@kortyx/agent 0.30.0 → 0.31.0

This diff represents the content of publicly available package versions that have been released to one of the supported registries. The information contained in this diff is provided for informational purposes only and reflects changes between package versions as they appear in their respective public registries.
@@ -1,5 +1,5 @@
1
- import { GetProviderFn, ProviderModelRef } from '@kortyx/providers';
2
1
  import { z } from 'zod';
2
+ import { GetProviderFn, ProviderModelRef } from '@kortyx/providers';
3
3
  import { KortyxError, FailureDescriptor } from '@kortyx/core/errors';
4
4
  import { FinalizedChatMessage, StreamChunk } from '@kortyx/stream';
5
5
  import { WorkflowDefinition, ExecutionLimits, TokenUsage, ExecutionLimitReached } from '@kortyx/core';
@@ -2273,7 +2273,1203 @@ type DefinedSuite<Params> = Omit<EvalSuite<Params>, "cases"> & {
2273
2273
  /** Author a suite with typed case parameters. Runtime validation belongs to createEvals. */
2274
2274
  declare function defineSuite<Params = EvalJson | undefined>(suite: DefinedSuite<Params>): DefinedSuite<Params>;
2275
2275
 
2276
+ declare const EvaluationMetadataSchema: z.ZodObject<{
2277
+ source: z.ZodDefault<z.ZodEnum<{
2278
+ manual: "manual";
2279
+ deployment: "deployment";
2280
+ schedule: "schedule";
2281
+ ci: "ci";
2282
+ }>>;
2283
+ commit: z.ZodOptional<z.ZodString>;
2284
+ deploymentUrl: z.ZodOptional<z.ZodURL>;
2285
+ }, z.core.$strict>;
2286
+ declare const StudioEvaluationStartRequestSchema: z.ZodObject<{
2287
+ targetId: z.ZodString;
2288
+ name: z.ZodOptional<z.ZodString>;
2289
+ selection: z.ZodEnum<{
2290
+ selected: "selected";
2291
+ all: "all";
2292
+ }>;
2293
+ suites: z.ZodArray<z.ZodObject<{
2294
+ suiteId: z.ZodString;
2295
+ suiteRevision: z.ZodString;
2296
+ caseIds: z.ZodOptional<z.ZodArray<z.ZodString>>;
2297
+ }, z.core.$strict>>;
2298
+ judge: z.ZodDefault<z.ZodEnum<{
2299
+ app: "app";
2300
+ studio: "studio";
2301
+ }>>;
2302
+ repetitions: z.ZodDefault<z.ZodNumber>;
2303
+ concurrency: z.ZodDefault<z.ZodNumber>;
2304
+ metadata: z.ZodDefault<z.ZodObject<{
2305
+ source: z.ZodDefault<z.ZodEnum<{
2306
+ manual: "manual";
2307
+ deployment: "deployment";
2308
+ schedule: "schedule";
2309
+ ci: "ci";
2310
+ }>>;
2311
+ commit: z.ZodOptional<z.ZodString>;
2312
+ deploymentUrl: z.ZodOptional<z.ZodURL>;
2313
+ }, z.core.$strict>>;
2314
+ idempotencyKey: z.ZodOptional<z.ZodString>;
2315
+ }, z.core.$strict>;
2316
+ type StudioEvaluationStartRequest = z.input<typeof StudioEvaluationStartRequestSchema>;
2317
+ declare const StudioEvaluationSummarySchema: z.ZodObject<{
2318
+ id: z.ZodUUID;
2319
+ name: z.ZodString;
2320
+ targetId: z.ZodString;
2321
+ targetName: z.ZodString;
2322
+ environment: z.ZodString;
2323
+ selection: z.ZodEnum<{
2324
+ selected: "selected";
2325
+ all: "all";
2326
+ }>;
2327
+ metadata: z.ZodObject<{
2328
+ source: z.ZodDefault<z.ZodEnum<{
2329
+ manual: "manual";
2330
+ deployment: "deployment";
2331
+ schedule: "schedule";
2332
+ ci: "ci";
2333
+ }>>;
2334
+ commit: z.ZodOptional<z.ZodString>;
2335
+ deploymentUrl: z.ZodOptional<z.ZodURL>;
2336
+ }, z.core.$strict>;
2337
+ judge: z.ZodOptional<z.ZodObject<{
2338
+ id: z.ZodString;
2339
+ version: z.ZodString;
2340
+ location: z.ZodOptional<z.ZodEnum<{
2341
+ app: "app";
2342
+ studio: "studio";
2343
+ }>>;
2344
+ }, z.core.$strict>>;
2345
+ status: z.ZodEnum<{
2346
+ cancelled: "cancelled";
2347
+ failed: "failed";
2348
+ error: "error";
2349
+ passed: "passed";
2350
+ queued: "queued";
2351
+ running: "running";
2352
+ }>;
2353
+ createdAt: z.ZodString;
2354
+ startedAt: z.ZodNullable<z.ZodString>;
2355
+ endedAt: z.ZodNullable<z.ZodString>;
2356
+ cancelRequestedAt: z.ZodNullable<z.ZodString>;
2357
+ suiteCount: z.ZodNumber;
2358
+ completedSuites: z.ZodNumber;
2359
+ totalAttempts: z.ZodNumber;
2360
+ completedAttempts: z.ZodNumber;
2361
+ counts: z.ZodObject<{
2362
+ passed: z.ZodNumber;
2363
+ failed: z.ZodNumber;
2364
+ error: z.ZodNumber;
2365
+ cancelled: z.ZodNumber;
2366
+ }, z.core.$strip>;
2367
+ costs: z.ZodOptional<z.ZodObject<{
2368
+ workflow: z.ZodObject<{
2369
+ amount: z.ZodNullable<z.ZodNumber>;
2370
+ currency: z.ZodNullable<z.ZodString>;
2371
+ status: z.ZodEnum<{
2372
+ complete: "complete";
2373
+ partial: "partial";
2374
+ unavailable: "unavailable";
2375
+ }>;
2376
+ calls: z.ZodNumber;
2377
+ unpricedCalls: z.ZodNumber;
2378
+ estimated: z.ZodBoolean;
2379
+ }, z.core.$strip>;
2380
+ judge: z.ZodObject<{
2381
+ amount: z.ZodNullable<z.ZodNumber>;
2382
+ currency: z.ZodNullable<z.ZodString>;
2383
+ status: z.ZodEnum<{
2384
+ complete: "complete";
2385
+ partial: "partial";
2386
+ unavailable: "unavailable";
2387
+ }>;
2388
+ calls: z.ZodNumber;
2389
+ unpricedCalls: z.ZodNumber;
2390
+ estimated: z.ZodBoolean;
2391
+ }, z.core.$strip>;
2392
+ total: z.ZodObject<{
2393
+ amount: z.ZodNullable<z.ZodNumber>;
2394
+ currency: z.ZodNullable<z.ZodString>;
2395
+ status: z.ZodEnum<{
2396
+ complete: "complete";
2397
+ partial: "partial";
2398
+ unavailable: "unavailable";
2399
+ }>;
2400
+ calls: z.ZodNumber;
2401
+ unpricedCalls: z.ZodNumber;
2402
+ estimated: z.ZodBoolean;
2403
+ }, z.core.$strip>;
2404
+ }, z.core.$strip>>;
2405
+ legacy: z.ZodOptional<z.ZodBoolean>;
2406
+ }, z.core.$strip>;
2407
+ declare const StudioEvaluationHistorySchema: z.ZodObject<{
2408
+ runs: z.ZodArray<z.ZodObject<{
2409
+ id: z.ZodUUID;
2410
+ name: z.ZodString;
2411
+ targetId: z.ZodString;
2412
+ targetName: z.ZodString;
2413
+ environment: z.ZodString;
2414
+ selection: z.ZodEnum<{
2415
+ selected: "selected";
2416
+ all: "all";
2417
+ }>;
2418
+ metadata: z.ZodObject<{
2419
+ source: z.ZodDefault<z.ZodEnum<{
2420
+ manual: "manual";
