@meetkai/mka1 0.55.33 → 0.55.34

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 (170) hide show
  1. package/bin/mcp-server.js +122 -10
  2. package/bin/mcp-server.js.map +23 -21
  3. package/dist/commonjs/funcs/llmEvalsCreateRun.d.ts +1 -1
  4. package/dist/commonjs/funcs/llmEvalsCreateRun.js +1 -1
  5. package/dist/commonjs/lib/config.d.ts +2 -2
  6. package/dist/commonjs/lib/config.js +2 -2
  7. package/dist/commonjs/mcp-server/mcp-server.js +1 -1
  8. package/dist/commonjs/mcp-server/server.js +1 -1
  9. package/dist/commonjs/mcp-server/tools/llmEvalsCreateRun.js +1 -1
  10. package/dist/commonjs/mcp-server/tools/llmEvalsCreateRun.js.map +1 -1
  11. package/dist/commonjs/models/components/createevalrunrequest.d.ts +8 -2
  12. package/dist/commonjs/models/components/createevalrunrequest.d.ts.map +1 -1
  13. package/dist/commonjs/models/components/createevalrunrequest.js +5 -2
  14. package/dist/commonjs/models/components/createevalrunrequest.js.map +1 -1
  15. package/dist/commonjs/models/components/evalcandidate.d.ts +45 -0
  16. package/dist/commonjs/models/components/evalcandidate.d.ts.map +1 -0
  17. package/dist/commonjs/models/components/evalcandidate.js +107 -0
  18. package/dist/commonjs/models/components/evalcandidate.js.map +1 -0
  19. package/dist/commonjs/models/components/evalcandidateobject.d.ts +46 -0
  20. package/dist/commonjs/models/components/evalcandidateobject.d.ts.map +1 -0
  21. package/dist/commonjs/models/components/evalcandidateobject.js +108 -0
  22. package/dist/commonjs/models/components/evalcandidateobject.js.map +1 -0
  23. package/dist/commonjs/models/components/evalmetric.d.ts +5 -0
  24. package/dist/commonjs/models/components/evalmetric.d.ts.map +1 -1
  25. package/dist/commonjs/models/components/evalmetric.js +4 -0
  26. package/dist/commonjs/models/components/evalmetric.js.map +1 -1
  27. package/dist/commonjs/models/components/index.d.ts +2 -0
  28. package/dist/commonjs/models/components/index.d.ts.map +1 -1
  29. package/dist/commonjs/models/components/index.js +2 -0
  30. package/dist/commonjs/models/components/index.js.map +1 -1
  31. package/dist/commonjs/models/operations/appendagentevaltrials.d.ts +5 -0
  32. package/dist/commonjs/models/operations/appendagentevaltrials.d.ts.map +1 -1
  33. package/dist/commonjs/models/operations/appendagentevaltrials.js +2 -0
  34. package/dist/commonjs/models/operations/appendagentevaltrials.js.map +1 -1
  35. package/dist/commonjs/models/operations/cancelevalrun.d.ts +5 -0
  36. package/dist/commonjs/models/operations/cancelevalrun.d.ts.map +1 -1
  37. package/dist/commonjs/models/operations/cancelevalrun.js +2 -0
  38. package/dist/commonjs/models/operations/cancelevalrun.js.map +1 -1
  39. package/dist/commonjs/models/operations/createevalrun.d.ts +5 -0
  40. package/dist/commonjs/models/operations/createevalrun.d.ts.map +1 -1
  41. package/dist/commonjs/models/operations/createevalrun.js +2 -0
  42. package/dist/commonjs/models/operations/createevalrun.js.map +1 -1
