openpond-sdk 0.12.26 → 0.12.27

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 (200) hide show
  1. package/PRODUCT_INTEGRATION.md +13 -6
  2. package/README.md +2 -10
  3. package/SERVERLESS_SERVING.md +52 -35
  4. package/TASK_INGESTION.md +0 -4
  5. package/dist/advanced-refiner-evaluations.js +7 -7
  6. package/dist/ai-sdk.js +20 -83
  7. package/dist/ai-sdk.js.map +1 -1
  8. package/dist/candidate-evaluations.js +6 -6
  9. package/dist/{chunk-UZAVG2WK.js → chunk-3HOGOWKU.js} +8 -6
  10. package/dist/chunk-3HOGOWKU.js.map +1 -0
  11. package/dist/{chunk-YV6PPDXZ.js → chunk-3Z5IK2UK.js} +27 -2
  12. package/dist/chunk-3Z5IK2UK.js.map +1 -0
  13. package/dist/{chunk-4REK5TTE.js → chunk-4D4QSQC3.js} +2 -2
  14. package/dist/{chunk-AMZ2SWHR.js → chunk-534W2O2S.js} +2 -2
  15. package/dist/{chunk-6LVLWRBJ.js → chunk-54BJJL7G.js} +2 -2
  16. package/dist/{chunk-XWUTYFOU.js → chunk-5KJY44FB.js} +3 -3
  17. package/dist/{chunk-QKLZZOZD.js → chunk-7MLFWVDI.js} +29 -10
  18. package/dist/chunk-7MLFWVDI.js.map +1 -0
  19. package/dist/{chunk-4A4FNG4R.js → chunk-7ZV626YS.js} +2 -2
  20. package/dist/{chunk-57ROW3XF.js → chunk-ABOZV47D.js} +2 -2
  21. package/dist/{chunk-3BILL33A.js → chunk-BIQ5TG3T.js} +4 -4
  22. package/dist/{chunk-JINWKUKR.js → chunk-BT6CC6QB.js} +3 -3
  23. package/dist/{chunk-4PEK3YOW.js → chunk-CNFN6GHY.js} +2 -2
  24. package/dist/{chunk-X5AT253K.js → chunk-CVYGJUWM.js} +2 -2
  25. package/dist/{chunk-D26RJQU7.js → chunk-CX7ETU2F.js} +2 -2
  26. package/dist/{chunk-4CDLPMJK.js → chunk-D7LV3USE.js} +2 -2
  27. package/dist/{chunk-WZKKQNYY.js → chunk-GA7OSSEA.js} +2 -2
  28. package/dist/{chunk-EL6KZO43.js → chunk-GFWXDWEE.js} +3 -3
  29. package/dist/{chunk-XYK3MAIC.js → chunk-JYWIKBGD.js} +122 -112
  30. package/dist/chunk-JYWIKBGD.js.map +1 -0
  31. package/dist/{chunk-AQV7RQZH.js → chunk-K4WU4JOP.js} +2 -2
  32. package/dist/{chunk-CAOMKM55.js → chunk-L7MVGGBH.js} +2 -2
  33. package/dist/{chunk-P56JWQYV.js → chunk-LRFVISLA.js} +3 -3
  34. package/dist/{chunk-2U6BQ5C4.js → chunk-MJHQMQEW.js} +3 -3
  35. package/dist/chunk-ON3KSXWR.js +639 -0
  36. package/dist/chunk-ON3KSXWR.js.map +1 -0
  37. package/dist/{chunk-PMVSO23F.js → chunk-RMXZISHK.js} +10 -1
  38. package/dist/chunk-RMXZISHK.js.map +1 -0
  39. package/dist/{chunk-JWEIAEQ6.js → chunk-TMNEDWA7.js} +10 -3
  40. package/dist/chunk-TMNEDWA7.js.map +1 -0
  41. package/dist/{chunk-C7I4ZG6Q.js → chunk-TX552FQG.js} +3 -3
  42. package/dist/{chunk-4CRLQMJ2.js → chunk-VIDRZCPQ.js} +6 -6
  43. package/dist/{chunk-MN4FE52A.js → chunk-VRUAIMM3.js} +7 -7
  44. package/dist/chunk-VRUAIMM3.js.map +1 -0
  45. package/dist/{chunk-H5KHLBC5.js → chunk-WJZHSIGD.js} +6 -8
  46. package/dist/{chunk-H5KHLBC5.js.map → chunk-WJZHSIGD.js.map} +1 -1
  47. package/dist/{chunk-HCUZYGUK.js → chunk-WPTFJZ4L.js} +3 -3
  48. package/dist/connected-evidence.js +7 -7
  49. package/dist/dataset-workspaces.js +10 -10
  50. package/dist/errors.js +1 -1
  51. package/dist/experiment-evaluation-schedules.js +16 -16
  52. package/dist/experiment-improvements.js +7 -7
  53. package/dist/experiments.js +15 -15
  54. package/dist/index.js +23 -23
  55. package/dist/learning.js +15 -4
  56. package/dist/learning.js.map +1 -1
  57. package/dist/model-projects.js +1 -1
  58. package/dist/model-starter-attempts.js +2 -2
  59. package/dist/model-starter-catalog.js +4 -4
  60. package/dist/model-starters.js +4 -4
  61. package/dist/model-taskset-runs.js +3 -3
  62. package/dist/post-training.js +6 -6
  63. package/dist/taskset-drafts.js +6 -6
  64. package/dist/taskset-packages.js +12 -10
  65. package/dist/training-bundle.js +46 -20
  66. package/dist/training-bundle.js.map +1 -1
  67. package/dist/training-hardware.js +2 -2
  68. package/dist/training.js +5620 -4342
  69. package/dist/training.js.map +1 -1
  70. package/dist/types/packages/sdk/src/ai-sdk.d.ts.map +1 -1
  71. package/dist/types/packages/sdk/src/alphaedit-model-serving.d.ts +8 -4
  72. package/dist/types/packages/sdk/src/alphaedit-model-serving.d.ts.map +1 -1
  73. package/dist/types/packages/sdk/src/candidate-evaluation-client.d.ts +162 -162
  74. package/dist/types/packages/sdk/src/decision-model-serving.d.ts +4 -2
  75. package/dist/types/packages/sdk/src/decision-model-serving.d.ts.map +1 -1
  76. package/dist/types/packages/sdk/src/engram-model-serving.d.ts +3 -1
  77. package/dist/types/packages/sdk/src/engram-model-serving.d.ts.map +1 -1
  78. package/dist/types/packages/sdk/src/experiment-client.d.ts +699 -691
  79. package/dist/types/packages/sdk/src/experiment-client.d.ts.map +1 -1
  80. package/dist/types/packages/sdk/src/experiment-draft-storage.d.ts +34 -28
  81. package/dist/types/packages/sdk/src/experiment-draft-storage.d.ts.map +1 -1
  82. package/dist/types/packages/sdk/src/experiment-improvement-client.d.ts +1889 -1889
  83. package/dist/types/packages/sdk/src/experiment-run-details.d.ts +7 -7
  84. package/dist/types/packages/sdk/src/experiment-scoring-contracts.d.ts +21 -3
  85. package/dist/types/packages/sdk/src/experiment-scoring-contracts.d.ts.map +1 -1
  86. package/dist/types/packages/sdk/src/learning.d.ts +1 -0
  87. package/dist/types/packages/sdk/src/learning.d.ts.map +1 -1
  88. package/dist/types/packages/sdk/src/managed-alphaedit-executions.d.ts +272 -16
  89. package/dist/types/packages/sdk/src/managed-alphaedit-executions.d.ts.map +1 -1
  90. package/dist/types/packages/sdk/src/managed-compute-price.d.ts +53 -1
  91. package/dist/types/packages/sdk/src/managed-compute-price.d.ts.map +1 -1
  92. package/dist/types/packages/sdk/src/managed-decision-executions.d.ts +612 -356
  93. package/dist/types/packages/sdk/src/managed-decision-executions.d.ts.map +1 -1
  94. package/dist/types/packages/sdk/src/managed-engram-executions.d.ts +272 -16
  95. package/dist/types/packages/sdk/src/managed-engram-executions.d.ts.map +1 -1
  96. package/dist/types/packages/sdk/src/managed-modification-executions.d.ts +208 -16
  97. package/dist/types/packages/sdk/src/managed-modification-executions.d.ts.map +1 -1
  98. package/dist/types/packages/sdk/src/managed-neuron-training-executions.d.ts +272 -16
  99. package/dist/types/packages/sdk/src/managed-neuron-training-executions.d.ts.map +1 -1
  100. package/dist/types/packages/sdk/src/managed-serving-usage.d.ts +3 -0
  101. package/dist/types/packages/sdk/src/managed-serving-usage.d.ts.map +1 -1
  102. package/dist/types/packages/sdk/src/managed-training-preparation.d.ts +25 -25
  103. package/dist/types/packages/sdk/src/managed-training-preparation.d.ts.map +1 -1
  104. package/dist/types/packages/sdk/src/model-learning-capabilities.d.ts +4 -0
  105. package/dist/types/packages/sdk/src/model-learning-capabilities.d.ts.map +1 -0
  106. package/dist/types/packages/sdk/src/model-learning-policy.d.ts.map +1 -1
  107. package/dist/types/packages/sdk/src/model-supervised-training-defaults.d.ts.map +1 -1
  108. package/dist/types/packages/sdk/src/model-training-defaults.d.ts +5 -2
  109. package/dist/types/packages/sdk/src/model-training-defaults.d.ts.map +1 -1
  110. package/dist/types/packages/sdk/src/neuron-training-model-serving.d.ts +8 -4
  111. package/dist/types/packages/sdk/src/neuron-training-model-serving.d.ts.map +1 -1
  112. package/dist/types/packages/sdk/src/post-training.d.ts +7 -7
  113. package/dist/types/packages/sdk/src/product-model-response.d.ts +6 -2
  114. package/dist/types/packages/sdk/src/product-model-response.d.ts.map +1 -1