2421
+ deployment: "deployment";
2422
+ schedule: "schedule";
2423
+ ci: "ci";
2424
+ }>>;
2425
+ commit: z.ZodOptional<z.ZodString>;
2426
+ deploymentUrl: z.ZodOptional<z.ZodURL>;
2427
+ }, z.core.$strict>;
2428
+ judge: z.ZodOptional<z.ZodObject<{
2429
+ id: z.ZodString;
2430
+ version: z.ZodString;
2431
+ location: z.ZodOptional<z.ZodEnum<{
2432
+ app: "app";
2433
+ studio: "studio";
2434
+ }>>;
2435
+ }, z.core.$strict>>;
2436
+ status: z.ZodEnum<{
2437
+ cancelled: "cancelled";
2438
+ failed: "failed";
2439
+ error: "error";
2440
+ passed: "passed";
2441
+ queued: "queued";
2442
+ running: "running";
2443
+ }>;
2444
+ createdAt: z.ZodString;
2445
+ startedAt: z.ZodNullable<z.ZodString>;
2446
+ endedAt: z.ZodNullable<z.ZodString>;
2447
+ cancelRequestedAt: z.ZodNullable<z.ZodString>;
2448
+ suiteCount: z.ZodNumber;
2449
+ completedSuites: z.ZodNumber;
2450
+ totalAttempts: z.ZodNumber;
2451
+ completedAttempts: z.ZodNumber;
2452
+ counts: z.ZodObject<{
2453
+ passed: z.ZodNumber;
2454
+ failed: z.ZodNumber;
2455
+ error: z.ZodNumber;
2456
+ cancelled: z.ZodNumber;
2457
+ }, z.core.$strip>;
2458
+ costs: z.ZodOptional<z.ZodObject<{
2459
+ workflow: z.ZodObject<{
2460
+ amount: z.ZodNullable<z.ZodNumber>;
2461
+ currency: z.ZodNullable<z.ZodString>;
2462
+ status: z.ZodEnum<{
2463
+ complete: "complete";
2464
+ partial: "partial";
2465
+ unavailable: "unavailable";
2466
+ }>;
2467
+ calls: z.ZodNumber;
2468
+ unpricedCalls: z.ZodNumber;
2469
+ estimated: z.ZodBoolean;
2470
+ }, z.core.$strip>;
2471
+ judge: z.ZodObject<{
2472
+ amount: z.ZodNullable<z.ZodNumber>;
2473
+ currency: z.ZodNullable<z.ZodString>;
2474
+ status: z.ZodEnum<{
2475
+ complete: "complete";
2476
+ partial: "partial";
2477
+ unavailable: "unavailable";
2478
+ }>;
2479
+ calls: z.ZodNumber;
2480
+ unpricedCalls: z.ZodNumber;
2481
+ estimated: z.ZodBoolean;
2482
+ }, z.core.$strip>;
2483
+ total: z.ZodObject<{
2484
+ amount: z.ZodNullable<z.ZodNumber>;
2485
+ currency: z.ZodNullable<z.ZodString>;
2486
+ status: z.ZodEnum<{
2487
+ complete: "complete";
2488
+ partial: "partial";
2489
+ unavailable: "unavailable";
2490
+ }>;
2491
+ calls: z.ZodNumber;
2492
+ unpricedCalls: z.ZodNumber;
2493
+ estimated: z.ZodBoolean;
2494
+ }, z.core.$strip>;
2495
+ }, z.core.$strip>>;
2496
+ legacy: z.ZodOptional<z.ZodBoolean>;
2497
+ }, z.core.$strip>>;
2498
+ }, z.core.$strip>;
2499
+ declare const StudioEvaluationDetailSchema: z.ZodObject<{
2500
+ run: z.ZodObject<{
2501
+ id: z.ZodUUID;
2502
+ name: z.ZodString;
2503
+ targetId: z.ZodString;
2504
+ targetName: z.ZodString;
2505
+ environment: z.ZodString;
2506
+ selection: z.ZodEnum<{
2507
+ selected: "selected";
2508
+ all: "all";
2509
+ }>;
2510
+ metadata: z.ZodObject<{
2511
+ source: z.ZodDefault<z.ZodEnum<{
2512
+ manual: "manual";
2513
+ deployment: "deployment";
2514
+ schedule: "schedule";
2515
+ ci: "ci";
2516
+ }>>;
2517
+ commit: z.ZodOptional<z.ZodString>;
2518
+ deploymentUrl: z.ZodOptional<z.ZodURL>;
2519
+ }, z.core.$strict>;
2520
+ judge: z.ZodOptional<z.ZodObject<{
2521
+ id: z.ZodString;
2522
+ version: z.ZodString;
2523
+ location: z.ZodOptional<z.ZodEnum<{
2524
+ app: "app";
2525
+ studio: "studio";
2526
+ }>>;
2527
+ }, z.core.$strict>>;
2528
+ status: z.ZodEnum<{
2529
+ cancelled: "cancelled";
2530
+ failed: "failed";
2531
+ error: "error";
2532
+ passed: "passed";
2533
+ queued: "queued";
2534
+ running: "running";
2535
+ }>;
2536
+ createdAt: z.ZodString;
2537
+ startedAt: z.ZodNullable<z.ZodString>;
2538
+ endedAt: z.ZodNullable<z.ZodString>;
2539
+ cancelRequestedAt: z.ZodNullable<z.ZodString>;
2540
+ suiteCount: z.ZodNumber;
2541
+ completedSuites: z.ZodNumber;
2542
+ totalAttempts: z.ZodNumber;
2543
+ completedAttempts: z.ZodNumber;
2544
+ counts: z.ZodObject<{
2545
+ passed: z.ZodNumber;
2546
+ failed: z.ZodNumber;
2547
+ error: z.ZodNumber;
2548
+ cancelled: z.ZodNumber;
2549
+ }, z.core.$strip>;
2550
+ costs: z.ZodOptional<z.ZodObject<{
2551
+ workflow: z.ZodObject<{
2552
+ amount: z.ZodNullable<z.ZodNumber>;
2553
+ currency: z.ZodNullable<z.ZodString>;
2554
+ status: z.ZodEnum<{
2555
+ complete: "complete";
2556
+ partial: "partial";
2557
+ unavailable: "unavailable";
2558
+ }>;
2559
+ calls: z.ZodNumber;
2560
+ unpricedCalls: z.ZodNumber;
2561
+ estimated: z.ZodBoolean;
2562
+ }, z.core.$strip>;
2563
+ judge: z.ZodObject<{
2564
+ amount: z.ZodNullable<z.ZodNumber>;
2565
+ currency: z.ZodNullable<z.ZodString>;
2566
+ status: z.ZodEnum<{
2567
+ complete: "complete";
2568
+ partial: "partial";
2569
+ unavailable: "unavailable";
2570
+ }>;
2571
+ calls: z.ZodNumber;
2572
+ unpricedCalls: z.ZodNumber;
2573
+ estimated: z.ZodBoolean;
2574
+ }, z.core.$strip>;
2575
+ total: z.ZodObject<{
2576
+ amount: z.ZodNullable<z.ZodNumber>;
2577
+ currency: z.ZodNullable<z.ZodString>;
2578
+ status: z.ZodEnum<{
2579
+ complete: "complete";
2580
+ partial: "partial";
2581
+ unavailable: "unavailable";
2582
+ }>;
2583
+ calls: z.ZodNumber;
2584
+ unpricedCalls: z.ZodNumber;
2585
+ estimated: z.ZodBoolean;
2586
+ }, z.core.$strip>;
2587
+ }, z.core.$strip>>;
2588
+ legacy: z.ZodOptional<z.ZodBoolean>;
2589
+ suites: z.ZodArray<z.ZodObject<{
2590
+ id: z.ZodUUID;
2591
+ evaluationId: z.ZodOptional<z.ZodNullable<z.ZodUUID>>;
2592
+ targetId: z.ZodString;
2593
+ targetName: z.ZodString;
2594
+ environment: z.ZodString;
2595
+ suiteId: z.ZodString;
2596
+ suiteRevision: z.ZodString;
2597
+ status: z.ZodEnum<{
2598
+ cancelled: "cancelled";
2599
+ failed: "failed";
2600
+ error: "error";
2601
+ passed: "passed";
2602
+ queued: "queued";
2603
+ running: "running";
2604
+ }>;
2605
+ createdAt: z.ZodString;
2606
+ startedAt: z.ZodNullable<z.ZodString>;
2607
+ endedAt: z.ZodNullable<z.ZodString>;
2608
