@kortyx/agent 0.35.0 → 0.37.0
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/CHANGELOG.md +45 -0
- package/dist/evals.d.ts +113 -12
- package/dist/evals.js +37 -1
- package/dist/evals.js.map +1 -1
- package/dist/index.d.ts +17 -5
- package/dist/index.js +1351 -429
- package/dist/index.js.map +1 -1
- package/dist/{studio-judge-0yb-Cf2q.d.ts → studio-judge-Brk9R5oB.d.ts} +373 -40
- package/package.json +10 -8
|
@@ -1,4 +1,5 @@
|
|
|
1
1
|
import { z } from 'zod';
|
|
2
|
+
import * as _kortyx_prompts from '@kortyx/prompts';
|
|
2
3
|
import { GetProviderFn, ProviderModelRef } from '@kortyx/providers';
|
|
3
4
|
import { KortyxError, FailureDescriptor } from '@kortyx/core/errors';
|
|
4
5
|
import { FinalizedChatMessage, StreamChunk } from '@kortyx/stream';
|
|
@@ -161,6 +162,7 @@ interface AgentProjectTopologyOptions {
|
|
|
161
162
|
service: AgentTopologySnapshot["service"];
|
|
162
163
|
}
|
|
163
164
|
interface CreateAgentArgs {
|
|
165
|
+
prompts?: _kortyx_prompts.PromptManager | undefined;
|
|
164
166
|
limits?: ExecutionLimits | undefined;
|
|
165
167
|
getProvider?: GetProviderFn | undefined;
|
|
166
168
|
workflows?: WorkflowDefinition[];
|
|
@@ -171,6 +173,12 @@ interface CreateAgentArgs {
|
|
|
171
173
|
telemetry?: ExecutionRuntimeConfig["telemetry"];
|
|
172
174
|
}
|
|
173
175
|
interface Agent extends ReturnType<typeof createInterruptDiscovery> {
|
|
176
|
+
describePromptContracts?: () => {
|
|
177
|
+
id: string;
|
|
178
|
+
format: "system-user" | "chat";
|
|
179
|
+
variablesSchema: Record<string, unknown>;
|
|
180
|
+
configSchema: Record<string, unknown>;
|
|
181
|
+
}[];
|
|
174
182
|
execute<W extends ExecutableWorkflow>(args: ExecuteOptions<W>): Promise<ExecutionResult<z.output<W["outputSchema"]>>>;
|
|
175
183
|
execute(args: {
|
|
176
184
|
limits?: ExecutionLimits;
|
|
@@ -316,6 +324,7 @@ type EvalInterrupt = {
|
|
|
316
324
|
};
|
|
317
325
|
/** Public output and emitted execution evidence; private runtime state is excluded. */
|
|
318
326
|
type EvalObservation = {
|
|
327
|
+
promptUsage?: _kortyx_prompts.PromptUsageReceipt[];
|
|
319
328
|
type: "answer" | "interrupt" | "error" | "cancelled";
|
|
320
329
|
text: string;
|
|
321
330
|
structured: readonly EvalJson[];
|
|
@@ -475,7 +484,7 @@ type EvalDefaults = {
|
|
|
475
484
|
cleanupTimeoutMs?: number;
|
|
476
485
|
};
|
|
477
486
|
type CreateEvalsOptions<Params = EvalJson | undefined, Prepared = undefined> = {
|
|
478
|
-
agent: Pick<Agent, "streamChat">;
|
|
487
|
+
agent: Pick<Agent, "streamChat" | "describePromptContracts">;
|
|
479
488
|
suites: readonly EvalSuite[];
|
|
480
489
|
paramsSchema?: z.ZodType<Params>;
|
|
481
490
|
setup?: (input: EvalSetupContext<Params>) => Prepared | Promise<Prepared>;
|
|
@@ -533,6 +542,7 @@ type EvalProgress = {
|
|
|
533
542
|
result: EvalCaseResult;
|
|
534
543
|
};
|
|
535
544
|
type EvalRunOptions = {
|
|
545
|
+
promptSnapshot?: _kortyx_prompts.PromptSnapshot;
|
|
536
546
|
/** Studio captures execution without invoking the code judge. Local runs default to app. */
|
|
537
547
|
grading?: "app" | "studio";
|
|
538
548
|
/** Selected identity pinned by the Studio server when enqueuing a run. */
|
|
@@ -659,6 +669,14 @@ declare const EvalVerdictSchema: z.ZodObject<{
|
|
|
659
669
|
evidence: z.ZodArray<z.ZodString>;
|
|
660
670
|
}, z.core.$strict>;
|
|
661
671
|
declare const EvalObservationSchema: z.ZodObject<{
|
|
672
|
+
promptUsage: z.ZodOptional<z.ZodArray<z.ZodObject<{
|
|
673
|
+
id: z.ZodString;
|
|
674
|
+
version: z.ZodNumber;
|
|
675
|
+
hash: z.ZodString;
|
|
676
|
+
environment: z.ZodString;
|
|
677
|
+
snapshotRevision: z.ZodString;
|
|
678
|
+
model: z.ZodOptional<z.ZodString>;
|
|
679
|
+
}, z.core.$strict>>>;
|
|
662
680
|
type: z.ZodEnum<{
|
|
663
681
|
interrupt: "interrupt";
|
|
664
682
|
error: "error";
|
|
@@ -722,6 +740,14 @@ declare const EvalStepResultSchema: z.ZodObject<{
|
|
|
722
740
|
version: z.ZodLiteral<"compact-v1">;
|
|
723
741
|
history: z.ZodBoolean;
|
|
724
742
|
observation: z.ZodObject<{
|
|
743
|
+
promptUsage: z.ZodOptional<z.ZodArray<z.ZodObject<{
|
|
744
|
+
id: z.ZodString;
|
|
745
|
+
version: z.ZodNumber;
|
|
746
|
+
hash: z.ZodString;
|
|
747
|
+
environment: z.ZodString;
|
|
748
|
+
snapshotRevision: z.ZodString;
|
|
749
|
+
model: z.ZodOptional<z.ZodString>;
|
|
750
|
+
}, z.core.$strict>>>;
|
|
725
751
|
type: z.ZodEnum<{
|
|
726
752
|
interrupt: "interrupt";
|
|
727
753
|
error: "error";
|
|
@@ -788,6 +814,14 @@ declare const EvalStepResultSchema: z.ZodObject<{
|
|
|
788
814
|
reference: z.ZodOptional<z.ZodJSONSchema>;
|
|
789
815
|
}, z.core.$strict>;
|
|
790
816
|
observation: z.ZodObject<{
|
|
817
|
+
promptUsage: z.ZodOptional<z.ZodArray<z.ZodObject<{
|
|
818
|
+
id: z.ZodString;
|
|
819
|
+
version: z.ZodNumber;
|
|
820
|
+
hash: z.ZodString;
|
|
821
|
+
environment: z.ZodString;
|
|
822
|
+
snapshotRevision: z.ZodString;
|
|
823
|
+
model: z.ZodOptional<z.ZodString>;
|
|
824
|
+
}, z.core.$strict>>>;
|
|
791
825
|
type: z.ZodEnum<{
|
|
792
826
|
interrupt: "interrupt";
|
|
793
827
|
error: "error";
|
|
@@ -823,8 +857,8 @@ declare const EvalStepResultSchema: z.ZodObject<{
|
|
|
823
857
|
error: "error";
|
|
824
858
|
cancelled: "cancelled";
|
|
825
859
