@meetkai/mka1 0.54.67 → 0.54.69

This diff represents the content of publicly available package versions that have been released to one of the supported registries. The information contained in this diff is provided for informational purposes only and reflects changes between package versions as they appear in their respective public registries.
Files changed (153) hide show
  1. package/bin/mcp-server.js +197 -17
  2. package/bin/mcp-server.js.map +21 -21
  3. package/dist/commonjs/lib/config.d.ts +2 -2
  4. package/dist/commonjs/lib/config.js +2 -2
  5. package/dist/commonjs/mcp-server/mcp-server.js +1 -1
  6. package/dist/commonjs/mcp-server/server.js +1 -1
  7. package/dist/commonjs/models/components/appendagenttrialsrequest.d.ts +8 -0
  8. package/dist/commonjs/models/components/appendagenttrialsrequest.d.ts.map +1 -1
  9. package/dist/commonjs/models/components/appendagenttrialsrequest.js +4 -0
  10. package/dist/commonjs/models/components/appendagenttrialsrequest.js.map +1 -1
  11. package/dist/commonjs/models/components/computelog.d.ts +2 -2
  12. package/dist/commonjs/models/components/evalleaderboardrow.d.ts +35 -0
  13. package/dist/commonjs/models/components/evalleaderboardrow.d.ts.map +1 -1
  14. package/dist/commonjs/models/components/evalleaderboardrow.js +26 -0
  15. package/dist/commonjs/models/components/evalleaderboardrow.js.map +1 -1
  16. package/dist/commonjs/models/components/evalrunoverview.d.ts +5 -0
  17. package/dist/commonjs/models/components/evalrunoverview.d.ts.map +1 -1
  18. package/dist/commonjs/models/components/evalrunoverview.js +2 -0
  19. package/dist/commonjs/models/components/evalrunoverview.js.map +1 -1
  20. package/dist/commonjs/models/components/evalsampleobject.d.ts +5 -0
  21. package/dist/commonjs/models/components/evalsampleobject.d.ts.map +1 -1
  22. package/dist/commonjs/models/components/evalsampleobject.js +4 -0
  23. package/dist/commonjs/models/components/evalsampleobject.js.map +1 -1
  24. package/dist/commonjs/models/components/importagentrunrequest.d.ts +8 -0
  25. package/dist/commonjs/models/components/importagentrunrequest.d.ts.map +1 -1
  26. package/dist/commonjs/models/components/importagentrunrequest.js +4 -0
  27. package/dist/commonjs/models/components/importagentrunrequest.js.map +1 -1
  28. package/dist/commonjs/models/operations/appendagentevaltrials.d.ts +36 -0
  29. package/dist/commonjs/models/operations/appendagentevaltrials.d.ts.map +1 -1
  30. package/dist/commonjs/models/operations/appendagentevaltrials.js +25 -1
  31. package/dist/commonjs/models/operations/appendagentevaltrials.js.map +1 -1
  32. package/dist/commonjs/models/operations/cancelevalrun.d.ts +36 -0
  33. package/dist/commonjs/models/operations/cancelevalrun.d.ts.map +1 -1
  34. package/dist/commonjs/models/operations/cancelevalrun.js +25 -1
  35. package/dist/commonjs/models/operations/cancelevalrun.js.map +1 -1
  36. package/dist/commonjs/models/operations/createevalrun.d.ts +36 -0
  37. package/dist/commonjs/models/operations/createevalrun.d.ts.map +1 -1
  38. package/dist/commonjs/models/operations/createevalrun.js +26 -1
  39. package/dist/commonjs/models/operations/createevalrun.js.map +1 -1
  40. package/dist/commonjs/models/operations/getevalrun.d.ts +36 -0
  41. package/dist/commonjs/models/operations/getevalrun.d.ts.map +1 -1
  42. package/dist/commonjs/models/operations/getevalrun.js +25 -1
  43. package/dist/commonjs/models/operations/getevalrun.js.map +1 -1
  44. package/dist/commonjs/models/operations/importagentevalrun.d.ts +36 -0
  45. package/dist/commonjs/models/operations/importagentevalrun.d.ts.map +1 -1
  46. package/dist/commonjs/models/operations/importagentevalrun.js +25 -1
  47. package/dist/commonjs/models/operations/importagentevalrun.js.map +1 -1
  48. package/dist/commonjs/models/operations/listevalruns.d.ts +36 -0
  49. package/dist/commonjs/models/operations/listevalruns.d.ts.map +1 -1
  50. package/dist/commonjs/models/operations/listevalruns.js +25 -1
  51. package/dist/commonjs/models/operations/listevalruns.js.map +1 -1
  52. package/dist/commonjs/models/operations/listevalscheduleruns.d.ts +36 -0
  53. package/dist/commonjs/models/operations/listevalscheduleruns.d.ts.map +1 -1
  54. package/dist/commonjs/models/operations/listevalscheduleruns.js +25 -1
  55. package/dist/commonjs/models/operations/listevalscheduleruns.js.map +1 -1
  56. package/dist/commonjs/models/operations/listevalsuiteleaderboardruns.d.ts +36 -0
  57. package/dist/commonjs/models/operations/listevalsuiteleaderboardruns.d.ts.map +1 -1
  58. package/dist/commonjs/models/operations/listevalsuiteleaderboardruns.js +25 -1
  59. package/dist/commonjs/models/operations/listevalsuiteleaderboardruns.js.map +1 -1
  60. package/dist/commonjs/models/operations/rerunfailedevalsamples.d.ts +36 -0
  61. package/dist/commonjs/models/operations/rerunfailedevalsamples.d.ts.map +1 -1
  62. package/dist/commonjs/models/operations/rerunfailedevalsamples.js +25 -1
  63. package/dist/commonjs/models/operations/rerunfailedevalsamples.js.map +1 -1
  64. package/dist/commonjs/models/operations/retryfailedevalrun.d.ts +36 -0
  65. package/dist/commonjs/models/operations/retryfailedevalrun.d.ts.map +1 -1
  66. package/dist/commonjs/models/operations/retryfailedevalrun.js +25 -1
  67. package/dist/commonjs/models/operations/retryfailedevalrun.js.map +1 -1
  68. package/dist/esm/lib/config.d.ts +2 -2
  69. package/dist/esm/lib/config.js +2 -2
  70. package/dist/esm/mcp-server/mcp-server.js +1 -1
  71. package/dist/esm/mcp-server/server.js +1 -1
  72. package/dist/esm/models/components/appendagenttrialsrequest.d.ts +8 -0
  73. package/dist/esm/models/components/appendagenttrialsrequest.d.ts.map +1 -1
  74. package/dist/esm/models/components/appendagenttrialsrequest.js +4 -0
