |
| 8293 | let thread_id = thread.id.clone(); |
| 8294 | emitters.push(tokio::spawn(async move { |
| 8295 | emitter |
| 8296 | .emit_event( |
| 8297 | &thread_id, |
| 8298 | None, |
| 8299 | None, |
| 8300 | "test.concurrent", |
| 8301 | json!({ "index": index }), |
| 8302 | ) |
| 8303 | .await |
| 8304 | })); |
| 8305 | } |
| 8306 | for emitter in emitters { |
| 8307 | emitter.await??; |
| 8308 | } |
| 8309 | |
| 8310 | let mut live = Vec::new(); |
| 8311 | for _ in 0..24 { |
| 8312 | live.push(tokio::time::timeout(Duration::from_secs(2), live_rx.recv()).await??); |
| 8313 | } |
| 8314 | assert!(live.windows(2).all(|pair| pair[0].seq < pair[1].seq)); |
| 8315 | |
| 8316 | let durable: Vec<_> = manager |
| 8317 | .events_since(&thread.id, None)? |
| 8318 | .into_iter() |
| 8319 | .filter(|event| event.event == "test.concurrent") |
| 8320 | .collect(); |
| 8321 | assert_eq!(durable.len(), 24); |
| 8322 | assert_eq!( |
| 8323 | live.iter() |
| 8324 | .map(|event| (event.seq, event.payload.clone())) |
| 8325 | .collect::<Vec<_>>(), |
| 8326 | durable |
| 8327 | .iter() |
| 8328 | .map(|event| (event.seq, event.payload.clone())) |
| 8329 | .collect::<Vec<_>>(), |
| 8330 | "broadcast order must exactly match append order" |
| 8331 | ); |
| 8332 | Ok(()) |
| 8333 | } |
| 8334 | |
| 8335 | #[tokio::test] |
| 8336 | async fn closed_engine_event_stream_fails_turn_items_and_evicts_engine() -> Result<()> { |
| 8337 | let manager = test_manager(test_runtime_dir())?; |
| 8338 | let thread = manager |
| 8339 | .create_thread(CreateThreadRequest::default()) |
| 8340 | .await?; |
| 8341 | let harness = install_mock_engine(&manager, &thread.id).await; |
| 8342 | let mut rx_op = harness.rx_op; |
| 8343 | let tx_event = harness.tx_event; |
| 8344 | |
| 8345 | let turn = manager |
| 8346 | .start_turn( |
| 8347 | &thread.id, |
| 8348 | StartTurnRequest { |
| 8349 | prompt: "engine stream will close".to_string(), |
| 8350 | ..Default::default() |
| 8351 | }, |
| 8352 | ) |
| 8353 | .await?; |
| 8354 | assert!(matches!(rx_op.recv().await, Some(Op::SendMessage { .. }))); |
| 8355 | drop(tx_event); |
| 8356 | |
| 8357 | let terminal = wait_for_terminal_turn(&manager, &turn.id, Duration::from_secs(2)).await?; |
| 8358 | assert_eq!(terminal.status, RuntimeTurnStatus::Failed); |
| 8359 | let terminal_error = terminal.error.as_deref().unwrap_or_default(); |
| 8360 | assert!( |
| 8361 | terminal.error.as_deref().is_some_and(|error| { |
| 8362 | error.contains("Failed to monitor") || error.contains("without producing any output") |
| 8363 | }), |
| 8364 | "unexpected terminal error: {terminal_error:?}" |
| 8365 | ); |
| 8366 | assert!( |
| 8367 | manager |
| 8368 | .store |
| 8369 | .list_items_for_turn(&turn.id)? |
| 8370 | .iter() |
| 8371 | .all(|item| !matches!( |
| 8372 | item.status, |
| 8373 | TurnItemLifecycleStatus::Queued | TurnItemLifecycleStatus::InProgress |
| 8374 | )) |
| 8375 | ); |
| 8376 | let deadline = Instant::now() + Duration::from_secs(2); |
| 8377 | loop { |
| 8378 | if !manager.active.lock().await.engines.contains_key(&thread.id) { |
| 8379 | break; |
| 8380 | } |
| 8381 | if Instant::now() >= deadline { |
| 8382 | bail!("failed engine was not evicted"); |
| 8383 | } |
| 8384 | tokio::task::yield_now().await; |
| 8385 | } |
| 8386 | assert!(matches!(rx_op.recv().await, Some(Op::Shutdown))); |
| 8387 | Ok(()) |
| 8388 | } |
| 8389 | |
| 8390 | #[tokio::test] |
| 8391 | async fn failed_turn_cancels_pending_user_input_and_clears_snapshot() -> Result<()> { |
| 8392 | let manager = test_manager(test_runtime_dir())?; |
| 8393 | let thread = manager |
| 8394 | .create_thread(CreateThreadRequest::default()) |
| 8395 | .await?; |
| 8396 | let mut harness = install_mock_engine(&manager, &thread.id).await; |
| 8397 | let turn = manager |
| 8398 | .start_turn( |
| 8399 | &thread.id, |
| 8400 | StartTurnRequest { |
| 8401 | prompt: "input required, then the engine stream closes".to_string(), |
| 8402 | ..Default::default() |
| 8403 | }, |
| 8404 | ) |
| 8405 | .await?; |
| 8406 | assert!(matches!( |
| 8407 | harness.rx_op.recv().await, |
| 8408 | Some(Op::SendMessage { .. }) |
| 8409 | )); |
| 8410 | harness |
| 8411 | .tx_event |
| 8412 | .send(EngineEvent::UserInputRequired { |
| 8413 | id: "input_failed_turn".to_string(), |
| 8414 | request: crate::tools::user_input::UserInputRequest { |
| 8415 | questions: vec![crate::tools::user_input::UserInputQuestion { |
| 8416 | header: "Continue".to_string(), |
| 8417 | id: "continue".to_string(), |
| 8418 | question: "Continue?".to_string(), |
| 8419 | options: vec![crate::tools::user_input::UserInputOption { |
| 8420 | label: "Yes".to_string(), |
| 8421 | description: "Continue now".to_string(), |
| 8422 | }], |
| 8423 | allow_free_text: false, |
| 8424 | multi_select: false, |
| 8425 | }], |
| 8426 | }, |
| 8427 | }) |
| 8428 | .await?; |
| 8429 | |
| 8430 | let deadline = Instant::now() + Duration::from_secs(2); |
| 8431 | loop { |
