| 1 | import { assert, it } from "@effect/vitest"; |
| 2 | import { |
| 3 | CommandId, |
| 4 | NodeId, |
| 5 | ProviderSessionId, |
| 6 | ProviderThreadId, |
| 7 | ProviderTurnId, |
| 8 | RunAttemptId, |
| 9 | RunId, |
| 10 | ThreadId, |
| 11 | } from "@t3tools/contracts"; |
| 12 | import * as Cause from "effect/Cause"; |
| 13 | import * as DateTime from "effect/DateTime"; |
| 14 | import * as Deferred from "effect/Deferred"; |
| 15 | import * as Effect from "effect/Effect"; |
| 16 | import * as Exit from "effect/Exit"; |
| 17 | import * as Layer from "effect/Layer"; |
| 18 | import * as Option from "effect/Option"; |
| 19 | import * as Queue from "effect/Queue"; |
| 20 | import * as Ref from "effect/Ref"; |
| 21 | import * as TestClock from "effect/testing/TestClock"; |
| 22 | |
| 23 | import * as CheckpointRollbackService from "./CheckpointRollbackService.ts"; |
| 24 | import * as EffectOutbox from "./EffectOutbox.ts"; |
| 25 | import * as EffectWorker from "./EffectWorker.ts"; |
| 26 | import * as RunFinalizationService from "./RunFinalizationService.ts"; |
| 27 | import * as ProviderSessionManager from "./ProviderSessionManager.ts"; |
| 28 | import * as ProviderTurnControlService from "./ProviderTurnControlService.ts"; |
| 29 | import * as ProviderTurnStartService from "./ProviderTurnStartService.ts"; |
| 30 | import * as RuntimeRequestService from "./RuntimeRequestService.ts"; |
| 31 | import * as ThreadTitleRegenerationService from "./ThreadTitleRegenerationService.ts"; |
| 32 | import * as ThreadManagementService from "./ThreadManagementService.ts"; |
| 33 | import * as ServerSettings from "../serverSettings.ts"; |
| 34 | |
| 35 | const threadId = ThreadId.make("thread:effect-worker-restart"); |
| 36 | const oldSessionId = ProviderSessionId.make("provider-session:effect-worker-restart:old"); |
| 37 | const replacementSessionId = ProviderSessionId.make( |
| 38 | "provider-session:effect-worker-restart:replacement", |
| 39 | ); |
| 40 | const providerThreadId = ProviderThreadId.make("provider-thread:effect-worker-restart"); |
| 41 | const providerTurnId = ProviderTurnId.make("provider-turn:effect-worker-restart"); |
| 42 | const attemptId = RunAttemptId.make("run-attempt:effect-worker-restart"); |
| 43 | const runId = RunId.make("run:effect-worker-restart"); |
| 44 | |
| 45 | function restartEffect( |
| 46 | now: DateTime.Utc, |
| 47 | sessionTransition: NonNullable< |
| 48 | Extract< |
| 49 | EffectOutbox.OrchestrationEffectV2["request"], |
| 50 | { readonly type: "provider-turn.restart" } |
| 51 | >["sessionTransition"] |
| 52 | >, |
| 53 | ): EffectOutbox.OrchestrationEffectV2 { |
| 54 | const timestamp = DateTime.formatIso(now); |
| 55 | return { |
| 56 | id: `effect:restart:${sessionTransition.type}`, |
| 57 | commandId: CommandId.make(`command:restart:${sessionTransition.type}`), |
| 58 | threadId, |
| 59 | request: { |
| 60 | type: "provider-turn.restart", |
| 61 | providerSessionId: oldSessionId, |
| 62 | providerThreadId, |
| 63 | providerTurnId, |
| 64 | interruptedAttemptId: attemptId, |
| 65 | runId, |
| 66 | sessionTransition, |
| 67 | }, |
| 68 | status: "running", |
| 69 | attemptCount: 1, |
| 70 | availableAt: timestamp, |
| 71 | leaseOwner: "test-worker", |
| 72 | leaseExpiresAt: timestamp, |
| 73 | createdAt: timestamp, |
| 74 | updatedAt: timestamp, |
| 75 | completedAt: null, |
| 76 | lastError: null, |
| 77 | }; |
| 78 | } |
| 79 | |
| 80 | function layerExecutorFor(input: { |
| 81 | readonly events: Ref.Ref<ReadonlyArray<string>>; |
| 82 | readonly failFirstStart?: Ref.Ref<boolean>; |
| 83 | readonly threads?: Partial<ThreadManagementService.ThreadManagementService["Service"]>; |
| 84 | readonly continueAfterRestart?: boolean; |
| 85 | readonly interrupt?: ProviderTurnControlService.ProviderTurnControlServiceV2Shape["interrupt"]; |
| 86 | }) { |
| 87 | const record = (event: string) => Ref.update(input.events, (events) => [...events, event]); |
| 88 | const layerDependencies = Layer.mergeAll( |
| 89 | Layer.succeed( |
| 90 | ProviderTurnControlService.ProviderTurnControlServiceV2, |
| 91 | ProviderTurnControlService.ProviderTurnControlServiceV2.of({ |
| 92 | interrupt: input.interrupt ?? (() => Effect.void), |
| 93 | steer: () => Effect.void, |
| 94 | interruptAndAwaitTerminal: (request) => |
| 95 | record( |
| 96 | request.replacementProviderSessionId === undefined |
| 97 | ? "interrupt" |
| 98 | : `interrupt:${request.replacementProviderSessionId}`, |
| 99 | ), |
| 100 | }), |
| 101 | ), |
| 102 | Layer.succeed( |
| 103 | ProviderSessionManager.ProviderSessionManagerV2, |
| 104 | ProviderSessionManager.ProviderSessionManagerV2.of({ |
| 105 | shutdown: Effect.void, |