  75. package/dist/esm/models/components/appendagenttrialsrequest.js.map +1 -1
  76. package/dist/esm/models/components/computelog.d.ts +2 -2
  77. package/dist/esm/models/components/evalleaderboardrow.d.ts +35 -0
  78. package/dist/esm/models/components/evalleaderboardrow.d.ts.map +1 -1
  79. package/dist/esm/models/components/evalleaderboardrow.js +26 -0
  80. package/dist/esm/models/components/evalleaderboardrow.js.map +1 -1
  81. package/dist/esm/models/components/evalrunoverview.d.ts +5 -0
  82. package/dist/esm/models/components/evalrunoverview.d.ts.map +1 -1
  83. package/dist/esm/models/components/evalrunoverview.js +2 -0
  84. package/dist/esm/models/components/evalrunoverview.js.map +1 -1
  85. package/dist/esm/models/components/evalsampleobject.d.ts +5 -0
  86. package/dist/esm/models/components/evalsampleobject.d.ts.map +1 -1
  87. package/dist/esm/models/components/evalsampleobject.js +4 -0
  88. package/dist/esm/models/components/evalsampleobject.js.map +1 -1
  89. package/dist/esm/models/components/importagentrunrequest.d.ts +8 -0
  90. package/dist/esm/models/components/importagentrunrequest.d.ts.map +1 -1
  91. package/dist/esm/models/components/importagentrunrequest.js +4 -0
  92. package/dist/esm/models/components/importagentrunrequest.js.map +1 -1
  93. package/dist/esm/models/operations/appendagentevaltrials.d.ts +36 -0
  94. package/dist/esm/models/operations/appendagentevaltrials.d.ts.map +1 -1
  95. package/dist/esm/models/operations/appendagentevaltrials.js +22 -0
  96. package/dist/esm/models/operations/appendagentevaltrials.js.map +1 -1
  97. package/dist/esm/models/operations/cancelevalrun.d.ts +36 -0
  98. package/dist/esm/models/operations/cancelevalrun.d.ts.map +1 -1
  99. package/dist/esm/models/operations/cancelevalrun.js +22 -0
  100. package/dist/esm/models/operations/cancelevalrun.js.map +1 -1
  101. package/dist/esm/models/operations/createevalrun.d.ts +36 -0
  102. package/dist/esm/models/operations/createevalrun.d.ts.map +1 -1
  103. package/dist/esm/models/operations/createevalrun.js +23 -0
  104. package/dist/esm/models/operations/createevalrun.js.map +1 -1
  105. package/dist/esm/models/operations/getevalrun.d.ts +36 -0
  106. package/dist/esm/models/operations/getevalrun.d.ts.map +1 -1
  107. package/dist/esm/models/operations/getevalrun.js +22 -0
  108. package/dist/esm/models/operations/getevalrun.js.map +1 -1
  109. package/dist/esm/models/operations/importagentevalrun.d.ts +36 -0
  110. package/dist/esm/models/operations/importagentevalrun.d.ts.map +1 -1
  111. package/dist/esm/models/operations/importagentevalrun.js +22 -0
  112. package/dist/esm/models/operations/importagentevalrun.js.map +1 -1
  113. package/dist/esm/models/operations/listevalruns.d.ts +36 -0
  114. package/dist/esm/models/operations/listevalruns.d.ts.map +1 -1
  115. package/dist/esm/models/operations/listevalruns.js +22 -0
  116. package/dist/esm/models/operations/listevalruns.js.map +1 -1
  117. package/dist/esm/models/operations/listevalscheduleruns.d.ts +36 -0
  118. package/dist/esm/models/operations/listevalscheduleruns.d.ts.map +1 -1
  119. package/dist/esm/models/operations/listevalscheduleruns.js +22 -0
  120. package/dist/esm/models/operations/listevalscheduleruns.js.map +1 -1
  121. package/dist/esm/models/operations/listevalsuiteleaderboardruns.d.ts +36 -0
  122. package/dist/esm/models/operations/listevalsuiteleaderboardruns.d.ts.map +1 -1
  123. package/dist/esm/models/operations/listevalsuiteleaderboardruns.js +22 -0
  124. package/dist/esm/models/operations/listevalsuiteleaderboardruns.js.map +1 -1
  125. package/dist/esm/models/operations/rerunfailedevalsamples.d.ts +36 -0
  126. package/dist/esm/models/operations/rerunfailedevalsamples.d.ts.map +1 -1
  127. package/dist/esm/models/operations/rerunfailedevalsamples.js +22 -0
  128. package/dist/esm/models/operations/rerunfailedevalsamples.js.map +1 -1
  129. package/dist/esm/models/operations/retryfailedevalrun.d.ts +36 -0
  130. package/dist/esm/models/operations/retryfailedevalrun.d.ts.map +1 -1
  131. package/dist/esm/models/operations/retryfailedevalrun.js +22 -0
  132. package/dist/esm/models/operations/retryfailedevalrun.js.map +1 -1
  133. package/jsr.json +1 -1
  134. package/package.json +1 -1
  135. package/src/lib/config.ts +2 -2
  136. package/src/mcp-server/mcp-server.ts +1 -1
  137. package/src/mcp-server/server.ts +1 -1
  138. package/src/models/components/appendagenttrialsrequest.ts +12 -0
  139. package/src/models/components/computelog.ts +2 -2
  140. package/src/models/components/evalleaderboardrow.ts +61 -0
  141. package/src/models/components/evalrunoverview.ts +7 -0
  142. package/src/models/components/evalsampleobject.ts +9 -0
  143. package/src/models/components/importagentrunrequest.ts +12 -0
  144. package/src/models/operations/appendagentevaltrials.ts +76 -0
  145. package/src/models/operations/cancelevalrun.ts +74 -0
  146. package/src/models/operations/createevalrun.ts +67 -0
  147. package/src/models/operations/getevalrun.ts +74 -0
  148. package/src/models/operations/importagentevalrun.ts +74 -0
  149. package/src/models/operations/listevalruns.ts +74 -0
  150. package/src/models/operations/listevalscheduleruns.ts +74 -0
  151. package/src/models/operations/listevalsuiteleaderboardruns.ts +77 -0
  152. package/src/models/operations/rerunfailedevalsamples.ts +76 -0
  153. package/src/models/operations/retryfailedevalrun.ts +74 -0
@@ -83,6 +83,25 @@ export type GetEvalRunRequestCounts = {
83
83
  failed: number;
84
84
  };
85
85
 
86
+ /**
87
+ * Token totals for the run, summed over its samples when it finalized. Each is null when nothing reported that figure, which is a different fact from zero and must not render as free.
