openpond-sdk 0.9.3 → 0.9.5

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.
Files changed (29) hide show
  1. package/README.md +8 -0
  2. package/dist/{chunk-ZV2BLO3D.js → chunk-62LWLACP.js} +66 -34
  3. package/dist/chunk-62LWLACP.js.map +7 -0
  4. package/dist/{chunk-DNBAJUVR.js → chunk-6D4JIOOC.js} +8 -3
  5. package/dist/{chunk-DNBAJUVR.js.map → chunk-6D4JIOOC.js.map} +2 -2
  6. package/dist/{chunk-KLX6OVHR.js → chunk-EMANJDJR.js} +37 -3
  7. package/dist/chunk-EMANJDJR.js.map +7 -0
  8. package/dist/dataset-workspaces.js +1 -1
  9. package/dist/experiments.js +9 -2
  10. package/dist/index.js +2 -2
  11. package/dist/model-taskset-runs.js +5 -1
  12. package/dist/model-taskset-runs.js.map +1 -1
  13. package/dist/taskset-packages.js +1 -1
  14. package/dist/types/packages/sdk/src/experiment-client.d.ts +460 -5
  15. package/dist/types/packages/sdk/src/experiment-client.d.ts.map +1 -1
  16. package/dist/types/packages/sdk/src/experiment-contracts.d.ts +314 -0
  17. package/dist/types/packages/sdk/src/experiment-contracts.d.ts.map +1 -1
  18. package/dist/types/packages/sdk/src/experiments.d.ts +1 -0
  19. package/dist/types/packages/sdk/src/experiments.d.ts.map +1 -1
  20. package/dist/types/packages/sdk/src/harness-experiment-policy.d.ts +44 -0
  21. package/dist/types/packages/sdk/src/harness-experiment-policy.d.ts.map +1 -0
  22. package/dist/types/packages/sdk/src/model-taskset-runs-contracts.d.ts +159 -0
  23. package/dist/types/packages/sdk/src/model-taskset-runs-contracts.d.ts.map +1 -1
  24. package/dist/types/packages/sdk/src/model-taskset-runs.d.ts +129 -5
  25. package/dist/types/packages/sdk/src/model-taskset-runs.d.ts.map +1 -1
  26. package/dist/types/packages/sdk/src/taskset-authored-files.d.ts.map +1 -1
  27. package/package.json +2 -2
  28. package/dist/chunk-KLX6OVHR.js.map +0 -7
  29. package/dist/chunk-ZV2BLO3D.js.map +0 -7
@@ -89,6 +89,44 @@ export declare const SaveExperimentSchema: z.ZodObject<{
89
89
  topP: z.ZodDefault<z.ZodNumber>;
90
90
  }, z.core.$strict>, z.ZodObject<{
91
91
  kind: z.ZodLiteral<"fixture">;
92
+ }, z.core.$strict>, z.ZodObject<{
93
+ kind: z.ZodLiteral<"hosted_harness">;
94
+ modelId: z.ZodString;
95
+ reasoningEffort: z.ZodOptional<z.ZodEnum<{
96
+ high: "high";
97
+ low: "low";
98
+ off: "off";
99
+ medium: "medium";
100
+ xhigh: "xhigh";
101
+ max: "max";
102
+ }>>;
103
+ profileRepositoryId: z.ZodString;
104
+ source: z.ZodObject<{
105
+ profileId: z.ZodString;
106
+ sourceRevision: z.ZodString;
107
+ harnessRelease: z.ZodObject<{
108
+ id: z.ZodString;
109
+ contentHash: z.ZodString;
110
+ }, z.core.$strict>;
111
+ catalogHash: z.ZodString;
112
+ definitionId: z.ZodString;
113
+ definitionHash: z.ZodString;
114
+ target: z.ZodDiscriminatedUnion<[z.ZodObject<{
115
+ kind: z.ZodLiteral<"profile">;
116
+ }, z.core.$strict>, z.ZodObject<{
117
+ kind: z.ZodLiteral<"workflow">;
118
+ workflowId: z.ZodString;
119
+ }, z.core.$strict>, z.ZodObject<{
120
+ kind: z.ZodLiteral<"skill">;
121
+ skillPath: z.ZodString;
122
+ }, z.core.$strict>, z.ZodObject<{
123
+ kind: z.ZodLiteral<"agent_action">;
124
+ actionId: z.ZodString;
125
+ }, z.core.$strict>], "kind">;
126
+ environmentHash: z.ZodString;
127
+ }, z.core.$strict>;
128
+ modelConfigurationHash: z.ZodString;
129
+ packageHash: z.ZodString;
92
130
  }, z.core.$strict>], "kind">;
93
131
  population: z.ZodArray<z.ZodObject<{
94
132
  receiptId: z.ZodString;
@@ -158,6 +196,44 @@ export declare const ExperimentDefinitionSchema: z.ZodObject<{
158
196
  topP: z.ZodDefault<z.ZodNumber>;
159
197
  }, z.core.$strict>, z.ZodObject<{
160
198
  kind: z.ZodLiteral<"fixture">;
199
+ }, z.core.$strict>, z.ZodObject<{
200
+ kind: z.ZodLiteral<"hosted_harness">;