+ error: z.ZodNullable<z.ZodString>;
2609
+ cancelRequestedAt: z.ZodNullable<z.ZodString>;
2610
+ costs: z.ZodOptional<z.ZodObject<{
2611
+ workflow: z.ZodObject<{
2612
+ amount: z.ZodNullable<z.ZodNumber>;
2613
+ currency: z.ZodNullable<z.ZodString>;
2614
+ status: z.ZodEnum<{
2615
+ complete: "complete";
2616
+ partial: "partial";
2617
+ unavailable: "unavailable";
2618
+ }>;
2619
+ calls: z.ZodNumber;
2620
+ unpricedCalls: z.ZodNumber;
2621
+ estimated: z.ZodBoolean;
2622
+ }, z.core.$strip>;
2623
+ judge: z.ZodObject<{
2624
+ amount: z.ZodNullable<z.ZodNumber>;
2625
+ currency: z.ZodNullable<z.ZodString>;
2626
+ status: z.ZodEnum<{
2627
+ complete: "complete";
2628
+ partial: "partial";
2629
+ unavailable: "unavailable";
2630
+ }>;
2631
+ calls: z.ZodNumber;
2632
+ unpricedCalls: z.ZodNumber;
2633
+ estimated: z.ZodBoolean;
2634
+ }, z.core.$strip>;
2635
+ total: z.ZodObject<{
2636
+ amount: z.ZodNullable<z.ZodNumber>;
2637
+ currency: z.ZodNullable<z.ZodString>;
2638
+ status: z.ZodEnum<{
2639
+ complete: "complete";
2640
+ partial: "partial";
2641
+ unavailable: "unavailable";
2642
+ }>;
2643
+ calls: z.ZodNumber;
2644
+ unpricedCalls: z.ZodNumber;
2645
+ estimated: z.ZodBoolean;
2646
+ }, z.core.$strip>;
2647
+ }, z.core.$strip>>;
2648
+ counts: z.ZodOptional<z.ZodNullable<z.ZodObject<{
2649
+ passed: z.ZodNumber;
2650
+ failed: z.ZodNumber;
2651
+ error: z.ZodNumber;
2652
+ cancelled: z.ZodNumber;
2653
+ }, z.core.$strip>>>;
2654
+ suiteName: z.ZodOptional<z.ZodNullable<z.ZodString>>;
2655
+ totalAttempts: z.ZodOptional<z.ZodNumber>;
2656
+ completedAttempts: z.ZodOptional<z.ZodNumber>;
2657
+ phase: z.ZodOptional<z.ZodString>;
2658
+ }, z.core.$strip>>;
2659
+ }, z.core.$strip>;
2660
+ }, z.core.$strip>;
2661
+ declare const StudioEvaluationResultsSchema: z.ZodObject<{
2662
+ run: z.ZodObject<{
2663
+ id: z.ZodUUID;
2664
+ name: z.ZodString;
2665
+ targetId: z.ZodString;
2666
+ targetName: z.ZodString;
2667
+ environment: z.ZodString;
2668
+ selection: z.ZodEnum<{
2669
+ selected: "selected";
2670
+ all: "all";
2671
+ }>;
2672
+ metadata: z.ZodObject<{
2673
+ source: z.ZodDefault<z.ZodEnum<{
2674
+ manual: "manual";
2675
+ deployment: "deployment";
2676
+ schedule: "schedule";
2677
+ ci: "ci";
2678
+ }>>;
2679
+ commit: z.ZodOptional<z.ZodString>;
2680
+ deploymentUrl: z.ZodOptional<z.ZodURL>;
2681
+ }, z.core.$strict>;
2682
+ judge: z.ZodOptional<z.ZodObject<{
2683
+ id: z.ZodString;
2684
+ version: z.ZodString;
2685
+ location: z.ZodOptional<z.ZodEnum<{
2686
+ app: "app";
2687
+ studio: "studio";
2688
+ }>>;
2689
+ }, z.core.$strict>>;
2690
+ status: z.ZodEnum<{
2691
+ cancelled: "cancelled";
2692
+ failed: "failed";
2693
+ error: "error";
2694
+ passed: "passed";
2695
+ queued: "queued";
2696
+ running: "running";
2697
+ }>;
2698
+ createdAt: z.ZodString;
2699
+ startedAt: z.ZodNullable<z.ZodString>;
2700
+ endedAt: z.ZodNullable<z.ZodString>;
2701
+ cancelRequestedAt: z.ZodNullable<z.ZodString>;
2702
+ suiteCount: z.ZodNumber;
2703
+ completedSuites: z.ZodNumber;
2704
+ totalAttempts: z.ZodNumber;
2705
+ completedAttempts: z.ZodNumber;
2706
+ counts: z.ZodObject<{
2707
+ passed: z.ZodNumber;
2708
+ failed: z.ZodNumber;
2709
+ error: z.ZodNumber;
2710
+ cancelled: z.ZodNumber;
2711
+ }, z.core.$strip>;
2712
+ costs: z.ZodOptional<z.ZodObject<{
2713
+ workflow: z.ZodObject<{
2714
+ amount: z.ZodNullable<z.ZodNumber>;
2715
+ currency: z.ZodNullable<z.ZodString>;
2716
+ status: z.ZodEnum<{
2717
+ complete: "complete";
2718
+ partial: "partial";
2719
+ unavailable: "unavailable";
2720
+ }>;
2721
+ calls: z.ZodNumber;
2722
+ unpricedCalls: z.ZodNumber;
2723
+ estimated: z.ZodBoolean;
2724
+ }, z.core.$strip>;
2725
+ judge: z.ZodObject<{
2726
+ amount: z.ZodNullable<z.ZodNumber>;
2727
+ currency: z.ZodNullable<z.ZodString>;
2728
+ status: z.ZodEnum<{
2729
+ complete: "complete";
2730
+ partial: "partial";
2731
+ unavailable: "unavailable";
2732
+ }>;
2733
+ calls: z.ZodNumber;
2734
+ unpricedCalls: z.ZodNumber;
2735
+ estimated: z.ZodBoolean;
2736
+ }, z.core.$strip>;
2737
+ total: z.ZodObject<{
2738
+ amount: z.ZodNullable<z.ZodNumber>;
2739
+ currency: z.ZodNullable<z.ZodString>;
2740
+ status: z.ZodEnum<{
2741
+ complete: "complete";
2742
+ partial: "partial";
2743
+ unavailable: "unavailable";
2744
+ }>;
2745
+ calls: z.ZodNumber;
2746
+ unpricedCalls: z.ZodNumber;
2747
+ estimated: z.ZodBoolean;
2748
+ }, z.core.$strip>;
2749
+ }, z.core.$strip>>;
2750
+ legacy: z.ZodOptional<z.ZodBoolean>;
2751
+ suites: z.ZodArray<z.ZodObject<{
2752
+ id: z.ZodUUID;
2753
+ evaluationId: z.ZodOptional<z.ZodNullable<z.ZodUUID>>;
2754
+ targetId: z.ZodString;
2755
+ targetName: z.ZodString;
2756
+ environment: z.ZodString;
2757
+ suiteId: z.ZodString;
2758
+ suiteRevision: z.ZodString;
2759
+ status: z.ZodEnum<{
2760
+ cancelled: "cancelled";
2761
+ failed: "failed";
2762
+ error: "error";
2763
+ passed: "passed";
2764
+ queued: "queued";
2765
+ running: "running";
2766
+ }>;
2767
+ createdAt: z.ZodString;
2768
+ startedAt: z.ZodNullable<z.ZodString>;
2769
+ endedAt: z.ZodNullable<z.ZodString>;
2770
+ error: z.ZodNullable<z.ZodString>;
2771
+ cancelRequestedAt: z.ZodNullable<z.ZodString>;
2772
+ costs: z.ZodOptional<z.ZodObject<{
2773
+ workflow: z.ZodObject<{
2774
+ amount: z.ZodNullable<z.ZodNumber>;
2775
+ currency: z.ZodNullable<z.ZodString>;
2776
+ status: z.ZodEnum<{
2777
+ complete: "complete";
2778
+ partial: "partial";
2779
+ unavailable: "unavailable";
2780
+ }>;
2781
+ calls: z.ZodNumber;
2782
+ unpricedCalls: z.ZodNumber;
2783
+ estimated: z.ZodBoolean;
2784
+ }, z.core.$strip>;
2785
+ judge: z.ZodObject<{
2786
+ amount: z.ZodNullable<z.ZodNumber>;
2787
+ currency: z.ZodNullable<z.ZodString>;
2788
+ status: z.ZodEnum<{
