| 1 | import * as ScriptProgress from "@t3tools/provider-core/server/ScriptProgress"; |
| 2 | import * as SourceControlProviderRegistry from "../sourceControl/SourceControlProviderRegistry.ts"; |
| 3 | import * as McpProviderSessions from "@t3tools/provider-core/server/McpProviderSessions"; |
| 4 | import * as NodeServices from "@effect/platform-node/NodeServices"; |
| 5 | import { assert, it } from "@effect/vitest"; |
| 6 | import { |
| 7 | CommandId, |
| 8 | EventId, |
| 9 | MessageId, |
| 10 | type ModelSelection, |
| 11 | NodeId, |
| 12 | type OrchestrationV2Run, |
| 13 | ProjectId, |
| 14 | ProviderDriverKind, |
| 15 | ProviderInstanceId, |
| 16 | ProviderThreadId, |
| 17 | RunId, |
| 18 | ThreadId, |
| 19 | TurnItemId, |
| 20 | } from "@t3tools/contracts"; |
| 21 | import * as DateTime from "effect/DateTime"; |
| 22 | import * as Effect from "effect/Effect"; |
| 23 | import * as Layer from "effect/Layer"; |
| 24 | import * as Stream from "effect/Stream"; |
| 25 | import * as Tracer from "effect/Tracer"; |
| 26 | |
| 27 | import * as CheckpointStore from "../checkpointing/CheckpointStore.ts"; |
| 28 | import * as ServerConfig from "../config.ts"; |
| 29 | import * as McpSessionRegistryTestkit from "../mcp/McpSessionRegistry.testkit.ts"; |
| 30 | import * as SqlitePersistence from "../persistence/Sqlite.ts"; |
| 31 | import * as ProjectEnrichmentService from "../project/ProjectEnrichmentService.ts"; |
| 32 | import * as ProjectService from "../project/ProjectService.ts"; |
| 33 | import type { ProviderInstance } from "@t3tools/provider-core/server/driver"; |
| 34 | import * as ProviderInstanceRegistry from "../provider/ProviderInstanceRegistry.ts"; |
| 35 | import * as ServerSettings from "../serverSettings.ts"; |
| 36 | import * as VcsDriverRegistry from "../vcs/VcsDriverRegistry.ts"; |
| 37 | import * as VcsProcess from "../vcs/VcsProcess.ts"; |
| 38 | import * as WorkspacePaths from "../workspace/WorkspacePaths.ts"; |
| 39 | import { CodexProviderCapabilitiesV2 } from "./Adapters/CodexAdapterV2.ts"; |
| 40 | import * as EventSink from "./EventSink.ts"; |
| 41 | import * as Orchestrator from "./Orchestrator.ts"; |
| 42 | import { continueRestartedRun } from "./RestartContinuation.ts"; |
| 43 | import * as RuntimeLayer from "./runtimeLayer.ts"; |
| 44 | import * as ProviderTurnStartServiceTestkit from "./ProviderTurnStartService.testkit.ts"; |
| 45 | import type * as ProviderAdapter from "@t3tools/provider-core/server/ProviderAdapter"; |
| 46 | import * as ProviderContinuationRequests from "@t3tools/provider-core/server/ProviderContinuationRequests"; |
| 47 | |
| 48 | const layerPlatformTest = Layer.merge( |
| 49 | NodeServices.layer, |
| 50 | Layer.mock(SourceControlProviderRegistry.SourceControlProviderRegistry)({ |
| 51 | resolveLink: () => Effect.die("unused title link"), |
| 52 | }), |
| 53 | ); |
| 54 | |
| 55 | const layerServerConfig = ServerConfig.layerTest(process.cwd(), { |
| 56 | prefix: "t3-orchestration-v2-delegated-completion-", |
| 57 | }); |
| 58 | |
| 59 | const modelSelection = { |
| 60 | instanceId: ProviderInstanceId.make("codex"), |
| 61 | model: "gpt-5.4", |
| 62 | } satisfies ModelSelection; |
| 63 | |
| 64 | const layerVcsDriverRegistryTest = VcsDriverRegistry.layer.pipe( |
| 65 | Layer.provide(VcsProcess.layer), |
| 66 | Layer.provide(layerServerConfig), |
| 67 | Layer.provide(layerPlatformTest), |
| 68 | ); |
| 69 | |
| 70 | const layerCheckpointStoreTest = CheckpointStore.layer.pipe( |
| 71 | Layer.provide(layerVcsDriverRegistryTest), |
| 72 | ); |
| 73 | |
| 74 | const driver = ProviderDriverKind.make("codex"); |
| 75 | const orchestrationAdapter = { |
| 76 | instanceId: modelSelection.instanceId, |
| 77 | driver, |
| 78 | getCapabilities: () => Effect.succeed(CodexProviderCapabilitiesV2), |
| 79 | planSelectionTransition: () => Effect.succeed({ type: "apply_on_next_turn" }), |
| 80 | openSession: () => Effect.die("sessions are not used by delegated completion tests"), |
| 81 | } as ProviderAdapter.ProviderAdapterV2["Service"]; |
| 82 | const providerInstance = { |
| 83 | instanceId: modelSelection.instanceId, |
| 84 | driverKind: driver, |
| 85 | continuationIdentity: { |
| 86 | driverKind: driver, |
| 87 | continuationKey: "codex:test", |
| 88 | }, |
| 89 | displayName: "Codex test", |
| 90 | enabled: true, |
| 91 | // No supportedRuntimeModes: every runtime mode runs as stored. |
| 92 | snapshot: { getSnapshot: Effect.succeed({}) } as unknown as ProviderInstance["snapshot"], |
| 93 | orchestrationAdapter, |
| 94 | textGeneration: {} as ProviderInstance["textGeneration"], |
| 95 | } satisfies ProviderInstance; |
| 96 | |
| 97 | const layerTestProviderInstanceRegistry = Layer.succeed( |
| 98 | ProviderInstanceRegistry.ProviderInstanceRegistry, |
| 99 | { |
| 100 | getInstance: (instanceId) => |
| 101 | Effect.succeed(instanceId === providerInstance.instanceId ? providerInstance : undefined), |
| 102 | listInstances: Effect.succeed([providerInstance]), |
| 103 | listUnavailable: Effect.succeed([]), |
| 104 | streamChanges: Stream.empty, |
| 105 | subscribeChanges: Effect.never, |
| 106 | }, |
| 107 | ); |
| 108 | |
| 109 | const layerTest = Layer.mergeAll(RuntimeLayer.layer, RuntimeLayer.layerEventSink).pipe( |
| 110 | Layer.provideMerge(RuntimeLayer.layerProjectService), |
| 111 | Layer.provide( |
| 112 | Layer.mock(WorkspacePaths.WorkspacePaths)({ |
| 113 | normalizeWorkspaceRoot: (workspaceRoot) => Effect.succeed(workspaceRoot), |
| 114 | }), |
| 115 | ), |
| 116 | Layer.provide(ProviderTurnStartServiceTestkit.layer), |
| 117 | Layer.provide( |
| 118 | Layer.succeed(ProjectEnrichmentService.ProjectEnrichmentService, { |
| 119 | peek: () => |
| 120 | Effect.succeed({ |
| 121 | repositoryIdentity: null, |
| 122 | faviconPath: null, |
| 123 | repositoryIdentityResolved: false, |
| 124 | }), |
| 125 | request: () => Effect.void, |
| 126 | getAvailable: () => |
| 127 | Effect.succeed({ |
| 128 | repositoryIdentity: null, |
| 129 | faviconPath: null, |
| 130 | repositoryIdentityResolved: false, |
| 131 | }), |
| 132 | invalidate: () => Effect.void, |
| 133 | subscribeChanges: Effect.never, |
| 134 | }), |
| 135 | ), |
| 136 | Layer.provide(McpSessionRegistryTestkit.layer), |
| 137 | Layer.provide(Layer.merge(McpProviderSessions.layer, ScriptProgress.layerUnavailable)), |
| 138 | Layer.provide(SqlitePersistence.layerMemory), |
| 139 | Layer.provide(layerCheckpointStoreTest), |
| 140 | Layer.provide(layerServerConfig), |
| 141 | Layer.provide(ServerSettings.layerTest()), |
| 142 | Layer.provide(layerTestProviderInstanceRegistry), |
| 143 | Layer.provide(layerPlatformTest), |
| 144 | ); |
| 145 | |
| 146 | const seedParentWithTerminalTask = (input: { |
| 147 | readonly threadId: ThreadId; |
| 148 | readonly projectId: ProjectId; |
| 149 | readonly runId: RunId; |
| 150 | readonly rootNodeId: NodeId; |
| 151 | readonly taskId: NodeId; |
| 152 | readonly deliveryState: "delivered" | "claimed" | "acknowledged" | "disposed"; |
| 153 | readonly completionWake?: "always" | "settled_only"; |
| 154 | readonly deliveryTaskIds?: ReadonlyArray<NodeId>; |
| 155 | readonly now: DateTime.Utc; |
| 156 | }) => |
| 157 | Effect.gen(function* () { |
| 158 | const projects = yield* ProjectService.ProjectService; |
| 159 | const orchestrator = yield* Orchestrator.OrchestratorV2; |
| 160 | const eventSink = yield* EventSink.EventSinkV2; |
| 161 | const providerThreadId = ProviderThreadId.make( |
| 162 | `provider-thread:${String(input.threadId).replace("thread:", "")}`, |
| 163 | ); |
| 164 | |
| 165 | yield* projects.create({ |
| 166 | commandId: CommandId.make(`command:seed-project:${input.threadId}`), |
| 167 | projectId: input.projectId, |
| 168 | title: "Delegated completion delivery", |
| 169 | workspaceRoot: `/workspace/${input.projectId}`, |
| 170 | }); |
| 171 | |
| 172 | yield* orchestrator.dispatch({ |
| 173 | type: "thread.create", |
| 174 | createdBy: "user", |
| 175 | creationSource: "web", |
| 176 | commandId: CommandId.make(`command:seed-create:${input.threadId}`), |
| 177 | threadId: input.threadId, |
| 178 | projectId: input.projectId, |
| 179 | title: "Delegated completion delivery", |
| 180 | modelSelection, |
| 181 | runtimeMode: "full-access", |
| 182 | interactionMode: "default", |
| 183 | branch: null, |
| 184 | worktreePath: null, |
| 185 | }); |
| 186 | |
| 187 | yield* eventSink.write({ |
| 188 | commandId: CommandId.make(`command:seed-projection:${input.threadId}`), |
| 189 | events: [ |
| 190 | { |
| 191 | id: EventId.make(`event:seed-provider-thread:${input.threadId}`), |
| 192 | type: "provider-thread.updated", |
| 193 | threadId: input.threadId, |
| 194 | driver, |
| 195 | providerInstanceId: modelSelection.instanceId, |
| 196 | occurredAt: input.now, |
| 197 | payload: { |
| 198 | id: providerThreadId, |
| 199 | driver, |
| 200 | providerInstanceId: modelSelection.instanceId, |
| 201 | providerSessionId: null, |
| 202 | appThreadId: input.threadId, |
| 203 | ownerNodeId: input.rootNodeId, |
| 204 | nativeThreadRef: { |
| 205 | driver, |
| 206 | nativeId: `native:${input.threadId}`, |
| 207 | strength: "strong", |
| 208 | }, |
| 209 | nativeConversationHeadRef: null, |
| 210 | status: "active", |
| 211 | firstRunOrdinal: 1, |
| 212 | lastRunOrdinal: 1, |
| 213 | handoffIds: [], |
| 214 | forkedFrom: null, |
| 215 | createdAt: input.now, |
| 216 | updatedAt: input.now, |
| 217 | }, |
| 218 | }, |
| 219 | { |
| 220 | id: EventId.make(`event:seed-run:${input.threadId}`), |
| 221 | type: "run.updated", |
| 222 | threadId: input.threadId, |
| 223 | runId: input.runId, |
| 224 | nodeId: input.rootNodeId, |
| 225 | providerInstanceId: modelSelection.instanceId, |
| 226 | occurredAt: input.now, |
| 227 | payload: { |
| 228 | id: input.runId, |
| 229 | threadId: input.threadId, |
| 230 | ordinal: 1, |
| 231 | providerInstanceId: modelSelection.instanceId, |
| 232 | modelSelection, |
| 233 | providerThreadId, |