201
+ modelId: z.ZodString;
202
+ reasoningEffort: z.ZodOptional<z.ZodEnum<{
203
+ high: "high";
204
+ low: "low";
205
+ off: "off";
206
+ medium: "medium";
207
+ xhigh: "xhigh";
208
+ max: "max";
209
+ }>>;
210
+ profileRepositoryId: z.ZodString;
211
+ source: z.ZodObject<{
212
+ profileId: z.ZodString;
213
+ sourceRevision: z.ZodString;
214
+ harnessRelease: z.ZodObject<{
215
+ id: z.ZodString;
216
+ contentHash: z.ZodString;
217
+ }, z.core.$strict>;
218
+ catalogHash: z.ZodString;
219
+ definitionId: z.ZodString;
220
+ definitionHash: z.ZodString;
221
+ target: z.ZodDiscriminatedUnion<[z.ZodObject<{
222
+ kind: z.ZodLiteral<"profile">;
223
+ }, z.core.$strict>, z.ZodObject<{
224
+ kind: z.ZodLiteral<"workflow">;
225
+ workflowId: z.ZodString;
226
+ }, z.core.$strict>, z.ZodObject<{
227
+ kind: z.ZodLiteral<"skill">;
228
+ skillPath: z.ZodString;
229
+ }, z.core.$strict>, z.ZodObject<{
230
+ kind: z.ZodLiteral<"agent_action">;
231
+ actionId: z.ZodString;
232
+ }, z.core.$strict>], "kind">;
233
+ environmentHash: z.ZodString;
234
+ }, z.core.$strict>;
235
+ modelConfigurationHash: z.ZodString;
236
+ packageHash: z.ZodString;
161
237
  }, z.core.$strict>], "kind">;
162
238
  population: z.ZodArray<z.ZodObject<{
163
239
  receiptId: z.ZodString;
@@ -251,6 +327,244 @@ export declare const ExperimentExecutionContextSchema: z.ZodObject<{
251
327
  }, z.core.$strict>>;
252
328
  }, z.core.$strict>;
253
329
  export type ExperimentExecutionContext = z.infer<typeof ExperimentExecutionContextSchema>;
330
+ export declare const PrepareHarnessExperimentSchema: z.ZodObject<{
331
+ operationId: z.ZodString;
332
+ profileRepositoryId: z.ZodString;
333
+ definitionId: z.ZodString;
334
+ modelId: z.ZodString;
335
+ reasoningEffort: z.ZodOptional<z.ZodEnum<{
336
+ high: "high";
337
+ low: "low";
338
+ off: "off";
339
+ medium: "medium";
340
+ xhigh: "xhigh";
341
+ max: "max";
342
+ }>>;
343
+ maximumCostUsd: z.ZodNumber;
344
+ }, z.core.$strict>;
345
+ export declare const PreparedHarnessExperimentSchema: z.ZodObject<{
346
+ request: z.ZodObject<{
347
+ schemaVersion: z.ZodLiteral<"openpond.modelTasksetRunRequest.v1">;
348
+ operationId: z.ZodString;
349
+ teamId: z.ZodString;
350
+ modelProjectId: z.ZodNullable<z.ZodString>;
351
+ name: z.ZodOptional<z.ZodString>;
352
+ project: z.ZodOptional<z.ZodObject<{
353
+ id: z.ZodString;
354
+ revision: z.ZodNumber;
355
+ contentHash: z.ZodString;
356
+ targetId: z.ZodNullable<z.ZodString>;
357
+ }, z.core.$strict>>;
358
+ taskset: z.ZodObject<{
359
+ id: z.ZodString;
360
+ contentHash: z.ZodString;
361
+ revision: z.ZodNumber;
362
+ }, z.core.$strict>;
363
+ policy: z.ZodDiscriminatedUnion<[z.ZodObject<{
364
+ kind: z.ZodLiteral<"hosted_chat">;
365
+ modelId: z.ZodString;
366
+ messages: z.ZodOptional<z.ZodArray<z.ZodObject<{
367
+ role: z.ZodEnum<{
368
+ user: "user";
369
+ system: "system";
370
+ assistant: "assistant";
371
+ }>;
372
+ content: z.ZodString;
373
+ }, z.core.$strict>>>;
374
+ maxOutputTokens: z.ZodDefault<z.ZodNumber>;
375
+ temperature: z.ZodDefault<z.ZodNumber>;
376
+ topP: z.ZodDefault<z.ZodNumber>;
377
+ }, z.core.$strict>, z.ZodObject<{
378
+ kind: z.ZodLiteral<"fixture">;
379
+ }, z.core.$strict>, z.ZodObject<{
380
+ kind: z.ZodLiteral<"hosted_harness">;
381
+ modelId: z.ZodString;
382
+ reasoningEffort: z.ZodOptional<z.ZodEnum<{
383
+ high: "high";
384
+ low: "low";
385
+ off: "off";
386
+ medium: "medium";
387
+ xhigh: "xhigh";
388
+ max: "max";
389
+ }>>;
390
+ profileRepositoryId: z.ZodString;
391
+ source: z.ZodObject<{
392
+ profileId: z.ZodString;
393
+ sourceRevision: z.ZodString;
394
+ harnessRelease: z.ZodObject<{
395
+ id: z.ZodString;
396
+ contentHash: z.ZodString;
397