| 8432 | if !manager |
| 8433 | .get_thread_detail(&thread.id) |
| 8434 | .await? |
| 8435 | .pending_user_inputs |
| 8436 | .is_empty() |
| 8437 | { |
| 8438 | break; |
| 8439 | } |
| 8440 | if Instant::now() >= deadline { |
| 8441 | bail!("pending user input did not reach the canonical snapshot"); |
| 8442 | } |
| 8443 | sleep(Duration::from_millis(20)).await; |
| 8444 | } |
| 8445 | |
| 8446 | // Fail the turn mid-prompt: closing the event stream settles the turn as |
| 8447 | // failed through the monitor-failure path. |
| 8448 | harness.close_event_stream(); |
| 8449 | |
| 8450 | let canceled = tokio::time::timeout( |
| 8451 | Duration::from_secs(2), |
| 8452 | harness.recv_user_input_cancellation(), |
| 8453 | ) |
| 8454 | .await |
| 8455 | .expect("failure-path user-input cancellation timed out"); |
| 8456 | assert_eq!(canceled.as_deref(), Some("input_failed_turn")); |
| 8457 | |
| 8458 | let terminal = wait_for_terminal_turn(&manager, &turn.id, Duration::from_secs(2)).await?; |
| 8459 | assert_eq!(terminal.status, RuntimeTurnStatus::Failed); |
| 8460 | |
| 8461 | let deadline = Instant::now() + Duration::from_secs(2); |
| 8462 | loop { |
| 8463 | let detail = manager.get_thread_detail(&thread.id).await?; |
| 8464 | if detail.pending_user_inputs.is_empty() |
| 8465 | && manager.events_since(&thread.id, None)?.iter().any(|event| { |
| 8466 | event.event == "user_input.canceled" |
| 8467 | && event.turn_id.as_deref() == Some(turn.id.as_str()) |
| 8468 | && event.payload.get("input_id").and_then(Value::as_str) |
| 8469 | == Some("input_failed_turn") |
| 8470 | && event.payload.get("terminal").and_then(Value::as_bool) == Some(true) |
| 8471 | }) |
| 8472 | { |
| 8473 | break; |
| 8474 | } |
| 8475 | if Instant::now() >= deadline { |
| 8476 | bail!("failed turn left a stale pending user input in the snapshot"); |
| 8477 | } |
| 8478 | sleep(Duration::from_millis(20)).await; |
| 8479 | } |
| 8480 | Ok(()) |
| 8481 | } |
| 8482 | |
| 8483 | #[tokio::test] |
| 8484 | async fn compaction_lifecycle_emits_item_events_with_compaction_counts() -> Result<()> { |
| 8485 | let manager = test_manager(test_runtime_dir())?; |
| 8486 | let thread = manager |
| 8487 | .create_thread(CreateThreadRequest { |
| 8488 | model: None, |
| 8489 | workspace: None, |
| 8490 | mode: None, |
| 8491 | allow_shell: None, |
| 8492 | trust_mode: None, |
| 8493 | auto_approve: None, |
| 8494 | archived: false, |
| 8495 | system_prompt: None, |
| 8496 | task_id: None, |
| 8497 | ..Default::default() |
| 8498 | }) |
| 8499 | .await?; |
| 8500 | |
| 8501 | let harness = install_mock_engine(&manager, &thread.id).await; |
| 8502 | let mut rx_op = harness.rx_op; |
| 8503 | let tx_event = harness.tx_event; |
| 8504 | tokio::spawn(async move { |
| 8505 | let mut op_count = 0usize; |
| 8506 | while let Some(op) = rx_op.recv().await { |
| 8507 | match op { |
| 8508 | Op::SendMessage { .. } => { |
| 8509 | op_count = op_count.saturating_add(1); |
| 8510 | let _ = tx_event |
| 8511 | .send(EngineEvent::TurnStarted { |
| 8512 | turn_id: "engine_turn_auto".to_string(), |
| 8513 | created_at: chrono::Utc::now(), |
| 8514 | route: None, |
| 8515 | }) |
| 8516 | .await; |
| 8517 | let _ = tx_event |
| 8518 | .send(EngineEvent::CompactionStarted { |
| 8519 | id: "auto_compact_1".to_string(), |
| 8520 | auto: true, |
| 8521 | message: "auto compact begin".to_string(), |
| 8522 | }) |
| 8523 | .await; |
| 8524 | let _ = tx_event |
| 8525 | .send(EngineEvent::CompactionCompleted { |
| 8526 | id: "auto_compact_1".to_string(), |
| 8527 | auto: true, |
| 8528 | message: "auto compact done".to_string(), |
| 8529 | messages_before: Some(7), |
| 8530 | messages_after: Some(3), |
| 8531 | summary_prompt: None, |
| 8532 | }) |
| 8533 | .await; |
| 8534 | let _ = tx_event |
| 8535 | .send(EngineEvent::TurnComplete { |
| 8536 | usage: Usage { |
| 8537 | input_tokens: 3, |
| 8538 | output_tokens: 3, |
| 8539 | ..Usage::default() |
| 8540 | }, |
| 8541 | status: TurnOutcomeStatus::Completed, |
| 8542 | error: None, |
| 8543 | tool_catalog: None, |
| 8544 | base_url: None, |
| 8545 | }) |
| 8546 | .await; |
| 8547 | } |
| 8548 | Op::CompactContext { .. } => { |
| 8549 | op_count = op_count.saturating_add(1); |
| 8550 | let _ = tx_event |
| 8551 | .send(EngineEvent::CompactionStarted { |
| 8552 | id: "manual_compact_1".to_string(), |
| 8553 | auto: false, |
| 8554 | message: "manual compact begin".to_string(), |
| 8555 | }) |
| 8556 | .await; |
| 8557 | let _ = tx_event |
| 8558 | .send(EngineEvent::CompactionCompleted { |
| 8559 | id: "manual_compact_1".to_string(), |
| 8560 | auto: false, |
| 8561 | message: "manual compact done".to_string(), |
| 8562 | messages_before: Some(5), |
| 8563 | messages_after: Some(2), |
| 8564 | summary_prompt: Some( |
| 8565 | "## 📋 Conversation Summary (Auto-Generated)\n\nkey facts." |
| 8566 | .to_string(), |
| 8567 | ), |
| 8568 | }) |
| 8569 | .await; |
| 8570 | let _ = tx_event |