| 106 | open: () => Effect.die("unused open"), |
| 107 | get: () => Effect.succeed(Option.none()), |
| 108 | isLive: () => Effect.succeed(false), |
| 109 | close: () => Effect.void, |
| 110 | closeInstance: () => Effect.void, |
| 111 | release: () => record("release"), |
| 112 | detach: () => record("detach"), |
| 113 | }), |
| 114 | ), |
| 115 | Layer.succeed( |
| 116 | ProviderTurnStartService.ProviderTurnStartServiceV2, |
| 117 | ProviderTurnStartService.ProviderTurnStartServiceV2.of({ |
| 118 | start: () => |
| 119 | Effect.gen(function* () { |
| 120 | yield* record("start"); |
| 121 | if ( |
| 122 | input.failFirstStart !== undefined && |
| 123 | (yield* Ref.getAndSet(input.failFirstStart, false)) |
| 124 | ) { |
| 125 | return yield* new ProviderTurnStartService.ProviderTurnStartError({ |
| 126 | runId, |
| 127 | cause: "simulated first start failure", |
| 128 | }); |
| 129 | } |
| 130 | }), |
| 131 | }), |
| 132 | ), |
| 133 | Layer.succeed( |
| 134 | RunFinalizationService.RunFinalizationService, |
| 135 | RunFinalizationService.RunFinalizationService.of({ finalize: () => Effect.void }), |
| 136 | ), |
| 137 | Layer.succeed( |
| 138 | CheckpointRollbackService.CheckpointRollbackServiceV2, |
| 139 | CheckpointRollbackService.CheckpointRollbackServiceV2.of({ execute: () => Effect.void }), |
| 140 | ), |
| 141 | Layer.succeed( |
| 142 | RuntimeRequestService.RuntimeRequestServiceV2, |
| 143 | RuntimeRequestService.RuntimeRequestServiceV2.of({ respond: () => Effect.void }), |
| 144 | ), |
| 145 | Layer.succeed( |
| 146 | ThreadTitleRegenerationService.ThreadTitleRegenerationService, |
| 147 | ThreadTitleRegenerationService.ThreadTitleRegenerationService.of({ |
| 148 | execute: () => Effect.void, |
| 149 | }), |
| 150 | ), |
| 151 | ); |
| 152 | return EffectWorker.layerExecutor.pipe( |
| 153 | Layer.provide( |
| 154 | Layer.mergeAll( |
| 155 | layerDependencies, |
| 156 | Layer.mock(ThreadManagementService.ThreadManagementService)(input.threads ?? {}), |
| 157 | ServerSettings.layerTest( |
| 158 | input.continueAfterRestart === true ? { continueThreadsAfterServerUpdate: true } : {}, |
| 159 | ), |
| 160 | ), |
| 161 | ), |
| 162 | ); |
| 163 | } |
| 164 | |
| 165 | it("does not retry pure interrupt races where the turn is already gone", () => { |
| 166 | assert.isTrue( |
| 167 | EffectWorker.isNonRetryableProviderTurnControlFailure( |
| 168 | "provider-turn.interrupt", |
| 169 | "ProviderAdapterInterruptError: ... ACP provider turn provider-turn:x is not active", |
| 170 | ), |
| 171 | ); |
| 172 | assert.isTrue( |
| 173 | EffectWorker.isNonRetryableProviderTurnControlFailure( |
| 174 | "provider-turn.interrupt", |
| 175 | "Provider session provider-session:x is not active.", |
| 176 | ), |
| 177 | ); |
| 178 | // Restart is compound (interrupt + detach + start). Do not swallow start failures. |
| 179 | assert.isFalse( |
| 180 | EffectWorker.isNonRetryableProviderTurnControlFailure( |
| 181 | "provider-turn.restart", |
| 182 | "Provider session provider-session:x is not active.", |
| 183 | ), |
| 184 | ); |
| 185 | assert.isFalse( |
| 186 | EffectWorker.isNonRetryableProviderTurnControlFailure( |
| 187 | "provider-turn.start", |
| 188 | "Provider session provider-session:x is not active.", |
| 189 | ), |
| 190 | ); |
| 191 | assert.isFalse( |
| 192 | EffectWorker.isNonRetryableProviderTurnControlFailure( |
| 193 | "provider-turn.interrupt", |
| 194 | "ACP hard teardown failed unexpectedly; the session is poisoned", |
| 195 | ), |
| 196 | ); |
| 197 | }); |
| 198 | |
| 199 | it.effect("settles a stopped run when its adapter has already lost the native turn", () => |
| 200 | Effect.gen(function* () { |
| 201 | const now = yield* DateTime.now; |
| 202 | const events = yield* Ref.make<ReadonlyArray<string>>([]); |
| 203 | const layer = layerExecutorFor({ |
| 204 | events, |
| 205 | interrupt: () => |
| 206 | new ProviderTurnControlService.ProviderTurnControlError({ |
| 207 | threadId, |
| 208 | operation: "interrupt", |
| 209 | providerTurnId, |
| 210 | cause: "Provider turn is not active.", |
| 211 | }), |
| 212 | threads: { |
| 213 | dispatch: (command) => |
| 214 | Ref.update(events, (current) => [...current, command.type]).pipe( |
| 215 | Effect.as({ sequence: 1, storedEvents: [] }), |
| 216 | ), |
| 217 | }, |
| 218 | }); |
| 219 | yield* Effect.gen(function* () { |
| 220 | const executor = yield* EffectWorker.OrchestrationEffectExecutorV2; |
| 221 | yield* executor.execute({ |
| 222 | ...restartEffect(now, { type: "detach" }), |
| 223 | request: { |
| 224 | type: "provider-turn.interrupt", |
| 225 | providerSessionId: oldSessionId, |
| 226 | providerThreadId, |
| 227 | providerTurnId, |
| 228 | }, |
| 229 | }); |