  115. package/dist/types/packages/sdk/src/product-model-transport.d.ts +99 -0
  116. package/dist/types/packages/sdk/src/product-model-transport.d.ts.map +1 -0
  117. package/dist/types/packages/sdk/src/product-model-wire.d.ts +104 -0
  118. package/dist/types/packages/sdk/src/product-model-wire.d.ts.map +1 -0
  119. package/dist/types/packages/sdk/src/product-models.d.ts +18 -15
  120. package/dist/types/packages/sdk/src/product-models.d.ts.map +1 -1
  121. package/dist/types/packages/sdk/src/serverless-serving.d.ts +416 -9
  122. package/dist/types/packages/sdk/src/serverless-serving.d.ts.map +1 -1
  123. package/dist/types/packages/sdk/src/steering-model-serving.d.ts +2144 -703
  124. package/dist/types/packages/sdk/src/steering-model-serving.d.ts.map +1 -1
  125. package/dist/types/packages/sdk/src/steering-profile-publication.d.ts +631 -0
  126. package/dist/types/packages/sdk/src/steering-profile-publication.d.ts.map +1 -1
  127. package/dist/types/packages/sdk/src/taskset-draft-client.d.ts +120 -120
  128. package/dist/types/packages/sdk/src/taskset-package-projection.d.ts +20 -20
  129. package/dist/types/packages/sdk/src/taskset-packages.d.ts +1 -0
  130. package/dist/types/packages/sdk/src/taskset-packages.d.ts.map +1 -1
  131. package/dist/types/packages/sdk/src/taskset-text-policy.d.ts +11 -0
  132. package/dist/types/packages/sdk/src/taskset-text-policy.d.ts.map +1 -0
  133. package/dist/types/packages/sdk/src/training-compute-selection.d.ts +6 -19
  134. package/dist/types/packages/sdk/src/training-compute-selection.d.ts.map +1 -1
  135. package/dist/types/packages/sdk/src/training-estimates.d.ts +1 -1
  136. package/dist/types/packages/sdk/src/training-experiment-setup.d.ts +4 -1
  137. package/dist/types/packages/sdk/src/training-experiment-setup.d.ts.map +1 -1
  138. package/dist/types/packages/sdk/src/training-harness-sources.d.ts +2 -2
  139. package/dist/types/packages/sdk/src/training-learning-batch.d.ts +40 -40
  140. package/dist/types/packages/sdk/src/training-model-serving-control.d.ts +406 -8
  141. package/dist/types/packages/sdk/src/training-model-serving-control.d.ts.map +1 -1
  142. package/dist/types/packages/sdk/src/training-model-serving.d.ts +1136 -30
  143. package/dist/types/packages/sdk/src/training-model-serving.d.ts.map +1 -1
  144. package/dist/types/packages/sdk/src/training-operations.d.ts +11197 -0
  145. package/dist/types/packages/sdk/src/training-operations.d.ts.map +1 -0
  146. package/dist/types/packages/sdk/src/training-qualification-preparation.d.ts +25 -25
  147. package/dist/types/packages/sdk/src/training-supervised.d.ts +13 -0
  148. package/dist/types/packages/sdk/src/training-supervised.d.ts.map +1 -1
  149. package/dist/types/packages/sdk/src/training-transport.d.ts +4 -1
  150. package/dist/types/packages/sdk/src/training-transport.d.ts.map +1 -1
  151. package/dist/types/packages/sdk/src/training.d.ts +15885 -4662
  152. package/dist/types/packages/sdk/src/training.d.ts.map +1 -1
  153. package/dist/types/packages/sdk/src/weight-inference.d.ts +8 -5
  154. package/dist/types/packages/sdk/src/weight-inference.d.ts.map +1 -1
  155. package/dist/types/packages/sdk/src/weight-model-versions.d.ts +2089 -398
  156. package/dist/types/packages/sdk/src/weight-model-versions.d.ts.map +1 -1
  157. package/dist/types/packages/sdk/src/weight-space-artifact.d.ts +433 -0
  158. package/dist/types/packages/sdk/src/weight-space-artifact.d.ts.map +1 -0
  159. package/dist/types/packages/sdk/src/weight-space-evaluation.d.ts +1437 -0
  160. package/dist/types/packages/sdk/src/weight-space-evaluation.d.ts.map +1 -0
  161. package/dist/types/packages/sdk/src/weight-space-model-serving.d.ts +2910 -0
  162. package/dist/types/packages/sdk/src/weight-space-model-serving.d.ts.map +1 -0
  163. package/dist/types/packages/sdk/src/weight-space-programming.d.ts +845 -0
  164. package/dist/types/packages/sdk/src/weight-space-programming.d.ts.map +1 -0
  165. package/dist/types/packages/sdk/src/weight-space-sessions.d.ts +4600 -0
  166. package/dist/types/packages/sdk/src/weight-space-sessions.d.ts.map +1 -0
  167. package/fixtures/python-foundation/operation-matrix.json +369 -0
  168. package/fixtures/python-foundation/schemas.json +34304 -27905
  169. package/fixtures/python-foundation/training-estimate-http-workflow.json +1 -1
  170. package/package.json +2 -2
  171. package/dist/chunk-JWEIAEQ6.js.map +0 -1
  172. package/dist/chunk-MN4FE52A.js.map +0 -1
  173. package/dist/chunk-PMVSO23F.js.map +0 -1
  174. package/dist/chunk-QKLZZOZD.js.map +0 -1
  175. package/dist/chunk-UZAVG2WK.js.map +0 -1
  176. package/dist/chunk-XYK3MAIC.js.map +0 -1
  177. package/dist/chunk-YV6PPDXZ.js.map +0 -1
  178. package/dist/chunk-ZG4HVCJB.js +0 -248
  179. package/dist/chunk-ZG4HVCJB.js.map +0 -1
  180. /package/dist/{chunk-4REK5TTE.js.map → chunk-4D4QSQC3.js.map} +0 -0
  181. /package/dist/{chunk-AMZ2SWHR.js.map → chunk-534W2O2S.js.map} +0 -0
  182. /package/dist/{chunk-6LVLWRBJ.js.map → chunk-54BJJL7G.js.map} +0 -0
  183. /package/dist/{chunk-XWUTYFOU.js.map → chunk-5KJY44FB.js.map} +0 -0
  184. /package/dist/{chunk-4A4FNG4R.js.map → chunk-7ZV626YS.js.map} +0 -0
  185. /package/dist/{chunk-57ROW3XF.js.map → chunk-ABOZV47D.js.map} +0 -0
  186. /package/dist/{chunk-3BILL33A.js.map → chunk-BIQ5TG3T.js.map} +0 -0
  187. /package/dist/{chunk-JINWKUKR.js.map → chunk-BT6CC6QB.js.map} +0 -0
  188. /package/dist/{chunk-4PEK3YOW.js.map → chunk-CNFN6GHY.js.map} +0 -0
  189. /package/dist/{chunk-X5AT253K.js.map → chunk-CVYGJUWM.js.map} +0 -0
  190. /package/dist/{chunk-D26RJQU7.js.map → chunk-CX7ETU2F.js.map} +0 -0
  191. /package/dist/{chunk-4CDLPMJK.js.map → chunk-D7LV3USE.js.map} +0 -0
  192. /package/dist/{chunk-WZKKQNYY.js.map → chunk-GA7OSSEA.js.map} +0 -0
  193. /package/dist/{chunk-EL6KZO43.js.map → chunk-GFWXDWEE.js.map} +0 -0
  194. /package/dist/{chunk-AQV7RQZH.js.map → chunk-K4WU4JOP.js.map} +0 -0
  195. /package/dist/{chunk-CAOMKM55.js.map → chunk-L7MVGGBH.js.map} +0 -0
  196. /package/dist/{chunk-P56JWQYV.js.map → chunk-LRFVISLA.js.map} +0 -0
  197. /package/dist/{chunk-2U6BQ5C4.js.map → chunk-MJHQMQEW.js.map} +0 -0
  198. /package/dist/{chunk-C7I4ZG6Q.js.map → chunk-TX552FQG.js.map} +0 -0
  199. /package/dist/{chunk-4CRLQMJ2.js.map → chunk-VIDRZCPQ.js.map} +0 -0
  200. /package/dist/{chunk-HCUZYGUK.js.map → chunk-WPTFJZ4L.js.map} +0 -0
@@ -98,6 +98,60 @@ export declare class OpenPondExperimentsClient {
98
98
  revision: number;
99
99
  };
100
100
  policy: {
101
+ kind: "hosted_chat";
102
+ modelId: string;
103
+ localRuntime?: {
104
+ providerId: "custom-openai-compatible";
105
+ configurationHash: string;
106
+ maximumRequestCostUsd: number;
107
+ requestTimeoutMs: number;
108
+ } | {
109
+ providerId: "claude-code";
110
+ configurationHash: string;
111
+ executableHash: string;
112
+ version: string;
113
+ capabilityHash: string;
114
+ maximumRequestCostUsd: number;
115
+ requestTimeoutMs: number;
116
+ maximumTurns: number;
117
+ holdOpen: boolean;
118
+ } | {
119
+ providerId: "codex";
120
+ configurationHash: string;
121
+ accountHash: string;
122
+ reasoningEffort: "high" | "low" | "medium" | "xhigh";
123
+ maximumRequestCostUsd: number;
124
+ requestTimeoutMs: number;
125
+ } | undefined;
126
+ candidate?: {
127
+ artifactId: string;
128
+ contentHash: string;
129
+ preparationId: string;
130
+ dispatchId: string;
131
+ baseProfileId: string;
132
+ workerImageDigest: string;
133
+ } | undefined;
134
+ harness?: {
135