| 234 | userMessageId: MessageId.make(`message:seed-user:${input.threadId}`), |
| 235 | rootNodeId: input.rootNodeId, |
| 236 | activeAttemptId: null, |
| 237 | status: "running", |
| 238 | requestedAt: input.now, |
| 239 | startedAt: input.now, |
| 240 | completedAt: null, |
| 241 | checkpointId: null, |
| 242 | contextHandoffId: null, |
| 243 | delegatedCompletion: { |
| 244 | disposition: "open", |
| 245 | nextGeneration: 2, |
| 246 | delivery: |
| 247 | input.deliveryTaskIds === undefined |
| 248 | ? null |
| 249 | : { |
| 250 | generation: 1, |
| 251 | messageId: MessageId.make(`message:delegated-delivery:${input.threadId}`), |
| 252 | taskIds: input.deliveryTaskIds, |
| 253 | }, |
| 254 | }, |
| 255 | }, |
| 256 | }, |
| 257 | { |
| 258 | id: EventId.make(`event:seed-task:${input.threadId}`), |
| 259 | type: "subagent.updated", |
| 260 | threadId: input.threadId, |
| 261 | runId: input.runId, |
| 262 | nodeId: input.taskId, |
| 263 | driver, |
| 264 | providerInstanceId: modelSelection.instanceId, |
| 265 | occurredAt: input.now, |
| 266 | payload: { |
| 267 | id: input.taskId, |
| 268 | threadId: input.threadId, |
| 269 | runId: input.runId, |
| 270 | parentNodeId: input.rootNodeId, |
| 271 | origin: "app_owned", |
| 272 | createdBy: "agent", |
| 273 | driver, |
| 274 | providerInstanceId: modelSelection.instanceId, |
| 275 | providerThreadId: null, |
| 276 | childThreadId: null, |
| 277 | nativeTaskRef: null, |
| 278 | prompt: "Inspect the delivered ownership edge.", |
| 279 | title: null, |
| 280 | model: null, |
| 281 | completionWake: input.completionWake ?? "settled_only", |
| 282 | completionDelivery: { |
| 283 | state: input.deliveryState, |
| 284 | observedByRunId: input.deliveryState === "acknowledged" ? input.runId : null, |
| 285 | }, |
| 286 | status: "completed", |
| 287 | result: "child finished", |
| 288 | startedAt: input.now, |
| 289 | completedAt: input.now, |
| 290 | updatedAt: input.now, |
| 291 | }, |
| 292 | }, |
| 293 | ], |
| 294 | }); |
| 295 | }); |
| 296 | |
| 297 | it.layer(layerTest)("delegated completion delivery repairs", (it) => { |
| 298 | it.effect("acceptance batches pending siblings without acknowledging their results", () => |
| 299 | Effect.gen(function* () { |
| 300 | const orchestrator = yield* Orchestrator.OrchestratorV2; |
| 301 | const sink = yield* EventSink.EventSinkV2; |
| 302 | const now = yield* DateTime.now; |
| 303 | const threadId = ThreadId.make("mailbox-batch"); |
| 304 | const runId = RunId.make("mailbox-parent"); |
| 305 | const taskId = NodeId.make("mailbox-first"); |
| 306 | const messageId = MessageId.make(`message:delegated-delivery:${threadId}`); |
| 307 | yield* seedParentWithTerminalTask({ |
| 308 | threadId, |
| 309 | runId, |
| 310 | projectId: ProjectId.make("mailbox-project"), |
| 311 | rootNodeId: NodeId.make("mailbox-root"), |
| 312 | taskId, |
| 313 | deliveryState: "claimed", |
| 314 | completionWake: "always", |
| 315 | deliveryTaskIds: [taskId], |
| 316 | now, |
| 317 | }); |
| 318 | const projection = yield* orchestrator.getThreadProjection(threadId); |
| 319 | const task = projection.subagents[0]!; |
| 320 | const pendingIds = [NodeId.make("mailbox-second"), NodeId.make("mailbox-third")]; |
| 321 | yield* sink.write({ |
| 322 | events: [ |
| 323 | { |
| 324 | id: EventId.make("mailbox-message"), |
| 325 | type: "message.updated", |
| 326 | threadId, |
| 327 | runId, |
| 328 | occurredAt: now, |
| 329 | payload: { |
| 330 | id: messageId, |
| 331 | threadId, |
| 332 | runId, |
| 333 | nodeId: task.parentNodeId, |
| 334 | role: "user", |
| 335 | text: "Background task finished", |
| 336 | attachments: [], |
| 337 | streaming: false, |
| 338 | createdBy: "agent", |
| 339 | creationSource: "server", |
| 340 | createdAt: now, |
| 341 | updatedAt: now, |
| 342 | delegatedCompletion: { parentRunId: runId, generation: 1, taskIds: [taskId] }, |
| 343 | }, |
| 344 | }, |
| 345 | ...pendingIds.map((id) => ({ |
| 346 | id: EventId.make(`event:${id}`), |
| 347 | type: "subagent.updated" as const, |
| 348 | threadId, |
| 349 | runId, |
| 350 | nodeId: id, |
| 351 | occurredAt: now, |
| 352 | payload: { |
| 353 | ...task, |
| 354 | id, |
| 355 | completionDelivery: { state: "pending" as const, observedByRunId: null }, |
| 356 | }, |
| 357 | })), |
| 358 | ], |
| 359 | }); |
| 360 | yield* orchestrator.dispatch({ |
| 361 | type: "notification.delivery.accept", |
| 362 | commandId: CommandId.make("accept-first"), |
| 363 | threadId, |
| 364 | messageId, |
| 365 | }); |
| 366 | const accepted = yield* orchestrator.getThreadProjection(threadId); |
| 367 | assert.equal( |
| 368 | accepted.subagents.find((row) => row.id === taskId)?.completionDelivery?.state, |
| 369 | "delivered", |
| 370 | ); |
| 371 | const cohort = accepted.runs.find((row) => row.id === runId)?.delegatedCompletion; |
| 372 | assert.deepEqual(cohort?.delivery?.taskIds, pendingIds); |
| 373 | assert.equal(cohort?.delivery?.generation, 2); |
| 374 | for (const id of pendingIds) { |
| 375 | assert.deepEqual(accepted.subagents.find((row) => row.id === id)?.completionDelivery, { |
| 376 | state: "claimed", |
| 377 | observedByRunId: null, |
| 378 | }); |
| 379 | } |
| 380 | yield* orchestrator.dispatch({ |
| 381 | type: "notification.delivery.accept", |
| 382 | commandId: CommandId.make("repeat-old-acceptance"), |
| 383 | threadId, |
| 384 | messageId, |
| 385 | }); |
| 386 | const duplicate = yield* orchestrator.getThreadProjection(threadId); |
| 387 | assert.deepEqual(duplicate.runs.find((row) => row.id === runId)?.delegatedCompletion, cohort); |
| 388 | }), |
| 389 | ); |
| 390 | |
| 391 | it.effect("acceptance batches a settled_only sibling once its spawning run ended", () => |
| 392 | Effect.gen(function* () { |
| 393 | const orchestrator = yield* Orchestrator.OrchestratorV2; |
| 394 | const sink = yield* EventSink.EventSinkV2; |
| 395 | const now = yield* DateTime.now; |
| 396 | const threadId = ThreadId.make("mailbox-mixed"); |
| 397 | const runId = RunId.make("mailbox-mixed-parent"); |
| 398 | const taskId = NodeId.make("mailbox-mixed-always"); |
| 399 | const siblingId = NodeId.make("mailbox-mixed-settled-only"); |
| 400 | const messageId = MessageId.make(`message:delegated-delivery:${threadId}`); |
| 401 | yield* seedParentWithTerminalTask({ |
| 402 | threadId, |
| 403 | runId, |
| 404 | projectId: ProjectId.make("mailbox-mixed-project"), |
| 405 | rootNodeId: NodeId.make("mailbox-mixed-root"), |
| 406 | taskId, |
| 407 | deliveryState: "claimed", |
| 408 | completionWake: "always", |
| 409 | deliveryTaskIds: [taskId], |
| 410 | now, |
| 411 | }); |
| 412 | const projection = yield* orchestrator.getThreadProjection(threadId); |
| 413 | const task = projection.subagents[0]!; |
| 414 | const spawningRun = projection.runs.find((row) => row.id === runId)!; |
| 415 | // A restart cut the spawning run; its continuation is the live run now. |
| 416 | yield* sink.write({ |
| 417 | events: [ |
| 418 | { |
| 419 | ...runEvent({ threadId, runId, ordinal: 1, status: "cancelled", now }), |
| 420 | payload: { ...spawningRun, status: "cancelled" as const, completedAt: now }, |
| 421 | }, |
| 422 | runEvent({ |
| 423 | threadId, |
| 424 | runId: RunId.make("mailbox-mixed-continuation"), |
| 425 | ordinal: 2, |
| 426 | status: "running", |
| 427 | now, |
| 428 | }), |
| 429 | { |
| 430 | id: EventId.make("mailbox-mixed-message"), |
| 431 | type: "message.updated", |
| 432 | threadId, |
| 433 | runId, |
| 434 | occurredAt: now, |
| 435 | payload: { |
| 436 | id: messageId, |
| 437 | threadId, |
| 438 | runId, |
| 439 | nodeId: task.parentNodeId, |
| 440 | role: "user", |
| 441 | text: "Background task finished", |
| 442 | attachments: [], |
| 443 | streaming: false, |
| 444 | createdBy: "agent", |
| 445 | creationSource: "server", |
| 446 | createdAt: now, |
| 447 | updatedAt: now, |
| 448 | delegatedCompletion: { parentRunId: runId, generation: 1, taskIds: [taskId] }, |
| 449 | }, |
| 450 | }, |
| 451 | { |
| 452 | id: EventId.make(`event:${siblingId}`), |
| 453 | type: "subagent.updated" as const, |
| 454 | threadId, |
| 455 | runId, |
| 456 | nodeId: siblingId, |
| 457 | occurredAt: now, |
| 458 | payload: { |
| 459 | ...task, |
| 460 | id: siblingId, |
| 461 | completionWake: "settled_only" as const, |
| 462 | completionDelivery: { state: "pending" as const, observedByRunId: null }, |
| 463 | }, |
| 464 | }, |
| 465 | ], |
| 466 | }); |
| 467 | yield* orchestrator.dispatch({ |
| 468 | type: "notification.delivery.accept", |
| 469 | commandId: CommandId.make("accept-mixed"), |
| 470 | threadId, |
| 471 | messageId, |
| 472 | }); |
| 473 | const accepted = yield* orchestrator.getThreadProjection(threadId); |
| 474 | const cohort = accepted.runs.find((row) => row.id === runId)?.delegatedCompletion; |
| 475 | assert.deepEqual(cohort?.delivery?.taskIds, [siblingId]); |
| 476 | assert.equal( |
| 477 | accepted.subagents.find((row) => row.id === siblingId)?.completionDelivery?.state, |
| 478 | "claimed", |
| 479 | ); |
| 480 | }), |
| 481 | ); |
| 482 | |
| 483 | it.effect("builds completion text and metadata from the same live cohort", () => |
| 484 | Effect.gen(function* () { |
| 485 | const orchestrator = yield* Orchestrator.OrchestratorV2; |
| 486 | const now = yield* DateTime.now; |
| 487 | const threadId = ThreadId.make("thread:delegated-delivery-live-cohort"); |
| 488 | const projectId = ProjectId.make("project:delegated-delivery-live-cohort"); |
| 489 | const runId = RunId.make("run:delegated-delivery-live-cohort"); |
| 490 | const rootNodeId = NodeId.make("node:delegated-delivery-live-cohort-root"); |
| 491 | const firstTaskId = NodeId.make("node:delegated-delivery-live-cohort-first"); |
| 492 | const secondTaskId = NodeId.make("node:delegated-delivery-live-cohort-second"); |
| 493 | const messageId = MessageId.make(`message:delegated-delivery:${threadId}`); |
| 494 | |
| 495 | yield* seedParentWithTerminalTask({ |
| 496 | threadId, |
| 497 | projectId, |
| 498 | runId, |