| 230 | }).pipe(Effect.provide(layer)); |
| 231 | assert.deepEqual(yield* Ref.get(events), ["thread.background-work.settle"]); |
| 232 | }), |
| 233 | ); |
| 234 | |
| 235 | it.effect("requeues a claim when a pre-execution worker check fails", () => |
| 236 | Effect.gen(function* () { |
| 237 | const now = DateTime.formatIso(yield* DateTime.now); |
| 238 | const effectId = "effect:worker-pre-execution-failure"; |
| 239 | const workerId = "worker-pre-execution-failure"; |
| 240 | const claimedEffect: EffectOutbox.OrchestrationEffectV2 = { |
| 241 | id: effectId, |
| 242 | commandId: CommandId.make("command:worker-pre-execution-failure"), |
| 243 | threadId: ThreadId.make("thread:worker-pre-execution-failure"), |
| 244 | request: { type: "terminal.cleanup" }, |
| 245 | status: "running", |
| 246 | attemptCount: 1, |
| 247 | availableAt: now, |
| 248 | leaseOwner: workerId, |
| 249 | leaseExpiresAt: now, |
| 250 | createdAt: now, |
| 251 | updatedAt: now, |
| 252 | completedAt: null, |
| 253 | lastError: null, |
| 254 | }; |
| 255 | const retries = yield* Ref.make< |
| 256 | ReadonlyArray<{ |
| 257 | readonly effectId: string; |
| 258 | readonly workerId: string; |
| 259 | readonly error: string; |
| 260 | readonly delayMs: number; |
| 261 | }> |
| 262 | >([]); |
| 263 | const executionCount = yield* Ref.make(0); |
| 264 | const layerOutbox = Layer.mock(EffectOutbox.EffectOutboxV2)({ |
| 265 | claimNext: () => Effect.succeed(Option.some(claimedEffect)), |
| 266 | get: () => |
| 267 | Effect.fail( |
| 268 | new EffectOutbox.EffectOutboxError({ |
| 269 | operation: "get", |
| 270 | effectId, |
| 271 | cause: "simulated cancellation-state read failure", |
| 272 | }), |
| 273 | ), |
| 274 | retry: (input) => |
| 275 | Ref.update(retries, (existing) => [...existing, input]).pipe(Effect.as(true)), |
| 276 | }); |
| 277 | const layerExecutor = Layer.succeed( |
| 278 | EffectWorker.OrchestrationEffectExecutorV2, |
| 279 | EffectWorker.OrchestrationEffectExecutorV2.of({ |
| 280 | execute: () => Ref.update(executionCount, (count) => count + 1), |
| 281 | }), |
| 282 | ); |
| 283 | const layerWorker = EffectWorker.layerWithOptions({ workerId }).pipe( |
| 284 | Layer.provide(Layer.merge(layerOutbox, layerExecutor)), |
| 285 | ); |
| 286 | |
| 287 | const exit = yield* EffectWorker.OrchestrationEffectWorkerV2.pipe( |
| 288 | Effect.flatMap((worker) => worker.runOnce), |
| 289 | Effect.provide(layerWorker), |
| 290 | Effect.exit, |
| 291 | ); |
| 292 | |
| 293 | assert.isTrue(Exit.isFailure(exit)); |
| 294 | if (Exit.isFailure(exit)) { |
| 295 | assert.include(Cause.pretty(exit.cause), "simulated cancellation-state read failure"); |
| 296 | } |
| 297 | assert.equal(yield* Ref.get(executionCount), 0); |
| 298 | const retry = (yield* Ref.get(retries))[0]; |
| 299 | assert.isDefined(retry); |
| 300 | assert.equal(retry.effectId, effectId); |
| 301 | assert.equal(retry.workerId, workerId); |
| 302 | assert.equal(retry.delayMs, 0); |
| 303 | assert.include(retry.error, "simulated cancellation-state read failure"); |
| 304 | }), |
| 305 | ); |
| 306 | |
| 307 | it.effect("arms cancellation before the durable pre-execution check", () => |
| 308 | Effect.gen(function* () { |
| 309 | const now = DateTime.formatIso(yield* DateTime.now); |
| 310 | const effectId = "effect:worker-cancellation-registration-race"; |
| 311 | const workerId = "worker-cancellation-registration-race"; |
| 312 | const claimedEffect: EffectOutbox.OrchestrationEffectV2 = { |
| 313 | id: effectId, |
| 314 | commandId: CommandId.make("command:worker-cancellation-registration-race"), |
| 315 | threadId: ThreadId.make("thread:worker-cancellation-registration-race"), |
| 316 | request: { type: "terminal.cleanup" }, |
| 317 | status: "running", |
| 318 | attemptCount: 1, |
| 319 | availableAt: now, |
| 320 | leaseOwner: workerId, |
| 321 | leaseExpiresAt: now, |
| 322 | createdAt: now, |
| 323 | updatedAt: now, |
| 324 | completedAt: null, |
| 325 | lastError: null, |
| 326 | }; |
| 327 | const signal = yield* Deferred.make<void>(); |
| 328 | let cancellationArmed = false; |
| 329 | const executionCount = yield* Ref.make(0); |
| 330 | const settlementCount = yield* Ref.make(0); |
| 331 | const layerOutbox = Layer.mock(EffectOutbox.EffectOutboxV2)({ |
| 332 | claimNext: () => Effect.succeed(Option.some(claimedEffect)), |
| 333 | awaitCancellation: () => { |
| 334 | cancellationArmed = true; |
| 335 | return Deferred.await(signal); |
| 336 | }, |
| 337 | get: () => |
| 338 | Effect.gen(function* () { |
| 339 | // Model a cancellation commit immediately after this durable read |
| 340 | // took its snapshot. Its process-local signal is only delivered when |