  43. package/dist/commonjs/models/operations/getevalrun.d.ts +5 -0
  44. package/dist/commonjs/models/operations/getevalrun.d.ts.map +1 -1
  45. package/dist/commonjs/models/operations/getevalrun.js +2 -0
  46. package/dist/commonjs/models/operations/getevalrun.js.map +1 -1
  47. package/dist/commonjs/models/operations/importagentevalrun.d.ts +5 -0
  48. package/dist/commonjs/models/operations/importagentevalrun.d.ts.map +1 -1
  49. package/dist/commonjs/models/operations/importagentevalrun.js +2 -0
  50. package/dist/commonjs/models/operations/importagentevalrun.js.map +1 -1
  51. package/dist/commonjs/models/operations/listevalruns.d.ts +5 -0
  52. package/dist/commonjs/models/operations/listevalruns.d.ts.map +1 -1
  53. package/dist/commonjs/models/operations/listevalruns.js +2 -0
  54. package/dist/commonjs/models/operations/listevalruns.js.map +1 -1
  55. package/dist/commonjs/models/operations/listevalscheduleruns.d.ts +5 -0
  56. package/dist/commonjs/models/operations/listevalscheduleruns.d.ts.map +1 -1
  57. package/dist/commonjs/models/operations/listevalscheduleruns.js +2 -0
  58. package/dist/commonjs/models/operations/listevalscheduleruns.js.map +1 -1
  59. package/dist/commonjs/models/operations/listevalsuiteleaderboardruns.d.ts +5 -0
  60. package/dist/commonjs/models/operations/listevalsuiteleaderboardruns.d.ts.map +1 -1
  61. package/dist/commonjs/models/operations/listevalsuiteleaderboardruns.js +2 -0
  62. package/dist/commonjs/models/operations/listevalsuiteleaderboardruns.js.map +1 -1
  63. package/dist/commonjs/models/operations/rerunfailedevalsamples.d.ts +5 -0
  64. package/dist/commonjs/models/operations/rerunfailedevalsamples.d.ts.map +1 -1
  65. package/dist/commonjs/models/operations/rerunfailedevalsamples.js +2 -0
  66. package/dist/commonjs/models/operations/rerunfailedevalsamples.js.map +1 -1
  67. package/dist/commonjs/models/operations/retryfailedevalrun.d.ts +5 -0
  68. package/dist/commonjs/models/operations/retryfailedevalrun.d.ts.map +1 -1
  69. package/dist/commonjs/models/operations/retryfailedevalrun.js +2 -0
  70. package/dist/commonjs/models/operations/retryfailedevalrun.js.map +1 -1
  71. package/dist/commonjs/react-query/llmEvalsCreateRun.d.ts +1 -1
  72. package/dist/commonjs/react-query/llmEvalsCreateRun.js +1 -1
  73. package/dist/commonjs/sdk/evals.d.ts +1 -1
  74. package/dist/commonjs/sdk/evals.js +1 -1
  75. package/dist/esm/funcs/llmEvalsCreateRun.d.ts +1 -1
  76. package/dist/esm/funcs/llmEvalsCreateRun.js +1 -1
  77. package/dist/esm/lib/config.d.ts +2 -2
  78. package/dist/esm/lib/config.js +2 -2
  79. package/dist/esm/mcp-server/mcp-server.js +1 -1
  80. package/dist/esm/mcp-server/server.js +1 -1
  81. package/dist/esm/mcp-server/tools/llmEvalsCreateRun.js +1 -1
  82. package/dist/esm/mcp-server/tools/llmEvalsCreateRun.js.map +1 -1
  83. package/dist/esm/models/components/createevalrunrequest.d.ts +8 -2
  84. package/dist/esm/models/components/createevalrunrequest.d.ts.map +1 -1