| 499 | rootNodeId, |
| 500 | taskId: firstTaskId, |
| 501 | deliveryState: "claimed", |
| 502 | completionWake: "always", |
| 503 | deliveryTaskIds: [firstTaskId, secondTaskId], |
| 504 | now, |
| 505 | }); |
| 506 | |
| 507 | yield* orchestrator.dispatch({ |
| 508 | type: "message.dispatch", |
| 509 | commandId: CommandId.make("command:delegated-delivery-live-cohort"), |
| 510 | threadId, |
| 511 | messageId, |
| 512 | text: `Delegated task ${firstTaskId} reached a terminal state.`, |
| 513 | attachments: [], |
| 514 | dispatchMode: { type: "queue_after_active" }, |
| 515 | createdBy: "agent", |
| 516 | creationSource: "server", |
| 517 | delegatedCompletion: { |
| 518 | parentRunId: runId, |
| 519 | generation: 1, |
| 520 | taskIds: [firstTaskId], |
| 521 | }, |
| 522 | }); |
| 523 | |
| 524 | const projection = yield* orchestrator.getThreadProjection(threadId); |
| 525 | const message = projection.messages.find((candidate) => candidate.id === messageId); |
| 526 | assert.deepEqual(message?.delegatedCompletion?.taskIds, [firstTaskId, secondTaskId]); |
| 527 | assert.include(message?.text ?? "", String(firstTaskId)); |
| 528 | assert.include(message?.text ?? "", String(secondTaskId)); |
| 529 | assert.include(message?.text ?? "", "task_status"); |
| 530 | }), |
| 531 | ); |
| 532 | |
| 533 | it.effect("does not re-offer when wake-policy upgrades after delivered ownership settled", () => |
| 534 | Effect.gen(function* () { |
| 535 | const orchestrator = yield* Orchestrator.OrchestratorV2; |
| 536 | const now = yield* DateTime.now; |
| 537 | const threadId = ThreadId.make("thread:delegated-delivery-a1"); |
| 538 | const projectId = ProjectId.make("project:delegated-delivery-a1"); |
| 539 | const runId = RunId.make("run:delegated-delivery-a1"); |
| 540 | const rootNodeId = NodeId.make("node:delegated-delivery-a1-root"); |
| 541 | const taskId = NodeId.make("node:delegated-delivery-a1-task"); |
| 542 | |
| 543 | yield* seedParentWithTerminalTask({ |
| 544 | threadId, |
| 545 | projectId, |
| 546 | runId, |
| 547 | rootNodeId, |
| 548 | taskId, |
| 549 | deliveryState: "delivered", |
| 550 | completionWake: "settled_only", |
| 551 | now, |
| 552 | }); |
| 553 | |
| 554 | const upgrade = yield* orchestrator.dispatch({ |
| 555 | type: "delegated_task.wake-policy", |
| 556 | commandId: CommandId.make("command:delegated-delivery-a1:wake-policy"), |
| 557 | parentThreadId: threadId, |
| 558 | taskId, |
| 559 | completionWake: "always", |
| 560 | }); |
| 561 | |
| 562 | const projection = yield* orchestrator.getThreadProjection(threadId); |
| 563 | const task = projection.subagents.find((candidate) => candidate.id === taskId); |
| 564 | const parentRun = projection.runs.find((candidate) => candidate.id === runId); |
| 565 | |
| 566 | assert.equal(task?.completionWake, "always"); |
| 567 | assert.deepEqual(task?.completionDelivery, { |
| 568 | state: "delivered", |
| 569 | observedByRunId: null, |
| 570 | }); |
| 571 | assert.deepEqual(parentRun?.delegatedCompletion, { |
| 572 | disposition: "open", |
| 573 | nextGeneration: 2, |
| 574 | delivery: null, |
| 575 | }); |
| 576 | assert.isFalse( |
| 577 | upgrade.storedEvents.some( |
| 578 | (stored) => |
| 579 | stored.event.type === "subagent.updated" && |
| 580 | stored.event.payload.id === taskId && |
| 581 | stored.event.payload.completionDelivery?.state === "claimed", |
| 582 | ), |
| 583 | ); |
| 584 | assert.isFalse( |
| 585 | upgrade.storedEvents.some( |
| 586 | (stored) => |
| 587 | stored.event.type === "run.updated" && |
| 588 | stored.event.payload.id === runId && |
| 589 | stored.event.payload.delegatedCompletion?.delivery !== null && |
| 590 | stored.event.payload.delegatedCompletion?.delivery !== undefined, |
| 591 | ), |
| 592 | ); |
| 593 | }), |
| 594 | ); |
| 595 | |
| 596 | it.effect( |
| 597 | "treats repeated acknowledge and dispose with distinct command IDs as successful no-ops", |
| 598 | () => |
| 599 | Effect.gen(function* () { |
| 600 | const orchestrator = yield* Orchestrator.OrchestratorV2; |
| 601 | const now = yield* DateTime.now; |
| 602 | const threadId = ThreadId.make("thread:delegated-delivery-a2"); |
| 603 | const projectId = ProjectId.make("project:delegated-delivery-a2"); |
| 604 | const runId = RunId.make("run:delegated-delivery-a2"); |
| 605 | const rootNodeId = NodeId.make("node:delegated-delivery-a2-root"); |
| 606 | const taskId = NodeId.make("node:delegated-delivery-a2-task"); |
| 607 | |
| 608 | yield* seedParentWithTerminalTask({ |
| 609 | threadId, |
| 610 | projectId, |
| 611 | runId, |
| 612 | rootNodeId, |
| 613 | taskId, |
| 614 | deliveryState: "delivered", |
| 615 | completionWake: "always", |
| 616 | now, |
| 617 | }); |
| 618 | |
| 619 | // Distinct command IDs mirror task_status vs t3_thread_read racing after |
| 620 | // their shared read preflight saw delivered ownership. |
| 621 | const firstAck = yield* orchestrator.dispatch({ |
| 622 | type: "delegated_task.completion-delivery.acknowledge", |
| 623 | commandId: CommandId.make("command:delegated-delivery-a2:ack-task-status"), |
| 624 | parentThreadId: threadId, |
| 625 | taskId, |
| 626 | observedByRunId: runId, |
| 627 | }); |
| 628 | const secondAck = yield* orchestrator.dispatch({ |
| 629 | type: "delegated_task.completion-delivery.acknowledge", |
| 630 | commandId: CommandId.make("command:delegated-delivery-a2:ack-thread-read"), |
| 631 | parentThreadId: threadId, |
| 632 | taskId, |
| 633 | observedByRunId: runId, |
| 634 | }); |
| 635 | |
| 636 | const firstAckTask = firstAck.storedEvents.find( |
| 637 | (stored) => |
| 638 | stored.event.type === "subagent.updated" && stored.event.payload.id === taskId, |
| 639 | ); |
| 640 | const secondAckTask = secondAck.storedEvents.find( |
| 641 | (stored) => |
| 642 | stored.event.type === "subagent.updated" && stored.event.payload.id === taskId, |
| 643 | ); |
| 644 | assert.isDefined(firstAckTask); |
| 645 | assert.isDefined(secondAckTask); |
| 646 | if ( |
| 647 | firstAckTask?.event.type !== "subagent.updated" || |
| 648 | secondAckTask?.event.type !== "subagent.updated" |
| 649 | ) { |
| 650 | return yield* Effect.die(new Error("Acknowledge events missing.")); |
| 651 | } |
| 652 | assert.equal(firstAckTask.event.payload.completionDelivery?.state, "acknowledged"); |
| 653 | assert.equal(secondAckTask.event.payload.completionDelivery?.state, "acknowledged"); |
| 654 | // Idempotent replay keeps the first observation's ownership and timestamp. |
| 655 | assert.deepEqual( |
| 656 | secondAckTask.event.payload.completionDelivery, |
| 657 | firstAckTask.event.payload.completionDelivery, |
| 658 | ); |
| 659 | assert.deepEqual( |
| 660 | secondAckTask.event.payload.updatedAt, |
| 661 | firstAckTask.event.payload.updatedAt, |
| 662 | ); |
| 663 | assert.equal(secondAck.storedEvents.length, 1); |
| 664 | |
| 665 | const afterAck = yield* orchestrator.getThreadProjection(threadId); |
| 666 | assert.deepEqual( |
| 667 | afterAck.subagents.find((candidate) => candidate.id === taskId)?.completionDelivery, |
| 668 | { |
| 669 | state: "acknowledged", |
| 670 | observedByRunId: runId, |
| 671 | }, |
| 672 | ); |
| 673 | |
| 674 | const firstDispose = yield* orchestrator.dispatch({ |
| 675 | type: "delegated_task.completion-delivery.dispose", |
| 676 | commandId: CommandId.make("command:delegated-delivery-a2:dispose-task-status"), |
| 677 | parentThreadId: threadId, |
| 678 | taskId, |
| 679 | }); |
| 680 | const secondDispose = yield* orchestrator.dispatch({ |
| 681 | type: "delegated_task.completion-delivery.dispose", |
| 682 | commandId: CommandId.make("command:delegated-delivery-a2:dispose-thread-read"), |
| 683 | parentThreadId: threadId, |
| 684 | taskId, |
| 685 | }); |
| 686 | |
| 687 | const firstDisposeTask = firstDispose.storedEvents.find( |
| 688 | (stored) => |
| 689 | stored.event.type === "subagent.updated" && stored.event.payload.id === taskId, |
| 690 | ); |
| 691 | const secondDisposeTask = secondDispose.storedEvents.find( |
| 692 | (stored) => |
| 693 | stored.event.type === "subagent.updated" && stored.event.payload.id === taskId, |
| 694 | ); |
| 695 | assert.isDefined(firstDisposeTask); |
| 696 | assert.isDefined(secondDisposeTask); |
| 697 | if ( |
| 698 | firstDisposeTask?.event.type !== "subagent.updated" || |
| 699 | secondDisposeTask?.event.type !== "subagent.updated" |
| 700 | ) { |
| 701 | return yield* Effect.die(new Error("Dispose events missing.")); |
| 702 | } |
| 703 | assert.equal(firstDisposeTask.event.payload.completionDelivery?.state, "disposed"); |
| 704 | assert.equal(secondDisposeTask.event.payload.completionDelivery?.state, "disposed"); |
| 705 | assert.deepEqual( |
| 706 | secondDisposeTask.event.payload.completionDelivery, |
| 707 | firstDisposeTask.event.payload.completionDelivery, |
| 708 | ); |
| 709 | assert.equal(secondDispose.storedEvents.length, 1); |
| 710 | |
| 711 | const afterDispose = yield* orchestrator.getThreadProjection(threadId); |
| 712 | assert.deepEqual( |
| 713 | afterDispose.subagents.find((candidate) => candidate.id === taskId)?.completionDelivery, |
| 714 | { |
| 715 | state: "disposed", |
| 716 | observedByRunId: null, |
| 717 | }, |
| 718 | ); |
| 719 | |
| 720 | const acknowledgeAfterDispose = yield* orchestrator.dispatch({ |
| 721 | type: "delegated_task.completion-delivery.acknowledge", |
| 722 | commandId: CommandId.make("command:delegated-delivery-a2:ack-after-dispose"), |
| 723 | parentThreadId: threadId, |
| 724 | taskId, |
| 725 | observedByRunId: runId, |
| 726 | }); |
| 727 | const acknowledgedTask = acknowledgeAfterDispose.storedEvents.find( |
| 728 | (stored) => stored.event.type === "subagent.updated", |
| 729 | ); |
| 730 | if (acknowledgedTask?.event.type !== "subagent.updated") { |
| 731 | return yield* Effect.die(new Error("Acknowledge-after-dispose event missing.")); |
| 732 | } |
| 733 | assert.deepEqual(acknowledgedTask.event.payload.completionDelivery, { |
| 734 | state: "disposed", |