88
+ */
89
+ export type GetEvalRunTokens = {
90
+ /**
91
+ * Prompt tokens over the run's samples, `cached` included.
92
+ */
93
+ input: number | null;
94
+ output: number | null;
95
+ /**
96
+ * The share of `input` the provider served from its prompt cache. Can exceed `input` on an older harness import, where the two numbers came from different accountings — treat the pair as inconsistent rather than deriving a hit rate from it.
97
+ */
98
+ cached: number | null;
99
+ /**
100
+ * Reasoning tokens inside `output`, summed over the samples whose provider reported the split. Null when none did; the per-task figures in `metrics.by_task` estimate the split from the stored text in that case.
101
+ */
102
+ reasoning: number | null;
103
+ };
104
+
86
105
  /**
87
106
  * OK
88
107
  */
@@ -127,6 +146,10 @@ export type GetEvalRunResponseBody = {
127
146
  * Total spend for the run in USD. Null when the harness did not report cost.
128
147
  */
129
148
  costUsd: number | null;
149
+ /**
150
+ * Token totals for the run, summed over its samples when it finalized. Each is null when nothing reported that figure, which is a different fact from zero and must not render as free.
151
+ */
152
+ tokens: GetEvalRunTokens;
130
153
  createdAt: number;
131
154
  startedAt: number | null;
132
155
  completedAt: number | null;
@@ -487,6 +510,54 @@ export function getEvalRunRequestCountsFromJSON(
487
510
  );
488
511
  }
489
512
 
513
+ /** @internal */
514
+ export const GetEvalRunTokens$inboundSchema: z.ZodType<
515
+ GetEvalRunTokens,
516
+ z.ZodTypeDef,
517
+ unknown
518
+ > = z.object({
519
+ input: z.nullable(z.number().int()),
520
+ output: z.nullable(z.number().int()),
521
+ cached: z.nullable(z.number().int()),
522
+ reasoning: z.nullable(z.number().int()),
523
+ });
524
+ /** @internal */
525
+ export type GetEvalRunTokens$Outbound = {
526
+ input: number | null;
527
+ output: number | null;
528
+ cached: number | null;
529
+ reasoning: number | null;
530
+ };
531
+
532
+ /** @internal */
533
+ export const GetEvalRunTokens$outboundSchema: z.ZodType<
534
+ GetEvalRunTokens$Outbound,
535
+ z.ZodTypeDef,
536
+ GetEvalRunTokens
537
+ > = z.object({
538
+ input: z.nullable(z.number().int()),
539
+ output: z.nullable(z.number().int()),
540
+ cached: z.nullable(z.number().int()),
541
+ reasoning: z.nullable(z.number().int()),
542
+ });
543
+
544
+ export function getEvalRunTokensToJSON(
545
+ getEvalRunTokens: GetEvalRunTokens,
546
+ ): string {
547
+ return JSON.stringify(
548
+ GetEvalRunTokens$outboundSchema.parse(getEvalRunTokens),
549
+ );
550
+ }
551
+ export function getEvalRunTokensFromJSON(
552
+ jsonString: string,
553
+ ): SafeParseResult<GetEvalRunTokens, SDKValidationError> {
554
+ return safeParse(
555
+ jsonString,
556
+ (x) => GetEvalRunTokens$inboundSchema.parse(JSON.parse(x)),
557
+ `Failed to parse 'GetEvalRunTokens' from JSON`,
558
+ );
559
+ }
560
+
490
561
  /** @internal */
491
562
  export const GetEvalRunResponseBody$inboundSchema: z.ZodType<
492
563
  GetEvalRunResponseBody,
@@ -515,6 +586,7 @@ export const GetEvalRunResponseBody$inboundSchema: z.ZodType<
515
586
  agent_version: z.nullable(z.string()),
516
587
  agent_effort: z.nullable(z.string()),
517
588
  cost_usd: z.nullable(z.number()),
589
+ tokens: z.lazy(() => GetEvalRunTokens$inboundSchema),
518
590
  created_at: z.number().int(),
519
591
  started_at: z.nullable(z.number().int()),
520
592
  completed_at: z.nullable(z.number().int()),
@@ -567,6 +639,7 @@ export type GetEvalRunResponseBody$Outbound = {
567
639
  agent_version: string | null;
568
640
  agent_effort: string | null;
569
641
  cost_usd: number | null;
642
+ tokens: GetEvalRunTokens$Outbound;
570
643
  created_at: number;
571
644
  started_at: number | null;
572
645
  completed_at: number | null;
@@ -602,6 +675,7 @@ export const GetEvalRunResponseBody$outboundSchema: z.ZodType<
602
675
  agentVersion: z.nullable(z.string()),
603
676
  agentEffort: z.nullable(z.string()),
604
677
  costUsd: z.nullable(z.number()),
678
+ tokens: z.lazy(() => GetEvalRunTokens$outboundSchema),
605
679
  createdAt: z.number().int(),
606
680
  startedAt: z.nullable(z.number().int()),
607
681
  completedAt: z.nullable(z.number().int()),
@@ -87,6 +87,25 @@ export type ImportAgentEvalRunRequestCounts = {
87
87
  failed: number;
88
88
  };
89
89
 
90
+ /**
91
+ * Token totals for the run, summed over its samples when it finalized. Each is null when nothing reported that figure, which is a different fact from zero and must not render as free.