  85. package/dist/esm/models/components/createevalrunrequest.js +5 -2
  86. package/dist/esm/models/components/createevalrunrequest.js.map +1 -1
  87. package/dist/esm/models/components/evalcandidate.d.ts +45 -0
  88. package/dist/esm/models/components/evalcandidate.d.ts.map +1 -0
  89. package/dist/esm/models/components/evalcandidate.js +65 -0
  90. package/dist/esm/models/components/evalcandidate.js.map +1 -0
  91. package/dist/esm/models/components/evalcandidateobject.d.ts +46 -0
  92. package/dist/esm/models/components/evalcandidateobject.d.ts.map +1 -0
  93. package/dist/esm/models/components/evalcandidateobject.js +66 -0
  94. package/dist/esm/models/components/evalcandidateobject.js.map +1 -0
  95. package/dist/esm/models/components/evalmetric.d.ts +5 -0
  96. package/dist/esm/models/components/evalmetric.d.ts.map +1 -1
  97. package/dist/esm/models/components/evalmetric.js +4 -0
  98. package/dist/esm/models/components/evalmetric.js.map +1 -1
  99. package/dist/esm/models/components/index.d.ts +2 -0
  100. package/dist/esm/models/components/index.d.ts.map +1 -1
  101. package/dist/esm/models/components/index.js +2 -0
  102. package/dist/esm/models/components/index.js.map +1 -1
  103. package/dist/esm/models/operations/appendagentevaltrials.d.ts +5 -0
  104. package/dist/esm/models/operations/appendagentevaltrials.d.ts.map +1 -1
  105. package/dist/esm/models/operations/appendagentevaltrials.js +2 -0
  106. package/dist/esm/models/operations/appendagentevaltrials.js.map +1 -1
  107. package/dist/esm/models/operations/cancelevalrun.d.ts +5 -0
  108. package/dist/esm/models/operations/cancelevalrun.d.ts.map +1 -1
  109. package/dist/esm/models/operations/cancelevalrun.js +2 -0
  110. package/dist/esm/models/operations/cancelevalrun.js.map +1 -1
  111. package/dist/esm/models/operations/createevalrun.d.ts +5 -0
  112. package/dist/esm/models/operations/createevalrun.d.ts.map +1 -1
  113. package/dist/esm/models/operations/createevalrun.js +2 -0
  114. package/dist/esm/models/operations/createevalrun.js.map +1 -1
  115. package/dist/esm/models/operations/getevalrun.d.ts +5 -0
  116. package/dist/esm/models/operations/getevalrun.d.ts.map +1 -1
  117. package/dist/esm/models/operations/getevalrun.js +2 -0
  118. package/dist/esm/models/operations/getevalrun.js.map +1 -1
  119. package/dist/esm/models/operations/importagentevalrun.d.ts +5 -0
  120. package/dist/esm/models/operations/importagentevalrun.d.ts.map +1 -1
  121. package/dist/esm/models/operations/importagentevalrun.js +2 -0
  122. package/dist/esm/models/operations/importagentevalrun.js.map +1 -1
  123. package/dist/esm/models/operations/listevalruns.d.ts +5 -0
  124. package/dist/esm/models/operations/listevalruns.d.ts.map +1 -1
  125. package/dist/esm/models/operations/listevalruns.js +2 -0
  126. package/dist/esm/models/operations/listevalruns.js.map +1 -1
  127. package/dist/esm/models/operations/listevalscheduleruns.d.ts +5 -0
  128. package/dist/esm/models/operations/listevalscheduleruns.d.ts.map +1 -1