| 735 | observedByRunId: null, |
| 736 | }); |
| 737 | assert.equal(acknowledgeAfterDispose.storedEvents.length, 1); |
| 738 | |
| 739 | const afterStaleAcknowledge = yield* orchestrator.getThreadProjection(threadId); |
| 740 | assert.deepEqual( |
| 741 | afterStaleAcknowledge.subagents.find((candidate) => candidate.id === taskId) |
| 742 | ?.completionDelivery, |
| 743 | { |
| 744 | state: "disposed", |
| 745 | observedByRunId: null, |
| 746 | }, |
| 747 | ); |
| 748 | }), |
| 749 | ); |
| 750 | }); |
| 751 | |
| 752 | // Runtime reconciliation writes restart cancellations under this command |
| 753 | // prefix, which the live terminal-run listener skips. |
| 754 | // Collects the SQL text of every statement run under the returned tracer. |
| 755 | const traceSqlStatements = () => { |
| 756 | const statements: Array<string> = []; |
| 757 | const tracer = Tracer.make({ |
| 758 | span(options) { |
| 759 | const span = new Tracer.NativeSpan(options); |
| 760 | const end = span.end.bind(span); |
| 761 | span.end = (endTime, exit) => { |
| 762 | end(endTime, exit); |
| 763 | const query = span.attributes.get("db.query.text"); |
| 764 | if (typeof query === "string") statements.push(query); |
| 765 | }; |
| 766 | return span; |
| 767 | }, |
| 768 | }); |
| 769 | return { statements, tracer }; |
| 770 | }; |
| 771 | |
| 772 | const reconcileCommandId = (name: string) => CommandId.make(`command:runtime-reconcile:${name}`); |
| 773 | |
| 774 | const parentProviderThreadId = (threadId: ThreadId) => |
| 775 | ProviderThreadId.make(`provider-thread:${String(threadId).replace("thread:", "")}`); |
| 776 | |
| 777 | const runEvent = (input: { |
| 778 | readonly threadId: ThreadId; |
| 779 | readonly runId: RunId; |
| 780 | readonly ordinal: number; |
| 781 | readonly status: OrchestrationV2Run["status"]; |
| 782 | readonly now: DateTime.Utc; |
| 783 | readonly providerThreadId?: ProviderThreadId; |
| 784 | readonly userMessageId?: MessageId; |
| 785 | readonly delegatedCompletion?: OrchestrationV2Run["delegatedCompletion"]; |
| 786 | }) => ({ |
| 787 | id: EventId.make(`event:${input.runId}:${input.status}`), |
| 788 | type: "run.updated" as const, |
| 789 | threadId: input.threadId, |
| 790 | runId: input.runId, |
| 791 | providerInstanceId: modelSelection.instanceId, |
| 792 | occurredAt: input.now, |
| 793 | payload: { |
| 794 | id: input.runId, |
| 795 | threadId: input.threadId, |
| 796 | ordinal: input.ordinal, |
| 797 | providerInstanceId: modelSelection.instanceId, |
| 798 | modelSelection, |
| 799 | providerThreadId: input.providerThreadId ?? null, |
| 800 | userMessageId: input.userMessageId ?? MessageId.make(`message:${input.runId}`), |
| 801 | rootNodeId: null, |
| 802 | activeAttemptId: null, |
| 803 | status: input.status, |
| 804 | requestedAt: input.now, |
| 805 | startedAt: input.now, |
| 806 | completedAt: input.status === "running" ? null : input.now, |
| 807 | checkpointId: null, |
| 808 | contextHandoffId: null, |
| 809 | ...(input.delegatedCompletion === undefined |
| 810 | ? {} |
| 811 | : { delegatedCompletion: input.delegatedCompletion }), |
| 812 | }, |
| 813 | }); |
| 814 | |
| 815 | /** A running app-owned task whose child thread's first run the restart cancelled. */ |
| 816 | const seedRestartCancelledChild = (input: { |
| 817 | readonly parentThreadId: ThreadId; |
| 818 | readonly projectId: ProjectId; |
| 819 | readonly parentRunId: RunId; |
| 820 | readonly rootNodeId: NodeId; |
| 821 | readonly name: string; |
| 822 | readonly completionWake: "always" | "settled_only"; |
| 823 | readonly continuationPending: boolean; |
| 824 | /** "completed" seeds a settled turn whose background work the restart cancelled. */ |
| 825 | readonly runStatus?: "cancelled" | "completed"; |
| 826 | /** Background work the restart cancelled on the child's run. */ |
| 827 | readonly lostWork?: OrchestrationV2Run["restartCancelledBackgroundWork"]; |
| 828 | readonly now: DateTime.Utc; |
| 829 | }) => |
| 830 | Effect.gen(function* () { |
| 831 | const eventSink = yield* EventSink.EventSinkV2; |
| 832 | const taskId = NodeId.make(`node:${input.name}`); |
| 833 | const childThreadId = ThreadId.make(`thread:${input.name}`); |
| 834 | const childRunId = RunId.make(`run:${input.name}:1`); |
| 835 | yield* eventSink.write({ |
| 836 | commandId: CommandId.make(`command:seed-child:${input.name}`), |
| 837 | events: [ |
| 838 | { |
| 839 | id: EventId.make(`event:${input.name}:thread`), |
| 840 | type: "thread.created", |
| 841 | threadId: childThreadId, |
| 842 | occurredAt: input.now, |
| 843 | payload: { |
| 844 | createdBy: "agent", |
| 845 | creationSource: "server", |
| 846 | id: childThreadId, |
| 847 | projectId: input.projectId, |
| 848 | title: input.name, |
| 849 | providerInstanceId: modelSelection.instanceId, |
| 850 | modelSelection, |
| 851 | runtimeMode: "full-access", |
| 852 | interactionMode: "default", |
| 853 | branch: null, |
| 854 | worktreePath: null, |
| 855 | activeProviderThreadId: null, |
| 856 | lineage: { |
| 857 | parentThreadId: input.parentThreadId, |
| 858 | relationshipToParent: "subagent", |
| 859 | rootThreadId: input.parentThreadId, |
| 860 | }, |
| 861 | forkedFrom: { type: "node", nodeId: taskId }, |
| 862 | createdAt: input.now, |
| 863 | updatedAt: input.now, |
| 864 | archivedAt: null, |
| 865 | settledOverride: null, |
| 866 | settledAt: null, |
| 867 | lastVisitedAt: null, |
| 868 | deletedAt: null, |
| 869 | }, |
| 870 | }, |
| 871 | { |
| 872 | id: EventId.make(`event:${input.name}:task`), |
| 873 | type: "subagent.updated", |
| 874 | threadId: input.parentThreadId, |
| 875 | runId: input.parentRunId, |
| 876 | nodeId: taskId, |
| 877 | driver, |
| 878 | providerInstanceId: modelSelection.instanceId, |
| 879 | occurredAt: input.now, |
| 880 | payload: { |
| 881 | id: taskId, |
| 882 | threadId: input.parentThreadId, |
| 883 | runId: input.parentRunId, |
| 884 | parentNodeId: input.rootNodeId, |
| 885 | origin: "app_owned", |
| 886 | createdBy: "agent", |
| 887 | driver, |
| 888 | providerInstanceId: modelSelection.instanceId, |
| 889 | providerThreadId: null, |
| 890 | childThreadId, |
| 891 | nativeTaskRef: null, |
| 892 | prompt: `Run ${input.name}.`, |
| 893 | title: null, |
| 894 | model: null, |
| 895 | completionWake: input.completionWake, |
| 896 | status: "running", |
| 897 | result: null, |
| 898 | startedAt: input.now, |
| 899 | completedAt: null, |
| 900 | updatedAt: input.now, |
| 901 | }, |
| 902 | }, |
| 903 | ], |
| 904 | }); |
| 905 | const childRun = runEvent({ |
| 906 | threadId: childThreadId, |
| 907 | runId: childRunId, |
| 908 | ordinal: 1, |
| 909 | status: input.runStatus ?? "cancelled", |
| 910 | now: input.now, |
| 911 | }); |
| 912 | yield* eventSink.writeWithEffects({ |
| 913 | commandId: reconcileCommandId(input.name), |
| 914 | events: [ |
| 915 | input.lostWork === undefined |
| 916 | ? childRun |
| 917 | : { |
| 918 | ...childRun, |
| 919 | payload: { ...childRun.payload, restartCancelledBackgroundWork: input.lostWork }, |
| 920 | }, |
| 921 | ], |
| 922 | effects: input.continuationPending |
| 923 | ? [ |
| 924 | { |
| 925 | id: `effect:restart-continuation:${childRunId}`, |
| 926 | commandId: reconcileCommandId(input.name), |
| 927 | threadId: childThreadId, |
| 928 | request: { type: "provider-runtime.continue", sourceRunId: childRunId }, |
| 929 | }, |
| 930 | ] |
| 931 | : [], |
| 932 | }); |
| 933 | return { taskId, childThreadId, childRunId }; |
| 934 | }); |
| 935 | |
| 936 | it.layer(Layer.merge(layerTest, ProviderContinuationRequests.layer))( |
| 937 | "subagent follow-ups", |
| 938 | (it) => { |
| 939 | const outcomes = [ |
| 940 | "completed", |
| 941 | "failed", |
| 942 | "interrupted", |
| 943 | "restarted", |
| 944 | "stopped", |
| 945 | "overlapping", |
| 946 | ] as const; |
| 947 | it.effect.each(outcomes)( |
| 948 | "delivers a %s follow-up independently of the original task", |
| 949 | (outcome) => |
| 950 | Effect.gen(function* () { |
| 951 | const orchestrator = yield* Orchestrator.OrchestratorV2; |
| 952 | const eventSink = yield* EventSink.EventSinkV2; |
| 953 | const requests = yield* ProviderContinuationRequests.ProviderContinuationRequests; |
| 954 | const now = yield* DateTime.now; |
| 955 | const parentThreadId = ThreadId.make(`thread:followup-${outcome}`); |
| 956 | const projectId = ProjectId.make(`project:followup-${outcome}`); |
| 957 | const parentRunId = RunId.make(`run:followup-${outcome}`); |
| 958 | const rootNodeId = NodeId.make(`node:followup-${outcome}`); |
| 959 | yield* seedParentWithTerminalTask({ |
| 960 | threadId: parentThreadId, |
| 961 | projectId, |
| 962 | runId: parentRunId, |
| 963 | rootNodeId, |
| 964 | taskId: NodeId.make(`node:seed-${outcome}`), |
| 965 | deliveryState: "acknowledged", |
| 966 | now, |
| 967 | }); |
| 968 | const child = yield* seedRestartCancelledChild({ |
| 969 | parentThreadId, |
| 970 | projectId, |
| 971 | parentRunId, |
| 972 | rootNodeId, |
| 973 | name: `followup-child-${outcome}`, |
| 974 | completionWake: "always", |
| 975 | continuationPending: false, |
| 976 | runStatus: "completed", |
| 977 | now, |
| 978 | }); |
| 979 | yield* orchestrator.recoverDelegatedTasks; |
| 980 | let firstWake = yield* requests.take; |
| 981 | while (firstWake.threadId !== parentThreadId) firstWake = yield* requests.take; |
| 982 | assert.equal(firstWake.delegatedCompletion?.parentRunId, parentRunId); |
| 983 | yield* orchestrator.dispatch({ |
| 984 | type: "delegated_task.completion-delivery.acknowledge", |
| 985 | commandId: CommandId.make(`command:followup-ack-${outcome}`), |
| 986 | parentThreadId, |
| 987 | taskId: child.taskId, |
| 988 | observedByRunId: parentRunId, |
| 989 | }); |
| 990 | const original = (yield* orchestrator.getThreadProjection(parentThreadId)).subagents.find( |
| 991 | (task) => task.id === child.taskId, |
| 992 | )!; |
| 993 | const parentRun = (yield* orchestrator.getThreadProjection(parentThreadId)).runs.find( |