| 341 | // the worker registered the waiter before starting the read. |
| 342 | if (cancellationArmed) { |
| 343 | yield* Deferred.succeed(signal, undefined); |
| 344 | } |
| 345 | return Option.some(claimedEffect); |
| 346 | }), |
| 347 | clearCancellation: () => Effect.void, |
| 348 | succeed: () => Ref.update(settlementCount, (count) => count + 1).pipe(Effect.as(true)), |
| 349 | }); |
| 350 | const layerExecutor = Layer.succeed( |
| 351 | EffectWorker.OrchestrationEffectExecutorV2, |
| 352 | EffectWorker.OrchestrationEffectExecutorV2.of({ |
| 353 | execute: () => |
| 354 | Effect.yieldNow.pipe(Effect.andThen(Ref.update(executionCount, (count) => count + 1))), |
| 355 | }), |
| 356 | ); |
| 357 | const layerWorker = EffectWorker.layerWithOptions({ workerId }).pipe( |
| 358 | Layer.provide(Layer.merge(layerOutbox, layerExecutor)), |
| 359 | ); |
| 360 | |
| 361 | const exit = yield* EffectWorker.OrchestrationEffectWorkerV2.pipe( |
| 362 | Effect.flatMap((worker) => worker.runOnce), |
| 363 | Effect.provide(layerWorker), |
| 364 | Effect.exit, |
| 365 | ); |
| 366 | |
| 367 | if (Exit.isFailure(exit)) { |
| 368 | assert.fail(Cause.pretty(exit.cause)); |
| 369 | } |
| 370 | assert.equal(yield* Ref.get(executionCount), 0); |
| 371 | assert.equal(yield* Ref.get(settlementCount), 0); |
| 372 | }), |
| 373 | ); |
| 374 | |
| 375 | it.effect("terminalizes a process-bound claim when success settlement fails", () => |
| 376 | Effect.gen(function* () { |
| 377 | const now = DateTime.formatIso(yield* DateTime.now); |
| 378 | const effectId = "effect:worker-process-bound-settlement-failure"; |
| 379 | const workerId = "worker-process-bound-settlement-failure"; |
| 380 | const claimedEffect: EffectOutbox.OrchestrationEffectV2 = { |
| 381 | id: effectId, |
| 382 | commandId: CommandId.make("command:worker-process-bound-settlement-failure"), |
| 383 | threadId: ThreadId.make("thread:worker-process-bound-settlement-failure"), |
| 384 | request: { |
| 385 | type: "provider-turn.start", |
| 386 | runId: RunId.make("run:worker-process-bound-settlement-failure"), |
| 387 | }, |
| 388 | status: "running", |
| 389 | attemptCount: 1, |
| 390 | availableAt: now, |
| 391 | leaseOwner: workerId, |
| 392 | leaseExpiresAt: now, |
| 393 | createdAt: now, |
| 394 | updatedAt: now, |
| 395 | completedAt: null, |
| 396 | lastError: null, |
| 397 | }; |
| 398 | const retries = yield* Ref.make(0); |
| 399 | const terminalErrors = yield* Ref.make<ReadonlyArray<string>>([]); |
| 400 | const executionCount = yield* Ref.make(0); |
| 401 | const layerOutbox = Layer.mock(EffectOutbox.EffectOutboxV2)({ |
| 402 | claimNext: () => Effect.succeed(Option.some(claimedEffect)), |
| 403 | get: () => Effect.succeed(Option.some(claimedEffect)), |
| 404 | awaitCancellation: () => Effect.never, |
| 405 | clearCancellation: () => Effect.void, |
| 406 | succeed: () => |
| 407 | Effect.fail( |
| 408 | new EffectOutbox.EffectOutboxError({ |
| 409 | operation: "succeed", |
| 410 | effectId, |
| 411 | cause: "simulated success settlement failure", |
| 412 | }), |
| 413 | ), |
| 414 | retry: () => Ref.update(retries, (count) => count + 1).pipe(Effect.as(true)), |
| 415 | fail: ({ error }) => |
| 416 | Ref.update(terminalErrors, (existing) => [...existing, error]).pipe(Effect.as(true)), |
| 417 | }); |
| 418 | const layerExecutor = Layer.succeed( |
| 419 | EffectWorker.OrchestrationEffectExecutorV2, |
| 420 | EffectWorker.OrchestrationEffectExecutorV2.of({ |
| 421 | execute: () => Ref.update(executionCount, (count) => count + 1), |
| 422 | }), |
| 423 | ); |
| 424 | const layerWorker = EffectWorker.layerWithOptions({ workerId }).pipe( |
| 425 | Layer.provide(Layer.merge(layerOutbox, layerExecutor)), |
| 426 | ); |
| 427 | |
| 428 | const exit = yield* EffectWorker.OrchestrationEffectWorkerV2.pipe( |
| 429 | Effect.flatMap((worker) => worker.runOnce), |
| 430 | Effect.provide(layerWorker), |
| 431 | Effect.exit, |
| 432 | ); |
| 433 | |
| 434 | assert.isTrue(Exit.isFailure(exit)); |
| 435 | assert.equal(yield* Ref.get(executionCount), 1); |
| 436 | assert.equal(yield* Ref.get(retries), 0); |
| 437 | const terminalError = (yield* Ref.get(terminalErrors))[0]; |
| 438 | assert.isDefined(terminalError); |
| 439 | assert.include(terminalError, "after execution started"); |
| 440 | assert.include(terminalError, "simulated success settlement failure"); |
| 441 | }), |
| 442 | ); |
| 443 | |
| 444 | it.effect("requeues a replay-safe claim when success settlement fails", () => |
| 445 | Effect.gen(function* () { |
| 446 | const now = DateTime.formatIso(yield* DateTime.now); |
| 447 | const effectId = "effect:worker-replay-safe-settlement-failure"; |
| 448 | const workerId = "worker-replay-safe-settlement-failure"; |