  129. package/dist/esm/models/operations/listevalscheduleruns.js +2 -0
  130. package/dist/esm/models/operations/listevalscheduleruns.js.map +1 -1
  131. package/dist/esm/models/operations/listevalsuiteleaderboardruns.d.ts +5 -0
  132. package/dist/esm/models/operations/listevalsuiteleaderboardruns.d.ts.map +1 -1
  133. package/dist/esm/models/operations/listevalsuiteleaderboardruns.js +2 -0
  134. package/dist/esm/models/operations/listevalsuiteleaderboardruns.js.map +1 -1
  135. package/dist/esm/models/operations/rerunfailedevalsamples.d.ts +5 -0
  136. package/dist/esm/models/operations/rerunfailedevalsamples.d.ts.map +1 -1
  137. package/dist/esm/models/operations/rerunfailedevalsamples.js +2 -0
  138. package/dist/esm/models/operations/rerunfailedevalsamples.js.map +1 -1
  139. package/dist/esm/models/operations/retryfailedevalrun.d.ts +5 -0
  140. package/dist/esm/models/operations/retryfailedevalrun.d.ts.map +1 -1
  141. package/dist/esm/models/operations/retryfailedevalrun.js +2 -0
  142. package/dist/esm/models/operations/retryfailedevalrun.js.map +1 -1
  143. package/dist/esm/react-query/llmEvalsCreateRun.d.ts +1 -1
  144. package/dist/esm/react-query/llmEvalsCreateRun.js +1 -1
  145. package/dist/esm/sdk/evals.d.ts +1 -1
  146. package/dist/esm/sdk/evals.js +1 -1
  147. package/jsr.json +1 -1
  148. package/package.json +1 -1
  149. package/src/funcs/llmEvalsCreateRun.ts +1 -1
  150. package/src/lib/config.ts +2 -2
  151. package/src/mcp-server/mcp-server.ts +1 -1
  152. package/src/mcp-server/server.ts +1 -1
  153. package/src/mcp-server/tools/llmEvalsCreateRun.ts +1 -1
  154. package/src/models/components/createevalrunrequest.ts +17 -4
  155. package/src/models/components/evalcandidate.ts +146 -0
  156. package/src/models/components/evalcandidateobject.ts +163 -0
  157. package/src/models/components/evalmetric.ts +9 -0
  158. package/src/models/components/index.ts +2 -0
  159. package/src/models/operations/appendagentevaltrials.ts +7 -0
  160. package/src/models/operations/cancelevalrun.ts +7 -0
  161. package/src/models/operations/createevalrun.ts +7 -0
  162. package/src/models/operations/getevalrun.ts +7 -0
  163. package/src/models/operations/importagentevalrun.ts +7 -0
  164. package/src/models/operations/listevalruns.ts +7 -0
  165. package/src/models/operations/listevalscheduleruns.ts +7 -0
  166. package/src/models/operations/listevalsuiteleaderboardruns.ts +7 -0
  167. package/src/models/operations/rerunfailedevalsamples.ts +7 -0
  168. package/src/models/operations/retryfailedevalrun.ts +7 -0
  169. package/src/react-query/llmEvalsCreateRun.ts +1 -1
  170. package/src/sdk/evals.ts +1 -1
package/src/models/operations/createevalrun.ts CHANGED
@@ -121,6 +121,10 @@ export type CreateEvalRunResponseBody = {
121
121
  teamId: string;
122
122
  status: components.EvalRunStatus;
123
123
  models: Array<string>;
124
+ /**
125
+ * Typed candidates, including pinned saved-agent versions. models remains the legacy storage/filter key list.