| 994 | (run) => run.id === parentRunId, |
| 995 | )!; |
| 996 | yield* eventSink.write({ |
| 997 | commandId: reconcileCommandId(`followup-parent-finished-${outcome}`), |
| 998 | events: [ |
| 999 | { |
| 1000 | id: EventId.make(`event:followup-parent-finished-${outcome}`), |
| 1001 | type: "run.updated", |
| 1002 | threadId: parentThreadId, |
| 1003 | runId: parentRunId, |
| 1004 | occurredAt: now, |
| 1005 | payload: { ...parentRun, status: "completed", completedAt: now }, |
| 1006 | }, |
| 1007 | ], |
| 1008 | }); |
| 1009 | const messageId = MessageId.make(`message:followup-${outcome}`); |
| 1010 | yield* orchestrator.dispatch({ |
| 1011 | type: "message.dispatch", |
| 1012 | createdBy: outcome === "failed" ? "user" : "agent", |
| 1013 | creationSource: outcome === "failed" ? "web" : "mcp", |
| 1014 | commandId: CommandId.make(`command:followup-send-${outcome}`), |
| 1015 | threadId: child.childThreadId, |
| 1016 | senderThreadId: parentThreadId, |
| 1017 | messageId, |
| 1018 | text: "Check the same work again.", |
| 1019 | attachments: [], |
| 1020 | dispatchMode: { type: "start_immediately" }, |
| 1021 | }); |
| 1022 | const childProjection = yield* orchestrator.getThreadProjection(child.childThreadId); |
| 1023 | let run = childProjection.runs.find((row) => row.userMessageId === messageId)!; |
| 1024 | assert.isDefined(run.delegatedTaskId); |
| 1025 | const workingParent = yield* orchestrator.getShellSnapshot(); |
| 1026 | const backgroundTask = workingParent.threads |
| 1027 | .find((thread) => thread.id === parentThreadId) |
| 1028 | ?.pendingBackgroundTasks?.find((task) => task.kind === "subagent"); |
| 1029 | assert.equal(backgroundTask?.childThreadId, child.childThreadId); |
| 1030 | if (outcome === "overlapping") { |
| 1031 | yield* orchestrator.dispatch({ |
| 1032 | type: "message.dispatch", |
| 1033 | createdBy: "agent", |
| 1034 | creationSource: "mcp", |
| 1035 | commandId: CommandId.make("command:overlapping-followup"), |
| 1036 | threadId: child.childThreadId, |
| 1037 | senderThreadId: parentThreadId, |
| 1038 | messageId: MessageId.make("message:overlapping-followup"), |
| 1039 | text: "Keep working on another follow-up.", |
| 1040 | attachments: [], |
| 1041 | dispatchMode: { type: "queue_after_active" }, |
| 1042 | }); |
| 1043 | } |
| 1044 | if (outcome === "restarted") { |
| 1045 | const cut = { ...run, status: "cancelled" as const, startedAt: now, completedAt: now }; |
| 1046 | yield* eventSink.writeWithEffects({ |
| 1047 | commandId: reconcileCommandId("followup-restart-cut"), |
| 1048 | events: [ |
| 1049 | { |
| 1050 | id: EventId.make("event:followup-restart-cut"), |
| 1051 | type: "run.updated", |
| 1052 | threadId: child.childThreadId, |
| 1053 | runId: run.id, |
| 1054 | occurredAt: now, |
| 1055 | payload: cut, |
| 1056 | }, |
| 1057 | ], |
| 1058 | effects: [ |
| 1059 | { |
| 1060 | id: `effect:restart-continuation:${run.id}`, |
| 1061 | commandId: reconcileCommandId("followup-restart-cut"), |
| 1062 | threadId: child.childThreadId, |
| 1063 | request: { type: "provider-runtime.continue", sourceRunId: run.id }, |
| 1064 | }, |
| 1065 | ], |
| 1066 | }); |
| 1067 | yield* orchestrator.recoverDelegatedTasks; |
| 1068 | assert.isNull( |
| 1069 | (yield* orchestrator.getThreadProjection(parentThreadId)).subagents.find( |
| 1070 | (task) => task.id === run.delegatedTaskId, |
| 1071 | )!.result, |
| 1072 | ); |
| 1073 | const continuationMessageId = MessageId.make("message:followup-restart-continuation"); |
| 1074 | yield* orchestrator.dispatch({ |
| 1075 | type: "message.dispatch", |
| 1076 | createdBy: "agent", |
| 1077 | creationSource: "server", |
| 1078 | commandId: CommandId.make("command:followup-restart-continuation"), |
| 1079 | threadId: child.childThreadId, |
| 1080 | restartContinuationOfRunId: run.id, |
| 1081 | messageId: continuationMessageId, |
| 1082 | text: "Continue where you left off.", |
| 1083 | attachments: [], |
| 1084 | dispatchMode: { type: "start_immediately" }, |
| 1085 | }); |
| 1086 | const continuation = (yield* orchestrator.getThreadProjection( |
| 1087 | child.childThreadId, |
| 1088 | )).runs.find((row) => row.userMessageId === continuationMessageId)!; |
| 1089 | assert.equal(continuation.delegatedTaskId, run.delegatedTaskId); |
| 1090 | run = continuation; |
| 1091 | } |
| 1092 | if (outcome === "stopped") { |
| 1093 | yield* orchestrator.dispatch({ |
| 1094 | type: "thread.stop", |
| 1095 | threadId: parentThreadId, |
| 1096 | commandId: CommandId.make(`command:followup-stop-${outcome}`), |
| 1097 | }); |
| 1098 | } |
| 1099 | // Reconciliation skips the live terminal listener; recovery must finish this recorded task. |
| 1100 | yield* eventSink.write({ |
| 1101 | commandId: reconcileCommandId(`followup-finish-${outcome}`), |
| 1102 | events: [ |
| 1103 | { |
| 1104 | id: EventId.make(`event:followup-finish-${outcome}`), |
| 1105 | type: "run.updated", |
| 1106 | threadId: child.childThreadId, |
| 1107 | runId: run.id, |
| 1108 | occurredAt: now, |
| 1109 | payload: { |
| 1110 | ...run, |
| 1111 | status: |
| 1112 | outcome === "stopped" |
| 1113 | ? "interrupted" |
| 1114 | : outcome === "restarted" || outcome === "overlapping" |
| 1115 | ? "completed" |
| 1116 | : outcome, |
| 1117 | startedAt: now, |
| 1118 | completedAt: now, |
| 1119 | }, |
| 1120 | }, |
| 1121 | ], |
| 1122 | }); |
| 1123 | yield* orchestrator.recoverDelegatedTasks; |
| 1124 | const finished = yield* orchestrator.getThreadProjection(parentThreadId); |
| 1125 | assert.deepEqual( |
| 1126 | finished.subagents.find((task) => task.id === child.taskId), |
| 1127 | original, |
| 1128 | ); |
| 1129 | const followup = finished.subagents.find((task) => task.id === run.delegatedTaskId)!; |
| 1130 | if (outcome === "overlapping") { |
| 1131 | const later = finished.subagents.find( |
| 1132 | (task) => |
| 1133 | task.childThreadId === child.childThreadId && |
| 1134 | task.id !== child.taskId && |
| 1135 | task.id !== followup.id, |
| 1136 | )!; |
| 1137 | assert.equal(later.status, "pending"); |
| 1138 | assert.isNull(later.result); |
| 1139 | } |
| 1140 | assert.equal( |
| 1141 | followup.status, |
| 1142 | outcome === "stopped" |
| 1143 | ? "interrupted" |
| 1144 | : outcome === "restarted" || outcome === "overlapping" |
| 1145 | ? "completed" |
| 1146 | : outcome, |
| 1147 | ); |
| 1148 | assert.equal( |
| 1149 | followup.completionDelivery?.state, |
| 1150 | outcome === "stopped" ? "disposed" : "claimed", |
| 1151 | ); |
| 1152 | if (outcome !== "stopped") { |
| 1153 | let wake = yield* requests.take; |
| 1154 | while ( |
| 1155 | wake.threadId !== parentThreadId || |
| 1156 | wake.delegatedCompletion?.messageId === firstWake.delegatedCompletion?.messageId |
| 1157 | ) { |
| 1158 | wake = yield* requests.take; |
| 1159 | } |
| 1160 | assert.equal(wake.delegatedCompletion?.parentRunId, parentRunId); |
| 1161 | assert.notEqual( |
| 1162 | wake.delegatedCompletion?.messageId, |
| 1163 | firstWake.delegatedCompletion?.messageId, |
| 1164 | ); |
| 1165 | } |
| 1166 | assert.equal( |
| 1167 | finished.contextTransfers.filter( |
| 1168 | (transfer) => |
| 1169 | transfer.type === "subagent_result" && |
| 1170 | transfer.sourceThreadId === child.childThreadId, |
| 1171 | ).length, |
| 1172 | 2, |
| 1173 | ); |
| 1174 | yield* orchestrator.recoverDelegatedTasks; |
| 1175 | const recovered = yield* orchestrator.getThreadProjection(parentThreadId); |
| 1176 | assert.equal(recovered.contextTransfers.length, finished.contextTransfers.length); |
| 1177 | if (outcome === "completed") { |
| 1178 | const automaticMessageId = MessageId.make("message:followup-automatic-wake"); |
| 1179 | yield* orchestrator.dispatch({ |
| 1180 | type: "message.dispatch", |
| 1181 | createdBy: "agent", |
| 1182 | creationSource: "server", |
| 1183 | commandId: CommandId.make("command:followup-automatic-wake"), |
| 1184 | threadId: child.childThreadId, |
| 1185 | messageId: automaticMessageId, |
| 1186 | text: "A monitor reported new output.", |
| 1187 | attachments: [], |
| 1188 | notification: { |
| 1189 | source: { kind: "monitor" }, |
| 1190 | outcome: "updated", |
| 1191 | summary: "Monitor updated", |
| 1192 | }, |
| 1193 | dispatchMode: { type: "queue_after_active" }, |
| 1194 | }); |
| 1195 | const automatic = (yield* orchestrator.getThreadProjection( |
| 1196 | child.childThreadId, |
| 1197 | )).runs.find((run) => run.userMessageId === automaticMessageId)!; |
| 1198 | assert.isUndefined(automatic.delegatedTaskId); |
| 1199 | yield* eventSink.write({ |
| 1200 | commandId: reconcileCommandId("followup-automatic-finished"), |
| 1201 | events: [ |
| 1202 | { |
| 1203 | id: EventId.make("event:followup-automatic-finished"), |
| 1204 | type: "run.updated", |
| 1205 | threadId: child.childThreadId, |
| 1206 | runId: automatic.id, |
| 1207 | occurredAt: now, |
| 1208 | payload: { ...automatic, status: "completed", startedAt: now, completedAt: now }, |
| 1209 | }, |
| 1210 | ], |
| 1211 | }); |
| 1212 | yield* orchestrator.recoverDelegatedTasks; |
| 1213 | const afterAutomatic = yield* orchestrator.getThreadProjection(parentThreadId); |
| 1214 | assert.deepEqual(afterAutomatic.subagents, recovered.subagents); |
| 1215 | assert.equal(afterAutomatic.contextTransfers.length, recovered.contextTransfers.length); |
| 1216 | const control = yield* seedRestartCancelledChild({ |
| 1217 | parentThreadId, |
| 1218 | projectId, |
| 1219 | parentRunId, |
| 1220 | rootNodeId, |
| 1221 | name: "followup-control-child", |
| 1222 | completionWake: "always", |
| 1223 | continuationPending: false, |
| 1224 | runStatus: "completed", |
| 1225 | now, |
| 1226 | }); |
| 1227 | yield* orchestrator.recoverDelegatedTasks; |
| 1228 | const afterControl = yield* orchestrator.getThreadProjection(parentThreadId); |
| 1229 | assert.equal( |
| 1230 | afterControl.subagents.find((task) => task.id === control.taskId)?.completionDelivery |
| 1231 | ?.state, |
| 1232 | "claimed", |
| 1233 | ); |
| 1234 | assert.equal( |
