openpond-sdk 0.10.0 → 0.11.0
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.
- package/dist/advanced-refiner-evaluations.js +42 -0
- package/dist/advanced-refiner-evaluations.js.map +7 -0
- package/dist/candidate-evaluations.js +25 -0
- package/dist/candidate-evaluations.js.map +7 -0
- package/dist/{chunk-W5TXWVFS.js → chunk-2JKV5OJO.js} +88 -59
- package/dist/chunk-2JKV5OJO.js.map +7 -0
- package/dist/{chunk-2ZJRG3KL.js → chunk-2NSNBED7.js} +7 -3
- package/dist/chunk-2NSNBED7.js.map +7 -0
- package/dist/chunk-5RZAPZWI.js +445 -0
- package/dist/chunk-5RZAPZWI.js.map +7 -0
- package/dist/{chunk-JCPLAIZL.js → chunk-5U25KAJM.js} +106 -76
- package/dist/chunk-5U25KAJM.js.map +7 -0
- package/dist/chunk-7Q7I6FAT.js +139 -0
- package/dist/chunk-7Q7I6FAT.js.map +7 -0
- package/dist/chunk-A64VYVXN.js +64 -0
- package/dist/chunk-A64VYVXN.js.map +7 -0
- package/dist/{chunk-DT7FRNIZ.js → chunk-CSWVE3E6.js} +2 -2
- package/dist/chunk-DJWHN7NU.js +394 -0
- package/dist/chunk-DJWHN7NU.js.map +7 -0
- package/dist/chunk-GDJRGRHA.js +144 -0
- package/dist/chunk-GDJRGRHA.js.map +7 -0
- package/dist/chunk-HOAYWW33.js +587 -0
- package/dist/chunk-HOAYWW33.js.map +7 -0
- package/dist/{chunk-Y4NGWGTA.js → chunk-I2MBVOVX.js} +265 -264
- package/dist/chunk-I2MBVOVX.js.map +7 -0
- package/dist/{chunk-I3TQTXKB.js → chunk-IXBSFJCX.js} +4 -4
- package/dist/chunk-JMQ4HBWP.js +160 -0
- package/dist/chunk-JMQ4HBWP.js.map +7 -0
- package/dist/chunk-KQWCXW3A.js +294 -0
- package/dist/chunk-KQWCXW3A.js.map +7 -0
- package/dist/{chunk-I2YM3UY4.js → chunk-NHCGOFQW.js} +2 -2
- package/dist/{chunk-32H2EJRA.js → chunk-R3M6BIOK.js} +2 -2
- package/dist/chunk-SLS3GFPG.js +125 -0
- package/dist/chunk-SLS3GFPG.js.map +7 -0
- package/dist/chunk-SVU4RZ4K.js +70 -0
- package/dist/chunk-SVU4RZ4K.js.map +7 -0
- package/dist/chunk-T4APHDES.js +74 -0
- package/dist/chunk-T4APHDES.js.map +7 -0
- package/dist/chunk-VC5OV5AS.js +174 -0
- package/dist/chunk-VC5OV5AS.js.map +7 -0
- package/dist/{chunk-7B44TQZD.js → chunk-WX22WN33.js} +13 -7
- package/dist/chunk-WX22WN33.js.map +7 -0
- package/dist/chunk-XEUOJ5WF.js +31 -0
- package/dist/chunk-XEUOJ5WF.js.map +7 -0
- package/dist/{chunk-LGEF4S5I.js → chunk-XJDWHTH5.js} +390 -362
- package/dist/chunk-XJDWHTH5.js.map +7 -0
- package/dist/chunk-YCQVVFE5.js +37 -0
- package/dist/chunk-YCQVVFE5.js.map +7 -0
- package/dist/chunk-ZVJU5CJ5.js +267 -0
- package/dist/chunk-ZVJU5CJ5.js.map +7 -0
- package/dist/connected-evidence.js +81 -0
- package/dist/connected-evidence.js.map +7 -0
- package/dist/dataset-workspaces.js +44 -79
- package/dist/dataset-workspaces.js.map +4 -4
- package/dist/experiment-evaluation-schedules.js +48 -0
- package/dist/experiment-evaluation-schedules.js.map +7 -0
- package/dist/experiment-improvements.js +52 -0
- package/dist/experiment-improvements.js.map +7 -0
- package/dist/experiments.js +48 -90
- package/dist/experiments.js.map +3 -3
- package/dist/human-review.js +13 -0
- package/dist/human-review.js.map +7 -0
- package/dist/index.js +208 -17
- package/dist/index.js.map +2 -2
- package/dist/learning.js +1 -1
- package/dist/model-starter-attempts.js +2 -2
- package/dist/model-starter-catalog.js +7 -2
- package/dist/model-starter-catalog.js.map +1 -1
- package/dist/model-starters.js +7 -2
- package/dist/model-taskset-runs.js +3 -3
- package/dist/post-training.js +47 -0
- package/dist/post-training.js.map +7 -0
- package/dist/taskset-drafts.js +11 -4
- package/dist/taskset-drafts.js.map +2 -2
- package/dist/taskset-packages.js +26 -556
- package/dist/taskset-packages.js.map +4 -4
- package/dist/training-bundle.js +58 -17
- package/dist/training-bundle.js.map +2 -2
- package/dist/training-projects.js +1 -1
- package/dist/training.js +7 -2
- package/dist/types/packages/sdk/src/advanced-refiner-evaluations.d.ts +2673 -0
- package/dist/types/packages/sdk/src/advanced-refiner-evaluations.d.ts.map +1 -0
- package/dist/types/packages/sdk/src/candidate-evaluation-client.d.ts +1079 -0
- package/dist/types/packages/sdk/src/candidate-evaluation-client.d.ts.map +1 -0
- package/dist/types/packages/sdk/src/candidate-evaluations.d.ts +845 -0
- package/dist/types/packages/sdk/src/candidate-evaluations.d.ts.map +1 -0
- package/dist/types/packages/sdk/src/claude-code-runtime.d.ts +51 -0
- package/dist/types/packages/sdk/src/claude-code-runtime.d.ts.map +1 -0
- package/dist/types/packages/sdk/src/connected-dataset-publication.d.ts +15 -0
- package/dist/types/packages/sdk/src/connected-dataset-publication.d.ts.map +1 -0
- package/dist/types/packages/sdk/src/connected-dataset-source.d.ts +48 -0
- package/dist/types/packages/sdk/src/connected-dataset-source.d.ts.map +1 -0
- package/dist/types/packages/sdk/src/connected-evidence-client.d.ts +2460 -0
- package/dist/types/packages/sdk/src/connected-evidence-client.d.ts.map +1 -0
- package/dist/types/packages/sdk/src/connected-evidence-contracts.d.ts +215 -0
- package/dist/types/packages/sdk/src/connected-evidence-contracts.d.ts.map +1 -0
- package/dist/types/packages/sdk/src/connected-evidence-http.d.ts +6 -0
- package/dist/types/packages/sdk/src/connected-evidence-http.d.ts.map +1 -0
- package/dist/types/packages/sdk/src/connected-evidence.d.ts +6 -0
- package/dist/types/packages/sdk/src/connected-evidence.d.ts.map +1 -0
- package/dist/types/packages/sdk/src/connected-recorded-execution.d.ts +690 -0
- package/dist/types/packages/sdk/src/connected-recorded-execution.d.ts.map +1 -0
- package/dist/types/packages/sdk/src/dataset-workspace-client.d.ts +771 -120
- package/dist/types/packages/sdk/src/dataset-workspace-client.d.ts.map +1 -1
- package/dist/types/packages/sdk/src/dataset-workspace-contracts.d.ts +278 -38
- package/dist/types/packages/sdk/src/dataset-workspace-contracts.d.ts.map +1 -1
- package/dist/types/packages/sdk/src/dataset-workspace-operations.d.ts +280 -40
- package/dist/types/packages/sdk/src/dataset-workspace-operations.d.ts.map +1 -1
- package/dist/types/packages/sdk/src/evaluation-coordination-http.d.ts +17 -0
- package/dist/types/packages/sdk/src/evaluation-coordination-http.d.ts.map +1 -0
- package/dist/types/packages/sdk/src/experiment-client.d.ts +3513 -52
- package/dist/types/packages/sdk/src/experiment-client.d.ts.map +1 -1
- package/dist/types/packages/sdk/src/experiment-contracts.d.ts +1017 -1
- package/dist/types/packages/sdk/src/experiment-contracts.d.ts.map +1 -1
- package/dist/types/packages/sdk/src/experiment-evaluation-schedules.d.ts +2780 -0
- package/dist/types/packages/sdk/src/experiment-evaluation-schedules.d.ts.map +1 -0
- package/dist/types/packages/sdk/src/experiment-grading-contracts.d.ts +8 -0
- package/dist/types/packages/sdk/src/experiment-grading-contracts.d.ts.map +1 -1
- package/dist/types/packages/sdk/src/experiment-improvement-client.d.ts +5724 -0
- package/dist/types/packages/sdk/src/experiment-improvement-client.d.ts.map +1 -0
- package/dist/types/packages/sdk/src/experiment-improvement-contracts.d.ts +13024 -0
- package/dist/types/packages/sdk/src/experiment-improvement-contracts.d.ts.map +1 -0
- package/dist/types/packages/sdk/src/experiment-improvement-hosted-adoption.d.ts +1361 -0
- package/dist/types/packages/sdk/src/experiment-improvement-hosted-adoption.d.ts.map +1 -0
- package/dist/types/packages/sdk/src/experiment-improvements.d.ts +4 -0
- package/dist/types/packages/sdk/src/experiment-improvements.d.ts.map +1 -0
- package/dist/types/packages/sdk/src/experiment-run-contracts.d.ts +262 -0
- package/dist/types/packages/sdk/src/experiment-run-contracts.d.ts.map +1 -1
- package/dist/types/packages/sdk/src/experiment-run-details.d.ts +1388 -1
- package/dist/types/packages/sdk/src/experiment-run-details.d.ts.map +1 -1
- package/dist/types/packages/sdk/src/experiment-scoring-contracts.d.ts +5 -2
- package/dist/types/packages/sdk/src/experiment-scoring-contracts.d.ts.map +1 -1
- package/dist/types/packages/sdk/src/experiments.d.ts +3 -0
- package/dist/types/packages/sdk/src/experiments.d.ts.map +1 -1
- package/dist/types/packages/sdk/src/grader-inspection.d.ts +307 -5
- package/dist/types/packages/sdk/src/grader-inspection.d.ts.map +1 -1
- package/dist/types/packages/sdk/src/harness-experiment-policy.d.ts +238 -0
- package/dist/types/packages/sdk/src/harness-experiment-policy.d.ts.map +1 -1
- package/dist/types/packages/sdk/src/human-review-client.d.ts +1267 -0
- package/dist/types/packages/sdk/src/human-review-client.d.ts.map +1 -0
- package/dist/types/packages/sdk/src/human-review.d.ts +3 -0
- package/dist/types/packages/sdk/src/human-review.d.ts.map +1 -0
- package/dist/types/packages/sdk/src/index.d.ts +10 -0
- package/dist/types/packages/sdk/src/index.d.ts.map +1 -1
- package/dist/types/packages/sdk/src/managed-training-preparation.d.ts +181 -0
- package/dist/types/packages/sdk/src/managed-training-preparation.d.ts.map +1 -1
- package/dist/types/packages/sdk/src/model-batch-review-contracts.d.ts +20 -4
- package/dist/types/packages/sdk/src/model-batch-review-contracts.d.ts.map +1 -1
- package/dist/types/packages/sdk/src/model-batch-review-preparation.d.ts +79 -5
- package/dist/types/packages/sdk/src/model-batch-review-preparation.d.ts.map +1 -1
- package/dist/types/packages/sdk/src/model-batch-review.d.ts +15 -13
- package/dist/types/packages/sdk/src/model-batch-review.d.ts.map +1 -1
- package/dist/types/packages/sdk/src/model-learning-policy.d.ts +359 -1
- package/dist/types/packages/sdk/src/model-learning-policy.d.ts.map +1 -1
- package/dist/types/packages/sdk/src/model-starter-attempts.d.ts +288 -19
- package/dist/types/packages/sdk/src/model-starter-attempts.d.ts.map +1 -1
- package/dist/types/packages/sdk/src/model-starter-catalog.d.ts +234 -12
- package/dist/types/packages/sdk/src/model-starter-catalog.d.ts.map +1 -1
- package/dist/types/packages/sdk/src/model-starter-execution.d.ts +78 -1
- package/dist/types/packages/sdk/src/model-starter-execution.d.ts.map +1 -1
- package/dist/types/packages/sdk/src/model-starters.d.ts +466 -13
- package/dist/types/packages/sdk/src/model-starters.d.ts.map +1 -1
- package/dist/types/packages/sdk/src/model-taskset-derivation.d.ts +312 -4
- package/dist/types/packages/sdk/src/model-taskset-derivation.d.ts.map +1 -1
- package/dist/types/packages/sdk/src/model-taskset-resources.d.ts +78 -1
- package/dist/types/packages/sdk/src/model-taskset-resources.d.ts.map +1 -1
- package/dist/types/packages/sdk/src/model-taskset-runs-contracts.d.ts +1282 -4
- package/dist/types/packages/sdk/src/model-taskset-runs-contracts.d.ts.map +1 -1
- package/dist/types/packages/sdk/src/model-taskset-runs.d.ts +1285 -5
- package/dist/types/packages/sdk/src/model-taskset-runs.d.ts.map +1 -1
- package/dist/types/packages/sdk/src/post-training.d.ts +3518 -0
- package/dist/types/packages/sdk/src/post-training.d.ts.map +1 -0
- package/dist/types/packages/sdk/src/profile-evaluation-discovery.d.ts +1127 -0
- package/dist/types/packages/sdk/src/profile-evaluation-discovery.d.ts.map +1 -0
- package/dist/types/packages/sdk/src/profile-external-dataset-package.d.ts +1369 -0
- package/dist/types/packages/sdk/src/profile-external-dataset-package.d.ts.map +1 -0
- package/dist/types/packages/sdk/src/taskset-authored-contracts.d.ts +142 -22
- package/dist/types/packages/sdk/src/taskset-authored-contracts.d.ts.map +1 -1
- package/dist/types/packages/sdk/src/taskset-authored-files.d.ts +111 -18
- package/dist/types/packages/sdk/src/taskset-authored-files.d.ts.map +1 -1
- package/dist/types/packages/sdk/src/taskset-draft-api-contracts.d.ts +278 -38
- package/dist/types/packages/sdk/src/taskset-draft-api-contracts.d.ts.map +1 -1
- package/dist/types/packages/sdk/src/taskset-draft-client.d.ts +661 -103
- package/dist/types/packages/sdk/src/taskset-draft-client.d.ts.map +1 -1
- package/dist/types/packages/sdk/src/taskset-draft-core.d.ts +217 -20
- package/dist/types/packages/sdk/src/taskset-draft-core.d.ts.map +1 -1
- package/dist/types/packages/sdk/src/taskset-draft-dataset-artifacts.d.ts +12 -12
- package/dist/types/packages/sdk/src/taskset-draft-dataset-sources.d.ts +17 -17
- package/dist/types/packages/sdk/src/taskset-draft-document.d.ts +140 -20
- package/dist/types/packages/sdk/src/taskset-draft-document.d.ts.map +1 -1
- package/dist/types/packages/sdk/src/taskset-draft-package-compiler.d.ts +638 -42
- package/dist/types/packages/sdk/src/taskset-draft-package-compiler.d.ts.map +1 -1
- package/dist/types/packages/sdk/src/taskset-draft-source.d.ts +249 -36
- package/dist/types/packages/sdk/src/taskset-draft-source.d.ts.map +1 -1
- package/dist/types/packages/sdk/src/taskset-draft-workspace.d.ts +278 -38
- package/dist/types/packages/sdk/src/taskset-draft-workspace.d.ts.map +1 -1
- package/dist/types/packages/sdk/src/taskset-drafts.d.ts +1 -0
- package/dist/types/packages/sdk/src/taskset-drafts.d.ts.map +1 -1
- package/dist/types/packages/sdk/src/taskset-package-client.d.ts +1322 -26
- package/dist/types/packages/sdk/src/taskset-package-client.d.ts.map +1 -1
- package/dist/types/packages/sdk/src/taskset-package-contracts.d.ts +660 -12
- package/dist/types/packages/sdk/src/taskset-package-contracts.d.ts.map +1 -1
- package/dist/types/packages/sdk/src/taskset-package-grader-projection.d.ts +77 -3
- package/dist/types/packages/sdk/src/taskset-package-grader-projection.d.ts.map +1 -1
- package/dist/types/packages/sdk/src/taskset-package-learning.d.ts +29 -11
- package/dist/types/packages/sdk/src/taskset-package-learning.d.ts.map +1 -1
- package/dist/types/packages/sdk/src/taskset-package-projection.d.ts +430 -39
- package/dist/types/packages/sdk/src/taskset-package-projection.d.ts.map +1 -1
- package/dist/types/packages/sdk/src/training-acceptance.d.ts +10 -10
- package/dist/types/packages/sdk/src/training-candidate-decisions.d.ts +2 -2
- package/dist/types/packages/sdk/src/training-evaluation-results.d.ts +1 -1
- package/dist/types/packages/sdk/src/training-learning-batch.d.ts +376 -42
- package/dist/types/packages/sdk/src/training-learning-batch.d.ts.map +1 -1
- package/dist/types/packages/sdk/src/training-preparations.d.ts +3930 -0
- package/dist/types/packages/sdk/src/training-preparations.d.ts.map +1 -1
- package/dist/types/packages/sdk/src/training-project-client.d.ts +776 -8
- package/dist/types/packages/sdk/src/training-project-client.d.ts.map +1 -1
- package/dist/types/packages/sdk/src/training-project-contracts.d.ts +1161 -11
- package/dist/types/packages/sdk/src/training-project-contracts.d.ts.map +1 -1
- package/dist/types/packages/sdk/src/training.d.ts +2655 -24
- package/dist/types/packages/sdk/src/training.d.ts.map +1 -1
- package/package.json +30 -2
- package/dist/chunk-2ZJRG3KL.js.map +0 -7
- package/dist/chunk-7B44TQZD.js.map +0 -7
- package/dist/chunk-JCPLAIZL.js.map +0 -7
- package/dist/chunk-LGEF4S5I.js.map +0 -7
- package/dist/chunk-PRRUBH2B.js +0 -314
- package/dist/chunk-PRRUBH2B.js.map +0 -7
- package/dist/chunk-W5TXWVFS.js.map +0 -7
- package/dist/chunk-Y4NGWGTA.js.map +0 -7
- /package/dist/{chunk-DT7FRNIZ.js.map → chunk-CSWVE3E6.js.map} +0 -0
- /package/dist/{chunk-I3TQTXKB.js.map → chunk-IXBSFJCX.js.map} +0 -0
- /package/dist/{chunk-I2YM3UY4.js.map → chunk-NHCGOFQW.js.map} +0 -0
- /package/dist/{chunk-32H2EJRA.js.map → chunk-R3M6BIOK.js.map} +0 -0
|
@@ -60,6 +60,30 @@ export declare class OpenPondExperimentsClient {
|
|
|
60
60
|
maxOutputTokens: number;
|
|
61
61
|
temperature: number;
|
|
62
62
|
topP: number;
|
|
63
|
+
localRuntime?: {
|
|
64
|
+
providerId: "claude-code";
|
|
65
|
+
configurationHash: string;
|
|
66
|
+
executableHash: string;
|
|
67
|
+
version: string;
|
|
68
|
+
capabilityHash: string;
|
|
69
|
+
maximumRequestCostUsd: number;
|
|
70
|
+
requestTimeoutMs: number;
|
|
71
|
+
maximumTurns: number;
|
|
72
|
+
holdOpen: boolean;
|
|
73
|
+
} | {
|
|
74
|
+
providerId: "custom-openai-compatible";
|
|
75
|
+
configurationHash: string;
|
|
76
|
+
maximumRequestCostUsd: number;
|
|
77
|
+
requestTimeoutMs: number;
|
|
78
|
+
} | undefined;
|
|
79
|
+
candidate?: {
|
|
80
|
+
artifactId: string;
|
|
81
|
+
contentHash: string;
|
|
82
|
+
preparationId: string;
|
|
83
|
+
dispatchId: string;
|
|
84
|
+
baseProfileId: string;
|
|
85
|
+
workerImageDigest: string;
|
|
86
|
+
} | undefined;
|
|
63
87
|
harness?: {
|
|
64
88
|
harnessRelease: {
|
|
65
89
|
id: string;
|
|
@@ -102,10 +126,210 @@ export declare class OpenPondExperimentsClient {
|
|
|
102
126
|
actionId: string;
|
|
103
127
|
};
|
|
104
128
|
environmentHash: string;
|
|
129
|
+
externalDatasetBinding?: {
|
|
130
|
+
schemaVersion: "openpond.profileExternalDatasetBinding.v1";
|
|
131
|
+
dataset: {
|
|
132
|
+
id: string;
|
|
133
|
+
contentHash: string;
|
|
134
|
+
revision: number;
|
|
135
|
+
};
|
|
136
|
+
packageHash: string;
|
|
137
|
+
profileId: string;
|
|
138
|
+
target: {
|
|
139
|
+
kind: "profile";
|
|
140
|
+
} | {
|
|
141
|
+
kind: "workflow";
|
|
142
|
+
workflowId: string;
|
|
143
|
+
} | {
|
|
144
|
+
kind: "skill";
|
|
145
|
+
skillPath: string;
|
|
146
|
+
} | {
|
|
147
|
+
kind: "agent_action";
|
|
148
|
+
actionId: string;
|
|
149
|
+
};
|
|
150
|
+
declaredProfileCatalogHash: string | null;
|
|
151
|
+
protectedProfileClosureHash: string;
|
|
152
|
+
split: "train" | "validation" | "test" | "frozen_eval";
|
|
153
|
+
population: {
|
|
154
|
+
taskId: string;
|
|
155
|
+
seed: string;
|
|
156
|
+
fixtureId: string | null;
|
|
157
|
+
}[];
|
|
158
|
+
evaluators: {
|
|
159
|
+
release: {
|
|
160
|
+
id: string;
|
|
161
|
+
contentHash: string;
|
|
162
|
+
revision: string | number;
|
|
163
|
+
};
|
|
164
|
+
feedbackKey: string;
|
|
165
|
+
configurationHash: string;
|
|
166
|
+
output: "boolean" | "score" | "category";
|
|
167
|
+
categories: string[];
|
|
168
|
+
}[];
|
|
169
|
+
fieldMappingsHash: string;
|
|
170
|
+
fieldMappings: {
|
|
171
|
+
graderId: string;
|
|
172
|
+
fields: {
|
|
173
|
+
destination: "input" | "output" | "expectedOutput" | "evaluatorContext";
|
|
174
|
+
path: string[];
|
|
175
|
+
source: "input" | "output" | "expectedOutput" | "evaluatorContext";
|
|
176
|
+
sourcePath: string[];
|
|
177
|
+
}[];
|
|
178
|
+
}[];
|
|
179
|
+
environmentHash: string;
|
|
180
|
+
privateDatasetClosureHash: string;
|
|
181
|
+
criterion: {
|
|
182
|
+
minimumPassRate: number;
|
|
183
|
+
requireComplete: boolean;
|
|
184
|
+
};
|
|
185