2789
+ complete: "complete";
2790
+ partial: "partial";
2791
+ unavailable: "unavailable";
2792
+ }>;
2793
+ calls: z.ZodNumber;
2794
+ unpricedCalls: z.ZodNumber;
2795
+ estimated: z.ZodBoolean;
2796
+ }, z.core.$strip>;
2797
+ total: z.ZodObject<{
2798
+ amount: z.ZodNullable<z.ZodNumber>;
2799
+ currency: z.ZodNullable<z.ZodString>;
2800
+ status: z.ZodEnum<{
2801
+ complete: "complete";
2802
+ partial: "partial";
2803
+ unavailable: "unavailable";
2804
+ }>;
2805
+ calls: z.ZodNumber;
2806
+ unpricedCalls: z.ZodNumber;
2807
+ estimated: z.ZodBoolean;
2808
+ }, z.core.$strip>;
2809
+ }, z.core.$strip>>;
2810
+ counts: z.ZodOptional<z.ZodNullable<z.ZodObject<{
2811
+ passed: z.ZodNumber;
2812
+ failed: z.ZodNumber;
2813
+ error: z.ZodNumber;
2814
+ cancelled: z.ZodNumber;
2815
+ }, z.core.$strip>>>;
2816
+ suiteName: z.ZodOptional<z.ZodNullable<z.ZodString>>;
2817
+ totalAttempts: z.ZodOptional<z.ZodNumber>;
2818
+ completedAttempts: z.ZodOptional<z.ZodNumber>;
2819
+ phase: z.ZodOptional<z.ZodString>;
2820
+ caseCosts: z.ZodOptional<z.ZodRecord<z.ZodString, z.ZodObject<{
2821
+ workflow: z.ZodObject<{
2822
+ amount: z.ZodNullable<z.ZodNumber>;
2823
+ currency: z.ZodNullable<z.ZodString>;
2824
+ status: z.ZodEnum<{
2825
+ complete: "complete";
2826
+ partial: "partial";
2827
+ unavailable: "unavailable";
2828
+ }>;
2829
+ calls: z.ZodNumber;
2830
+ unpricedCalls: z.ZodNumber;
2831
+ estimated: z.ZodBoolean;
2832
+ }, z.core.$strip>;
2833
+ judge: z.ZodObject<{
2834
+ amount: z.ZodNullable<z.ZodNumber>;
2835
+ currency: z.ZodNullable<z.ZodString>;
2836
+ status: z.ZodEnum<{
2837
+ complete: "complete";
2838
+ partial: "partial";
2839
+ unavailable: "unavailable";
2840
+ }>;
2841
+ calls: z.ZodNumber;
2842
+ unpricedCalls: z.ZodNumber;
2843
+ estimated: z.ZodBoolean;
2844
+ }, z.core.$strip>;
2845
+ total: z.ZodObject<{
2846
+ amount: z.ZodNullable<z.ZodNumber>;
2847
+ currency: z.ZodNullable<z.ZodString>;
2848
+ status: z.ZodEnum<{
2849
+ complete: "complete";
2850
+ partial: "partial";
2851
+ unavailable: "unavailable";
2852
+ }>;
2853
+ calls: z.ZodNumber;
2854
+ unpricedCalls: z.ZodNumber;
2855
+ estimated: z.ZodBoolean;
2856
+ }, z.core.$strip>;
2857
+ }, z.core.$strip>>>;
2858
+ request: z.ZodOptional<z.ZodObject<{
2859
+ caseIds: z.ZodOptional<z.ZodArray<z.ZodString>>;
2860
+ grading: z.ZodOptional<z.ZodEnum<{
2861
+ app: "app";
2862
+ studio: "studio";
2863
+ }>>;
2864
+ judge: z.ZodOptional<z.ZodObject<{
2865
+ id: z.ZodString;
2866
+ version: z.ZodString;
2867
+ location: z.ZodOptional<z.ZodEnum<{
2868
+ app: "app";
2869
+ studio: "studio";
2870
+ }>>;
2871
+ }, z.core.$strict>>;
2872
+ repetitions: z.ZodNumber;
2873
+ concurrency: z.ZodOptional<z.ZodNumber>;
2874
+ }, z.core.$strip>>;
2875
+ suite: z.ZodObject<{
2876
+ id: z.ZodString;
2877
+ name: z.ZodOptional<z.ZodString>;
2878
+ cases: z.ZodArray<z.ZodObject<{
2879
+ id: z.ZodString;
2880
+ name: z.ZodOptional<z.ZodString>;
2881
+ params: z.ZodOptional<z.ZodJSONSchema>;
2882
+ workflowId: z.ZodOptional<z.ZodString>;
2883
+ steps: z.ZodArray<z.ZodUnion<readonly [z.ZodObject<{
2884
+ message: z.ZodString;
2885
+ expect: z.ZodObject<{
2886
+ type: z.ZodEnum<{
2887
+ interrupt: "interrupt";
2888
+ answer: "answer";
2889
+ }>;
2890
+ schemaId: z.ZodOptional<z.ZodString>;
2891
+ schemaVersion: z.ZodOptional<z.ZodString>;
2892
+ outputs: z.ZodOptional<z.ZodArray<z.ZodObject<{
2893
+ schemaId: z.ZodString;
2894
+ schemaVersion: z.ZodOptional<z.ZodString>;
2895
+ }, z.core.$strict>>>;
2896
+ criteria: z.ZodOptional<z.ZodArray<z.ZodUnion<readonly [z.ZodString, z.ZodObject<{
2897
+ id: z.ZodString;
2898
+ text: z.ZodString;
2899
+ }, z.core.$strict>]>>>;
2900
+ reference: z.ZodOptional<z.ZodJSONSchema>;
2901
+ }, z.core.$strict>;
2902
+ }, z.core.$strict>, z.ZodObject<{
2903
+ resume: z.ZodUnion<readonly [z.ZodDiscriminatedUnion<[z.ZodObject<{
2904
+ type: z.ZodLiteral<"text">;
2905
+ text: z.ZodString;
2906
+ }, z.core.$strict>, z.ZodObject<{
2907
+ type: z.ZodLiteral<"select">;
2908
+ ids: z.ZodArray<z.ZodString>;
2909
+ }, z.core.$strict>, z.ZodObject<{
2910
+ type: z.ZodLiteral<"value">;
2911
+ value: z.ZodJSONSchema;
2912
+ }, z.core.$strict>, z.ZodObject<{
2913
+ type: z.ZodLiteral<"cancel">;
2914
+ }, z.core.$strict>], "type">, z.ZodObject<{
2915
+ using: z.ZodString;
2916
+ params: z.ZodOptional<z.ZodJSONSchema>;
2917
+ }, z.core.$strict>]>;
2918
+ expect: z.ZodObject<{
2919
+ type: z.ZodEnum<{
2920
+ interrupt: "interrupt";
2921
+ answer: "answer";
2922
+ }>;
2923
+ schemaId: z.ZodOptional<z.ZodString>;
2924
+ schemaVersion: z.ZodOptional<z.ZodString>;
2925
+ outputs: z.ZodOptional<z.ZodArray<z.ZodObject<{
2926
+ schemaId: z.ZodString;
2927
+ schemaVersion: z.ZodOptional<z.ZodString>;
2928
+ }, z.core.$strict>>>;
2929
+ criteria: z.ZodOptional<z.ZodArray<z.ZodUnion<readonly [z.ZodString, z.ZodObject<{
2930
+ id: z.ZodString;
2931
+ text: z.ZodString;
2932
+ }, z.core.$strict>]>>>;
2933
+ reference: z.ZodOptional<z.ZodJSONSchema>;
2934
+ }, z.core.$strict>;
2935
+ }, z.core.$strict>]>>;
2936
+ }, z.core.$strict>>;
2937
+ }, z.core.$strict>;
2938
+ result: z.ZodNullable<z.ZodObject<{
2939
+ id: z.ZodUUID;
2940
+ suiteId: z.ZodString;
2941
+ suiteRevision: z.ZodString;
2942
+ suite: z.ZodObject<{
2943
+ id: z.ZodString;
2944
+ name: z.ZodOptional<z.ZodString>;
2945
+ cases: z.ZodArray<z.ZodObject<{
2946
+ id: z.ZodString;
2947
+ name: z.ZodOptional<z.ZodString>;
2948
+ params: z.ZodOptional<z.ZodJSONSchema>;
2949
+ workflowId: z.ZodOptional<z.ZodString>;
2950
+ steps: z.ZodArray<z.ZodUnion<readonly [z.ZodObject<{
2951
+ message: z.ZodString;
2952
+ expect: z.ZodObject<{
2953
+ type: z.ZodEnum<{
2954
+ interrupt: "interrupt";
2955
+ answer: "answer";
2956
+ }>;