+ }, z.core.$strict>;
398
+ catalogHash: z.ZodString;
399
+ definitionId: z.ZodString;
400
+ definitionHash: z.ZodString;
401
+ target: z.ZodDiscriminatedUnion<[z.ZodObject<{
402
+ kind: z.ZodLiteral<"profile">;
403
+ }, z.core.$strict>, z.ZodObject<{
404
+ kind: z.ZodLiteral<"workflow">;
405
+ workflowId: z.ZodString;
406
+ }, z.core.$strict>, z.ZodObject<{
407
+ kind: z.ZodLiteral<"skill">;
408
+ skillPath: z.ZodString;
409
+ }, z.core.$strict>, z.ZodObject<{
410
+ kind: z.ZodLiteral<"agent_action">;
411
+ actionId: z.ZodString;
412
+ }, z.core.$strict>], "kind">;
413
+ environmentHash: z.ZodString;
414
+ }, z.core.$strict>;
415
+ modelConfigurationHash: z.ZodString;
416
+ packageHash: z.ZodString;
417
+ }, z.core.$strict>], "kind">;
418
+ population: z.ZodArray<z.ZodObject<{
419
+ receiptId: z.ZodString;
420
+ taskId: z.ZodString;
421
+ seed: z.ZodString;
422
+ fixtureId: z.ZodNullable<z.ZodString>;
423
+ }, z.core.$strict>>;
424
+ }, z.core.$strict>;
425
+ manifest: z.ZodObject<{
426
+ schemaVersion: z.ZodLiteral<"openpond.tasksetRunManifest.v1">;
427
+ id: z.ZodString;
428
+ tasksetRelease: z.ZodObject<{
429
+ id: z.ZodString;
430
+ contentHash: z.ZodString;
431
+ }, z.core.$strict>;
432
+ packageHash: z.ZodString;
433
+ execution: z.ZodDiscriminatedUnion<[z.ZodObject<{
434
+ kind: z.ZodLiteral<"taskset">;
435
+ environmentRelease: z.ZodObject<{
436
+ id: z.ZodString;
437
+ contentHash: z.ZodString;
438
+ }, z.core.$strict>;
439
+ verifierSetRelease: z.ZodObject<{
440
+ id: z.ZodString;
441
+ contentHash: z.ZodString;
442
+ }, z.core.$strict>;
443
+ policyHash: z.ZodString;
444
+ }, z.core.$strict>, z.ZodObject<{
445
+ kind: z.ZodLiteral<"harness">;
446
+ harnessRelease: z.ZodObject<{
447
+ id: z.ZodString;
448
+ contentHash: z.ZodString;
449
+ }, z.core.$strict>;
450
+ }, z.core.$strict>], "kind">;
451
+ profileEvaluation: z.ZodOptional<z.ZodObject<{
452
+ profileId: z.ZodString;
453
+ sourceRevision: z.ZodString;
454
+ harnessRelease: z.ZodObject<{
455
+ id: z.ZodString;
456
+ contentHash: z.ZodString;
457
+ }, z.core.$strict>;
458
+ catalogHash: z.ZodString;
459
+ definitionId: z.ZodString;
460
+ definitionHash: z.ZodString;
461
+ target: z.ZodDiscriminatedUnion<[z.ZodObject<{
462
+ kind: z.ZodLiteral<"profile">;
463
+ }, z.core.$strict>, z.ZodObject<{
464
+ kind: z.ZodLiteral<"workflow">;
465
+ workflowId: z.ZodString;
466
+ }, z.core.$strict>, z.ZodObject<{
467
+ kind: z.ZodLiteral<"skill">;
468
+ skillPath: z.ZodString;
469
+ }, z.core.$strict>, z.ZodObject<{
470
+ kind: z.ZodLiteral<"agent_action">;
471
+ actionId: z.ZodString;
472
+ }, z.core.$strict>], "kind">;
473
+ environmentHash: z.ZodString;
474
+ }, z.core.$strict>>;
475
+ policy: z.ZodDiscriminatedUnion<[z.ZodObject<{
476
+ kind: z.ZodLiteral<"model">;
477
+ model: z.ZodObject<{
478
+ provider: z.ZodString;
479
+ model: z.ZodString;
480
+ revision: z.ZodDefault<z.ZodNullable<z.ZodString>>;
481
+ artifactHash: z.ZodDefault<z.ZodNullable<z.ZodString>>;
482
+ tokenizerRevision: z.ZodDefault<z.ZodNullable<z.ZodString>>;
483
+ chatTemplateHash: z.ZodDefault<z.ZodNullable<z.ZodString>>;
484
+ }, z.core.$strict>;
485
+ configurationHash: z.ZodString;
486
+ }, z.core.$strict>, z.ZodObject<{
487
+ kind: z.ZodLiteral<"fixture">;
488
+ }, z.core.$strict>], "kind">;
489
+ gradingRole: z.ZodLiteral<"evaluation">;
490
+ metricPolicy: z.ZodObject<{
491
+ schemaVersion: z.ZodLiteral<"openpond.tasksetMetricPolicy.v1">;
492
+ primaryMetric: z.ZodString;
493
+ aggregation: z.ZodEnum<{
494
+ mean_score: "mean_score";
495
+ pass_rate: "pass_rate";
496
+ weighted_mean: "weighted_mean";
497
+ custom: "custom";
498
+ }>;
499
+ missingReward: z.ZodEnum<{
500
+ zero: "zero";
501
+ exclude: "exclude";
502
+ }>;
