@effect-agent/testing 0.0.1-beta.6 → 0.1.0-beta.8
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/index.d.mts +113 -83
- package/dist/index.mjs +5 -0
- package/dist/index.mjs.map +1 -1
- package/package.json +18 -18
- package/src/certification.ts +5 -0
- package/src/fixtures/travel-planner/phase6.ts +4 -0
package/dist/index.d.mts
CHANGED
|
@@ -141,7 +141,7 @@ declare const ChaosPlan_base: Schema.Class<ChaosPlan, Schema.Struct<{
|
|
|
141
141
|
readonly lanes: Schema.Int;
|
|
142
142
|
readonly submissions: Schema.NonEmptyArray<typeof ChaosSubmissionSpec>;
|
|
143
143
|
/** Coordinator failpoint arms, consumed one per round (each fails every hit that round). */
|
|
144
|
-
readonly failpointArms: Schema.$Array<Schema.Literals<readonly ["submit:after-admit", "submit:after-materialize", "claim:after-claim", "input:after-canonical-append", "turn:after-canonical-append", "turn:after-response-append", "turn:after-results-append", "tools:after-prepared-append", "step:after-step-append", "approval:after-request-append", "approval:after-suspend", "join:after-claim", "join:after-canonical-append", "resolve:after-intent", "terminalize:after-reserve", "terminalize:after-canonical-append", "abort:after-intent", "subagent:after-reserve", "subagent:after-request-append", "subagent:after-admit", "subagent:after-child-ready", "subagent:after-start-append", "subagent:after-sibling-settle", "subagent:after-suspend", "subagent:after-child-abort-intent", "subagent:after-join-append", "subagent:after-release-pending", "subagent:after-release"]>>;
|
|
144
|
+
readonly failpointArms: Schema.$Array<Schema.Literals<readonly ["submit:after-admit", "submit:after-materialize", "claim:after-claim", "input:after-canonical-append", "turn:after-canonical-append", "turn:after-response-append", "turn:after-results-append", "compaction:after-canonical-append", "tools:after-prepared-append", "step:after-step-append", "approval:after-request-append", "approval:after-suspend", "join:after-claim", "join:after-canonical-append", "resolve:after-intent", "terminalize:after-reserve", "terminalize:after-canonical-append", "abort:after-intent", "subagent:after-reserve", "subagent:after-request-append", "subagent:after-admit", "subagent:after-child-ready", "subagent:after-start-append", "subagent:after-sibling-settle", "subagent:after-suspend", "subagent:after-child-abort-intent", "subagent:after-join-append", "subagent:after-release-pending", "subagent:after-release"]>>;
|
|
145
145
|
/** Adapter-owned failpoint arms (e.g. SQLite `ledger:*`/`append:*` locations). */
|
|
146
146
|
readonly adapterArms: Schema.$Array<Schema.String>;
|
|
147
147
|
/** Flattened submission indices to abort mid-plan (modulo the submission count). */
|
|
@@ -365,7 +365,7 @@ declare const delegateDocumentSummary: import("@effect-agent/capabilities").Suba
|
|
|
365
365
|
readonly failure: typeof DocumentUnavailable;
|
|
366
366
|
readonly failureMode: "error";
|
|
367
367
|
}, DocumentLibrary>;
|
|
368
|
-
}, typeof SummaryRequest, typeof SummaryFinding, typeof DocumentSummaryFailed, never, never>;
|
|
368
|
+
}, typeof SummaryRequest, typeof SummaryFinding, typeof DocumentSummaryFailed, never, never, "error">;
|
|
369
369
|
/** Total mapping from every expected child Run failure to the declared Tool failure (SUB-028). */
|
|
370
370
|
declare const mapSummaryChildFailure: (failure: {
|
|
371
371
|
readonly _tag: string;
|
|