|
+
contentHash: string;
|
|
186
|
+
recordedOrigin?: {
|
|
187
|
+
execution: {
|
|
188
|
+
id: string;
|
|
189
|
+
contentHash: string;
|
|
190
|
+
};
|
|
191
|
+
originalEvidence: {
|
|
192
|
+
id: string;
|
|
193
|
+
contentHash: string;
|
|
194
|
+
};
|
|
195
|
+
members: {
|
|
196
|
+
caseId: string;
|
|
197
|
+
seed: string;
|
|
198
|
+
fixtureId: string;
|
|
199
|
+
sourceId: string;
|
|
200
|
+
snapshotHash: string;
|
|
201
|
+
boundaryRevisionHash: string;
|
|
202
|
+
}[];
|
|
203
|
+
} | undefined;
|
|
204
|
+
gradingSource?: {
|
|
205
|
+
pass: {
|
|
206
|
+
id: string;
|
|
207
|
+
contentHash: string;
|
|
208
|
+
};
|
|
209
|
+
evidence: {
|
|
210
|
+
id: string;
|
|
211
|
+
contentHash: string;
|
|
212
|
+
};
|
|
213
|
+
} | undefined;
|
|
214
|
+
} | undefined;
|
|
215
|
+
sourceCandidate?: {
|
|
216
|
+
candidateId: string;
|
|
217
|
+
frozenRevision: number;
|
|
218
|
+
frozenHash: string;
|
|
219
|
+
harnessRelease: {
|
|
220
|
+
id: string;
|
|
221
|
+
contentHash: string;
|
|
222
|
+
};
|
|
223
|
+
protectedClosureHash: string;
|
|
224
|
+
} | undefined;
|
|
105
225
|
};
|
|
106
226
|
modelConfigurationHash: string;
|
|
107
227
|
packageHash: string;
|
|
228
|
+
candidate?: {
|
|
229
|
+
artifactId: string;
|
|
230
|
+
contentHash: string;
|
|
231
|
+
preparationId: string;
|
|
232
|
+
dispatchId: string;
|
|
233
|
+
baseProfileId: string;
|
|
234
|
+
workerImageDigest: string;
|
|
235
|
+
} | undefined;
|
|
236
|
+
sourceCandidate?: {
|
|
237
|
+
candidateId: string;
|
|
238
|
+
frozenRevision: number;
|
|
239
|
+
frozenHash: string;
|
|
240
|
+
harnessRelease: {
|
|
241
|
+
id: string;
|
|
242
|
+
contentHash: string;
|
|
243
|
+
};
|
|
244
|
+
protectedClosureHash: string;
|
|
245
|
+
} | undefined;
|
|
108
246
|
reasoningEffort?: "high" | "low" | "off" | "medium" | "xhigh" | "max" | undefined;
|
|
247
|
+
externalDatasetBinding?: {
|
|
248
|
+
schemaVersion: "openpond.profileExternalDatasetBinding.v1";
|
|
249
|
+
dataset: {
|
|
250
|
+
id: string;
|
|
251
|
+
contentHash: string;
|
|
252
|
+
revision: number;
|
|
253
|
+
};
|
|
254
|
+
packageHash: string;
|
|
255
|
+
profileId: string;
|
|
256
|
+
target: {
|
|
257
|
+
kind: "profile";
|
|
258
|
+
} | {
|
|
259
|
+
kind: "workflow";
|
|
260
|
+
workflowId: string;
|
|
261
|
+
} | {
|
|
262
|
+
kind: "skill";
|
|
263
|
+
skillPath: string;
|
|
264
|
+
} | {
|
|
265
|
+
kind: "agent_action";
|
|
266
|
+
actionId: string;
|
|
267
|
+
};
|
|
268
|
+
declaredProfileCatalogHash: string | null;
|
|
269
|
+
protectedProfileClosureHash: string;
|
|
270
|
+
split: "train" | "validation" | "test" | "frozen_eval";
|
|
271
|
+
population: {
|
|
272
|
+
taskId: string;
|
|
273
|
+
seed: string;
|
|
274
|
+
fixtureId: string | null;
|
|
275
|
+
}[];
|
|
276
|
+
evaluators: {
|
|
277
|
+
release: {
|
|
278
|
+
id: string;
|
|
279
|
+
contentHash: string;
|
|
280
|
+
revision: string | number;
|
|
281
|
+
};
|
|
282
|
+
feedbackKey: string;
|
|
283
|
+
configurationHash: string;
|
|
284
|
+
output: "boolean" | "score" | "category";
|
|
285
|
+
categories: string[];
|
|
286
|
+
}[];
|
|
287
|
+
fieldMappingsHash: string;
|
|
288
|
+
fieldMappings: {
|
|
289
|
+
graderId: string;
|
|
290
|
+
fields: {
|
|
291
|
+
destination: "input" | "output" | "expectedOutput" | "evaluatorContext";
|
|
292
|
+
path: string[];
|
|
293
|
+
source: "input" | "output" | "expectedOutput" | "evaluatorContext";
|
|
294
|
+
sourcePath: string[];
|
|
295
|
+
}[];
|
|
296
|
+
}[];
|
|
297
|
+
environmentHash: string;
|
|
298
|
+
privateDatasetClosureHash: string;
|
|
299
|
+
criterion: {
|
|
300
|
+
minimumPassRate: number;
|
|
301
|
+
requireComplete: boolean;
|
|
302
|
+
};
|
|
303
|
+
contentHash: string;
|
|
304
|
+
recordedOrigin?: {
|
|
305
|
+
execution: {
|
|
306
|
+
id: string;
|
|
307
|
+
contentHash: string;
|
|
308
|
+
};
|
|
309
|
+
originalEvidence: {
|
|
310
|
+
id: string;
|
|
311
|
+
contentHash: string;
|
|
312
|
+
};
|
|
313
|
+
members: {
|
|
314
|
+
caseId: string;
|
|
315
|
+
seed: string;
|
|
316
|
+
fixtureId: string;
|
|
317
|
+
sourceId: string;
|
|
318
|
+
snapshotHash: string;
|
|
319
|
+
boundaryRevisionHash: string;
|
|
320
|
+
}[];
|
|
321
|
+
} | undefined;
|
|
322
|
+
gradingSource?: {
|
|
323
|
+
pass: {
|
|
324
|
+
id: string;
|
|
325
|
+
contentHash: string;
|
|
326
|
+
};
|
|
327
|
+
evidence: {
|
|
328
|
+
id: string;
|
|
329
|
+
contentHash: string;
|
|
330
|
+
};
|
|
331
|
+
} | undefined;
|
|
332
|
+
} | undefined;
|
|
109
333
|
} | {
|
|
110
334
|
kind: "fixture";
|
|
111
335
|
};
|
|
@@ -222,6 +446,102 @@ export declare class OpenPondExperimentsClient {
|
|
|
222
446
|
actionId: string;
|
|
223
447
|
};
|
|
224
448
|
environmentHash: string;
|
|
449
|
+
externalDatasetBinding?: {
|
|
450
|
+
schemaVersion: "openpond.profileExternalDatasetBinding.v1";
|
|
451
|
+
dataset: {
|
|
452
|
+
id: string;
|
|
453
|
+
contentHash: string;
|
|
454
|
+
revision: number;
|
|
455
|
+
};
|
|
456
|
+
packageHash: string;
|
|
457
|
+
profileId: string;
|
|
458
|
+
target: {
|
|
459
|
+
kind: "profile";
|
|
460
|
+
} | {
|
|
461
|
+
kind: "workflow";
|
|
462
|
+
workflowId: string;
|
|
463
|
+
} | {
|
|
464
|
+
kind: "skill";
|
|
465
|
+
skillPath: string;
|
|
466
|
+
} | {
|
|
467
|
+
kind: "agent_action";
|
|
468
|
+
actionId: string;
|
|
469
|
+
};
|
|
470
|
+
declaredProfileCatalogHash: string | null;
|
|
471
|
+
protectedProfileClosureHash: string;
|
|
472
|
+
split: "train" | "validation" | "test" | "frozen_eval";
|
|
473
|
+
population: {
|
|
474
|
+
taskId: string;
|
|
475
|
+
seed: string;
|
|
476
|
+
fixtureId: string | null;
|
|
477
|
+
}[];
|
|
478
|
+
evaluators: {
|
|
479
|
+
release: {
|
|
480
|
+
id: string;
|
|
481
|
+
contentHash: string;
|
|
482
|
+
revision: string | number;
|
|
483
|
+
};
|
|
484
|
+
feedbackKey: string;
|
|
485
|
+
configurationHash: string;
|
|
486
|
+
output: "boolean" | "score" | "category";
|
|
487
|
+
categories: string[];
|
|
488
|
+
}[];
|
|
489
|
+
fieldMappingsHash: string;
|
|
490
|
+
fieldMappings: {
|
|
491
|
+
graderId: string;
|
|
492
|
+
fields: {
|
|
493
|
+
destination: "input" | "output" | "expectedOutput" | "evaluatorContext";
|
|
494
|
+
path: string[];
|
|
495
|
+
source: "input" | "output" | "expectedOutput" | "evaluatorContext";
|
|
496
|
+
sourcePath: string[];
|
|
497
|
+
}[];
|
|
498
|
+
}[];
|
|
499
|
+
environmentHash: string;
|
|
500
|
+
privateDatasetClosureHash: string;
|
|
501
|
+
criterion: {
|
|
502
|
+
minimumPassRate: number;
|
|
503
|
+
requireComplete: boolean;
|
|
504
|
+
};
|
|
505
|
+
contentHash: string;
|
|
506
|
+
recordedOrigin?: {
|
|
507
|
+
execution: {
|
|
508
|
+
id: string;
|
|
509
|
+
contentHash: string;
|
|
510
|
+
};
|
|
511
|
+
originalEvidence: {
|
|
512
|
+
id: string;
|
|
513
|
+
contentHash: string;
|
|
514
|
+
};
|
|
515
|
+
members: {
|
|
516
|
+
caseId: string;
|
|
517
|
+
seed: string;
|
|
518
|
+
fixtureId: string;
|
|
519
|
+
sourceId: string;
|
|
520
|
+
snapshotHash: string;
|
|
521
|
+
boundaryRevisionHash: string;
|
|
522
|
+
}[];
|
|
523
|
+
} | undefined;
|
|
524
|
+
gradingSource?: {
|
|
525
|
+
pass: {
|
|
526
|
+
id: string;
|
|
527
|
+
contentHash: string;
|
|
528
|
+
};
|
|
529
|
+
evidence: {
|
|
530
|
+
id: string;
|
|
531
|
+
contentHash: string;
|
|
532
|
+
};
|
|
533
|
+
} | undefined;
|
|
534
|
+
} | undefined;
|
|
535
|
+
sourceCandidate?: {
|
|
536
|
+
candidateId: string;
|
|
537
|
+
frozenRevision: number;
|
|
538
|
+
frozenHash: string;
|
|
539
|
+
harnessRelease: {
|
|
540
|
+
id: string;
|
|
541
|
+
contentHash: string;
|
|
542
|
+
};
|
|
543
|
+
protectedClosureHash: string;
|
|
544
|
+
} | undefined;
|
|
225
545
|
} | undefined;
|
|
226
546
|
};
|
|
227
547
|
maximumCostUsd: number;
|
|
@@ -259,7 +579,7 @@ export declare class OpenPondExperimentsClient {
|
|
|
259
579
|
};
|
|
260
580
|
policyKind: "hosted_chat" | "fixture" | "hosted_harness";
|
|
261
581
|
manifestHash: string;
|
|
262
|
-
status: "
|
|
582
|
+
status: "cancelled" | "failed" | "queued" | "running" | "cancelling" | "completed";
|
|
263
583
|
totalCount: number;
|
|
264
584
|
counts: {
|
|
265
585
|
pending: number;
|
|
@@ -303,6 +623,30 @@ export declare class OpenPondExperimentsClient {
|
|
|
303
623
|
maxOutputTokens: number;
|
|
304
624
|
temperature: number;
|
|
305
625
|
topP: number;
|
|
626
|
+
localRuntime?: {
|
|
627
|
+
providerId: "claude-code";
|
|
628
|
+
configurationHash: string;
|
|
629
|
+
executableHash: string;
|
|
630
|
+
version: string;
|
|
631
|
+
capabilityHash: string;
|
|
632
|
+
maximumRequestCostUsd: number;
|
|
633
|
+
requestTimeoutMs: number;
|
|
634
|
+
maximumTurns: number;
|
|
635
|
+
holdOpen: boolean;
|
|
636
|
+
} | {
|
|
637
|
+
providerId: "custom-openai-compatible";
|
|
638
|
+
configurationHash: string;
|
|
639
|
+
maximumRequestCostUsd: number;
|
|
640
|
+
requestTimeoutMs: number;
|
|
641
|
+
} | undefined;
|
|
642
|
+
candidate?: {
|
|
643
|
+
artifactId: string;
|
|
644
|
+
contentHash: string;
|
|
645
|
+
preparationId: string;
|
|
646
|
+
dispatchId: string;
|
|
647
|
+
baseProfileId: string;
|
|
648
|
+
workerImageDigest: string;
|
|
649
|
+
} | undefined;
|
|
306
650
|
harness?: {
|
|
307
651
|
harnessRelease: {
|
|
308
652
|
id: string;
|
|
@@ -345,10 +689,210 @@ export declare class OpenPondExperimentsClient {
|
|
|
345
689
|
actionId: string;
|
|
346
690
|
};
|
|
347
691
|
environmentHash: string;
|
|
692
|
+
externalDatasetBinding?: {
|
|
693
|
+
schemaVersion: "openpond.profileExternalDatasetBinding.v1";
|
|
694
|
+
dataset: {
|
|
695
|
+
id: string;
|
|
696
|
+
contentHash: string;
|
|
697
|
+
revision: number;
|
|
698
|
+
};
|
|
699
|
+
packageHash: string;
|
|
700
|
+
profileId: string;
|
|
701
|
+
target: {
|
|
702
|
+
kind: "profile";
|
|
703
|
+
} | {
|
|
704
|
+
kind: "workflow";
|
|
705
|
+
workflowId: string;
|
|
706
|
+
} | {
|
|
707
|
+
kind: "skill";
|
|
708
|
+
skillPath: string;
|
|
709
|
+
} | {
|
|
710
|
+
kind: "agent_action";
|
|
711
|
+
actionId: string;
|
|
712
|
+
};
|
|
713
|
+
declaredProfileCatalogHash: string | null;
|
|
714
|
+
protectedProfileClosureHash: string;
|
|
715
|
+
split: "train" | "validation" | "test" | "frozen_eval";
|
|
716
|
+
population: {
|
|
717
|
+
taskId: string;
|
|
718
|
+
seed: string;
|
|
719
|
+
fixtureId: string | null;
|
|
720
|
+
}[];
|
|
721
|
+
evaluators: {
|
|
722
|
+
release: {
|
|
723
|
+
id: string;
|
|
724
|
+
contentHash: string;
|
|
725
|
+
revision: string | number;
|
|
726
|
+
};
|
|
727
|
+
feedbackKey: string;
|
|
728
|
+
configurationHash: string;
|
|
729
|
+
output: "boolean" | "score" | "category";
|
|
730
|
+
categories: string[];
|
|
731
|
+
}[];
|
|
732
|
+
fieldMappingsHash: string;
|
|
733
|
+
fieldMappings: {
|
|
734
|
+
graderId: string;
|
|
735
|
+
fields: {
|
|
736
|
+
destination: "input" | "output" | "expectedOutput" | "evaluatorContext";
|
|
737
|
+
path: string[];
|
|
738
|
+
source: "input" | "output" | "expectedOutput" | "evaluatorContext";
|
|
739
|
+
sourcePath: string[];
|
|
740
|
+
}[];
|
|
741
|
+
}[];
|
|
742
|
+
environmentHash: string;
|
|
743
|
+
privateDatasetClosureHash: string;
|
|
744
|
+
criterion: {
|
|
745
|
+
minimumPassRate: number;
|
|
746
|
+
requireComplete: boolean;
|
|
747
|
+
};
|
|
748
|
+
contentHash: string;
|
|
749
|
+
recordedOrigin?: {
|
|
750
|
+
execution: {
|
|
751
|
+
id: string;
|
|
752
|
+
contentHash: string;
|
|
753
|
+
};
|
|
754
|
+
originalEvidence: {
|
|
755
|
+
id: string;
|
|
756
|
+
contentHash: string;
|
|
757
|
+
};
|
|
758
|
+
members: {
|
|
759
|
+
caseId: string;
|
|
760
|
+
seed: string;
|
|
761
|
+
fixtureId: string;
|
|
762
|
+
sourceId: string;
|
|
763
|
+
snapshotHash: string;
|
|
764
|
+
boundaryRevisionHash: string;
|
|
765
|
+
}[];
|
|
766
|
+
} | undefined;
|
|
767
|
+
gradingSource?: {
|
|
768
|
+
pass: {
|
|
769
|
+
id: string;
|
|
770
|
+
contentHash: string;
|
|
771
|
+
};
|
|
772
|
+
evidence: {
|
|
773
|
+
id: string;
|
|
774
|
+
contentHash: string;
|
|
775
|
+
};
|
|
776
|
+
} | undefined;
|
|
777
|
+
} | undefined;
|
|
778
|
+
sourceCandidate?: {
|
|
779
|
+
candidateId: string;
|
|
780
|
+
frozenRevision: number;
|
|
781
|
+
frozenHash: string;
|
|
782
|
+
harnessRelease: {
|
|
783
|
+
id: string;
|
|
784
|
+
contentHash: string;
|
|
785
|
+
};
|
|
786
|
+
protectedClosureHash: string;
|
|
787
|
+
} | undefined;
|
|
348
788
|
};
|
|
349
789
|
modelConfigurationHash: string;
|
|
350
790
|
packageHash: string;
|
|
791
|
+
candidate?: {
|
|
792
|
+
artifactId: string;
|
|
793
|
+
contentHash: string;
|
|
794
|
+
preparationId: string;
|
|
795
|
+
dispatchId: string;
|
|
796
|
+
baseProfileId: string;
|
|
797
|
+
workerImageDigest: string;
|
|
798
|
+
} | undefined;
|
|
799
|
+
sourceCandidate?: {
|
|
800
|
+
candidateId: string;
|
|
801
|
+
frozenRevision: number;
|
|
802
|
+
frozenHash: string;
|
|
803
|
+
harnessRelease: {
|
|
804
|
+
id: string;
|
|
805
|
+
contentHash: string;
|
|
806
|
+
};
|
|
807
|
+
protectedClosureHash: string;
|
|
808
|
+
} | undefined;
|
|
351
809
|
reasoningEffort?: "high" | "low" | "off" | "medium" | "xhigh" | "max" | undefined;
|
|
810
|
+
externalDatasetBinding?: {
|
|
811
|
+
schemaVersion: "openpond.profileExternalDatasetBinding.v1";
|
|
812
|
+
dataset: {
|
|
813
|
+
id: string;
|
|
814
|
+
contentHash: string;
|
|
815
|
+
revision: number;
|
|
816
|
+
};
|
|
817
|
+
packageHash: string;
|
|
818
|
+
profileId: string;
|
|
819
|
+
target: {
|
|
820
|
+
kind: "profile";
|
|
821
|
+
} | {
|
|
822
|
+
kind: "workflow";
|
|
823
|
+
workflowId: string;
|
|
824
|
+
} | {
|
|
825
|
+
kind: "skill";
|
|
826
|
+
skillPath: string;
|
|
827
|
+
} | {
|
|
828
|
+
kind: "agent_action";
|
|
829
|
+
actionId: string;
|
|
830
|
+
};
|
|
831
|
+
declaredProfileCatalogHash: string | null;
|
|
832
|
+
protectedProfileClosureHash: string;
|
|
833
|
+
split: "train" | "validation" | "test" | "frozen_eval";
|
|
834
|
+
population: {
|
|
835
|
+
taskId: string;
|
|
836
|
+
seed: string;
|
|
837
|
+
fixtureId: string | null;
|
|
838
|
+
}[];
|
|
839
|
+
evaluators: {
|
|
840
|
+
release: {
|
|
841
|
+
id: string;
|
|
842
|
+
contentHash: string;
|
|
843
|
+
revision: string | number;
|
|
844
|
+
};
|
|
845
|
+
feedbackKey: string;
|
|
846
|
+
configurationHash: string;
|
|
847
|
+
output: "boolean" | "score" | "category";
|
|
848
|
+
categories: string[];
|
|
849
|
+
}[];
|
|
850
|
+
fieldMappingsHash: string;
|
|
851
|
+
fieldMappings: {
|
|
852
|
+
graderId: string;
|
|
853
|
+
fields: {
|
|
854
|
+
destination: "input" | "output" | "expectedOutput" | "evaluatorContext";
|
|
855
|
+
path: string[];
|
|
856
|
+
source: "input" | "output" | "expectedOutput" | "evaluatorContext";
|
|
857
|
+
sourcePath: string[];
|
|
858
|
+
}[];
|
|
859
|
+
}[];
|
|
860
|
+
environmentHash: string;
|
|
861
|
+
privateDatasetClosureHash: string;
|
|
862
|
+
criterion: {
|
|
863
|
+
minimumPassRate: number;
|
|
864
|
+
requireComplete: boolean;
|
|
865
|
+
};
|
|
866
|
+
contentHash: string;
|
|
867
|
+
recordedOrigin?: {
|
|
868
|
+
execution: {
|
|
869
|
+
id: string;
|
|
870
|
+
contentHash: string;
|
|
871
|
+
};
|
|
872
|
+
originalEvidence: {
|
|
873
|
+
id: string;
|
|
874
|
+
contentHash: string;
|
|
875
|
+
};
|
|
876
|
+
members: {
|
|
877
|
+
caseId: string;
|
|
878
|
+
seed: string;
|
|
879
|
+
fixtureId: string;
|
|
880
|
+
sourceId: string;
|
|
881
|
+
snapshotHash: string;
|
|
882
|
+
boundaryRevisionHash: string;
|
|
883
|
+
}[];
|
|
884
|
+
} | undefined;
|
|
885
|
+
gradingSource?: {
|
|
886
|
+
pass: {
|
|
887
|
+
id: string;
|
|
888
|
+
contentHash: string;
|
|
889
|
+
};
|
|
890
|
+
evidence: {
|
|
891
|
+
id: string;
|
|
892
|
+
contentHash: string;
|
|
893
|
+
};
|
|
894
|
+
} | undefined;
|
|
895
|
+
} | undefined;
|
|
352
896
|
} | {
|
|
353
897
|
kind: "fixture";
|
|
354
898
|
};
|
|
@@ -465,6 +1009,102 @@ export declare class OpenPondExperimentsClient {
|
|
|
465
1009
|
actionId: string;
|
|
466
1010
|
};
|
|
467
1011
|
environmentHash: string;
|
|
1012
|
+
externalDatasetBinding?: {
|
|
1013
|
+
schemaVersion: "openpond.profileExternalDatasetBinding.v1";
|
|
1014
|
+
dataset: {
|
|
1015
|
+
id: string;
|
|
1016
|
+
contentHash: string;
|
|
1017
|
+
revision: number;
|
|
1018
|
+
};
|
|
1019
|
+
packageHash: string;
|
|
1020
|
+
profileId: string;
|
|
1021
|
+
target: {
|
|
1022
|
+
kind: "profile";
|
|
1023
|
+
} | {
|
|
1024
|
+
kind: "workflow";
|
|
1025
|
+
workflowId: string;
|
|
1026
|
+
} | {
|
|
1027
|
+
kind: "skill";
|
|
1028
|
+
skillPath: string;
|
|
1029
|
+
} | {
|
|
1030
|
+
kind: "agent_action";
|
|
1031
|
+
actionId: string;
|
|
1032
|
+
};
|
|
1033
|
+
declaredProfileCatalogHash: string | null;
|
|
1034
|
+
protectedProfileClosureHash: string;
|
|
1035
|
+
split: "train" | "validation" | "test" | "frozen_eval";
|
|
1036
|
+
population: {
|
|
1037
|
+
taskId: string;
|
|
1038
|
+
seed: string;
|
|
1039
|
+
fixtureId: string | null;
|
|
1040
|
+
}[];
|
|
1041
|
+
evaluators: {
|
|
1042
|
+
release: {
|
|
1043
|
+
id: string;
|
|
1044
|
+
contentHash: string;
|
|
1045
|
+
revision: string | number;
|
|
1046
|
+
};
|
|
1047
|
+
feedbackKey: string;
|
|
1048
|
+
configurationHash: string;
|
|
1049
|
+
output: "boolean" | "score" | "category";
|
|
1050
|
+
categories: string[];
|
|
1051
|
+
}[];
|
|
1052
|
+
fieldMappingsHash: string;
|
|
1053
|
+
fieldMappings: {
|
|
1054
|
+
graderId: string;
|
|
1055
|
+
fields: {
|
|
1056
|
+
destination: "input" | "output" | "expectedOutput" | "evaluatorContext";
|
|
1057
|
+
path: string[];
|
|
1058
|
+
source: "input" | "output" | "expectedOutput" | "evaluatorContext";
|
|
1059
|
+
sourcePath: string[];
|
|
1060
|
+
}[];
|
|
1061
|
+
}[];
|
|
1062
|
+
environmentHash: string;
|
|
1063
|
+
privateDatasetClosureHash: string;
|
|
1064
|
+
criterion: {
|
|
1065
|
+
minimumPassRate: number;
|
|
1066
|
+
requireComplete: boolean;
|
|
1067
|
+
};
|
|
1068
|
+
contentHash: string;
|
|
1069
|
+
recordedOrigin?: {
|
|
1070
|
+
execution: {
|
|
1071
|
+
id: string;
|
|
1072
|
+
contentHash: string;
|
|
1073
|
+
};
|
|
1074
|
+
originalEvidence: {
|
|
1075
|
+
id: string;
|
|
1076
|
+
contentHash: string;
|
|
1077
|
+
};
|
|
1078
|
+
members: {
|
|
1079
|
+
caseId: string;
|
|
1080
|
+
seed: string;
|
|
1081
|
+
fixtureId: string;
|
|
1082
|
+