|
failed: "failed";
|
|
826
|
-
passed: "passed";
|
|
827
860
|
ungraded: "ungraded";
|
|
861
|
+
passed: "passed";
|
|
828
862
|
}>;
|
|
829
863
|
reason: z.ZodOptional<z.ZodString>;
|
|
830
864
|
judgeCalls: z.ZodOptional<z.ZodNumber>;
|
|
@@ -876,8 +910,8 @@ declare const EvalCaseResultSchema: z.ZodObject<{
|
|
|
876
910
|
error: "error";
|
|
877
911
|
cancelled: "cancelled";
|
|
878
912
|
failed: "failed";
|
|
879
|
-
passed: "passed";
|
|
880
913
|
ungraded: "ungraded";
|
|
914
|
+
passed: "passed";
|
|
881
915
|
}>;
|
|
882
916
|
durationMs: z.ZodNumber;
|
|
883
917
|
steps: z.ZodArray<z.ZodObject<{
|
|
@@ -885,6 +919,14 @@ declare const EvalCaseResultSchema: z.ZodObject<{
|
|
|
885
919
|
version: z.ZodLiteral<"compact-v1">;
|
|
886
920
|
history: z.ZodBoolean;
|
|
887
921
|
observation: z.ZodObject<{
|
|
922
|
+
promptUsage: z.ZodOptional<z.ZodArray<z.ZodObject<{
|
|
923
|
+
id: z.ZodString;
|
|
924
|
+
version: z.ZodNumber;
|
|
925
|
+
hash: z.ZodString;
|
|
926
|
+
environment: z.ZodString;
|
|
927
|
+
snapshotRevision: z.ZodString;
|
|
928
|
+
model: z.ZodOptional<z.ZodString>;
|
|
929
|
+
}, z.core.$strict>>>;
|
|
888
930
|
type: z.ZodEnum<{
|
|
889
931
|
interrupt: "interrupt";
|
|
890
932
|
error: "error";
|
|
@@ -951,6 +993,14 @@ declare const EvalCaseResultSchema: z.ZodObject<{
|
|
|
951
993
|
reference: z.ZodOptional<z.ZodJSONSchema>;
|
|
952
994
|
}, z.core.$strict>;
|
|
953
995
|
observation: z.ZodObject<{
|
|
996
|
+
promptUsage: z.ZodOptional<z.ZodArray<z.ZodObject<{
|
|
997
|
+
id: z.ZodString;
|
|
998
|
+
version: z.ZodNumber;
|
|
999
|
+
hash: z.ZodString;
|
|
1000
|
+
environment: z.ZodString;
|
|
1001
|
+
snapshotRevision: z.ZodString;
|
|
1002
|
+
model: z.ZodOptional<z.ZodString>;
|
|
1003
|
+
}, z.core.$strict>>>;
|
|
954
1004
|
type: z.ZodEnum<{
|
|
955
1005
|
interrupt: "interrupt";
|
|
956
1006
|
error: "error";
|
|
@@ -986,8 +1036,8 @@ declare const EvalCaseResultSchema: z.ZodObject<{
|
|
|
986
1036
|
error: "error";
|
|
987
1037
|
cancelled: "cancelled";
|
|
988
1038
|
failed: "failed";
|
|
989
|
-
passed: "passed";
|
|
990
1039
|
ungraded: "ungraded";
|
|
1040
|
+
passed: "passed";
|
|
991
1041
|
}>;
|
|
992
1042
|
reason: z.ZodOptional<z.ZodString>;
|
|
993
1043
|
judgeCalls: z.ZodOptional<z.ZodNumber>;
|
|
@@ -1045,8 +1095,8 @@ declare const EvalJudgeIdentitySchema: z.ZodObject<{
|
|
|
1045
1095
|
id: z.ZodString;
|
|
1046
1096
|
version: z.ZodString;
|
|
1047
1097
|
location: z.ZodOptional<z.ZodEnum<{
|
|
1048
|
-
app: "app";
|
|
1049
1098
|
studio: "studio";
|
|
1099
|
+
app: "app";
|
|
1050
1100
|
}>>;
|
|
1051
1101
|
}, z.core.$strict>;
|
|
1052
1102
|
declare const StudioEvalJudgeRequestSchema: z.ZodObject<{
|
|
@@ -1055,8 +1105,8 @@ declare const StudioEvalJudgeRequestSchema: z.ZodObject<{
|
|
|
1055
1105
|
id: z.ZodString;
|
|
1056
1106
|
version: z.ZodString;
|
|
1057
1107
|
location: z.ZodOptional<z.ZodEnum<{
|
|
1058
|
-
app: "app";
|
|
1059
1108
|
studio: "studio";
|
|
1109
|
+
app: "app";
|
|
1060
1110
|
}>>;
|
|
1061
1111
|
}, z.core.$strict>;
|
|
1062
1112
|
criterion: z.ZodObject<{
|
|
@@ -1080,6 +1130,14 @@ declare const StudioEvalJudgeRequestSchema: z.ZodObject<{
|
|
|
1080
1130
|
}, z.core.$strict>], "type">;
|
|
1081
1131
|
}, z.core.$strict>]>;
|
|
1082
1132
|
observation: z.ZodObject<{
|
|
1133
|
+
promptUsage: z.ZodOptional<z.ZodArray<z.ZodObject<{
|
|
1134
|
+
id: z.ZodString;
|
|
1135
|
+
version: z.ZodNumber;
|
|
1136
|
+
hash: z.ZodString;
|
|
1137
|
+
environment: z.ZodString;
|
|
1138
|
+
snapshotRevision: z.ZodString;
|
|
1139
|
+
model: z.ZodOptional<z.ZodString>;
|
|
1140
|
+
}, z.core.$strict>>>;
|
|
1083
1141
|
type: z.ZodEnum<{
|
|
1084
1142
|
interrupt: "interrupt";
|
|
1085
1143
|
error: "error";
|
|
@@ -1116,6 +1174,14 @@ declare const StudioEvalJudgeRequestSchema: z.ZodObject<{
|
|
|
1116
1174
|
version: z.ZodLiteral<"compact-v1">;
|
|
1117
1175
|
history: z.ZodBoolean;
|
|
1118
1176
|
observation: z.ZodObject<{
|
|
1177
|
+
promptUsage: z.ZodOptional<z.ZodArray<z.ZodObject<{
|
|
1178
|
+
id: z.ZodString;
|
|
1179
|
+
version: z.ZodNumber;
|
|
1180
|
+
hash: z.ZodString;
|
|
1181
|
+
environment: z.ZodString;
|
|
1182
|
+
snapshotRevision: z.ZodString;
|
|
1183
|
+
model: z.ZodOptional<z.ZodString>;
|
|
1184
|
+
}, z.core.$strict>>>;
|
|
1119
1185
|
type: z.ZodEnum<{
|
|
1120
1186
|
interrupt: "interrupt";
|
|
1121
1187
|
error: "error";
|
|
@@ -1182,6 +1248,14 @@ declare const StudioEvalJudgeRequestSchema: z.ZodObject<{
|
|
|
1182
1248
|
reference: z.ZodOptional<z.ZodJSONSchema>;
|
|
1183
1249
|
}, z.core.$strict>;
|
|
1184
1250
|
observation: z.ZodObject<{
|
|
1251
|
+
promptUsage: z.ZodOptional<z.ZodArray<z.ZodObject<{
|
|
1252
|
+
id: z.ZodString;
|
|
1253
|
+
version: z.ZodNumber;
|
|
1254
|
+
hash: z.ZodString;
|
|
1255
|
+
environment: z.ZodString;
|
|
1256
|
+
snapshotRevision: z.ZodString;
|
|
1257
|
+
model: z.ZodOptional<z.ZodString>;
|
|
1258
|
+
}, z.core.$strict>>>;
|
|
1185
1259
|
type: z.ZodEnum<{
|
|
1186
1260
|
interrupt: "interrupt";
|
|
1187
1261
|
error: "error";
|
|
@@ -1217,8 +1291,8 @@ declare const StudioEvalJudgeRequestSchema: z.ZodObject<{
|
|
|
1217
1291
|
error: "error";
|
|
1218
1292
|
cancelled: "cancelled";
|
|
1219
1293
|
failed: "failed";