| 449 | const claimedEffect: EffectOutbox.OrchestrationEffectV2 = { |
| 450 | id: effectId, |
| 451 | commandId: CommandId.make("command:worker-replay-safe-settlement-failure"), |
| 452 | threadId: ThreadId.make("thread:worker-replay-safe-settlement-failure"), |
| 453 | request: { type: "terminal.cleanup" }, |
| 454 | status: "running", |
| 455 | attemptCount: 1, |
| 456 | availableAt: now, |
| 457 | leaseOwner: workerId, |
| 458 | leaseExpiresAt: now, |
| 459 | createdAt: now, |
| 460 | updatedAt: now, |
| 461 | completedAt: null, |
| 462 | lastError: null, |
| 463 | }; |
| 464 | const retries = yield* Ref.make(0); |
| 465 | const terminalizations = yield* Ref.make(0); |
| 466 | const layerOutbox = Layer.mock(EffectOutbox.EffectOutboxV2)({ |
| 467 | claimNext: () => Effect.succeed(Option.some(claimedEffect)), |
| 468 | get: () => Effect.succeed(Option.some(claimedEffect)), |
| 469 | awaitCancellation: () => Effect.never, |
| 470 | clearCancellation: () => Effect.void, |
| 471 | succeed: () => |
| 472 | Effect.fail( |
| 473 | new EffectOutbox.EffectOutboxError({ |
| 474 | operation: "succeed", |
| 475 | effectId, |
| 476 | cause: "simulated replay-safe settlement failure", |
| 477 | }), |
| 478 | ), |
| 479 | retry: () => Ref.update(retries, (count) => count + 1).pipe(Effect.as(true)), |
| 480 | fail: () => Ref.update(terminalizations, (count) => count + 1).pipe(Effect.as(true)), |
| 481 | }); |
| 482 | const layerExecutor = Layer.succeed( |
| 483 | EffectWorker.OrchestrationEffectExecutorV2, |
| 484 | EffectWorker.OrchestrationEffectExecutorV2.of({ execute: () => Effect.void }), |
| 485 | ); |
| 486 | const layerWorker = EffectWorker.layerWithOptions({ workerId }).pipe( |
| 487 | Layer.provide(Layer.merge(layerOutbox, layerExecutor)), |
| 488 | ); |
| 489 | |
| 490 | const exit = yield* EffectWorker.OrchestrationEffectWorkerV2.pipe( |
| 491 | Effect.flatMap((worker) => worker.runOnce), |
| 492 | Effect.provide(layerWorker), |
| 493 | Effect.exit, |
| 494 | ); |
| 495 | |
| 496 | assert.isTrue(Exit.isFailure(exit)); |
| 497 | assert.equal(yield* Ref.get(retries), 1); |
| 498 | assert.equal(yield* Ref.get(terminalizations), 0); |
| 499 | }), |
| 500 | ); |
| 501 | |
| 502 | it.effect("keeps a process-bound executor failure retryable when retry settlement fails", () => |
| 503 | Effect.gen(function* () { |
| 504 | const now = DateTime.formatIso(yield* DateTime.now); |
| 505 | const effectId = "effect:worker-process-bound-retry-settlement-failure"; |
| 506 | const workerId = "worker-process-bound-retry-settlement-failure"; |
| 507 | const claimedEffect: EffectOutbox.OrchestrationEffectV2 = { |
| 508 | id: effectId, |
| 509 | commandId: CommandId.make("command:worker-process-bound-retry-settlement-failure"), |
| 510 | threadId: ThreadId.make("thread:worker-process-bound-retry-settlement-failure"), |
| 511 | request: { |
| 512 | type: "provider-turn.start", |
| 513 | runId: RunId.make("run:worker-process-bound-retry-settlement-failure"), |
| 514 | }, |
| 515 | status: "running", |
| 516 | attemptCount: 1, |
| 517 | availableAt: now, |
| 518 | leaseOwner: workerId, |
| 519 | leaseExpiresAt: now, |
| 520 | createdAt: now, |
| 521 | updatedAt: now, |
| 522 | completedAt: null, |
| 523 | lastError: null, |
| 524 | }; |
| 525 | const retryAttempts = yield* Ref.make(0); |
| 526 | const terminalizations = yield* Ref.make(0); |
| 527 | const layerOutbox = Layer.mock(EffectOutbox.EffectOutboxV2)({ |
| 528 | claimNext: () => Effect.succeed(Option.some(claimedEffect)), |
| 529 | get: () => Effect.succeed(Option.some(claimedEffect)), |
| 530 | awaitCancellation: () => Effect.never, |
| 531 | clearCancellation: () => Effect.void, |
| 532 | retry: () => |
| 533 | Ref.updateAndGet(retryAttempts, (count) => count + 1).pipe( |
| 534 | Effect.flatMap((attempt) => |
| 535 | attempt === 1 |
| 536 | ? Effect.fail( |
| 537 | new EffectOutbox.EffectOutboxError({ |
| 538 | operation: "retry", |
| 539 | effectId, |
| 540 | cause: "simulated retry settlement failure", |
| 541 | }), |
| 542 | ) |
| 543 | : Effect.succeed(true), |
| 544 | ), |
| 545 | ), |
| 546 | fail: () => Ref.update(terminalizations, (count) => count + 1).pipe(Effect.as(true)), |
| 547 | }); |
| 548 | const layerExecutor = Layer.succeed( |
| 549 | EffectWorker.OrchestrationEffectExecutorV2, |
| 550 | EffectWorker.OrchestrationEffectExecutorV2.of({ |
| 551 | execute: () => |
| 552 | Effect.fail( |
| 553 | new EffectWorker.OrchestrationEffectExecutionError({ |
| 554 | effectId, |
| 555 | effectType: claimedEffect.request.type, |
| 556 | cause: "simulated provider execution failure", |
| 557 | }), |
| 558 | ), |
| 559 | }), |
| 560 | ); |