92
+ */
93
+ export type ImportAgentEvalRunTokens = {
94
+ /**
95
+ * Prompt tokens over the run's samples, `cached` included.
96
+ */
97
+ input: number | null;
98
+ output: number | null;
99
+ /**
100
+ * The share of `input` the provider served from its prompt cache. Can exceed `input` on an older harness import, where the two numbers came from different accountings — treat the pair as inconsistent rather than deriving a hit rate from it.
101
+ */
102
+ cached: number | null;
103
+ /**
104
+ * Reasoning tokens inside `output`, summed over the samples whose provider reported the split. Null when none did; the per-task figures in `metrics.by_task` estimate the split from the stored text in that case.
105
+ */
106
+ reasoning: number | null;
107
+ };
108
+
90
109
  /**
91
110
  * OK
92
111
  */
@@ -131,6 +150,10 @@ export type ImportAgentEvalRunResponseBody = {
131
150
  * Total spend for the run in USD. Null when the harness did not report cost.
132
151
  */
133
152
  costUsd: number | null;
153
+ /**
154
+ * Token totals for the run, summed over its samples when it finalized. Each is null when nothing reported that figure, which is a different fact from zero and must not render as free.
155
+ */
156
+ tokens: ImportAgentEvalRunTokens;
134
157
  createdAt: number;
135
158
  startedAt: number | null;
136
159
  completedAt: number | null;
@@ -501,6 +524,54 @@ export function importAgentEvalRunRequestCountsFromJSON(
501
524
  );
502
525
  }
503
526
 
527
+ /** @internal */
528
+ export const ImportAgentEvalRunTokens$inboundSchema: z.ZodType<
529
+ ImportAgentEvalRunTokens,
530
+ z.ZodTypeDef,
531
+ unknown
532
+ > = z.object({
533
+ input: z.nullable(z.number().int()),
534
+ output: z.nullable(z.number().int()),
535
+ cached: z.nullable(z.number().int()),
536
+ reasoning: z.nullable(z.number().int()),
537
+ });
538
+ /** @internal */
539
+ export type ImportAgentEvalRunTokens$Outbound = {
540
+ input: number | null;
541
+ output: number | null;
542
+ cached: number | null;
543
+ reasoning: number | null;
544
+ };
545
+
546
+ /** @internal */
547
+ export const ImportAgentEvalRunTokens$outboundSchema: z.ZodType<
548
+ ImportAgentEvalRunTokens$Outbound,
549
+ z.ZodTypeDef,
550
+ ImportAgentEvalRunTokens
551
+ > = z.object({
552
+ input: z.nullable(z.number().int()),
553
+ output: z.nullable(z.number().int()),
554
+ cached: z.nullable(z.number().int()),
555
+ reasoning: z.nullable(z.number().int()),
556
+ });
557
+
558
+ export function importAgentEvalRunTokensToJSON(
559
+ importAgentEvalRunTokens: ImportAgentEvalRunTokens,
560
+ ): string {
561
+ return JSON.stringify(
562
+ ImportAgentEvalRunTokens$outboundSchema.parse(importAgentEvalRunTokens),
563
+ );
564
+ }
565
+ export function importAgentEvalRunTokensFromJSON(
566
+ jsonString: string,
567
+ ): SafeParseResult<ImportAgentEvalRunTokens, SDKValidationError> {
568
+ return safeParse(
569
+ jsonString,
570
+ (x) => ImportAgentEvalRunTokens$inboundSchema.parse(JSON.parse(x)),
571
+ `Failed to parse 'ImportAgentEvalRunTokens' from JSON`,
572
+ );
573
+ }
574
+
504
575
  /** @internal */
505
576
  export const ImportAgentEvalRunResponseBody$inboundSchema: z.ZodType<
506
577
  ImportAgentEvalRunResponseBody,
@@ -529,6 +600,7 @@ export const ImportAgentEvalRunResponseBody$inboundSchema: z.ZodType<
529
600
  agent_version: z.nullable(z.string()),
530
601
  agent_effort: z.nullable(z.string()),
531
602
  cost_usd: z.nullable(z.number()),
603
+ tokens: z.lazy(() => ImportAgentEvalRunTokens$inboundSchema),
532
604
  created_at: z.number().int(),
533
605
  started_at: z.nullable(z.number().int()),
534
606
  completed_at: z.nullable(z.number().int()),
@@ -581,6 +653,7 @@ export type ImportAgentEvalRunResponseBody$Outbound = {
581
653
  agent_version: string | null;
582
654
  agent_effort: string | null;
583
655
  cost_usd: number | null;
656
+ tokens: ImportAgentEvalRunTokens$Outbound;
584
657
  created_at: number;
585
658
  started_at: number | null;
586
659
  completed_at: number | null;
@@ -616,6 +689,7 @@ export const ImportAgentEvalRunResponseBody$outboundSchema: z.ZodType<
616
689
  agentVersion: z.nullable(z.string()),
617
690
  agentEffort: z.nullable(z.string()),
618
691
  costUsd: z.nullable(z.number()),
692
+ tokens: z.lazy(() => ImportAgentEvalRunTokens$outboundSchema),
619
693
  createdAt: z.number().int(),
620
694
  startedAt: z.nullable(z.number().int()),
621
695
  completedAt: z.nullable(z.number().int()),
@@ -102,6 +102,25 @@ export type ListEvalRunsRequestCounts = {
102
102
  failed: number;
103
103
  };
104
104
 
105
+ /**
106
+ * Token totals for the run, summed over its samples when it finalized. Each is null when nothing reported that figure, which is a different fact from zero and must not render as free.