|
|
1220
|
-
passed: "passed";
|
|
1221
1294
|
ungraded: "ungraded";
|
|
1295
|
+
passed: "passed";
|
|
1222
1296
|
}>;
|
|
1223
1297
|
reason: z.ZodOptional<z.ZodString>;
|
|
1224
1298
|
judgeCalls: z.ZodOptional<z.ZodNumber>;
|
|
@@ -1263,8 +1337,8 @@ declare const StudioEvalJudgeResponseSchema: z.ZodObject<{
|
|
|
1263
1337
|
id: z.ZodString;
|
|
1264
1338
|
version: z.ZodString;
|
|
1265
1339
|
location: z.ZodOptional<z.ZodEnum<{
|
|
1266
|
-
app: "app";
|
|
1267
1340
|
studio: "studio";
|
|
1341
|
+
app: "app";
|
|
1268
1342
|
}>>;
|
|
1269
1343
|
}, z.core.$strict>;
|
|
1270
1344
|
verdict: z.ZodObject<{
|
|
@@ -1397,8 +1471,8 @@ declare const EvalRunResultSchema: z.ZodObject<{
|
|
|
1397
1471
|
id: z.ZodString;
|
|
1398
1472
|
version: z.ZodString;
|
|
1399
1473
|
location: z.ZodOptional<z.ZodEnum<{
|
|
1400
|
-
app: "app";
|
|
1401
1474
|
studio: "studio";
|
|
1475
|
+
app: "app";
|
|
1402
1476
|
}>>;
|
|
1403
1477
|
}, z.core.$strict>>;
|
|
1404
1478
|
startedAt: z.ZodISODateTime;
|
|
@@ -1407,8 +1481,8 @@ declare const EvalRunResultSchema: z.ZodObject<{
|
|
|
1407
1481
|
error: "error";
|
|
1408
1482
|
cancelled: "cancelled";
|
|
1409
1483
|
failed: "failed";
|
|
1410
|
-
passed: "passed";
|
|
1411
1484
|
ungraded: "ungraded";
|
|
1485
|
+
passed: "passed";
|
|
1412
1486
|
}>;
|
|
1413
1487
|
counts: z.ZodObject<{
|
|
1414
1488
|
passed: z.ZodNumber;
|
|
@@ -1430,8 +1504,8 @@ declare const EvalRunResultSchema: z.ZodObject<{
|
|
|
1430
1504
|
error: "error";
|
|
1431
1505
|
cancelled: "cancelled";
|
|
1432
1506
|
failed: "failed";
|
|
1433
|
-
passed: "passed";
|
|
1434
1507
|
ungraded: "ungraded";
|
|
1508
|
+
passed: "passed";
|
|
1435
1509
|
}>;
|
|
1436
1510
|
durationMs: z.ZodNumber;
|
|
1437
1511
|
steps: z.ZodArray<z.ZodObject<{
|
|
@@ -1439,6 +1513,14 @@ declare const EvalRunResultSchema: z.ZodObject<{
|
|
|
1439
1513
|
version: z.ZodLiteral<"compact-v1">;
|
|
1440
1514
|
history: z.ZodBoolean;
|
|
1441
1515
|
observation: z.ZodObject<{
|
|
1516
|
+
promptUsage: z.ZodOptional<z.ZodArray<z.ZodObject<{
|
|
1517
|
+
id: z.ZodString;
|
|
1518
|
+
version: z.ZodNumber;
|
|
1519
|
+
hash: z.ZodString;
|
|
1520
|
+
environment: z.ZodString;
|
|
1521
|
+
snapshotRevision: z.ZodString;
|
|
1522
|
+
model: z.ZodOptional<z.ZodString>;
|
|
1523
|
+
}, z.core.$strict>>>;
|
|
1442
1524
|
type: z.ZodEnum<{
|
|
1443
1525
|
interrupt: "interrupt";
|
|
1444
1526
|
error: "error";
|
|
@@ -1505,6 +1587,14 @@ declare const EvalRunResultSchema: z.ZodObject<{
|
|
|
1505
1587
|
reference: z.ZodOptional<z.ZodJSONSchema>;
|
|
1506
1588
|
}, z.core.$strict>;
|
|
1507
1589
|
observation: z.ZodObject<{
|
|
1590
|
+
promptUsage: z.ZodOptional<z.ZodArray<z.ZodObject<{
|
|
1591
|
+
id: z.ZodString;
|
|
1592
|
+
version: z.ZodNumber;
|
|
1593
|
+
hash: z.ZodString;
|
|
1594
|
+
environment: z.ZodString;
|
|
1595
|
+
snapshotRevision: z.ZodString;
|
|
1596
|
+
model: z.ZodOptional<z.ZodString>;
|
|
1597
|
+
}, z.core.$strict>>>;
|
|
1508
1598
|
type: z.ZodEnum<{
|
|
1509
1599
|
interrupt: "interrupt";
|
|
1510
1600
|
error: "error";
|
|
@@ -1540,8 +1630,8 @@ declare const EvalRunResultSchema: z.ZodObject<{
|
|
|
1540
1630
|
error: "error";
|
|
1541
1631
|
cancelled: "cancelled";
|
|
1542
1632
|
failed: "failed";
|
|
1543
|
-
passed: "passed";
|
|
1544
1633
|
ungraded: "ungraded";
|
|
1634
|
+
passed: "passed";
|
|
1545
1635
|
}>;
|
|
1546
1636
|
reason: z.ZodOptional<z.ZodString>;
|
|
1547
1637
|
judgeCalls: z.ZodOptional<z.ZodNumber>;
|
|
@@ -1613,6 +1703,15 @@ declare const EvalRunResultSchema: z.ZodObject<{
|
|
|
1613
1703
|
declare const EvalManifestSchema: z.ZodObject<{
|
|
1614
1704
|
schemaVersion: z.ZodLiteral<1>;
|
|
1615
1705
|
studioJudging: z.ZodOptional<z.ZodLiteral<true>>;
|
|
1706
|
+
promptContracts: z.ZodOptional<z.ZodArray<z.ZodObject<{
|
|
1707
|
+
id: z.ZodString;
|
|
1708
|
+
format: z.ZodEnum<{
|
|
1709
|
+
"system-user": "system-user";
|
|
1710
|
+
chat: "chat";
|
|
1711
|
+
}>;
|
|
1712
|
+
variablesSchema: z.ZodRecord<z.ZodString, z.ZodUnknown>;
|
|
1713
|
+
configSchema: z.ZodRecord<z.ZodString, z.ZodUnknown>;
|
|
1714
|
+
}, z.core.$strict>>>;
|
|
1616
1715
|
attemptScheduling: z.ZodOptional<z.ZodLiteral<true>>;
|
|
1617
1716
|
suites: z.ZodArray<z.ZodObject<{
|
|
1618
1717
|
evidence: z.ZodOptional<z.ZodObject<{
|
|
@@ -1714,8 +1813,8 @@ declare const EvalManifestSchema: z.ZodObject<{
|
|
|
1714
1813
|
id: z.ZodString;
|
|
1715
1814
|
version: z.ZodString;
|
|
1716
1815
|
location: z.ZodOptional<z.ZodEnum<{
|
|
1717
|
-
app: "app";
|
|
1718
1816
|
studio: "studio";
|
|
1817
|
+
app: "app";
|
|
1719
1818
|
}>>;
|
|
1720
1819
|
}, z.core.$strict>>;
|
|
1721
1820
|
}, z.core.$strict>;
|
|
@@ -1763,6 +1862,14 @@ declare const EvalProgressSchema: z.ZodDiscriminatedUnion<[z.ZodObject<{
|
|
|
1763
1862
|
version: z.ZodLiteral<"compact-v1">;
|
|
1764
1863
|
history: z.ZodBoolean;
|
|
1765
1864
|
observation: z.ZodObject<{
|
|
1865
|
+
promptUsage: z.ZodOptional<z.ZodArray<z.ZodObject<{
|
|
1866
|
+
id: z.ZodString;
|
|
1867
|
+
version: z.ZodNumber;
|
|
1868
|
+
hash: z.ZodString;
|
|
1869
|
+
environment: z.ZodString;
|
|
1870
|
+
snapshotRevision: z.ZodString;
|
|
1871
|
+
model: z.ZodOptional<z.ZodString>;
|
|
1872
|
+