| 561 | const layerWorker = EffectWorker.layerWithOptions({ workerId }).pipe( |
| 562 | Layer.provide(Layer.merge(layerOutbox, layerExecutor)), |
| 563 | ); |
| 564 | |
| 565 | const exit = yield* EffectWorker.OrchestrationEffectWorkerV2.pipe( |
| 566 | Effect.flatMap((worker) => worker.runOnce), |
| 567 | Effect.provide(layerWorker), |
| 568 | Effect.exit, |
| 569 | ); |
| 570 | |
| 571 | assert.isTrue(Exit.isFailure(exit)); |
| 572 | assert.equal(yield* Ref.get(retryAttempts), 2); |
| 573 | assert.equal(yield* Ref.get(terminalizations), 0); |
| 574 | }), |
| 575 | ); |
| 576 | |
| 577 | it.effect("keeps a max-attempt replay-safe failure terminal when fail settlement fails", () => |
| 578 | Effect.gen(function* () { |
| 579 | const now = DateTime.formatIso(yield* DateTime.now); |
| 580 | const effectId = "effect:worker-replay-safe-terminal-settlement-failure"; |
| 581 | const workerId = "worker-replay-safe-terminal-settlement-failure"; |
| 582 | const claimedEffect: EffectOutbox.OrchestrationEffectV2 = { |
| 583 | id: effectId, |
| 584 | commandId: CommandId.make("command:worker-replay-safe-terminal-settlement-failure"), |
| 585 | threadId: ThreadId.make("thread:worker-replay-safe-terminal-settlement-failure"), |
| 586 | request: { type: "terminal.cleanup" }, |
| 587 | status: "running", |
| 588 | attemptCount: 5, |
| 589 | availableAt: now, |
| 590 | leaseOwner: workerId, |
| 591 | leaseExpiresAt: now, |
| 592 | createdAt: now, |
| 593 | updatedAt: now, |
| 594 | completedAt: null, |
| 595 | lastError: null, |
| 596 | }; |
| 597 | const failAttempts = yield* Ref.make(0); |
| 598 | const retries = yield* Ref.make(0); |
| 599 | const layerOutbox = Layer.mock(EffectOutbox.EffectOutboxV2)({ |
| 600 | claimNext: () => Effect.succeed(Option.some(claimedEffect)), |
| 601 | get: () => Effect.succeed(Option.some(claimedEffect)), |
| 602 | awaitCancellation: () => Effect.never, |
| 603 | clearCancellation: () => Effect.void, |
| 604 | fail: () => |
| 605 | Ref.updateAndGet(failAttempts, (count) => count + 1).pipe( |
| 606 | Effect.flatMap((attempt) => |
| 607 | attempt === 1 |
| 608 | ? Effect.fail( |
| 609 | new EffectOutbox.EffectOutboxError({ |
| 610 | operation: "fail", |
| 611 | effectId, |
| 612 | cause: "simulated terminal settlement failure", |
| 613 | }), |
| 614 | ) |
| 615 | : Effect.succeed(true), |
| 616 | ), |
| 617 | ), |
| 618 | retry: () => Ref.update(retries, (count) => count + 1).pipe(Effect.as(true)), |
| 619 | }); |
| 620 | const layerExecutor = Layer.succeed( |
| 621 | EffectWorker.OrchestrationEffectExecutorV2, |
| 622 | EffectWorker.OrchestrationEffectExecutorV2.of({ |
| 623 | execute: () => |
| 624 | Effect.fail( |
| 625 | new EffectWorker.OrchestrationEffectExecutionError({ |
| 626 | effectId, |
| 627 | effectType: claimedEffect.request.type, |
| 628 | cause: "simulated terminal cleanup failure", |
| 629 | }), |
| 630 | ), |
| 631 | }), |
| 632 | ); |
| 633 | const layerWorker = EffectWorker.layerWithOptions({ workerId, maxAttempts: 5 }).pipe( |
| 634 | Layer.provide(Layer.merge(layerOutbox, layerExecutor)), |
| 635 | ); |
| 636 | |
| 637 | const exit = yield* EffectWorker.OrchestrationEffectWorkerV2.pipe( |
| 638 | Effect.flatMap((worker) => worker.runOnce), |
| 639 | Effect.provide(layerWorker), |
| 640 | Effect.exit, |
| 641 | ); |
| 642 | |
| 643 | assert.isTrue(Exit.isFailure(exit)); |
| 644 | assert.equal(yield* Ref.get(failAttempts), 2); |
| 645 | assert.equal(yield* Ref.get(retries), 0); |
| 646 | }), |
| 647 | ); |
| 648 | |
| 649 | it.effect("uses durable deadlines, notifications, and a slow liveness poll", () => |
| 650 | Effect.gen(function* () { |
| 651 | const attempts = yield* Ref.make(0); |
| 652 | const available = yield* Queue.unbounded<void>(); |
| 653 | const now = yield* DateTime.now; |
| 654 | const nextClaimableAt = yield* Ref.make<Option.Option<DateTime.Utc>>( |
| 655 | Option.some(DateTime.add(now, { milliseconds: 100 })), |
| 656 | ); |
| 657 | const worker = EffectWorker.OrchestrationEffectWorkerV2.of({ |
| 658 | awaitWork: Queue.take(available), |
| 659 | runRecoveryOnce: Effect.succeed(false), |
| 660 | runOnce: Effect.gen(function* () { |
| 661 | const count = yield* Ref.updateAndGet(attempts, (current) => current + 1); |
| 662 | if (count === 2) { |
| 663 | yield* Ref.set(nextClaimableAt, Option.some(DateTime.add(now, { milliseconds: 5_000 }))); |
| 664 | } |
| 665 | if (count === 3) yield* Ref.set(nextClaimableAt, Option.none()); |
| 666 | return false; |
| 667 | }), |
| 668 | nextClaimableAt: Ref.get(nextClaimableAt), |
| 669 | drain: () => Effect.succeed(0), |
| 670 | }); |
| 671 | const awaitAttempts = Effect.fnUntraced(function* (expected: number) { |
| 672 | while ((yield* Ref.get(attempts)) < expected) { |