107
+ */
108
+ export type ListEvalRunsTokens = {
109
+ /**
110
+ * Prompt tokens over the run's samples, `cached` included.
111
+ */
112
+ input: number | null;
113
+ output: number | null;
114
+ /**
115
+ * The share of `input` the provider served from its prompt cache. Can exceed `input` on an older harness import, where the two numbers came from different accountings — treat the pair as inconsistent rather than deriving a hit rate from it.
116
+ */
117
+ cached: number | null;
118
+ /**
119
+ * Reasoning tokens inside `output`, summed over the samples whose provider reported the split. Null when none did; the per-task figures in `metrics.by_task` estimate the split from the stored text in that case.
120
+ */
121
+ reasoning: number | null;
122
+ };
123
+
105
124
  export type ListEvalRunsData = {
106
125
  id: string;
107
126
  object: "eval.run";
@@ -143,6 +162,10 @@ export type ListEvalRunsData = {
143
162
  * Total spend for the run in USD. Null when the harness did not report cost.
144
163
  */
145
164
  costUsd: number | null;
165
+ /**
166
+ * Token totals for the run, summed over its samples when it finalized. Each is null when nothing reported that figure, which is a different fact from zero and must not render as free.
167
+ */
168
+ tokens: ListEvalRunsTokens;
146
169
  createdAt: number;
147
170
  startedAt: number | null;
148
171
  completedAt: number | null;
@@ -541,6 +564,54 @@ export function listEvalRunsRequestCountsFromJSON(
541
564
  );
542
565
  }
543
566
 
567
+ /** @internal */
568
+ export const ListEvalRunsTokens$inboundSchema: z.ZodType<
569
+ ListEvalRunsTokens,
570
+ z.ZodTypeDef,
571
+ unknown
572
+ > = z.object({
573
+ input: z.nullable(z.number().int()),
574
+ output: z.nullable(z.number().int()),
575
+ cached: z.nullable(z.number().int()),
576
+ reasoning: z.nullable(z.number().int()),
577
+ });
578
+ /** @internal */
579
+ export type ListEvalRunsTokens$Outbound = {
580
+ input: number | null;
581
+ output: number | null;
582
+ cached: number | null;
583
+ reasoning: number | null;
584
+ };
585
+
586
+ /** @internal */
587
+ export const ListEvalRunsTokens$outboundSchema: z.ZodType<
588
+ ListEvalRunsTokens$Outbound,
589
+ z.ZodTypeDef,
590
+ ListEvalRunsTokens
591
+ > = z.object({
592
+ input: z.nullable(z.number().int()),
593
+ output: z.nullable(z.number().int()),
594
+ cached: z.nullable(z.number().int()),
595
+ reasoning: z.nullable(z.number().int()),
596
+ });
597
+
598
+ export function listEvalRunsTokensToJSON(
599
+ listEvalRunsTokens: ListEvalRunsTokens,
600
+ ): string {
601
+ return JSON.stringify(
602
+ ListEvalRunsTokens$outboundSchema.parse(listEvalRunsTokens),
603
+ );
604
+ }
605
+ export function listEvalRunsTokensFromJSON(
606
+ jsonString: string,
607
+ ): SafeParseResult<ListEvalRunsTokens, SDKValidationError> {
608
+ return safeParse(
609
+ jsonString,
610
+ (x) => ListEvalRunsTokens$inboundSchema.parse(JSON.parse(x)),
611
+ `Failed to parse 'ListEvalRunsTokens' from JSON`,
612
+ );
613
+ }
614
+
544
615
  /** @internal */
545
616
  export const ListEvalRunsData$inboundSchema: z.ZodType<
546
617
  ListEvalRunsData,
@@ -569,6 +640,7 @@ export const ListEvalRunsData$inboundSchema: z.ZodType<
569
640
  agent_version: z.nullable(z.string()),
570
641
  agent_effort: z.nullable(z.string()),
571
642
  cost_usd: z.nullable(z.number()),
643
+ tokens: z.lazy(() => ListEvalRunsTokens$inboundSchema),
572
644
  created_at: z.number().int(),
573
645
  started_at: z.nullable(z.number().int()),
574
646
  completed_at: z.nullable(z.number().int()),
@@ -621,6 +693,7 @@ export type ListEvalRunsData$Outbound = {
621
693
  agent_version: string | null;
622
694
  agent_effort: string | null;
623
695
  cost_usd: number | null;
696
+ tokens: ListEvalRunsTokens$Outbound;
624
697
  created_at: number;
625
698
  started_at: number | null;
626
699
  completed_at: number | null;
@@ -656,6 +729,7 @@ export const ListEvalRunsData$outboundSchema: z.ZodType<
656
729
  agentVersion: z.nullable(z.string()),
657
730
  agentEffort: z.nullable(z.string()),
658
731
  costUsd: z.nullable(z.number()),
732
+ tokens: z.lazy(() => ListEvalRunsTokens$outboundSchema),
659
733
  createdAt: z.number().int(),
660
734
  startedAt: z.nullable(z.number().int()),
661
735
  completedAt: z.nullable(z.number().int()),
@@ -105,6 +105,25 @@ export type ListEvalScheduleRunsRequestCounts = {
105
105
  failed: number;
106
106
  };
107
107
 
108
+ /**
109
+ * Token totals for the run, summed over its samples when it finalized. Each is null when nothing reported that figure, which is a different fact from zero and must not render as free.