}, z.core.$strict>>>;
|
|
1766
1873
|
type: z.ZodEnum<{
|
|
1767
1874
|
interrupt: "interrupt";
|
|
1768
1875
|
error: "error";
|
|
@@ -1829,6 +1936,14 @@ declare const EvalProgressSchema: z.ZodDiscriminatedUnion<[z.ZodObject<{
|
|
|
1829
1936
|
reference: z.ZodOptional<z.ZodJSONSchema>;
|
|
1830
1937
|
}, z.core.$strict>;
|
|
1831
1938
|
observation: z.ZodObject<{
|
|
1939
|
+
promptUsage: z.ZodOptional<z.ZodArray<z.ZodObject<{
|
|
1940
|
+
id: z.ZodString;
|
|
1941
|
+
version: z.ZodNumber;
|
|
1942
|
+
hash: z.ZodString;
|
|
1943
|
+
environment: z.ZodString;
|
|
1944
|
+
snapshotRevision: z.ZodString;
|
|
1945
|
+
model: z.ZodOptional<z.ZodString>;
|
|
1946
|
+
}, z.core.$strict>>>;
|
|
1832
1947
|
type: z.ZodEnum<{
|
|
1833
1948
|
interrupt: "interrupt";
|
|
1834
1949
|
error: "error";
|
|
@@ -1864,8 +1979,8 @@ declare const EvalProgressSchema: z.ZodDiscriminatedUnion<[z.ZodObject<{
|
|
|
1864
1979
|
error: "error";
|
|
1865
1980
|
cancelled: "cancelled";
|
|
1866
1981
|
failed: "failed";
|
|
1867
|
-
passed: "passed";
|
|
1868
1982
|
ungraded: "ungraded";
|
|
1983
|
+
passed: "passed";
|
|
1869
1984
|
}>;
|
|
1870
1985
|
reason: z.ZodOptional<z.ZodString>;
|
|
1871
1986
|
judgeCalls: z.ZodOptional<z.ZodNumber>;
|
|
@@ -1919,8 +2034,8 @@ declare const EvalProgressSchema: z.ZodDiscriminatedUnion<[z.ZodObject<{
|
|
|
1919
2034
|
error: "error";
|
|
1920
2035
|
cancelled: "cancelled";
|
|
1921
2036
|
failed: "failed";
|
|
1922
|
-
passed: "passed";
|
|
1923
2037
|
ungraded: "ungraded";
|
|
2038
|
+
passed: "passed";
|
|
1924
2039
|
}>;
|
|
1925
2040
|
durationMs: z.ZodNumber;
|
|
1926
2041
|
steps: z.ZodArray<z.ZodObject<{
|
|
@@ -1928,6 +2043,14 @@ declare const EvalProgressSchema: z.ZodDiscriminatedUnion<[z.ZodObject<{
|
|
|
1928
2043
|
version: z.ZodLiteral<"compact-v1">;
|
|
1929
2044
|
history: z.ZodBoolean;
|
|
1930
2045
|
observation: z.ZodObject<{
|
|
2046
|
+
promptUsage: z.ZodOptional<z.ZodArray<z.ZodObject<{
|
|
2047
|
+
id: z.ZodString;
|
|
2048
|
+
version: z.ZodNumber;
|
|
2049
|
+
hash: z.ZodString;
|
|
2050
|
+
environment: z.ZodString;
|
|
2051
|
+
snapshotRevision: z.ZodString;
|
|
2052
|
+
model: z.ZodOptional<z.ZodString>;
|
|
2053
|
+
}, z.core.$strict>>>;
|
|
1931
2054
|
type: z.ZodEnum<{
|
|
1932
2055
|
interrupt: "interrupt";
|
|
1933
2056
|
error: "error";
|
|
@@ -1994,6 +2117,14 @@ declare const EvalProgressSchema: z.ZodDiscriminatedUnion<[z.ZodObject<{
|
|
|
1994
2117
|
reference: z.ZodOptional<z.ZodJSONSchema>;
|
|
1995
2118
|
}, z.core.$strict>;
|
|
1996
2119
|
observation: z.ZodObject<{
|
|
2120
|
+
promptUsage: z.ZodOptional<z.ZodArray<z.ZodObject<{
|
|
2121
|
+
id: z.ZodString;
|
|
2122
|
+
version: z.ZodNumber;
|
|
2123
|
+
hash: z.ZodString;
|
|
2124
|
+
environment: z.ZodString;
|
|
2125
|
+
snapshotRevision: z.ZodString;
|
|
2126
|
+
model: z.ZodOptional<z.ZodString>;
|
|
2127
|
+
}, z.core.$strict>>>;
|
|
1997
2128
|
type: z.ZodEnum<{
|
|
1998
2129
|
interrupt: "interrupt";
|
|
1999
2130
|
error: "error";
|
|
@@ -2029,8 +2160,8 @@ declare const EvalProgressSchema: z.ZodDiscriminatedUnion<[z.ZodObject<{
|
|
|
2029
2160
|
error: "error";
|
|
2030
2161
|
cancelled: "cancelled";
|
|
2031
2162
|
failed: "failed";
|
|
2032
|
-
passed: "passed";
|
|
2033
2163
|
ungraded: "ungraded";
|
|
2164
|
+
passed: "passed";
|
|
2034
2165
|
}>;
|
|
2035
2166
|
reason: z.ZodOptional<z.ZodString>;
|
|
2036
2167
|
judgeCalls: z.ZodOptional<z.ZodNumber>;
|
|
@@ -2086,18 +2217,61 @@ declare const EvalProgressSchema: z.ZodDiscriminatedUnion<[z.ZodObject<{
|
|
|
2086
2217
|
}, z.core.$strict>;
|
|
2087
2218
|
}, z.core.$strict>], "type">;
|
|
2088
2219
|
declare const EvalRemoteRunRequestSchema: z.ZodObject<{
|
|
2220
|
+
promptSnapshot: z.ZodOptional<z.ZodObject<{
|
|
2221
|
+
schemaVersion: z.ZodLiteral<1>;
|
|
2222
|
+
environment: z.ZodString;
|
|
2223
|
+
tag: z.ZodOptional<z.ZodString>;
|
|
2224
|
+
revision: z.ZodString;
|
|
2225
|
+
versions: z.ZodRecord<z.ZodString, z.ZodObject<{
|
|
2226
|
+
id: z.ZodString;
|
|
2227
|
+
version: z.ZodNumber;
|
|
2228
|
+
hash: z.ZodString;
|
|
2229
|
+
content: z.ZodObject<{
|
|
2230
|
+
format: z.ZodEnum<{
|
|
2231
|
+
"system-user": "system-user";
|
|
2232
|
+
chat: "chat";
|
|
2233
|
+
}>;
|
|
2234
|
+
messages: z.ZodArray<z.ZodObject<{
|
|
2235
|
+
role: z.ZodEnum<{
|
|
2236
|
+
system: "system";
|
|
2237
|
+
user: "user";
|
|
2238
|
+
assistant: "assistant";
|
|
2239
|
+
}>;
|
|
2240
|
+
content: z.ZodString;
|
|
2241
|
+
}, z.core.$strict>>;
|
|
2242
|
+
variablesSchema: z.ZodRecord<z.ZodString, z.ZodJSONSchema>;
|
|
2243
|
+
configSchema: z.ZodRecord<z.ZodString, z.ZodJSONSchema>;
|
|
2244
|
+
config: z.ZodRecord<z.ZodString, z.ZodJSONSchema>;
|
|
2245
|
+
dependencies: z.ZodDefault<z.ZodArray<z.ZodObject<{
|
|
2246
|
+
id: z.ZodString;
|
|
2247
|
+
version: z.ZodNumber;
|
|
2248
|
+
hash: z.ZodString;
|
|
2249
|
+
}, z.core.$strict>>>;
|
|
2250
|
+
}, z.core.$strict>;
|
|
2251
|
+
}, z.core.$strip>>;
|
|
2252
|
+
source: z.ZodEnum<{
|
|
2253
|
+
studio: "studio";
|
|
2254
|
+
local: "local";
|
|
2255
|
+
cache: "cache";
|
|
2256
|
+
fallback: "fallback";
|
|
2257
|
+
eval: "eval";
|
|
2258
|
+
}>;
|
|
2259
|
+
resolvedAt: z.ZodISODateTime;
|
|
2260
|
+
requestedIds: z.ZodOptional<z.ZodArray<z.ZodString>>;