+ harnessRelease: {
136
+ id: string;
137
+ contentHash: string;
138
+ };
139
+ agentSnapshot: {
140
+ id: string;
141
+ contentHash: string;
142
+ };
143
+ sourcePackageHash: string;
144
+ } | undefined;
145
+ messages?: {
146
+ role: "assistant" | "system" | "user";
147
+ content: string;
148
+ }[] | undefined;
149
+ maxOutputTokens: number;
150
+ temperature: number;
151
+ topP: number;
152
+ } | {
153
+ kind: "fixture";
154
+ } | {
101
155
  kind: "hosted_harness";
102
156
  modelId: string;
103
157
  maxOutputTokens?: number | undefined;
@@ -339,60 +393,6 @@ export declare class OpenPondExperimentsClient {
339
393
  };
340
394
  contentHash: string;
341
395
  } | undefined;
342
- } | {
343
- kind: "hosted_chat";
344
- modelId: string;
345
- localRuntime?: {
346
- providerId: "claude-code";
347
- configurationHash: string;
348
- executableHash: string;
349
- version: string;
350
- capabilityHash: string;
351
- maximumRequestCostUsd: number;
352
- requestTimeoutMs: number;
353
- maximumTurns: number;
354
- holdOpen: boolean;
355
- } | {
356
- providerId: "codex";
357
- configurationHash: string;
358
- accountHash: string;
359
- reasoningEffort: "high" | "low" | "medium" | "xhigh";
360
- maximumRequestCostUsd: number;
361
- requestTimeoutMs: number;
362
- } | {
363
- providerId: "custom-openai-compatible";
364
- configurationHash: string;
365
- maximumRequestCostUsd: number;
366
- requestTimeoutMs: number;
367
- } | undefined;
368
- candidate?: {
369
- artifactId: string;
370
- contentHash: string;
371
- preparationId: string;
372
- dispatchId: string;
373
- baseProfileId: string;
374
- workerImageDigest: string;
375
- } | undefined;
376
- harness?: {
377
- harnessRelease: {
378
- id: string;
379
- contentHash: string;
380
- };
381
- agentSnapshot: {
382
- id: string;
383
- contentHash: string;
384
- };
385
- sourcePackageHash: string;
386
- } | undefined;
387
- messages?: {
388
- role: "assistant" | "system" | "user";
389
- content: string;
390
- }[] | undefined;
391
- maxOutputTokens: number;
392
- temperature: number;
393
- topP: number;
394
- } | {
395
- kind: "fixture";
396
396
  };
397
397
  population: {
398
398
  receiptId: string;
@@ -681,6 +681,60 @@ export declare class OpenPondExperimentsClient {
681
681
  revision: number;
682
682
  };
683
683
  policy: {
684
+ kind: "hosted_chat";
685
+ modelId: string;
686
+ localRuntime?: {
687
+ providerId: "custom-openai-compatible";
688
+ configurationHash: string;
689
+ maximumRequestCostUsd: number;
690
+ requestTimeoutMs: number;
691
+ } | {
692
+ providerId: "claude-code";
693
+ configurationHash: string;
694
+ executableHash: string;
695
+ version: string;
696
+ capabilityHash: string;
697
+ maximumRequestCostUsd: number;
698
+ requestTimeoutMs: number;
699
+ maximumTurns: number;
700
+ holdOpen: boolean;
701
+ } | {
702
+ providerId: "codex";
703
+ configurationHash: string;
704
+ accountHash: string;
705
+ reasoningEffort: "high" | "low" | "medium" | "xhigh";
706
+ maximumRequestCostUsd: number;
707
+ requestTimeoutMs: number;
708
+ } | undefined;
709
+ candidate?: {
710
+ artifactId: string;
711
+ contentHash: string;
712
+ preparationId: string;
713
+ dispatchId: string;
714
+ baseProfileId: string;
715
+ workerImageDigest: string;
716
+ } | undefined;
717
+ harness?: {
718
+ harnessRelease: {
719
+ id: string;
720
+ contentHash: string;
721
+ };
722
+ agentSnapshot: {
723
+ id: string;
724
+ contentHash: string;
725
+ };
726
+ sourcePackageHash: string;
727
+ } | undefined;
728
+ messages?: {
729
+ role: "assistant" | "system" | "user";
730
+ content: string;
731
+ }[] | undefined;
732
+ maxOutputTokens: number;
733
+ temperature: number;
734
+ topP: number;
735
+ } | {
736
+ kind: "fixture";
737
+ } | {
684
738
  kind: "hosted_harness";
685
739
  modelId: string;
686
740
  maxOutputTokens?: number | undefined;
@@ -922,60 +976,6 @@ export declare class OpenPondExperimentsClient {
922
976
  };
923
977
  contentHash: string;
924
978
  } | undefined;
925
- } | {
926
- kind: "hosted_chat";
927
- modelId: string;
928
- localRuntime?: {
929
- providerId: "claude-code";
930
- configurationHash: string;
931
- executableHash: string;
932
- version: string;
933
- capabilityHash: string;
934
- maximumRequestCostUsd: number;
935
- requestTimeoutMs: number;
936
- maximumTurns: number;
937
- holdOpen: boolean;
938
- } | {
939
- providerId: "codex";
940
- configurationHash: string;
941
- accountHash: string;
942
- reasoningEffort: "high" | "low" | "medium" | "xhigh";
943
- maximumRequestCostUsd: number;
944
- requestTimeoutMs: number;
945
- } | {
946
- providerId: "custom-openai-compatible";
947
- configurationHash: string;
948
- maximumRequestCostUsd: number;
949
- requestTimeoutMs: number;
950
- } | undefined;
951
- candidate?: {
952
- artifactId: string;
953
- contentHash: string;
954
- preparationId: string;
955
- dispatchId: string;
956
- baseProfileId: string;
957
- workerImageDigest: string;
958
- } | undefined;
959
- harness?: {
960
- harnessRelease: {
961
- id: string;
962
- contentHash: string;
963
- };
964
- agentSnapshot: {
965
- id: string;
966
- contentHash: string;
967
- };
968
- sourcePackageHash: string;
969
- } | undefined;
970
- messages?: {
971
- role: "assistant" | "system" | "user";
972
- content: string;
973
- }[] | undefined;
974
- maxOutputTokens: number;
975
- temperature: number;
976
- topP: number;
977
- } | {
978
- kind: "fixture";
979
979
  };
980
980
  population: {
981
981
  receiptId: string;
@@ -1206,12 +1206,24 @@ export declare class OpenPondExperimentsClient {
1206
1206
  revision: number;
1207
1207
  };
1208
1208
  policy: {
1209
- kind: "hosted_harness";
1209
+ kind: "hosted_chat";
1210
1210
  modelId: string;
1211
- maxOutputTokens?: number | undefined;
1212
- temperature?: number | undefined;
1213
- topP?: number | undefined;
1214
1211
  localRuntime?: {
1212
+ providerId: "custom-openai-compatible";
1213
+ configurationHash: string;
1214
+ maximumRequestCostUsd: number;
1215
+ requestTimeoutMs: number;
1216
+ } | {
1217
+ providerId: "claude-code";
1218
+ configurationHash: string;
1219
+ executableHash: string;
1220
+ version: string;
1221
+ capabilityHash: string;
1222
+ maximumRequestCostUsd: number;
1223
+ requestTimeoutMs: number;
1224
+ maximumTurns: number;
1225
+ holdOpen: boolean;
1226
+ } | {
1215
1227
  providerId: "codex";
1216
1228
  configurationHash: string;
1217
1229
  accountHash: string;
@@ -1227,31 +1239,73 @@ export declare class OpenPondExperimentsClient {
1227
1239
  baseProfileId: string;
1228
1240
  workerImageDigest: string;
1229
1241
  } | undefined;
1230
- sourceCandidate?: {
1231
- candidateId: string;
1232
- frozenRevision: number;
1233
- frozenHash: string;
1242
+ harness?: {
1234
1243
  harnessRelease: {
1235
1244
  id: string;
1236
1245
  contentHash: string;
1237
1246
  };
1238
- protectedClosureHash: string;
1239
- } | undefined;
1240
- reasoningEffort?: "high" | "low" | "max" | "medium" | "off" | "xhigh" | undefined;
1241
- profileRepositoryId: string;
1242
- source: {
1243
- profileId: string;
1244
- sourceRevision: string;
1245
- harnessRelease: {
1247
+ agentSnapshot: {
1246
1248
  id: string;
1247
1249
  contentHash: string;
1248
1250
  };
1249
- catalogHash: string;
1250
- definitionId: string;
1251
- definitionHash: string;
1252
- target: {
1253
- kind: "profile";
1254
- } | {
1251
+ sourcePackageHash: string;
1252
+ } | undefined;
1253
+ messages?: {
1254
+ role: "assistant" | "system" | "user";
1255