sourceId: string;
|
|
1083
|
+
snapshotHash: string;
|
|
1084
|
+
boundaryRevisionHash: string;
|
|
1085
|
+
}[];
|
|
1086
|
+
} | undefined;
|
|
1087
|
+
gradingSource?: {
|
|
1088
|
+
pass: {
|
|
1089
|
+
id: string;
|
|
1090
|
+
contentHash: string;
|
|
1091
|
+
};
|
|
1092
|
+
evidence: {
|
|
1093
|
+
id: string;
|
|
1094
|
+
contentHash: string;
|
|
1095
|
+
};
|
|
1096
|
+
} | undefined;
|
|
1097
|
+
} | undefined;
|
|
1098
|
+
sourceCandidate?: {
|
|
1099
|
+
candidateId: string;
|
|
1100
|
+
frozenRevision: number;
|
|
1101
|
+
frozenHash: string;
|
|
1102
|
+
harnessRelease: {
|
|
1103
|
+
id: string;
|
|
1104
|
+
contentHash: string;
|
|
1105
|
+
};
|
|
1106
|
+
protectedClosureHash: string;
|
|
1107
|
+
} | undefined;
|
|
468
1108
|
} | undefined;
|
|
469
1109
|
};
|
|
470
1110
|
policySnapshot: {
|
|
@@ -490,6 +1130,30 @@ export declare class OpenPondExperimentsClient {
|
|
|
490
1130
|
maxOutputTokens: number;
|
|
491
1131
|
temperature: number;
|
|
492
1132
|
topP: number;
|
|
1133
|
+
localRuntime?: {
|
|
1134
|
+
providerId: "claude-code";
|
|
1135
|
+
configurationHash: string;
|
|
1136
|
+
executableHash: string;
|
|
1137
|
+
version: string;
|
|
1138
|
+
capabilityHash: string;
|
|
1139
|
+
maximumRequestCostUsd: number;
|
|
1140
|
+
requestTimeoutMs: number;
|
|
1141
|
+
maximumTurns: number;
|
|
1142
|
+
holdOpen: boolean;
|
|
1143
|
+
} | {
|
|
1144
|
+
providerId: "custom-openai-compatible";
|
|
1145
|
+
configurationHash: string;
|
|
1146
|
+
maximumRequestCostUsd: number;
|
|
1147
|
+
requestTimeoutMs: number;
|
|
1148
|
+
} | undefined;
|
|
1149
|
+
candidate?: {
|
|
1150
|
+
artifactId: string;
|
|
1151
|
+
contentHash: string;
|
|
1152
|
+
preparationId: string;
|
|
1153
|
+
dispatchId: string;
|
|
1154
|
+
baseProfileId: string;
|
|
1155
|
+
workerImageDigest: string;
|
|
1156
|
+
} | undefined;
|
|
493
1157
|
harness?: {
|
|
494
1158
|
harnessRelease: {
|
|
495
1159
|
id: string;
|
|
@@ -532,10 +1196,210 @@ export declare class OpenPondExperimentsClient {
|
|
|
532
1196
|
actionId: string;
|
|
533
1197
|
};
|
|
534
1198
|
environmentHash: string;
|
|
1199
|
+
externalDatasetBinding?: {
|
|
1200
|
+
schemaVersion: "openpond.profileExternalDatasetBinding.v1";
|
|
1201
|
+
dataset: {
|
|
1202
|
+
id: string;
|
|
1203
|
+
contentHash: string;
|
|
1204
|
+
revision: number;
|
|
1205
|
+
};
|
|
1206
|
+
packageHash: string;
|
|
1207
|
+
profileId: string;
|
|
1208
|
+
target: {
|
|
1209
|
+
kind: "profile";
|
|
1210
|
+
} | {
|
|
1211
|
+
kind: "workflow";
|
|
1212
|
+
workflowId: string;
|
|
1213
|
+
} | {
|
|
1214
|
+
kind: "skill";
|
|
1215
|
+
skillPath: string;
|
|
1216
|
+
} | {
|
|
1217
|
+
kind: "agent_action";
|
|
1218
|
+
actionId: string;
|
|
1219
|
+
};
|
|
1220
|
+
declaredProfileCatalogHash: string | null;
|
|
1221
|
+
protectedProfileClosureHash: string;
|
|
1222
|
+
split: "train" | "validation" | "test" | "frozen_eval";
|
|
1223
|
+
population: {
|
|
1224
|
+
taskId: string;
|
|
1225
|
+
seed: string;
|
|
1226
|
+
fixtureId: string | null;
|
|
1227
|
+
}[];
|
|
1228
|
+
evaluators: {
|
|
1229
|
+
release: {
|
|
1230
|
+
id: string;
|
|
1231
|
+
contentHash: string;
|
|
1232
|
+
revision: string | number;
|
|
1233
|
+
};
|
|
1234
|
+
feedbackKey: string;
|
|
1235
|
+
configurationHash: string;
|
|
1236
|
+
output: "boolean" | "score" | "category";
|
|
1237
|
+
categories: string[];
|
|
1238
|
+
}[];
|
|
1239
|
+
fieldMappingsHash: string;
|
|
1240
|
+
fieldMappings: {
|
|
1241
|
+
graderId: string;
|
|
1242
|
+
fields: {
|
|
1243
|
+
destination: "input" | "output" | "expectedOutput" | "evaluatorContext";
|
|
1244
|
+
path: string[];
|
|
1245
|
+
source: "input" | "output" | "expectedOutput" | "evaluatorContext";
|
|
1246
|
+
sourcePath: string[];
|
|
1247
|
+
}[];
|
|
1248
|
+
}[];
|
|
1249
|
+
environmentHash: string;
|
|
1250
|
+
privateDatasetClosureHash: string;
|
|
1251
|
+
criterion: {
|
|
1252
|
+
minimumPassRate: number;
|
|
1253
|
+
requireComplete: boolean;
|
|
1254
|
+
};
|
|
1255
|
+
contentHash: string;
|
|
1256
|
+
recordedOrigin?: {
|
|
1257
|
+
execution: {
|
|
1258
|
+
id: string;
|
|
1259
|
+
contentHash: string;
|
|
1260
|
+
};
|
|
1261
|
+
originalEvidence: {
|
|
1262
|
+
id: string;
|
|
1263
|
+
contentHash: string;
|
|
1264
|
+
};
|
|
1265
|
+
members: {
|
|
1266
|
+
caseId: string;
|
|
1267
|
+
seed: string;
|
|
1268
|
+
fixtureId: string;
|
|
1269
|
+
sourceId: string;
|
|
1270
|
+
snapshotHash: string;
|
|
1271
|
+
boundaryRevisionHash: string;
|
|
1272
|
+
}[];
|
|
1273
|
+
} | undefined;
|
|
1274
|
+
gradingSource?: {
|
|
1275
|
+
pass: {
|
|
1276
|
+
id: string;
|
|
1277
|
+
contentHash: string;
|
|
1278
|
+
};
|
|
1279
|
+
evidence: {
|
|
1280
|
+
id: string;
|
|
1281
|
+
contentHash: string;
|
|
1282
|
+
};
|
|
1283
|
+
} | undefined;
|
|
1284
|
+
} | undefined;
|
|
1285
|
+
sourceCandidate?: {
|
|
1286
|
+
candidateId: string;
|
|
1287
|
+
frozenRevision: number;
|
|
1288
|
+
frozenHash: string;
|
|
1289
|
+
harnessRelease: {
|
|
1290
|
+
id: string;
|
|
1291
|
+
contentHash: string;
|
|
1292
|
+
};
|
|
1293
|
+
protectedClosureHash: string;
|
|
1294
|
+
} | undefined;
|
|
535
1295
|
};
|
|
536
1296
|
modelConfigurationHash: string;
|
|
537
1297
|
packageHash: string;
|
|
1298
|
+
candidate?: {
|
|
1299
|
+
artifactId: string;
|
|
1300
|
+
contentHash: string;
|
|
1301
|
+
preparationId: string;
|
|
1302
|
+
dispatchId: string;
|
|
1303
|
+
baseProfileId: string;
|
|
1304
|
+
workerImageDigest: string;
|
|
1305
|
+
} | undefined;
|
|
1306
|
+
sourceCandidate?: {
|
|
1307
|
+
candidateId: string;
|
|
1308
|
+
frozenRevision: number;
|
|
1309
|
+
frozenHash: string;
|
|
1310
|
+
harnessRelease: {
|
|
1311
|
+
id: string;
|
|
1312
|
+
contentHash: string;
|
|
1313
|
+
};
|
|
1314
|
+
protectedClosureHash: string;
|
|
1315
|
+
} | undefined;
|
|
538
1316
|
reasoningEffort?: "high" | "low" | "off" | "medium" | "xhigh" | "max" | undefined;
|
|
1317
|
+
externalDatasetBinding?: {
|
|
1318
|
+
schemaVersion: "openpond.profileExternalDatasetBinding.v1";
|
|
1319
|
+
dataset: {
|
|
1320
|
+
id: string;
|
|
1321
|
+
contentHash: string;
|
|
1322
|
+
revision: number;
|
|
1323
|
+
};
|
|
1324
|
+
packageHash: string;
|
|
1325
|
+
profileId: string;
|
|
1326
|
+
target: {
|
|
1327
|
+
kind: "profile";
|
|
1328
|
+
} | {
|
|
1329
|
+
kind: "workflow";
|
|
1330
|
+
workflowId: string;
|
|
1331
|
+
} | {
|
|
1332
|
+
kind: "skill";
|
|
1333
|
+
skillPath: string;
|
|
1334
|
+
} | {
|
|
1335
|
+
kind: "agent_action";
|
|
1336
|
+
actionId: string;
|
|
1337
|
+
};
|
|
1338
|
+
declaredProfileCatalogHash: string | null;
|
|
1339
|
+
protectedProfileClosureHash: string;
|
|
1340
|
+
split: "train" | "validation" | "test" | "frozen_eval";
|
|
1341
|
+
population: {
|
|
1342
|
+
taskId: string;
|
|
1343
|
+
seed: string;
|
|
1344
|
+
fixtureId: string | null;
|
|
1345
|
+
}[];
|
|
1346
|
+
evaluators: {
|
|
1347
|
+
release: {
|
|
1348
|
+
id: string;
|
|
1349
|
+
contentHash: string;
|
|
1350
|
+
revision: string | number;
|
|
1351
|
+
};
|
|
1352
|
+
feedbackKey: string;
|
|
1353
|
+
configurationHash: string;
|
|
1354
|
+
output: "boolean" | "score" | "category";
|
|
1355
|
+
categories: string[];
|
|
1356
|
+
}[];
|
|
1357
|
+
fieldMappingsHash: string;
|
|
1358
|
+
fieldMappings: {
|
|
1359
|
+
graderId: string;
|
|
1360
|
+
fields: {
|
|
1361
|
+
destination: "input" | "output" | "expectedOutput" | "evaluatorContext";
|
|
1362
|
+
path: string[];
|
|
1363
|
+
source: "input" | "output" | "expectedOutput" | "evaluatorContext";
|
|
1364
|
+
sourcePath: string[];
|
|
1365
|
+
}[];
|
|
1366
|
+
}[];
|
|
1367
|
+
environmentHash: string;
|
|
1368
|
+
privateDatasetClosureHash: string;
|
|
1369
|
+
criterion: {
|
|
1370
|
+
minimumPassRate: number;
|
|
1371
|
+
requireComplete: boolean;
|
|
1372
|
+
};
|
|
1373
|
+
contentHash: string;
|
|
1374
|
+
recordedOrigin?: {
|
|
1375
|
+
execution: {
|
|
1376
|
+
id: string;
|
|
1377
|
+
contentHash: string;
|
|
1378
|
+
};
|
|
1379
|
+
originalEvidence: {
|
|
1380
|
+
id: string;
|
|
1381
|
+
contentHash: string;
|
|
1382
|
+
};
|
|
1383
|
+
members: {
|
|
1384
|
+
caseId: string;
|
|
1385
|
+
seed: string;
|
|
1386
|
+
fixtureId: string;
|
|
1387
|
+
sourceId: string;
|
|
1388
|
+
snapshotHash: string;
|
|
1389
|
+
boundaryRevisionHash: string;
|
|
1390
|
+
}[];
|
|
1391
|
+
} | undefined;
|
|
1392
|
+
gradingSource?: {
|
|
1393
|
+
pass: {
|
|
1394
|
+
id: string;
|
|
1395
|
+
contentHash: string;
|
|
1396
|
+
};
|
|
1397
|
+
evidence: {
|
|
1398
|
+
id: string;
|
|
1399
|
+
contentHash: string;
|
|
1400
|
+
};
|
|
1401
|
+
} | undefined;
|
|
1402
|
+
} | undefined;
|
|
539
1403
|
} | {
|
|
540
1404
|
kind: "fixture";
|
|
541
1405
|
};
|
|
@@ -594,7 +1458,7 @@ export declare class OpenPondExperimentsClient {
|
|
|
594
1458
|
};
|
|
595
1459
|
policyKind: "hosted_chat" | "fixture" | "hosted_harness";
|
|
596
1460
|
manifestHash: string;
|
|
597
|
-
status: "
|
|
1461
|
+
status: "cancelled" | "failed" | "queued" | "running" | "cancelling" | "completed";
|
|
598
1462
|
totalCount: number;
|
|
599
1463
|
counts: {
|
|
600
1464
|
pending: number;
|
|
@@ -638,35 +1502,199 @@ export declare class OpenPondExperimentsClient {
|
|
|
638
1502
|
maxOutputTokens: number;
|
|
639
1503
|
temperature: number;
|
|
640
1504
|
topP: number;
|
|
1505
|
+
localRuntime?: {
|
|
1506
|
+
providerId: "claude-code";
|
|
1507
|
+
configurationHash: string;
|
|
1508
|
+
executableHash: string;
|
|
1509
|
+
version: string;
|
|
1510
|
+
capabilityHash: string;
|
|
1511
|
+
maximumRequestCostUsd: number;
|
|
1512
|
+
requestTimeoutMs: number;
|
|
1513
|
+
maximumTurns: number;
|
|
1514
|
+
holdOpen: boolean;
|
|
1515
|
+
} | {
|
|
1516
|
+
providerId: "custom-openai-compatible";
|
|
1517
|
+
configurationHash: string;
|
|
1518
|
+
maximumRequestCostUsd: number;
|
|
1519
|
+
requestTimeoutMs: number;
|
|
1520
|
+
} | undefined;
|
|
1521
|
+
candidate?: {
|
|
1522
|
+
artifactId: string;
|
|
1523
|
+
contentHash: string;
|
|
1524
|
+
preparationId: string;
|
|
1525
|
+
dispatchId: string;
|
|
1526
|
+
baseProfileId: string;
|
|
1527
|
+
workerImageDigest: string;
|
|
1528
|
+
} | undefined;
|
|
641
1529
|
harness?: {
|
|
642
1530
|
harnessRelease: {
|
|
643
1531
|
id: string;
|
|
644
1532
|
contentHash: string;
|
|
645
1533
|
};
|
|
646
|
-
agentSnapshot: {
|
|
647
|
-
id: string;
|
|
648
|
-
contentHash: string;
|
|
649
|
-
};
|
|
650
|
-
sourcePackageHash: string;
|
|
1534
|
+
agentSnapshot: {
|
|
1535
|
+
id: string;
|
|
1536
|
+
contentHash: string;
|
|
1537
|
+
};
|
|
1538
|
+
sourcePackageHash: string;
|
|
1539
|
+
} | undefined;
|
|
1540
|
+
messages?: {
|
|
1541
|
+
role: "system" | "user" | "assistant";
|
|
1542
|
+
content: string;
|
|
1543
|
+
}[] | undefined;
|
|
1544
|
+
} | {
|
|
1545
|
+
kind: "hosted_harness";
|
|
1546
|
+
modelId: string;
|
|
1547
|
+
profileRepositoryId: string;
|
|
1548
|
+
source: {
|
|
1549
|
+
profileId: string;
|
|
1550
|
+
sourceRevision: string;
|
|
1551
|
+
harnessRelease: {
|
|
1552
|
+
id: string;
|
|
1553
|
+
contentHash: string;
|
|
1554
|
+
};
|
|
1555
|
+
catalogHash: string;
|
|
1556
|
+
definitionId: string;
|
|
1557
|
+
definitionHash: string;
|
|
1558
|
+
target: {
|
|
1559
|
+
kind: "profile";
|
|
1560
|
+
} | {
|
|
1561
|
+
kind: "workflow";
|
|
1562
|
+
workflowId: string;
|
|
1563
|
+
} | {
|
|
1564
|
+
kind: "skill";
|
|
1565
|
+
skillPath: string;
|
|
1566
|
+
} | {
|
|
1567
|
+
kind: "agent_action";
|
|
1568
|
+
actionId: string;
|
|
1569
|
+
};
|
|
1570
|
+
environmentHash: string;
|
|
1571
|
+
externalDatasetBinding?: {
|
|
1572
|
+
schemaVersion: "openpond.profileExternalDatasetBinding.v1";
|
|
1573
|
+
dataset: {
|
|
1574
|
+
id: string;
|
|
1575
|
+
contentHash: string;
|
|
1576
|
+
revision: number;
|
|
1577
|
+
};
|
|
1578
|
+
packageHash: string;
|
|
1579
|
+
profileId: string;
|
|
1580
|
+
target: {
|
|
1581
|
+
kind: "profile";
|
|
1582
|
+
} | {
|
|
1583
|
+
kind: "workflow";
|
|
1584
|
+
workflowId: string;
|
|
1585
|
+
} | {
|
|
1586
|
+
kind: "skill";
|
|
1587
|
+
skillPath: string;
|
|
1588
|
+
} | {
|
|
1589
|
+
kind: "agent_action";
|
|
1590
|
+
actionId: string;
|
|
1591
|
+
};
|
|
1592
|
+
declaredProfileCatalogHash: string | null;
|
|
1593
|
+
protectedProfileClosureHash: string;
|
|
1594
|
+
split: "train" | "validation" | "test" | "frozen_eval";
|
|
1595
|
+
population: {
|
|
1596
|
+
taskId: string;
|
|
1597
|
+
seed: string;
|
|
1598
|
+
fixtureId: string | null;
|
|
1599
|
+
}[];
|
|
1600
|
+
evaluators: {
|
|
1601
|
+
release: {
|
|
1602
|
+
id: string;
|
|
1603
|
+
contentHash: string;
|
|
1604
|
+
revision: string | number;
|
|
1605
|
+
};
|
|
1606
|
+
feedbackKey: string;
|
|
1607
|
+
configurationHash: string;
|
|
1608
|
+
output: "boolean" | "score" | "category";
|
|
1609
|
+
categories: string[];
|
|
1610
|
+
}[];
|
|
1611
|
+
fieldMappingsHash: string;
|
|
1612
|
+
fieldMappings: {
|
|
1613
|
+
graderId: string;
|
|
1614
|
+
fields: {
|
|
1615
|
+
destination: "input" | "output" | "expectedOutput" | "evaluatorContext";
|
|
1616
|
+
path: string[];
|
|
1617
|
+
source: "input" | "output" | "expectedOutput" | "evaluatorContext";
|
|
1618
|
+
sourcePath: string[];
|
|
1619
|
+
}[];
|
|
1620
|
+
}[];
|
|
1621
|
+
environmentHash: string;
|
|
1622
|
+
privateDatasetClosureHash: string;
|
|
1623
|
+
criterion: {
|
|
1624
|
+
minimumPassRate: number;
|
|
1625
|
+
requireComplete: boolean;
|
|
1626
|
+
};
|
|
1627
|
+
contentHash: string;
|
|
1628
|
+
recordedOrigin?: {
|
|
1629
|
+
execution: {
|
|
1630
|
+
id: string;
|
|
1631
|
+
contentHash: string;
|
|
1632
|
+
};
|
|
1633
|
+
originalEvidence: {
|
|
1634
|
+
id: string;
|
|
1635
|
+
contentHash: string;
|
|
1636
|
+
};
|
|
1637
|
+
members: {
|
|
1638
|
+
caseId: string;
|
|
1639
|
+
seed: string;
|
|
1640
|
+
fixtureId: string;
|
|
1641
|
+
sourceId: string;
|
|
1642
|
+
snapshotHash: string;
|
|
1643
|
+
boundaryRevisionHash: string;
|
|
1644
|
+
}[];
|
|
1645
|
+
} | undefined;
|
|
1646
|
+
gradingSource?: {
|
|
1647
|
+
pass: {
|
|
1648
|
+
id: string;
|
|
1649
|
+
contentHash: string;
|
|
1650
|
+
};
|
|
1651
|
+
evidence: {
|
|
1652
|
+
id: string;
|
|
1653
|
+
contentHash: string;
|
|
1654
|
+
};
|
|
1655
|
+
} | undefined;
|
|
1656
|
+
} | undefined;
|
|
1657
|
+
sourceCandidate?: {
|
|
1658
|
+
candidateId: string;
|
|
1659
|
+
frozenRevision: number;
|
|
1660
|
+
frozenHash: string;
|
|
1661
|
+
harnessRelease: {
|
|
1662
|
+
id: string;
|
|
1663
|
+
contentHash: string;
|
|
1664
|
+
};
|
|
1665
|
+
protectedClosureHash: string;
|
|
1666
|
+
} | undefined;
|
|
1667
|
+
};
|
|
1668
|
+
modelConfigurationHash: string;
|
|
1669
|
+
packageHash: string;
|
|
1670
|
+
candidate?: {
|
|
1671
|
+
artifactId: string;
|
|
1672
|
+
contentHash: string;
|
|
1673
|
+
preparationId: string;
|
|
1674
|
+
dispatchId: string;
|
|
1675
|
+
baseProfileId: string;
|
|
1676
|
+
workerImageDigest: string;
|
|
1677
|
+
} | undefined;
|
|
1678
|
+
sourceCandidate?: {
|
|
1679
|
+
candidateId: string;
|
|
1680
|
+
frozenRevision: number;
|
|
1681
|
+
frozenHash: string;
|
|
1682
|
+
harnessRelease: {
|
|
1683
|
+
id: string;
|
|
1684
|
+
contentHash: string;
|
|
1685
|
+
};
|
|
1686
|
+
protectedClosureHash: string;
|
|
651
1687
|
} | undefined;
|
|
652
|
-
|
|
653
|
-
|
|
654
|
-
|
|
655
|
-
|
|
656
|
-
} | {
|
|
657
|
-
kind: "hosted_harness";
|
|
658
|
-
modelId: string;
|
|
659
|
-
profileRepositoryId: string;
|
|
660
|
-
source: {
|
|
661
|
-
profileId: string;
|
|
662
|
-
sourceRevision: string;
|
|
663
|
-
harnessRelease: {
|
|
1688
|
+
reasoningEffort?: "high" | "low" | "off" | "medium" | "xhigh" | "max" | undefined;
|
|
1689
|
+
externalDatasetBinding?: {
|
|
1690
|
+
schemaVersion: "openpond.profileExternalDatasetBinding.v1";
|
|
1691
|
+
dataset: {
|
|
664
1692
|
id: string;
|
|
665
1693
|
contentHash: string;
|
|
1694
|
+
revision: number;
|
|
666
1695
|
};
|
|
667
|
-
|
|
668
|
-
|
|
669
|
-
definitionHash: string;
|
|
1696
|
+
packageHash: string;
|
|
1697
|
+
profileId: string;
|
|
670
1698
|
target: {
|
|
671
1699
|
kind: "profile";
|
|
672
1700
|
} | {
|
|
@@ -679,11 +1707,71 @@ export declare class OpenPondExperimentsClient {
|
|
|
679
1707
|
kind: "agent_action";
|
|
680
1708
|
actionId: string;
|
|
681
1709
|
};
|
|
1710
|
+
declaredProfileCatalogHash: string | null;
|
|
1711
|
+
protectedProfileClosureHash: string;
|
|
1712
|
+
split: "train" | "validation" | "test" | "frozen_eval";
|
|
1713
|
+
population: {
|
|
1714
|
+
taskId: string;
|
|
1715
|
+
seed: string;
|
|
1716
|
+
fixtureId: string | null;
|
|
1717
|
+
}[];
|
|
1718
|
+
evaluators: {
|
|
1719
|
+
release: {
|
|
1720
|
+
id: string;
|
|
1721
|
+
contentHash: string;
|
|
1722
|
+
revision: string | number;
|
|
1723
|
+
};
|
|
1724
|
+
feedbackKey: string;
|
|
1725
|
+
configurationHash: string;
|
|
1726
|
+
output: "boolean" | "score" | "category";
|
|
1727
|
+
categories: string[];
|
|
1728
|
+
}[];
|
|
1729
|
+
fieldMappingsHash: string;
|
|
1730
|
+
fieldMappings: {
|
|
1731
|
+
graderId: string;
|
|
1732
|
+
fields: {
|
|
1733
|
+
destination: "input" | "output" | "expectedOutput" | "evaluatorContext";
|
|
1734
|
+
path: string[];
|
|
1735
|
+
source: "input" | "output" | "expectedOutput" | "evaluatorContext";
|
|
1736
|
+
sourcePath: string[];
|
|
1737
|
+
}[];
|
|
1738
|
+
}[];
|
|
682
1739
|
environmentHash: string;
|
|
683
|
-
|
|
684
|
-
|
|
685
|
-
|
|
686
|
-
|
|
1740
|
+
privateDatasetClosureHash: string;
|
|
1741
|
+
criterion: {
|
|
1742
|
+
minimumPassRate: number;
|
|
1743
|
+
requireComplete: boolean;
|
|
1744
|
+
};
|
|
1745
|
+
contentHash: string;
|
|
1746
|
+
recordedOrigin?: {
|
|
1747
|
+
execution: {
|
|
1748
|
+
id: string;
|
|
1749
|
+
contentHash: string;
|
|
1750
|
+
};
|
|
1751
|
+
originalEvidence: {
|
|
1752
|
+
id: string;
|
|
1753
|
+
contentHash: string;
|
|
1754
|
+
};
|
|
1755
|
+
members: {
|
|
1756
|
+
caseId: string;
|
|
1757
|
+
seed: string;
|
|
1758
|
+
fixtureId: string;
|
|
1759
|
+
sourceId: string;
|
|
1760
|
+
snapshotHash: string;
|
|
1761
|
+
boundaryRevisionHash: string;
|
|
1762
|
+
}[];
|
|
1763
|
+
} | undefined;
|
|
1764
|
+
gradingSource?: {
|
|
1765
|
+
pass: {
|
|
1766
|
+
id: string;
|
|
1767
|
+
contentHash: string;
|
|
1768
|
+
};
|