|
|
2261
|
+
}, z.core.$strict>>;
|
|
2262
|
+
promptGroupName: z.ZodOptional<z.ZodString>;
|
|
2089
2263
|
suiteId: z.ZodString;
|
|
2090
2264
|
suiteRevision: z.ZodString;
|
|
2091
2265
|
grading: z.ZodOptional<z.ZodEnum<{
|
|
2092
|
-
app: "app";
|
|
2093
2266
|
studio: "studio";
|
|
2267
|
+
app: "app";
|
|
2094
2268
|
}>>;
|
|
2095
2269
|
judge: z.ZodOptional<z.ZodObject<{
|
|
2096
2270
|
id: z.ZodString;
|
|
2097
2271
|
version: z.ZodString;
|
|
2098
2272
|
location: z.ZodOptional<z.ZodEnum<{
|
|
2099
|
-
app: "app";
|
|
2100
2273
|
studio: "studio";
|
|
2274
|
+
app: "app";
|
|
2101
2275
|
}>>;
|
|
2102
2276
|
}, z.core.$strict>>;
|
|
2103
2277
|
caseIds: z.ZodOptional<z.ZodArray<z.ZodString>>;
|
|
@@ -2115,10 +2289,20 @@ declare const StudioEvalStartRequestSchema: z.ZodObject<{
|
|
|
2115
2289
|
caseIds: z.ZodOptional<z.ZodArray<z.ZodString>>;
|
|
2116
2290
|
repetitions: z.ZodDefault<z.ZodNumber>;
|
|
2117
2291
|
concurrency: z.ZodDefault<z.ZodNumber>;
|
|
2292
|
+
promptSelection: z.ZodOptional<z.ZodDiscriminatedUnion<[z.ZodObject<{
|
|
2293
|
+
type: z.ZodLiteral<"live">;
|
|
2294
|
+
}, z.core.$strict>, z.ZodObject<{
|
|
2295
|
+
type: z.ZodLiteral<"single">;
|
|
2296
|
+
id: z.ZodString;
|
|
2297
|
+
version: z.ZodNumber;
|
|
2298
|
+
}, z.core.$strict>, z.ZodObject<{
|
|
2299
|
+
type: z.ZodLiteral<"group">;
|
|
2300
|
+
groupId: z.ZodUUID;
|
|
2301
|
+
}, z.core.$strict>], "type">>;
|
|
2118
2302
|
targetId: z.ZodString;
|
|
2119
2303
|
judge: z.ZodDefault<z.ZodEnum<{
|
|
2120
|
-
app: "app";
|
|
2121
2304
|
studio: "studio";
|
|
2305
|
+
app: "app";
|
|
2122
2306
|
}>>;
|
|
2123
2307
|
}, z.core.$strict>;
|
|
2124
2308
|
declare const EvalWireEventSchema: z.ZodDiscriminatedUnion<[z.ZodObject<{
|
|
@@ -2167,6 +2351,14 @@ declare const EvalWireEventSchema: z.ZodDiscriminatedUnion<[z.ZodObject<{
|
|
|
2167
2351
|
version: z.ZodLiteral<"compact-v1">;
|
|
2168
2352
|
history: z.ZodBoolean;
|
|
2169
2353
|
observation: z.ZodObject<{
|
|
2354
|
+
promptUsage: z.ZodOptional<z.ZodArray<z.ZodObject<{
|
|
2355
|
+
id: z.ZodString;
|
|
2356
|
+
version: z.ZodNumber;
|
|
2357
|
+
hash: z.ZodString;
|
|
2358
|
+
environment: z.ZodString;
|
|
2359
|
+
snapshotRevision: z.ZodString;
|
|
2360
|
+
model: z.ZodOptional<z.ZodString>;
|
|
2361
|
+
}, z.core.$strict>>>;
|
|
2170
2362
|
type: z.ZodEnum<{
|
|
2171
2363
|
interrupt: "interrupt";
|
|
2172
2364
|
error: "error";
|
|
@@ -2233,6 +2425,14 @@ declare const EvalWireEventSchema: z.ZodDiscriminatedUnion<[z.ZodObject<{
|
|
|
2233
2425
|
reference: z.ZodOptional<z.ZodJSONSchema>;
|
|
2234
2426
|
}, z.core.$strict>;
|
|
2235
2427
|
observation: z.ZodObject<{
|
|
2428
|
+
promptUsage: z.ZodOptional<z.ZodArray<z.ZodObject<{
|
|
2429
|
+
id: z.ZodString;
|
|
2430
|
+
version: z.ZodNumber;
|
|
2431
|
+
hash: z.ZodString;
|
|
2432
|
+
environment: z.ZodString;
|
|
2433
|
+
snapshotRevision: z.ZodString;
|
|
2434
|
+
model: z.ZodOptional<z.ZodString>;
|
|
2435
|
+
}, z.core.$strict>>>;
|
|
2236
2436
|
type: z.ZodEnum<{
|
|
2237
2437
|
interrupt: "interrupt";
|
|
2238
2438
|
error: "error";
|
|
@@ -2268,8 +2468,8 @@ declare const EvalWireEventSchema: z.ZodDiscriminatedUnion<[z.ZodObject<{
|
|
|
2268
2468
|
error: "error";
|
|
2269
2469
|
cancelled: "cancelled";
|
|
2270
2470
|
failed: "failed";
|
|
2271
|
-
passed: "passed";
|
|
2272
2471
|
ungraded: "ungraded";
|
|
2472
|
+
passed: "passed";
|
|
2273
2473
|
}>;
|
|
2274
2474
|
reason: z.ZodOptional<z.ZodString>;
|
|
2275
2475
|
judgeCalls: z.ZodOptional<z.ZodNumber>;
|
|
@@ -2323,8 +2523,8 @@ declare const EvalWireEventSchema: z.ZodDiscriminatedUnion<[z.ZodObject<{
|
|
|
2323
2523
|
error: "error";
|
|
2324
2524
|
cancelled: "cancelled";
|
|
2325
2525
|
failed: "failed";
|
|
2326
|
-
passed: "passed";
|
|
2327
2526
|
ungraded: "ungraded";
|
|
2527
|
+
passed: "passed";
|
|
2328
2528
|
}>;
|
|
2329
2529
|
durationMs: z.ZodNumber;
|
|
2330
2530
|
steps: z.ZodArray<z.ZodObject<{
|
|
@@ -2332,6 +2532,14 @@ declare const EvalWireEventSchema: z.ZodDiscriminatedUnion<[z.ZodObject<{
|
|
|
2332
2532
|
version: z.ZodLiteral<"compact-v1">;
|
|
2333
2533
|
history: z.ZodBoolean;
|
|
2334
2534
|
observation: z.ZodObject<{
|
|
2535
|
+
promptUsage: z.ZodOptional<z.ZodArray<z.ZodObject<{
|
|
2536
|
+
id: z.ZodString;
|
|
2537
|
+
version: z.ZodNumber;
|
|
2538
|
+
hash: z.ZodString;
|
|
2539
|
+
environment: z.ZodString;
|
|
2540
|
+
snapshotRevision: z.ZodString;
|
|
2541
|
+
model: z.ZodOptional<z.ZodString>;
|
|
2542
|
+
}, z.core.$strict>>>;
|
|
2335
2543
|
type: z.ZodEnum<{
|
|
2336
2544
|
interrupt: "interrupt";
|
|
2337
2545
|
error: "error";
|
|
@@ -2398,6 +2606,14 @@ declare const EvalWireEventSchema: z.ZodDiscriminatedUnion<[z.ZodObject<{
|
|
|
2398
2606
|
reference: z.ZodOptional<z.ZodJSONSchema>;
|
|
2399
2607
|
}, z.core.$strict>;
|
|
2400
2608
|
observation: z.ZodObject<{
|
|
2609
|
+
promptUsage: z.ZodOptional<z.ZodArray<z.ZodObject<{
|
|
2610
|
+
id: z.ZodString;
|
|
2611
|
+
version: z.ZodNumber;
|
|
2612
|
+
hash: z.ZodString;
|
|
2613
|
+
environment: z.ZodString;
|
|
2614
|
+
snapshotRevision: z.ZodString;
|
|
2615
|
+
model: z.ZodOptional<z.ZodString>;
|
|
2616
|
+
}, z.core.$strict>>>;
|
|
2401
2617
|
type: z.ZodEnum<{
|