@@ -401,10 +401,20 @@ declare class ResearchDigest extends ResearchDigest_base {}
|
|
|
401
401
|
declare const docsCoordinatorConfidentialMarker = "docs-coordinator-vault-19x";
|
|
402
402
|
declare const docsMissionConfidentialMarker = "docs-mission-dossier-42f";
|
|
403
403
|
declare const DocsResearcherToolkit: Toolkit.Toolkit<{
|
|
404
|
-
readonly delegate_document_summary:
|
|
404
|
+
readonly delegate_document_summary: Tool.Tool<"delegate_document_summary", {
|
|
405
|
+
readonly parameters: typeof SummaryRequest;
|
|
406
|
+
readonly success: typeof SummaryFinding;
|
|
407
|
+
readonly failure: import("@effect-agent/capabilities").SubagentToolFailure<typeof DocumentSummaryFailed>;
|
|
408
|
+
readonly failureMode: "error";
|
|
409
|
+
}, import("@effect-agent/engine").AgentSpawner | import("@effect-agent/core").IdGenerator | import("@effect-agent/engine").RunEventSink | import("@effect-agent/engine").SubagentDurability>;
|
|
405
410
|
}>;
|
|
406
411
|
declare const DocsResearcher: import("@effect-agent/core").Definition<typeof ResearchRequest, typeof ResearchDigest, string, Toolkit.Toolkit<{
|
|
407
|
-
readonly delegate_document_summary:
|
|
412
|
+
readonly delegate_document_summary: Tool.Tool<"delegate_document_summary", {
|
|
413
|
+
readonly parameters: typeof SummaryRequest;
|
|
414
|
+
readonly success: typeof SummaryFinding;
|
|
415
|
+
readonly failure: import("@effect-agent/capabilities").SubagentToolFailure<typeof DocumentSummaryFailed>;
|
|
416
|
+
readonly failureMode: "error";
|
|
417
|
+
}, import("@effect-agent/engine").AgentSpawner | import("@effect-agent/core").IdGenerator | import("@effect-agent/engine").RunEventSink | import("@effect-agent/engine").SubagentDurability>;
|
|
408
418
|
}>>;
|
|
409
419
|
/** The default two-document research mission. */
|
|
410
420
|
declare const researchMissionRequest: ResearchRequest;
|
|
@@ -983,6 +993,42 @@ declare const phase3TravelPlannerEncodedFixture: readonly {
|
|
|
983
993
|
readonly createdAt: string;
|
|
984
994
|
readonly deploymentId: string;
|
|
985
995
|
readonly payload: {
|
|
996
|
+
readonly _tag: "AbortRequested";
|
|
997
|
+
readonly submissionId: string;
|
|
998
|
+
readonly author: string;
|
|
999
|
+
readonly reason: string;
|
|
1000
|
+
} | {
|
|
1001
|
+
readonly _tag: "ToolCallResolved";
|
|
1002
|
+
readonly runId: string;
|
|
1003
|
+
readonly toolCallId: string;
|
|
1004
|
+
readonly resolution: "completed-with-result" | "failed-with-error" | "never-started" | "safe-retry";
|
|
1005
|
+
readonly author: string;
|
|
1006
|
+
readonly reason: string;
|
|
1007
|
+
} | {
|
|
1008
|
+
readonly _tag: "ToolCallPrepared";
|
|
1009
|
+
readonly runId: string;
|
|
1010
|
+
readonly turnId: string;
|
|
1011
|
+
readonly turn: number;
|
|
1012
|
+
readonly toolCallId: string;
|
|
1013
|
+
readonly toolName: string;
|
|
1014
|
+
readonly parameters: Schema.Json;
|
|
1015
|
+
readonly parametersDigest: string;
|
|
1016
|
+
} | {
|
|
1017
|
+
readonly _tag: "ToolCallUnknown";
|
|
1018
|
+
readonly runId: string;
|
|
1019
|
+
readonly turn: number;
|
|
1020
|
+
readonly toolCallId: string;
|
|
1021
|
+
readonly toolName: string;
|
|
1022
|
+
readonly reason: string;
|
|
1023
|
+
} | {
|
|
1024
|
+
readonly _tag: "ToolApprovalDecided";
|
|
1025
|
+
readonly runId: string;
|
|
1026
|
+
readonly turn: number;
|
|
1027
|
+
readonly toolCallId: string;
|
|
1028
|
+
readonly decision: "approved" | "denied";