|
1769
|
+
evidence: {
|
|
1770
|
+
id: string;
|
|
1771
|
+
contentHash: string;
|
|
1772
|
+
};
|
|
1773
|
+
} | undefined;
|
|
1774
|
+
} | undefined;
|
|
687
1775
|
} | {
|
|
688
1776
|
kind: "fixture";
|
|
689
1777
|
};
|
|
@@ -800,6 +1888,102 @@ export declare class OpenPondExperimentsClient {
|
|
|
800
1888
|
actionId: string;
|
|
801
1889
|
};
|
|
802
1890
|
environmentHash: string;
|
|
1891
|
+
externalDatasetBinding?: {
|
|
1892
|
+
schemaVersion: "openpond.profileExternalDatasetBinding.v1";
|
|
1893
|
+
dataset: {
|
|
1894
|
+
id: string;
|
|
1895
|
+
contentHash: string;
|
|
1896
|
+
revision: number;
|
|
1897
|
+
};
|
|
1898
|
+
packageHash: string;
|
|
1899
|
+
profileId: string;
|
|
1900
|
+
target: {
|
|
1901
|
+
kind: "profile";
|
|
1902
|
+
} | {
|
|
1903
|
+
kind: "workflow";
|
|
1904
|
+
workflowId: string;
|
|
1905
|
+
} | {
|
|
1906
|
+
kind: "skill";
|
|
1907
|
+
skillPath: string;
|
|
1908
|
+
} | {
|
|
1909
|
+
kind: "agent_action";
|
|
1910
|
+
actionId: string;
|
|
1911
|
+
};
|
|
1912
|
+
declaredProfileCatalogHash: string | null;
|
|
1913
|
+
protectedProfileClosureHash: string;
|
|
1914
|
+
split: "train" | "validation" | "test" | "frozen_eval";
|
|
1915
|
+
population: {
|
|
1916
|
+
taskId: string;
|
|
1917
|
+
seed: string;
|
|
1918
|
+
fixtureId: string | null;
|
|
1919
|
+
}[];
|
|
1920
|
+
evaluators: {
|
|
1921
|
+
release: {
|
|
1922
|
+
id: string;
|
|
1923
|
+
contentHash: string;
|
|
1924
|
+
revision: string | number;
|
|
1925
|
+
};
|
|
1926
|
+
feedbackKey: string;
|
|
1927
|
+
configurationHash: string;
|
|
1928
|
+
output: "boolean" | "score" | "category";
|
|
1929
|
+
categories: string[];
|
|
1930
|
+
}[];
|
|
1931
|
+
fieldMappingsHash: string;
|
|
1932
|
+
fieldMappings: {
|
|
1933
|
+
graderId: string;
|
|
1934
|
+
fields: {
|
|
1935
|
+
destination: "input" | "output" | "expectedOutput" | "evaluatorContext";
|
|
1936
|
+
path: string[];
|
|
1937
|
+
source: "input" | "output" | "expectedOutput" | "evaluatorContext";
|
|
1938
|
+
sourcePath: string[];
|
|
1939
|
+
}[];
|
|
1940
|
+
}[];
|
|
1941
|
+
environmentHash: string;
|
|
1942
|
+
privateDatasetClosureHash: string;
|
|
1943
|
+
criterion: {
|
|
1944
|
+
minimumPassRate: number;
|
|
1945
|
+
requireComplete: boolean;
|
|
1946
|
+
};
|
|
1947
|
+
contentHash: string;
|
|
1948
|
+
recordedOrigin?: {
|
|
1949
|
+
execution: {
|
|
1950
|
+
id: string;
|
|
1951
|
+
contentHash: string;
|
|
1952
|
+
};
|
|
1953
|
+
originalEvidence: {
|
|
1954
|
+
id: string;
|
|
1955
|
+
contentHash: string;
|
|
1956
|
+
};
|
|
1957
|
+
members: {
|
|
1958
|
+
caseId: string;
|
|
1959
|
+
seed: string;
|
|
1960
|
+
fixtureId: string;
|
|
1961
|
+
sourceId: string;
|
|
1962
|
+
snapshotHash: string;
|
|
1963
|
+
boundaryRevisionHash: string;
|
|
1964
|
+
}[];
|
|
1965
|
+
} | undefined;
|
|
1966
|
+
gradingSource?: {
|
|
1967
|
+
pass: {
|
|
1968
|
+
id: string;
|
|
1969
|
+
contentHash: string;
|
|
1970
|
+
};
|
|
1971
|
+
evidence: {
|
|
1972
|
+
id: string;
|
|
1973
|
+
contentHash: string;
|
|
1974
|
+
};
|
|
1975
|
+
} | undefined;
|
|
1976
|
+
} | undefined;
|
|
1977
|
+
sourceCandidate?: {
|
|
1978
|
+
candidateId: string;
|
|
1979
|
+
frozenRevision: number;
|
|
1980
|
+
frozenHash: string;
|
|
1981
|
+
harnessRelease: {
|
|
1982
|
+
id: string;
|
|
1983
|
+
contentHash: string;
|
|
1984
|
+
};
|
|
1985
|
+
protectedClosureHash: string;
|
|
1986
|
+
} | undefined;
|
|
803
1987
|
} | undefined;
|
|
804
1988
|
};
|
|
805
1989
|
policySnapshot: {
|
|
@@ -825,6 +2009,30 @@ export declare class OpenPondExperimentsClient {
|
|
|
825
2009
|
maxOutputTokens: number;
|
|
826
2010
|
temperature: number;
|
|
827
2011
|
topP: number;
|
|
2012
|
+
localRuntime?: {
|
|
2013
|
+
providerId: "claude-code";
|
|
2014
|
+
configurationHash: string;
|
|
2015
|
+
executableHash: string;
|
|
2016
|
+
version: string;
|
|
2017
|
+
capabilityHash: string;
|
|
2018
|
+
maximumRequestCostUsd: number;
|
|
2019
|
+
requestTimeoutMs: number;
|
|
2020
|
+
maximumTurns: number;
|
|
2021
|
+
holdOpen: boolean;
|
|
2022
|
+
} | {
|
|
2023
|
+
providerId: "custom-openai-compatible";
|
|
2024
|
+
configurationHash: string;
|
|
2025
|
+
maximumRequestCostUsd: number;
|
|
2026
|
+
requestTimeoutMs: number;
|
|
2027
|
+
} | undefined;
|
|
2028
|
+
candidate?: {
|
|
2029
|
+
artifactId: string;
|
|
2030
|
+
contentHash: string;
|
|
2031
|
+
preparationId: string;
|
|
2032
|
+
dispatchId: string;
|
|
2033
|
+
baseProfileId: string;
|
|
2034
|
+
workerImageDigest: string;
|
|
2035
|
+
} | undefined;
|
|
828
2036
|
harness?: {
|
|
829
2037
|
harnessRelease: {
|
|
830
2038
|
id: string;
|
|
@@ -867,10 +2075,210 @@ export declare class OpenPondExperimentsClient {
|
|
|
867
2075
|
actionId: string;
|
|
868
2076
|
};
|
|
869
2077
|
environmentHash: string;
|
|
2078
|
+
externalDatasetBinding?: {
|
|
2079
|
+
schemaVersion: "openpond.profileExternalDatasetBinding.v1";
|
|
2080
|
+
dataset: {
|
|
2081
|
+
id: string;
|
|
2082
|
+
contentHash: string;
|
|
2083
|
+
revision: number;
|
|
2084
|
+
};
|
|
2085
|
+
packageHash: string;
|
|
2086
|
+
profileId: string;
|
|
2087
|
+
target: {
|
|
2088
|
+
kind: "profile";
|
|
2089
|
+
} | {
|
|
2090
|
+
kind: "workflow";
|
|
2091
|
+
workflowId: string;
|
|
2092
|
+
} | {
|
|
2093
|
+
kind: "skill";
|
|
2094
|
+
skillPath: string;
|
|
2095
|
+
} | {
|
|
2096
|
+
kind: "agent_action";
|
|
2097
|
+
actionId: string;
|
|
2098
|
+
};
|
|
2099
|
+
declaredProfileCatalogHash: string | null;
|
|
2100
|
+
protectedProfileClosureHash: string;
|
|
2101
|
+
split: "train" | "validation" | "test" | "frozen_eval";
|
|
2102
|
+
population: {
|
|
2103
|
+
taskId: string;
|
|
2104
|
+
seed: string;
|
|
2105
|
+
fixtureId: string | null;
|
|
2106
|
+
}[];
|
|
2107
|
+
evaluators: {
|
|
2108
|
+
release: {
|
|
2109
|
+
id: string;
|
|
2110
|
+
contentHash: string;
|
|
2111
|
+
revision: string | number;
|
|
2112
|
+
};
|
|
2113
|
+
feedbackKey: string;
|
|
2114
|
+
configurationHash: string;
|
|
2115
|
+
output: "boolean" | "score" | "category";
|
|
2116
|
+
categories: string[];
|
|
2117
|
+
}[];
|
|
2118
|
+
fieldMappingsHash: string;
|
|
2119
|
+
fieldMappings: {
|
|
2120
|
+
graderId: string;
|
|
2121
|
+
fields: {
|
|
2122
|
+
destination: "input" | "output" | "expectedOutput" | "evaluatorContext";
|
|
2123
|
+
path: string[];
|
|
2124
|
+
source: "input" | "output" | "expectedOutput" | "evaluatorContext";
|
|
2125
|
+
sourcePath: string[];
|
|
2126
|
+
}[];
|
|
2127
|
+
}[];
|
|
2128
|
+
environmentHash: string;
|
|
2129
|
+
privateDatasetClosureHash: string;
|
|
2130
|
+
criterion: {
|
|
2131
|
+
minimumPassRate: number;
|
|
2132
|
+
requireComplete: boolean;
|
|
2133
|
+
};
|
|
2134
|
+
contentHash: string;
|
|
2135
|
+
recordedOrigin?: {
|
|
2136
|
+
execution: {
|
|
2137
|
+
id: string;
|
|
2138
|
+
contentHash: string;
|
|
2139
|
+
};
|
|
2140
|
+
originalEvidence: {
|
|
2141
|
+
id: string;
|
|
2142
|
+
contentHash: string;
|
|
2143
|
+
};
|
|
2144
|
+
members: {
|
|
2145
|
+
caseId: string;
|
|
2146
|
+
seed: string;
|
|
2147
|
+
fixtureId: string;
|
|
2148
|
+
sourceId: string;
|
|
2149
|
+
snapshotHash: string;
|
|
2150
|
+
boundaryRevisionHash: string;
|
|
2151
|
+
}[];
|
|
2152
|
+
} | undefined;
|
|
2153
|
+
gradingSource?: {
|
|
2154
|
+
pass: {
|
|
2155
|
+
id: string;
|
|
2156
|
+
contentHash: string;
|
|
2157
|
+
};
|
|
2158
|
+
evidence: {
|
|
2159
|
+
id: string;
|
|
2160
|
+
contentHash: string;
|
|
2161
|
+
};
|
|
2162
|
+
} | undefined;
|
|
2163
|
+
} | undefined;
|
|
2164
|
+
sourceCandidate?: {
|
|
2165
|
+
candidateId: string;
|
|
2166
|
+
frozenRevision: number;
|
|
2167
|
+
frozenHash: string;
|
|
2168
|
+
harnessRelease: {
|
|
2169
|
+
id: string;
|
|
2170
|
+
contentHash: string;
|
|
2171
|
+
};
|
|
2172
|
+
protectedClosureHash: string;
|
|
2173
|
+
} | undefined;
|
|
870
2174
|
};
|
|
871
2175
|
modelConfigurationHash: string;
|
|
872
2176
|
packageHash: string;
|
|
2177
|
+
candidate?: {
|
|
2178
|
+
artifactId: string;
|
|
2179
|
+
contentHash: string;
|
|
2180
|
+
preparationId: string;
|
|
2181
|
+
dispatchId: string;
|
|
2182
|
+
baseProfileId: string;
|
|
2183
|
+
workerImageDigest: string;
|
|
2184
|
+
} | undefined;
|
|
2185
|
+
sourceCandidate?: {
|
|
2186
|
+
candidateId: string;
|
|
2187
|
+
frozenRevision: number;
|
|
2188
|
+
frozenHash: string;
|
|
2189
|
+
harnessRelease: {
|
|
2190
|
+
id: string;
|
|
2191
|
+
contentHash: string;
|
|
2192
|
+
};
|
|
2193
|
+
protectedClosureHash: string;
|
|
2194
|
+
} | undefined;
|
|
873
2195
|
reasoningEffort?: "high" | "low" | "off" | "medium" | "xhigh" | "max" | undefined;
|
|
2196
|
+
externalDatasetBinding?: {
|
|
2197
|
+
schemaVersion: "openpond.profileExternalDatasetBinding.v1";
|
|
2198
|
+
dataset: {
|
|
2199
|
+
id: string;
|
|
2200
|
+
contentHash: string;
|
|
2201
|
+
revision: number;
|
|
2202
|
+
};
|
|
2203
|
+
packageHash: string;
|
|
2204
|
+
profileId: string;
|
|
2205
|
+
target: {
|
|
2206
|
+
kind: "profile";
|
|
2207
|
+
} | {
|
|
2208
|
+
kind: "workflow";
|
|
2209
|
+
workflowId: string;
|
|
2210
|
+
} | {
|
|
2211
|
+
kind: "skill";
|
|
2212
|
+
skillPath: string;
|
|
2213
|
+
} | {
|
|
2214
|
+
kind: "agent_action";
|
|
2215
|
+
actionId: string;
|
|
2216
|
+
};
|
|
2217
|
+
declaredProfileCatalogHash: string | null;
|
|
2218
|
+
protectedProfileClosureHash: string;
|
|
2219
|
+
split: "train" | "validation" | "test" | "frozen_eval";
|
|
2220
|
+
population: {
|
|
2221
|
+
taskId: string;
|
|
2222
|
+
seed: string;
|
|
2223
|
+
fixtureId: string | null;
|
|
2224
|
+
}[];
|
|
2225
|
+
evaluators: {
|
|
2226
|
+
release: {
|
|
2227
|
+
id: string;
|
|
2228
|
+
contentHash: string;
|
|
2229
|
+
revision: string | number;
|
|
2230
|
+
};
|
|
2231
|
+
feedbackKey: string;
|
|
2232
|
+
configurationHash: string;
|
|
2233
|
+
output: "boolean" | "score" | "category";
|
|
2234
|
+
categories: string[];
|
|
2235
|
+
}[];
|
|
2236
|
+
fieldMappingsHash: string;
|
|
2237
|
+
fieldMappings: {
|
|
2238
|
+
graderId: string;
|
|
2239
|
+
fields: {
|
|
2240
|
+
destination: "input" | "output" | "expectedOutput" | "evaluatorContext";
|
|
2241
|
+
path: string[];
|
|
2242
|
+
source: "input" | "output" | "expectedOutput" | "evaluatorContext";
|
|
2243
|
+
sourcePath: string[];
|
|
2244
|
+
}[];
|
|
2245
|
+
}[];
|
|
2246
|
+
environmentHash: string;
|
|
2247
|
+
privateDatasetClosureHash: string;
|
|
2248
|
+
criterion: {
|
|
2249
|
+
minimumPassRate: number;
|
|
2250
|
+
requireComplete: boolean;
|
|
2251
|
+
};
|
|
2252
|
+
contentHash: string;
|
|
2253
|
+
recordedOrigin?: {
|
|
2254
|
+
execution: {
|
|
2255
|
+
id: string;
|
|
2256
|
+
contentHash: string;
|
|
2257
|
+
};
|
|
2258
|
+
originalEvidence: {
|
|
2259
|
+
id: string;
|
|
2260
|
+
contentHash: string;
|
|
2261
|
+
};
|
|
2262
|
+
members: {
|
|
2263
|
+
caseId: string;
|
|
2264
|
+
seed: string;
|
|
2265
|
+
fixtureId: string;
|
|
2266
|
+
sourceId: string;
|
|
2267
|
+
snapshotHash: string;
|
|
2268
|
+
boundaryRevisionHash: string;
|
|
2269
|
+
}[];
|
|
2270
|
+
} | undefined;
|
|
2271
|
+
gradingSource?: {
|
|
2272
|
+
pass: {
|
|
2273
|
+
id: string;
|
|
2274
|
+
contentHash: string;
|
|
2275
|
+
};
|
|
2276
|
+
evidence: {
|
|
2277
|
+
id: string;
|
|
2278
|
+
contentHash: string;
|
|
2279
|
+
};
|
|
2280
|
+
} | undefined;
|
|
2281
|
+
} | undefined;
|
|
874
2282
|
} | {
|
|
875
2283
|
kind: "fixture";
|
|
876
2284
|
};
|
|
@@ -928,7 +2336,7 @@ export declare class OpenPondExperimentsClient {
|
|
|
928
2336
|
};
|
|
929
2337
|
policyKind: "hosted_chat" | "fixture" | "hosted_harness";
|
|
930
2338
|
manifestHash: string;
|
|
931
|
-
status: "
|
|
2339
|
+
status: "cancelled" | "failed" | "queued" | "running" | "cancelling" | "completed";
|
|
932
2340
|
totalCount: number;
|
|
933
2341
|
counts: {
|
|
934
2342
|
pending: number;
|
|
@@ -972,6 +2380,30 @@ export declare class OpenPondExperimentsClient {
|
|
|
972
2380
|
maxOutputTokens: number;
|
|
973
2381
|
temperature: number;
|
|
974
2382
|
topP: number;
|
|
2383
|
+
localRuntime?: {
|
|
2384
|
+
providerId: "claude-code";
|
|
2385
|
+
configurationHash: string;
|
|
2386
|
+
executableHash: string;
|
|
2387
|
+
version: string;
|
|
2388
|
+
capabilityHash: string;
|
|
2389
|
+
maximumRequestCostUsd: number;
|
|
2390
|
+
requestTimeoutMs: number;
|
|
2391
|
+
maximumTurns: number;
|
|
2392
|
+
holdOpen: boolean;
|
|
2393
|
+
} | {
|
|
2394
|
+
providerId: "custom-openai-compatible";
|
|
2395
|
+
configurationHash: string;
|
|
2396
|
+
maximumRequestCostUsd: number;
|
|
2397
|
+
requestTimeoutMs: number;
|
|
2398
|
+
} | undefined;
|
|
2399
|
+
candidate?: {
|
|
2400
|
+
artifactId: string;
|
|
2401
|
+
contentHash: string;
|
|
2402
|
+
preparationId: string;
|
|
2403
|
+
dispatchId: string;
|
|
2404
|
+
baseProfileId: string;
|
|
2405
|
+
workerImageDigest: string;
|
|
2406
|
+
} | undefined;
|
|
975
2407
|
harness?: {
|
|
976
2408
|
harnessRelease: {
|
|
977
2409
|
id: string;
|
|
@@ -998,9 +2430,149 @@ export declare class OpenPondExperimentsClient {
|
|
|
998
2430
|
id: string;
|
|
999
2431
|
contentHash: string;
|
|
1000
2432
|
};
|
|
1001
|
-
catalogHash: string;
|
|
1002
|
-
definitionId: string;
|
|
1003
|
-
definitionHash: string;
|
|
2433
|
+
catalogHash: string;
|
|
2434
|
+
definitionId: string;
|
|
2435
|
+
definitionHash: string;
|
|
2436
|
+
target: {
|
|
2437
|
+
kind: "profile";
|
|
2438
|
+
} | {
|
|
2439
|
+
kind: "workflow";
|
|
2440
|
+
workflowId: string;
|
|
2441
|
+
} | {
|
|
2442
|
+
kind: "skill";
|
|
2443
|
+
skillPath: string;
|
|
2444
|
+
} | {
|
|
2445
|
+
kind: "agent_action";
|
|
2446
|
+
actionId: string;
|
|
2447
|
+
};
|
|
2448
|
+
environmentHash: string;
|
|
2449
|
+
externalDatasetBinding?: {
|
|
2450
|
+
schemaVersion: "openpond.profileExternalDatasetBinding.v1";
|
|
2451
|
+
dataset: {
|
|
2452
|
+
id: string;
|
|
2453
|
+
contentHash: string;
|
|
2454
|
+
revision: number;
|
|
2455
|
+
};
|
|
2456
|
+
packageHash: string;
|
|
2457
|
+
profileId: string;
|
|
2458
|
+
target: {
|
|
2459
|
+
kind: "profile";
|
|
2460
|
+
} | {
|
|
2461
|
+
kind: "workflow";
|
|
2462
|
+
workflowId: string;
|
|
2463
|
+
} | {
|
|
2464
|
+
kind: "skill";
|
|
2465
|
+
skillPath: string;
|
|
2466
|
+
} | {
|
|
2467
|
+
kind: "agent_action";
|
|
2468
|
+
actionId: string;
|
|
2469
|
+
};
|
|
2470
|
+
declaredProfileCatalogHash: string | null;
|
|
2471
|
+
protectedProfileClosureHash: string;
|
|
2472
|
+
split: "train" | "validation" | "test" | "frozen_eval";
|
|
2473
|
+
population: {
|
|
2474
|
+
taskId: string;
|
|
2475
|
+
seed: string;
|
|
2476
|
+
fixtureId: string | null;
|
|
2477
|
+
}[];
|
|
2478
|
+
evaluators: {
|
|
2479
|
+
release: {
|
|
2480
|
+
id: string;
|
|
2481
|
+
contentHash: string;
|
|
2482
|
+
revision: string | number;
|
|
2483
|
+
};
|
|
2484
|
+
feedbackKey: string;
|
|
2485
|
+
configurationHash: string;
|
|
2486
|
+
output: "boolean" | "score" | "category";
|
|
2487
|
+
categories: string[];
|
|
2488
|
+
}[];
|
|
2489
|
+
fieldMappingsHash: string;
|
|
2490
|
+
fieldMappings: {
|
|
2491
|
+
graderId: string;
|
|
2492
|
+
fields: {
|
|
2493
|
+
destination: "input" | "output" | "expectedOutput" | "evaluatorContext";
|
|
2494
|
+
path: string[];
|
|
2495
|
+
source: "input" | "output" | "expectedOutput" | "evaluatorContext";
|
|
2496
|
+
sourcePath: string[];
|
|
2497
|
+
}[];
|
|
2498
|
+
}[];
|
|
2499
|
+
environmentHash: string;
|
|
2500
|
+
privateDatasetClosureHash: string;
|
|
2501
|
+
criterion: {
|
|
2502
|
+
minimumPassRate: number;
|
|
2503
|
+
requireComplete: boolean;
|
|
2504
|
+
};
|
|
2505
|
+
contentHash: string;
|
|
2506
|
+
recordedOrigin?: {
|
|
2507
|
+
execution: {
|
|
2508
|
+
id: string;
|
|
2509
|
+
contentHash: string;
|
|
2510
|
+
};
|
|
2511
|
+
originalEvidence: {
|
|
2512
|
+
id: string;
|
|
2513
|
+
contentHash: string;
|
|
2514
|
+
};
|
|
2515
|
+
members: {
|
|
2516
|
+
caseId: string;
|
|
2517
|
+
seed: string;
|
|
2518
|
+
fixtureId: string;
|
|
2519
|
+
sourceId: string;
|
|
2520
|
+
snapshotHash: string;
|
|
2521
|
+
boundaryRevisionHash: string;
|
|
2522
|
+
}[];
|
|
2523
|
+
} | undefined;
|
|
2524
|
+
gradingSource?: {
|
|
2525
|
+
pass: {
|
|
2526
|
+
id: string;
|
|
2527
|
+
contentHash: string;
|
|
2528
|
+
};
|
|
2529
|
+
evidence: {
|
|
2530
|
+
id: string;
|
|
2531
|
+
contentHash: string;
|
|
2532
|
+
};
|
|
2533
|
+
} | undefined;
|
|
2534
|
+
} | undefined;
|
|
2535
|
+
sourceCandidate?: {
|
|
2536
|
+
candidateId: string;
|
|
2537
|
+
frozenRevision: number;
|
|
2538
|
+
frozenHash: string;
|
|
2539
|
+
harnessRelease: {
|
|
2540
|
+
id: string;
|
|
2541
|
+
contentHash: string;
|
|
2542
|
+
};
|
|
2543
|
+
protectedClosureHash: string;
|
|
2544
|
+
} | undefined;
|
|
2545
|
+
};
|
|
2546
|
+
modelConfigurationHash: string;
|
|
2547
|
+
packageHash: string;
|
|
2548
|
+
candidate?: {
|
|
2549
|
+
artifactId: string;
|
|
2550
|
+
contentHash: string;
|
|
2551
|
+
preparationId: string;
|
|
2552
|
+
dispatchId: string;
|
|
2553
|
+
baseProfileId: string;
|
|
2554
|
+
workerImageDigest: string;
|
|
2555
|
+
} | undefined;
|
|
2556
|
+
sourceCandidate?: {
|
|
2557
|
+
candidateId: string;
|
|
2558
|
+
frozenRevision: number;
|
|
2559
|
+
frozenHash: string;
|
|
2560
|
+
harnessRelease: {
|
|
2561
|
+
id: string;
|
|
2562
|
+
contentHash: string;
|
|
2563
|
+
};
|
|
2564
|
+
protectedClosureHash: string;
|
|
2565
|
+
} | undefined;
|
|
2566
|
+
reasoningEffort?: "high" | "low" | "off" | "medium" | "xhigh" | "max" | undefined;
|
|
2567
|
+
externalDatasetBinding?: {
|
|
2568
|
+
schemaVersion: "openpond.profileExternalDatasetBinding.v1";
|
|
2569
|
+
dataset: {
|
|
2570
|
+
id: string;
|
|
2571
|
+
contentHash: string;
|
|
2572
|
+
revision: number;
|
|
2573
|
+
};
|
|
2574
|
+
packageHash: string;
|
|
2575
|
+
profileId: string;
|
|
1004
2576
|
target: {
|
|
1005
2577
|
kind: "profile";
|
|
1006
2578
|
} | {
|
|
@@ -1013,11 +2585,71 @@ export declare class OpenPondExperimentsClient {
|
|
|
1013
2585
|
kind: "agent_action";
|
|
1014
2586
|
actionId: string;
|
|
1015
2587
|
};
|
|
2588
|
+
declaredProfileCatalogHash: string | null;
|
|
2589
|
+
protectedProfileClosureHash: string;
|
|
2590
|
+
split: "train" | "validation" | "test" | "frozen_eval";
|
|
2591
|
+
population: {
|
|
2592
|
+