503
+ taskWeights: z.ZodOptional<z.ZodRecord<z.ZodString, z.ZodNumber>>;
504
+ customAggregator: z.ZodNullable<z.ZodObject<{
505
+ module: z.ZodString;
506
+ exportName: z.ZodString;
507
+ contentHash: z.ZodString;
508
+ timeoutMs: z.ZodNumber;
509
+ networkPolicy: z.ZodLiteral<"none">;
510
+ }, z.core.$strict>>;
511
+ }, z.core.$strict>;
512
+ population: z.ZodArray<z.ZodObject<{
513
+ receiptId: z.ZodString;
514
+ taskId: z.ZodString;
515
+ seed: z.ZodString;
516
+ fixtureId: z.ZodNullable<z.ZodString>;
517
+ }, z.core.$strict>>;
518
+ runtimeTarget: z.ZodObject<{
519
+ adapterId: z.ZodString;
520
+ placement: z.ZodEnum<{
521
+ local: "local";
522
+ remote: "remote";
523
+ colocated: "colocated";
524
+ provider_native: "provider_native";
525
+ }>;
526
+ runtimeVersion: z.ZodString;
527
+ capabilityReceipt: z.ZodString;
528
+ }, z.core.$strict>;
529
+ limits: z.ZodObject<{
530
+ maxTurns: z.ZodNumber;
531
+ timeoutMs: z.ZodNumber;
532
+ maxOutputBytes: z.ZodNumber;
533
+ maximumSpendUsd: z.ZodNullable<z.ZodNumber>;
534
+ }, z.core.$strict>;
535
+ createdAt: z.ZodString;
536
+ metadata: z.ZodDefault<z.ZodRecord<z.ZodString, z.ZodUnknown>>;
537
+ contentHash: z.ZodString;
538
+ }, z.core.$strict>;
539
+ maximumCostUsd: z.ZodNumber;
540
+ graders: z.ZodArray<z.ZodObject<{
541
+ id: z.ZodString;
542
+ version: z.ZodString;
543
+ contentHash: z.ZodString;
544
+ feedbackKey: z.ZodString;
545
+ release: z.ZodNullable<z.ZodObject<{
546
+ id: z.ZodString;
547
+ revision: z.ZodNumber;
548
+ contentHash: z.ZodString;
549
+ }, z.core.$strict>>;
550
+ mappings: z.ZodOptional<z.ZodArray<z.ZodObject<{
551
+ destination: z.ZodEnum<{
552
+ input: "input";
553
+ output: "output";
554
+ expectedOutput: "expectedOutput";
555
+ evaluatorContext: "evaluatorContext";
556
+ }>;
557
+ path: z.ZodArray<z.ZodString>;
558
+ source: z.ZodEnum<{
559
+ input: "input";
560
+ output: "output";
561
+ expectedOutput: "expectedOutput";
562
+ evaluatorContext: "evaluatorContext";
563
+ }>;
564
+ sourcePath: z.ZodArray<z.ZodString>;
565
+ }, z.core.$strict>>>;
566
+ }, z.core.$strict>>;
567
+ }, z.core.$strict>;
254
568
  export declare function verifyExperimentDefinition(value: unknown): ExperimentDefinition;
255
569
  export declare const experimentDefinitionRef: (value: ExperimentDefinition) => ExperimentDefinitionRef;
256
570
  //# sourceMappingURL=experiment-contracts.d.ts.map
@@ -1 +1 @@
1
- {"version":3,"file":"experiment-contracts.d.ts","sourceRoot":"","sources":["../../../../../src/experiment-contracts.ts"],"names":[],"mappings":"AAGA,OAAO,EAAE,CAAC,EAAE,MAAM,KAAK,CAAC;AAKxB,eAAO,MAAM,6BAA6B;;;;kBAE/B,CAAC;AACZ,eAAO,MAAM,yBAAyB;;;;;;;;;;;;;;;;;;;;;;;;;;kBAK3B,CAAC;AACZ,eAAO,MAAM,+BAA+B;;;;;;;;;;;;;;;;;;;;kBAAiI,CAAC;AAC9K,eAAO,MAAM,oBAAoB;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;kBAW/B,CAAC;AASH,eAAO,MAAM,0BAA0B;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;kBAAkD,CAAC;AAC1F,MAAM,MAAM,oBAAoB,GAAG,CAAC,CAAC,KAAK,CAAC,OAAO,0BAA0B,CAAC,CAAC;AAC9E,MAAM,MAAM,uBAAuB,GAAG,CAAC,CAAC,KAAK,CAAC,OAAO,6BAA6B,CAAC,CAAC;AACpF,eAAO,MAAM,qBAAqB;;;;;;;kBAAoF,CAAC;AACvH,eAAO,MAAM,yBAAyB;;;;;;kBAG3B,CAAC;AACZ,eAAO,MAAM,gCAAgC;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;kBAGlC,CAAC;AACZ,MAAM,MAAM,0BAA0B,GAAG,CAAC,CAAC,KAAK,CAAC,OAAO,gCAAgC,CAAC,CAAC;AAE1F,wBAAgB,0BAA0B,CAAC,KAAK,EAAE,OAAO,GAAG,oBAAoB,CAQ/E;AAED,eAAO,MAAM,uBAAuB,GAAI,OAAO,oBAAoB,KAAG,uBAEpE,CAAC"}
1