+ content: string;
1256
+ }[] | undefined;
1257
+ maxOutputTokens: number;
1258
+ temperature: number;
1259
+ topP: number;
1260
+ } | {
1261
+ kind: "fixture";
1262
+ } | {
1263
+ kind: "hosted_harness";
1264
+ modelId: string;
1265
+ maxOutputTokens?: number | undefined;
1266
+ temperature?: number | undefined;
1267
+ topP?: number | undefined;
1268
+ localRuntime?: {
1269
+ providerId: "codex";
1270
+ configurationHash: string;
1271
+ accountHash: string;
1272
+ reasoningEffort: "high" | "low" | "medium" | "xhigh";
1273
+ maximumRequestCostUsd: number;
1274
+ requestTimeoutMs: number;
1275
+ } | undefined;
1276
+ candidate?: {
1277
+ artifactId: string;
1278
+ contentHash: string;
1279
+ preparationId: string;
1280
+ dispatchId: string;
1281
+ baseProfileId: string;
1282
+ workerImageDigest: string;
1283
+ } | undefined;
1284
+ sourceCandidate?: {
1285
+ candidateId: string;
1286
+ frozenRevision: number;
1287
+ frozenHash: string;
1288
+ harnessRelease: {
1289
+ id: string;
1290
+ contentHash: string;
1291
+ };
1292
+ protectedClosureHash: string;
1293
+ } | undefined;
1294
+ reasoningEffort?: "high" | "low" | "max" | "medium" | "off" | "xhigh" | undefined;
1295
+ profileRepositoryId: string;
1296
+ source: {
1297
+ profileId: string;
1298
+ sourceRevision: string;
1299
+ harnessRelease: {
1300
+ id: string;
1301
+ contentHash: string;
1302
+ };
1303
+ catalogHash: string;
1304
+ definitionId: string;
1305
+ definitionHash: string;
1306
+ target: {
1307
+ kind: "profile";
1308
+ } | {
1255
1309
  kind: "workflow";
1256
1310
  workflowId: string;
1257
1311
  } | {
@@ -1447,60 +1501,6 @@ export declare class OpenPondExperimentsClient {
1447
1501
  };
1448
1502
  contentHash: string;
1449
1503
  } | undefined;
1450
- } | {
1451
- kind: "hosted_chat";
1452
- modelId: string;
1453
- localRuntime?: {
1454
- providerId: "claude-code";
1455
- configurationHash: string;
1456
- executableHash: string;
1457
- version: string;
1458
- capabilityHash: string;
1459
- maximumRequestCostUsd: number;
1460
- requestTimeoutMs: number;
1461
- maximumTurns: number;
1462
- holdOpen: boolean;
1463
- } | {
1464
- providerId: "codex";
1465
- configurationHash: string;
1466
- accountHash: string;
1467
- reasoningEffort: "high" | "low" | "medium" | "xhigh";
1468
- maximumRequestCostUsd: number;
1469
- requestTimeoutMs: number;
1470
- } | {
1471
- providerId: "custom-openai-compatible";
1472
- configurationHash: string;
1473
- maximumRequestCostUsd: number;
1474
- requestTimeoutMs: number;
1475
- } | undefined;
1476
- candidate?: {
1477
- artifactId: string;
1478
- contentHash: string;
1479
- preparationId: string;
1480
- dispatchId: string;
1481
- baseProfileId: string;
1482
- workerImageDigest: string;
1483
- } | undefined;
1484
- harness?: {
1485
- harnessRelease: {
1486
- id: string;
1487
- contentHash: string;
1488
- };
1489
- agentSnapshot: {
1490
- id: string;
1491
- contentHash: string;
1492
- };
1493
- sourcePackageHash: string;
1494
- } | undefined;
1495
- messages?: {
1496
- role: "assistant" | "system" | "user";
1497
- content: string;
1498
- }[] | undefined;
1499
- maxOutputTokens: number;
1500
- temperature: number;
1501
- topP: number;
1502
- } | {
1503
- kind: "fixture";
1504
1504
  };
1505
1505
  population: {
1506
1506
  receiptId: string;
@@ -1596,6 +1596,60 @@ export declare class OpenPondExperimentsClient {
1596
1596
  revision: number;
1597
1597
  };
1598
1598
  policy: {
1599
+ kind: "hosted_chat";
1600
+ modelId: string;
1601
+ localRuntime?: {
1602
+ providerId: "custom-openai-compatible";
1603
+ configurationHash: string;
1604
+ maximumRequestCostUsd: number;
1605
+ requestTimeoutMs: number;
1606
+ } | {
1607
+ providerId: "claude-code";
1608
+ configurationHash: string;
1609
+ executableHash: string;
1610
+ version: string;
1611
+ capabilityHash: string;
1612
+ maximumRequestCostUsd: number;
1613
+ requestTimeoutMs: number;
1614
+ maximumTurns: number;
1615
+ holdOpen: boolean;
1616
+ } | {
1617
+ providerId: "codex";
1618
+ configurationHash: string;
1619
+ accountHash: string;
1620
+ reasoningEffort: "high" | "low" | "medium" | "xhigh";
1621
+ maximumRequestCostUsd: number;
1622
+ requestTimeoutMs: number;
1623
+ } | undefined;
1624
+ candidate?: {
1625
+ artifactId: string;
1626
+ contentHash: string;
1627
+ preparationId: string;
1628
+ dispatchId: string;
1629
+ baseProfileId: string;
1630
+ workerImageDigest: string;
1631
+ } | undefined;
1632
+ harness?: {
1633
+ harnessRelease: {
1634
+ id: string;
1635
+ contentHash: string;
1636
+ };
1637
+ agentSnapshot: {
1638
+ id: string;
1639
+ contentHash: string;
1640
+ };
1641
+ sourcePackageHash: string;
1642
+ } | undefined;
1643
+ messages?: {
1644
+ role: "assistant" | "system" | "user";
1645
+ content: string;
1646
+ }[] | undefined;
1647
+ maxOutputTokens: number;
1648
+ temperature: number;
1649
+ topP: number;
1650
+ } | {
1651
+ kind: "fixture";
1652
+ } | {
1599
1653
  kind: "hosted_harness";
1600
1654
  modelId: string;
1601
1655
  maxOutputTokens?: number | undefined;
@@ -1837,60 +1891,6 @@ export declare class OpenPondExperimentsClient {
1837
1891
  };
1838
1892
  contentHash: string;
1839
1893
  } | undefined;
1840
- } | {
1841
- kind: "hosted_chat";
1842
- modelId: string;
1843
- localRuntime?: {
1844
- providerId: "claude-code";
1845
- configurationHash: string;
1846
- executableHash: string;
1847
- version: string;
1848
- capabilityHash: string;
1849
- maximumRequestCostUsd: number;
1850
- requestTimeoutMs: number;
1851
- maximumTurns: number;
1852
- holdOpen: boolean;
1853
- } | {
1854
- providerId: "codex";
1855
- configurationHash: string;
1856
- accountHash: string;
1857
- reasoningEffort: "high" | "low" | "medium" | "xhigh";
1858
- maximumRequestCostUsd: number;
1859
- requestTimeoutMs: number;
1860
- } | {
1861
- providerId: "custom-openai-compatible";
1862
- configurationHash: string;
1863
- maximumRequestCostUsd: number;
1864
- requestTimeoutMs: number;
1865
- } | undefined;
1866
- candidate?: {
1867
- artifactId: string;
1868
- contentHash: string;
1869
- preparationId: string;
1870
- dispatchId: string;
1871
- baseProfileId: string;
1872
- workerImageDigest: string;
1873
- } | undefined;
1874
- harness?: {
1875
- harnessRelease: {
1876
- id: string;
1877
- contentHash: string;
1878
- };
1879
- agentSnapshot: {
1880
- id: string;
1881
- contentHash: string;
1882
- };
1883
- sourcePackageHash: string;
1884
- } | undefined;
1885
- messages?: {
1886
- role: "assistant" | "system" | "user";
1887
- content: string;
1888
- }[] | undefined;
1889
- maxOutputTokens: number;
1890
- temperature: number;
1891
- topP: number;
1892
- } | {
1893
- kind: "fixture";
1894
1894
  };
1895
1895
  population: {
1896
1896
  receiptId: string;
@@ -2121,6 +2121,60 @@ export declare class OpenPondExperimentsClient {
2121
2121
  revision: number;
2122
2122
  };
2123
2123
  policy: {
2124
+ kind: "hosted_chat";
2125
+ modelId: string;
2126
+ localRuntime?: {
2127
+ providerId: "custom-openai-compatible";
2128
+ configurationHash: string;
2129
+ maximumRequestCostUsd: number;
2130
+ requestTimeoutMs: number;
2131
+ } | {
2132
+ providerId: "claude-code";
2133
+ configurationHash: string;
2134
+ executableHash: string;
2135
+ version: string;
2136
+ capabilityHash: string;
2137
+ maximumRequestCostUsd: number;
2138
+ requestTimeoutMs: number;
2139
+ maximumTurns: number;
2140
+ holdOpen: boolean;
2141
+ } | {
2142
+ providerId: "codex";
2143
+ configurationHash: string;
2144
+ accountHash: string;