110
+ */
111
+ export type ListEvalScheduleRunsTokens = {
112
+ /**
113
+ * Prompt tokens over the run's samples, `cached` included.
114
+ */
115
+ input: number | null;
116
+ output: number | null;
117
+ /**
118
+ * The share of `input` the provider served from its prompt cache. Can exceed `input` on an older harness import, where the two numbers came from different accountings — treat the pair as inconsistent rather than deriving a hit rate from it.
119
+ */
120
+ cached: number | null;
121
+ /**
122
+ * Reasoning tokens inside `output`, summed over the samples whose provider reported the split. Null when none did; the per-task figures in `metrics.by_task` estimate the split from the stored text in that case.
123
+ */
124
+ reasoning: number | null;
125
+ };
126
+
108
127
  export type ListEvalScheduleRunsData = {
109
128
  id: string;
110
129
  object: "eval.run";
@@ -146,6 +165,10 @@ export type ListEvalScheduleRunsData = {
146
165
  * Total spend for the run in USD. Null when the harness did not report cost.
147
166
  */
148
167
  costUsd: number | null;
168
+ /**
169
+ * Token totals for the run, summed over its samples when it finalized. Each is null when nothing reported that figure, which is a different fact from zero and must not render as free.
170
+ */
171
+ tokens: ListEvalScheduleRunsTokens;
149
172
  createdAt: number;
150
173
  startedAt: number | null;
151
174
  completedAt: number | null;
@@ -558,6 +581,54 @@ export function listEvalScheduleRunsRequestCountsFromJSON(
558
581
  );
559
582
  }
560
583
 
584
+ /** @internal */
585
+ export const ListEvalScheduleRunsTokens$inboundSchema: z.ZodType<
586
+ ListEvalScheduleRunsTokens,
587
+ z.ZodTypeDef,
588
+ unknown
589
+ > = z.object({
590
+ input: z.nullable(z.number().int()),
591
+ output: z.nullable(z.number().int()),
592
+ cached: z.nullable(z.number().int()),
593
+ reasoning: z.nullable(z.number().int()),
594
+ });
595
+ /** @internal */
596
+ export type ListEvalScheduleRunsTokens$Outbound = {
597
+ input: number | null;
598
+ output: number | null;
599
+ cached: number | null;
600
+ reasoning: number | null;
601
+ };
602
+
603
+ /** @internal */
604
+ export const ListEvalScheduleRunsTokens$outboundSchema: z.ZodType<
605
+ ListEvalScheduleRunsTokens$Outbound,
606
+ z.ZodTypeDef,
607
+ ListEvalScheduleRunsTokens
608
+ > = z.object({
609
+ input: z.nullable(z.number().int()),
610
+ output: z.nullable(z.number().int()),
611
+ cached: z.nullable(z.number().int()),
612
+ reasoning: z.nullable(z.number().int()),
613
+ });
614
+
615
+ export function listEvalScheduleRunsTokensToJSON(
616
+ listEvalScheduleRunsTokens: ListEvalScheduleRunsTokens,
617
+ ): string {
618
+ return JSON.stringify(
619
+ ListEvalScheduleRunsTokens$outboundSchema.parse(listEvalScheduleRunsTokens),
620
+ );
621
+ }
622
+ export function listEvalScheduleRunsTokensFromJSON(
623
+ jsonString: string,
624
+ ): SafeParseResult<ListEvalScheduleRunsTokens, SDKValidationError> {
625
+ return safeParse(
626
+ jsonString,
627
+ (x) => ListEvalScheduleRunsTokens$inboundSchema.parse(JSON.parse(x)),
628
+ `Failed to parse 'ListEvalScheduleRunsTokens' from JSON`,
629
+ );
630
+ }
631
+
561
632
  /** @internal */
562
633
  export const ListEvalScheduleRunsData$inboundSchema: z.ZodType<
563
634
  ListEvalScheduleRunsData,
@@ -586,6 +657,7 @@ export const ListEvalScheduleRunsData$inboundSchema: z.ZodType<
586
657
  agent_version: z.nullable(z.string()),
587
658
  agent_effort: z.nullable(z.string()),
588
659
  cost_usd: z.nullable(z.number()),
660
+ tokens: z.lazy(() => ListEvalScheduleRunsTokens$inboundSchema),
589
661
  created_at: z.number().int(),
590
662
  started_at: z.nullable(z.number().int()),
591
663
  completed_at: z.nullable(z.number().int()),
@@ -638,6 +710,7 @@ export type ListEvalScheduleRunsData$Outbound = {
638
710
  agent_version: string | null;
639
711
  agent_effort: string | null;
640
712
  cost_usd: number | null;
713
+ tokens: ListEvalScheduleRunsTokens$Outbound;
641
714
  created_at: number;
642
715
  started_at: number | null;
643
716
  completed_at: number | null;
@@ -673,6 +746,7 @@ export const ListEvalScheduleRunsData$outboundSchema: z.ZodType<
673
746
  agentVersion: z.nullable(z.string()),
674
747
  agentEffort: z.nullable(z.string()),
675
748
  costUsd: z.nullable(z.number()),
749
+ tokens: z.lazy(() => ListEvalScheduleRunsTokens$outboundSchema),
676
750
  createdAt: z.number().int(),
677
751
  startedAt: z.nullable(z.number().int()),
678
752
  completedAt: z.nullable(z.number().int()),
@@ -91,6 +91,25 @@ export type ListEvalSuiteLeaderboardRunsRequestCounts = {
91
91
  failed: number;
92
92
  };
93
93
 
94
+ /**
95
+ * Token totals for the run, summed over its samples when it finalized. Each is null when nothing reported that figure, which is a different fact from zero and must not render as free.