|
|
1029
|
+
readonly resolver: string;
|
|
1030
|
+
readonly reason: string;
|
|
1031
|
+
} | {
|
|
986
1032
|
readonly _tag: "ConversationCreated";
|
|
987
1033
|
readonly agentId: string;
|
|
988
1034
|
readonly definitions: {
|
|
@@ -1007,15 +1053,8 @@ declare const phase3TravelPlannerEncodedFixture: readonly {
|
|
|
1007
1053
|
readonly turn: number;
|
|
1008
1054
|
readonly messages: Schema.Json;
|
|
1009
1055
|
readonly messagesDigest: string;
|
|
1010
|
-
|
|
1011
|
-
readonly
|
|
1012
|
-
readonly runId: string;
|
|
1013
|
-
readonly turnId: string;
|
|
1014
|
-
readonly turn: number;
|
|
1015
|
-
readonly toolCallId: string;
|
|
1016
|
-
readonly toolName: string;
|
|
1017
|
-
readonly parameters: Schema.Json;
|
|
1018
|
-
readonly parametersDigest: string;
|
|
1056
|
+
readonly inputTokens?: number | undefined;
|
|
1057
|
+
readonly outputTokens?: number | undefined;
|
|
1019
1058
|
} | {
|
|
1020
1059
|
readonly _tag: "ToolCallSettled";
|
|
1021
1060
|
readonly runId: string;
|
|
@@ -1023,20 +1062,6 @@ declare const phase3TravelPlannerEncodedFixture: readonly {
|
|
|
1023
1062
|
readonly toolName: string;
|
|
1024
1063
|
readonly result: Schema.Json;
|
|
1025
1064
|
readonly isFailure: boolean;
|
|
1026
|
-
} | {
|
|
1027
|
-
readonly _tag: "ToolCallUnknown";
|
|
1028
|
-
readonly runId: string;
|
|
1029
|
-
readonly turn: number;
|
|
1030
|
-
readonly toolCallId: string;
|
|
1031
|
-
readonly toolName: string;
|
|
1032
|
-
readonly reason: string;
|
|
1033
|
-
} | {
|
|
1034
|
-
readonly _tag: "ToolCallResolved";
|
|
1035
|
-
readonly runId: string;
|
|
1036
|
-
readonly toolCallId: string;
|
|
1037
|
-
readonly resolution: "completed-with-result" | "failed-with-error" | "never-started" | "safe-retry";
|
|
1038
|
-
readonly author: string;
|
|
1039
|
-
readonly reason: string;
|
|
1040
1065
|
} | {
|
|
1041
1066
|
readonly _tag: "ToolStepSettled";
|
|
1042
1067
|
readonly runId: string;
|
|
@@ -1052,14 +1077,6 @@ declare const phase3TravelPlannerEncodedFixture: readonly {
|
|
|
1052
1077
|
readonly toolCallId: string;
|
|
1053
1078
|
readonly toolName: string;
|
|
1054
1079
|
readonly parametersDigest: string;
|
|
1055
|
-
} | {
|
|
1056
|
-
readonly _tag: "ToolApprovalDecided";
|
|
1057
|
-
readonly runId: string;
|
|
1058
|
-
readonly turn: number;
|
|
1059
|
-
readonly toolCallId: string;
|
|
1060
|
-
readonly decision: "approved" | "denied";
|
|
1061
|
-
readonly resolver: string;
|
|
1062
|
-
readonly reason: string;
|
|
1063
1080
|
} | {
|
|
1064
1081
|
readonly _tag: "ModelResponseInterrupted";
|
|
1065
1082
|
readonly runId: string;
|
|
@@ -1069,8 +1086,10 @@ declare const phase3TravelPlannerEncodedFixture: readonly {
|
|
|
1069
1086
|
} | {
|
|
1070
1087
|
readonly _tag: "CompactionCreated";
|
|
1071
1088
|
readonly runId: string;
|
|
1072
|
-
readonly
|
|
1073
|
-
readonly
|
|
1089
|
+
readonly turn: number;
|
|
1090
|
+
readonly kind: "clear-tool-results" | "summarize";
|
|
1091
|
+
readonly coversThrough: number;
|
|
1092
|
+
readonly summary?: string | undefined;
|
|
1074
1093
|
} | {
|
|
1075
1094
|
readonly _tag: "RunFailed";
|
|
1076
1095
|
readonly runId: string;
|
|
@@ -1079,11 +1098,6 @@ declare const phase3TravelPlannerEncodedFixture: readonly {
|
|
|
1079
1098
|
readonly _tag: "RunCompleted";
|
|
1080
1099
|
readonly runId: string;
|
|
1081
1100
|
readonly output: Schema.Json;
|
|
1082
|
-
} | {
|
|
1083
|
-