| 8571 | .send(EngineEvent::TurnComplete { |
| 8572 | usage: Usage { |
| 8573 | input_tokens: 1, |
| 8574 | output_tokens: 1, |
| 8575 | ..Usage::default() |
| 8576 | }, |
| 8577 | status: TurnOutcomeStatus::Completed, |
| 8578 | error: None, |
| 8579 | tool_catalog: None, |
| 8580 | base_url: None, |
| 8581 | }) |
| 8582 | .await; |
| 8583 | } |
| 8584 | _ => {} |
| 8585 | } |
| 8586 | if op_count >= 2 { |
| 8587 | break; |
| 8588 | } |
| 8589 | } |
| 8590 | }); |
| 8591 | |
| 8592 | let auto_turn = manager |
| 8593 | .start_turn( |
| 8594 | &thread.id, |
| 8595 | StartTurnRequest { |
| 8596 | prompt: "trigger auto".to_string(), |
| 8597 | input_summary: None, |
| 8598 | model: None, |
| 8599 | mode: None, |
| 8600 | allow_shell: None, |
| 8601 | trust_mode: None, |
| 8602 | auto_approve: None, |
| 8603 | ..Default::default() |
| 8604 | }, |
| 8605 | ) |
| 8606 | .await?; |
| 8607 | let auto_turn = wait_for_terminal_turn(&manager, &auto_turn.id, Duration::from_secs(2)).await?; |
| 8608 | assert_eq!(auto_turn.status, RuntimeTurnStatus::Completed); |
| 8609 | |
| 8610 | let manual_turn = manager |
| 8611 | .compact_thread( |
| 8612 | &thread.id, |
| 8613 | CompactThreadRequest { |
| 8614 | reason: Some("manual request".to_string()), |
| 8615 | }, |
| 8616 | ) |
| 8617 | .await?; |
| 8618 | let manual_turn = |
| 8619 | wait_for_terminal_turn(&manager, &manual_turn.id, Duration::from_secs(2)).await?; |
| 8620 | assert_eq!(manual_turn.status, RuntimeTurnStatus::Completed); |
| 8621 | |
| 8622 | let events = manager.events_since(&thread.id, None)?; |
| 8623 | assert!(events.iter().any(|ev| { |
| 8624 | ev.event == "item.started" |
| 8625 | && ev |
| 8626 | .payload |
| 8627 | .get("item") |
| 8628 | .and_then(|item| item.get("kind")) |
| 8629 | .and_then(Value::as_str) |
| 8630 | == Some("context_compaction") |
| 8631 | && ev.payload.get("auto").and_then(Value::as_bool) == Some(true) |
| 8632 | })); |
| 8633 | assert!(events.iter().any(|ev| { |
| 8634 | ev.event == "item.completed" |
| 8635 | && ev |
| 8636 | .payload |
| 8637 | .get("item") |
| 8638 | .and_then(|item| item.get("kind")) |
| 8639 | .and_then(Value::as_str) |
| 8640 | == Some("context_compaction") |
| 8641 | && ev.payload.get("auto").and_then(Value::as_bool) == Some(true) |
| 8642 | && ev.payload.get("messages_before").and_then(Value::as_u64) == Some(7) |
| 8643 | && ev.payload.get("messages_after").and_then(Value::as_u64) == Some(3) |
| 8644 | })); |
| 8645 | assert!(events.iter().any(|ev| { |
| 8646 | ev.event == "item.completed" |
| 8647 | && ev |
| 8648 | .payload |
| 8649 | .get("item") |
| 8650 | .and_then(|item| item.get("kind")) |
| 8651 | .and_then(Value::as_str) |
| 8652 | == Some("context_compaction") |
| 8653 | && ev.payload.get("auto").and_then(Value::as_bool) == Some(false) |
| 8654 | && ev.payload.get("messages_before").and_then(Value::as_u64) == Some(5) |
| 8655 | && ev.payload.get("messages_after").and_then(Value::as_u64) == Some(2) |
| 8656 | })); |
| 8657 | |
| 8658 | // The manual compact carried a summary_prompt → it must be persisted into |
| 8659 | // the thread record so engine reloads restore it. The auto compact carried |
| 8660 | // None → exactly one summary section, from the manual pass. |
| 8661 | let record = manager.get_thread(&thread.id).await?; |
| 8662 | let record_prompt = record.system_prompt.expect("record keeps a system prompt"); |
| 8663 | assert!(record_prompt.contains(COMPACTION_SUMMARY_BEGIN)); |
| 8664 | assert!(record_prompt.contains("Conversation Summary (Auto-Generated)")); |
| 8665 | assert!(record_prompt.contains("key facts.")); |
| 8666 | assert_eq!(record_prompt.matches(COMPACTION_SUMMARY_BEGIN).count(), 1); |
| 8667 | Ok(()) |
| 8668 | } |
| 8669 | |
| 8670 | #[test] |
| 8671 | fn summarize_text_truncates() { |
| 8672 | let out = summarize_text("abcdefghijklmnopqrstuvwxyz", 10); |
| 8673 | assert_eq!(out, "abcdefg..."); |
| 8674 | } |
| 8675 | |
| 8676 | #[test] |
| 8677 | fn approval_decision_requires_auto_approve_and_trust_for_full_access() { |
| 8678 | assert_eq!( |
| 8679 | RuntimeThreadManager::approval_decision(false, false, false), |
| 8680 | RuntimeApprovalDecision::DenyTool |
| 8681 | ); |
| 8682 | assert_eq!( |
| 8683 | RuntimeThreadManager::approval_decision(false, true, false), |
| 8684 | RuntimeApprovalDecision::DenyTool |
| 8685 | ); |
| 8686 | assert_eq!( |
| 8687 | RuntimeThreadManager::approval_decision(true, false, false), |
| 8688 | RuntimeApprovalDecision::ApproveTool |
| 8689 | ); |
| 8690 | assert_eq!( |
| 8691 | RuntimeThreadManager::approval_decision(true, false, true), |
| 8692 | RuntimeApprovalDecision::DenyTool |
| 8693 | ); |
| 8694 | assert_eq!( |
| 8695 | RuntimeThreadManager::approval_decision(true, true, true), |
| 8696 | RuntimeApprovalDecision::RetryWithFullAccess |
| 8697 | ); |
| 8698 | } |
| 8699 | |
| 8700 | #[test] |
| 8701 | fn opening_manager_recovers_stale_queued_and_in_progress_work() -> Result<()> { |