+ {"version":3,"file":"experiment-contracts.d.ts","sourceRoot":"","sources":["../../../../../src/experiment-contracts.ts"],"names":[],"mappings":"AAIA,OAAO,EAAE,CAAC,EAAE,MAAM,KAAK,CAAC;AAKxB,eAAO,MAAM,6BAA6B;;;;kBAE/B,CAAC;AACZ,eAAO,MAAM,yBAAyB;;;;;;;;;;;;;;;;;;;;;;;;;;kBAK3B,CAAC;AACZ,eAAO,MAAM,+BAA+B;;;;;;;;;;;;;;;;;;;;kBAAiI,CAAC;AAC9K,eAAO,MAAM,oBAAoB;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;kBAW/B,CAAC;AASH,eAAO,MAAM,0BAA0B;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;kBAAkD,CAAC;AAC1F,MAAM,MAAM,oBAAoB,GAAG,CAAC,CAAC,KAAK,CAAC,OAAO,0BAA0B,CAAC,CAAC;AAC9E,MAAM,MAAM,uBAAuB,GAAG,CAAC,CAAC,KAAK,CAAC,OAAO,6BAA6B,CAAC,CAAC;AACpF,eAAO,MAAM,qBAAqB;;;;;;;kBAAoF,CAAC;AACvH,eAAO,MAAM,yBAAyB;;;;;;kBAG3B,CAAC;AACZ,eAAO,MAAM,gCAAgC;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;kBAGlC,CAAC;AACZ,MAAM,MAAM,0BAA0B,GAAG,CAAC,CAAC,KAAK,CAAC,OAAO,gCAAgC,CAAC,CAAC;AAE1F,eAAO,MAAM,8BAA8B;;;;;;;;;;;;;;kBAIhC,CAAC;AACZ,eAAO,MAAM,+BAA+B;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;kBAIjC,CAAC;AAEZ,wBAAgB,0BAA0B,CAAC,KAAK,EAAE,OAAO,GAAG,oBAAoB,CAQ/E;AAED,eAAO,MAAM,uBAAuB,GAAI,OAAO,oBAAoB,KAAG,uBAEpE,CAAC"}
@@ -3,4 +3,5 @@ export * from "./experiment-client.js";
3
3
  export * from "./experiment-scoring-contracts.js";
4
4
  export * from "./experiment-field-mappings.js";
5
5
  export * from "./experiment-grading-contracts.js";
6
+ export * from "./harness-experiment-policy.js";
6
7
  //# sourceMappingURL=experiments.d.ts.map
@@ -1 +1 @@
1
- {"version":3,"file":"experiments.d.ts","sourceRoot":"","sources":["../../../../../src/experiments.ts"],"names":[],"mappings":"AAAA,cAAc,2BAA2B,CAAC;AAC1C,cAAc,wBAAwB,CAAC;AACvC,cAAc,mCAAmC,CAAC;AAClD,cAAc,gCAAgC,CAAC;AAC/C,cAAc,mCAAmC,CAAC"}
1
+ {"version":3,"file":"experiments.d.ts","sourceRoot":"","sources":["../../../../../src/experiments.ts"],"names":[],"mappings":"AAAA,cAAc,2BAA2B,CAAC;AAC1C,cAAc,wBAAwB,CAAC;AACvC,cAAc,mCAAmC,CAAC;AAClD,cAAc,gCAAgC,CAAC;AAC/C,cAAc,mCAAmC,CAAC;AAClD,cAAc,gCAAgC,CAAC"}
@@ -0,0 +1,44 @@
1
+ import { z } from "zod";
2
+ /** The host admits this released component through its native Harness engine.
3
+ * Chat prompt/settings are owned by the released source, not replaced here. */
4
+ export declare const HarnessExperimentPolicySchema: z.ZodObject<{
5
+ kind: z.ZodLiteral<"hosted_harness">;
6
+ modelId: z.ZodString;
7
+ reasoningEffort: z.ZodOptional<z.ZodEnum<{
8
+ high: "high";
9
+ low: "low";
10
+ off: "off";
11
+ medium: "medium";
12
+ xhigh: "xhigh";
13
+ max: "max";
14
+ }>>;
15
+ profileRepositoryId: z.ZodString;
16
+ source: z.ZodObject<{
17
+ profileId: z.ZodString;
18
+ sourceRevision: z.ZodString;
19
+ harnessRelease: z.ZodObject<{
20
+ id: z.ZodString;
21
+ contentHash: z.ZodString;
22
+ }, z.core.$strict>;
23
+ catalogHash: z.ZodString;
24
+ definitionId: z.ZodString;
25
+ definitionHash: z.ZodString;
26
+ target: z.ZodDiscriminatedUnion<[z.ZodObject<{
27
+ kind: z.ZodLiteral<"profile">;
28
+ }, z.core.$strict>, z.ZodObject<{
29
+ kind: z.ZodLiteral<"workflow">;
30
+ workflowId: z.ZodString;
31
+ }, z.core.$strict>, z.ZodObject<{
32
+ kind: z.ZodLiteral<"skill">;
33
+ skillPath: z.ZodString;
34
+ }, z.core.$strict>, z.ZodObject<{
35
+ kind: z.ZodLiteral<"agent_action">;
36
+ actionId: z.ZodString;
37
+ }, z.core.$strict>], "kind">;
38
+ environmentHash: z.ZodString;
39
+ }, z.core.$strict>;
40
+ modelConfigurationHash: z.ZodString;
41
+ packageHash: z.ZodString;
42
+ }, z.core.$strict>;
43
+ export type HarnessExperimentPolicy = z.infer<typeof HarnessExperimentPolicySchema>;
44
+ //# sourceMappingURL=harness-experiment-policy.d.ts.map
@@ -0,0 +1 @@
1
+ {"version":3,"file":"harness-experiment-policy.d.ts","sourceRoot":"","sources":["../../../../../src/harness-experiment-policy.ts"],"names":[],"mappings":"AAAA,OAAO,EAAE,CAAC,EAAE,MAAM,KAAK,CAAC;AAMxB;+EAC+E;AAC/E,eAAO,MAAM,6BAA6B;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;kBAQ/B,CAAC;AACZ,MAAM,MAAM,uBAAuB,GAAG,CAAC,CAAC,KAAK,CAAC,OAAO,6BAA6B,CAAC,CAAC"}
@@ -1,4 +1,6 @@
1
1
  import { z } from "zod";