readonly _tag: "AbortRequested";
|
|
1084
|
-
readonly submissionId: string;
|
|
1085
|
-
readonly author: string;
|
|
1086
|
-
readonly reason: string;
|
|
1087
1101
|
} | {
|
|
1088
1102
|
readonly _tag: "SubmissionSettled";
|
|
1089
1103
|
readonly submissionId: string;
|
|
@@ -1092,6 +1106,7 @@ declare const phase3TravelPlannerEncodedFixture: readonly {
|
|
|
1092
1106
|
readonly outcome: "aborted" | "completed" | "failed";
|
|
1093
1107
|
readonly runId?: string | undefined;
|
|
1094
1108
|
readonly result?: Schema.Json | undefined;
|
|
1109
|
+
readonly finishReason?: "budget-exhausted" | undefined;
|
|
1095
1110
|
} | {
|
|
1096
1111
|
readonly _tag: "SubagentRequested";
|
|
1097
1112
|
readonly runId: string;
|
|
@@ -1163,6 +1178,42 @@ declare const phase3TravelPlannerEncodedFixture: readonly {
|
|
|
1163
1178
|
readonly createdAt: string;
|
|
1164
1179
|
readonly deploymentId: string;
|
|
1165
1180
|
readonly payload: {
|
|
1181
|
+
readonly _tag: "AbortRequested";
|
|
1182
|
+
readonly submissionId: string;
|
|
1183
|
+
readonly author: string;
|
|
1184
|
+
readonly reason: string;
|
|
1185
|
+
} | {
|
|
1186
|
+
readonly _tag: "ToolCallResolved";
|
|
1187
|
+
readonly runId: string;
|
|
1188
|
+
readonly toolCallId: string;
|
|
1189
|
+
readonly resolution: "completed-with-result" | "failed-with-error" | "never-started" | "safe-retry";
|
|
1190
|
+
readonly author: string;
|
|
1191
|
+
readonly reason: string;
|
|
1192
|
+
} | {
|
|
1193
|
+
readonly _tag: "ToolCallPrepared";
|
|
1194
|
+
readonly runId: string;
|
|
1195
|
+
readonly turnId: string;
|
|
1196
|
+
readonly turn: number;
|
|
1197
|
+
readonly toolCallId: string;
|
|
1198
|
+
readonly toolName: string;
|
|
1199
|
+
readonly parameters: Schema.Json;
|
|
1200
|
+
readonly parametersDigest: string;
|
|
1201
|
+
} | {
|
|
1202
|
+
readonly _tag: "ToolCallUnknown";
|
|
1203
|
+
readonly runId: string;
|
|
1204
|
+
readonly turn: number;
|
|
1205
|
+
readonly toolCallId: string;
|
|
1206
|
+
readonly toolName: string;
|
|
1207
|
+
readonly reason: string;
|
|
1208
|
+
} | {
|
|
1209
|
+
readonly _tag: "ToolApprovalDecided";
|
|
1210
|
+
readonly runId: string;
|
|
1211
|
+
readonly turn: number;
|
|
1212
|
+
readonly toolCallId: string;
|
|
1213
|
+
readonly decision: "approved" | "denied";
|
|
1214
|
+
readonly resolver: string;
|
|
1215
|
+
readonly reason: string;
|
|
1216
|
+
} | {
|
|
1166
1217
|
readonly _tag: "ConversationCreated";
|
|
1167
1218
|
readonly agentId: string;
|
|
1168
1219
|
readonly definitions: {
|
|
@@ -1187,15 +1238,8 @@ declare const phase3TravelPlannerEncodedFixture: readonly {
|
|
|
1187
1238
|
readonly turn: number;
|
|
1188
1239
|
readonly messages: Schema.Json;
|
|
1189
1240
|
readonly messagesDigest: string;
|
|
1190
|
-
|
|
1191
|
-
readonly
|
|
1192
|
-
readonly runId: string;
|
|
1193
|
-
readonly turnId: string;
|
|
1194
|
-
readonly turn: number;
|
|
1195
|
-
readonly toolCallId: string;
|
|
1196
|
-
readonly toolName: string;
|
|
1197
|
-
readonly parameters: Schema.Json;
|
|
1198
|
-
readonly parametersDigest: string;
|
|
1241
|
+
readonly inputTokens?: number | undefined;
|
|
1242
|
+
readonly outputTokens?: number | undefined;
|
|
1199
1243
|
} | {
|
|
1200
1244
|
readonly _tag: "ToolCallSettled";
|
|
1201
1245
|
readonly runId: string;
|
|
@@ -1203,20 +1247,6 @@ declare const phase3TravelPlannerEncodedFixture: readonly {
|
|
|
1203
1247
|
readonly toolName: string;