2145
+ reasoningEffort: "high" | "low" | "medium" | "xhigh";
2146
+ maximumRequestCostUsd: number;
2147
+ requestTimeoutMs: number;
2148
+ } | undefined;
2149
+ candidate?: {
2150
+ artifactId: string;
2151
+ contentHash: string;
2152
+ preparationId: string;
2153
+ dispatchId: string;
2154
+ baseProfileId: string;
2155
+ workerImageDigest: string;
2156
+ } | undefined;
2157
+ harness?: {
2158
+ harnessRelease: {
2159
+ id: string;
2160
+ contentHash: string;
2161
+ };
2162
+ agentSnapshot: {
2163
+ id: string;
2164
+ contentHash: string;
2165
+ };
2166
+ sourcePackageHash: string;
2167
+ } | undefined;
2168
+ messages?: {
2169
+ role: "assistant" | "system" | "user";
2170
+ content: string;
2171
+ }[] | undefined;
2172
+ maxOutputTokens: number;
2173
+ temperature: number;
2174
+ topP: number;
2175
+ } | {
2176
+ kind: "fixture";
2177
+ } | {
2124
2178
  kind: "hosted_harness";
2125
2179
  modelId: string;
2126
2180
  maxOutputTokens?: number | undefined;
@@ -2330,92 +2384,38 @@ export declare class OpenPondExperimentsClient {
2330
2384
  split: "frozen_eval" | "test" | "train" | "validation";
2331
2385
  population: {
2332
2386
  taskId: string;
2333
- seed: string;
2334
- fixtureId: string | null;
2335
- }[];
2336
- evaluators: {
2337
- release: {
2338
- id: string;
2339
- contentHash: string;
2340
- revision: string | number;
2341
- };
2342
- feedbackKey: string;
2343
- configurationHash: string;
2344
- output: "boolean" | "category" | "score";
2345
- categories: string[];
2346
- }[];
2347
- fieldMappingsHash: string;
2348
- fieldMappings: {
2349
- graderId: string;
2350
- fields: {
2351
- destination: "evaluatorContext" | "expectedOutput" | "input" | "output";
2352
- path: string[];
2353
- source: "evaluatorContext" | "expectedOutput" | "input" | "output";
2354
- sourcePath: string[];
2355
- }[];
2356
- }[];
2357
- environmentHash: string;
2358
- privateDatasetClosureHash: string;
2359
- criterion: {
2360
- minimumPassRate: number;
2361
- requireComplete: boolean;
2362
- };
2363
- contentHash: string;
2364
- } | undefined;
2365
- } | {
2366
- kind: "hosted_chat";
2367
- modelId: string;
2368
- localRuntime?: {
2369
- providerId: "claude-code";
2370
- configurationHash: string;
2371
- executableHash: string;
2372
- version: string;
2373
- capabilityHash: string;
2374
- maximumRequestCostUsd: number;
2375
- requestTimeoutMs: number;
2376
- maximumTurns: number;
2377
- holdOpen: boolean;
2378
- } | {
2379
- providerId: "codex";
2380
- configurationHash: string;
2381
- accountHash: string;
2382
- reasoningEffort: "high" | "low" | "medium" | "xhigh";
2383
- maximumRequestCostUsd: number;
2384
- requestTimeoutMs: number;
2385
- } | {
2386
- providerId: "custom-openai-compatible";
2387
- configurationHash: string;
2388
- maximumRequestCostUsd: number;
2389
- requestTimeoutMs: number;
2390
- } | undefined;
2391
- candidate?: {
2392
- artifactId: string;
2393
- contentHash: string;
2394
- preparationId: string;
2395
- dispatchId: string;
2396
- baseProfileId: string;
2397
- workerImageDigest: string;
2398
- } | undefined;
2399
- harness?: {
2400
- harnessRelease: {
2401
- id: string;
2402
- contentHash: string;
2403
- };
2404
- agentSnapshot: {
2405
- id: string;
2406
- contentHash: string;
2387
+ seed: string;
2388
+ fixtureId: string | null;
2389
+ }[];
2390
+ evaluators: {
2391
+ release: {
2392
+ id: string;
2393
+ contentHash: string;
2394
+ revision: string | number;
2395
+ };
2396
+ feedbackKey: string;
2397
+ configurationHash: string;
2398
+ output: "boolean" | "category" | "score";
2399
+ categories: string[];
2400
+ }[];
2401
+ fieldMappingsHash: string;
2402
+ fieldMappings: {
2403
+ graderId: string;
2404
+ fields: {
2405
+ destination: "evaluatorContext" | "expectedOutput" | "input" | "output";
2406
+ path: string[];
2407
+ source: "evaluatorContext" | "expectedOutput" | "input" | "output";
2408
+ sourcePath: string[];
2409
+ }[];
2410
+ }[];
2411
+ environmentHash: string;
2412
+ privateDatasetClosureHash: string;
2413
+ criterion: {
2414
+ minimumPassRate: number;
2415
+ requireComplete: boolean;
2407
2416
  };
2408
- sourcePackageHash: string;
2417
+ contentHash: string;
2409
2418
  } | undefined;
2410
- messages?: {
2411
- role: "assistant" | "system" | "user";
2412
- content: string;
2413
- }[] | undefined;
2414
- maxOutputTokens: number;
2415
- temperature: number;
2416
- topP: number;
2417
- } | {
2418
- kind: "fixture";
2419
2419
  };
2420
2420
  population: {
2421
2421
  receiptId: string;
@@ -2577,6 +2577,60 @@ export declare class OpenPondExperimentsClient {
2577
2577
  revision: number;
2578
2578
  };
2579
2579
  policy: {
2580
+ kind: "hosted_chat";
2581
+ modelId: string;
2582
+ localRuntime?: {
2583
+ providerId: "custom-openai-compatible";
2584
+ configurationHash: string;
2585
+ maximumRequestCostUsd: number;
2586
+ requestTimeoutMs: number;
2587
+ } | {
2588
+ providerId: "claude-code";
2589
+ configurationHash: string;
2590
+ executableHash: string;
2591
+ version: string;
2592
+ capabilityHash: string;
2593
+ maximumRequestCostUsd: number;
2594
+ requestTimeoutMs: number;
2595
+ maximumTurns: number;
2596
+ holdOpen: boolean;
2597
+ } | {
2598
+ providerId: "codex";
2599
+ configurationHash: string;
2600
+ accountHash: string;
2601
+ reasoningEffort: "high" | "low" | "medium" | "xhigh";
2602
+ maximumRequestCostUsd: number;
2603
+ requestTimeoutMs: number;
2604
+ } | undefined;
2605
+ candidate?: {
2606
+ artifactId: string;
2607
+ contentHash: string;
2608
+ preparationId: string;
2609
+ dispatchId: string;
2610
+ baseProfileId: string;
2611
+ workerImageDigest: string;
2612
+ } | undefined;
2613
+ harness?: {
2614
+ harnessRelease: {
2615
+ id: string;
2616
+ contentHash: string;
2617
+ };
2618
+ agentSnapshot: {
2619
+ id: string;
2620
+ contentHash: string;
2621
+ };
2622
+ sourcePackageHash: string;
2623
+ } | undefined;
2624
+ messages?: {
2625
+ role: "assistant" | "system" | "user";
2626
+ content: string;
2627
+ }[] | undefined;
2628
+ maxOutputTokens: number;
2629
+ temperature: number;
2630
+ topP: number;
2631
+ } | {
2632
+ kind: "fixture";
2633
+ } | {
2580
2634
  kind: "hosted_harness";
2581
2635
  modelId: string;
2582
2636
  maxOutputTokens?: number | undefined;
@@ -2818,60 +2872,6 @@ export declare class OpenPondExperimentsClient {
2818
2872
  };
2819
2873
  contentHash: string;
2820
2874
  } | undefined;
2821
- } | {
2822
- kind: "hosted_chat";
2823
- modelId: string;
2824
- localRuntime?: {
2825
- providerId: "claude-code";
2826
- configurationHash: string;
2827
- executableHash: string;
2828
- version: string;
2829
- capabilityHash: string;
2830
- maximumRequestCostUsd: number;
2831
- requestTimeoutMs: number;
2832
- maximumTurns: number;
2833
- holdOpen: boolean;
2834
- } | {
2835
- providerId: "codex";
2836
- configurationHash: string;
2837
- accountHash: string;
2838
- reasoningEffort: "high" | "low" | "medium" | "xhigh";
2839
- maximumRequestCostUsd: number;
2840
- requestTimeoutMs: number;
2841
- } | {
2842
- providerId: "custom-openai-compatible";
2843
- configurationHash: string;
2844
- maximumRequestCostUsd: number;
2845
- requestTimeoutMs: number;
2846
- } | undefined;
2847
- candidate?: {
2848
- artifactId: string;
2849
- contentHash: string;
2850
- preparationId: string;
2851
- dispatchId: string;
2852
- baseProfileId: string;
2853
- workerImageDigest: string;
2854
- } | undefined;
2855
- harness?: {
2856
- harnessRelease: {
2857
- id: string;
2858
- contentHash: string;