|
2402
2618
|
interrupt: "interrupt";
|
|
2403
2619
|
error: "error";
|
|
@@ -2433,8 +2649,8 @@ declare const EvalWireEventSchema: z.ZodDiscriminatedUnion<[z.ZodObject<{
|
|
|
2433
2649
|
error: "error";
|
|
2434
2650
|
cancelled: "cancelled";
|
|
2435
2651
|
failed: "failed";
|
|
2436
|
-
passed: "passed";
|
|
2437
2652
|
ungraded: "ungraded";
|
|
2653
|
+
passed: "passed";
|
|
2438
2654
|
}>;
|
|
2439
2655
|
reason: z.ZodOptional<z.ZodString>;
|
|
2440
2656
|
judgeCalls: z.ZodOptional<z.ZodNumber>;
|
|
@@ -2588,8 +2804,8 @@ declare const EvalWireEventSchema: z.ZodDiscriminatedUnion<[z.ZodObject<{
|
|
|
2588
2804
|
id: z.ZodString;
|
|
2589
2805
|
version: z.ZodString;
|
|
2590
2806
|
location: z.ZodOptional<z.ZodEnum<{
|
|
2591
|
-
app: "app";
|
|
2592
2807
|
studio: "studio";
|
|
2808
|
+
app: "app";
|
|
2593
2809
|
}>>;
|
|
2594
2810
|
}, z.core.$strict>>;
|
|
2595
2811
|
startedAt: z.ZodISODateTime;
|
|
@@ -2598,8 +2814,8 @@ declare const EvalWireEventSchema: z.ZodDiscriminatedUnion<[z.ZodObject<{
|
|
|
2598
2814
|
error: "error";
|
|
2599
2815
|
cancelled: "cancelled";
|
|
2600
2816
|
failed: "failed";
|
|
2601
|
-
passed: "passed";
|
|
2602
2817
|
ungraded: "ungraded";
|
|
2818
|
+
passed: "passed";
|
|
2603
2819
|
}>;
|
|
2604
2820
|
counts: z.ZodObject<{
|
|
2605
2821
|
passed: z.ZodNumber;
|
|
@@ -2621,8 +2837,8 @@ declare const EvalWireEventSchema: z.ZodDiscriminatedUnion<[z.ZodObject<{
|
|
|
2621
2837
|
error: "error";
|
|
2622
2838
|
cancelled: "cancelled";
|
|
2623
2839
|
failed: "failed";
|
|
2624
|
-
passed: "passed";
|
|
2625
2840
|
ungraded: "ungraded";
|
|
2841
|
+
passed: "passed";
|
|
2626
2842
|
}>;
|
|
2627
2843
|
durationMs: z.ZodNumber;
|
|
2628
2844
|
steps: z.ZodArray<z.ZodObject<{
|
|
@@ -2630,6 +2846,14 @@ declare const EvalWireEventSchema: z.ZodDiscriminatedUnion<[z.ZodObject<{
|
|
|
2630
2846
|
version: z.ZodLiteral<"compact-v1">;
|
|
2631
2847
|
history: z.ZodBoolean;
|
|
2632
2848
|
observation: z.ZodObject<{
|
|
2849
|
+
promptUsage: z.ZodOptional<z.ZodArray<z.ZodObject<{
|
|
2850
|
+
id: z.ZodString;
|
|
2851
|
+
version: z.ZodNumber;
|
|
2852
|
+
hash: z.ZodString;
|
|
2853
|
+
environment: z.ZodString;
|
|
2854
|
+
snapshotRevision: z.ZodString;
|
|
2855
|
+
model: z.ZodOptional<z.ZodString>;
|
|
2856
|
+
}, z.core.$strict>>>;
|
|
2633
2857
|
type: z.ZodEnum<{
|
|
2634
2858
|
interrupt: "interrupt";
|
|
2635
2859
|
error: "error";
|
|
@@ -2696,6 +2920,14 @@ declare const EvalWireEventSchema: z.ZodDiscriminatedUnion<[z.ZodObject<{
|
|
|
2696
2920
|
reference: z.ZodOptional<z.ZodJSONSchema>;
|
|
2697
2921
|
}, z.core.$strict>;
|
|
2698
2922
|
observation: z.ZodObject<{
|
|
2923
|
+
promptUsage: z.ZodOptional<z.ZodArray<z.ZodObject<{
|
|
2924
|
+
id: z.ZodString;
|
|
2925
|
+
version: z.ZodNumber;
|
|
2926
|
+
hash: z.ZodString;
|
|
2927
|
+
environment: z.ZodString;
|
|
2928
|
+
snapshotRevision: z.ZodString;
|
|
2929
|
+
model: z.ZodOptional<z.ZodString>;
|
|
2930
|
+
}, z.core.$strict>>>;
|
|
2699
2931
|
type: z.ZodEnum<{
|
|
2700
2932
|
interrupt: "interrupt";
|
|
2701
2933
|
error: "error";
|
|
@@ -2731,8 +2963,8 @@ declare const EvalWireEventSchema: z.ZodDiscriminatedUnion<[z.ZodObject<{
|
|
|
2731
2963
|
error: "error";
|
|
2732
2964
|
cancelled: "cancelled";
|
|
2733
2965
|
failed: "failed";
|
|
2734
|
-
passed: "passed";
|
|
2735
2966
|
ungraded: "ungraded";
|
|
2967
|
+
passed: "passed";
|
|
2736
2968
|
}>;
|
|
2737
2969
|
reason: z.ZodOptional<z.ZodString>;
|
|
2738
2970
|
judgeCalls: z.ZodOptional<z.ZodNumber>;
|
|
@@ -2836,6 +3068,16 @@ declare const EvaluationMetadataSchema: z.ZodObject<{
|
|
|
2836
3068
|
declare const StudioEvaluationStartRequestSchema: z.ZodObject<{
|
|
2837
3069
|
targetId: z.ZodString;
|
|
2838
3070
|
name: z.ZodOptional<z.ZodString>;
|
|
3071
|
+
promptSelection: z.ZodOptional<z.ZodDiscriminatedUnion<[z.ZodObject<{
|
|
3072
|
+
type: z.ZodLiteral<"live">;
|
|
3073
|
+
}, z.core.$strict>, z.ZodObject<{
|
|
3074
|
+
type: z.ZodLiteral<"single">;
|
|
3075
|
+
id: z.ZodString;
|
|
3076
|
+
version: z.ZodNumber;
|
|
3077
|
+
}, z.core.$strict>, z.ZodObject<{
|
|
3078
|
+
type: z.ZodLiteral<"group">;
|
|
3079
|
+
groupId: z.ZodUUID;
|
|
3080
|
+
}, z.core.$strict>], "type">>;
|
|
2839
3081
|
selection: z.ZodEnum<{
|
|
2840
3082
|
selected: "selected";
|
|
2841
3083
|
all: "all";
|
|
@@ -2846,8 +3088,8 @@ declare const StudioEvaluationStartRequestSchema: z.ZodObject<{
|
|
|
2846
3088
|
caseIds: z.ZodOptional<z.ZodArray<z.ZodString>>;
|
|
2847
3089
|
}, z.core.$strict>>;
|
|
2848
3090
|
judge: z.ZodDefault<z.ZodEnum<{
|
|
2849
|
-
app: "app";
|
|
2850
3091
|
studio: "studio";
|
|
3092
|
+
app: "app";
|
|
2851
3093
|
}>>;
|
|
2852
3094
|
repetitions: z.ZodDefault<z.ZodNumber>;
|
|
2853
3095
|
concurrency: z.ZodDefault<z.ZodNumber>;
|
|
@@ -2888,8 +3130,8 @@ declare const StudioEvaluationSummarySchema: z.ZodObject<{
|
|
|
2888
3130
|
id: z.ZodString;
|
|
2889
3131
|
version: z.ZodString;
|
|
2890
3132
|
location: z.ZodOptional<z.ZodEnum<{
|
|
2891
|
-
app: "app";
|
|
2892
3133
|
studio: "studio";
|
|
3134
|
+
app: "app";
|
|
2893
3135
|
}>>;
|
|
2894
3136
|
}, z.core.$strict>>;
|
|
2895
3137
|
status: z.ZodEnum<{
|
|
@@ -2979,8 +3221,8 @@ declare const StudioEvaluationHistorySchema: z.ZodObject<{