2
+ import { TasksetRunManifestSchema } from "@openpond/evals/metrics";
3
+ export { HarnessExperimentPolicySchema, type HarnessExperimentPolicy } from "./harness-experiment-policy.js";
2
4
  export declare const ModelTasksetRunPolicySchema: z.ZodDiscriminatedUnion<[z.ZodObject<{
3
5
  kind: z.ZodLiteral<"hosted_chat">;
4
6
  modelId: z.ZodString;
@@ -15,6 +17,44 @@ export declare const ModelTasksetRunPolicySchema: z.ZodDiscriminatedUnion<[z.Zod
15
17
  topP: z.ZodDefault<z.ZodNumber>;
16
18
  }, z.core.$strict>, z.ZodObject<{
17
19
  kind: z.ZodLiteral<"fixture">;
20
+ }, z.core.$strict>, z.ZodObject<{
21
+ kind: z.ZodLiteral<"hosted_harness">;
22
+ modelId: z.ZodString;
23
+ reasoningEffort: z.ZodOptional<z.ZodEnum<{
24
+ high: "high";
25
+ low: "low";
26
+ off: "off";
27
+ medium: "medium";
28
+ xhigh: "xhigh";
29
+ max: "max";
30
+ }>>;
31
+ profileRepositoryId: z.ZodString;
32
+ source: z.ZodObject<{
33
+ profileId: z.ZodString;
34
+ sourceRevision: z.ZodString;
35
+ harnessRelease: z.ZodObject<{
36
+ id: z.ZodString;
37
+ contentHash: z.ZodString;
38
+ }, z.core.$strict>;
39
+ catalogHash: z.ZodString;
40
+ definitionId: z.ZodString;
41
+ definitionHash: z.ZodString;
42
+ target: z.ZodDiscriminatedUnion<[z.ZodObject<{
43
+ kind: z.ZodLiteral<"profile">;
44
+ }, z.core.$strict>, z.ZodObject<{
45
+ kind: z.ZodLiteral<"workflow">;
46
+ workflowId: z.ZodString;
47
+ }, z.core.$strict>, z.ZodObject<{
48
+ kind: z.ZodLiteral<"skill">;
49
+ skillPath: z.ZodString;
50
+ }, z.core.$strict>, z.ZodObject<{
51
+ kind: z.ZodLiteral<"agent_action">;
52
+ actionId: z.ZodString;
53
+ }, z.core.$strict>], "kind">;
54
+ environmentHash: z.ZodString;
55
+ }, z.core.$strict>;
56
+ modelConfigurationHash: z.ZodString;
57
+ packageHash: z.ZodString;
18
58
  }, z.core.$strict>], "kind">;
19
59
  export declare const ModelTasksetRunRequestSchema: z.ZodObject<{
20
60
  schemaVersion: z.ZodLiteral<"openpond.modelTasksetRunRequest.v1">;
@@ -49,6 +89,44 @@ export declare const ModelTasksetRunRequestSchema: z.ZodObject<{
49
89
  topP: z.ZodDefault<z.ZodNumber>;
50
90
  }, z.core.$strict>, z.ZodObject<{
51
91
  kind: z.ZodLiteral<"fixture">;
92
+ }, z.core.$strict>, z.ZodObject<{
93
+ kind: z.ZodLiteral<"hosted_harness">;
94
+ modelId: z.ZodString;
95
+ reasoningEffort: z.ZodOptional<z.ZodEnum<{
96
+ high: "high";
97
+ low: "low";
98
+ off: "off";
99
+ medium: "medium";
100
+ xhigh: "xhigh";
101
+ max: "max";
102
+ }>>;
103
+ profileRepositoryId: z.ZodString;
104
+ source: z.ZodObject<{
105
+ profileId: z.ZodString;
106
+ sourceRevision: z.ZodString;
107
+ harnessRelease: z.ZodObject<{
108
+ id: z.ZodString;
109
+ contentHash: z.ZodString;
110
+ }, z.core.$strict>;
111
+ catalogHash: z.ZodString;
112
+ definitionId: z.ZodString;
113
+ definitionHash: z.ZodString;
114
+ target: z.ZodDiscriminatedUnion<[z.ZodObject<{
115
+ kind: z.ZodLiteral<"profile">;
116
+ }, z.core.$strict>, z.ZodObject<{
117
+ kind: z.ZodLiteral<"workflow">;
118
+ workflowId: z.ZodString;
119
+ }, z.core.$strict>, z.ZodObject<{
120
+ kind: z.ZodLiteral<"skill">;
121
+ skillPath: z.ZodString;
122
+ }, z.core.$strict>, z.ZodObject<{
123
+ kind: z.ZodLiteral<"agent_action">;
124
+ actionId: z.ZodString;
125
+ }, z.core.$strict>], "kind">;
126
+ environmentHash: z.ZodString;
127
+ }, z.core.$strict>;
128
+ modelConfigurationHash: z.ZodString;
129
+ packageHash: z.ZodString;
52
130
  }, z.core.$strict>], "kind">;
53
131
  population: z.ZodArray<z.ZodObject<{
54
132
  receiptId: z.ZodString;
@@ -80,6 +158,7 @@ export declare const ModelTasksetRunSummarySchema: z.ZodObject<{
80
158
  policyKind: z.ZodEnum<{
81
159
  fixture: "fixture";
82
160
  hosted_chat: "hosted_chat";
161
+ hosted_harness: "hosted_harness";