126
+ */
127
+ candidates?: Array<components.EvalCandidateObject> | undefined;
124
128
  taskIds: Array<string> | null;
125
129
  judgeModel: string | null;
126
130
  embeddingModel: string | null;
@@ -544,6 +548,7 @@ export const CreateEvalRunResponseBody$inboundSchema: z.ZodType<
544
548
  team_id: z.string(),
545
549
  status: components.EvalRunStatus$inboundSchema,
546
550
  models: z.array(z.string()),
551
+ candidates: z.array(components.EvalCandidateObject$inboundSchema).optional(),
547
552
  task_ids: z.nullable(z.array(z.string())),
548
553
  judge_model: z.nullable(z.string()),
549
554
  embedding_model: z.nullable(z.string()),
@@ -597,6 +602,7 @@ export type CreateEvalRunResponseBody$Outbound = {
597
602
  team_id: string;
598
603
  status: string;
599
604
  models: Array<string>;
605
+ candidates?: Array<components.EvalCandidateObject$Outbound> | undefined;
600
606
  task_ids: Array<string> | null;
601
607
  judge_model: string | null;
602
608
  embedding_model: string | null;
@@ -633,6 +639,7 @@ export const CreateEvalRunResponseBody$outboundSchema: z.ZodType<
633
639
  teamId: z.string(),
634
640
  status: components.EvalRunStatus$outboundSchema,
635
641
  models: z.array(z.string()),
642
+ candidates: z.array(components.EvalCandidateObject$outboundSchema).optional(),
636
643
  taskIds: z.nullable(z.array(z.string())),
637
644
  judgeModel: z.nullable(z.string()),
638
645
  embeddingModel: z.nullable(z.string()),
package/src/models/operations/getevalrun.ts CHANGED
@@ -121,6 +121,10 @@ export type GetEvalRunResponseBody = {
121
121
  teamId: string;
122
122
  status: components.EvalRunStatus;
123
123
  models: Array<string>;
124
+ /**
125
+ * Typed candidates, including pinned saved-agent versions. models remains the legacy storage/filter key list.
126
+ */
127
+ candidates?: Array<components.EvalCandidateObject> | undefined;
124
128
  taskIds: Array<string> | null;
125
129
  judgeModel: string | null;
126
130
  embeddingModel: string | null;
@@ -573,6 +577,7 @@ export const GetEvalRunResponseBody$inboundSchema: z.ZodType<
573
577
  team_id: z.string(),
574
578
  status: components.EvalRunStatus$inboundSchema,
575
579
  models: z.array(z.string()),
580
+ candidates: z.array(components.EvalCandidateObject$inboundSchema).optional(),
576
581
  task_ids: z.nullable(z.array(z.string())),
577
582
  judge_model: z.nullable(z.string()),
578
583
  embedding_model: z.nullable(z.string()),
@@ -626,6 +631,7 @@ export type GetEvalRunResponseBody$Outbound = {
626
631
  team_id: string;
627
632
  status: string;
628
633
  models: Array<string>;
634
+ candidates?: Array<components.EvalCandidateObject$Outbound> | undefined;
629
635
  task_ids: Array<string> | null;
630
636
  judge_model: string | null;
631
637
  embedding_model: string | null;
@@ -662,6 +668,7 @@ export const GetEvalRunResponseBody$outboundSchema: z.ZodType<
662
668
  teamId: z.string(),
663
669
  status: components.EvalRunStatus$outboundSchema,
664
670
  models: z.array(z.string()),
671
+ candidates: z.array(components.EvalCandidateObject$outboundSchema).optional(),
665
672
  taskIds: z.nullable(z.array(z.string())),
666
673
  judgeModel: z.nullable(z.string()),
667
674
  embeddingModel: z.nullable(z.string()),
package/src/models/operations/importagentevalrun.ts CHANGED
@@ -125,6 +125,10 @@ export type ImportAgentEvalRunResponseBody = {
125
125
  teamId: string;
126
126
  status: components.EvalRunStatus;
127
127
  models: Array<string>;
128
+ /**
129
+ * Typed candidates, including pinned saved-agent versions. models remains the legacy storage/filter key list.