96
+ */
97
+ export type ListEvalSuiteLeaderboardRunsTokens = {
98
+ /**
99
+ * Prompt tokens over the run's samples, `cached` included.
100
+ */
101
+ input: number | null;
102
+ output: number | null;
103
+ /**
104
+ * The share of `input` the provider served from its prompt cache. Can exceed `input` on an older harness import, where the two numbers came from different accountings — treat the pair as inconsistent rather than deriving a hit rate from it.
105
+ */
106
+ cached: number | null;
107
+ /**
108
+ * Reasoning tokens inside `output`, summed over the samples whose provider reported the split. Null when none did; the per-task figures in `metrics.by_task` estimate the split from the stored text in that case.
109
+ */
110
+ reasoning: number | null;
111
+ };
112
+
94
113
  export type ListEvalSuiteLeaderboardRunsData = {
95
114
  id: string;
96
115
  object: "eval.run";
@@ -132,6 +151,10 @@ export type ListEvalSuiteLeaderboardRunsData = {
132
151
  * Total spend for the run in USD. Null when the harness did not report cost.
133
152
  */
134
153
  costUsd: number | null;
154
+ /**
155
+ * Token totals for the run, summed over its samples when it finalized. Each is null when nothing reported that figure, which is a different fact from zero and must not render as free.
156
+ */
157
+ tokens: ListEvalSuiteLeaderboardRunsTokens;
135
158
  createdAt: number;
136
159
  startedAt: number | null;
137
160
  completedAt: number | null;
@@ -568,6 +591,57 @@ export function listEvalSuiteLeaderboardRunsRequestCountsFromJSON(
568
591
  );
569
592
  }
570
593
 
594
+ /** @internal */
595
+ export const ListEvalSuiteLeaderboardRunsTokens$inboundSchema: z.ZodType<
596
+ ListEvalSuiteLeaderboardRunsTokens,
597
+ z.ZodTypeDef,
598
+ unknown
599
+ > = z.object({
600
+ input: z.nullable(z.number().int()),
601
+ output: z.nullable(z.number().int()),
602
+ cached: z.nullable(z.number().int()),
603
+ reasoning: z.nullable(z.number().int()),
604
+ });
605
+ /** @internal */
606
+ export type ListEvalSuiteLeaderboardRunsTokens$Outbound = {
607
+ input: number | null;
608
+ output: number | null;
609
+ cached: number | null;
610
+ reasoning: number | null;
611
+ };
612
+
613
+ /** @internal */
614
+ export const ListEvalSuiteLeaderboardRunsTokens$outboundSchema: z.ZodType<
615
+ ListEvalSuiteLeaderboardRunsTokens$Outbound,
616
+ z.ZodTypeDef,
617
+ ListEvalSuiteLeaderboardRunsTokens
618
+ > = z.object({
619
+ input: z.nullable(z.number().int()),
620
+ output: z.nullable(z.number().int()),
621
+ cached: z.nullable(z.number().int()),
622
+ reasoning: z.nullable(z.number().int()),
623
+ });
624
+
625
+ export function listEvalSuiteLeaderboardRunsTokensToJSON(
626
+ listEvalSuiteLeaderboardRunsTokens: ListEvalSuiteLeaderboardRunsTokens,
627
+ ): string {
628
+ return JSON.stringify(
629
+ ListEvalSuiteLeaderboardRunsTokens$outboundSchema.parse(
630
+ listEvalSuiteLeaderboardRunsTokens,
631
+ ),
632
+ );
633
+ }
634
+ export function listEvalSuiteLeaderboardRunsTokensFromJSON(
635
+ jsonString: string,
636
+ ): SafeParseResult<ListEvalSuiteLeaderboardRunsTokens, SDKValidationError> {
637
+ return safeParse(
638
+ jsonString,
639
+ (x) =>
640
+ ListEvalSuiteLeaderboardRunsTokens$inboundSchema.parse(JSON.parse(x)),
641
+ `Failed to parse 'ListEvalSuiteLeaderboardRunsTokens' from JSON`,
642
+ );
643
+ }
644
+
571
645
  /** @internal */
572
646
  export const ListEvalSuiteLeaderboardRunsData$inboundSchema: z.ZodType<
573
647
  ListEvalSuiteLeaderboardRunsData,
@@ -600,6 +674,7 @@ export const ListEvalSuiteLeaderboardRunsData$inboundSchema: z.ZodType<
600
674
  agent_version: z.nullable(z.string()),
601
675
  agent_effort: z.nullable(z.string()),
602
676
  cost_usd: z.nullable(z.number()),
677
+ tokens: z.lazy(() => ListEvalSuiteLeaderboardRunsTokens$inboundSchema),
603
678
  created_at: z.number().int(),
604
679
  started_at: z.nullable(z.number().int()),
605
680
  completed_at: z.nullable(z.number().int()),
@@ -652,6 +727,7 @@ export type ListEvalSuiteLeaderboardRunsData$Outbound = {
652
727
  agent_version: string | null;
653
728
  agent_effort: string | null;
654
729
  cost_usd: number | null;
730
+ tokens: ListEvalSuiteLeaderboardRunsTokens$Outbound;
655
731
  created_at: number;
656
732
  started_at: number | null;
657
733
  completed_at: number | null;
@@ -691,6 +767,7 @@ export const ListEvalSuiteLeaderboardRunsData$outboundSchema: z.ZodType<
691
767
  agentVersion: z.nullable(z.string()),
692
768
  agentEffort: z.nullable(z.string()),
693
769
  costUsd: z.nullable(z.number()),
770
+ tokens: z.lazy(() => ListEvalSuiteLeaderboardRunsTokens$outboundSchema),
694
771
  createdAt: z.number().int(),
695
772
  startedAt: z.nullable(z.number().int()),
696
773
  completedAt: z.nullable(z.number().int()),
@@ -87,6 +87,25 @@ export type RerunFailedEvalSamplesRequestCounts = {
87
87
  failed: number;
88
88
  };
89
89
 
90
+ /**
91
+ * Token totals for the run, summed over its samples when it finalized. Each is null when nothing reported that figure, which is a different fact from zero and must not render as free.