| 673 | yield* Effect.yieldNow; |
| 674 | } |
| 675 | }); |
| 676 | |
| 677 | yield* EffectWorker.runDaemonWithOptions({ |
| 678 | concurrency: 1, |
| 679 | livenessPollIntervalMs: 1_000, |
| 680 | }).pipe( |
| 681 | Effect.provideService(EffectWorker.OrchestrationEffectWorkerV2, worker), |
| 682 | Effect.forkScoped, |
| 683 | ); |
| 684 | |
| 685 | yield* awaitAttempts(1); |
| 686 | yield* TestClock.adjust("99 millis"); |
| 687 | assert.equal(yield* Ref.get(attempts), 1); |
| 688 | |
| 689 | yield* TestClock.adjust("1 millis"); |
| 690 | yield* awaitAttempts(2); |
| 691 | yield* TestClock.adjust("999 millis"); |
| 692 | assert.equal(yield* Ref.get(attempts), 2); |
| 693 | |
| 694 | yield* Queue.offer(available, undefined); |
| 695 | yield* awaitAttempts(3); |
| 696 | yield* TestClock.adjust("999 millis"); |
| 697 | assert.equal(yield* Ref.get(attempts), 3); |
| 698 | |
| 699 | yield* TestClock.adjust("1 millis"); |
| 700 | yield* awaitAttempts(4); |
| 701 | }).pipe(Effect.provide(TestClock.layer())), |
| 702 | ); |
| 703 | |
| 704 | it.effect("does not hot-loop when a claim fails", () => |
| 705 | Effect.gen(function* () { |
| 706 | const attempts = yield* Ref.make(0); |
| 707 | const now = yield* DateTime.now; |
| 708 | const worker = EffectWorker.OrchestrationEffectWorkerV2.of({ |
| 709 | awaitWork: Effect.never, |
| 710 | runRecoveryOnce: Effect.succeed(false), |
| 711 | runOnce: Ref.update(attempts, (count) => count + 1).pipe( |
| 712 | Effect.andThen( |
| 713 | new EffectWorker.OrchestrationEffectWorkerError({ |
| 714 | operation: "claim", |
| 715 | cause: "simulated database failure", |
| 716 | }), |
| 717 | ), |
| 718 | ), |
| 719 | nextClaimableAt: Effect.succeed(Option.some(now)), |
| 720 | drain: () => Effect.succeed(0), |
| 721 | }); |
| 722 | |
| 723 | yield* EffectWorker.runDaemonWithOptions({ |
| 724 | concurrency: 1, |
| 725 | livenessPollIntervalMs: 1_000, |
| 726 | }).pipe( |
| 727 | Effect.provideService(EffectWorker.OrchestrationEffectWorkerV2, worker), |
| 728 | Effect.forkScoped, |
| 729 | ); |
| 730 | |
| 731 | while ((yield* Ref.get(attempts)) < 1) yield* Effect.yieldNow; |
| 732 | yield* TestClock.adjust("999 millis"); |
| 733 | assert.equal(yield* Ref.get(attempts), 1); |
| 734 | yield* TestClock.adjust("1 millis"); |
| 735 | while ((yield* Ref.get(attempts)) < 2) yield* Effect.yieldNow; |
| 736 | }).pipe(Effect.provide(TestClock.layer())), |
| 737 | ); |
| 738 | |
| 739 | it.effect("backs off briefly when a due deadline loses a claim race", () => |
| 740 | Effect.gen(function* () { |
| 741 | const attempts = yield* Ref.make(0); |
| 742 | const now = yield* DateTime.now; |
| 743 | const worker = EffectWorker.OrchestrationEffectWorkerV2.of({ |
| 744 | awaitWork: Effect.never, |
| 745 | runRecoveryOnce: Effect.succeed(false), |
| 746 | runOnce: Ref.update(attempts, (count) => count + 1).pipe(Effect.as(false)), |
| 747 | nextClaimableAt: Effect.succeed(Option.some(now)), |
| 748 | drain: () => Effect.succeed(0), |
| 749 | }); |
| 750 | |
| 751 | yield* EffectWorker.runDaemonWithOptions({ |
| 752 | concurrency: 1, |
| 753 | livenessPollIntervalMs: 1_000, |
| 754 | }).pipe( |
| 755 | Effect.provideService(EffectWorker.OrchestrationEffectWorkerV2, worker), |
| 756 | Effect.forkScoped, |
| 757 | ); |
| 758 | |
| 759 | while ((yield* Ref.get(attempts)) < 1) yield* Effect.yieldNow; |
| 760 | yield* TestClock.adjust("24 millis"); |
| 761 | assert.equal(yield* Ref.get(attempts), 1); |
| 762 | yield* TestClock.adjust("1 millis"); |
| 763 | while ((yield* Ref.get(attempts)) < 2) yield* Effect.yieldNow; |
| 764 | }).pipe(Effect.provide(TestClock.layer())), |
| 765 | ); |
| 766 | |
| 767 | it.effect("safely retries after replacement cleanup succeeds and start fails", () => |
| 768 | Effect.gen(function* () { |
| 769 | const now = yield* DateTime.now; |
| 770 | const events = yield* Ref.make<ReadonlyArray<string>>([]); |
| 771 | const failFirstStart = yield* Ref.make(true); |
| 772 | const effect = restartEffect(now, { |
| 773 | type: "replace", |
| 774 | replacementProviderSessionId: replacementSessionId, |
| 775 | }); |
| 776 | const layer = layerExecutorFor({ events, failFirstStart }); |
| 777 | |
| 778 | const first = yield* Effect.gen(function* () { |
| 779 | const executor = yield* EffectWorker.OrchestrationEffectExecutorV2; |
| 780 | return yield* Effect.exit(executor.execute(effect)); |
| 781 | }).pipe(Effect.provide(layer)); |
| 782 | assert.isTrue(Exit.isFailure(first)); |
| 783 | |
| 784 | yield* Effect.gen(function* () { |
| 785 | const executor = yield* EffectWorker.OrchestrationEffectExecutorV2; |
| 786 | yield* executor.execute(effect); |
| 787 | }).pipe(Effect.provide(layer)); |