130
+ */
131
+ candidates?: Array<components.EvalCandidateObject> | undefined;
128
132
  taskIds: Array<string> | null;
129
133
  judgeModel: string | null;
130
134
  embeddingModel: string | null;
@@ -587,6 +591,7 @@ export const ImportAgentEvalRunResponseBody$inboundSchema: z.ZodType<
587
591
  team_id: z.string(),
588
592
  status: components.EvalRunStatus$inboundSchema,
589
593
  models: z.array(z.string()),
594
+ candidates: z.array(components.EvalCandidateObject$inboundSchema).optional(),
590
595
  task_ids: z.nullable(z.array(z.string())),
591
596
  judge_model: z.nullable(z.string()),
592
597
  embedding_model: z.nullable(z.string()),
@@ -640,6 +645,7 @@ export type ImportAgentEvalRunResponseBody$Outbound = {
640
645
  team_id: string;
641
646
  status: string;
642
647
  models: Array<string>;
648
+ candidates?: Array<components.EvalCandidateObject$Outbound> | undefined;
643
649
  task_ids: Array<string> | null;
644
650
  judge_model: string | null;
645
651
  embedding_model: string | null;
@@ -676,6 +682,7 @@ export const ImportAgentEvalRunResponseBody$outboundSchema: z.ZodType<
676
682
  teamId: z.string(),
677
683
  status: components.EvalRunStatus$outboundSchema,
678
684
  models: z.array(z.string()),
685
+ candidates: z.array(components.EvalCandidateObject$outboundSchema).optional(),
679
686
  taskIds: z.nullable(z.array(z.string())),
680
687
  judgeModel: z.nullable(z.string()),
681
688
  embeddingModel: z.nullable(z.string()),
package/src/models/operations/listevalruns.ts CHANGED
@@ -138,6 +138,10 @@ export type ListEvalRunsData = {
138
138
  teamId: string;
139
139
  status: components.EvalRunStatus;
140
140
  models: Array<string>;
141
+ /**
142
+ * Typed candidates, including pinned saved-agent versions. models remains the legacy storage/filter key list.
143
+ */
144
+ candidates?: Array<components.EvalCandidateObject> | undefined;
141
145
  taskIds: Array<string> | null;
142
146
  judgeModel: string | null;
143
147
  embeddingModel: string | null;
@@ -633,6 +637,7 @@ export const ListEvalRunsData$inboundSchema: z.ZodType<
633
637
  team_id: z.string(),
634
638
  status: components.EvalRunStatus$inboundSchema,
635
639
  models: z.array(z.string()),
640
+ candidates: z.array(components.EvalCandidateObject$inboundSchema).optional(),
636
641
  task_ids: z.nullable(z.array(z.string())),
637
642
  judge_model: z.nullable(z.string()),
638
643
  embedding_model: z.nullable(z.string()),
@@ -686,6 +691,7 @@ export type ListEvalRunsData$Outbound = {
686
691
  team_id: string;
687
692
  status: string;
688
693
  models: Array<string>;
694
+ candidates?: Array<components.EvalCandidateObject$Outbound> | undefined;
689
695
  task_ids: Array<string> | null;
690
696
  judge_model: string | null;
691
697
  embedding_model: string | null;
@@ -722,6 +728,7 @@ export const ListEvalRunsData$outboundSchema: z.ZodType<
722
728
  teamId: z.string(),
723
729
  status: components.EvalRunStatus$outboundSchema,
724
730
  models: z.array(z.string()),
731
+ candidates: z.array(components.EvalCandidateObject$outboundSchema).optional(),
725
732
  taskIds: z.nullable(z.array(z.string())),
726
733
  judgeModel: z.nullable(z.string()),
727
734
  embeddingModel: z.nullable(z.string()),
package/src/models/operations/listevalscheduleruns.ts CHANGED
@@ -140,6 +140,10 @@ export type ListEvalScheduleRunsData = {
140
140
  teamId: string;
141
141
  status: components.EvalRunStatus;
142
142
  models: Array<string>;
143
+ /**
144
+ * Typed candidates, including pinned saved-agent versions. models remains the legacy storage/filter key list.