2957
+ schemaId: z.ZodOptional<z.ZodString>;
2958
+ schemaVersion: z.ZodOptional<z.ZodString>;
2959
+ outputs: z.ZodOptional<z.ZodArray<z.ZodObject<{
2960
+ schemaId: z.ZodString;
2961
+ schemaVersion: z.ZodOptional<z.ZodString>;
2962
+ }, z.core.$strict>>>;
2963
+ criteria: z.ZodOptional<z.ZodArray<z.ZodUnion<readonly [z.ZodString, z.ZodObject<{
2964
+ id: z.ZodString;
2965
+ text: z.ZodString;
2966
+ }, z.core.$strict>]>>>;
2967
+ reference: z.ZodOptional<z.ZodJSONSchema>;
2968
+ }, z.core.$strict>;
2969
+ }, z.core.$strict>, z.ZodObject<{
2970
+ resume: z.ZodUnion<readonly [z.ZodDiscriminatedUnion<[z.ZodObject<{
2971
+ type: z.ZodLiteral<"text">;
2972
+ text: z.ZodString;
2973
+ }, z.core.$strict>, z.ZodObject<{
2974
+ type: z.ZodLiteral<"select">;
2975
+ ids: z.ZodArray<z.ZodString>;
2976
+ }, z.core.$strict>, z.ZodObject<{
2977
+ type: z.ZodLiteral<"value">;
2978
+ value: z.ZodJSONSchema;
2979
+ }, z.core.$strict>, z.ZodObject<{
2980
+ type: z.ZodLiteral<"cancel">;
2981
+ }, z.core.$strict>], "type">, z.ZodObject<{
2982
+ using: z.ZodString;
2983
+ params: z.ZodOptional<z.ZodJSONSchema>;
2984
+ }, z.core.$strict>]>;
2985
+ expect: z.ZodObject<{
2986
+ type: z.ZodEnum<{
2987
+ interrupt: "interrupt";
2988
+ answer: "answer";
2989
+ }>;
2990
+ schemaId: z.ZodOptional<z.ZodString>;
2991
+ schemaVersion: z.ZodOptional<z.ZodString>;
2992
+ outputs: z.ZodOptional<z.ZodArray<z.ZodObject<{
2993
+ schemaId: z.ZodString;
2994
+ schemaVersion: z.ZodOptional<z.ZodString>;
2995
+ }, z.core.$strict>>>;
2996
+ criteria: z.ZodOptional<z.ZodArray<z.ZodUnion<readonly [z.ZodString, z.ZodObject<{
2997
+ id: z.ZodString;
2998
+ text: z.ZodString;
2999
+ }, z.core.$strict>]>>>;
3000
+ reference: z.ZodOptional<z.ZodJSONSchema>;
3001
+ }, z.core.$strict>;
3002
+ }, z.core.$strict>]>>;
3003
+ }, z.core.$strict>>;
3004
+ }, z.core.$strict>;
3005
+ judge: z.ZodOptional<z.ZodObject<{
3006
+ id: z.ZodString;
3007
+ version: z.ZodString;
3008
+ location: z.ZodOptional<z.ZodEnum<{
3009
+ app: "app";
3010
+ studio: "studio";
3011
+ }>>;
3012
+ }, z.core.$strict>>;
3013
+ startedAt: z.ZodISODateTime;
3014
+ durationMs: z.ZodNumber;
3015
+ status: z.ZodEnum<{
3016
+ cancelled: "cancelled";
3017
+ failed: "failed";
3018
+ error: "error";
3019
+ passed: "passed";
3020
+ ungraded: "ungraded";
3021
+ }>;
3022
+ counts: z.ZodObject<{
3023
+ passed: z.ZodNumber;
3024
+ failed: z.ZodNumber;
3025
+ error: z.ZodNumber;
3026
+ cancelled: z.ZodNumber;
3027
+ ungraded: z.ZodOptional<z.ZodNumber>;
3028
+ }, z.core.$strict>;
3029
+ cases: z.ZodArray<z.ZodObject<{
3030
+ caseId: z.ZodString;
3031
+ repetition: z.ZodNumber;
3032
+ sessionId: z.ZodString;
3033
+ status: z.ZodEnum<{
3034
+ cancelled: "cancelled";
3035
+ failed: "failed";
3036
+ error: "error";
3037
+ passed: "passed";
3038
+ ungraded: "ungraded";
3039
+ }>;
3040
+ durationMs: z.ZodNumber;
3041
+ steps: z.ZodArray<z.ZodObject<{
3042
+ index: z.ZodNumber;
3043
+ input: z.ZodUnion<readonly [z.ZodObject<{
3044
+ message: z.ZodString;
3045
+ }, z.core.$strict>, z.ZodObject<{
3046
+ resume: z.ZodDiscriminatedUnion<[z.ZodObject<{
3047
+ type: z.ZodLiteral<"text">;
3048
+ text: z.ZodString;
3049
+ }, z.core.$strict>, z.ZodObject<{
3050
+ type: z.ZodLiteral<"select">;
3051
+ ids: z.ZodArray<z.ZodString>;
3052
+ }, z.core.$strict>, z.ZodObject<{
3053
+ type: z.ZodLiteral<"value">;
3054
+ value: z.ZodJSONSchema;
3055
+ }, z.core.$strict>, z.ZodObject<{
3056
+ type: z.ZodLiteral<"cancel">;
3057
+ }, z.core.$strict>], "type">;
3058
+ }, z.core.$strict>]>;
3059
+ expectation: z.ZodObject<{
3060
+ type: z.ZodEnum<{
3061
+ interrupt: "interrupt";
3062
+ answer: "answer";
3063
+ }>;
3064
+ schemaId: z.ZodOptional<z.ZodString>;
3065
+ schemaVersion: z.ZodOptional<z.ZodString>;
3066
+ outputs: z.ZodOptional<z.ZodArray<z.ZodObject<{
3067
+ schemaId: z.ZodString;
3068
+ schemaVersion: z.ZodOptional<z.ZodString>;
3069
+ }, z.core.$strict>>>;
3070
+ criteria: z.ZodOptional<z.ZodArray<z.ZodUnion<readonly [z.ZodString, z.ZodObject<{
3071
+ id: z.ZodString;
3072
+ text: z.ZodString;
3073
+ }, z.core.$strict>]>>>;
3074
+ reference: z.ZodOptional<z.ZodJSONSchema>;
3075
+ }, z.core.$strict>;
3076
+ observation: z.ZodObject<{
3077
+ type: z.ZodEnum<{
3078
+ cancelled: "cancelled";
3079
+ interrupt: "interrupt";
3080
+ error: "error";
3081
+ answer: "answer";
3082
+ }>;
3083
+ text: z.ZodString;
3084
+ structured: z.ZodArray<z.ZodJSONSchema>;
3085
+ events: z.ZodOptional<z.ZodArray<z.ZodJSONSchema>>;
3086
+ runId: z.ZodOptional<z.ZodString>;
3087
+ checkpointId: z.ZodOptional<z.ZodString>;
3088
+ interrupt: z.ZodOptional<z.ZodObject<{
3089
+ requestId: z.ZodString;
3090
+ kind: z.ZodEnum<{
3091
+ text: "text";
3092
+ custom: "custom";
3093
+ "multi-choice": "multi-choice";
3094
+ choice: "choice";
3095
+ }>;
3096
+ question: z.ZodOptional<z.ZodString>;
3097
+ schemaId: z.ZodOptional<z.ZodString>;
3098
+ schemaVersion: z.ZodOptional<z.ZodString>;
3099
+ options: z.ZodArray<z.ZodObject<{
3100
+ id: z.ZodString;
3101
+ label: z.ZodString;
3102
+ description: z.ZodOptional<z.ZodString>;
3103
+ }, z.core.$strict>>;
3104
+ request: z.ZodOptional<z.ZodJSONSchema>;
3105
+ }, z.core.$strict>>;
3106
+ }, z.core.$strict>;
3107
+ reference: z.ZodOptional<z.ZodJSONSchema>;
3108
+ status: z.ZodEnum<{
3109
+ cancelled: "cancelled";
3110
+ failed: "failed";
3111
+ error: "error";
3112
+ passed: "passed";
3113
+ ungraded: "ungraded";
3114
+ }>;
3115
+ reason: z.ZodOptional<z.ZodString>;
3116
+ judgeCalls: z.ZodOptional<z.ZodNumber>;
3117
+ judgeUsage: z.ZodOptional<z.ZodArray<z.ZodObject<{
3118
+ provider: z.ZodString;
3119
+ model: z.ZodString;
3120
+ occurredAt: z.ZodISODateTime;
3121
+ usage: z.ZodOptional<z.ZodObject<{
3122