2859
- };
2860
- agentSnapshot: {
2861
- id: string;
2862
- contentHash: string;
2863
- };
2864
- sourcePackageHash: string;
2865
- } | undefined;
2866
- messages?: {
2867
- role: "assistant" | "system" | "user";
2868
- content: string;
2869
- }[] | undefined;
2870
- maxOutputTokens: number;
2871
- temperature: number;
2872
- topP: number;
2873
- } | {
2874
- kind: "fixture";
2875
2875
  };
2876
2876
  population: {
2877
2877
  receiptId: string;
@@ -3102,6 +3102,60 @@ export declare class OpenPondExperimentsClient {
3102
3102
  revision: number;
3103
3103
  };
3104
3104
  policy: {
3105
+ kind: "hosted_chat";
3106
+ modelId: string;
3107
+ localRuntime?: {
3108
+ providerId: "custom-openai-compatible";
3109
+ configurationHash: string;
3110
+ maximumRequestCostUsd: number;
3111
+ requestTimeoutMs: number;
3112
+ } | {
3113
+ providerId: "claude-code";
3114
+ configurationHash: string;
3115
+ executableHash: string;
3116
+ version: string;
3117
+ capabilityHash: string;
3118
+ maximumRequestCostUsd: number;
3119
+ requestTimeoutMs: number;
3120
+ maximumTurns: number;
3121
+ holdOpen: boolean;
3122
+ } | {
3123
+ providerId: "codex";
3124
+ configurationHash: string;
3125
+ accountHash: string;
3126
+ reasoningEffort: "high" | "low" | "medium" | "xhigh";
3127
+ maximumRequestCostUsd: number;
3128
+ requestTimeoutMs: number;
3129
+ } | undefined;
3130
+ candidate?: {
3131
+ artifactId: string;
3132
+ contentHash: string;
3133
+ preparationId: string;
3134
+ dispatchId: string;
3135
+ baseProfileId: string;
3136
+ workerImageDigest: string;
3137
+ } | undefined;
3138
+ harness?: {
3139
+ harnessRelease: {
3140
+ id: string;
3141
+ contentHash: string;
3142
+ };
3143
+ agentSnapshot: {
3144
+ id: string;
3145
+ contentHash: string;
3146
+ };
3147
+ sourcePackageHash: string;
3148
+ } | undefined;
3149
+ messages?: {
3150
+ role: "assistant" | "system" | "user";
3151
+ content: string;
3152
+ }[] | undefined;
3153
+ maxOutputTokens: number;
3154
+ temperature: number;
3155
+ topP: number;
3156
+ } | {
3157
+ kind: "fixture";
3158
+ } | {
3105
3159
  kind: "hosted_harness";
3106
3160
  modelId: string;
3107
3161
  maxOutputTokens?: number | undefined;
@@ -3331,72 +3385,18 @@ export declare class OpenPondExperimentsClient {
3331
3385
  fields: {
3332
3386
  destination: "evaluatorContext" | "expectedOutput" | "input" | "output";
3333
3387
  path: string[];
3334
- source: "evaluatorContext" | "expectedOutput" | "input" | "output";
3335
- sourcePath: string[];
3336
- }[];
3337
- }[];
3338
- environmentHash: string;
3339
- privateDatasetClosureHash: string;
3340
- criterion: {
3341
- minimumPassRate: number;
3342
- requireComplete: boolean;
3343
- };
3344
- contentHash: string;
3345
- } | undefined;
3346
- } | {
3347
- kind: "hosted_chat";
3348
- modelId: string;
3349
- localRuntime?: {
3350
- providerId: "claude-code";
3351
- configurationHash: string;
3352
- executableHash: string;
3353
- version: string;
3354
- capabilityHash: string;
3355
- maximumRequestCostUsd: number;
3356
- requestTimeoutMs: number;
3357
- maximumTurns: number;
3358
- holdOpen: boolean;
3359
- } | {
3360
- providerId: "codex";
3361
- configurationHash: string;
3362
- accountHash: string;
3363
- reasoningEffort: "high" | "low" | "medium" | "xhigh";
3364
- maximumRequestCostUsd: number;
3365
- requestTimeoutMs: number;
3366
- } | {
3367
- providerId: "custom-openai-compatible";
3368
- configurationHash: string;
3369
- maximumRequestCostUsd: number;
3370
- requestTimeoutMs: number;
3371
- } | undefined;
3372
- candidate?: {
3373
- artifactId: string;
3374
- contentHash: string;
3375
- preparationId: string;
3376
- dispatchId: string;
3377
- baseProfileId: string;
3378
- workerImageDigest: string;
3379
- } | undefined;
3380
- harness?: {
3381
- harnessRelease: {
3382
- id: string;
3383
- contentHash: string;
3384
- };
3385
- agentSnapshot: {
3386
- id: string;
3387
- contentHash: string;
3388
+ source: "evaluatorContext" | "expectedOutput" | "input" | "output";
3389
+ sourcePath: string[];
3390
+ }[];
3391
+ }[];
3392
+ environmentHash: string;
3393
+ privateDatasetClosureHash: string;
3394
+ criterion: {
3395
+ minimumPassRate: number;
3396
+ requireComplete: boolean;
3388
3397
  };
3389
- sourcePackageHash: string;
3398
+ contentHash: string;
3390
3399
  } | undefined;
3391
- messages?: {
3392
- role: "assistant" | "system" | "user";
3393
- content: string;
3394
- }[] | undefined;
3395
- maxOutputTokens: number;
3396
- temperature: number;
3397
- topP: number;
3398
- } | {
3399
- kind: "fixture";
3400
3400
  };
3401
3401
  population: {
3402
3402
  receiptId: string;
@@ -3763,6 +3763,60 @@ export declare class OpenPondExperimentsClient {
3763
3763
  revision: number;
3764
3764
  };
3765
3765
  policy: {
3766
+ kind: "hosted_chat";
3767
+ modelId: string;
3768
+ localRuntime?: {
3769
+ providerId: "custom-openai-compatible";
3770
+ configurationHash: string;
3771
+ maximumRequestCostUsd: number;
3772
+ requestTimeoutMs: number;
3773
+ } | {
3774
+ providerId: "claude-code";
3775
+ configurationHash: string;
3776
+ executableHash: string;
3777
+ version: string;
3778
+ capabilityHash: string;
3779
+ maximumRequestCostUsd: number;
3780
+ requestTimeoutMs: number;
3781
+ maximumTurns: number;
3782
+ holdOpen: boolean;
3783
+ } | {
3784
+ providerId: "codex";
3785
+ configurationHash: string;
3786
+ accountHash: string;
3787
+ reasoningEffort: "high" | "low" | "medium" | "xhigh";
3788
+ maximumRequestCostUsd: number;
3789
+ requestTimeoutMs: number;
3790
+ } | undefined;
3791
+ candidate?: {
3792
+ artifactId: string;
3793
+ contentHash: string;
3794
+ preparationId: string;
3795
+ dispatchId: string;
3796
+ baseProfileId: string;
3797
+ workerImageDigest: string;
3798
+ } | undefined;
3799
+ harness?: {
3800
+ harnessRelease: {
3801
+ id: string;
3802
+ contentHash: string;
3803
+ };
3804
+ agentSnapshot: {
3805
+ id: string;
3806
+ contentHash: string;
3807
+ };
3808
+ sourcePackageHash: string;
3809
+ } | undefined;
3810
+ messages?: {
3811
+ role: "assistant" | "system" | "user";
3812
+ content: string;
3813
+ }[] | undefined;
3814
+ maxOutputTokens: number;
3815
+ temperature: number;
3816
+ topP: number;
3817
+ } | {
3818
+ kind: "fixture";
3819
+ } | {
3766
3820
  kind: "hosted_harness";
3767
3821
  modelId: string;
3768
3822
  maxOutputTokens?: number | undefined;
@@ -4004,60 +4058,6 @@ export declare class OpenPondExperimentsClient {
4004
4058
  };
4005
4059
  contentHash: string;
4006
4060
  } | undefined;
4007
- } | {
4008
- kind: "hosted_chat";
4009
- modelId: string;
4010
- localRuntime?: {
4011
- providerId: "claude-code";
4012
- configurationHash: string;
4013
- executableHash: string;
4014
- version: string;
4015
- capabilityHash: string;
4016
- maximumRequestCostUsd: number;
4017
- requestTimeoutMs: number;
4018
- maximumTurns: number;
4019
- holdOpen: boolean;
4020
- } | {
4021
- providerId: "codex";
4022
- configurationHash: string;
4023
- accountHash: string;
4024
- reasoningEffort: "high" | "low" | "medium" | "xhigh";
4025
- maximumRequestCostUsd: number;
4026
- requestTimeoutMs: number;
4027
- } | {
4028
- providerId: "custom-openai-compatible";
4029
- configurationHash: string;
4030
- maximumRequestCostUsd: number;
4031
- requestTimeoutMs: number;
4032
- } | undefined;
4033
- candidate?: {
4034
- artifactId: string;
4035
- contentHash: string;
4036
- preparationId: string;
4037
- dispatchId: string;
4038
- baseProfileId: string;
4039