| 8702 | let data_dir = test_runtime_dir(); |
| 8703 | let manager = test_manager(data_dir.clone())?; |
| 8704 | let started_at = Utc::now() - chrono::Duration::seconds(5); |
| 8705 | let created_at = started_at - chrono::Duration::seconds(1); |
| 8706 | |
| 8707 | let thread = ThreadRecord { |
| 8708 | schema_version: CURRENT_RUNTIME_SCHEMA_VERSION, |
| 8709 | id: "thr_restart".to_string(), |
| 8710 | created_at, |
| 8711 | updated_at: created_at, |
| 8712 | model: DEFAULT_TEXT_MODEL.to_string(), |
| 8713 | model_provider: None, |
| 8714 | model_provider_id: None, |
| 8715 | workspace: PathBuf::from("."), |
| 8716 | mode: "agent".to_string(), |
| 8717 | permission_posture: None, |
| 8718 | allow_shell: false, |
| 8719 | trust_mode: false, |
| 8720 | auto_approve: false, |
| 8721 | latest_turn_id: Some("turn_in_progress".to_string()), |
| 8722 | latest_response_bookmark: None, |
| 8723 | archived: false, |
| 8724 | system_prompt: None, |
| 8725 | task_id: None, |
| 8726 | title: None, |
| 8727 | session_id: None, |
| 8728 | }; |
| 8729 | manager.store.save_thread(&thread)?; |
| 8730 | |
| 8731 | let completed_item = TurnItemRecord { |
| 8732 | schema_version: CURRENT_RUNTIME_SCHEMA_VERSION, |
| 8733 | id: "item_completed".to_string(), |
| 8734 | turn_id: "turn_in_progress".to_string(), |
| 8735 | kind: TurnItemKind::Status, |
| 8736 | status: TurnItemLifecycleStatus::Completed, |
| 8737 | summary: "done".to_string(), |
| 8738 | detail: None, |
| 8739 | metadata: None, |
| 8740 | artifact_refs: Vec::new(), |
| 8741 | started_at: Some(started_at), |
| 8742 | ended_at: Some(started_at + chrono::Duration::seconds(1)), |
| 8743 | }; |
| 8744 | let in_progress_item = TurnItemRecord { |
| 8745 | schema_version: CURRENT_RUNTIME_SCHEMA_VERSION, |
| 8746 | id: "item_in_progress".to_string(), |
| 8747 | turn_id: "turn_in_progress".to_string(), |
| 8748 | kind: TurnItemKind::ToolCall, |
| 8749 | status: TurnItemLifecycleStatus::InProgress, |
| 8750 | summary: "running".to_string(), |
| 8751 | detail: None, |
| 8752 | metadata: None, |
| 8753 | artifact_refs: Vec::new(), |
| 8754 | started_at: Some(started_at), |
| 8755 | ended_at: None, |
| 8756 | }; |
| 8757 | let queued_item = TurnItemRecord { |
| 8758 | schema_version: CURRENT_RUNTIME_SCHEMA_VERSION, |
| 8759 | id: "item_queued".to_string(), |
| 8760 | turn_id: "turn_queued".to_string(), |
| 8761 | kind: TurnItemKind::ToolCall, |
| 8762 | status: TurnItemLifecycleStatus::Queued, |
| 8763 | summary: "queued".to_string(), |
| 8764 | detail: None, |
| 8765 | metadata: None, |
| 8766 | artifact_refs: Vec::new(), |
| 8767 | started_at: None, |
| 8768 | ended_at: None, |
| 8769 | }; |
| 8770 | manager.store.save_item(&completed_item)?; |
| 8771 | manager.store.save_item(&in_progress_item)?; |
| 8772 | manager.store.save_item(&queued_item)?; |
| 8773 | |
| 8774 | manager.store.save_turn(&TurnRecord { |
| 8775 | schema_version: CURRENT_RUNTIME_SCHEMA_VERSION, |
| 8776 | id: "turn_in_progress".to_string(), |
| 8777 | thread_id: thread.id.clone(), |
| 8778 | status: RuntimeTurnStatus::InProgress, |
| 8779 | input_summary: "hello".to_string(), |
| 8780 | created_at, |
| 8781 | started_at: Some(started_at), |
| 8782 | ended_at: None, |
| 8783 | duration_ms: None, |
| 8784 | usage: None, |
| 8785 | permission_posture: None, |
| 8786 | effective_provider: None, |
| 8787 | effective_provider_id: None, |
| 8788 | effective_billing_surface: None, |
| 8789 | effective_endpoint_fingerprint: None, |
| 8790 | effective_billing_mode: None, |
| 8791 | effective_dispatched_at: None, |
| 8792 | effective_model: None, |
| 8793 | routed_usage: Vec::new(), |
| 8794 | routed_usage_source_ids: Vec::new(), |
| 8795 | routed_usage_dropped_records: 0, |
| 8796 | error: None, |
| 8797 | item_ids: vec![completed_item.id.clone(), in_progress_item.id.clone()], |
| 8798 | steer_count: 0, |
| 8799 | })?; |
| 8800 | manager.store.save_turn(&TurnRecord { |
| 8801 | schema_version: CURRENT_RUNTIME_SCHEMA_VERSION, |
| 8802 | id: "turn_queued".to_string(), |
| 8803 | thread_id: thread.id.clone(), |
| 8804 | status: RuntimeTurnStatus::Queued, |
| 8805 | input_summary: "later".to_string(), |
| 8806 | created_at, |
| 8807 | started_at: None, |
| 8808 | ended_at: None, |
| 8809 | duration_ms: None, |
| 8810 | usage: None, |
| 8811 | permission_posture: None, |
| 8812 | effective_provider: None, |
| 8813 | effective_provider_id: None, |
| 8814 | effective_billing_surface: None, |
| 8815 | effective_endpoint_fingerprint: None, |
| 8816 | effective_billing_mode: None, |
| 8817 | effective_dispatched_at: None, |
| 8818 | effective_model: None, |
| 8819 | routed_usage: Vec::new(), |
| 8820 | routed_usage_source_ids: Vec::new(), |
| 8821 | routed_usage_dropped_records: 0, |
| 8822 | error: None, |
| 8823 | item_ids: vec![queued_item.id.clone()], |
| 8824 | steer_count: 0, |
| 8825 | })?; |
| 8826 | drop(manager); |
| 8827 | |
| 8828 | let recovered = test_manager(data_dir)?; |