+ input: z.ZodOptional<z.ZodNumber>;
3123
+ output: z.ZodOptional<z.ZodNumber>;
3124
+ total: z.ZodOptional<z.ZodNumber>;
3125
+ reasoning: z.ZodOptional<z.ZodNumber>;
3126
+ cacheRead: z.ZodOptional<z.ZodNumber>;
3127
+ cacheWrite: z.ZodOptional<z.ZodNumber>;
3128
+ cacheWrite1h: z.ZodOptional<z.ZodNumber>;
3129
+ outputIncludesReasoning: z.ZodOptional<z.ZodBoolean>;
3130
+ inputIncludesCacheRead: z.ZodOptional<z.ZodBoolean>;
3131
+ inputIncludesCacheWrite: z.ZodOptional<z.ZodBoolean>;
3132
+ }, z.core.$strict>>;
3133
+ pricingContext: z.ZodOptional<z.ZodObject<{
3134
+ serviceTier: z.ZodOptional<z.ZodString>;
3135
+ inferenceGeo: z.ZodOptional<z.ZodString>;
3136
+ speed: z.ZodOptional<z.ZodString>;
3137
+ }, z.core.$strict>>;
3138
+ pricing: z.ZodOptional<z.ZodObject<{
3139
+ source: z.ZodLiteral<"provider">;
3140
+ currency: z.ZodLiteral<"USD">;
3141
+ actualCostMicros: z.ZodNumber;
3142
+ }, z.core.$strict>>;
3143
+ }, z.core.$strict>>>;
3144
+ criteria: z.ZodArray<z.ZodObject<{
3145
+ passed: z.ZodBoolean;
3146
+ reason: z.ZodString;
3147
+ evidence: z.ZodArray<z.ZodString>;
3148
+ id: z.ZodString;
3149
+ text: z.ZodString;
3150
+ }, z.core.$strict>>;
3151
+ }, z.core.$strict>>;
3152
+ errors: z.ZodArray<z.ZodObject<{
3153
+ phase: z.ZodEnum<{
3154
+ params: "params";
3155
+ execute: "execute";
3156
+ setup: "setup";
3157
+ responder: "responder";
3158
+ reference: "reference";
3159
+ grading: "grading";
3160
+ cleanup: "cleanup";
3161
+ reporting: "reporting";
3162
+ }>;
3163
+ code: z.ZodString;
3164
+ message: z.ZodString;
3165
+ }, z.core.$strict>>;
3166
+ }, z.core.$strict>>;
3167
+ errors: z.ZodArray<z.ZodObject<{
3168
+ phase: z.ZodEnum<{
3169
+ params: "params";
3170
+ execute: "execute";
3171
+ setup: "setup";
3172
+ responder: "responder";
3173
+ reference: "reference";
3174
+ grading: "grading";
3175
+ cleanup: "cleanup";
3176
+ reporting: "reporting";
3177
+ }>;
3178
+ code: z.ZodString;
3179
+ message: z.ZodString;
3180
+ }, z.core.$strict>>;
3181
+ }, z.core.$strict>>;
3182
+ events: z.ZodArray<z.ZodObject<{
3183
+ id: z.ZodNumber;
3184
+ event: z.ZodDiscriminatedUnion<[z.ZodObject<{
3185
+ type: z.ZodLiteral<"run-started">;
3186
+ caseIds: z.ZodArray<z.ZodString>;
3187
+ repetitions: z.ZodNumber;
3188
+ concurrency: z.ZodNumber;
3189
+ }, z.core.$strict>, z.ZodObject<{
3190
+ type: z.ZodLiteral<"case-progress">;
3191
+ caseId: z.ZodString;
3192
+ repetition: z.ZodNumber;
3193
+ phase: z.ZodEnum<{
3194
+ params: "params";
3195
+ execute: "execute";
3196
+ setup: "setup";
3197
+ responder: "responder";
3198
+ reference: "reference";
3199
+ grading: "grading";
3200
+ cleanup: "cleanup";
3201
+ reporting: "reporting";
3202
+ }>;
3203
+ stepIndex: z.ZodOptional<z.ZodNumber>;
3204
+ }, z.core.$strict>, z.ZodObject<{
3205
+ type: z.ZodLiteral<"case-started">;
3206
+ caseId: z.ZodString;
3207
+ repetition: z.ZodNumber;
3208
+ sessionId: z.ZodString;
3209
+ }, z.core.$strict>, z.ZodObject<{
3210
+ type: z.ZodLiteral<"step-completed">;
3211
+ caseId: z.ZodString;
3212
+ repetition: z.ZodNumber;
3213
+ step: z.ZodObject<{
3214
+ index: z.ZodNumber;
3215
+ input: z.ZodUnion<readonly [z.ZodObject<{
3216
+ message: z.ZodString;
3217
+ }, z.core.$strict>, z.ZodObject<{
3218
+ resume: z.ZodDiscriminatedUnion<[z.ZodObject<{
3219
+ type: z.ZodLiteral<"text">;
3220
+ text: z.ZodString;
3221
+ }, z.core.$strict>, z.ZodObject<{
3222
+ type: z.ZodLiteral<"select">;
3223
+ ids: z.ZodArray<z.ZodString>;
3224
+ }, z.core.$strict>, z.ZodObject<{
3225
+ type: z.ZodLiteral<"value">;
3226
+ value: z.ZodJSONSchema;
3227
+ }, z.core.$strict>, z.ZodObject<{
3228
+ type: z.ZodLiteral<"cancel">;
3229
+ }, z.core.$strict>], "type">;
3230
+ }, z.core.$strict>]>;
3231
+ expectation: z.ZodObject<{
3232
+ type: z.ZodEnum<{
3233
+ interrupt: "interrupt";
3234
+ answer: "answer";
3235
+ }>;
3236
+ schemaId: z.ZodOptional<z.ZodString>;
3237
+ schemaVersion: z.ZodOptional<z.ZodString>;
3238
+ outputs: z.ZodOptional<z.ZodArray<z.ZodObject<{
3239
+ schemaId: z.ZodString;
3240
+ schemaVersion: z.ZodOptional<z.ZodString>;
3241
+ }, z.core.$strict>>>;
3242
+ criteria: z.ZodOptional<z.ZodArray<z.ZodUnion<readonly [z.ZodString, z.ZodObject<{
3243
+ id: z.ZodString;
3244
+ text: z.ZodString;
3245
+ }, z.core.$strict>]>>>;
3246
+ reference: z.ZodOptional<z.ZodJSONSchema>;
3247
+ }, z.core.$strict>;
3248
+ observation: z.ZodObject<{
3249
+ type: z.ZodEnum<{
3250
+ cancelled: "cancelled";
3251
+ interrupt: "interrupt";
3252
+ error: "error";
3253
+ answer: "answer";
3254
+ }>;
3255
+ text: z.ZodString;
3256
+ structured: z.ZodArray<z.ZodJSONSchema>;
3257
+ events: z.ZodOptional<z.ZodArray<z.ZodJSONSchema>>;
3258
+ runId: z.ZodOptional<z.ZodString>;
3259
+ checkpointId: z.ZodOptional<z.ZodString>;
3260
+ interrupt: z.ZodOptional<z.ZodObject<{
3261
+ requestId: z.ZodString;
3262
+ kind: z.ZodEnum<{
3263
+ text: "text";
3264
+ custom: "custom";
3265
+ "multi-choice": "multi-choice";
3266
+ choice: "choice";
3267
+ }>;
3268
+ question: z.ZodOptional<z.ZodString>;
3269
+ schemaId: z.ZodOptional<z.ZodString>;
3270
+ schemaVersion: z.ZodOptional<z.ZodString>;
3271
+ options: z.ZodArray<z.ZodObject<{
3272
+ id: z.ZodString;
3273
+ label: z.ZodString;
3274
+ description: z.ZodOptional<z.ZodString>;
3275
+ }, z.core.$strict>>;
3276
+ request: z.ZodOptional<z.ZodJSONSchema>;
3277
+ }, z.core.$strict>>;
3278
+ }, z.core.$strict>;
3279
+ reference: z.ZodOptional<z.ZodJSONSchema>;
3280
+ status: z.ZodEnum<{
3281
+ cancelled: "cancelled";
3282
+ failed: "failed";
3283
+ error: "error";
3284
+ passed: "passed";
3285
+ ungraded: "ungraded";
3286
+ }>;
3287
+ reason: z.ZodOptional<z.ZodString>;
3288
+ judgeCalls: z.ZodOptional<z.ZodNumber>;
3289
+ judgeUsage: z.ZodOptional<z.ZodArray<z.ZodObject<{