83
162
  }>;
84
163
  manifestHash: z.ZodString;
85
164
  status: z.ZodEnum<{
@@ -134,6 +213,7 @@ export declare const ModelTasksetRunDetailsSchema: z.ZodObject<{
134
213
  policyKind: z.ZodEnum<{
135
214
  fixture: "fixture";
136
215
  hosted_chat: "hosted_chat";
216
+ hosted_harness: "hosted_harness";
137
217
  }>;
138
218
  manifestHash: z.ZodString;
139
219
  status: z.ZodEnum<{
@@ -197,6 +277,44 @@ export declare const ModelTasksetRunDetailsSchema: z.ZodObject<{
197
277
  topP: z.ZodDefault<z.ZodNumber>;
198
278
  }, z.core.$strict>, z.ZodObject<{
199
279
  kind: z.ZodLiteral<"fixture">;
280
+ }, z.core.$strict>, z.ZodObject<{
281
+ kind: z.ZodLiteral<"hosted_harness">;
282
+ modelId: z.ZodString;
283
+ reasoningEffort: z.ZodOptional<z.ZodEnum<{
284
+ high: "high";
285
+ low: "low";
286
+ off: "off";
287
+ medium: "medium";
288
+ xhigh: "xhigh";
289
+ max: "max";
290
+ }>>;
291
+ profileRepositoryId: z.ZodString;
292
+ source: z.ZodObject<{
293
+ profileId: z.ZodString;
294
+ sourceRevision: z.ZodString;
295
+ harnessRelease: z.ZodObject<{
296
+ id: z.ZodString;
297
+ contentHash: z.ZodString;
298
+ }, z.core.$strict>;
299
+ catalogHash: z.ZodString;
300
+ definitionId: z.ZodString;
301
+ definitionHash: z.ZodString;
302
+ target: z.ZodDiscriminatedUnion<[z.ZodObject<{
303
+ kind: z.ZodLiteral<"profile">;
304
+ }, z.core.$strict>, z.ZodObject<{
305
+ kind: z.ZodLiteral<"workflow">;
306
+ workflowId: z.ZodString;
307
+ }, z.core.$strict>, z.ZodObject<{
308
+ kind: z.ZodLiteral<"skill">;
309
+ skillPath: z.ZodString;
310
+ }, z.core.$strict>, z.ZodObject<{
311
+ kind: z.ZodLiteral<"agent_action">;
312
+ actionId: z.ZodString;
313
+ }, z.core.$strict>], "kind">;
314
+ environmentHash: z.ZodString;
315
+ }, z.core.$strict>;
316
+ modelConfigurationHash: z.ZodString;
317
+ packageHash: z.ZodString;
200
318
  }, z.core.$strict>], "kind">;
201
319
  population: z.ZodArray<z.ZodObject<{
202
320
  receiptId: z.ZodString;
@@ -356,6 +474,7 @@ export declare const ModelTasksetRunPageSchema: z.ZodObject<{
356
474
  policyKind: z.ZodEnum<{
357
475
  fixture: "fixture";
358
476
  hosted_chat: "hosted_chat";
477
+ hosted_harness: "hosted_harness";
359
478
  }>;
360
479
  manifestHash: z.ZodString;
361
480
  status: z.ZodEnum<{
@@ -413,6 +532,7 @@ export declare const ModelTasksetRunResultSchema: z.ZodObject<{
413
532
  policyKind: z.ZodEnum<{
414
533
  fixture: "fixture";
415
534
  hosted_chat: "hosted_chat";
535
+ hosted_harness: "hosted_harness";
416
536
  }>;
417
537
  manifestHash: z.ZodString;
418
538
  status: z.ZodEnum<{
@@ -476,6 +596,44 @@ export declare const ModelTasksetRunResultSchema: z.ZodObject<{
476
596
  topP: z.ZodDefault<z.ZodNumber>;
477
597
  }, z.core.$strict>, z.ZodObject<{
478
598
  kind: z.ZodLiteral<"fixture">;
599
+ }, z.core.$strict>, z.ZodObject<{
600
+ kind: z.ZodLiteral<"hosted_harness">;
601
+ modelId: z.ZodString;
602
+ reasoningEffort: z.ZodOptional<z.ZodEnum<{
603
+ high: "high";
604
+ low: "low";
605
+ off: "off";
606
+ medium: "medium";
607
+ xhigh: "xhigh";
608
+ max: "max";
609
+ }>>;
610
+ profileRepositoryId: z.ZodString;
611
+ source: z.ZodObject<{
612
+ profileId: z.ZodString;
613
+ sourceRevision: z.ZodString;
614
+ harnessRelease: z.ZodObject<{
615
+ id: z.ZodString;
616
+ contentHash: z.ZodString;
617
+ }, z.core.$strict>;
618
+ catalogHash: z.ZodString;
619
+ definitionId: z.ZodString;
620
+ definitionHash: z.ZodString;
621
+ target: z.ZodDiscriminatedUnion<[z.ZodObject<{
622
+ kind: z.ZodLiteral<"profile">;
623
+ }, z.core.$strict>, z.ZodObject<{
624
+ kind: z.ZodLiteral<"workflow">;
625
+ workflowId: z.ZodString;
626
+ }, z.core.$strict>, z.ZodObject<{
627
+ kind: z.ZodLiteral<"skill">;
628