| 8829 | |
| 8830 | let recovered_thread = recovered.store.load_thread(&thread.id)?; |
| 8831 | assert!(recovered_thread.updated_at >= thread.updated_at); |
| 8832 | |
| 8833 | let recovered_in_progress_turn = recovered.store.load_turn("turn_in_progress")?; |
| 8834 | assert_eq!( |
| 8835 | recovered_in_progress_turn.status, |
| 8836 | RuntimeTurnStatus::Interrupted |
| 8837 | ); |
| 8838 | assert_eq!( |
| 8839 | recovered_in_progress_turn.error.as_deref(), |
| 8840 | Some(RUNTIME_RESTART_REASON) |
| 8841 | ); |
| 8842 | assert!(recovered_in_progress_turn.ended_at.is_some()); |
| 8843 | assert!( |
| 8844 | recovered_in_progress_turn |
| 8845 | .duration_ms |
| 8846 | .is_some_and(|duration| duration >= 5_000) |
| 8847 | ); |
| 8848 | |
| 8849 | let recovered_queued_turn = recovered.store.load_turn("turn_queued")?; |
| 8850 | assert_eq!(recovered_queued_turn.status, RuntimeTurnStatus::Interrupted); |
| 8851 | assert_eq!( |
| 8852 | recovered_queued_turn.error.as_deref(), |
| 8853 | Some(RUNTIME_RESTART_REASON) |
| 8854 | ); |
| 8855 | assert!(recovered_queued_turn.ended_at.is_some()); |
| 8856 | assert_eq!(recovered_queued_turn.duration_ms, None); |
| 8857 | |
| 8858 | assert_eq!( |
| 8859 | recovered.store.load_item(&completed_item.id)?.status, |
| 8860 | TurnItemLifecycleStatus::Completed |
| 8861 | ); |
| 8862 | let recovered_in_progress_item = recovered.store.load_item(&in_progress_item.id)?; |
| 8863 | assert_eq!( |
| 8864 | recovered_in_progress_item.status, |
| 8865 | TurnItemLifecycleStatus::Interrupted |
| 8866 | ); |
| 8867 | assert!(recovered_in_progress_item.ended_at.is_some()); |
| 8868 | |
| 8869 | let recovered_queued_item = recovered.store.load_item(&queued_item.id)?; |
| 8870 | assert_eq!( |
| 8871 | recovered_queued_item.status, |
| 8872 | TurnItemLifecycleStatus::Interrupted |
| 8873 | ); |
| 8874 | assert!(recovered_queued_item.ended_at.is_some()); |
| 8875 | |
| 8876 | Ok(()) |
| 8877 | } |
| 8878 | |
| 8879 | #[test] |
| 8880 | fn parse_mode_defaults_to_agent() { |
| 8881 | assert_eq!(parse_mode("unknown"), AppMode::Agent); |
| 8882 | assert_eq!(parse_mode("plan"), AppMode::Plan); |
| 8883 | } |
| 8884 | |
| 8885 | #[test] |
| 8886 | fn mode_only_override_preserves_legacy_full_access_posture() -> Result<()> { |
| 8887 | let mut thread = sample_thread("thr_legacy_full_access"); |
| 8888 | thread.permission_posture = None; |
| 8889 | thread.auto_approve = true; |
| 8890 | |
| 8891 | let policy = runtime_policy_with_overrides(&thread, Some("plan"), None, None)?; |
| 8892 | assert_eq!(policy.mode, AppMode::Plan); |
| 8893 | assert_eq!(policy.permission_wire(), "full_access"); |
| 8894 | |
| 8895 | let ask = runtime_policy_with_overrides(&thread, Some("act"), None, Some(false))?; |
| 8896 | assert_eq!(ask.permission_wire(), "ask"); |
| 8897 | Ok(()) |
| 8898 | } |
| 8899 | |
| 8900 | #[test] |
| 8901 | fn parse_mode_opt_resolves_explicit_tokens_and_aliases() { |
| 8902 | assert_eq!(parse_mode_opt("agent"), Some(AppMode::Agent)); |
| 8903 | assert_eq!(parse_mode_opt("1"), Some(AppMode::Agent)); |
| 8904 | assert_eq!(parse_mode_opt("plan"), Some(AppMode::Plan)); |
| 8905 | assert_eq!(parse_mode_opt("2"), Some(AppMode::Plan)); |
| 8906 | assert_eq!(parse_mode_opt("auto"), Some(AppMode::Agent)); |
| 8907 | assert_eq!(parse_mode_opt("operate"), Some(AppMode::Operate)); |
| 8908 | assert_eq!(parse_mode_opt("3"), Some(AppMode::Operate)); |
| 8909 | assert_eq!(parse_mode_opt("yolo"), Some(AppMode::Yolo)); |
| 8910 | assert_eq!(parse_mode_opt("4"), Some(AppMode::Yolo)); |
| 8911 | assert_eq!(parse_mode_opt(" PLAN "), Some(AppMode::Plan)); |
| 8912 | } |
| 8913 | |
| 8914 | #[test] |
| 8915 | fn parse_mode_opt_rejects_prompt_fragments() { |
| 8916 | for input in [ |
| 8917 | "plan a trip to Tokyo", |
| 8918 | "switch the agent on", |
| 8919 | "enter yolo mode", |
| 8920 | "agent of chaos", |
| 8921 | "mode", |
| 8922 | ] { |
| 8923 | assert_eq!(parse_mode_opt(input), None); |
| 8924 | } |
| 8925 | } |
| 8926 | |
| 8927 | #[test] |
| 8928 | fn parse_mode_wrapper_defaults_and_resolves_numeric_aliases() { |
| 8929 | assert_eq!(parse_mode("plan a trip to Tokyo"), AppMode::Agent); |
| 8930 | assert_eq!(parse_mode("auto"), AppMode::Agent); |
| 8931 | assert_eq!(parse_mode("1"), AppMode::Agent); |
| 8932 | assert_eq!(parse_mode("2"), AppMode::Plan); |
| 8933 | assert_eq!(parse_mode("3"), AppMode::Operate); |
| 8934 | assert_eq!(parse_mode("4"), AppMode::Yolo); |
| 8935 | } |
| 8936 | |
| 8937 | fn rebind_event(event: &str, agent_id: &str, seq: u64) -> RuntimeEventRecord { |
| 8938 | RuntimeEventRecord { |
| 8939 | schema_version: CURRENT_RUNTIME_SCHEMA_VERSION, |
| 8940 | seq, |
| 8941 | timestamp: Utc::now(), |
| 8942 | thread_id: "thr_test".to_string(), |
| 8943 | turn_id: Some("turn_test".to_string()), |
| 8944 | item_id: None, |
| 8945 | event: event.to_string(), |
| 8946 | payload: json!({ "agent_id": agent_id }), |
| 8947 | } |
| 8948 | } |