- workerImageDigest: string;
4040
- } | undefined;
4041
- harness?: {
4042
- harnessRelease: {
4043
- id: string;
4044
- contentHash: string;
4045
- };
4046
- agentSnapshot: {
4047
- id: string;
4048
- contentHash: string;
4049
- };
4050
- sourcePackageHash: string;
4051
- } | undefined;
4052
- messages?: {
4053
- role: "assistant" | "system" | "user";
4054
- content: string;
4055
- }[] | undefined;
4056
- maxOutputTokens: number;
4057
- temperature: number;
4058
- topP: number;
4059
- } | {
4060
- kind: "fixture";
4061
4061
  };
4062
4062
  population: {
4063
4063
  receiptId: string;
@@ -4288,6 +4288,60 @@ export declare class OpenPondExperimentsClient {
4288
4288
  revision: number;
4289
4289
  };
4290
4290
  policy: {
4291
+ kind: "hosted_chat";
4292
+ modelId: string;
4293
+ localRuntime?: {
4294
+ providerId: "custom-openai-compatible";
4295
+ configurationHash: string;
4296
+ maximumRequestCostUsd: number;
4297
+ requestTimeoutMs: number;
4298
+ } | {
4299
+ providerId: "claude-code";
4300
+ configurationHash: string;
4301
+ executableHash: string;
4302
+ version: string;
4303
+ capabilityHash: string;
4304
+ maximumRequestCostUsd: number;
4305
+ requestTimeoutMs: number;
4306
+ maximumTurns: number;
4307
+ holdOpen: boolean;
4308
+ } | {
4309
+ providerId: "codex";
4310
+ configurationHash: string;
4311
+ accountHash: string;
4312
+ reasoningEffort: "high" | "low" | "medium" | "xhigh";
4313
+ maximumRequestCostUsd: number;
4314
+ requestTimeoutMs: number;
4315
+ } | undefined;
4316
+ candidate?: {
4317
+ artifactId: string;
4318
+ contentHash: string;
4319
+ preparationId: string;
4320
+ dispatchId: string;
4321
+ baseProfileId: string;
4322
+ workerImageDigest: string;
4323
+ } | undefined;
4324
+ harness?: {
4325
+ harnessRelease: {
4326
+ id: string;
4327
+ contentHash: string;
4328
+ };
4329
+ agentSnapshot: {
4330
+ id: string;
4331
+ contentHash: string;
4332
+ };
4333
+ sourcePackageHash: string;
4334
+ } | undefined;
4335
+ messages?: {
4336
+ role: "assistant" | "system" | "user";
4337
+ content: string;
4338
+ }[] | undefined;
4339
+ maxOutputTokens: number;
4340
+ temperature: number;
4341
+ topP: number;
4342
+ } | {
4343
+ kind: "fixture";
4344
+ } | {
4291
4345
  kind: "hosted_harness";
4292
4346
  modelId: string;
4293
4347
  maxOutputTokens?: number | undefined;
@@ -4497,92 +4551,38 @@ export declare class OpenPondExperimentsClient {
4497
4551
  split: "frozen_eval" | "test" | "train" | "validation";
4498
4552
  population: {
4499
4553
  taskId: string;
4500
- seed: string;
4501
- fixtureId: string | null;
4502
- }[];
4503
- evaluators: {
4504
- release: {
4505
- id: string;
4506
- contentHash: string;
4507
- revision: string | number;
4508
- };
4509
- feedbackKey: string;
4510
- configurationHash: string;
4511
- output: "boolean" | "category" | "score";
4512
- categories: string[];
4513
- }[];
4514
- fieldMappingsHash: string;
4515
- fieldMappings: {
4516
- graderId: string;
4517
- fields: {
4518
- destination: "evaluatorContext" | "expectedOutput" | "input" | "output";
4519
- path: string[];
4520
- source: "evaluatorContext" | "expectedOutput" | "input" | "output";
4521
- sourcePath: string[];
4522
- }[];
4523
- }[];
4524
- environmentHash: string;
4525
- privateDatasetClosureHash: string;
4526
- criterion: {
4527
- minimumPassRate: number;
4528
- requireComplete: boolean;
4529
- };
4530
- contentHash: string;
4531
- } | undefined;
4532
- } | {
4533
- kind: "hosted_chat";
4534
- modelId: string;
4535
- localRuntime?: {
4536
- providerId: "claude-code";
4537
- configurationHash: string;
4538
- executableHash: string;
4539
- version: string;
4540
- capabilityHash: string;
4541
- maximumRequestCostUsd: number;
4542
- requestTimeoutMs: number;
4543
- maximumTurns: number;
4544
- holdOpen: boolean;
4545
- } | {
4546
- providerId: "codex";
4547
- configurationHash: string;
4548
- accountHash: string;
4549
- reasoningEffort: "high" | "low" | "medium" | "xhigh";
4550
- maximumRequestCostUsd: number;
4551
- requestTimeoutMs: number;
4552
- } | {
4553
- providerId: "custom-openai-compatible";
4554
- configurationHash: string;
4555
- maximumRequestCostUsd: number;
4556
- requestTimeoutMs: number;
4557
- } | undefined;
4558
- candidate?: {
4559
- artifactId: string;
4560
- contentHash: string;
4561
- preparationId: string;
4562
- dispatchId: string;
4563
- baseProfileId: string;
4564
- workerImageDigest: string;
4565
- } | undefined;
4566
- harness?: {
4567
- harnessRelease: {
4568
- id: string;
4569
- contentHash: string;
4570
- };
4571
- agentSnapshot: {
4572
- id: string;
4573
- contentHash: string;
4554
+ seed: string;
4555
+ fixtureId: string | null;
4556
+ }[];
4557
+ evaluators: {
4558
+ release: {
4559
+ id: string;
4560
+ contentHash: string;
4561
+ revision: string | number;
4562
+ };
4563
+ feedbackKey: string;
4564
+ configurationHash: string;
4565
+ output: "boolean" | "category" | "score";
4566
+ categories: string[];
4567
+ }[];
4568
+ fieldMappingsHash: string;
4569
+ fieldMappings: {
4570
+ graderId: string;
4571
+ fields: {
4572
+ destination: "evaluatorContext" | "expectedOutput" | "input" | "output";
4573
+ path: string[];
4574
+ source: "evaluatorContext" | "expectedOutput" | "input" | "output";
4575
+ sourcePath: string[];
4576
+ }[];
4577
+ }[];
4578
+ environmentHash: string;
4579
+ privateDatasetClosureHash: string;
4580
+ criterion: {
4581
+ minimumPassRate: number;
4582
+ requireComplete: boolean;
4574
4583
  };
4575
- sourcePackageHash: string;
4584
+ contentHash: string;
4576
4585
  } | undefined;
4577
- messages?: {
4578
- role: "assistant" | "system" | "user";
4579
- content: string;
4580
- }[] | undefined;
4581
- maxOutputTokens: number;
4582
- temperature: number;
4583
- topP: number;
4584
- } | {
4585
- kind: "fixture";
4586
4586
  };
4587
4587
  population: {
4588
4588
  receiptId: string;
@@ -4677,6 +4677,60 @@ export declare class OpenPondExperimentsClient {
4677
4677
  revision: number;
4678
4678
  };
4679
4679
  policy: {
4680
+ kind: "hosted_chat";
4681
+ modelId: string;
4682
+ localRuntime?: {
4683
+ providerId: "custom-openai-compatible";
4684
+ configurationHash: string;
4685
+ maximumRequestCostUsd: number;
4686
+ requestTimeoutMs: number;
4687
+ } | {
4688
+ providerId: "claude-code";
4689
+ configurationHash: string;
4690
+ executableHash: string;
4691
+ version: string;
4692
+ capabilityHash: string;
4693
+ maximumRequestCostUsd: number;
4694
+ requestTimeoutMs: number;
4695
+ maximumTurns: number;
4696
+ holdOpen: boolean;
4697
+ } | {
4698
+ providerId: "codex";
4699
+ configurationHash: string;
4700
+ accountHash: string;
4701
+ reasoningEffort: "high" | "low" | "medium" | "xhigh";
4702
+ maximumRequestCostUsd: number;
4703
+ requestTimeoutMs: number;
4704
+ } | undefined;
4705
+ candidate?: {
4706
+ artifactId: string;
4707
+ contentHash: string;
4708
+ preparationId: string;
4709
+ dispatchId: string;
4710
+ baseProfileId: string;
4711
+ workerImageDigest: string;
4712
+ } | undefined;
4713
+ harness?: {
4714
+ harnessRelease: {
4715
+ id: string;
4716
+ contentHash: string;
4717
+ };
4718
+ agentSnapshot: {
4719
+ id: string;
4720
+ contentHash: string;
4721
+ };
4722
+ sourcePackageHash: string;
4723
+ } | undefined;
4724
+ messages?: {
4725
+ role: "assistant" | "system" | "user";
4726
+ content: string;
4727
+ }[] | undefined;
4728
+ maxOutputTokens: number;
4729
+ temperature: number;
4730
+ topP: number;
4731
+ } | {
4732
+ kind: "fixture";
4733
+ } | {
4680
4734