| 1235 | afterControl.contextTransfers.filter( |
| 1236 | (transfer) => |
| 1237 | transfer.type === "subagent_result" && |
| 1238 | transfer.sourceThreadId === control.childThreadId, |
| 1239 | ).length, |
| 1240 | 1, |
| 1241 | ); |
| 1242 | } |
| 1243 | }), |
| 1244 | ); |
| 1245 | }, |
| 1246 | ); |
| 1247 | |
| 1248 | it.layer(layerTest)("delegated tasks across a server restart", (it) => { |
| 1249 | it.effect("holds a restart-cancelled child for its continuation's result", () => |
| 1250 | Effect.gen(function* () { |
| 1251 | const orchestrator = yield* Orchestrator.OrchestratorV2; |
| 1252 | const eventSink = yield* EventSink.EventSinkV2; |
| 1253 | const now = yield* DateTime.now; |
| 1254 | const threadId = ThreadId.make("thread:restart-parent"); |
| 1255 | const projectId = ProjectId.make("project:restart-parent"); |
| 1256 | const runId = RunId.make("run:restart-parent"); |
| 1257 | const rootNodeId = NodeId.make("node:restart-parent-root"); |
| 1258 | yield* seedParentWithTerminalTask({ |
| 1259 | threadId, |
| 1260 | projectId, |
| 1261 | runId, |
| 1262 | rootNodeId, |
| 1263 | taskId: NodeId.make("node:restart-parent-settled"), |
| 1264 | deliveryState: "delivered", |
| 1265 | now, |
| 1266 | }); |
| 1267 | const child = ( |
| 1268 | name: string, |
| 1269 | continuationPending: boolean, |
| 1270 | runStatus?: "cancelled" | "completed", |
| 1271 | ) => |
| 1272 | seedRestartCancelledChild({ |
| 1273 | parentThreadId: threadId, |
| 1274 | projectId, |
| 1275 | parentRunId: runId, |
| 1276 | rootNodeId, |
| 1277 | name, |
| 1278 | completionWake: "always", |
| 1279 | continuationPending, |
| 1280 | ...(runStatus === undefined ? {} : { runStatus }), |
| 1281 | now, |
| 1282 | }); |
| 1283 | const resumed = yield* child("restart-resumed-child", true); |
| 1284 | const stopped = yield* child("restart-stopped-child", false); |
| 1285 | // Settled with only background work left: its interim reply is not the result. |
| 1286 | const backgrounded = yield* child("restart-backgrounded-child", true, "completed"); |
| 1287 | // A second restart cut the first continuation before it started. |
| 1288 | const recut = yield* child("restart-recut-child", false); |
| 1289 | const recutContinuationId = RunId.make("run:restart-recut-child:2"); |
| 1290 | const recutCommandId = CommandId.make("command:restart-recut-child:reconcile"); |
| 1291 | const recutRun = runEvent({ |
| 1292 | threadId: recut.childThreadId, |
| 1293 | runId: recutContinuationId, |
| 1294 | ordinal: 2, |
| 1295 | status: "cancelled", |
| 1296 | now, |
| 1297 | }); |
| 1298 | yield* eventSink.writeWithEffects({ |
| 1299 | commandId: recutCommandId, |
| 1300 | events: [ |
| 1301 | { |
| 1302 | ...recutRun, |
| 1303 | payload: { |
| 1304 | ...recutRun.payload, |
| 1305 | startedAt: null, |
| 1306 | restartContinuationOfRunId: recut.childRunId, |
| 1307 | }, |
| 1308 | }, |
| 1309 | ], |
| 1310 | effects: [ |
| 1311 | { |
| 1312 | id: `effect:restart-continuation:${recutContinuationId}`, |
| 1313 | commandId: recutCommandId, |
| 1314 | threadId: recut.childThreadId, |
| 1315 | request: { type: "provider-runtime.continue", sourceRunId: recutContinuationId }, |
| 1316 | }, |
| 1317 | ], |
| 1318 | }); |
| 1319 | |
| 1320 | yield* orchestrator.recoverDelegatedTasks; |
| 1321 | |
| 1322 | const recovered = yield* orchestrator.getThreadProjection(threadId); |
| 1323 | const task = (id: NodeId) => recovered.subagents.find((row) => row.id === id); |
| 1324 | assert.equal(task(stopped.taskId)?.status, "cancelled"); |
| 1325 | assert.equal(task(stopped.taskId)?.completionDelivery?.state, "claimed"); |
| 1326 | assert.equal(task(resumed.taskId)?.status, "running"); |
| 1327 | assert.isNull(task(resumed.taskId)?.result ?? null); |
| 1328 | assert.equal(task(backgrounded.taskId)?.status, "running"); |
| 1329 | assert.isNull(task(backgrounded.taskId)?.result ?? null); |
| 1330 | assert.equal(task(recut.taskId)?.status, "running"); |
| 1331 | assert.isTrue(yield* orchestrator.delegatedTaskResultPending(recut.childThreadId)); |
| 1332 | assert.isTrue(yield* orchestrator.delegatedTaskResultPending(resumed.childThreadId)); |
| 1333 | assert.isFalse(yield* orchestrator.delegatedTaskResultPending(stopped.childThreadId)); |
| 1334 | // A replayed first continuation settling must not release the second one's hold. |
| 1335 | yield* orchestrator.recoverDelegatedTask(recut.childThreadId, recut.childRunId); |
| 1336 | const replayed = yield* orchestrator.getThreadProjection(threadId); |
| 1337 | assert.equal(replayed.subagents.find((row) => row.id === recut.taskId)?.status, "running"); |
| 1338 | // A caller that read the cancelled run before the child resumed sees it as pending. |
| 1339 | yield* eventSink.write({ |
| 1340 | commandId: CommandId.make("command:restart-stopped-child:resumed"), |
| 1341 | events: [ |
| 1342 | runEvent({ |
| 1343 | threadId: stopped.childThreadId, |
| 1344 | runId: RunId.make("run:restart-stopped-child:2"), |
| 1345 | ordinal: 2, |
| 1346 | status: "running", |
| 1347 | now, |
| 1348 | }), |
| 1349 | ], |
| 1350 | }); |
| 1351 | assert.isTrue(yield* orchestrator.delegatedTaskResultPending(stopped.childThreadId)); |
| 1352 | assert.isFalse( |
| 1353 | recovered.contextTransfers.some( |
| 1354 | (transfer) => transfer.sourceThreadId === resumed.childThreadId, |
| 1355 | ), |
| 1356 | ); |
| 1357 | |
| 1358 | // The continuation's own run finishing settles the task with its result. |
| 1359 | const afterSequence = yield* eventSink.latestSequence(); |
| 1360 | const continuationRunId = RunId.make("run:restart-resumed-child:2"); |
| 1361 | yield* eventSink.write({ |
| 1362 | commandId: CommandId.make("command:restart-resumed-child:completed"), |
| 1363 | events: [ |
| 1364 | { |
| 1365 | id: EventId.make("event:restart-resumed-child:result"), |
| 1366 | type: "message.updated", |
| 1367 | threadId: resumed.childThreadId, |
| 1368 | runId: continuationRunId, |
| 1369 | occurredAt: now, |
| 1370 | payload: { |
| 1371 | id: MessageId.make("message:restart-resumed-child:result"), |
| 1372 | threadId: resumed.childThreadId, |
| 1373 | runId: continuationRunId, |
| 1374 | nodeId: null, |
| 1375 | role: "assistant", |
| 1376 | text: "Finished after the restart.", |
| 1377 | attachments: [], |
| 1378 | streaming: false, |
| 1379 | createdBy: "agent", |
| 1380 | creationSource: "server", |
| 1381 | createdAt: now, |
| 1382 | updatedAt: now, |
| 1383 | }, |
| 1384 | }, |
| 1385 | runEvent({ |
| 1386 | threadId: resumed.childThreadId, |
| 1387 | runId: continuationRunId, |
| 1388 | ordinal: 2, |
| 1389 | status: "completed", |
| 1390 | now, |
| 1391 | }), |
| 1392 | ], |
| 1393 | }); |
| 1394 | const settled = yield* eventSink |
| 1395 | .stream({ afterSequence, eventType: "subagent.updated" }) |
| 1396 | .pipe( |
| 1397 | Stream.filter( |
| 1398 | (stored) => |
| 1399 | stored.event.type === "subagent.updated" && |
| 1400 | stored.event.payload.id === resumed.taskId, |
| 1401 | ), |
| 1402 | Stream.take(1), |
| 1403 | Stream.runHead, |
| 1404 | ); |
| 1405 | assert.isTrue(settled._tag === "Some"); |
| 1406 | const finished = yield* orchestrator.getThreadProjection(threadId); |
| 1407 | const finishedTask = finished.subagents.find((row) => row.id === resumed.taskId); |
| 1408 | assert.equal(finishedTask?.status, "completed"); |
| 1409 | assert.equal(finishedTask?.result, "Finished after the restart."); |
| 1410 | }), |
| 1411 | ); |
| 1412 | |
| 1413 | it.effect("settles a restart-cancelled child whose continuation declines to start", () => |
| 1414 | Effect.gen(function* () { |
| 1415 | const orchestrator = yield* Orchestrator.OrchestratorV2; |
| 1416 | const now = yield* DateTime.now; |
| 1417 | const threadId = ThreadId.make("thread:restart-declined-parent"); |
| 1418 | const projectId = ProjectId.make("project:restart-declined-parent"); |
| 1419 | const runId = RunId.make("run:restart-declined-parent"); |
| 1420 | const rootNodeId = NodeId.make("node:restart-declined-parent-root"); |
| 1421 | yield* seedParentWithTerminalTask({ |
| 1422 | threadId, |
| 1423 | projectId, |
| 1424 | runId, |
| 1425 | rootNodeId, |
| 1426 | taskId: NodeId.make("node:restart-declined-parent-settled"), |
| 1427 | deliveryState: "delivered", |
| 1428 | now, |
| 1429 | }); |
| 1430 | const child = yield* seedRestartCancelledChild({ |
| 1431 | parentThreadId: threadId, |
| 1432 | projectId, |
| 1433 | parentRunId: runId, |
| 1434 | rootNodeId, |
| 1435 | name: "restart-declined-child", |
| 1436 | completionWake: "always", |
| 1437 | continuationPending: true, |
| 1438 | now, |
| 1439 | }); |
| 1440 | yield* orchestrator.recoverDelegatedTasks; |
| 1441 | const held = yield* orchestrator.getThreadProjection(threadId); |
| 1442 | assert.equal(held.subagents.find((row) => row.id === child.taskId)?.status, "running"); |
| 1443 | |
| 1444 | yield* continueRestartedRun({ |
| 1445 | threadId: child.childThreadId, |
| 1446 | sourceRunId: child.childRunId, |
| 1447 | }).pipe( |
| 1448 | Effect.provide(ServerSettings.layerTest({ continueThreadsAfterServerUpdate: false })), |
| 1449 | ); |
| 1450 | |
| 1451 | const settled = yield* orchestrator.getThreadProjection(threadId); |
| 1452 | const task = settled.subagents.find((row) => row.id === child.taskId); |
| 1453 | assert.equal(task?.status, "cancelled"); |
| 1454 | assert.equal(task?.completionDelivery?.state, "claimed"); |
| 1455 | }), |
| 1456 | ); |
| 1457 | |
| 1458 | it.effect("reports a child whose monitor the restart cut as interrupted, not completed", () => |
| 1459 | Effect.gen(function* () { |
| 1460 | const orchestrator = yield* Orchestrator.OrchestratorV2; |
| 1461 | const now = yield* DateTime.now; |
| 1462 | const threadId = ThreadId.make("thread:restart-monitor-parent"); |
| 1463 | const projectId = ProjectId.make("project:restart-monitor-parent"); |
| 1464 | const runId = RunId.make("run:restart-monitor-parent"); |
| 1465 | const rootNodeId = NodeId.make("node:restart-monitor-parent-root"); |