| 788 | |
| 789 | assert.deepEqual(yield* Ref.get(events), [ |
| 790 | `interrupt:${replacementSessionId}`, |
| 791 | "detach", |
| 792 | "start", |
| 793 | `interrupt:${replacementSessionId}`, |
| 794 | "detach", |
| 795 | "start", |
| 796 | ]); |
| 797 | }), |
| 798 | ); |
| 799 | |
| 800 | it.effect("settles a delegated child once its restart continuation fails for good", () => |
| 801 | Effect.gen(function* () { |
| 802 | const timestamp = DateTime.formatIso(yield* DateTime.now); |
| 803 | const events = yield* Ref.make<ReadonlyArray<string>>([]); |
| 804 | const recovered = yield* Ref.make<ReadonlyArray<ThreadId>>([]); |
| 805 | const layer = layerExecutorFor({ |
| 806 | events, |
| 807 | continueAfterRestart: true, |
| 808 | threads: { |
| 809 | getThreadRecords: () => Effect.fail(new Error("provider instance removed") as never), |
| 810 | recoverDelegatedTask: (childThreadId) => |
| 811 | Ref.update(recovered, (ids) => [...ids, childThreadId]), |
| 812 | }, |
| 813 | }); |
| 814 | const effect: EffectOutbox.OrchestrationEffectV2 = { |
| 815 | id: `effect:restart-continuation:${runId}`, |
| 816 | commandId: CommandId.make("command:restart-continuation-failure"), |
| 817 | threadId, |
| 818 | request: { type: "provider-runtime.continue", sourceRunId: runId }, |
| 819 | status: "running", |
| 820 | attemptCount: 1, |
| 821 | availableAt: timestamp, |
| 822 | leaseOwner: "test-worker", |
| 823 | leaseExpiresAt: timestamp, |
| 824 | createdAt: timestamp, |
| 825 | updatedAt: timestamp, |
| 826 | completedAt: null, |
| 827 | lastError: null, |
| 828 | }; |
| 829 | yield* Effect.gen(function* () { |
| 830 | const executor = yield* EffectWorker.OrchestrationEffectExecutorV2; |
| 831 | assert.isTrue( |
| 832 | Exit.isFailure(yield* Effect.exit(executor.execute(effect, { willRetry: true }))), |
| 833 | ); |
| 834 | assert.deepEqual(yield* Ref.get(recovered), []); |
| 835 | assert.isTrue( |
| 836 | Exit.isFailure(yield* Effect.exit(executor.execute(effect, { willRetry: false }))), |
| 837 | ); |
| 838 | assert.deepEqual(yield* Ref.get(recovered), [threadId]); |
| 839 | }).pipe(Effect.provide(layer)); |
| 840 | }), |
| 841 | ); |
| 842 | |
| 843 | it.effect.each(["confirmed", "unconfirmed", "failed"] as const)( |
| 844 | "a native subagent stop only repairs %s provider evidence", |
| 845 | (outcome) => |
| 846 | Effect.gen(function* () { |
| 847 | const now = yield* DateTime.now; |
| 848 | const events = yield* Ref.make<ReadonlyArray<string>>([]); |
| 849 | const layer = layerExecutorFor({ |
| 850 | events, |
| 851 | interrupt: (input) => { |
| 852 | assert.equal(input.subagent?.nativeTaskId, "native-task"); |
| 853 | return Ref.update(events, (current) => [...current, "stop-child"]).pipe( |
| 854 | Effect.andThen( |
| 855 | outcome === "failed" |
| 856 | ? new ProviderTurnControlService.ProviderTurnControlError({ |
| 857 | threadId, |
| 858 | providerTurnId, |
| 859 | operation: "interrupt", |
| 860 | cause: "SDK stop failed", |
| 861 | }) |
| 862 | : Effect.succeed( |
| 863 | outcome === "confirmed" |
| 864 | ? { |
| 865 | status: "cancelled" as const, |
| 866 | completedAt: now, |
| 867 | result: "Stopped by Claude", |
| 868 | } |
| 869 | : undefined, |
| 870 | ), |
| 871 | ), |
| 872 | ); |
| 873 | }, |
| 874 | threads: { |
| 875 | dispatch: (command) => { |
| 876 | assert.equal(command.type, "thread.background-work.settle"); |
| 877 | if (command.type === "thread.background-work.settle") { |
| 878 | assert.equal(command.subagent?.id, "subagent"); |
| 879 | assert.deepEqual(command.subagent?.outcome, { |
| 880 | status: "cancelled", |
| 881 | completedAt: now, |
| 882 | result: "Stopped by Claude", |
| 883 | }); |
| 884 | } |
| 885 | return Ref.update(events, (current) => [...current, "settle-child"]).pipe( |
| 886 | Effect.as({ sequence: 1, storedEvents: [] }), |
| 887 | ); |
| 888 | }, |
| 889 | }, |
| 890 | }); |
| 891 | const execution = yield* Effect.gen(function* () { |
| 892 | const executor = yield* EffectWorker.OrchestrationEffectExecutorV2; |
| 893 | yield* executor.execute({ |
| 894 | ...restartEffect(now, { type: "detach" }), |
| 895 | request: { |
| 896 | type: "provider-turn.interrupt", |
| 897 | providerSessionId: oldSessionId, |
| 898 | providerThreadId, |
| 899 | providerTurnId, |
| 900 | subagent: { id: NodeId.make("subagent"), nativeTaskId: "native-task" }, |
| 901 | }, |
| 902 | }); |
| 903 | }).pipe(Effect.provide(layer), Effect.exit); |
| 904 | assert.equal(Exit.isFailure(execution), outcome === "failed"); |
| 905 | assert.deepEqual( |
| 906 | yield* Ref.get(events), |
| 907 | outcome === "confirmed" ? ["stop-child", "settle-child"] : ["stop-child"], |
| 908 | ); |
| 909 | }), |
| 910 | ); |