taskId: string;
|
|
2593
|
+
seed: string;
|
|
2594
|
+
fixtureId: string | null;
|
|
2595
|
+
}[];
|
|
2596
|
+
evaluators: {
|
|
2597
|
+
release: {
|
|
2598
|
+
id: string;
|
|
2599
|
+
contentHash: string;
|
|
2600
|
+
revision: string | number;
|
|
2601
|
+
};
|
|
2602
|
+
feedbackKey: string;
|
|
2603
|
+
configurationHash: string;
|
|
2604
|
+
output: "boolean" | "score" | "category";
|
|
2605
|
+
categories: string[];
|
|
2606
|
+
}[];
|
|
2607
|
+
fieldMappingsHash: string;
|
|
2608
|
+
fieldMappings: {
|
|
2609
|
+
graderId: string;
|
|
2610
|
+
fields: {
|
|
2611
|
+
destination: "input" | "output" | "expectedOutput" | "evaluatorContext";
|
|
2612
|
+
path: string[];
|
|
2613
|
+
source: "input" | "output" | "expectedOutput" | "evaluatorContext";
|
|
2614
|
+
sourcePath: string[];
|
|
2615
|
+
}[];
|
|
2616
|
+
}[];
|
|
1016
2617
|
environmentHash: string;
|
|
1017
|
-
|
|
1018
|
-
|
|
1019
|
-
|
|
1020
|
-
|
|
2618
|
+
privateDatasetClosureHash: string;
|
|
2619
|
+
criterion: {
|
|
2620
|
+
minimumPassRate: number;
|
|
2621
|
+
requireComplete: boolean;
|
|
2622
|
+
};
|
|
2623
|
+
contentHash: string;
|
|
2624
|
+
recordedOrigin?: {
|
|
2625
|
+
execution: {
|
|
2626
|
+
id: string;
|
|
2627
|
+
contentHash: string;
|
|
2628
|
+
};
|
|
2629
|
+
originalEvidence: {
|
|
2630
|
+
id: string;
|
|
2631
|
+
contentHash: string;
|
|
2632
|
+
};
|
|
2633
|
+
members: {
|
|
2634
|
+
caseId: string;
|
|
2635
|
+
seed: string;
|
|
2636
|
+
fixtureId: string;
|
|
2637
|
+
sourceId: string;
|
|
2638
|
+
snapshotHash: string;
|
|
2639
|
+
boundaryRevisionHash: string;
|
|
2640
|
+
}[];
|
|
2641
|
+
} | undefined;
|
|
2642
|
+
gradingSource?: {
|
|
2643
|
+
pass: {
|
|
2644
|
+
id: string;
|
|
2645
|
+
contentHash: string;
|
|
2646
|
+
};
|
|
2647
|
+
evidence: {
|
|
2648
|
+
id: string;
|
|
2649
|
+
contentHash: string;
|
|
2650
|
+
};
|
|
2651
|
+
} | undefined;
|
|
2652
|
+
} | undefined;
|
|
1021
2653
|
} | {
|
|
1022
2654
|
kind: "fixture";
|
|
1023
2655
|
};
|
|
@@ -1134,6 +2766,102 @@ export declare class OpenPondExperimentsClient {
|
|
|
1134
2766
|
actionId: string;
|
|
1135
2767
|
};
|
|
1136
2768
|
environmentHash: string;
|
|
2769
|
+
externalDatasetBinding?: {
|
|
2770
|
+
schemaVersion: "openpond.profileExternalDatasetBinding.v1";
|
|
2771
|
+
dataset: {
|
|
2772
|
+
id: string;
|
|
2773
|
+
contentHash: string;
|
|
2774
|
+
revision: number;
|
|
2775
|
+
};
|
|
2776
|
+
packageHash: string;
|
|
2777
|
+
profileId: string;
|
|
2778
|
+
target: {
|
|
2779
|
+
kind: "profile";
|
|
2780
|
+
} | {
|
|
2781
|
+
kind: "workflow";
|
|
2782
|
+
workflowId: string;
|
|
2783
|
+
} | {
|
|
2784
|
+
kind: "skill";
|
|
2785
|
+
skillPath: string;
|
|
2786
|
+
} | {
|
|
2787
|
+
kind: "agent_action";
|
|
2788
|
+
actionId: string;
|
|
2789
|
+
};
|
|
2790
|
+
declaredProfileCatalogHash: string | null;
|
|
2791
|
+
protectedProfileClosureHash: string;
|
|
2792
|
+
split: "train" | "validation" | "test" | "frozen_eval";
|
|
2793
|
+
population: {
|
|
2794
|
+
taskId: string;
|
|
2795
|
+
seed: string;
|
|
2796
|
+
fixtureId: string | null;
|
|
2797
|
+
}[];
|
|
2798
|
+
evaluators: {
|
|
2799
|
+
release: {
|
|
2800
|
+
id: string;
|
|
2801
|
+
contentHash: string;
|
|
2802
|
+
revision: string | number;
|
|
2803
|
+
};
|
|
2804
|
+
feedbackKey: string;
|
|
2805
|
+
configurationHash: string;
|
|
2806
|
+
output: "boolean" | "score" | "category";
|
|
2807
|
+
categories: string[];
|
|
2808
|
+
}[];
|
|
2809
|
+
fieldMappingsHash: string;
|
|
2810
|
+
fieldMappings: {
|
|
2811
|
+
graderId: string;
|
|
2812
|
+
fields: {
|
|
2813
|
+
destination: "input" | "output" | "expectedOutput" | "evaluatorContext";
|
|
2814
|
+
path: string[];
|
|
2815
|
+
source: "input" | "output" | "expectedOutput" | "evaluatorContext";
|
|
2816
|
+
sourcePath: string[];
|
|
2817
|
+
}[];
|
|
2818
|
+
}[];
|
|
2819
|
+
environmentHash: string;
|
|
2820
|
+
privateDatasetClosureHash: string;
|
|
2821
|
+
criterion: {
|
|
2822
|
+
minimumPassRate: number;
|
|
2823
|
+
requireComplete: boolean;
|
|
2824
|
+
};
|
|
2825
|
+
contentHash: string;
|
|
2826
|
+
recordedOrigin?: {
|
|
2827
|
+
execution: {
|
|
2828
|
+
id: string;
|
|
2829
|
+
contentHash: string;
|
|
2830
|
+
};
|
|
2831
|
+
originalEvidence: {
|
|
2832
|
+
id: string;
|
|
2833
|
+
contentHash: string;
|
|
2834
|
+
};
|
|
2835
|
+
members: {
|
|
2836
|
+
caseId: string;
|
|
2837
|
+
seed: string;
|
|
2838
|
+
fixtureId: string;
|
|
2839
|
+
sourceId: string;
|
|
2840
|
+
snapshotHash: string;
|
|
2841
|
+
boundaryRevisionHash: string;
|
|
2842
|
+
}[];
|
|
2843
|
+
} | undefined;
|
|
2844
|
+
gradingSource?: {
|
|
2845
|
+
pass: {
|
|
2846
|
+
id: string;
|
|
2847
|
+
contentHash: string;
|
|
2848
|
+
};
|
|
2849
|
+
evidence: {
|
|
2850
|
+
id: string;
|
|
2851
|
+
contentHash: string;
|
|
2852
|
+
};
|
|
2853
|
+
} | undefined;
|
|
2854
|
+
} | undefined;
|
|
2855
|
+
sourceCandidate?: {
|
|
2856
|
+
candidateId: string;
|
|
2857
|
+
frozenRevision: number;
|
|
2858
|
+
frozenHash: string;
|
|
2859
|
+
harnessRelease: {
|
|
2860
|
+
id: string;
|
|
2861
|
+
contentHash: string;
|
|
2862
|
+
};
|
|
2863
|
+
protectedClosureHash: string;
|
|
2864
|
+
} | undefined;
|
|
1137
2865
|
} | undefined;
|
|
1138
2866
|
};
|
|
1139
2867
|
policySnapshot: {
|
|
@@ -1159,6 +2887,30 @@ export declare class OpenPondExperimentsClient {
|
|
|
1159
2887
|
maxOutputTokens: number;
|
|
1160
2888
|
temperature: number;
|
|
1161
2889
|
topP: number;
|
|
2890
|
+
localRuntime?: {
|
|
2891
|
+
providerId: "claude-code";
|
|
2892
|
+
configurationHash: string;
|
|
2893
|
+
executableHash: string;
|
|
2894
|
+
version: string;
|
|
2895
|
+
capabilityHash: string;
|
|
2896
|
+
maximumRequestCostUsd: number;
|
|
2897
|
+
requestTimeoutMs: number;
|
|
2898
|
+
maximumTurns: number;
|
|
2899
|
+
holdOpen: boolean;
|
|
2900
|
+
} | {
|
|
2901
|
+
providerId: "custom-openai-compatible";
|
|
2902
|
+
configurationHash: string;
|
|
2903
|
+
maximumRequestCostUsd: number;
|
|
2904
|
+
requestTimeoutMs: number;
|
|
2905
|
+
} | undefined;
|
|
2906
|
+
candidate?: {
|
|
2907
|
+
artifactId: string;
|
|
2908
|
+
contentHash: string;
|
|
2909
|
+
preparationId: string;
|
|
2910
|
+
dispatchId: string;
|
|
2911
|
+
baseProfileId: string;
|
|
2912
|
+
workerImageDigest: string;
|
|
2913
|
+
} | undefined;
|
|
1162
2914
|
harness?: {
|
|
1163
2915
|
harnessRelease: {
|
|
1164
2916
|
id: string;
|
|
@@ -1201,10 +2953,210 @@ export declare class OpenPondExperimentsClient {
|
|
|
1201
2953
|
actionId: string;
|
|
1202
2954
|
};
|
|
1203
2955
|
environmentHash: string;
|
|
2956
|
+
externalDatasetBinding?: {
|
|
2957
|
+
schemaVersion: "openpond.profileExternalDatasetBinding.v1";
|
|
2958
|
+
dataset: {
|
|
2959
|
+
id: string;
|
|
2960
|
+
contentHash: string;
|
|
2961
|
+
revision: number;
|
|
2962
|
+
};
|
|
2963
|
+
packageHash: string;
|
|
2964
|
+
profileId: string;
|
|
2965
|
+
target: {
|
|
2966
|
+
kind: "profile";
|
|
2967
|
+
} | {
|
|
2968
|
+
kind: "workflow";
|
|
2969
|
+
workflowId: string;
|
|
2970
|
+
} | {
|
|
2971
|
+
kind: "skill";
|
|
2972
|
+
skillPath: string;
|
|
2973
|
+
} | {
|
|
2974
|
+
kind: "agent_action";
|
|
2975
|
+
actionId: string;
|
|
2976
|
+
};
|
|
2977
|
+
declaredProfileCatalogHash: string | null;
|
|
2978
|
+
protectedProfileClosureHash: string;
|
|
2979
|
+
split: "train" | "validation" | "test" | "frozen_eval";
|
|
2980
|
+
population: {
|
|
2981
|
+
taskId: string;
|
|
2982
|
+
seed: string;
|
|
2983
|
+
fixtureId: string | null;
|
|
2984
|
+
}[];
|
|
2985
|
+
evaluators: {
|
|
2986
|
+
release: {
|
|
2987
|
+
id: string;
|
|
2988
|
+
contentHash: string;
|
|
2989
|
+
revision: string | number;
|
|
2990
|
+
};
|
|
2991
|
+
feedbackKey: string;
|
|
2992
|
+
configurationHash: string;
|
|
2993
|
+
output: "boolean" | "score" | "category";
|
|
2994
|
+
categories: string[];
|
|
2995
|
+
}[];
|
|
2996
|
+
fieldMappingsHash: string;
|
|
2997
|
+
fieldMappings: {
|
|
2998
|
+
graderId: string;
|
|
2999
|
+
fields: {
|
|
3000
|
+
destination: "input" | "output" | "expectedOutput" | "evaluatorContext";
|
|
3001
|
+
path: string[];
|
|
3002
|
+
source: "input" | "output" | "expectedOutput" | "evaluatorContext";
|
|
3003
|
+
sourcePath: string[];
|
|
3004
|
+
}[];
|
|
3005
|
+
}[];
|
|
3006
|
+
environmentHash: string;
|
|
3007
|
+
privateDatasetClosureHash: string;
|
|
3008
|
+
criterion: {
|
|
3009
|
+
minimumPassRate: number;
|
|
3010
|
+
requireComplete: boolean;
|
|
3011
|
+
};
|
|
3012
|
+
contentHash: string;
|
|
3013
|
+
recordedOrigin?: {
|
|
3014
|
+
execution: {
|
|
3015
|
+
id: string;
|
|
3016
|
+
contentHash: string;
|
|
3017
|
+
};
|
|
3018
|
+
originalEvidence: {
|
|
3019
|
+
id: string;
|
|
3020
|
+
contentHash: string;
|
|
3021
|
+
};
|
|
3022
|
+
members: {
|
|
3023
|
+
caseId: string;
|
|
3024
|
+
seed: string;
|
|
3025
|
+
fixtureId: string;
|
|
3026
|
+
sourceId: string;
|
|
3027
|
+
snapshotHash: string;
|
|
3028
|
+
boundaryRevisionHash: string;
|
|
3029
|
+
}[];
|
|
3030
|
+
} | undefined;
|
|
3031
|
+
gradingSource?: {
|
|
3032
|
+
pass: {
|
|
3033
|
+
id: string;
|
|
3034
|
+
contentHash: string;
|
|
3035
|
+
};
|
|
3036
|
+
evidence: {
|
|
3037
|
+
id: string;
|
|
3038
|
+
contentHash: string;
|
|
3039
|
+
};
|
|
3040
|
+
} | undefined;
|
|
3041
|
+
} | undefined;
|
|
3042
|
+
sourceCandidate?: {
|
|
3043
|
+
candidateId: string;
|
|
3044
|
+
frozenRevision: number;
|
|
3045
|
+
frozenHash: string;
|
|
3046
|
+
harnessRelease: {
|
|
3047
|
+
id: string;
|
|
3048
|
+
contentHash: string;
|
|
3049
|
+
};
|
|
3050
|
+
protectedClosureHash: string;
|
|
3051
|
+
} | undefined;
|
|
1204
3052
|
};
|
|
1205
3053
|
modelConfigurationHash: string;
|
|
1206
3054
|
packageHash: string;
|
|
3055
|
+
candidate?: {
|
|
3056
|
+
artifactId: string;
|
|
3057
|
+
contentHash: string;
|
|
3058
|
+
preparationId: string;
|
|
3059
|
+
dispatchId: string;
|
|
3060
|
+
baseProfileId: string;
|
|
3061
|
+
workerImageDigest: string;
|
|
3062
|
+
} | undefined;
|
|
3063
|
+
sourceCandidate?: {
|
|
3064
|
+
candidateId: string;
|
|
3065
|
+
frozenRevision: number;
|
|
3066
|
+
frozenHash: string;
|
|
3067
|
+
harnessRelease: {
|
|
3068
|
+
id: string;
|
|
3069
|
+
contentHash: string;
|
|
3070
|
+
};
|
|
3071
|
+
protectedClosureHash: string;
|
|
3072
|
+
} | undefined;
|
|
1207
3073
|
reasoningEffort?: "high" | "low" | "off" | "medium" | "xhigh" | "max" | undefined;
|
|
3074
|
+
externalDatasetBinding?: {
|
|
3075
|
+
schemaVersion: "openpond.profileExternalDatasetBinding.v1";
|
|
3076
|
+
dataset: {
|
|
3077
|
+
id: string;
|
|
3078
|
+
contentHash: string;
|
|
3079
|
+
revision: number;
|
|
3080
|
+
};
|
|
3081
|
+
packageHash: string;
|
|
3082
|
+
profileId: string;
|
|
3083
|
+
target: {
|
|
3084
|
+
kind: "profile";
|
|
3085
|
+
} | {
|
|
3086
|
+
kind: "workflow";
|
|
3087
|
+
workflowId: string;
|
|
3088
|
+
} | {
|
|
3089
|
+
kind: "skill";
|
|
3090
|
+
skillPath: string;
|
|
3091
|
+
} | {
|
|
3092
|
+
kind: "agent_action";
|
|
3093
|
+
actionId: string;
|
|
3094
|
+
};
|
|
3095
|
+
declaredProfileCatalogHash: string | null;
|
|
3096
|
+
protectedProfileClosureHash: string;
|
|
3097
|
+
split: "train" | "validation" | "test" | "frozen_eval";
|
|
3098
|
+
population: {
|
|
3099
|
+
taskId: string;
|
|
3100
|
+
seed: string;
|
|
3101
|
+
fixtureId: string | null;
|
|
3102
|
+
}[];
|
|
3103
|
+
evaluators: {
|
|
3104
|
+
release: {
|
|
3105
|
+
id: string;
|
|
3106
|
+
contentHash: string;
|
|
3107
|
+
revision: string | number;
|
|
3108
|
+
};
|
|
3109
|
+
feedbackKey: string;
|
|
3110
|
+
configurationHash: string;
|
|
3111
|
+
output: "boolean" | "score" | "category";
|
|
3112
|
+
categories: string[];
|
|
3113
|
+
}[];
|
|
3114
|
+
fieldMappingsHash: string;
|
|
3115
|
+
fieldMappings: {
|
|
3116
|
+
graderId: string;
|
|
3117
|
+
fields: {
|
|
3118
|
+
destination: "input" | "output" | "expectedOutput" | "evaluatorContext";
|
|
3119
|
+
path: string[];
|
|
3120
|
+
source: "input" | "output" | "expectedOutput" | "evaluatorContext";
|
|
3121
|
+
sourcePath: string[];
|
|
3122
|
+
}[];
|
|
3123
|
+
}[];
|
|
3124
|
+
environmentHash: string;
|
|
3125
|
+
privateDatasetClosureHash: string;
|
|
3126
|
+
criterion: {
|
|
3127
|
+
minimumPassRate: number;
|
|
3128
|
+
requireComplete: boolean;
|
|
3129
|
+
};
|
|
3130
|
+
contentHash: string;
|
|
3131
|
+
recordedOrigin?: {
|
|
3132
|
+
execution: {
|
|
3133
|
+
id: string;
|
|
3134
|
+
contentHash: string;
|
|
3135
|
+
};
|
|
3136
|
+
originalEvidence: {
|
|
3137
|
+
id: string;
|
|
3138
|
+
contentHash: string;
|
|
3139
|
+
};
|
|
3140
|
+
members: {
|
|
3141
|
+
caseId: string;
|
|
3142
|
+
seed: string;
|
|
3143
|
+
fixtureId: string;
|
|
3144
|
+
sourceId: string;
|
|
3145
|
+
snapshotHash: string;
|
|
3146
|
+
boundaryRevisionHash: string;
|
|
3147
|
+
}[];
|
|
3148
|
+
} | undefined;
|
|
3149
|
+
gradingSource?: {
|
|
3150
|
+
pass: {
|
|
3151
|
+
id: string;
|
|
3152
|
+
contentHash: string;
|
|
3153
|
+
};
|
|
3154
|
+
evidence: {
|
|
3155
|
+
id: string;
|
|
3156
|
+
contentHash: string;
|
|
3157
|
+
};
|
|
3158
|
+
} | undefined;
|
|
3159
|
+
} | undefined;
|
|
1208
3160
|
} | {
|
|
1209
3161
|
kind: "fixture";
|
|
1210
3162
|
};
|
|
@@ -1292,6 +3244,102 @@ export declare class OpenPondExperimentsClient {
|
|
|
1292
3244
|
actionId: string;
|
|
1293
3245
|
};
|
|
1294
3246
|
environmentHash: string;
|
|
3247
|
+
externalDatasetBinding?: {
|
|
3248
|
+
schemaVersion: "openpond.profileExternalDatasetBinding.v1";
|
|
3249
|
+
dataset: {
|
|
3250
|
+
id: string;
|
|
3251
|
+
contentHash: string;
|
|
3252
|
+
revision: number;
|
|
3253
|
+
};
|
|
3254
|
+
packageHash: string;
|
|
3255
|
+
profileId: string;
|
|
3256
|
+
target: {
|
|
3257
|
+
kind: "profile";
|
|
3258
|
+
} | {
|
|
3259
|
+
kind: "workflow";
|
|
3260
|
+
workflowId: string;
|
|
3261
|
+
} | {
|
|
3262
|
+
kind: "skill";
|
|
3263
|
+
skillPath: string;
|
|
3264
|
+
} | {
|
|
3265
|
+
kind: "agent_action";
|
|
3266
|
+
actionId: string;
|
|
3267
|
+
};
|
|
3268
|
+
declaredProfileCatalogHash: string | null;
|
|
3269
|
+
protectedProfileClosureHash: string;
|
|
3270
|
+
split: "train" | "validation" | "test" | "frozen_eval";
|
|
3271
|
+
population: {
|
|
3272
|
+
taskId: string;
|
|
3273
|
+
seed: string;
|
|
3274
|
+
fixtureId: string | null;
|
|
3275
|
+
}[];
|
|
3276
|
+
evaluators: {
|
|
3277
|
+
release: {
|
|
3278
|
+
id: string;
|
|
3279
|
+
contentHash: string;
|
|
3280
|
+
revision: string | number;
|
|
3281
|
+
};
|
|
3282
|
+
feedbackKey: string;
|
|
3283
|
+
configurationHash: string;
|
|
3284
|
+
output: "boolean" | "score" | "category";
|
|
3285
|
+
categories: string[];
|
|
3286
|
+
}[];
|
|
3287
|
+
fieldMappingsHash: string;
|
|
3288
|
+
fieldMappings: {
|
|
3289
|
+
graderId: string;
|
|
3290
|
+
fields: {
|
|
3291
|
+
destination: "output" | "input" | "expectedOutput" | "evaluatorContext";
|
|
3292
|
+
path: string[];
|
|
3293
|
+
source: "output" | "input" | "expectedOutput" | "evaluatorContext";
|
|
3294
|
+
sourcePath: string[];
|
|
3295
|
+
}[];
|
|
3296
|
+
}[];
|
|
3297
|
+
environmentHash: string;
|
|
3298
|
+
privateDatasetClosureHash: string;
|
|
3299
|
+
criterion: {
|
|
3300
|
+
minimumPassRate: number;
|
|
3301
|
+
requireComplete: boolean;
|
|
3302
|
+
};
|
|
3303
|
+
contentHash: string;
|
|
3304
|
+
recordedOrigin?: {
|
|
3305
|
+
execution: {
|
|
3306
|
+
id: string;
|
|
3307
|
+
contentHash: string;
|
|
3308
|
+
};
|
|
3309
|
+
originalEvidence: {
|
|
3310
|
+
id: string;
|
|
3311
|
+
contentHash: string;
|
|
3312
|
+
};
|
|
3313
|
+
members: {
|
|
3314
|
+
caseId: string;
|
|
3315
|
+
seed: string;
|
|
3316
|
+
fixtureId: string;
|
|
3317
|
+
sourceId: string;
|
|
3318
|
+
snapshotHash: string;
|
|
3319
|
+
boundaryRevisionHash: string;
|
|
3320
|
+
}[];
|
|
3321
|
+
} | undefined;
|
|
3322
|
+
gradingSource?: {
|
|
3323
|
+
pass: {
|
|
3324
|
+
id: string;
|
|
3325
|
+
contentHash: string;
|
|
3326
|
+
};
|
|
3327
|
+
evidence: {
|
|
3328
|
+
id: string;
|
|
3329
|
+
contentHash: string;
|
|
3330
|
+
};
|
|
3331
|
+
} | undefined;
|
|
3332
|
+
} | undefined;
|
|
3333
|
+
sourceCandidate?: {
|
|
3334
|
+
candidateId: string;
|
|
3335
|
+
frozenRevision: number;
|
|
3336
|
+
frozenHash: string;
|
|
3337
|
+
harnessRelease: {
|
|
3338
|
+
id: string;
|
|
3339
|
+
contentHash: string;
|
|
3340
|
+
};
|
|
3341
|
+
protectedClosureHash: string;
|
|
3342
|
+
} | undefined;
|
|
1295
3343
|
};
|
|
1296
3344
|
model: {
|
|
1297
3345
|
modelId: string;
|
|
@@ -1356,6 +3404,14 @@ export declare class OpenPondExperimentsClient {
|
|
|
1356
3404
|
contentHash: string;
|
|
1357
3405
|
};
|
|
1358
3406
|
scoringPassId: string | null;
|
|
3407
|
+
humanProjection?: {
|
|
3408
|
+
originalManifest: {
|
|
3409
|
+
id: string;
|
|
3410
|
+
contentHash: string;
|
|
3411
|
+
};
|
|
3412
|
+
selectionHash: string;
|
|
3413
|
+
rulesHash: string;
|
|
3414
|
+
} | undefined;
|
|
1359
3415
|
} | undefined;
|
|
1360
3416
|
};
|
|
1361
3417
|
result: {
|
|
@@ -1364,14 +3420,14 @@ export declare class OpenPondExperimentsClient {
|
|
|
1364
3420
|
id: string;
|
|
1365
3421
|
contentHash: string;
|
|
1366
3422
|
};
|
|
1367
|
-
status: "
|
|
3423
|
+
status: "cancelled" | "failed" | "completed";
|
|
1368
3424
|
cases: {
|
|
1369
3425
|
identity: {
|
|
1370
3426
|
caseId: string;
|
|
1371
3427
|
seed: string;
|
|
1372
3428
|
fixtureId: string | null;
|
|
1373
3429
|
};
|
|
1374
|
-
status: "
|
|
3430
|
+
status: "cancelled" | "failed" | "completed" | "unavailable";
|
|
1375
3431
|
output: unknown;
|
|
1376
3432
|
error: {
|
|
1377
3433
|
code: string;
|
|
@@ -1430,7 +3486,7 @@ export declare class OpenPondExperimentsClient {
|
|
|
1430
3486
|
};
|
|
1431
3487
|
policyKind: "hosted_chat" | "fixture" | "hosted_harness";
|
|
1432
3488
|
manifestHash: string;
|
|
1433
|
-
status: "
|
|
3489
|
+
status: "cancelled" | "failed" | "queued" | "running" | "cancelling" | "completed";
|
|
1434
3490
|
totalCount: number;
|
|
1435
3491
|
counts: {
|
|
1436
3492
|
pending: number;
|
|
@@ -1474,6 +3530,30 @@ export declare class OpenPondExperimentsClient {
|
|
|
1474
3530
|
maxOutputTokens: number;
|
|
1475
3531
|
temperature: number;
|
|
1476
3532
|
topP: number;
|
|
3533
|
+
localRuntime?: {
|
|
3534
|
+
providerId: "claude-code";