| 1466 | yield* seedParentWithTerminalTask({ |
| 1467 | threadId, |
| 1468 | projectId, |
| 1469 | runId, |
| 1470 | rootNodeId, |
| 1471 | taskId: NodeId.make("node:restart-monitor-parent-settled"), |
| 1472 | deliveryState: "delivered", |
| 1473 | now, |
| 1474 | }); |
| 1475 | // The child ended its turn to wait on a gate monitor; the restart cut it. |
| 1476 | const child = yield* seedRestartCancelledChild({ |
| 1477 | parentThreadId: threadId, |
| 1478 | projectId, |
| 1479 | parentRunId: runId, |
| 1480 | rootNodeId, |
| 1481 | name: "restart-monitor-child", |
| 1482 | completionWake: "always", |
| 1483 | continuationPending: true, |
| 1484 | runStatus: "completed", |
| 1485 | lostWork: [{ kind: "monitor", label: "Gate attempt 9" }], |
| 1486 | now, |
| 1487 | }); |
| 1488 | yield* orchestrator.recoverDelegatedTasks; |
| 1489 | const held = yield* orchestrator.getThreadProjection(threadId); |
| 1490 | assert.equal(held.subagents.find((row) => row.id === child.taskId)?.status, "running"); |
| 1491 | |
| 1492 | // Continuation is off, so the cut run is the outcome. It still goes to the parent. |
| 1493 | yield* continueRestartedRun({ |
| 1494 | threadId: child.childThreadId, |
| 1495 | sourceRunId: child.childRunId, |
| 1496 | }).pipe( |
| 1497 | Effect.provide(ServerSettings.layerTest({ continueThreadsAfterServerUpdate: false })), |
| 1498 | ); |
| 1499 | |
| 1500 | const settled = yield* orchestrator.getThreadProjection(threadId); |
| 1501 | const task = settled.subagents.find((row) => row.id === child.taskId); |
| 1502 | assert.equal(task?.status, "interrupted"); |
| 1503 | assert.equal(task?.result, "Child task ended with status interrupted."); |
| 1504 | assert.equal(task?.completionDelivery?.state, "claimed"); |
| 1505 | }), |
| 1506 | ); |
| 1507 | |
| 1508 | // A running provider turn does not prove the provider consumed the delivery |
| 1509 | // (Claude marks the turn running before it reads the prompt), so a cut |
| 1510 | // delivery is offered again even when a continuation resumes that turn. |
| 1511 | it.effect("re-offers a cut delivery even when a continuation resumes its turn", () => |
| 1512 | Effect.gen(function* () { |
| 1513 | const orchestrator = yield* Orchestrator.OrchestratorV2; |
| 1514 | const eventSink = yield* EventSink.EventSinkV2; |
| 1515 | const now = yield* DateTime.now; |
| 1516 | const seed = (name: string, continuationPending: boolean) => |
| 1517 | Effect.gen(function* () { |
| 1518 | const threadId = ThreadId.make(`thread:${name}`); |
| 1519 | const runId = RunId.make(`run:${name}`); |
| 1520 | const taskId = NodeId.make(`node:${name}-task`); |
| 1521 | const messageId = MessageId.make(`message:delegated-delivery:${threadId}`); |
| 1522 | const deliveryRunId = RunId.make(`run:${name}:delivery`); |
| 1523 | yield* seedParentWithTerminalTask({ |
| 1524 | threadId, |
| 1525 | projectId: ProjectId.make(`project:${name}`), |
| 1526 | runId, |
| 1527 | rootNodeId: NodeId.make(`node:${name}-root`), |
| 1528 | taskId, |
| 1529 | deliveryState: "claimed", |
| 1530 | completionWake: "always", |
| 1531 | deliveryTaskIds: [taskId], |
| 1532 | now, |
| 1533 | }); |
| 1534 | yield* eventSink.writeWithEffects({ |
| 1535 | commandId: reconcileCommandId(name), |
| 1536 | events: [ |
| 1537 | { |
| 1538 | id: EventId.make(`event:${name}:delivery-message`), |
| 1539 | type: "message.updated", |
| 1540 | threadId, |
| 1541 | runId: deliveryRunId, |
| 1542 | occurredAt: now, |
| 1543 | payload: { |
| 1544 | id: messageId, |
| 1545 | threadId, |
| 1546 | runId: deliveryRunId, |
| 1547 | nodeId: null, |
| 1548 | role: "user", |
| 1549 | text: `Delegated task ${taskId} reached a terminal state.`, |
| 1550 | attachments: [], |
| 1551 | streaming: false, |
| 1552 | createdBy: "agent", |
| 1553 | creationSource: "server", |
| 1554 | createdAt: now, |
| 1555 | updatedAt: now, |
| 1556 | delegatedCompletion: { parentRunId: runId, generation: 1, taskIds: [taskId] }, |
| 1557 | }, |
| 1558 | }, |
| 1559 | runEvent({ |
| 1560 | threadId, |
| 1561 | runId: deliveryRunId, |
| 1562 | ordinal: 2, |
| 1563 | status: "cancelled", |
| 1564 | now, |
| 1565 | providerThreadId: parentProviderThreadId(threadId), |
| 1566 | userMessageId: messageId, |
| 1567 | }), |
| 1568 | ], |
| 1569 | effects: continuationPending |
| 1570 | ? [ |
| 1571 | { |
| 1572 | id: `effect:restart-continuation:${deliveryRunId}`, |
| 1573 | commandId: reconcileCommandId(name), |
| 1574 | threadId, |
| 1575 | request: { type: "provider-runtime.continue", sourceRunId: deliveryRunId }, |
| 1576 | }, |
| 1577 | ] |
| 1578 | : [], |
| 1579 | }); |
| 1580 | return { threadId, runId, taskId }; |
| 1581 | }); |
| 1582 | const cut = yield* seed("delivery-cut", false); |
| 1583 | const resumed = yield* seed("delivery-resumed", true); |
| 1584 | |
| 1585 | yield* orchestrator.recoverDelegatedTasks; |
| 1586 | |
| 1587 | for (const seeded of [cut, resumed]) { |
| 1588 | const projection = yield* orchestrator.getThreadProjection(seeded.threadId); |
| 1589 | assert.deepEqual( |
| 1590 | projection.subagents.find((row) => row.id === seeded.taskId)?.completionDelivery?.state, |
| 1591 | "claimed", |
| 1592 | ); |
| 1593 | const delivery = projection.runs.find((row) => row.id === seeded.runId)?.delegatedCompletion |
| 1594 | ?.delivery; |
| 1595 | assert.equal(delivery?.generation, 2); |
| 1596 | assert.deepEqual(delivery?.taskIds, [seeded.taskId]); |
| 1597 | } |
| 1598 | }), |
| 1599 | ); |
| 1600 | |
| 1601 | it.effect("settles only the current delivery of a parent with long delivery history", () => |
| 1602 | Effect.gen(function* () { |
| 1603 | const orchestrator = yield* Orchestrator.OrchestratorV2; |
| 1604 | const eventSink = yield* EventSink.EventSinkV2; |
| 1605 | const now = yield* DateTime.now; |
| 1606 | const threadId = ThreadId.make("thread:delivery-history"); |
| 1607 | const runId = RunId.make("run:delivery-history"); |
| 1608 | const taskId = NodeId.make("node:delivery-history-task"); |
| 1609 | const currentMessageId = MessageId.make(`message:delegated-delivery:${threadId}`); |
| 1610 | const currentRunId = RunId.make("run:delivery-history:current"); |
| 1611 | const HISTORY = 40; |
| 1612 | yield* seedParentWithTerminalTask({ |
| 1613 | threadId, |
| 1614 | projectId: ProjectId.make("project:delivery-history"), |
| 1615 | runId, |
| 1616 | rootNodeId: NodeId.make("node:delivery-history-root"), |
| 1617 | taskId, |
| 1618 | deliveryState: "claimed", |
| 1619 | completionWake: "always", |
| 1620 | deliveryTaskIds: [taskId], |
| 1621 | now, |
| 1622 | }); |
| 1623 | const deliveryMessage = (id: MessageId, deliveryRunId: RunId) => ({ |
| 1624 | id: EventId.make(`event:${deliveryRunId}:message`), |
| 1625 | type: "message.updated" as const, |
| 1626 | threadId, |
| 1627 | runId: deliveryRunId, |
| 1628 | occurredAt: now, |
| 1629 | payload: { |
| 1630 | id, |
| 1631 | threadId, |
| 1632 | runId: deliveryRunId, |
| 1633 | nodeId: null, |
| 1634 | role: "user" as const, |
| 1635 | text: `Delegated task ${taskId} reached a terminal state.`, |
| 1636 | attachments: [], |
| 1637 | streaming: false, |
| 1638 | createdBy: "agent" as const, |
| 1639 | creationSource: "server" as const, |
| 1640 | createdAt: now, |
| 1641 | updatedAt: now, |
| 1642 | delegatedCompletion: { parentRunId: runId, generation: 1, taskIds: [taskId] }, |
| 1643 | }, |
| 1644 | }); |
| 1645 | // Earlier deliveries that already settled. Each one still owns a |
| 1646 | // delegated message and a terminal run. |
| 1647 | const history = Array.from({ length: HISTORY }, (_, index) => { |
| 1648 | const historyRunId = RunId.make(`run:delivery-history:${index}`); |
| 1649 | const messageId = MessageId.make(`message:delivery-history:${index}`); |
| 1650 | return [ |
| 1651 | deliveryMessage(messageId, historyRunId), |
| 1652 | runEvent({ |
| 1653 | threadId, |
| 1654 | runId: historyRunId, |
| 1655 | ordinal: index + 2, |
| 1656 | status: "completed", |
| 1657 | now, |
| 1658 | providerThreadId: parentProviderThreadId(threadId), |
| 1659 | userMessageId: messageId, |
| 1660 | }), |
| 1661 | ]; |
| 1662 | }).flat(); |
| 1663 | yield* eventSink.write({ |
| 1664 | commandId: reconcileCommandId("delivery-history"), |
| 1665 | events: [ |
| 1666 | ...history, |
| 1667 | deliveryMessage(currentMessageId, currentRunId), |
| 1668 | runEvent({ |
| 1669 | threadId, |
| 1670 | runId: currentRunId, |
| 1671 | ordinal: HISTORY + 2, |
| 1672 | status: "completed", |
| 1673 | now, |
| 1674 | providerThreadId: parentProviderThreadId(threadId), |
| 1675 | userMessageId: currentMessageId, |
| 1676 | }), |
| 1677 | ], |
| 1678 | }); |
| 1679 | |
| 1680 | const { statements, tracer } = traceSqlStatements(); |
| 1681 | yield* orchestrator.recoverDelegatedTasks.pipe(Effect.withTracer(tracer)); |
| 1682 | |
| 1683 | const projection = yield* orchestrator.getThreadProjection(threadId); |
| 1684 | assert.equal( |
| 1685 | projection.subagents.find((row) => row.id === taskId)?.completionDelivery?.state, |
| 1686 | "delivered", |
| 1687 | ); |
| 1688 | assert.isNull(projection.runs.find((row) => row.id === runId)?.delegatedCompletion?.delivery); |
| 1689 | // Recovery reads the delivery message once, for the current delivery |
| 1690 | // run only. History must not cost one thread reload per past delivery. |
| 1691 | assert.lengthOf( |
| 1692 | statements.filter((statement) => |
| 1693 | statement.includes("FROM orchestration_v2_projection_messages AS message"), |
| 1694 | ), |
| 1695 | 1, |
| 1696 | ); |
| 1697 | |
| 1698 | // A second recovery finds nothing open on this thread and changes nothing. |
| 1699 | yield* orchestrator.recoverDelegatedTasks; |
| 1700 | const again = yield* orchestrator.getThreadProjection(threadId); |
| 1701 | assert.deepEqual(again.runs, projection.runs); |
| 1702 | assert.deepEqual(again.subagents, projection.subagents); |
| 1703 | }), |
| 1704 | ); |
| 1705 | |
| 1706 | it.effect("settles a cancelled child whose held wakes wait behind it", () => |