| 8949 | |
| 8950 | #[test] |
| 8951 | fn collect_agent_rebind_hints_resumes_a_mid_fanout_session() { |
| 8952 | // Mirror what runtime_threads persists during a real fanout: three |
| 8953 | // workers spawned, two finished, one still running when the session |
| 8954 | // was killed. The TUI re-attach must rebuild placeholders for the |
| 8955 | // running worker AND the two completed workers (the fanout card |
| 8956 | // tracks all of them so the dot-grid stays accurate post-resume). |
| 8957 | let events = vec![ |
| 8958 | rebind_event("agent.spawned", "agent_a", 1), |
| 8959 | rebind_event("agent.spawned", "agent_b", 2), |
| 8960 | rebind_event("agent.spawned", "agent_c", 3), |
| 8961 | rebind_event("agent.progress", "agent_a", 4), |
| 8962 | rebind_event("agent.completed", "agent_a", 5), |
| 8963 | rebind_event("agent.progress", "agent_b", 6), |
| 8964 | rebind_event("agent.completed", "agent_b", 7), |
| 8965 | rebind_event("agent.progress", "agent_c", 8), |
| 8966 | ]; |
| 8967 | let hints = collect_agent_rebind_hints(&events); |
| 8968 | assert_eq!(hints.len(), 3, "every fanout worker must be rebound"); |
| 8969 | let by_id: std::collections::BTreeMap<&str, AgentRebindStatus> = hints |
| 8970 | .iter() |
| 8971 | .map(|h| (h.agent_id.as_str(), h.status)) |
| 8972 | .collect(); |
| 8973 | assert_eq!(by_id.get("agent_a"), Some(&AgentRebindStatus::Completed)); |
| 8974 | assert_eq!(by_id.get("agent_b"), Some(&AgentRebindStatus::Completed)); |
| 8975 | assert_eq!( |
| 8976 | by_id.get("agent_c"), |
| 8977 | Some(&AgentRebindStatus::InProgress), |
| 8978 | "in-flight worker must rebind in InProgress, not downgrade" |
| 8979 | ); |
| 8980 | } |
| 8981 | |
| 8982 | #[test] |
| 8983 | fn collect_agent_rebind_hints_ignores_unrelated_events() { |
| 8984 | // Status / tool events should not produce phantom hints — only the |
| 8985 | // agent.* family carries the contract we re-bind from. |
| 8986 | let events = vec![ |
| 8987 | RuntimeEventRecord { |
| 8988 | schema_version: CURRENT_RUNTIME_SCHEMA_VERSION, |
| 8989 | seq: 1, |
| 8990 | timestamp: Utc::now(), |
| 8991 | thread_id: "thr".to_string(), |
| 8992 | turn_id: None, |
| 8993 | item_id: None, |
| 8994 | event: "tool.completed".to_string(), |
| 8995 | payload: json!({"name": "read_file"}), |
| 8996 | }, |
| 8997 | rebind_event("agent.spawned", "agent_x", 2), |
| 8998 | RuntimeEventRecord { |
| 8999 | schema_version: CURRENT_RUNTIME_SCHEMA_VERSION, |
| 9000 | seq: 3, |
| 9001 | timestamp: Utc::now(), |
| 9002 | thread_id: "thr".to_string(), |
| 9003 | turn_id: None, |
| 9004 | item_id: None, |
| 9005 | event: "compaction.completed".to_string(), |
| 9006 | payload: json!({"messages_after": 12}), |
| 9007 | }, |
| 9008 | ]; |
| 9009 | let hints = collect_agent_rebind_hints(&events); |
| 9010 | assert_eq!(hints.len(), 1); |
| 9011 | assert_eq!(hints[0].agent_id, "agent_x"); |
| 9012 | } |
| 9013 | |
| 9014 | #[test] |
| 9015 | fn collect_agent_rebind_hints_does_not_downgrade_completed_to_in_progress() { |
| 9016 | // Out-of-order replay: a stale `agent.progress` arriving after the |
| 9017 | // completed event must NOT clobber the terminal status. This matters |
| 9018 | // when an event log is concatenated from interrupted segments. |
| 9019 | let events = vec![ |
| 9020 | rebind_event("agent.spawned", "agent_y", 1), |
| 9021 | rebind_event("agent.completed", "agent_y", 2), |
| 9022 | rebind_event("agent.progress", "agent_y", 3), |
| 9023 | ]; |
| 9024 | let hints = collect_agent_rebind_hints(&events); |
| 9025 | assert_eq!(hints.len(), 1); |
| 9026 | assert_eq!(hints[0].status, AgentRebindStatus::Completed); |
| 9027 | } |
| 9028 | |
| 9029 | /// Helper for the `fork_at_user_message` tests: write a sequence of |
| 9030 | /// (user, assistant) turns under the given thread id. Each turn gets |
| 9031 | /// one UserMessage item carrying `user_text` in `detail` plus one |
| 9032 | /// AgentMessage item. Turn `created_at` is monotonically increasing |
| 9033 | /// so the chronological sort in `list_turns_for_thread` is stable. |
| 9034 | fn seed_turns_with_user_messages( |
| 9035 | manager: &RuntimeThreadManager, |
| 9036 | thread_id: &str, |
| 9037 | user_texts: &[&str], |
| 9038 | ) -> Result<Vec<String>> { |
| 9039 | let mut turn_ids = Vec::new(); |
| 9040 | let base = Utc::now(); |
| 9041 | for (offset, text) in user_texts.iter().enumerate() { |
| 9042 | let created_at = base + chrono::Duration::milliseconds(offset as i64); |
| 9043 | let turn_id = format!("turn_test_{offset}"); |
| 9044 | let user_item_id = format!("item_user_{offset}"); |
| 9045 | let asst_item_id = format!("item_asst_{offset}"); |
| 9046 | manager.store.save_item(&TurnItemRecord { |
| 9047 | schema_version: CURRENT_RUNTIME_SCHEMA_VERSION, |
| 9048 | id: user_item_id.clone(), |
| 9049 | turn_id: turn_id.clone(), |
| 9050 | kind: TurnItemKind::UserMessage, |
| 9051 | status: TurnItemLifecycleStatus::Completed, |