|
|
3535
|
+
configurationHash: string;
|
|
3536
|
+
executableHash: string;
|
|
3537
|
+
version: string;
|
|
3538
|
+
capabilityHash: string;
|
|
3539
|
+
maximumRequestCostUsd: number;
|
|
3540
|
+
requestTimeoutMs: number;
|
|
3541
|
+
maximumTurns: number;
|
|
3542
|
+
holdOpen: boolean;
|
|
3543
|
+
} | {
|
|
3544
|
+
providerId: "custom-openai-compatible";
|
|
3545
|
+
configurationHash: string;
|
|
3546
|
+
maximumRequestCostUsd: number;
|
|
3547
|
+
requestTimeoutMs: number;
|
|
3548
|
+
} | undefined;
|
|
3549
|
+
candidate?: {
|
|
3550
|
+
artifactId: string;
|
|
3551
|
+
contentHash: string;
|
|
3552
|
+
preparationId: string;
|
|
3553
|
+
dispatchId: string;
|
|
3554
|
+
baseProfileId: string;
|
|
3555
|
+
workerImageDigest: string;
|
|
3556
|
+
} | undefined;
|
|
1477
3557
|
harness?: {
|
|
1478
3558
|
harnessRelease: {
|
|
1479
3559
|
id: string;
|
|
@@ -1516,10 +3596,210 @@ export declare class OpenPondExperimentsClient {
|
|
|
1516
3596
|
actionId: string;
|
|
1517
3597
|
};
|
|
1518
3598
|
environmentHash: string;
|
|
1519
|
-
|
|
1520
|
-
|
|
1521
|
-
|
|
1522
|
-
|
|
3599
|
+
externalDatasetBinding?: {
|
|
3600
|
+
schemaVersion: "openpond.profileExternalDatasetBinding.v1";
|
|
3601
|
+
dataset: {
|
|
3602
|
+
id: string;
|
|
3603
|
+
contentHash: string;
|
|
3604
|
+
revision: number;
|
|
3605
|
+
};
|
|
3606
|
+
packageHash: string;
|
|
3607
|
+
profileId: string;
|
|
3608
|
+
target: {
|
|
3609
|
+
kind: "profile";
|
|
3610
|
+
} | {
|
|
3611
|
+
kind: "workflow";
|
|
3612
|
+
workflowId: string;
|
|
3613
|
+
} | {
|
|
3614
|
+
kind: "skill";
|
|
3615
|
+
skillPath: string;
|
|
3616
|
+
} | {
|
|
3617
|
+
kind: "agent_action";
|
|
3618
|
+
actionId: string;
|
|
3619
|
+
};
|
|
3620
|
+
declaredProfileCatalogHash: string | null;
|
|
3621
|
+
protectedProfileClosureHash: string;
|
|
3622
|
+
split: "train" | "validation" | "test" | "frozen_eval";
|
|
3623
|
+
population: {
|
|
3624
|
+
taskId: string;
|
|
3625
|
+
seed: string;
|
|
3626
|
+
fixtureId: string | null;
|
|
3627
|
+
}[];
|
|
3628
|
+
evaluators: {
|
|
3629
|
+
release: {
|
|
3630
|
+
id: string;
|
|
3631
|
+
contentHash: string;
|
|
3632
|
+
revision: string | number;
|
|
3633
|
+
};
|
|
3634
|
+
feedbackKey: string;
|
|
3635
|
+
configurationHash: string;
|
|
3636
|
+
output: "boolean" | "score" | "category";
|
|
3637
|
+
categories: string[];
|
|
3638
|
+
}[];
|
|
3639
|
+
fieldMappingsHash: string;
|
|
3640
|
+
fieldMappings: {
|
|
3641
|
+
graderId: string;
|
|
3642
|
+
fields: {
|
|
3643
|
+
destination: "input" | "output" | "expectedOutput" | "evaluatorContext";
|
|
3644
|
+
path: string[];
|
|
3645
|
+
source: "input" | "output" | "expectedOutput" | "evaluatorContext";
|
|
3646
|
+
sourcePath: string[];
|
|
3647
|
+
}[];
|
|
3648
|
+
}[];
|
|
3649
|
+
environmentHash: string;
|
|
3650
|
+
privateDatasetClosureHash: string;
|
|
3651
|
+
criterion: {
|
|
3652
|
+
minimumPassRate: number;
|
|
3653
|
+
requireComplete: boolean;
|
|
3654
|
+
};
|
|
3655
|
+
contentHash: string;
|
|
3656
|
+
recordedOrigin?: {
|
|
3657
|
+
execution: {
|
|
3658
|
+
id: string;
|
|
3659
|
+
contentHash: string;
|
|
3660
|
+
};
|
|
3661
|
+
originalEvidence: {
|
|
3662
|
+
id: string;
|
|
3663
|
+
contentHash: string;
|
|
3664
|
+
};
|
|
3665
|
+
members: {
|
|
3666
|
+
caseId: string;
|
|
3667
|
+
seed: string;
|
|
3668
|
+
fixtureId: string;
|
|
3669
|
+
sourceId: string;
|
|
3670
|
+
snapshotHash: string;
|
|
3671
|
+
boundaryRevisionHash: string;
|
|
3672
|
+
}[];
|
|
3673
|
+
} | undefined;
|
|
3674
|
+
gradingSource?: {
|
|
3675
|
+
pass: {
|
|
3676
|
+
id: string;
|
|
3677
|
+
contentHash: string;
|
|
3678
|
+
};
|
|
3679
|
+
evidence: {
|
|
3680
|
+
id: string;
|
|
3681
|
+
contentHash: string;
|
|
3682
|
+
};
|
|
3683
|
+
} | undefined;
|
|
3684
|
+
} | undefined;
|
|
3685
|
+
sourceCandidate?: {
|
|
3686
|
+
candidateId: string;
|
|
3687
|
+
frozenRevision: number;
|
|
3688
|
+
frozenHash: string;
|
|
3689
|
+
harnessRelease: {
|
|
3690
|
+
id: string;
|
|
3691
|
+
contentHash: string;
|
|
3692
|
+
};
|
|
3693
|
+
protectedClosureHash: string;
|
|
3694
|
+
} | undefined;
|
|
3695
|
+
};
|
|
3696
|
+
modelConfigurationHash: string;
|
|
3697
|
+
packageHash: string;
|
|
3698
|
+
candidate?: {
|
|
3699
|
+
artifactId: string;
|
|
3700
|
+
contentHash: string;
|
|
3701
|
+
preparationId: string;
|
|
3702
|
+
dispatchId: string;
|
|
3703
|
+
baseProfileId: string;
|
|
3704
|
+
workerImageDigest: string;
|
|
3705
|
+
} | undefined;
|
|
3706
|
+
sourceCandidate?: {
|
|
3707
|
+
candidateId: string;
|
|
3708
|
+
frozenRevision: number;
|
|
3709
|
+
frozenHash: string;
|
|
3710
|
+
harnessRelease: {
|
|
3711
|
+
id: string;
|
|
3712
|
+
contentHash: string;
|
|
3713
|
+
};
|
|
3714
|
+
protectedClosureHash: string;
|
|
3715
|
+
} | undefined;
|
|
3716
|
+
reasoningEffort?: "high" | "low" | "off" | "medium" | "xhigh" | "max" | undefined;
|
|
3717
|
+
externalDatasetBinding?: {
|
|
3718
|
+
schemaVersion: "openpond.profileExternalDatasetBinding.v1";
|
|
3719
|
+
dataset: {
|
|
3720
|
+
id: string;
|
|
3721
|
+
contentHash: string;
|
|
3722
|
+
revision: number;
|
|
3723
|
+
};
|
|
3724
|
+
packageHash: string;
|
|
3725
|
+
profileId: string;
|
|
3726
|
+
target: {
|
|
3727
|
+
kind: "profile";
|
|
3728
|
+
} | {
|
|
3729
|
+
kind: "workflow";
|
|
3730
|
+
workflowId: string;
|
|
3731
|
+
} | {
|
|
3732
|
+
kind: "skill";
|
|
3733
|
+
skillPath: string;
|
|
3734
|
+
} | {
|
|
3735
|
+
kind: "agent_action";
|
|
3736
|
+
actionId: string;
|
|
3737
|
+
};
|
|
3738
|
+
declaredProfileCatalogHash: string | null;
|
|
3739
|
+
protectedProfileClosureHash: string;
|
|
3740
|
+
split: "train" | "validation" | "test" | "frozen_eval";
|
|
3741
|
+
population: {
|
|
3742
|
+
taskId: string;
|
|
3743
|
+
seed: string;
|
|
3744
|
+
fixtureId: string | null;
|
|
3745
|
+
}[];
|
|
3746
|
+
evaluators: {
|
|
3747
|
+
release: {
|
|
3748
|
+
id: string;
|
|
3749
|
+
contentHash: string;
|
|
3750
|
+
revision: string | number;
|
|
3751
|
+
};
|
|
3752
|
+
feedbackKey: string;
|
|
3753
|
+
configurationHash: string;
|
|
3754
|
+
output: "boolean" | "score" | "category";
|
|
3755
|
+
categories: string[];
|
|
3756
|
+
}[];
|
|
3757
|
+
fieldMappingsHash: string;
|
|
3758
|
+
fieldMappings: {
|
|
3759
|
+
graderId: string;
|
|
3760
|
+
fields: {
|
|
3761
|
+
destination: "input" | "output" | "expectedOutput" | "evaluatorContext";
|
|
3762
|
+
path: string[];
|
|
3763
|
+
source: "input" | "output" | "expectedOutput" | "evaluatorContext";
|
|
3764
|
+
sourcePath: string[];
|
|
3765
|
+
}[];
|
|
3766
|
+
}[];
|
|
3767
|
+
environmentHash: string;
|
|
3768
|
+
privateDatasetClosureHash: string;
|
|
3769
|
+
criterion: {
|
|
3770
|
+
minimumPassRate: number;
|
|
3771
|
+
requireComplete: boolean;
|
|
3772
|
+
};
|
|
3773
|
+
contentHash: string;
|
|
3774
|
+
recordedOrigin?: {
|
|
3775
|
+
execution: {
|
|
3776
|
+
id: string;
|
|
3777
|
+
contentHash: string;
|
|
3778
|
+
};
|
|
3779
|
+
originalEvidence: {
|
|
3780
|
+
id: string;
|
|
3781
|
+
contentHash: string;
|
|
3782
|
+
};
|
|
3783
|
+
members: {
|
|
3784
|
+
caseId: string;
|
|
3785
|
+
seed: string;
|
|
3786
|
+
fixtureId: string;
|
|
3787
|
+
sourceId: string;
|
|
3788
|
+
snapshotHash: string;
|
|
3789
|
+
boundaryRevisionHash: string;
|
|
3790
|
+
}[];
|
|
3791
|
+
} | undefined;
|
|
3792
|
+
gradingSource?: {
|
|
3793
|
+
pass: {
|
|
3794
|
+
id: string;
|
|
3795
|
+
contentHash: string;
|
|
3796
|
+
};
|
|
3797
|
+
evidence: {
|
|
3798
|
+
id: string;
|
|
3799
|
+
contentHash: string;
|
|
3800
|
+
};
|
|
3801
|
+
} | undefined;
|
|
3802
|
+
} | undefined;
|
|
1523
3803
|
} | {
|
|
1524
3804
|
kind: "fixture";
|
|
1525
3805
|
};
|
|
@@ -1636,6 +3916,102 @@ export declare class OpenPondExperimentsClient {
|
|
|
1636
3916
|
actionId: string;
|
|
1637
3917
|
};
|
|
1638
3918
|
environmentHash: string;
|
|
3919
|
+
externalDatasetBinding?: {
|
|
3920
|
+
schemaVersion: "openpond.profileExternalDatasetBinding.v1";
|
|
3921
|
+
dataset: {
|
|
3922
|
+
id: string;
|
|
3923
|
+
contentHash: string;
|
|
3924
|
+
revision: number;
|
|
3925
|
+
};
|
|
3926
|
+
packageHash: string;
|
|
3927
|
+
profileId: string;
|
|
3928
|
+
target: {
|
|
3929
|
+
kind: "profile";
|
|
3930
|
+
} | {
|
|
3931
|
+
kind: "workflow";
|
|
3932
|
+
workflowId: string;
|
|
3933
|
+
} | {
|
|
3934
|
+
kind: "skill";
|
|
3935
|
+
skillPath: string;
|
|
3936
|
+
} | {
|
|
3937
|
+
kind: "agent_action";
|
|
3938
|
+
actionId: string;
|
|
3939
|
+
};
|
|
3940
|
+
declaredProfileCatalogHash: string | null;
|
|
3941
|
+
protectedProfileClosureHash: string;
|
|
3942
|
+
split: "train" | "validation" | "test" | "frozen_eval";
|
|
3943
|
+
population: {
|
|
3944
|
+
taskId: string;
|
|
3945
|
+
seed: string;
|
|
3946
|
+
fixtureId: string | null;
|
|
3947
|
+
}[];
|
|
3948
|
+
evaluators: {
|
|
3949
|
+
release: {
|
|
3950
|
+
id: string;
|
|
3951
|
+
contentHash: string;
|
|
3952
|
+
revision: string | number;
|
|
3953
|
+
};
|
|
3954
|
+
feedbackKey: string;
|
|
3955
|
+
configurationHash: string;
|
|
3956
|
+
output: "boolean" | "score" | "category";
|
|
3957
|
+
categories: string[];
|
|
3958
|
+
}[];
|
|
3959
|
+
fieldMappingsHash: string;
|
|
3960
|
+
fieldMappings: {
|
|
3961
|
+
graderId: string;
|
|
3962
|
+
fields: {
|
|
3963
|
+
destination: "input" | "output" | "expectedOutput" | "evaluatorContext";
|
|
3964
|
+
path: string[];
|
|
3965
|
+
source: "input" | "output" | "expectedOutput" | "evaluatorContext";
|
|
3966
|
+
sourcePath: string[];
|
|
3967
|
+
}[];
|
|
3968
|
+
}[];
|
|
3969
|
+
environmentHash: string;
|
|
3970
|
+
privateDatasetClosureHash: string;
|
|
3971
|
+
criterion: {
|
|
3972
|
+
minimumPassRate: number;
|
|
3973
|
+
requireComplete: boolean;
|
|
3974
|
+
};
|
|
3975
|
+
contentHash: string;
|
|
3976
|
+
recordedOrigin?: {
|
|
3977
|
+
execution: {
|
|
3978
|
+
id: string;
|
|
3979
|
+
contentHash: string;
|
|
3980
|
+
};
|
|
3981
|
+
originalEvidence: {
|
|
3982
|
+
id: string;
|
|
3983
|
+
contentHash: string;
|
|
3984
|
+
};
|
|
3985
|
+
members: {
|
|
3986
|
+
caseId: string;
|
|
3987
|
+
seed: string;
|
|
3988
|
+
fixtureId: string;
|
|
3989
|
+
sourceId: string;
|
|
3990
|
+
snapshotHash: string;
|
|
3991
|
+
boundaryRevisionHash: string;
|
|
3992
|
+
}[];
|
|
3993
|
+
} | undefined;
|
|
3994
|
+
gradingSource?: {
|
|
3995
|
+
pass: {
|
|
3996
|
+
id: string;
|
|
3997
|
+
contentHash: string;
|
|
3998
|
+
};
|
|
3999
|
+
evidence: {
|
|
4000
|
+
id: string;
|
|
4001
|
+
contentHash: string;
|
|
4002
|
+
};
|
|
4003
|
+
} | undefined;
|
|
4004
|
+
} | undefined;
|
|
4005
|
+
sourceCandidate?: {
|
|
4006
|
+
candidateId: string;
|
|
4007
|
+
frozenRevision: number;
|
|
4008
|
+
frozenHash: string;
|
|
4009
|
+
harnessRelease: {
|
|
4010
|
+
id: string;
|
|
4011
|
+
contentHash: string;
|
|
4012
|
+
};
|
|
4013
|
+
protectedClosureHash: string;
|
|
4014
|
+
} | undefined;
|
|
1639
4015
|
} | undefined;
|
|
1640
4016
|
};
|
|
1641
4017
|
policySnapshot: {
|
|
@@ -1661,6 +4037,30 @@ export declare class OpenPondExperimentsClient {
|
|
|
1661
4037
|
maxOutputTokens: number;
|
|
1662
4038
|
temperature: number;
|
|
1663
4039
|
topP: number;
|
|
4040
|
+
localRuntime?: {
|
|
4041
|
+
providerId: "claude-code";
|
|
4042
|
+
configurationHash: string;
|
|
4043
|
+
executableHash: string;
|
|
4044
|
+
version: string;
|
|
4045
|
+
capabilityHash: string;
|
|
4046
|
+
maximumRequestCostUsd: number;
|
|
4047
|
+
requestTimeoutMs: number;
|
|
4048
|
+
maximumTurns: number;
|
|
4049
|
+
holdOpen: boolean;
|
|
4050
|
+
} | {
|
|
4051
|
+
providerId: "custom-openai-compatible";
|
|
4052
|
+
configurationHash: string;
|
|
4053
|
+
maximumRequestCostUsd: number;
|
|
4054
|
+
requestTimeoutMs: number;
|
|
4055
|
+
} | undefined;
|
|
4056
|
+
candidate?: {
|
|
4057
|
+
artifactId: string;
|
|
4058
|
+
contentHash: string;
|
|
4059
|
+
preparationId: string;
|
|
4060
|
+
dispatchId: string;
|
|
4061
|
+
baseProfileId: string;
|
|
4062
|
+
workerImageDigest: string;
|
|
4063
|
+
} | undefined;
|
|
1664
4064
|
harness?: {
|
|
1665
4065
|
harnessRelease: {
|
|
1666
4066
|
id: string;
|
|
@@ -1703,10 +4103,210 @@ export declare class OpenPondExperimentsClient {
|
|
|
1703
4103
|
actionId: string;
|
|
1704
4104
|
};
|
|
1705
4105
|
environmentHash: string;
|
|
4106
|
+
externalDatasetBinding?: {
|
|
4107
|
+
schemaVersion: "openpond.profileExternalDatasetBinding.v1";
|
|
4108
|
+
dataset: {
|
|
4109
|
+
id: string;
|
|
4110
|
+
contentHash: string;
|
|
4111
|
+
revision: number;
|
|
4112
|
+
};
|
|
4113
|
+
packageHash: string;
|
|
4114
|
+
profileId: string;
|
|
4115
|
+
target: {
|
|
4116
|
+
kind: "profile";
|
|
4117
|
+
} | {
|
|
4118
|
+
kind: "workflow";
|
|
4119
|
+
workflowId: string;
|
|
4120
|
+
} | {
|
|
4121
|
+
kind: "skill";
|
|
4122
|
+
skillPath: string;
|
|
4123
|
+
} | {
|
|
4124
|
+
kind: "agent_action";
|
|
4125
|
+
actionId: string;
|
|
4126
|
+
};
|
|
4127
|
+
declaredProfileCatalogHash: string | null;
|
|
4128
|
+
protectedProfileClosureHash: string;
|
|
4129
|
+
split: "train" | "validation" | "test" | "frozen_eval";
|
|
4130
|
+
population: {
|
|
4131
|
+
taskId: string;
|
|
4132
|
+
seed: string;
|
|
4133
|
+
fixtureId: string | null;
|
|
4134
|
+
}[];
|
|
4135
|
+
evaluators: {
|
|
4136
|
+
release: {
|
|
4137
|
+
id: string;
|
|
4138
|
+
contentHash: string;
|
|
4139
|
+
revision: string | number;
|
|
4140
|
+
};
|
|
4141
|
+
feedbackKey: string;
|
|
4142
|
+
configurationHash: string;
|
|
4143
|
+
output: "boolean" | "score" | "category";
|
|
4144
|
+
categories: string[];
|
|
4145
|
+
}[];
|
|
4146
|
+
fieldMappingsHash: string;
|
|
4147
|
+
fieldMappings: {
|
|
4148
|
+
graderId: string;
|
|
4149
|
+
fields: {
|
|
4150
|
+
destination: "input" | "output" | "expectedOutput" | "evaluatorContext";
|
|
4151
|
+
path: string[];
|
|
4152
|
+
source: "input" | "output" | "expectedOutput" | "evaluatorContext";
|
|
4153
|
+
sourcePath: string[];
|
|
4154
|
+
}[];
|
|
4155
|
+
}[];
|
|
4156
|
+
environmentHash: string;
|
|
4157
|
+
privateDatasetClosureHash: string;
|
|
4158
|
+
criterion: {
|
|
4159
|
+
minimumPassRate: number;
|
|
4160
|
+
requireComplete: boolean;
|
|
4161
|
+
};
|
|
4162
|
+
contentHash: string;
|
|
4163
|
+
recordedOrigin?: {
|
|
4164
|
+
execution: {
|
|
4165
|
+
id: string;
|
|
4166
|
+
contentHash: string;
|
|
4167
|
+
};
|
|
4168
|
+
originalEvidence: {
|
|
4169
|
+
id: string;
|
|
4170
|
+
contentHash: string;
|
|
4171
|
+
};
|
|
4172
|
+
members: {
|
|
4173
|
+
caseId: string;
|
|
4174
|
+
seed: string;
|
|
4175
|
+
fixtureId: string;
|
|
4176
|
+
sourceId: string;
|
|
4177
|
+
snapshotHash: string;
|
|
4178
|
+
boundaryRevisionHash: string;
|
|
4179
|
+
}[];
|
|
4180
|
+
} | undefined;
|
|
4181
|
+
gradingSource?: {
|
|
4182
|
+
pass: {
|
|
4183
|
+
id: string;
|
|
4184
|
+
contentHash: string;
|
|
4185
|
+
};
|
|
4186
|
+
evidence: {
|
|
4187
|
+
id: string;
|
|
4188
|
+
contentHash: string;
|
|
4189
|
+
};
|
|
4190
|
+
} | undefined;
|
|
4191
|
+
} | undefined;
|
|
4192
|
+
sourceCandidate?: {
|
|
4193
|
+
candidateId: string;
|
|
4194
|
+
frozenRevision: number;
|
|
4195
|
+
frozenHash: string;
|
|
4196
|
+
harnessRelease: {
|
|
4197
|
+
id: string;
|
|
4198
|
+
contentHash: string;
|
|
4199
|
+
};
|
|
4200
|
+
protectedClosureHash: string;
|
|
4201
|
+
} | undefined;
|
|
1706
4202
|
};
|
|
1707
4203
|
modelConfigurationHash: string;
|
|
1708
4204
|
packageHash: string;
|
|
4205
|
+
candidate?: {
|
|
4206
|
+
artifactId: string;
|
|
4207
|
+
contentHash: string;
|
|
4208
|
+
preparationId: string;
|
|
4209
|
+
dispatchId: string;
|
|
4210
|
+
baseProfileId: string;
|
|
4211
|
+
workerImageDigest: string;
|
|
4212
|
+
} | undefined;
|
|
4213
|
+
sourceCandidate?: {
|
|
4214
|
+
candidateId: string;
|
|
4215
|
+
frozenRevision: number;
|
|
4216
|
+
frozenHash: string;
|
|
4217
|
+
harnessRelease: {
|
|
4218
|
+
id: string;
|
|
4219
|
+
contentHash: string;
|
|
4220
|
+
};
|
|
4221
|
+
protectedClosureHash: string;
|
|
4222
|
+
} | undefined;
|
|
1709
4223
|
reasoningEffort?: "high" | "low" | "off" | "medium" | "xhigh" | "max" | undefined;
|
|
4224
|
+
externalDatasetBinding?: {
|
|
4225
|
+
schemaVersion: "openpond.profileExternalDatasetBinding.v1";
|
|
4226
|
+
dataset: {
|
|
4227
|
+
id: string;
|
|
4228
|
+
contentHash: string;
|
|
4229
|
+
revision: number;
|
|
4230
|
+
};
|
|
4231
|
+
packageHash: string;
|
|
4232
|
+
profileId: string;
|
|
4233
|
+
target: {
|
|
4234
|
+
kind: "profile";
|
|
4235
|
+
} | {
|
|
4236
|
+
kind: "workflow";
|
|
4237
|
+
workflowId: string;
|
|
4238
|
+
} | {
|
|
4239
|
+
kind: "skill";
|
|
4240
|
+
skillPath: string;
|
|
4241
|
+
} | {
|
|
4242
|
+
kind: "agent_action";
|
|
4243
|
+
actionId: string;
|
|
4244
|
+
};
|
|
4245
|
+
declaredProfileCatalogHash: string | null;
|
|
4246
|
+
protectedProfileClosureHash: string;
|
|
4247
|
+
split: "train" | "validation" | "test" | "frozen_eval";
|
|
4248
|
+
population: {
|
|
4249
|
+
taskId: string;
|
|
4250
|
+
seed: string;
|
|
4251
|
+
fixtureId: string | null;
|
|
4252
|
+
}[];
|
|
4253
|
+
evaluators: {
|
|
4254
|
+
release: {
|
|
4255
|
+
id: string;
|
|
4256
|
+
contentHash: string;
|
|
4257
|
+
revision: string | number;
|
|
4258
|
+
};
|
|
4259
|
+
feedbackKey: string;
|
|
4260
|
+
configurationHash: string;
|
|
4261
|
+
output: "boolean" | "score" | "category";
|
|
4262
|
+
categories: string[];
|
|
4263
|
+
}[];
|
|
4264
|
+
fieldMappingsHash: string;
|
|
4265
|
+
fieldMappings: {
|
|
4266
|
+
graderId: string;
|
|
4267
|
+
fields: {
|
|
4268
|
+
destination: "input" | "output" | "expectedOutput" | "evaluatorContext";
|
|
4269
|
+
path: string[];
|
|
4270
|
+
source: "input" | "output" | "expectedOutput" | "evaluatorContext";
|
|
4271
|
+
sourcePath: string[];
|
|
4272
|
+
}[];
|
|
4273
|
+
}[];
|
|
4274
|
+
environmentHash: string;
|
|
4275
|
+
privateDatasetClosureHash: string;
|
|
4276
|
+
criterion: {
|
|
4277
|
+
minimumPassRate: number;
|
|
4278
|
+
requireComplete: boolean;
|
|
4279
|
+
};
|
|
4280
|
+
contentHash: string;
|
|
4281
|
+
recordedOrigin?: {
|
|
4282
|
+
execution: {
|
|
4283
|
+
id: string;
|
|
4284
|
+
contentHash: string;
|
|
4285
|
+
};
|
|
4286
|
+
originalEvidence: {
|
|
4287
|
+
id: string;
|
|
4288
|
+
contentHash: string;