3290
+ provider: z.ZodString;
3291
+ model: z.ZodString;
3292
+ occurredAt: z.ZodISODateTime;
3293
+ usage: z.ZodOptional<z.ZodObject<{
3294
+ input: z.ZodOptional<z.ZodNumber>;
3295
+ output: z.ZodOptional<z.ZodNumber>;
3296
+ total: z.ZodOptional<z.ZodNumber>;
3297
+ reasoning: z.ZodOptional<z.ZodNumber>;
3298
+ cacheRead: z.ZodOptional<z.ZodNumber>;
3299
+ cacheWrite: z.ZodOptional<z.ZodNumber>;
3300
+ cacheWrite1h: z.ZodOptional<z.ZodNumber>;
3301
+ outputIncludesReasoning: z.ZodOptional<z.ZodBoolean>;
3302
+ inputIncludesCacheRead: z.ZodOptional<z.ZodBoolean>;
3303
+ inputIncludesCacheWrite: z.ZodOptional<z.ZodBoolean>;
3304
+ }, z.core.$strict>>;
3305
+ pricingContext: z.ZodOptional<z.ZodObject<{
3306
+ serviceTier: z.ZodOptional<z.ZodString>;
3307
+ inferenceGeo: z.ZodOptional<z.ZodString>;
3308
+ speed: z.ZodOptional<z.ZodString>;
3309
+ }, z.core.$strict>>;
3310
+ pricing: z.ZodOptional<z.ZodObject<{
3311
+ source: z.ZodLiteral<"provider">;
3312
+ currency: z.ZodLiteral<"USD">;
3313
+ actualCostMicros: z.ZodNumber;
3314
+ }, z.core.$strict>>;
3315
+ }, z.core.$strict>>>;
3316
+ criteria: z.ZodArray<z.ZodObject<{
3317
+ passed: z.ZodBoolean;
3318
+ reason: z.ZodString;
3319
+ evidence: z.ZodArray<z.ZodString>;
3320
+ id: z.ZodString;
3321
+ text: z.ZodString;
3322
+ }, z.core.$strict>>;
3323
+ }, z.core.$strict>;
3324
+ }, z.core.$strict>, z.ZodObject<{
3325
+ type: z.ZodLiteral<"case-completed">;
3326
+ result: z.ZodObject<{
3327
+ caseId: z.ZodString;
3328
+ repetition: z.ZodNumber;
3329
+ sessionId: z.ZodString;
3330
+ status: z.ZodEnum<{
3331
+ cancelled: "cancelled";
3332
+ failed: "failed";
3333
+ error: "error";
3334
+ passed: "passed";
3335
+ ungraded: "ungraded";
3336
+ }>;
3337
+ durationMs: z.ZodNumber;
3338
+ steps: z.ZodArray<z.ZodObject<{
3339
+ index: z.ZodNumber;
3340
+ input: z.ZodUnion<readonly [z.ZodObject<{
3341
+ message: z.ZodString;
3342
+ }, z.core.$strict>, z.ZodObject<{
3343
+ resume: z.ZodDiscriminatedUnion<[z.ZodObject<{
3344
+ type: z.ZodLiteral<"text">;
3345
+ text: z.ZodString;
3346
+ }, z.core.$strict>, z.ZodObject<{
3347
+ type: z.ZodLiteral<"select">;
3348
+ ids: z.ZodArray<z.ZodString>;
3349
+ }, z.core.$strict>, z.ZodObject<{
3350
+ type: z.ZodLiteral<"value">;
3351
+ value: z.ZodJSONSchema;
3352
+ }, z.core.$strict>, z.ZodObject<{
3353
+ type: z.ZodLiteral<"cancel">;
3354
+ }, z.core.$strict>], "type">;
3355
+ }, z.core.$strict>]>;
3356
+ expectation: z.ZodObject<{
3357
+ type: z.ZodEnum<{
3358
+ interrupt: "interrupt";
3359
+ answer: "answer";
3360
+ }>;
3361
+ schemaId: z.ZodOptional<z.ZodString>;
3362
+ schemaVersion: z.ZodOptional<z.ZodString>;
3363
+ outputs: z.ZodOptional<z.ZodArray<z.ZodObject<{
3364
+ schemaId: z.ZodString;
3365
+ schemaVersion: z.ZodOptional<z.ZodString>;
3366
+ }, z.core.$strict>>>;
3367
+ criteria: z.ZodOptional<z.ZodArray<z.ZodUnion<readonly [z.ZodString, z.ZodObject<{
3368
+ id: z.ZodString;
3369
+ text: z.ZodString;
3370
+ }, z.core.$strict>]>>>;
3371
+ reference: z.ZodOptional<z.ZodJSONSchema>;
3372
+ }, z.core.$strict>;
3373
+ observation: z.ZodObject<{
3374
+ type: z.ZodEnum<{
3375
+ cancelled: "cancelled";
3376
+ interrupt: "interrupt";
3377
+ error: "error";
3378
+ answer: "answer";
3379
+ }>;
3380
+ text: z.ZodString;
3381
+ structured: z.ZodArray<z.ZodJSONSchema>;
3382
+ events: z.ZodOptional<z.ZodArray<z.ZodJSONSchema>>;
3383
+ runId: z.ZodOptional<z.ZodString>;
3384
+ checkpointId: z.ZodOptional<z.ZodString>;
3385
+ interrupt: z.ZodOptional<z.ZodObject<{
3386
+ requestId: z.ZodString;
3387
+ kind: z.ZodEnum<{
3388
+ text: "text";
3389
+ custom: "custom";
3390
+ "multi-choice": "multi-choice";
3391
+ choice: "choice";
3392
+ }>;
3393
+ question: z.ZodOptional<z.ZodString>;
3394
+ schemaId: z.ZodOptional<z.ZodString>;
3395
+ schemaVersion: z.ZodOptional<z.ZodString>;
3396
+ options: z.ZodArray<z.ZodObject<{
3397
+ id: z.ZodString;
3398
+ label: z.ZodString;
3399
+ description: z.ZodOptional<z.ZodString>;
3400
+ }, z.core.$strict>>;
3401
+ request: z.ZodOptional<z.ZodJSONSchema>;
3402
+ }, z.core.$strict>>;
3403
+ }, z.core.$strict>;
3404
+ reference: z.ZodOptional<z.ZodJSONSchema>;
3405
+ status: z.ZodEnum<{
3406
+ cancelled: "cancelled";
3407
+ failed: "failed";
3408
+ error: "error";
3409
+ passed: "passed";
3410
+ ungraded: "ungraded";
3411
+ }>;
3412
+ reason: z.ZodOptional<z.ZodString>;
3413
+ judgeCalls: z.ZodOptional<z.ZodNumber>;
3414
+ judgeUsage: z.ZodOptional<z.ZodArray<z.ZodObject<{
3415
+ provider: z.ZodString;
3416
+ model: z.ZodString;
3417
+ occurredAt: z.ZodISODateTime;
3418
+ usage: z.ZodOptional<z.ZodObject<{
3419
+ input: z.ZodOptional<z.ZodNumber>;
3420
+ output: z.ZodOptional<z.ZodNumber>;
3421
+ total: z.ZodOptional<z.ZodNumber>;
3422
+ reasoning: z.ZodOptional<z.ZodNumber>;
3423
+ cacheRead: z.ZodOptional<z.ZodNumber>;
3424
+ cacheWrite: z.ZodOptional<z.ZodNumber>;
3425
+ cacheWrite1h: z.ZodOptional<z.ZodNumber>;
3426
+ outputIncludesReasoning: z.ZodOptional<z.ZodBoolean>;
3427
+ inputIncludesCacheRead: z.ZodOptional<z.ZodBoolean>;
3428
+ inputIncludesCacheWrite: z.ZodOptional<z.ZodBoolean>;
3429
+ }, z.core.$strict>>;
3430
+ pricingContext: z.ZodOptional<z.ZodObject<{
3431
+ serviceTier: z.ZodOptional<z.ZodString>;
3432
+ inferenceGeo: z.ZodOptional<z.ZodString>;
3433
+ speed: z.ZodOptional<z.ZodString>;
3434
+ }, z.core.$strict>>;
3435
+ pricing: z.ZodOptional<z.ZodObject<{
3436
+ source: z.ZodLiteral<"provider">;
3437
+ currency: z.ZodLiteral<"USD">;
3438
+ actualCostMicros: z.ZodNumber;
3439
+ }, z.core.$strict>>;
3440
+ }, z.core.$strict>>>;
3441
+ criteria: z.ZodArray<z.ZodObject<{
3442
+ passed: z.ZodBoolean;
3443
+ reason: z.ZodString;
3444
+ evidence: z.ZodArray<z.ZodString>;
3445
+ id: z.ZodString;