|
|
|
2979
3221
|
id: z.ZodString;
|
|
2980
3222
|
version: z.ZodString;
|
|
2981
3223
|
location: z.ZodOptional<z.ZodEnum<{
|
|
2982
|
-
app: "app";
|
|
2983
3224
|
studio: "studio";
|
|
3225
|
+
app: "app";
|
|
2984
3226
|
}>>;
|
|
2985
3227
|
}, z.core.$strict>>;
|
|
2986
3228
|
status: z.ZodEnum<{
|
|
@@ -3071,8 +3313,8 @@ declare const StudioEvaluationDetailSchema: z.ZodObject<{
|
|
|
3071
3313
|
id: z.ZodString;
|
|
3072
3314
|
version: z.ZodString;
|
|
3073
3315
|
location: z.ZodOptional<z.ZodEnum<{
|
|
3074
|
-
app: "app";
|
|
3075
3316
|
studio: "studio";
|
|
3317
|
+
app: "app";
|
|
3076
3318
|
}>>;
|
|
3077
3319
|
}, z.core.$strict>>;
|
|
3078
3320
|
status: z.ZodEnum<{
|
|
@@ -3233,8 +3475,8 @@ declare const StudioEvaluationResultsSchema: z.ZodObject<{
|
|
|
3233
3475
|
id: z.ZodString;
|
|
3234
3476
|
version: z.ZodString;
|
|
3235
3477
|
location: z.ZodOptional<z.ZodEnum<{
|
|
3236
|
-
app: "app";
|
|
3237
3478
|
studio: "studio";
|
|
3479
|
+
app: "app";
|
|
3238
3480
|
}>>;
|
|
3239
3481
|
}, z.core.$strict>>;
|
|
3240
3482
|
status: z.ZodEnum<{
|
|
@@ -3408,19 +3650,62 @@ declare const StudioEvaluationResultsSchema: z.ZodObject<{
|
|
|
3408
3650
|
request: z.ZodOptional<z.ZodObject<{
|
|
3409
3651
|
caseIds: z.ZodOptional<z.ZodArray<z.ZodString>>;
|
|
3410
3652
|
grading: z.ZodOptional<z.ZodEnum<{
|
|
3411
|
-
app: "app";
|
|
3412
3653
|
studio: "studio";
|
|
3654
|
+
app: "app";
|
|
3413
3655
|
}>>;
|
|
3414
3656
|
judge: z.ZodOptional<z.ZodObject<{
|
|
3415
3657
|
id: z.ZodString;
|
|
3416
3658
|
version: z.ZodString;
|
|
3417
3659
|
location: z.ZodOptional<z.ZodEnum<{
|
|
3418
|
-
app: "app";
|
|
3419
3660
|
studio: "studio";
|
|
3661
|
+
app: "app";
|
|
3420
3662
|
}>>;
|
|
3421
3663
|
}, z.core.$strict>>;
|
|
3422
3664
|
repetitions: z.ZodNumber;
|
|
3423
3665
|
concurrency: z.ZodOptional<z.ZodNumber>;
|
|
3666
|
+
promptSnapshot: z.ZodOptional<z.ZodObject<{
|
|
3667
|
+
schemaVersion: z.ZodLiteral<1>;
|
|
3668
|
+
environment: z.ZodString;
|
|
3669
|
+
tag: z.ZodOptional<z.ZodString>;
|
|
3670
|
+
revision: z.ZodString;
|
|
3671
|
+
versions: z.ZodRecord<z.ZodString, z.ZodObject<{
|
|
3672
|
+
id: z.ZodString;
|
|
3673
|
+
version: z.ZodNumber;
|
|
3674
|
+
hash: z.ZodString;
|
|
3675
|
+
content: z.ZodObject<{
|
|
3676
|
+
format: z.ZodEnum<{
|
|
3677
|
+
"system-user": "system-user";
|
|
3678
|
+
chat: "chat";
|
|
3679
|
+
}>;
|
|
3680
|
+
messages: z.ZodArray<z.ZodObject<{
|
|
3681
|
+
role: z.ZodEnum<{
|
|
3682
|
+
system: "system";
|
|
3683
|
+
user: "user";
|
|
3684
|
+
assistant: "assistant";
|
|
3685
|
+
}>;
|
|
3686
|
+
content: z.ZodString;
|
|
3687
|
+
}, z.core.$strict>>;
|
|
3688
|
+
variablesSchema: z.ZodRecord<z.ZodString, z.ZodJSONSchema>;
|
|
3689
|
+
configSchema: z.ZodRecord<z.ZodString, z.ZodJSONSchema>;
|
|
3690
|
+
config: z.ZodRecord<z.ZodString, z.ZodJSONSchema>;
|
|
3691
|
+
dependencies: z.ZodDefault<z.ZodArray<z.ZodObject<{
|
|
3692
|
+
id: z.ZodString;
|
|
3693
|
+
version: z.ZodNumber;
|
|
3694
|
+
hash: z.ZodString;
|
|
3695
|
+
}, z.core.$strict>>>;
|
|
3696
|
+
}, z.core.$strict>;
|
|
3697
|
+
}, z.core.$strip>>;
|
|
3698
|
+
source: z.ZodEnum<{
|
|
3699
|
+
studio: "studio";
|
|
3700
|
+
local: "local";
|
|
3701
|
+
cache: "cache";
|
|
3702
|
+
fallback: "fallback";
|
|
3703
|
+
eval: "eval";
|
|
3704
|
+
}>;
|
|
3705
|
+
resolvedAt: z.ZodISODateTime;
|
|
3706
|
+
requestedIds: z.ZodOptional<z.ZodArray<z.ZodString>>;
|
|
3707
|
+
}, z.core.$strict>>;
|
|
3708
|
+
promptGroupName: z.ZodOptional<z.ZodString>;
|
|
3424
3709
|
}, z.core.$strip>>;
|
|
3425
3710
|
suite: z.ZodObject<{
|
|
3426
3711
|
evidence: z.ZodOptional<z.ZodObject<{
|
|
@@ -3608,8 +3893,8 @@ declare const StudioEvaluationResultsSchema: z.ZodObject<{
|
|
|
3608
3893
|
id: z.ZodString;
|
|
3609
3894
|
version: z.ZodString;
|
|
3610
3895
|
location: z.ZodOptional<z.ZodEnum<{
|
|
3611
|
-
app: "app";
|
|
3612
3896
|
studio: "studio";
|
|
3897
|
+
app: "app";
|
|
3613
3898
|
}>>;
|
|
3614
3899
|
}, z.core.$strict>>;
|
|
3615
3900
|
startedAt: z.ZodISODateTime;
|
|
@@ -3618,8 +3903,8 @@ declare const StudioEvaluationResultsSchema: z.ZodObject<{
|
|
|
3618
3903
|
error: "error";
|
|
3619
3904
|
cancelled: "cancelled";
|
|
3620
3905
|
failed: "failed";
|
|
3621
|
-
passed: "passed";
|
|
3622
3906
|
ungraded: "ungraded";
|
|
3907
|
+
passed: "passed";
|
|
3623
3908
|
}>;
|
|
3624
3909
|
counts: z.ZodObject<{
|
|
3625
3910
|
passed: z.ZodNumber;
|
|
@@ -3641,8 +3926,8 @@ declare const StudioEvaluationResultsSchema: z.ZodObject<{
|
|
|
3641
3926
|
error: "error";
|
|
3642
3927
|
cancelled: "cancelled";
|
|
3643
3928
|
failed: "failed";
|
|
3644
|
-
passed: "passed";
|
|
3645
3929
|
ungraded: "ungraded";
|
|
3930
|
+
passed: "passed";
|
|
3646
3931
|
}>;
|
|
3647
3932
|
durationMs: z.ZodNumber;
|
|
3648
3933
|
steps: z.ZodArray<z.ZodObject<{
|
|
@@ -3650,6 +3935,14 @@ declare const StudioEvaluationResultsSchema: z.ZodObject<{
|
|
|
3650
3935
|
version: z.ZodLiteral<"compact-v1">;
|
|
3651
3936
|
history: z.ZodBoolean;
|
|
3652
3937
|
observation: z.ZodObject<{
|
|
3938
|
+
promptUsage: z.ZodOptional<z.ZodArray<z.ZodObject<{
|
|
3939
|
+
id: z.ZodString;
|
|
3940
|
+
version: z.ZodNumber;
|
|
3941
|
+
hash: z.ZodString;