  kind: "hosted_harness";
4681
4735
  modelId: string;
4682
4736
  maxOutputTokens?: number | undefined;
@@ -4918,60 +4972,6 @@ export declare class OpenPondExperimentsClient {
4918
4972
  };
4919
4973
  contentHash: string;
4920
4974
  } | undefined;
4921
- } | {
4922
- kind: "hosted_chat";
4923
- modelId: string;
4924
- localRuntime?: {
4925
- providerId: "claude-code";
4926
- configurationHash: string;
4927
- executableHash: string;
4928
- version: string;
4929
- capabilityHash: string;
4930
- maximumRequestCostUsd: number;
4931
- requestTimeoutMs: number;
4932
- maximumTurns: number;
4933
- holdOpen: boolean;
4934
- } | {
4935
- providerId: "codex";
4936
- configurationHash: string;
4937
- accountHash: string;
4938
- reasoningEffort: "high" | "low" | "medium" | "xhigh";
4939
- maximumRequestCostUsd: number;
4940
- requestTimeoutMs: number;
4941
- } | {
4942
- providerId: "custom-openai-compatible";
4943
- configurationHash: string;
4944
- maximumRequestCostUsd: number;
4945
- requestTimeoutMs: number;
4946
- } | undefined;
4947
- candidate?: {
4948
- artifactId: string;
4949
- contentHash: string;
4950
- preparationId: string;
4951
- dispatchId: string;
4952
- baseProfileId: string;
4953
- workerImageDigest: string;
4954
- } | undefined;
4955
- harness?: {
4956
- harnessRelease: {
4957
- id: string;
4958
- contentHash: string;
4959
- };
4960
- agentSnapshot: {
4961
- id: string;
4962
- contentHash: string;
4963
- };
4964
- sourcePackageHash: string;
4965
- } | undefined;
4966
- messages?: {
4967
- role: "assistant" | "system" | "user";
4968
- content: string;
4969
- }[] | undefined;
4970
- maxOutputTokens: number;
4971
- temperature: number;
4972
- topP: number;
4973
- } | {
4974
- kind: "fixture";
4975
4975
  };
4976
4976
  population: {
4977
4977
  receiptId: string;
@@ -5202,6 +5202,60 @@ export declare class OpenPondExperimentsClient {
5202
5202
  revision: number;
5203
5203
  };
5204
5204
  policy: {
5205
+ kind: "hosted_chat";
5206
+ modelId: string;
5207
+ localRuntime?: {
5208
+ providerId: "custom-openai-compatible";
5209
+ configurationHash: string;
5210
+ maximumRequestCostUsd: number;
5211
+ requestTimeoutMs: number;
5212
+ } | {
5213
+ providerId: "claude-code";
5214
+ configurationHash: string;
5215
+ executableHash: string;
5216
+ version: string;
5217
+ capabilityHash: string;
5218
+ maximumRequestCostUsd: number;
5219
+ requestTimeoutMs: number;
5220
+ maximumTurns: number;
5221
+ holdOpen: boolean;
5222
+ } | {
5223
+ providerId: "codex";
5224
+ configurationHash: string;
5225
+ accountHash: string;
5226
+ reasoningEffort: "high" | "low" | "medium" | "xhigh";
5227
+ maximumRequestCostUsd: number;
5228
+ requestTimeoutMs: number;
5229
+ } | undefined;
5230
+ candidate?: {
5231
+ artifactId: string;
5232
+ contentHash: string;
5233
+ preparationId: string;
5234
+ dispatchId: string;
5235
+ baseProfileId: string;
5236
+ workerImageDigest: string;
5237
+ } | undefined;
5238
+ harness?: {
5239
+ harnessRelease: {
5240
+ id: string;
5241
+ contentHash: string;
5242
+ };
5243
+ agentSnapshot: {
5244
+ id: string;
5245
+ contentHash: string;
5246
+ };
5247
+ sourcePackageHash: string;
5248
+ } | undefined;
5249
+ messages?: {
5250
+ role: "assistant" | "system" | "user";
5251
+ content: string;
5252
+ }[] | undefined;
5253
+ maxOutputTokens: number;
5254
+ temperature: number;
5255
+ topP: number;
5256
+ } | {
5257
+ kind: "fixture";
5258
+ } | {
5205
5259
  kind: "hosted_harness";
5206
5260
  modelId: string;
5207
5261
  maxOutputTokens?: number | undefined;
@@ -5443,60 +5497,6 @@ export declare class OpenPondExperimentsClient {
5443
5497
  };
5444
5498
  contentHash: string;
5445
5499
  } | undefined;
5446
- } | {
5447
- kind: "hosted_chat";
5448
- modelId: string;
5449
- localRuntime?: {
5450
- providerId: "claude-code";
5451
- configurationHash: string;
5452
- executableHash: string;
5453
- version: string;
5454
- capabilityHash: string;
5455
- maximumRequestCostUsd: number;
5456
- requestTimeoutMs: number;
5457
- maximumTurns: number;
5458
- holdOpen: boolean;
5459
- } | {
5460
- providerId: "codex";
5461
- configurationHash: string;
5462
- accountHash: string;
5463
- reasoningEffort: "high" | "low" | "medium" | "xhigh";
5464
- maximumRequestCostUsd: number;
5465
- requestTimeoutMs: number;
5466
- } | {
5467
- providerId: "custom-openai-compatible";
5468
- configurationHash: string;
5469
- maximumRequestCostUsd: number;
5470
- requestTimeoutMs: number;
5471
- } | undefined;
5472
- candidate?: {
5473
- artifactId: string;
5474
- contentHash: string;
5475
- preparationId: string;
5476
- dispatchId: string;
5477
- baseProfileId: string;
5478
- workerImageDigest: string;
5479
- } | undefined;
5480
- harness?: {
5481
- harnessRelease: {
5482
- id: string;
5483
- contentHash: string;
5484
- };
5485
- agentSnapshot: {
5486
- id: string;
5487
- contentHash: string;
5488
- };
5489
- sourcePackageHash: string;
5490
- } | undefined;
5491
- messages?: {
5492
- role: "assistant" | "system" | "user";
5493
- content: string;
5494
- }[] | undefined;
5495
- maxOutputTokens: number;
5496
- temperature: number;
5497
- topP: number;
5498
- } | {
5499
- kind: "fixture";
5500
5500
  };
5501
5501
  population: {
5502
5502
  receiptId: string;
@@ -5540,7 +5540,9 @@ export declare class OpenPondExperimentsClient {
5540
5540
  id: string;
5541
5541
  contentHash: string;
5542
5542
  };
5543
- executionKind?: "recorded_evidence" | undefined;
5543
+ executionKind?: "recorded_evidence" | "weight_space" | undefined;
5544
+ weightSpaceOperationId?: string | undefined;
5545
+ weightSpaceTarget?: "baseline" | "candidate" | undefined;
5544
5546
  graders: {
5545
5547
  id: string;
5546
5548
  contentHash: string;
@@ -5606,7 +5608,9 @@ export declare class OpenPondExperimentsClient {
5606
5608
  id: string;
5607
5609
  contentHash: string;
5608
5610
  };
5609
- executionKind?: "recorded_evidence" | undefined;
5611
+ executionKind?: "recorded_evidence" | "weight_space" | undefined;
5612
+ weightSpaceOperationId?: string | undefined;
5613
+ weightSpaceTarget?: "baseline" | "candidate" | undefined;
5610
5614
  graders: {
5611
5615
  id: string;
5612
5616
  contentHash: string;
@@ -5663,7 +5667,7 @@ export declare class OpenPondExperimentsClient {
5663
5667
  }>;
5664
5668
  scoringPasses(executionId: string, options?: {
5665
5669
  afterId?: string;
5666
- executionKind?: "recorded_evidence";
5670
+ executionKind?: "recorded_evidence" | "weight_space";
5667
5671
  signal?: AbortSignal;
5668
5672
  }): Promise<{
5669
5673
  items: {
@@ -5677,7 +5681,9 @@ export declare class OpenPondExperimentsClient {
5677
5681
  id: string;
5678
5682
  contentHash: string;
5679
5683
  };
5680
- executionKind?: "recorded_evidence" | undefined;
5684
+ executionKind?: "recorded_evidence" | "weight_space" | undefined;
5685
+ weightSpaceOperationId?: string | undefined;
5686
+ weightSpaceTarget?: "baseline" | "candidate" | undefined;
5681
5687
  graders: {
5682
5688
  id: string;
5683
5689
  contentHash: string;
@@ -5745,7 +5751,9 @@ export declare class OpenPondExperimentsClient {
5745
5751
  id: string;
5746
5752
  contentHash: string;
5747
5753
  };
5748
- executionKind?: "recorded_evidence" | undefined;
5754
+ executionKind?: "recorded_evidence" | "weight_space" | undefined;
5755
+ weightSpaceOperationId?: string | undefined;
5756
+ weightSpaceTarget?: "baseline" | "candidate" | undefined;
5749
5757
  graders: {
5750
5758
  id: string;
5751
5759
  contentHash: string;