|
|
1204
1248
|
readonly result: Schema.Json;
|
|
1205
1249
|
readonly isFailure: boolean;
|
|
1206
|
-
} | {
|
|
1207
|
-
readonly _tag: "ToolCallUnknown";
|
|
1208
|
-
readonly runId: string;
|
|
1209
|
-
readonly turn: number;
|
|
1210
|
-
readonly toolCallId: string;
|
|
1211
|
-
readonly toolName: string;
|
|
1212
|
-
readonly reason: string;
|
|
1213
|
-
} | {
|
|
1214
|
-
readonly _tag: "ToolCallResolved";
|
|
1215
|
-
readonly runId: string;
|
|
1216
|
-
readonly toolCallId: string;
|
|
1217
|
-
readonly resolution: "completed-with-result" | "failed-with-error" | "never-started" | "safe-retry";
|
|
1218
|
-
readonly author: string;
|
|
1219
|
-
readonly reason: string;
|
|
1220
1250
|
} | {
|
|
1221
1251
|
readonly _tag: "ToolStepSettled";
|
|
1222
1252
|
readonly runId: string;
|
|
@@ -1232,14 +1262,6 @@ declare const phase3TravelPlannerEncodedFixture: readonly {
|
|
|
1232
1262
|
readonly toolCallId: string;
|
|
1233
1263
|
readonly toolName: string;
|
|
1234
1264
|
readonly parametersDigest: string;
|
|
1235
|
-
} | {
|
|
1236
|
-
readonly _tag: "ToolApprovalDecided";
|
|
1237
|
-
readonly runId: string;
|
|
1238
|
-
readonly turn: number;
|
|
1239
|
-
readonly toolCallId: string;
|
|
1240
|
-
readonly decision: "approved" | "denied";
|
|
1241
|
-
readonly resolver: string;
|
|
1242
|
-
readonly reason: string;
|
|
1243
1265
|
} | {
|
|
1244
1266
|
readonly _tag: "ModelResponseInterrupted";
|
|
1245
1267
|
readonly runId: string;
|
|
@@ -1249,8 +1271,10 @@ declare const phase3TravelPlannerEncodedFixture: readonly {
|
|
|
1249
1271
|
} | {
|
|
1250
1272
|
readonly _tag: "CompactionCreated";
|
|
1251
1273
|
readonly runId: string;
|
|
1252
|
-
readonly
|
|
1253
|
-
readonly
|
|
1274
|
+
readonly turn: number;
|
|
1275
|
+
readonly kind: "clear-tool-results" | "summarize";
|
|
1276
|
+
readonly coversThrough: number;
|
|
1277
|
+
readonly summary?: string | undefined;
|
|
1254
1278
|
} | {
|
|
1255
1279
|
readonly _tag: "RunFailed";
|
|
1256
1280
|
readonly runId: string;
|
|
@@ -1259,11 +1283,6 @@ declare const phase3TravelPlannerEncodedFixture: readonly {
|
|
|
1259
1283
|
readonly _tag: "RunCompleted";
|
|
1260
1284
|
readonly runId: string;
|
|
1261
1285
|
readonly output: Schema.Json;
|
|
1262
|
-
} | {
|
|
1263
|
-
readonly _tag: "AbortRequested";
|
|
1264
|
-
readonly submissionId: string;
|
|
1265
|
-
readonly author: string;
|
|
1266
|
-
readonly reason: string;
|
|
1267
1286
|
} | {
|
|
1268
1287
|
readonly _tag: "SubmissionSettled";
|
|
1269
1288
|
readonly submissionId: string;
|
|
@@ -1272,6 +1291,7 @@ declare const phase3TravelPlannerEncodedFixture: readonly {
|
|
|
1272
1291
|
readonly outcome: "aborted" | "completed" | "failed";
|
|
1273
1292
|
readonly runId?: string | undefined;
|
|
1274
1293
|
readonly result?: Schema.Json | undefined;
|
|
1294
|
+
readonly finishReason?: "budget-exhausted" | undefined;
|
|
1275
1295
|
} | {
|
|
1276
1296
|
readonly _tag: "SubagentRequested";
|
|
1277
1297
|
readonly runId: string;
|
|
@@ -2772,7 +2792,7 @@ declare const destinationResearchDelegation: import("@effect-agent/capabilities"
|
|
|
2772
2792
|
readonly failure: typeof DestinationGuideUnavailable;
|
|
2773
2793
|
readonly failureMode: "error";
|
|
2774
2794
|
}, DestinationGuide>;
|
|
2775
|
-
}, typeof DestinationResearchRequest, typeof DestinationResearchFindings, typeof DestinationResearchFailed, ResearchDispatchGate, never>;
|
|
2795