3446
+ text: z.ZodString;
3447
+ }, z.core.$strict>>;
3448
+ }, z.core.$strict>>;
3449
+ errors: z.ZodArray<z.ZodObject<{
3450
+ phase: z.ZodEnum<{
3451
+ params: "params";
3452
+ execute: "execute";
3453
+ setup: "setup";
3454
+ responder: "responder";
3455
+ reference: "reference";
3456
+ grading: "grading";
3457
+ cleanup: "cleanup";
3458
+ reporting: "reporting";
3459
+ }>;
3460
+ code: z.ZodString;
3461
+ message: z.ZodString;
3462
+ }, z.core.$strict>>;
3463
+ }, z.core.$strict>;
3464
+ }, z.core.$strict>], "type">;
3465
+ }, z.core.$strip>>;
3466
+ }, z.core.$strip>>;
3467
+ }, z.core.$strip>;
3468
+ }, z.core.$strip>;
3469
+ type StudioEvaluationSummary = z.infer<typeof StudioEvaluationSummarySchema>;
3470
+ type StudioEvaluationDetail = z.infer<typeof StudioEvaluationDetailSchema>["run"];
3471
+
2276
3472
  /** Discover and pin the Studio backend's judge; grading never runs in the browser. */
2277
3473
  declare function createStudioEvalJudge(options: StudioEvalJudgeOptions): Promise<EvalJudge>;
2278
3474
 
2279
- export { StudioEvalStartRequestSchema as $, type EvalOutputExpectation as A, type EvalPhase as B, type CreateEvalsOptions as C, type EvalProgress as D, type EvalCase as E, EvalProgressSchema as F, type EvalRemoteRunRequest as G, EvalRemoteRunRequestSchema as H, type EvalResponder as I, EvalResumeResponseSchema as J, type EvalRunOptions as K, type EvalRunResult as L, EvalRunResultSchema as M, type EvalSetupContext as N, type EvalStatus as O, type EvalStep as P, type EvalStepResult as Q, EvalStepResultSchema as R, type EvalSuite as S, EvalSuiteSchema as T, type EvalVerdict as U, EvalVerdictSchema as V, type EvalWireEvent as W, EvalWireEventSchema as X, type StudioEvalJudgeOptions as Y, StudioEvalJudgeRequestSchema as Z, StudioEvalJudgeResponseSchema as _, type EvalCaseResult as a, createStudioEvalJudge as a0, defineSuite as a1, parseEvalSuite as a2, type ChatLifecyclePhase as a3, ChatLifecycleHookError as a4, type Agent as a5, type ChatResponseFinalized as a6, type AgentProcessOptions as a7, type AgentProjectTopologyOptions as a8, type ChatResponseStatus as a9, type CreateAgentArgs as aa, type ExecutableWorkflow as ab, type ExecuteOptions as ac, type ExecutionInfo as ad, type ExecutionInterrupt as ae, ExecutionRequestError as af, type ExecutionResult as ag, type InterruptScope as ah, type ListInterruptsOptions as ai, type PendingInterrupt as aj, type ResumableInterrupt as ak, type ResumeHandle as al, type ResumeOptions as am, type ResumeResponse as an, createAgent as ao, projectWorkflowTopology as ap, EvalCaseResultSchema as b, type EvalCommand as c, EvalConfigurationError as d, type EvalContext as e, type EvalCriterion as f, type EvalCriterionResult as g, type EvalDefaults as h, type EvalExecution as i, type EvalExecutionStatus as j, type EvalExpectation as k, type EvalGradeInput as l, type EvalHandlerRef as m, type EvalInterrupt as n, type EvalIssue as o, type EvalJson as p, type EvalJudge as q, type EvalJudgeIdentity as r, EvalJudgeIdentitySchema as s, type EvalJudgeOptions as t, type EvalJudgeUsage as u, EvalJudgeUsageSchema as v, type EvalManifest as w, EvalManifestSchema as x, type EvalObservation as y, EvalObservationSchema as z };
3475
+ export { StudioEvalJudgeResponseSchema as $, type EvalOutputExpectation as A, type EvalPhase as B, type CreateEvalsOptions as C, type EvalProgress as D, type EvalCase as E, EvalProgressSchema as F, type EvalRemoteRunRequest as G, EvalRemoteRunRequestSchema as H, type EvalResponder as I, EvalResumeResponseSchema as J, type EvalRunOptions as K, type EvalRunResult as L, EvalRunResultSchema as M, type EvalSetupContext as N, type EvalStatus as O, type EvalStep as P, type EvalStepResult as Q, EvalStepResultSchema as R, type EvalSuite as S, EvalSuiteSchema as T, type EvalVerdict as U, EvalVerdictSchema as V, type EvalWireEvent as W, EvalWireEventSchema as X, EvaluationMetadataSchema as Y, type StudioEvalJudgeOptions as Z, StudioEvalJudgeRequestSchema as _, type EvalCaseResult as a, StudioEvalStartRequestSchema as a0, type StudioEvaluationDetail as a1, StudioEvaluationDetailSchema as a2, StudioEvaluationHistorySchema as a3, StudioEvaluationResultsSchema as a4, type StudioEvaluationStartRequest as a5, StudioEvaluationStartRequestSchema as a6, type StudioEvaluationSummary as a7, StudioEvaluationSummarySchema as a8, createStudioEvalJudge as a9, defineSuite as aa, parseEvalSuite as ab, type ChatLifecyclePhase as ac, ChatLifecycleHookError as ad, type Agent as ae, type ChatResponseFinalized as af, type AgentProcessOptions as ag, type AgentProjectTopologyOptions as ah, type ChatResponseStatus as ai, type CreateAgentArgs as aj, type ExecutableWorkflow as ak, type ExecuteOptions as al, type ExecutionInfo as am, type ExecutionInterrupt as an, ExecutionRequestError as ao, type ExecutionResult as ap, type InterruptScope as aq, type ListInterruptsOptions as ar, type PendingInterrupt as as, type ResumableInterrupt as at, type ResumeHandle as au, type ResumeOptions as av, type ResumeResponse as aw, createAgent as ax, projectWorkflowTopology as ay, EvalCaseResultSchema as b, type EvalCommand as c, EvalConfigurationError as d, type EvalContext as e, type EvalCriterion as f, type EvalCriterionResult as g, type EvalDefaults as h, type EvalExecution as i, type EvalExecutionStatus as j, type EvalExpectation as k, type EvalGradeInput as l, type EvalHandlerRef as m, type EvalInterrupt as n, type EvalIssue as o, type EvalJson as p, type EvalJudge as q, type EvalJudgeIdentity as r, EvalJudgeIdentitySchema as s, type EvalJudgeOptions as t, type EvalJudgeUsage as u, EvalJudgeUsageSchema as v, type EvalManifest as w, EvalManifestSchema as x, type EvalObservation as y, EvalObservationSchema as z };