| 9052 | summary: (*text).to_string(), |
| 9053 | detail: Some((*text).to_string()), |
| 9054 | metadata: None, |
| 9055 | artifact_refs: Vec::new(), |
| 9056 | started_at: Some(created_at), |
| 9057 | ended_at: Some(created_at), |
| 9058 | })?; |
| 9059 | manager.store.save_item(&TurnItemRecord { |
| 9060 | schema_version: CURRENT_RUNTIME_SCHEMA_VERSION, |
| 9061 | id: asst_item_id.clone(), |
| 9062 | turn_id: turn_id.clone(), |
| 9063 | kind: TurnItemKind::AgentMessage, |
| 9064 | status: TurnItemLifecycleStatus::Completed, |
| 9065 | summary: format!("reply {offset}"), |
| 9066 | detail: Some(format!("reply {offset}")), |
| 9067 | metadata: None, |
| 9068 | artifact_refs: Vec::new(), |
| 9069 | started_at: Some(created_at), |
| 9070 | ended_at: Some(created_at), |
| 9071 | })?; |
| 9072 | manager.store.save_turn(&TurnRecord { |
| 9073 | schema_version: CURRENT_RUNTIME_SCHEMA_VERSION, |
| 9074 | id: turn_id.clone(), |
| 9075 | thread_id: thread_id.to_string(), |
| 9076 | status: RuntimeTurnStatus::Completed, |
| 9077 | input_summary: (*text).to_string(), |
| 9078 | created_at, |
| 9079 | started_at: Some(created_at), |
| 9080 | ended_at: Some(created_at), |
| 9081 | duration_ms: Some(0), |
| 9082 | usage: None, |
| 9083 | permission_posture: None, |
| 9084 | effective_provider: None, |
| 9085 | effective_provider_id: None, |
| 9086 | effective_billing_surface: None, |
| 9087 | effective_endpoint_fingerprint: None, |
| 9088 | effective_billing_mode: None, |
| 9089 | effective_dispatched_at: None, |
| 9090 | effective_model: None, |
| 9091 | routed_usage: Vec::new(), |
| 9092 | routed_usage_source_ids: Vec::new(), |
| 9093 | routed_usage_dropped_records: 0, |
| 9094 | error: None, |
| 9095 | item_ids: vec![user_item_id, asst_item_id], |
| 9096 | steer_count: 0, |
| 9097 | })?; |
| 9098 | turn_ids.push(turn_id); |
| 9099 | } |
| 9100 | Ok(turn_ids) |
| 9101 | } |
| 9102 | |
| 9103 | #[tokio::test] |
| 9104 | async fn fork_at_user_message_drops_tail_and_returns_user_text() -> Result<()> { |
| 9105 | // Seed three completed user/assistant turns. Backtracking with |
| 9106 | // depth=0 should drop only the most recent turn ("third") and |
| 9107 | // hand back its original text so the caller can refill the |
| 9108 | // composer. |
| 9109 | let manager = test_manager(test_runtime_dir())?; |
| 9110 | let thread = manager |
| 9111 | .create_thread(CreateThreadRequest { |
| 9112 | model: None, |
| 9113 | workspace: None, |
| 9114 | mode: None, |
| 9115 | allow_shell: None, |
| 9116 | trust_mode: None, |
| 9117 | auto_approve: None, |
| 9118 | archived: false, |
| 9119 | system_prompt: None, |
| 9120 | task_id: None, |
| 9121 | ..Default::default() |
| 9122 | }) |
| 9123 | .await?; |
| 9124 | seed_turns_with_user_messages(&manager, &thread.id, &["first", "second", "third"])?; |
| 9125 | |
| 9126 | let (forked, original_text) = manager.fork_at_user_message(&thread.id, 0).await?; |
| 9127 | assert_eq!(original_text.as_deref(), Some("third")); |
| 9128 | assert_ne!(forked.id, thread.id); |
| 9129 | |
| 9130 | let forked_turns = manager.store.list_turns_for_thread(&forked.id)?; |
| 9131 | assert_eq!( |
| 9132 | forked_turns.len(), |
| 9133 | 2, |
| 9134 | "depth=0 should drop the most recent turn" |
| 9135 | ); |
| 9136 | let summaries: Vec<&str> = forked_turns |
| 9137 | .iter() |
| 9138 | .map(|t| t.input_summary.as_str()) |
| 9139 | .collect(); |
| 9140 | assert_eq!(summaries, vec!["first", "second"]); |
| 9141 | Ok(()) |
| 9142 | } |
| 9143 | |
| 9144 | #[tokio::test] |
| 9145 | async fn fork_at_user_message_depth_one_drops_two_turns() -> Result<()> { |
| 9146 | let manager = test_manager(test_runtime_dir())?; |
| 9147 | let thread = manager |
| 9148 | .create_thread(CreateThreadRequest { |
| 9149 | model: None, |
| 9150 | workspace: None, |
| 9151 | mode: None, |
| 9152 | allow_shell: None, |
| 9153 | trust_mode: None, |
| 9154 | auto_approve: None, |
| 9155 | archived: false, |
| 9156 | system_prompt: None, |
| 9157 | task_id: None, |
| 9158 | ..Default::default() |
| 9159 | }) |
| 9160 | .await?; |
| 9161 | seed_turns_with_user_messages(&manager, &thread.id, &["a", "b", "c", "d"])?; |
| 9162 | |
| 9163 | let (forked, original_text) = manager.fork_at_user_message(&thread.id, 1).await?; |
| 9164 | assert_eq!(original_text.as_deref(), Some("c")); |
| 9165 | let forked_turns = manager.store.list_turns_for_thread(&forked.id)?; |
| 9166 | let summaries: Vec<&str> = forked_turns |
| 9167 | .iter() |
| 9168 | .map(|t| t.input_summary.as_str()) |
| 9169 | .collect(); |
| 9170 | assert_eq!(summaries, vec!["a", "b"]); |
| 9171 | Ok(()) |
| 9172 | } |
| 9173 | |
| 9174 | #[tokio::test] |
| 9175 | async fn fork_at_user_message_out_of_range_errors() -> Result<()> { |
| 9176 | let manager = test_manager(test_runtime_dir())?; |
| 9177 | let thread = manager |
| 9178 | .create_thread(CreateThreadRequest { |
| 9179 | model: None, |
| 9180 | workspace: None, |
| 9181 | mode: None, |