|
|
4289
|
+
};
|
|
4290
|
+
members: {
|
|
4291
|
+
caseId: string;
|
|
4292
|
+
seed: string;
|
|
4293
|
+
fixtureId: string;
|
|
4294
|
+
sourceId: string;
|
|
4295
|
+
snapshotHash: string;
|
|
4296
|
+
boundaryRevisionHash: string;
|
|
4297
|
+
}[];
|
|
4298
|
+
} | undefined;
|
|
4299
|
+
gradingSource?: {
|
|
4300
|
+
pass: {
|
|
4301
|
+
id: string;
|
|
4302
|
+
contentHash: string;
|
|
4303
|
+
};
|
|
4304
|
+
evidence: {
|
|
4305
|
+
id: string;
|
|
4306
|
+
contentHash: string;
|
|
4307
|
+
};
|
|
4308
|
+
} | undefined;
|
|
4309
|
+
} | undefined;
|
|
1710
4310
|
} | {
|
|
1711
4311
|
kind: "fixture";
|
|
1712
4312
|
};
|
|
@@ -1764,7 +4364,7 @@ export declare class OpenPondExperimentsClient {
|
|
|
1764
4364
|
};
|
|
1765
4365
|
policyKind: "hosted_chat" | "fixture" | "hosted_harness";
|
|
1766
4366
|
manifestHash: string;
|
|
1767
|
-
status: "
|
|
4367
|
+
status: "cancelled" | "failed" | "queued" | "running" | "cancelling" | "completed";
|
|
1768
4368
|
totalCount: number;
|
|
1769
4369
|
counts: {
|
|
1770
4370
|
pending: number;
|
|
@@ -1808,6 +4408,30 @@ export declare class OpenPondExperimentsClient {
|
|
|
1808
4408
|
maxOutputTokens: number;
|
|
1809
4409
|
temperature: number;
|
|
1810
4410
|
topP: number;
|
|
4411
|
+
localRuntime?: {
|
|
4412
|
+
providerId: "claude-code";
|
|
4413
|
+
configurationHash: string;
|
|
4414
|
+
executableHash: string;
|
|
4415
|
+
version: string;
|
|
4416
|
+
capabilityHash: string;
|
|
4417
|
+
maximumRequestCostUsd: number;
|
|
4418
|
+
requestTimeoutMs: number;
|
|
4419
|
+
maximumTurns: number;
|
|
4420
|
+
holdOpen: boolean;
|
|
4421
|
+
} | {
|
|
4422
|
+
providerId: "custom-openai-compatible";
|
|
4423
|
+
configurationHash: string;
|
|
4424
|
+
maximumRequestCostUsd: number;
|
|
4425
|
+
requestTimeoutMs: number;
|
|
4426
|
+
} | undefined;
|
|
4427
|
+
candidate?: {
|
|
4428
|
+
artifactId: string;
|
|
4429
|
+
contentHash: string;
|
|
4430
|
+
preparationId: string;
|
|
4431
|
+
dispatchId: string;
|
|
4432
|
+
baseProfileId: string;
|
|
4433
|
+
workerImageDigest: string;
|
|
4434
|
+
} | undefined;
|
|
1811
4435
|
harness?: {
|
|
1812
4436
|
harnessRelease: {
|
|
1813
4437
|
id: string;
|
|
@@ -1850,10 +4474,210 @@ export declare class OpenPondExperimentsClient {
|
|
|
1850
4474
|
actionId: string;
|
|
1851
4475
|
};
|
|
1852
4476
|
environmentHash: string;
|
|
4477
|
+
externalDatasetBinding?: {
|
|
4478
|
+
schemaVersion: "openpond.profileExternalDatasetBinding.v1";
|
|
4479
|
+
dataset: {
|
|
4480
|
+
id: string;
|
|
4481
|
+
contentHash: string;
|
|
4482
|
+
revision: number;
|
|
4483
|
+
};
|
|
4484
|
+
packageHash: string;
|
|
4485
|
+
profileId: string;
|
|
4486
|
+
target: {
|
|
4487
|
+
kind: "profile";
|
|
4488
|
+
} | {
|
|
4489
|
+
kind: "workflow";
|
|
4490
|
+
workflowId: string;
|
|
4491
|
+
} | {
|
|
4492
|
+
kind: "skill";
|
|
4493
|
+
skillPath: string;
|
|
4494
|
+
} | {
|
|
4495
|
+
kind: "agent_action";
|
|
4496
|
+
actionId: string;
|
|
4497
|
+
};
|
|
4498
|
+
declaredProfileCatalogHash: string | null;
|
|
4499
|
+
protectedProfileClosureHash: string;
|
|
4500
|
+
split: "train" | "validation" | "test" | "frozen_eval";
|
|
4501
|
+
population: {
|
|
4502
|
+
taskId: string;
|
|
4503
|
+
seed: string;
|
|
4504
|
+
fixtureId: string | null;
|
|
4505
|
+
}[];
|
|
4506
|
+
evaluators: {
|
|
4507
|
+
release: {
|
|
4508
|
+
id: string;
|
|
4509
|
+
contentHash: string;
|
|
4510
|
+
revision: string | number;
|
|
4511
|
+
};
|
|
4512
|
+
feedbackKey: string;
|
|
4513
|
+
configurationHash: string;
|
|
4514
|
+
output: "boolean" | "score" | "category";
|
|
4515
|
+
categories: string[];
|
|
4516
|
+
}[];
|
|
4517
|
+
fieldMappingsHash: string;
|
|
4518
|
+
fieldMappings: {
|
|
4519
|
+
graderId: string;
|
|
4520
|
+
fields: {
|
|
4521
|
+
destination: "input" | "output" | "expectedOutput" | "evaluatorContext";
|
|
4522
|
+
path: string[];
|
|
4523
|
+
source: "input" | "output" | "expectedOutput" | "evaluatorContext";
|
|
4524
|
+
sourcePath: string[];
|
|
4525
|
+
}[];
|
|
4526
|
+
}[];
|
|
4527
|
+
environmentHash: string;
|
|
4528
|
+
privateDatasetClosureHash: string;
|
|
4529
|
+
criterion: {
|
|
4530
|
+
minimumPassRate: number;
|
|
4531
|
+
requireComplete: boolean;
|
|
4532
|
+
};
|
|
4533
|
+
contentHash: string;
|
|
4534
|
+
recordedOrigin?: {
|
|
4535
|
+
execution: {
|
|
4536
|
+
id: string;
|
|
4537
|
+
contentHash: string;
|
|
4538
|
+
};
|
|
4539
|
+
originalEvidence: {
|
|
4540
|
+
id: string;
|
|
4541
|
+
contentHash: string;
|
|
4542
|
+
};
|
|
4543
|
+
members: {
|
|
4544
|
+
caseId: string;
|
|
4545
|
+
seed: string;
|
|
4546
|
+
fixtureId: string;
|
|
4547
|
+
sourceId: string;
|
|
4548
|
+
snapshotHash: string;
|
|
4549
|
+
boundaryRevisionHash: string;
|
|
4550
|
+
}[];
|
|
4551
|
+
} | undefined;
|
|
4552
|
+
gradingSource?: {
|
|
4553
|
+
pass: {
|
|
4554
|
+
id: string;
|
|
4555
|
+
contentHash: string;
|
|
4556
|
+
};
|
|
4557
|
+
evidence: {
|
|
4558
|
+
id: string;
|
|
4559
|
+
contentHash: string;
|
|
4560
|
+
};
|
|
4561
|
+
} | undefined;
|
|
4562
|
+
} | undefined;
|
|
4563
|
+
sourceCandidate?: {
|
|
4564
|
+
candidateId: string;
|
|
4565
|
+
frozenRevision: number;
|
|
4566
|
+
frozenHash: string;
|
|
4567
|
+
harnessRelease: {
|
|
4568
|
+
id: string;
|
|
4569
|
+
contentHash: string;
|
|
4570
|
+
};
|
|
4571
|
+
protectedClosureHash: string;
|
|
4572
|
+
} | undefined;
|
|
1853
4573
|
};
|
|
1854
4574
|
modelConfigurationHash: string;
|
|
1855
4575
|
packageHash: string;
|
|
4576
|
+
candidate?: {
|
|
4577
|
+
artifactId: string;
|
|
4578
|
+
contentHash: string;
|
|
4579
|
+
preparationId: string;
|
|
4580
|
+
dispatchId: string;
|
|
4581
|
+
baseProfileId: string;
|
|
4582
|
+
workerImageDigest: string;
|
|
4583
|
+
} | undefined;
|
|
4584
|
+
sourceCandidate?: {
|
|
4585
|
+
candidateId: string;
|
|
4586
|
+
frozenRevision: number;
|
|
4587
|
+
frozenHash: string;
|
|
4588
|
+
harnessRelease: {
|
|
4589
|
+
id: string;
|
|
4590
|
+
contentHash: string;
|
|
4591
|
+
};
|
|
4592
|
+
protectedClosureHash: string;
|
|
4593
|
+
} | undefined;
|
|
1856
4594
|
reasoningEffort?: "high" | "low" | "off" | "medium" | "xhigh" | "max" | undefined;
|
|
4595
|
+
externalDatasetBinding?: {
|
|
4596
|
+
schemaVersion: "openpond.profileExternalDatasetBinding.v1";
|
|
4597
|
+
dataset: {
|
|
4598
|
+
id: string;
|
|
4599
|
+
contentHash: string;
|
|
4600
|
+
revision: number;
|
|
4601
|
+
};
|
|
4602
|
+
packageHash: string;
|
|
4603
|
+
profileId: string;
|
|
4604
|
+
target: {
|
|
4605
|
+
kind: "profile";
|
|
4606
|
+
} | {
|
|
4607
|
+
kind: "workflow";
|
|
4608
|
+
workflowId: string;
|
|
4609
|
+
} | {
|
|
4610
|
+
kind: "skill";
|
|
4611
|
+
skillPath: string;
|
|
4612
|
+
} | {
|
|
4613
|
+
kind: "agent_action";
|
|
4614
|
+
actionId: string;
|
|
4615
|
+
};
|
|
4616
|
+
declaredProfileCatalogHash: string | null;
|
|
4617
|
+
protectedProfileClosureHash: string;
|
|
4618
|
+
split: "train" | "validation" | "test" | "frozen_eval";
|
|
4619
|
+
population: {
|
|
4620
|
+
taskId: string;
|
|
4621
|
+
seed: string;
|
|
4622
|
+
fixtureId: string | null;
|
|
4623
|
+
}[];
|
|
4624
|
+
evaluators: {
|
|
4625
|
+
release: {
|
|
4626
|
+
id: string;
|
|
4627
|
+
contentHash: string;
|
|
4628
|
+
revision: string | number;
|
|
4629
|
+
};
|
|
4630
|
+
feedbackKey: string;
|
|
4631
|
+
configurationHash: string;
|
|
4632
|
+
output: "boolean" | "score" | "category";
|
|
4633
|
+
categories: string[];
|
|
4634
|
+
}[];
|
|
4635
|
+
fieldMappingsHash: string;
|
|
4636
|
+
fieldMappings: {
|
|
4637
|
+
graderId: string;
|
|
4638
|
+
fields: {
|
|
4639
|
+
destination: "input" | "output" | "expectedOutput" | "evaluatorContext";
|
|
4640
|
+
path: string[];
|
|
4641
|
+
source: "input" | "output" | "expectedOutput" | "evaluatorContext";
|
|
4642
|
+
sourcePath: string[];
|
|
4643
|
+
}[];
|
|
4644
|
+
}[];
|
|
4645
|
+
environmentHash: string;
|
|
4646
|
+
privateDatasetClosureHash: string;
|
|
4647
|
+
criterion: {
|
|
4648
|
+
minimumPassRate: number;
|
|
4649
|
+
requireComplete: boolean;
|
|
4650
|
+
};
|
|
4651
|
+
contentHash: string;
|
|
4652
|
+
recordedOrigin?: {
|
|
4653
|
+
execution: {
|
|
4654
|
+
id: string;
|
|
4655
|
+
contentHash: string;
|
|
4656
|
+
};
|
|
4657
|
+
originalEvidence: {
|
|
4658
|
+
id: string;
|
|
4659
|
+
contentHash: string;
|
|
4660
|
+
};
|
|
4661
|
+
members: {
|
|
4662
|
+
caseId: string;
|
|
4663
|
+
seed: string;
|
|
4664
|
+
fixtureId: string;
|
|
4665
|
+
sourceId: string;
|
|
4666
|
+
snapshotHash: string;
|
|
4667
|
+
boundaryRevisionHash: string;
|
|
4668
|
+
}[];
|
|
4669
|
+
} | undefined;
|
|
4670
|
+
gradingSource?: {
|
|
4671
|
+
pass: {
|
|
4672
|
+
id: string;
|
|
4673
|
+
contentHash: string;
|
|
4674
|
+
};
|
|
4675
|
+
evidence: {
|
|
4676
|
+
id: string;
|
|
4677
|
+
contentHash: string;
|
|
4678
|
+
};
|
|
4679
|
+
} | undefined;
|
|
4680
|
+
} | undefined;
|
|
1857
4681
|
} | {
|
|
1858
4682
|
kind: "fixture";
|
|
1859
4683
|
};
|
|
@@ -1970,6 +4794,102 @@ export declare class OpenPondExperimentsClient {
|
|
|
1970
4794
|
actionId: string;
|
|
1971
4795
|
};
|
|
1972
4796
|
environmentHash: string;
|
|
4797
|
+
externalDatasetBinding?: {
|
|
4798
|
+
schemaVersion: "openpond.profileExternalDatasetBinding.v1";
|
|
4799
|
+
dataset: {
|
|
4800
|
+
id: string;
|
|
4801
|
+
contentHash: string;
|
|
4802
|
+
revision: number;
|
|
4803
|
+
};
|
|
4804
|
+
packageHash: string;
|
|
4805
|
+
profileId: string;
|
|
4806
|
+
target: {
|
|
4807
|
+
kind: "profile";
|
|
4808
|
+
} | {
|
|
4809
|
+
kind: "workflow";
|
|
4810
|
+
workflowId: string;
|
|
4811
|
+
} | {
|
|
4812
|
+
kind: "skill";
|
|
4813
|
+
skillPath: string;
|
|
4814
|
+
} | {
|
|
4815
|
+
kind: "agent_action";
|
|
4816
|
+
actionId: string;
|
|
4817
|
+
};
|
|
4818
|
+
declaredProfileCatalogHash: string | null;
|
|
4819
|
+
protectedProfileClosureHash: string;
|
|
4820
|
+
split: "train" | "validation" | "test" | "frozen_eval";
|
|
4821
|
+
population: {
|
|
4822
|
+
taskId: string;
|
|
4823
|
+
seed: string;
|
|
4824
|
+
fixtureId: string | null;
|
|
4825
|
+
}[];
|
|
4826
|
+
evaluators: {
|
|
4827
|
+
release: {
|
|
4828
|
+
id: string;
|
|
4829
|
+
contentHash: string;
|
|
4830
|
+
revision: string | number;
|
|
4831
|
+
};
|
|
4832
|
+
feedbackKey: string;
|
|
4833
|
+
configurationHash: string;
|
|
4834
|
+
output: "boolean" | "score" | "category";
|
|
4835
|
+
categories: string[];
|
|
4836
|
+
}[];
|
|
4837
|
+
fieldMappingsHash: string;
|
|
4838
|
+
fieldMappings: {
|
|
4839
|
+
graderId: string;
|
|
4840
|
+
fields: {
|
|
4841
|
+
destination: "input" | "output" | "expectedOutput" | "evaluatorContext";
|
|
4842
|
+
path: string[];
|
|
4843
|
+
source: "input" | "output" | "expectedOutput" | "evaluatorContext";
|
|
4844
|
+
sourcePath: string[];
|
|
4845
|
+
}[];
|
|
4846
|
+
}[];
|
|
4847
|
+
environmentHash: string;
|
|
4848
|
+
privateDatasetClosureHash: string;
|
|
4849
|
+
criterion: {
|
|
4850
|
+
minimumPassRate: number;
|
|
4851
|
+
requireComplete: boolean;
|
|
4852
|
+
};
|
|
4853
|
+
contentHash: string;
|
|
4854
|
+
recordedOrigin?: {
|
|
4855
|
+
execution: {
|
|
4856
|
+
id: string;
|
|
4857
|
+
contentHash: string;
|
|
4858
|
+
};
|
|
4859
|
+
originalEvidence: {
|
|
4860
|
+
id: string;
|
|
4861
|
+
contentHash: string;
|
|
4862
|
+
};
|
|
4863
|
+
members: {
|
|
4864
|
+
caseId: string;
|
|
4865
|
+
seed: string;
|
|
4866
|
+
fixtureId: string;
|
|
4867
|
+
sourceId: string;
|
|
4868
|
+
snapshotHash: string;
|
|
4869
|
+
boundaryRevisionHash: string;
|
|
4870
|
+
}[];
|
|
4871
|
+
} | undefined;
|
|
4872
|
+
gradingSource?: {
|
|
4873
|
+
pass: {
|
|
4874
|
+
id: string;
|
|
4875
|
+
contentHash: string;
|
|
4876
|
+
};
|
|
4877
|
+
evidence: {
|
|
4878
|
+
id: string;
|
|
4879
|
+
contentHash: string;
|
|
4880
|
+
};
|
|
4881
|
+
} | undefined;
|
|
4882
|
+
} | undefined;
|
|
4883
|
+
sourceCandidate?: {
|
|
4884
|
+
candidateId: string;
|
|
4885
|
+
frozenRevision: number;
|
|
4886
|
+
frozenHash: string;
|
|
4887
|
+
harnessRelease: {
|
|
4888
|
+
id: string;
|
|
4889
|
+
contentHash: string;
|
|
4890
|
+
};
|
|
4891
|
+
protectedClosureHash: string;
|
|
4892
|
+
} | undefined;
|
|
1973
4893
|
} | undefined;
|
|
1974
4894
|
};
|
|
1975
4895
|
policySnapshot: {
|
|
@@ -1995,6 +4915,30 @@ export declare class OpenPondExperimentsClient {
|
|
|
1995
4915
|
maxOutputTokens: number;
|
|
1996
4916
|
temperature: number;
|
|
1997
4917
|
topP: number;
|
|
4918
|
+
localRuntime?: {
|
|
4919
|
+
providerId: "claude-code";
|
|
4920
|
+
configurationHash: string;
|
|
4921
|
+
executableHash: string;
|
|
4922
|
+
version: string;
|
|
4923
|
+
capabilityHash: string;
|
|
4924
|
+
maximumRequestCostUsd: number;
|
|
4925
|
+
requestTimeoutMs: number;
|
|
4926
|
+
maximumTurns: number;
|
|
4927
|
+
holdOpen: boolean;
|
|
4928
|
+
} | {
|
|
4929
|
+
providerId: "custom-openai-compatible";
|
|
4930
|
+
configurationHash: string;
|
|
4931
|
+
maximumRequestCostUsd: number;
|
|
4932
|
+
requestTimeoutMs: number;
|
|
4933
|
+
} | undefined;
|
|
4934
|
+
candidate?: {
|
|
4935
|
+
artifactId: string;
|
|
4936
|
+
contentHash: string;
|
|
4937
|
+
preparationId: string;
|
|
4938
|
+
dispatchId: string;
|
|
4939
|
+
baseProfileId: string;
|
|
4940
|
+
workerImageDigest: string;
|
|
4941
|
+
} | undefined;
|
|
1998
4942
|
harness?: {
|
|
1999
4943
|
harnessRelease: {
|
|
2000
4944
|
id: string;
|
|
@@ -2037,10 +4981,210 @@ export declare class OpenPondExperimentsClient {
|
|
|
2037
4981
|
actionId: string;
|
|
2038
4982
|
};
|
|
2039
4983
|
environmentHash: string;
|
|
4984
|
+
externalDatasetBinding?: {
|
|
4985
|
+
schemaVersion: "openpond.profileExternalDatasetBinding.v1";
|
|
4986
|
+
dataset: {
|
|
4987
|
+
id: string;
|
|
4988
|
+
contentHash: string;
|
|
4989
|
+
revision: number;
|
|
4990
|
+
};
|
|
4991
|
+
packageHash: string;
|
|
4992
|
+
profileId: string;
|
|
4993
|
+
target: {
|
|
4994
|
+
kind: "profile";
|
|
4995
|
+
} | {
|
|
4996
|
+
kind: "workflow";
|
|
4997
|
+
workflowId: string;
|
|
4998
|
+
} | {
|
|
4999
|
+
kind: "skill";
|
|
5000
|
+
skillPath: string;
|
|
5001
|
+
} | {
|
|
5002
|
+
kind: "agent_action";
|
|
5003
|
+
actionId: string;
|
|
5004
|
+
};
|
|
5005
|
+
declaredProfileCatalogHash: string | null;
|
|
5006
|
+
protectedProfileClosureHash: string;
|
|
5007
|
+
split: "train" | "validation" | "test" | "frozen_eval";
|
|
5008
|
+
population: {
|
|
5009
|
+
taskId: string;
|
|
5010
|
+
seed: string;
|
|
5011
|
+
fixtureId: string | null;
|
|
5012
|
+
}[];
|
|
5013
|
+
evaluators: {
|
|
5014
|
+
release: {
|
|
5015
|
+
id: string;
|
|
5016
|
+
contentHash: string;
|
|
5017
|
+
revision: string | number;
|
|
5018
|
+
};
|
|
5019
|
+
feedbackKey: string;
|
|
5020
|
+
configurationHash: string;
|
|
5021
|
+
output: "boolean" | "score" | "category";
|
|
5022
|
+
categories: string[];
|
|
5023
|
+
}[];
|
|
5024
|
+
fieldMappingsHash: string;
|
|
5025
|
+
fieldMappings: {
|
|
5026
|
+
graderId: string;
|
|
5027
|
+
fields: {
|
|
5028
|
+
destination: "input" | "output" | "expectedOutput" | "evaluatorContext";
|
|
5029
|
+
path: string[];
|
|
5030
|
+
source: "input" | "output" | "expectedOutput" | "evaluatorContext";
|
|
5031
|
+
sourcePath: string[];
|
|
5032
|
+
}[];
|
|
5033
|
+
}[];
|
|
5034
|
+
environmentHash: string;
|
|
5035
|
+
privateDatasetClosureHash: string;
|
|
5036
|
+
criterion: {
|
|
5037
|
+
minimumPassRate: number;
|
|
5038
|
+
requireComplete: boolean;
|
|
5039
|
+
};
|
|
5040
|
+
contentHash: string;
|
|
5041
|
+
recordedOrigin?: {
|
|
5042
|
+
execution: {
|
|
5043
|
+
id: string;
|
|
5044
|
+
contentHash: string;
|
|
5045
|
+
};
|
|
5046
|
+
originalEvidence: {
|
|
5047
|
+
id: string;
|
|
5048
|
+
contentHash: string;
|
|
5049
|
+
};
|
|
5050
|
+
members: {
|
|
5051
|
+
caseId: string;
|
|
5052
|
+
seed: string;
|
|
5053
|
+
fixtureId: string;
|
|
5054
|
+
sourceId: string;
|
|
5055
|
+
snapshotHash: string;
|
|
5056
|
+
boundaryRevisionHash: string;
|
|
5057
|
+
}[];
|
|
5058
|
+
} | undefined;
|
|
5059
|
+
gradingSource?: {
|
|
5060
|
+
pass: {
|
|
5061
|
+
id: string;
|
|
5062
|
+
contentHash: string;
|
|
5063
|
+
};
|
|
5064
|
+
evidence: {
|
|
5065
|
+
id: string;
|
|
5066
|
+
contentHash: string;
|
|
5067
|
+
};
|
|
5068
|
+
} | undefined;
|
|
5069
|
+
} | undefined;
|
|
5070
|
+
sourceCandidate?: {
|
|
5071
|
+
candidateId: string;
|
|
5072
|
+
frozenRevision: number;
|
|
5073
|
+
frozenHash: string;
|
|
5074
|
+
harnessRelease: {
|
|
5075
|
+
id: string;
|
|
5076
|
+
contentHash: string;
|
|
5077
|
+
};
|
|
5078
|
+
protectedClosureHash: string;
|
|
5079
|
+
} | undefined;
|
|
2040
5080
|
};
|
|
2041
5081
|
modelConfigurationHash: string;
|
|
2042
5082
|
packageHash: string;
|
|
5083
|
+
candidate?: {
|
|
5084
|
+
artifactId: string;
|
|
5085
|
+
contentHash: string;
|
|
5086
|
+
preparationId: string;
|
|
5087
|
+
dispatchId: string;
|
|
5088
|
+
baseProfileId: string;
|
|
5089
|
+
workerImageDigest: string;
|
|
5090
|
+
} | undefined;
|
|
5091
|
+
sourceCandidate?: {
|
|
5092
|
+
candidateId: string;
|
|
5093
|
+
frozenRevision: number;
|
|
5094
|
+
frozenHash: string;
|
|
5095
|
+
harnessRelease: {
|
|
5096
|
+
id: string;
|
|
5097
|
+
contentHash: string;
|
|
5098
|
+
};
|
|
5099
|
+
protectedClosureHash: string;
|
|
5100
|
+
} | undefined;
|
|
2043
5101
|
reasoningEffort?: "high" | "low" | "off" | "medium" | "xhigh" | "max" | undefined;
|
|
5102
|
+
externalDatasetBinding?: {
|
|
5103
|
+
schemaVersion: "openpond.profileExternalDatasetBinding.v1";
|
|
5104
|
+
dataset: {
|
|
5105
|
+
id: string;
|
|
5106
|
+
contentHash: string;
|
|
5107
|
+
revision: number;
|
|
5108
|
+
};
|
|
5109
|
+
packageHash: string;
|
|
5110
|
+
profileId: string;
|
|
5111
|
+
target: {
|
|
5112
|
+
kind: "profile";
|
|
5113
|
+
} | {
|
|
5114
|
+
kind: "workflow";
|
|
5115
|
+
workflowId: string;
|
|
5116
|
+
} | {
|
|
5117
|
+
kind: "skill";
|
|
5118
|
+
skillPath: string;
|
|
5119
|
+
} | {
|
|
5120
|
+
kind: "agent_action";
|
|
5121
|
+
actionId: string;
|
|
5122
|
+
};
|
|
5123
|
+
declaredProfileCatalogHash: string | null;
|
|
5124
|
+
protectedProfileClosureHash: string;
|
|
5125
|
+
split: "train" | "validation" | "test" | "frozen_eval";
|
|
5126
|
+
population: {
|
|
5127
|
+
taskId: string;