92
+ */
93
+ export type RerunFailedEvalSamplesTokens = {
94
+ /**
95
+ * Prompt tokens over the run's samples, `cached` included.
96
+ */
97
+ input: number | null;
98
+ output: number | null;
99
+ /**
100
+ * The share of `input` the provider served from its prompt cache. Can exceed `input` on an older harness import, where the two numbers came from different accountings — treat the pair as inconsistent rather than deriving a hit rate from it.
101
+ */
102
+ cached: number | null;
103
+ /**
104
+ * Reasoning tokens inside `output`, summed over the samples whose provider reported the split. Null when none did; the per-task figures in `metrics.by_task` estimate the split from the stored text in that case.
105
+ */
106
+ reasoning: number | null;
107
+ };
108
+
90
109
  /**
91
110
  * OK
92
111
  */
@@ -131,6 +150,10 @@ export type RerunFailedEvalSamplesResponseBody = {
131
150
  * Total spend for the run in USD. Null when the harness did not report cost.
132
151
  */
133
152
  costUsd: number | null;
153
+ /**
154
+ * Token totals for the run, summed over its samples when it finalized. Each is null when nothing reported that figure, which is a different fact from zero and must not render as free.
155
+ */
156
+ tokens: RerunFailedEvalSamplesTokens;
134
157
  createdAt: number;
135
158
  startedAt: number | null;
136
159
  completedAt: number | null;
@@ -505,6 +528,56 @@ export function rerunFailedEvalSamplesRequestCountsFromJSON(
505
528
  );
506
529
  }
507
530
 
531
+ /** @internal */
532
+ export const RerunFailedEvalSamplesTokens$inboundSchema: z.ZodType<
533
+ RerunFailedEvalSamplesTokens,
534
+ z.ZodTypeDef,
535
+ unknown
536
+ > = z.object({
537
+ input: z.nullable(z.number().int()),
538
+ output: z.nullable(z.number().int()),
539
+ cached: z.nullable(z.number().int()),
540
+ reasoning: z.nullable(z.number().int()),
541
+ });
542
+ /** @internal */
543
+ export type RerunFailedEvalSamplesTokens$Outbound = {
544
+ input: number | null;
545
+ output: number | null;
546
+ cached: number | null;
547
+ reasoning: number | null;
548
+ };
549
+
550
+ /** @internal */
551
+ export const RerunFailedEvalSamplesTokens$outboundSchema: z.ZodType<
552
+ RerunFailedEvalSamplesTokens$Outbound,
553
+ z.ZodTypeDef,
554
+ RerunFailedEvalSamplesTokens
555
+ > = z.object({
556
+ input: z.nullable(z.number().int()),
557
+ output: z.nullable(z.number().int()),
558
+ cached: z.nullable(z.number().int()),
559
+ reasoning: z.nullable(z.number().int()),
560
+ });
561
+
562
+ export function rerunFailedEvalSamplesTokensToJSON(
563
+ rerunFailedEvalSamplesTokens: RerunFailedEvalSamplesTokens,
564
+ ): string {
565
+ return JSON.stringify(
566
+ RerunFailedEvalSamplesTokens$outboundSchema.parse(
567
+ rerunFailedEvalSamplesTokens,
568
+ ),
569
+ );
570
+ }
571
+ export function rerunFailedEvalSamplesTokensFromJSON(
572
+ jsonString: string,
573
+ ): SafeParseResult<RerunFailedEvalSamplesTokens, SDKValidationError> {
574
+ return safeParse(
575
+ jsonString,
576
+ (x) => RerunFailedEvalSamplesTokens$inboundSchema.parse(JSON.parse(x)),
577
+ `Failed to parse 'RerunFailedEvalSamplesTokens' from JSON`,
578
+ );
579
+ }
580
+
508
581
  /** @internal */
509
582
  export const RerunFailedEvalSamplesResponseBody$inboundSchema: z.ZodType<
510
583
  RerunFailedEvalSamplesResponseBody,
@@ -535,6 +608,7 @@ export const RerunFailedEvalSamplesResponseBody$inboundSchema: z.ZodType<
535
608
  agent_version: z.nullable(z.string()),
536
609
  agent_effort: z.nullable(z.string()),
537
610
  cost_usd: z.nullable(z.number()),
611
+ tokens: z.lazy(() => RerunFailedEvalSamplesTokens$inboundSchema),
538
612
  created_at: z.number().int(),
539
613
  started_at: z.nullable(z.number().int()),
540
614
  completed_at: z.nullable(z.number().int()),
@@ -587,6 +661,7 @@ export type RerunFailedEvalSamplesResponseBody$Outbound = {
587
661
  agent_version: string | null;
588
662
  agent_effort: string | null;
589
663
  cost_usd: number | null;
664
+ tokens: RerunFailedEvalSamplesTokens$Outbound;
590
665
  created_at: number;
591
666
  started_at: number | null;
592
667
  completed_at: number | null;
@@ -624,6 +699,7 @@ export const RerunFailedEvalSamplesResponseBody$outboundSchema: z.ZodType<
624
699
  agentVersion: z.nullable(z.string()),
625
700
  agentEffort: z.nullable(z.string()),
626
701
  costUsd: z.nullable(z.number()),
702
+ tokens: z.lazy(() => RerunFailedEvalSamplesTokens$outboundSchema),
627
703
  createdAt: z.number().int(),
628
704
  startedAt: z.nullable(z.number().int()),
629
705
  completedAt: z.nullable(z.number().int()),