|
|
3942
|
+
environment: z.ZodString;
|
|
3943
|
+
snapshotRevision: z.ZodString;
|
|
3944
|
+
model: z.ZodOptional<z.ZodString>;
|
|
3945
|
+
}, z.core.$strict>>>;
|
|
3653
3946
|
type: z.ZodEnum<{
|
|
3654
3947
|
interrupt: "interrupt";
|
|
3655
3948
|
error: "error";
|
|
@@ -3716,6 +4009,14 @@ declare const StudioEvaluationResultsSchema: z.ZodObject<{
|
|
|
3716
4009
|
reference: z.ZodOptional<z.ZodJSONSchema>;
|
|
3717
4010
|
}, z.core.$strict>;
|
|
3718
4011
|
observation: z.ZodObject<{
|
|
4012
|
+
promptUsage: z.ZodOptional<z.ZodArray<z.ZodObject<{
|
|
4013
|
+
id: z.ZodString;
|
|
4014
|
+
version: z.ZodNumber;
|
|
4015
|
+
hash: z.ZodString;
|
|
4016
|
+
environment: z.ZodString;
|
|
4017
|
+
snapshotRevision: z.ZodString;
|
|
4018
|
+
model: z.ZodOptional<z.ZodString>;
|
|
4019
|
+
}, z.core.$strict>>>;
|
|
3719
4020
|
type: z.ZodEnum<{
|
|
3720
4021
|
interrupt: "interrupt";
|
|
3721
4022
|
error: "error";
|
|
@@ -3751,8 +4052,8 @@ declare const StudioEvaluationResultsSchema: z.ZodObject<{
|
|
|
3751
4052
|
error: "error";
|
|
3752
4053
|
cancelled: "cancelled";
|
|
3753
4054
|
failed: "failed";
|
|
3754
|
-
passed: "passed";
|
|
3755
4055
|
ungraded: "ungraded";
|
|
4056
|
+
passed: "passed";
|
|
3756
4057
|
}>;
|
|
3757
4058
|
reason: z.ZodOptional<z.ZodString>;
|
|
3758
4059
|
judgeCalls: z.ZodOptional<z.ZodNumber>;
|
|
@@ -3867,6 +4168,14 @@ declare const StudioEvaluationResultsSchema: z.ZodObject<{
|
|
|
3867
4168
|
version: z.ZodLiteral<"compact-v1">;
|
|
3868
4169
|
history: z.ZodBoolean;
|
|
3869
4170
|
observation: z.ZodObject<{
|
|
4171
|
+
promptUsage: z.ZodOptional<z.ZodArray<z.ZodObject<{
|
|
4172
|
+
id: z.ZodString;
|
|
4173
|
+
version: z.ZodNumber;
|
|
4174
|
+
hash: z.ZodString;
|
|
4175
|
+
environment: z.ZodString;
|
|
4176
|
+
snapshotRevision: z.ZodString;
|
|
4177
|
+
model: z.ZodOptional<z.ZodString>;
|
|
4178
|
+
}, z.core.$strict>>>;
|
|
3870
4179
|
type: z.ZodEnum<{
|
|
3871
4180
|
interrupt: "interrupt";
|
|
3872
4181
|
error: "error";
|
|
@@ -3933,6 +4242,14 @@ declare const StudioEvaluationResultsSchema: z.ZodObject<{
|
|
|
3933
4242
|
reference: z.ZodOptional<z.ZodJSONSchema>;
|
|
3934
4243
|
}, z.core.$strict>;
|
|
3935
4244
|
observation: z.ZodObject<{
|
|
4245
|
+
promptUsage: z.ZodOptional<z.ZodArray<z.ZodObject<{
|
|
4246
|
+
id: z.ZodString;
|
|
4247
|
+
version: z.ZodNumber;
|
|
4248
|
+
hash: z.ZodString;
|
|
4249
|
+
environment: z.ZodString;
|
|
4250
|
+
snapshotRevision: z.ZodString;
|
|
4251
|
+
model: z.ZodOptional<z.ZodString>;
|
|
4252
|
+
}, z.core.$strict>>>;
|
|
3936
4253
|
type: z.ZodEnum<{
|
|
3937
4254
|
interrupt: "interrupt";
|
|
3938
4255
|
error: "error";
|
|
@@ -3968,8 +4285,8 @@ declare const StudioEvaluationResultsSchema: z.ZodObject<{
|
|
|
3968
4285
|
error: "error";
|
|
3969
4286
|
cancelled: "cancelled";
|
|
3970
4287
|
failed: "failed";
|
|
3971
|
-
passed: "passed";
|
|
3972
4288
|
ungraded: "ungraded";
|
|
4289
|
+
passed: "passed";
|
|
3973
4290
|
}>;
|
|
3974
4291
|
reason: z.ZodOptional<z.ZodString>;
|
|
3975
4292
|
judgeCalls: z.ZodOptional<z.ZodNumber>;
|
|
@@ -4023,8 +4340,8 @@ declare const StudioEvaluationResultsSchema: z.ZodObject<{
|
|
|
4023
4340
|
error: "error";
|
|
4024
4341
|
cancelled: "cancelled";
|
|
4025
4342
|
failed: "failed";
|
|
4026
|
-
passed: "passed";
|
|
4027
4343
|
ungraded: "ungraded";
|
|
4344
|
+
passed: "passed";
|
|
4028
4345
|
}>;
|
|
4029
4346
|
durationMs: z.ZodNumber;
|
|
4030
4347
|
steps: z.ZodArray<z.ZodObject<{
|
|
@@ -4032,6 +4349,14 @@ declare const StudioEvaluationResultsSchema: z.ZodObject<{
|
|
|
4032
4349
|
version: z.ZodLiteral<"compact-v1">;
|
|
4033
4350
|
history: z.ZodBoolean;
|
|
4034
4351
|
observation: z.ZodObject<{
|
|
4352
|
+
promptUsage: z.ZodOptional<z.ZodArray<z.ZodObject<{
|
|
4353
|
+
id: z.ZodString;
|
|
4354
|
+
version: z.ZodNumber;
|
|
4355
|
+
hash: z.ZodString;
|
|
4356
|
+
environment: z.ZodString;
|
|
4357
|
+
snapshotRevision: z.ZodString;
|
|
4358
|
+
model: z.ZodOptional<z.ZodString>;
|
|
4359
|
+
}, z.core.$strict>>>;
|
|
4035
4360
|
type: z.ZodEnum<{
|
|
4036
4361
|
interrupt: "interrupt";
|
|
4037
4362
|
error: "error";
|
|
@@ -4098,6 +4423,14 @@ declare const StudioEvaluationResultsSchema: z.ZodObject<{
|
|
|
4098
4423
|
reference: z.ZodOptional<z.ZodJSONSchema>;
|
|
4099
4424
|
}, z.core.$strict>;
|
|
4100
4425
|
observation: z.ZodObject<{
|
|
4426
|
+
promptUsage: z.ZodOptional<z.ZodArray<z.ZodObject<{
|
|
4427
|
+
id: z.ZodString;
|
|
4428
|
+
version: z.ZodNumber;
|
|
4429
|
+
hash: z.ZodString;
|
|
4430
|
+
environment: z.ZodString;
|
|
4431
|
+
snapshotRevision: z.ZodString;
|
|
4432
|
+
model: z.ZodOptional<z.ZodString>;
|
|
4433
|
+
}, z.core.$strict>>>;
|
|
4101
4434
|
type: z.ZodEnum<{
|
|
4102
4435
|
interrupt: "interrupt";
|
|
4103
4436
|
error: "error";
|
|
@@ -4133,8 +4466,8 @@ declare const StudioEvaluationResultsSchema: z.ZodObject<{
|
|
|
4133
4466
|
error: "error";
|
|
4134
4467
|
cancelled: "cancelled";
|
|
4135
4468
|
failed: "failed";
|
|
4136
|
-
passed: "passed";
|
|
4137
4469
|
ungraded: "ungraded";
|
|
4470
|
+
passed: "passed";
|
|
4138
4471
|
}>;
|
|
4139
4472
|
reason: z.ZodOptional<z.ZodString>;
|
|
4140
4473
|
judgeCalls: z.ZodOptional<z.ZodNumber>;
|