|
|
5128
|
+
seed: string;
|
|
5129
|
+
fixtureId: string | null;
|
|
5130
|
+
}[];
|
|
5131
|
+
evaluators: {
|
|
5132
|
+
release: {
|
|
5133
|
+
id: string;
|
|
5134
|
+
contentHash: string;
|
|
5135
|
+
revision: string | number;
|
|
5136
|
+
};
|
|
5137
|
+
feedbackKey: string;
|
|
5138
|
+
configurationHash: string;
|
|
5139
|
+
output: "boolean" | "score" | "category";
|
|
5140
|
+
categories: string[];
|
|
5141
|
+
}[];
|
|
5142
|
+
fieldMappingsHash: string;
|
|
5143
|
+
fieldMappings: {
|
|
5144
|
+
graderId: string;
|
|
5145
|
+
fields: {
|
|
5146
|
+
destination: "input" | "output" | "expectedOutput" | "evaluatorContext";
|
|
5147
|
+
path: string[];
|
|
5148
|
+
source: "input" | "output" | "expectedOutput" | "evaluatorContext";
|
|
5149
|
+
sourcePath: string[];
|
|
5150
|
+
}[];
|
|
5151
|
+
}[];
|
|
5152
|
+
environmentHash: string;
|
|
5153
|
+
privateDatasetClosureHash: string;
|
|
5154
|
+
criterion: {
|
|
5155
|
+
minimumPassRate: number;
|
|
5156
|
+
requireComplete: boolean;
|
|
5157
|
+
};
|
|
5158
|
+
contentHash: string;
|
|
5159
|
+
recordedOrigin?: {
|
|
5160
|
+
execution: {
|
|
5161
|
+
id: string;
|
|
5162
|
+
contentHash: string;
|
|
5163
|
+
};
|
|
5164
|
+
originalEvidence: {
|
|
5165
|
+
id: string;
|
|
5166
|
+
contentHash: string;
|
|
5167
|
+
};
|
|
5168
|
+
members: {
|
|
5169
|
+
caseId: string;
|
|
5170
|
+
seed: string;
|
|
5171
|
+
fixtureId: string;
|
|
5172
|
+
sourceId: string;
|
|
5173
|
+
snapshotHash: string;
|
|
5174
|
+
boundaryRevisionHash: string;
|
|
5175
|
+
}[];
|
|
5176
|
+
} | undefined;
|
|
5177
|
+
gradingSource?: {
|
|
5178
|
+
pass: {
|
|
5179
|
+
id: string;
|
|
5180
|
+
contentHash: string;
|
|
5181
|
+
};
|
|
5182
|
+
evidence: {
|
|
5183
|
+
id: string;
|
|
5184
|
+
contentHash: string;
|
|
5185
|
+
};
|
|
5186
|
+
} | undefined;
|
|
5187
|
+
} | undefined;
|
|
2044
5188
|
} | {
|
|
2045
5189
|
kind: "fixture";
|
|
2046
5190
|
};
|
|
@@ -2099,6 +5243,7 @@ export declare class OpenPondExperimentsClient {
|
|
|
2099
5243
|
revision: number;
|
|
2100
5244
|
}[];
|
|
2101
5245
|
maximumCostUsd: number;
|
|
5246
|
+
executionKind?: "recorded_evidence" | undefined;
|
|
2102
5247
|
mappings?: {
|
|
2103
5248
|
graderId: string;
|
|
2104
5249
|
fields: {
|
|
@@ -2127,7 +5272,7 @@ export declare class OpenPondExperimentsClient {
|
|
|
2127
5272
|
sourcePath: string[];
|
|
2128
5273
|
}[] | undefined;
|
|
2129
5274
|
}[];
|
|
2130
|
-
status: "
|
|
5275
|
+
status: "cancelled" | "failed" | "queued" | "running" | "cancelling" | "completed";
|
|
2131
5276
|
totalCount: number;
|
|
2132
5277
|
counts: {
|
|
2133
5278
|
pending: number;
|
|
@@ -2164,6 +5309,7 @@ export declare class OpenPondExperimentsClient {
|
|
|
2164
5309
|
revision: number;
|
|
2165
5310
|
}[];
|
|
2166
5311
|
maximumCostUsd: number;
|
|
5312
|
+
executionKind?: "recorded_evidence" | undefined;
|
|
2167
5313
|
mappings?: {
|
|
2168
5314
|
graderId: string;
|
|
2169
5315
|
fields: {
|
|
@@ -2192,7 +5338,7 @@ export declare class OpenPondExperimentsClient {
|
|
|
2192
5338
|
sourcePath: string[];
|
|
2193
5339
|
}[] | undefined;
|
|
2194
5340
|
}[];
|
|
2195
|
-
status: "
|
|
5341
|
+
status: "cancelled" | "failed" | "queued" | "running" | "cancelling" | "completed";
|
|
2196
5342
|
totalCount: number;
|
|
2197
5343
|
counts: {
|
|
2198
5344
|
pending: number;
|
|
@@ -2214,6 +5360,7 @@ export declare class OpenPondExperimentsClient {
|
|
|
2214
5360
|
}>;
|
|
2215
5361
|
scoringPasses(executionId: string, options?: {
|
|
2216
5362
|
afterId?: string;
|
|
5363
|
+
executionKind?: "recorded_evidence";
|
|
2217
5364
|
signal?: AbortSignal;
|
|
2218
5365
|
}): Promise<{
|
|
2219
5366
|
items: {
|
|
@@ -2233,6 +5380,7 @@ export declare class OpenPondExperimentsClient {
|
|
|
2233
5380
|
revision: number;
|
|
2234
5381
|
}[];
|
|
2235
5382
|
maximumCostUsd: number;
|
|
5383
|
+
executionKind?: "recorded_evidence" | undefined;
|
|
2236
5384
|
mappings?: {
|
|
2237
5385
|
graderId: string;
|
|
2238
5386
|
fields: {
|
|
@@ -2261,7 +5409,7 @@ export declare class OpenPondExperimentsClient {
|
|
|
2261
5409
|
sourcePath: string[];
|
|
2262
5410
|
}[] | undefined;
|
|
2263
5411
|
}[];
|
|
2264
|
-
status: "
|
|
5412
|
+
status: "cancelled" | "failed" | "queued" | "running" | "cancelling" | "completed";
|
|
2265
5413
|
totalCount: number;
|
|
2266
5414
|
counts: {
|
|
2267
5415
|
pending: number;
|
|
@@ -2300,6 +5448,7 @@ export declare class OpenPondExperimentsClient {
|
|
|
2300
5448
|
revision: number;
|
|
2301
5449
|
}[];
|
|
2302
5450
|
maximumCostUsd: number;
|
|
5451
|
+
executionKind?: "recorded_evidence" | undefined;
|
|
2303
5452
|
mappings?: {
|
|
2304
5453
|
graderId: string;
|
|
2305
5454
|
fields: {
|
|
@@ -2328,7 +5477,7 @@ export declare class OpenPondExperimentsClient {
|
|
|
2328
5477
|
sourcePath: string[];
|
|
2329
5478
|
}[] | undefined;
|
|
2330
5479
|
}[];
|
|
2331
|
-
status: "
|
|
5480
|
+
status: "cancelled" | "failed" | "queued" | "running" | "cancelling" | "completed";
|
|
2332
5481
|
totalCount: number;
|
|
2333
5482
|
counts: {
|
|
2334
5483
|
pending: number;
|
|
@@ -2392,6 +5541,102 @@ export declare class OpenPondExperimentsClient {
|
|
|
2392
5541
|
actionId: string;
|
|
2393
5542
|
};
|
|
2394
5543
|
environmentHash: string;
|
|
5544
|
+
externalDatasetBinding?: {
|
|
5545
|
+
schemaVersion: "openpond.profileExternalDatasetBinding.v1";
|
|
5546
|
+
dataset: {
|
|
5547
|
+
id: string;
|
|
5548
|
+
contentHash: string;
|
|
5549
|
+
revision: number;
|
|
5550
|
+
};
|
|
5551
|
+
packageHash: string;
|
|
5552
|
+
profileId: string;
|
|
5553
|
+
target: {
|
|
5554
|
+
kind: "profile";
|
|
5555
|
+
} | {
|
|
5556
|
+
kind: "workflow";
|
|
5557
|
+
workflowId: string;
|
|
5558
|
+
} | {
|
|
5559
|
+
kind: "skill";
|
|
5560
|
+
skillPath: string;
|
|
5561
|
+
} | {
|
|
5562
|
+
kind: "agent_action";
|
|
5563
|
+
actionId: string;
|
|
5564
|
+
};
|
|
5565
|
+
declaredProfileCatalogHash: string | null;
|
|
5566
|
+
protectedProfileClosureHash: string;
|
|
5567
|
+
split: "train" | "validation" | "test" | "frozen_eval";
|
|
5568
|
+
population: {
|
|
5569
|
+
taskId: string;
|
|
5570
|
+
seed: string;
|
|
5571
|
+
fixtureId: string | null;
|
|
5572
|
+
}[];
|
|
5573
|
+
evaluators: {
|
|
5574
|
+
release: {
|
|
5575
|
+
id: string;
|
|
5576
|
+
contentHash: string;
|
|
5577
|
+
revision: string | number;
|
|
5578
|
+
};
|
|
5579
|
+
feedbackKey: string;
|
|
5580
|
+
configurationHash: string;
|
|
5581
|
+
output: "boolean" | "score" | "category";
|
|
5582
|
+
categories: string[];
|
|
5583
|
+
}[];
|
|
5584
|
+
fieldMappingsHash: string;
|
|
5585
|
+
fieldMappings: {
|
|
5586
|
+
graderId: string;
|
|
5587
|
+
fields: {
|
|
5588
|
+
destination: "output" | "input" | "expectedOutput" | "evaluatorContext";
|
|
5589
|
+
path: string[];
|
|
5590
|
+
source: "output" | "input" | "expectedOutput" | "evaluatorContext";
|
|
5591
|
+
sourcePath: string[];
|
|
5592
|
+
}[];
|
|
5593
|
+
}[];
|
|
5594
|
+
environmentHash: string;
|
|
5595
|
+
privateDatasetClosureHash: string;
|
|
5596
|
+
criterion: {
|
|
5597
|
+
minimumPassRate: number;
|
|
5598
|
+
requireComplete: boolean;
|
|
5599
|
+
};
|
|
5600
|
+
contentHash: string;
|
|
5601
|
+
recordedOrigin?: {
|
|
5602
|
+
execution: {
|
|
5603
|
+
id: string;
|
|
5604
|
+
contentHash: string;
|
|
5605
|
+
};
|
|
5606
|
+
originalEvidence: {
|
|
5607
|
+
id: string;
|
|
5608
|
+
contentHash: string;
|
|
5609
|
+
};
|
|
5610
|
+
members: {
|
|
5611
|
+
caseId: string;
|
|
5612
|
+
seed: string;
|
|
5613
|
+
fixtureId: string;
|
|
5614
|
+
sourceId: string;
|
|
5615
|
+
snapshotHash: string;
|
|
5616
|
+
boundaryRevisionHash: string;
|
|
5617
|
+
}[];
|
|
5618
|
+
} | undefined;
|
|
5619
|
+
gradingSource?: {
|
|
5620
|
+
pass: {
|
|
5621
|
+
id: string;
|
|
5622
|
+
contentHash: string;
|
|
5623
|
+
};
|
|
5624
|
+
evidence: {
|
|
5625
|
+
id: string;
|
|
5626
|
+
contentHash: string;
|
|
5627
|
+
};
|
|
5628
|
+
} | undefined;
|
|
5629
|
+
} | undefined;
|
|
5630
|
+
sourceCandidate?: {
|
|
5631
|
+
candidateId: string;
|
|
5632
|
+
frozenRevision: number;
|
|
5633
|
+
frozenHash: string;
|
|
5634
|
+
harnessRelease: {
|
|
5635
|
+
id: string;
|
|
5636
|
+
contentHash: string;
|
|
5637
|
+
};
|
|
5638
|
+
protectedClosureHash: string;
|
|
5639
|
+
} | undefined;
|
|
2395
5640
|
};
|
|
2396
5641
|
model: {
|
|
2397
5642
|
modelId: string;
|
|
@@ -2456,6 +5701,14 @@ export declare class OpenPondExperimentsClient {
|
|
|
2456
5701
|
contentHash: string;
|
|
2457
5702
|
};
|
|
2458
5703
|
scoringPassId: string | null;
|
|
5704
|
+
humanProjection?: {
|
|
5705
|
+
originalManifest: {
|
|
5706
|
+
id: string;
|
|
5707
|
+
contentHash: string;
|
|
5708
|
+
};
|
|
5709
|
+
selectionHash: string;
|
|
5710
|
+
rulesHash: string;
|
|
5711
|
+
} | undefined;
|
|
2459
5712
|
} | undefined;
|
|
2460
5713
|
};
|
|
2461
5714
|
result: {
|
|
@@ -2464,14 +5717,14 @@ export declare class OpenPondExperimentsClient {
|
|
|
2464
5717
|
id: string;
|
|
2465
5718
|
contentHash: string;
|
|
2466
5719
|
};
|
|
2467
|
-
status: "
|
|
5720
|
+
status: "cancelled" | "failed" | "completed";
|
|
2468
5721
|
cases: {
|
|
2469
5722
|
identity: {
|
|
2470
5723
|
caseId: string;
|
|
2471
5724
|
seed: string;
|
|
2472
5725
|
fixtureId: string | null;
|
|
2473
5726
|
};
|
|
2474
|
-
status: "
|
|
5727
|
+
status: "cancelled" | "failed" | "completed" | "unavailable";
|
|
2475
5728
|
output: unknown;
|
|
2476
5729
|
error: {
|
|
2477
5730
|
code: string;
|
|
@@ -2560,6 +5813,102 @@ export declare class OpenPondExperimentsClient {
|
|
|
2560
5813
|
actionId: string;
|
|
2561
5814
|
};
|
|
2562
5815
|
environmentHash: string;
|
|
5816
|
+
externalDatasetBinding?: {
|
|
5817
|
+
schemaVersion: "openpond.profileExternalDatasetBinding.v1";
|
|
5818
|
+
dataset: {
|
|
5819
|
+
id: string;
|
|
5820
|
+
contentHash: string;
|
|
5821
|
+
revision: number;
|
|
5822
|
+
};
|
|
5823
|
+
packageHash: string;
|
|
5824
|
+
profileId: string;
|
|
5825
|
+
target: {
|
|
5826
|
+
kind: "profile";
|
|
5827
|
+
} | {
|
|
5828
|
+
kind: "workflow";
|
|
5829
|
+
workflowId: string;
|
|
5830
|
+
} | {
|
|
5831
|
+
kind: "skill";
|
|
5832
|
+
skillPath: string;
|
|
5833
|
+
} | {
|
|
5834
|
+
kind: "agent_action";
|
|
5835
|
+
actionId: string;
|
|
5836
|
+
};
|
|
5837
|
+
declaredProfileCatalogHash: string | null;
|
|
5838
|
+
protectedProfileClosureHash: string;
|
|
5839
|
+
split: "train" | "validation" | "test" | "frozen_eval";
|
|
5840
|
+
population: {
|
|
5841
|
+
taskId: string;
|
|
5842
|
+
seed: string;
|
|
5843
|
+
fixtureId: string | null;
|
|
5844
|
+
}[];
|
|
5845
|
+
evaluators: {
|
|
5846
|
+
release: {
|
|
5847
|
+
id: string;
|
|
5848
|
+
contentHash: string;
|
|
5849
|
+
revision: string | number;
|
|
5850
|
+
};
|
|
5851
|
+
feedbackKey: string;
|
|
5852
|
+
configurationHash: string;
|
|
5853
|
+
output: "boolean" | "score" | "category";
|
|
5854
|
+
categories: string[];
|
|
5855
|
+
}[];
|
|
5856
|
+
fieldMappingsHash: string;
|
|
5857
|
+
fieldMappings: {
|
|
5858
|
+
graderId: string;
|
|
5859
|
+
fields: {
|
|
5860
|
+
destination: "output" | "input" | "expectedOutput" | "evaluatorContext";
|
|
5861
|
+
path: string[];
|
|
5862
|
+
source: "output" | "input" | "expectedOutput" | "evaluatorContext";
|
|
5863
|
+
sourcePath: string[];
|
|
5864
|
+
}[];
|
|
5865
|
+
}[];
|
|
5866
|
+
environmentHash: string;
|
|
5867
|
+
privateDatasetClosureHash: string;
|
|
5868
|
+
criterion: {
|
|
5869
|
+
minimumPassRate: number;
|
|
5870
|
+
requireComplete: boolean;
|
|
5871
|
+
};
|
|
5872
|
+
contentHash: string;
|
|
5873
|
+
recordedOrigin?: {
|
|
5874
|
+
execution: {
|
|
5875
|
+
id: string;
|
|
5876
|
+
contentHash: string;
|
|
5877
|
+
};
|
|
5878
|
+
originalEvidence: {
|
|
5879
|
+
id: string;
|
|
5880
|
+
contentHash: string;
|
|
5881
|
+
};
|
|
5882
|
+
members: {
|
|
5883
|
+
caseId: string;
|
|
5884
|
+
seed: string;
|
|
5885
|
+
fixtureId: string;
|
|
5886
|
+
sourceId: string;
|
|
5887
|
+
snapshotHash: string;
|
|
5888
|
+
boundaryRevisionHash: string;
|
|
5889
|
+
}[];
|
|
5890
|
+
} | undefined;
|
|
5891
|
+
gradingSource?: {
|
|
5892
|
+
pass: {
|
|
5893
|
+
id: string;
|
|
5894
|
+
contentHash: string;
|
|
5895
|
+
};
|
|
5896
|
+
evidence: {
|
|
5897
|
+
id: string;
|
|
5898
|
+
contentHash: string;
|
|
5899
|
+
};
|
|
5900
|
+
} | undefined;
|
|
5901
|
+
} | undefined;
|
|
5902
|
+
sourceCandidate?: {
|
|
5903
|
+
candidateId: string;
|
|
5904
|
+
frozenRevision: number;
|
|
5905
|
+
frozenHash: string;
|
|
5906
|
+
harnessRelease: {
|
|
5907
|
+
id: string;
|
|
5908
|
+
contentHash: string;
|
|
5909
|
+
};
|
|
5910
|
+
protectedClosureHash: string;
|
|
5911
|
+
} | undefined;
|
|
2563
5912
|
};
|
|
2564
5913
|
model: {
|
|
2565
5914
|
modelId: string;
|
|
@@ -2624,6 +5973,14 @@ export declare class OpenPondExperimentsClient {
|
|
|
2624
5973
|
contentHash: string;
|
|
2625
5974
|
};
|
|
2626
5975
|
scoringPassId: string | null;
|
|
5976
|
+
humanProjection?: {
|
|
5977
|
+
originalManifest: {
|
|
5978
|
+
id: string;
|
|
5979
|
+
contentHash: string;
|
|
5980
|
+
};
|
|
5981
|
+
selectionHash: string;
|
|
5982
|
+
rulesHash: string;
|
|
5983
|
+
} | undefined;
|
|
2627
5984
|
} | undefined;
|
|
2628
5985
|
};
|
|
2629
5986
|
result: {
|
|
@@ -2632,14 +5989,14 @@ export declare class OpenPondExperimentsClient {
|
|
|
2632
5989
|
id: string;
|
|
2633
5990
|
contentHash: string;
|
|
2634
5991
|
};
|
|
2635
|
-
status: "
|
|
5992
|
+
status: "cancelled" | "failed" | "completed";
|
|
2636
5993
|
cases: {
|
|
2637
5994
|
identity: {
|
|
2638
5995
|
caseId: string;
|
|
2639
5996
|
seed: string;
|
|
2640
5997
|
fixtureId: string | null;
|
|
2641
5998
|
};
|
|
2642
|
-
status: "
|
|
5999
|
+
status: "cancelled" | "failed" | "completed" | "unavailable";
|
|
2643
6000
|
output: unknown;
|
|
2644
6001
|
error: {
|
|
2645
6002
|
code: string;
|
|
@@ -2727,6 +6084,102 @@ export declare class OpenPondExperimentsClient {
|
|
|
2727
6084
|
actionId: string;
|
|
2728
6085
|
};
|
|
2729
6086
|
environmentHash: string;
|
|
6087
|
+
externalDatasetBinding?: {
|
|
6088
|
+
schemaVersion: "openpond.profileExternalDatasetBinding.v1";
|
|
6089
|
+
dataset: {
|
|
6090
|
+
id: string;
|
|
6091
|
+
contentHash: string;
|
|
6092
|
+
revision: number;
|
|
6093
|
+
};
|
|
6094
|
+
packageHash: string;
|
|
6095
|
+
profileId: string;
|
|
6096
|
+
target: {
|
|
6097
|
+
kind: "profile";
|
|
6098
|
+
} | {
|
|
6099
|
+
kind: "workflow";
|
|
6100
|
+
workflowId: string;
|
|
6101
|
+
} | {
|
|
6102
|
+
kind: "skill";
|
|
6103
|
+
skillPath: string;
|
|
6104
|
+
} | {
|
|
6105
|
+
kind: "agent_action";
|
|
6106
|
+
actionId: string;
|
|
6107
|
+
};
|
|
6108
|
+
declaredProfileCatalogHash: string | null;
|
|
6109
|
+
protectedProfileClosureHash: string;
|
|
6110
|
+
split: "train" | "validation" | "test" | "frozen_eval";
|
|
6111
|
+
population: {
|
|
6112
|
+
taskId: string;
|
|
6113
|
+
seed: string;
|
|
6114
|
+
fixtureId: string | null;
|
|
6115
|
+
}[];
|
|
6116
|
+
evaluators: {
|
|
6117
|
+
release: {
|
|
6118
|
+
id: string;
|
|
6119
|
+
contentHash: string;
|
|
6120
|
+
revision: string | number;
|
|
6121
|
+
};
|
|
6122
|
+
feedbackKey: string;
|
|
6123
|
+
configurationHash: string;
|
|
6124
|
+
output: "boolean" | "score" | "category";
|
|
6125
|
+
categories: string[];
|
|
6126
|
+
}[];
|
|
6127
|
+
fieldMappingsHash: string;
|
|
6128
|
+
fieldMappings: {
|
|
6129
|
+
graderId: string;
|
|
6130
|
+
fields: {
|
|
6131
|
+
destination: "output" | "input" | "expectedOutput" | "evaluatorContext";
|
|
6132
|
+
path: string[];
|
|
6133
|
+
source: "output" | "input" | "expectedOutput" | "evaluatorContext";
|
|
6134
|
+
sourcePath: string[];
|
|
6135
|
+
}[];
|
|
6136
|
+
}[];
|
|
6137
|
+
environmentHash: string;
|
|
6138
|
+
privateDatasetClosureHash: string;
|
|
6139
|
+
criterion: {
|
|
6140
|
+
minimumPassRate: number;
|
|
6141
|
+
requireComplete: boolean;
|
|
6142
|
+
};
|
|
6143
|
+
contentHash: string;
|
|
6144
|
+
recordedOrigin?: {
|
|
6145
|
+
execution: {
|
|
6146
|
+
id: string;
|
|
6147
|
+
contentHash: string;
|
|
6148
|
+
};
|
|
6149
|
+
originalEvidence: {
|
|
6150
|
+
id: string;
|
|
6151
|
+
contentHash: string;
|
|
6152
|
+
};
|
|
6153
|
+
members: {
|
|
6154
|
+
caseId: string;
|
|
6155
|
+
seed: string;
|
|
6156
|
+
fixtureId: string;
|
|
6157
|
+
sourceId: string;
|
|
6158
|
+
snapshotHash: string;
|
|
6159
|
+
boundaryRevisionHash: string;
|
|
6160
|
+
}[];
|
|
6161
|
+
} | undefined;
|
|
6162
|
+
gradingSource?: {
|
|
6163
|
+
pass: {
|
|
6164
|
+
id: string;
|
|
6165
|
+
contentHash: string;
|
|
6166
|
+
};
|
|
6167
|
+
evidence: {
|
|
6168
|
+
id: string;
|
|
6169
|
+
contentHash: string;
|
|
6170
|
+
};
|
|
6171
|
+
} | undefined;
|
|
6172
|
+
} | undefined;
|
|
6173
|
+
sourceCandidate?: {
|
|
6174
|
+
candidateId: string;
|
|
6175
|
+
frozenRevision: number;
|
|
6176
|
+
frozenHash: string;
|
|
6177
|
+
harnessRelease: {
|
|
6178
|
+
id: string;
|
|
6179
|
+
contentHash: string;
|
|
6180
|
+
};
|
|
6181
|
+
protectedClosureHash: string;
|
|
6182
|
+
} | undefined;
|
|
2730
6183
|
};
|
|
2731
6184
|
model: {
|
|
2732
6185
|
modelId: string;
|
|
@@ -2791,6 +6244,14 @@ export declare class OpenPondExperimentsClient {
|
|
|
2791
6244
|
contentHash: string;
|
|
2792
6245
|
};
|
|
2793
6246
|
scoringPassId: string | null;
|
|
6247
|
+
humanProjection?: {
|
|
6248
|
+
originalManifest: {
|
|
6249
|
+
id: string;
|
|
6250
|
+
contentHash: string;
|
|
6251
|
+
};
|
|
6252
|
+
selectionHash: string;
|
|
6253
|
+
rulesHash: string;
|
|
6254
|
+
} | undefined;
|
|
2794
6255
|
} | undefined;
|
|
2795
6256
|
};
|
|
2796
6257
|
result: {
|
|
@@ -2799,14 +6260,14 @@ export declare class OpenPondExperimentsClient {
|
|
|
2799
6260
|
id: string;
|
|
2800
6261
|
contentHash: string;
|
|
2801
6262
|
};
|
|
2802
|
-
status: "
|
|
6263
|
+
status: "cancelled" | "failed" | "completed";
|
|
2803
6264
|
cases: {
|
|
2804
6265
|
identity: {
|
|
2805
6266
|
caseId: string;
|
|
2806
6267
|
seed: string;
|
|
2807
6268
|
fixtureId: string | null;
|
|
2808
6269
|
};
|
|
2809
|
-
status: "
|
|
6270
|
+
status: "cancelled" | "failed" | "completed" | "unavailable";
|
|
2810
6271
|
output: unknown;
|
|
2811
6272
|
error: {
|
|
2812
6273
|
code: string;
|