| 9182 | allow_shell: None, |
| 9183 | trust_mode: None, |
| 9184 | auto_approve: None, |
| 9185 | archived: false, |
| 9186 | system_prompt: None, |
| 9187 | task_id: None, |
| 9188 | ..Default::default() |
| 9189 | }) |
| 9190 | .await?; |
| 9191 | seed_turns_with_user_messages(&manager, &thread.id, &["only"])?; |
| 9192 | |
| 9193 | let err = manager.fork_at_user_message(&thread.id, 5).await.err(); |
| 9194 | assert!(err.is_some(), "depth past the end should bail out"); |
| 9195 | Ok(()) |
| 9196 | } |
| 9197 | |
| 9198 | #[tokio::test] |
| 9199 | async fn fork_at_user_message_does_not_mutate_source() -> Result<()> { |
| 9200 | // The source thread must be untouched: turns still present, items |
| 9201 | // still present, latest_turn_id still pointing at the original |
| 9202 | // tail. Backtrack creates a sibling, never edits in place. |
| 9203 | let manager = test_manager(test_runtime_dir())?; |
| 9204 | let thread = manager |
| 9205 | .create_thread(CreateThreadRequest { |
| 9206 | model: None, |
| 9207 | workspace: None, |
| 9208 | mode: None, |
| 9209 | allow_shell: None, |
| 9210 | trust_mode: None, |
| 9211 | auto_approve: None, |
| 9212 | archived: false, |
| 9213 | system_prompt: None, |
| 9214 | task_id: None, |
| 9215 | ..Default::default() |
| 9216 | }) |
| 9217 | .await?; |
| 9218 | let turn_ids = seed_turns_with_user_messages(&manager, &thread.id, &["x", "y", "z"])?; |
| 9219 | |
| 9220 | let _ = manager.fork_at_user_message(&thread.id, 0).await?; |
| 9221 | |
| 9222 | let source_turns = manager.store.list_turns_for_thread(&thread.id)?; |
| 9223 | assert_eq!( |
| 9224 | source_turns.len(), |
| 9225 | 3, |
| 9226 | "source thread must still hold every turn after fork" |
| 9227 | ); |
| 9228 | for tid in &turn_ids { |
| 9229 | assert!( |
| 9230 | manager.store.load_turn(tid).is_ok(), |
| 9231 | "turn {tid} must remain on disk" |
| 9232 | ); |
| 9233 | } |
| 9234 | Ok(()) |
| 9235 | } |
| 9236 | |
| 9237 | // ── compaction summary persistence (merge_summary_into_prompt) ── |
| 9238 | |
| 9239 | #[test] |
| 9240 | fn summary_merge_appends_section_to_base_prompt() { |
| 9241 | let merged = merge_summary_into_prompt( |
| 9242 | Some("You are a helpful agent."), |
| 9243 | "## 📋 Conversation Summary (Auto-Generated)\n\nUser prefers lists.", |
| 9244 | ); |
| 9245 | assert!(merged.starts_with("You are a helpful agent.")); |
| 9246 | assert!(merged.contains(COMPACTION_SUMMARY_BEGIN)); |
| 9247 | assert!(merged.contains("User prefers lists.")); |
| 9248 | assert!(merged.ends_with(COMPACTION_SUMMARY_END)); |
| 9249 | // Reload restore keys on the marker: SyncSession maps the record to |
| 9250 | // SystemPrompt::Text and extract_compaction_summary_prompt checks |
| 9251 | // `contains("Conversation Summary (Auto-Generated)")`. |
| 9252 | assert!(merged.contains("Conversation Summary (Auto-Generated)")); |
| 9253 | } |
| 9254 | |
| 9255 | #[test] |
| 9256 | fn summary_merge_replaces_existing_section_idempotently() { |
| 9257 | let first = merge_summary_into_prompt(Some("Base prompt."), "summary v1"); |
| 9258 | let second = merge_summary_into_prompt(Some(&first), "summary v2"); |
| 9259 | assert!(second.contains("summary v2")); |
| 9260 | assert!(!second.contains("summary v1")); |
| 9261 | assert_eq!( |
| 9262 | second.matches(COMPACTION_SUMMARY_BEGIN).count(), |
| 9263 | 1, |
| 9264 | "repeated compactions must swap the section, not stack duplicates" |
| 9265 | ); |
| 9266 | assert!(second.starts_with("Base prompt.")); |
| 9267 | } |
| 9268 | |
| 9269 | #[test] |
| 9270 | fn summary_merge_handles_missing_base() { |
| 9271 | let merged = merge_summary_into_prompt(None, "only summary"); |
| 9272 | assert!(merged.starts_with(COMPACTION_SUMMARY_BEGIN)); |
| 9273 | assert!(merged.contains("only summary")); |
| 9274 | let empty_base = merge_summary_into_prompt(Some(""), "only summary"); |
| 9275 | assert!(empty_base.starts_with(COMPACTION_SUMMARY_BEGIN)); |
| 9276 | } |
| 9277 | |
| 9278 | #[test] |
| 9279 | fn summary_strip_preserves_text_after_section() { |
| 9280 | let with_tail = format!( |
| 9281 | "Base.\n\n{COMPACTION_SUMMARY_BEGIN}\nold summary\n{COMPACTION_SUMMARY_END}\n\nTrailing rules." |
| 9282 | ); |
| 9283 | let stripped = strip_summary_section(&with_tail); |
| 9284 | assert!(stripped.contains("Base.")); |
| 9285 | assert!(stripped.contains("Trailing rules.")); |
| 9286 | assert!(!stripped.contains("old summary")); |
| 9287 | // Re-merge keeps the tail intact. |
| 9288 | let merged = merge_summary_into_prompt(Some(&with_tail), "new summary"); |
| 9289 | assert!(merged.contains("Trailing rules.")); |
| 9290 | assert!(merged.contains("new summary")); |
| 9291 | } |
| 9292 | |
| 9293 | #[test] |
| 9294 | fn summary_strip_handles_missing_end_sentinel() { |
| 9295 | let broken = format!("Base.\n\n{COMPACTION_SUMMARY_BEGIN}\ntruncated…"); |
| 9296 | let stripped = strip_summary_section(&broken); |
| 9297 | assert_eq!(stripped, "Base."); |
| 9298 | } |
| 9299 |