| 1707 | Effect.gen(function* () { |
| 1708 | const orchestrator = yield* Orchestrator.OrchestratorV2; |
| 1709 | const eventSink = yield* EventSink.EventSinkV2; |
| 1710 | const now = yield* DateTime.now; |
| 1711 | const threadId = ThreadId.make("thread:held-wake-parent"); |
| 1712 | const projectId = ProjectId.make("project:held-wake-parent"); |
| 1713 | const runId = RunId.make("run:held-wake-parent"); |
| 1714 | const rootNodeId = NodeId.make("node:held-wake-parent-root"); |
| 1715 | yield* seedParentWithTerminalTask({ |
| 1716 | threadId, |
| 1717 | projectId, |
| 1718 | runId, |
| 1719 | rootNodeId, |
| 1720 | taskId: NodeId.make("node:held-wake-parent-settled"), |
| 1721 | deliveryState: "delivered", |
| 1722 | now, |
| 1723 | }); |
| 1724 | const child = yield* seedRestartCancelledChild({ |
| 1725 | parentThreadId: threadId, |
| 1726 | projectId, |
| 1727 | parentRunId: runId, |
| 1728 | rootNodeId, |
| 1729 | name: "held-wake-child", |
| 1730 | completionWake: "always", |
| 1731 | continuationPending: false, |
| 1732 | now, |
| 1733 | }); |
| 1734 | // Pull request watch wakes queued while the child ran; the restart held them. |
| 1735 | const held = runEvent({ |
| 1736 | threadId: child.childThreadId, |
| 1737 | runId: RunId.make("run:held-wake-child:2"), |
| 1738 | ordinal: 2, |
| 1739 | status: "queued", |
| 1740 | now, |
| 1741 | }); |
| 1742 | yield* eventSink.write({ |
| 1743 | commandId: CommandId.make("command:held-wake-child:held"), |
| 1744 | events: [ |
| 1745 | { |
| 1746 | ...held, |
| 1747 | payload: { ...held.payload, startedAt: null, completedAt: null, queueHeld: true }, |
| 1748 | }, |
| 1749 | ], |
| 1750 | }); |
| 1751 | |
| 1752 | yield* orchestrator.recoverDelegatedTasks; |
| 1753 | |
| 1754 | const recovered = yield* orchestrator.getThreadProjection(threadId); |
| 1755 | const task = recovered.subagents.find((row) => row.id === child.taskId); |
| 1756 | assert.equal(task?.status, "cancelled"); |
| 1757 | assert.isNotNull(task?.result ?? null); |
| 1758 | assert.isTrue( |
| 1759 | recovered.contextTransfers.some( |
| 1760 | (transfer) => |
| 1761 | transfer.type === "subagent_result" && transfer.sourceThreadId === child.childThreadId, |
| 1762 | ), |
| 1763 | ); |
| 1764 | assert.isFalse(yield* orchestrator.delegatedTaskResultPending(child.childThreadId)); |
| 1765 | // The held wake stays held for the user to resume or discard. |
| 1766 | const childProjection = yield* orchestrator.getThreadProjection(child.childThreadId); |
| 1767 | assert.deepEqual( |
| 1768 | childProjection.runs.map((run) => [run.status, run.queueHeld ?? false]), |
| 1769 | [ |
| 1770 | ["cancelled", false], |
| 1771 | ["queued", true], |
| 1772 | ], |
| 1773 | ); |
| 1774 | }), |
| 1775 | ); |
| 1776 | |
| 1777 | it.effect("settles a child whose provider failure holds its queued wakes", () => |
| 1778 | Effect.gen(function* () { |
| 1779 | const orchestrator = yield* Orchestrator.OrchestratorV2; |
| 1780 | const eventSink = yield* EventSink.EventSinkV2; |
| 1781 | const now = yield* DateTime.now; |
| 1782 | const threadId = ThreadId.make("thread:failed-hold-parent"); |
| 1783 | const projectId = ProjectId.make("project:failed-hold-parent"); |
| 1784 | const runId = RunId.make("run:failed-hold-parent"); |
| 1785 | const rootNodeId = NodeId.make("node:failed-hold-parent-root"); |
| 1786 | yield* seedParentWithTerminalTask({ |
| 1787 | threadId, |
| 1788 | projectId, |
| 1789 | runId, |
| 1790 | rootNodeId, |
| 1791 | taskId: NodeId.make("node:failed-hold-parent-settled"), |
| 1792 | deliveryState: "delivered", |
| 1793 | now, |
| 1794 | }); |
| 1795 | const child = yield* seedRestartCancelledChild({ |
| 1796 | parentThreadId: threadId, |
| 1797 | projectId, |
| 1798 | parentRunId: runId, |
| 1799 | rootNodeId, |
| 1800 | name: "failed-hold-child", |
| 1801 | completionWake: "always", |
| 1802 | continuationPending: false, |
| 1803 | now, |
| 1804 | }); |
| 1805 | // The child is resumed, and a wake queues behind its running turn. |
| 1806 | const activeRun = runEvent({ |
| 1807 | threadId: child.childThreadId, |
| 1808 | runId: RunId.make("run:failed-hold-child:2"), |
| 1809 | ordinal: 2, |
| 1810 | status: "running", |
| 1811 | now, |
| 1812 | }); |
| 1813 | const queued = runEvent({ |
| 1814 | threadId: child.childThreadId, |
| 1815 | runId: RunId.make("run:failed-hold-child:3"), |
| 1816 | ordinal: 3, |
| 1817 | status: "queued", |
| 1818 | now, |
| 1819 | }); |
| 1820 | yield* eventSink.write({ |
| 1821 | commandId: CommandId.make("command:failed-hold-child:resumed"), |
| 1822 | events: [ |
| 1823 | activeRun, |
| 1824 | { ...queued, payload: { ...queued.payload, startedAt: null, completedAt: null } }, |
| 1825 | ], |
| 1826 | }); |
| 1827 | const afterSequence = yield* eventSink.latestSequence(); |
| 1828 | const errorItemId = TurnItemId.make("turn-item:failed-hold-child:error"); |
| 1829 | yield* eventSink.write({ |
| 1830 | commandId: CommandId.make("command:failed-hold-child:failed"), |
| 1831 | events: [ |
| 1832 | { |
| 1833 | id: EventId.make("event:failed-hold-child:error"), |
| 1834 | type: "turn-item.updated", |
| 1835 | threadId: child.childThreadId, |
| 1836 | runId: activeRun.runId, |
| 1837 | providerInstanceId: modelSelection.instanceId, |
| 1838 | occurredAt: now, |
| 1839 | payload: { |
| 1840 | id: errorItemId, |
| 1841 | type: "error", |
| 1842 | threadId: child.childThreadId, |
| 1843 | runId: activeRun.runId, |
| 1844 | nodeId: null, |
| 1845 | providerThreadId: null, |
| 1846 | providerTurnId: null, |
| 1847 | nativeItemRef: null, |
| 1848 | parentItemId: null, |
| 1849 | ordinal: 1, |
| 1850 | status: "failed", |
| 1851 | title: "Provider failure", |
| 1852 | startedAt: now, |
| 1853 | completedAt: now, |
| 1854 | updatedAt: now, |
| 1855 | failure: { |
| 1856 | class: "provider_error", |
| 1857 | message: "Provider failed.", |
| 1858 | code: "provider_failed", |
| 1859 | retryable: null, |
| 1860 | }, |
| 1861 | }, |
| 1862 | }, |
| 1863 | { |
| 1864 | ...activeRun, |
| 1865 | id: EventId.make("event:failed-hold-child:failed"), |
| 1866 | payload: { ...activeRun.payload, status: "failed", completedAt: now }, |
| 1867 | }, |
| 1868 | ], |
| 1869 | }); |
| 1870 | |
| 1871 | // The failure holds the queue, which leaves the failure as the task's result. |
| 1872 | const settled = yield* eventSink |
| 1873 | .stream({ afterSequence, eventType: "subagent.updated" }) |
| 1874 | .pipe( |
| 1875 | Stream.filter( |
| 1876 | (stored) => |
| 1877 | stored.event.type === "subagent.updated" && stored.event.payload.id === child.taskId, |
| 1878 | ), |
| 1879 | Stream.take(1), |
| 1880 | Stream.runHead, |
| 1881 | ); |
| 1882 | assert.isTrue(settled._tag === "Some"); |
| 1883 | const recovered = yield* orchestrator.getThreadProjection(threadId); |
| 1884 | assert.equal(recovered.subagents.find((row) => row.id === child.taskId)?.status, "failed"); |
| 1885 | const childProjection = yield* orchestrator.getThreadProjection(child.childThreadId); |
| 1886 | assert.deepEqual( |
| 1887 | childProjection.runs.map((run) => [run.status, run.queueHeld ?? false]), |
| 1888 | [ |
| 1889 | ["cancelled", false], |
| 1890 | ["failed", false], |
| 1891 | ["queued", true], |
| 1892 | ], |
| 1893 | ); |
| 1894 | }), |
| 1895 | ); |
| 1896 | |
| 1897 | it.effect("wakes the parent for a settled_only task once its spawning turn ended", () => |
| 1898 | Effect.gen(function* () { |
| 1899 | const orchestrator = yield* Orchestrator.OrchestratorV2; |
| 1900 | const eventSink = yield* EventSink.EventSinkV2; |
| 1901 | const now = yield* DateTime.now; |
| 1902 | const threadId = ThreadId.make("thread:settled-only-restart"); |
| 1903 | const projectId = ProjectId.make("project:settled-only-restart"); |
| 1904 | const runId = RunId.make("run:settled-only-restart"); |
| 1905 | const rootNodeId = NodeId.make("node:settled-only-restart-root"); |
| 1906 | yield* seedParentWithTerminalTask({ |
| 1907 | threadId, |
| 1908 | projectId, |
| 1909 | runId, |
| 1910 | rootNodeId, |
| 1911 | taskId: NodeId.make("node:settled-only-restart-settled"), |
| 1912 | deliveryState: "delivered", |
| 1913 | now, |
| 1914 | }); |
| 1915 | // The restart cut the turn blocked in delegate_task(wait); its |
| 1916 | // continuation is a new, live run. |
| 1917 | yield* eventSink.write({ |
| 1918 | commandId: reconcileCommandId("settled-only-restart"), |
| 1919 | events: [ |
| 1920 | runEvent({ |
| 1921 | threadId, |
| 1922 | runId, |
| 1923 | ordinal: 1, |
| 1924 | status: "cancelled", |
| 1925 | now, |
| 1926 | providerThreadId: parentProviderThreadId(threadId), |
| 1927 | userMessageId: MessageId.make(`message:seed-user:${threadId}`), |
| 1928 | delegatedCompletion: { disposition: "open", nextGeneration: 2, delivery: null }, |
| 1929 | }), |
| 1930 | runEvent({ |
| 1931 | threadId, |
| 1932 | runId: RunId.make("run:settled-only-restart:continuation"), |
| 1933 | ordinal: 2, |
| 1934 | status: "running", |
| 1935 | now, |
| 1936 | providerThreadId: parentProviderThreadId(threadId), |
| 1937 | }), |
| 1938 | ], |
| 1939 | }); |
| 1940 | const child = yield* seedRestartCancelledChild({ |
| 1941 | parentThreadId: threadId, |
| 1942 | projectId, |
| 1943 | parentRunId: runId, |
| 1944 | rootNodeId, |
| 1945 | name: "settled-only-restart-child", |
| 1946 | completionWake: "settled_only", |
| 1947 | continuationPending: false, |
| 1948 | now, |
| 1949 | }); |
| 1950 | |
| 1951 | yield* orchestrator.recoverDelegatedTasks; |
| 1952 | |
| 1953 | const projection = yield* orchestrator.getThreadProjection(threadId); |
| 1954 | assert.equal( |
| 1955 | projection.subagents.find((row) => row.id === child.taskId)?.completionDelivery?.state, |
| 1956 | "claimed", |
| 1957 | ); |
| 1958 | assert.deepEqual( |
| 1959 | projection.runs.find((row) => row.id === runId)?.delegatedCompletion?.delivery?.taskIds, |
| 1960 | [child.taskId], |
| 1961 | ); |
| 1962 | }), |
| 1963 | ); |
| 1964 | }); |