+ skillPath: z.ZodString;
629
+ }, z.core.$strict>, z.ZodObject<{
630
+ kind: z.ZodLiteral<"agent_action">;
631
+ actionId: z.ZodString;
632
+ }, z.core.$strict>], "kind">;
633
+ environmentHash: z.ZodString;
634
+ }, z.core.$strict>;
635
+ modelConfigurationHash: z.ZodString;
636
+ packageHash: z.ZodString;
479
637
  }, z.core.$strict>], "kind">;
480
638
  population: z.ZodArray<z.ZodObject<{
481
639
  receiptId: z.ZodString;
@@ -691,6 +849,7 @@ export declare const ModelTasksetRunResultSchema: z.ZodObject<{
691
849
  contentHash: z.ZodString;
692
850
  }, z.core.$strict>;
693
851
  export type ModelTasksetRunResult = z.infer<typeof ModelTasksetRunResultSchema>;
852
+ export declare function verifyHarnessExperimentManifest(request: ModelTasksetRunRequest, manifest: z.infer<typeof TasksetRunManifestSchema>): void;
694
853
  /** Server artifact ownership authenticates the producer. These checks bind all
695
854
  * portable projections to the admitted request, population and policy. */
696
855
  export declare function verifyModelTasksetRunDetails(value: unknown): Promise<ModelTasksetRunDetails>;
@@ -1 +1 @@
1
- {"version":3,"file":"model-taskset-runs-contracts.d.ts","sourceRoot":"","sources":["../../../../../src/model-taskset-runs-contracts.ts"],"names":[],"mappings":"AAAA,OAAO,EAAE,CAAC,EAAE,MAAM,KAAK,CAAC;AAWxB,eAAO,MAAM,2BAA2B;;;;;;;;;;;;;;;;4BAGtC,CAAC;AACH,eAAO,MAAM,4BAA4B;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;kBAWvC,CAAC;AACH,MAAM,MAAM,sBAAsB,GAAG,CAAC,CAAC,KAAK,CAAC,OAAO,4BAA4B,CAAC,CAAC;AAClF,eAAO,MAAM,4BAA4B;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;kBAqBvC,CAAC;AACH,MAAM,MAAM,sBAAsB,GAAG,CAAC,CAAC,KAAK,CAAC,OAAO,4BAA4B,CAAC,CAAC;AAClF,eAAO,MAAM,4BAA4B;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;kBAK9B,CAAC;AACZ,MAAM,MAAM,sBAAsB,GAAG,CAAC,CAAC,KAAK,CAAC,OAAO,4BAA4B,CAAC,CAAC;AAClF,eAAO,MAAM,8BAA8B;;;;;kBAA+M,CAAC;AAC3P,eAAO,MAAM,yBAAyB;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;kBAAgH,CAAC;AACvJ,eAAO,MAAM,2BAA2B;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;kBAM7B,CAAC;AACZ,MAAM,MAAM,qBAAqB,GAAG,CAAC,CAAC,KAAK,CAAC,OAAO,2BAA2B,CAAC,CAAC;AAEhF;0EAC0E;AAC1E,wBAAsB,4BAA4B,CAAC,KAAK,EAAE,OAAO,GAAG,OAAO,CAAC,sBAAsB,CAAC,CAqBlG;AAED,wBAAsB,2BAA2B,CAAC,KAAK,EAAE,OAAO,GAAG,OAAO,CAAC,qBAAqB,CAAC,CAkBhG"}
1
+ {"version":3,"file":"model-taskset-runs-contracts.d.ts","sourceRoot":"","sources":["../../../../../src/model-taskset-runs-contracts.ts"],"names":[],"mappings":"AAAA,OAAO,EAAE,CAAC,EAAE,MAAM,KAAK,CAAC;AAGxB,OAAO,EAAE,wBAAwB,EAAyG,MAAM,yBAAyB,CAAC;AAM1K,OAAO,EAAE,6BAA6B,EAAE,KAAK,uBAAuB,EAAE,MAAM,gCAAgC,CAAC;AAI7G,eAAO,MAAM,2BAA2B;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;4BAItC,CAAC;AACH,eAAO,MAAM,4BAA4B;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;kBAgBvC,CAAC;AACH,MAAM,MAAM,sBAAsB,GAAG,CAAC,CAAC,KAAK,CAAC,OAAO,4BAA4B,CAAC,CAAC;AAClF,eAAO,MAAM,4BAA4B;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;kBAqBvC,CAAC;AACH,MAAM,MAAM,sBAAsB,GAAG,CAAC,CAAC,KAAK,CAAC,OAAO,4BAA4B,CAAC,CAAC;AAClF,eAAO,MAAM,4BAA4B;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;kBAK9B,CAAC;AACZ,MAAM,MAAM,sBAAsB,GAAG,CAAC,CAAC,KAAK,CAAC,OAAO,4BAA4B,CAAC,CAAC;AAClF,eAAO,MAAM,8BAA8B;;;;;kBAA+M,CAAC;AAC3P,eAAO,MAAM,yBAAyB;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;kBAAgH,CAAC;AACvJ,eAAO,MAAM,2BAA2B;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;kBAM7B,CAAC;AACZ,MAAM,MAAM,qBAAqB,GAAG,CAAC,CAAC,KAAK,CAAC,OAAO,2BAA2B,CAAC,CAAC;AAEhF,wBAAgB,+BAA+B,CAAC,OAAO,EAAE,sBAAsB,EAAE,QAAQ,EAAE,CAAC,CAAC,KAAK,CAAC,OAAO,wBAAwB,CAAC,QAgBlI;AAED;0EAC0E;AAC1E,wBAAsB,4BAA4B,CAAC,KAAK,EAAE,OAAO,GAAG,OAAO,CAAC,sBAAsB,CAAC,CAwBlG;AAED,wBAAsB,2BAA2B,CAAC,KAAK,EAAE,OAAO,GAAG,OAAO,CAAC,qBAAqB,CAAC,CAkBhG"}