|
+
}, typeof DestinationResearchRequest, typeof DestinationResearchFindings, typeof DestinationResearchFailed, ResearchDispatchGate, never, "error">;
|
|
2776
2796
|
/** Total mapping from every expected child Run failure to the declared Tool failure (SUB-028). */
|
|
2777
2797
|
declare const mapResearchChildFailure: (failure: {
|
|
2778
2798
|
readonly _tag: string;
|
|
@@ -2807,10 +2827,20 @@ declare class DestinationShortlist extends DestinationShortlist_base {}
|
|
|
2807
2827
|
declare const coordinatorConfidentialMarker = "coordinator-vault-7q42";
|
|
2808
2828
|
declare const missionConfidentialMarker = "traveler-dossier-19f";
|
|
2809
2829
|
declare const TravelCoordinatorToolkit: Toolkit.Toolkit<{
|
|
2810
|
-
readonly delegate_destination_research:
|
|
2830
|
+
readonly delegate_destination_research: Tool.Tool<"delegate_destination_research", {
|
|
2831
|
+
readonly parameters: typeof DestinationResearchRequest;
|
|
2832
|
+
readonly success: typeof DestinationResearchFindings;
|
|
2833
|
+
readonly failure: import("@effect-agent/capabilities").SubagentToolFailure<typeof DestinationResearchFailed>;
|
|
2834
|
+
readonly failureMode: "error";
|
|
2835
|
+
}, import("@effect-agent/engine").AgentSpawner | import("@effect-agent/core").IdGenerator | import("@effect-agent/engine").RunEventSink | import("@effect-agent/engine").SubagentDurability>;
|
|
2811
2836
|
}>;
|
|
2812
2837
|
declare const TravelCoordinator: import("@effect-agent/core").Definition<typeof ResearchMission, typeof DestinationShortlist, string, Toolkit.Toolkit<{
|
|
2813
|
-
readonly delegate_destination_research:
|
|
2838
|
+
readonly delegate_destination_research: Tool.Tool<"delegate_destination_research", {
|
|
2839
|
+
readonly parameters: typeof DestinationResearchRequest;
|
|
2840
|
+
readonly success: typeof DestinationResearchFindings;
|
|
2841
|
+
readonly failure: import("@effect-agent/capabilities").SubagentToolFailure<typeof DestinationResearchFailed>;
|
|
2842
|
+
readonly failureMode: "error";
|
|
2843
|
+
}, import("@effect-agent/engine").AgentSpawner | import("@effect-agent/core").IdGenerator | import("@effect-agent/engine").RunEventSink | import("@effect-agent/engine").SubagentDurability>;
|
|
2814
2844
|
}>>;
|
|
2815
2845
|
declare const researchMission: ResearchMission;
|
|
2816
2846
|
declare const expectedDestinationShortlist: DestinationShortlist;
|
package/dist/index.mjs
CHANGED
|
@@ -29,6 +29,7 @@ const CERTIFICATION_SCENARIOS = [
|
|
|
29
29
|
*/
|
|
30
30
|
const TIER2_UNREACHED_LOCATIONS = [
|
|
31
31
|
"abort:after-intent",
|
|
32
|
+
"compaction:after-canonical-append",
|
|
32
33
|
"resolve:after-intent",
|
|
33
34
|
"subagent:after-child-abort-intent"
|
|
34
35
|
];
|
|
@@ -4519,6 +4520,8 @@ const phase6TravelPlannerGoldenEvidence = [
|
|
|
4519
4520
|
runId: "run:{submissionId}",
|
|
4520
4521
|
turnId: "turn:run:{submissionId}:1",
|
|
4521
4522
|
turn: 1,
|
|
4523
|
+
inputTokens: 128,
|
|
4524
|
+
outputTokens: 96,
|
|
4522
4525
|
messages: { content: [
|
|
4523
4526
|
{
|
|
4524
4527
|
options: {},
|
|
@@ -4660,6 +4663,8 @@ const phase6TravelPlannerGoldenEvidence = [
|
|
|
4660
4663
|
runId: "run:{submissionId}",
|
|
4661
4664
|
turnId: "turn:run:{submissionId}:2",
|
|
4662
4665
|
turn: 2,
|
|
4666
|
+
inputTokens: 128,
|
|
4667
|
+
outputTokens: 96,
|
|
4663
4668
|
messages: { content: [{
|
|
4664
4669
|
options: {},
|
|
4665
4670
|
role: "assistant",
|