145
+ */
146
+ candidates?: Array<components.EvalCandidateObject> | undefined;
143
147
  taskIds: Array<string> | null;
144
148
  judgeModel: string | null;
145
149
  embeddingModel: string | null;
@@ -644,6 +648,7 @@ export const ListEvalScheduleRunsData$inboundSchema: z.ZodType<
644
648
  team_id: z.string(),
645
649
  status: components.EvalRunStatus$inboundSchema,
646
650
  models: z.array(z.string()),
651
+ candidates: z.array(components.EvalCandidateObject$inboundSchema).optional(),
647
652
  task_ids: z.nullable(z.array(z.string())),
648
653
  judge_model: z.nullable(z.string()),
649
654
  embedding_model: z.nullable(z.string()),
@@ -697,6 +702,7 @@ export type ListEvalScheduleRunsData$Outbound = {
697
702
  team_id: string;
698
703
  status: string;
699
704
  models: Array<string>;
705
+ candidates?: Array<components.EvalCandidateObject$Outbound> | undefined;
700
706
  task_ids: Array<string> | null;
701
707
  judge_model: string | null;
702
708
  embedding_model: string | null;
@@ -733,6 +739,7 @@ export const ListEvalScheduleRunsData$outboundSchema: z.ZodType<
733
739
  teamId: z.string(),
734
740
  status: components.EvalRunStatus$outboundSchema,
735
741
  models: z.array(z.string()),
742
+ candidates: z.array(components.EvalCandidateObject$outboundSchema).optional(),
736
743
  taskIds: z.nullable(z.array(z.string())),
737
744
  judgeModel: z.nullable(z.string()),
738
745
  embeddingModel: z.nullable(z.string()),
package/src/models/operations/listevalsuiteleaderboardruns.ts CHANGED
@@ -126,6 +126,10 @@ export type ListEvalSuiteLeaderboardRunsData = {
126
126
  teamId: string;
127
127
  status: components.EvalRunStatus;
128
128
  models: Array<string>;
129
+ /**
130
+ * Typed candidates, including pinned saved-agent versions. models remains the legacy storage/filter key list.
131
+ */
132
+ candidates?: Array<components.EvalCandidateObject> | undefined;
129
133
  taskIds: Array<string> | null;
130
134
  judgeModel: string | null;
131
135
  embeddingModel: string | null;
@@ -657,6 +661,7 @@ export const ListEvalSuiteLeaderboardRunsData$inboundSchema: z.ZodType<
657
661
  team_id: z.string(),
658
662
  status: components.EvalRunStatus$inboundSchema,
659
663
  models: z.array(z.string()),
664
+ candidates: z.array(components.EvalCandidateObject$inboundSchema).optional(),
660
665
  task_ids: z.nullable(z.array(z.string())),
661
666
  judge_model: z.nullable(z.string()),
662
667
  embedding_model: z.nullable(z.string()),
@@ -714,6 +719,7 @@ export type ListEvalSuiteLeaderboardRunsData$Outbound = {
714
719
  team_id: string;
715
720
  status: string;
716
721
  models: Array<string>;
722
+ candidates?: Array<components.EvalCandidateObject$Outbound> | undefined;
717
723
  task_ids: Array<string> | null;
718
724
  judge_model: string | null;
719
725
  embedding_model: string | null;
@@ -750,6 +756,7 @@ export const ListEvalSuiteLeaderboardRunsData$outboundSchema: z.ZodType<
750
756
  teamId: z.string(),
751
757
  status: components.EvalRunStatus$outboundSchema,
752
758
  models: z.array(z.string()),
759
+ candidates: z.array(components.EvalCandidateObject$outboundSchema).optional(),
753
760
  taskIds: z.nullable(z.array(z.string())),
754
761
  judgeModel: z.nullable(z.string()),
755
762
  embeddingModel: z.nullable(z.string()),
package/src/models/operations/rerunfailedevalsamples.ts CHANGED
@@ -125,6 +125,10 @@ export type RerunFailedEvalSamplesResponseBody = {
125
125
  teamId: string;
126
126
  status: components.EvalRunStatus;
127
127
  models: Array<string>;
128
+ /**
129
+ * Typed candidates, including pinned saved-agent versions. models remains the legacy storage/filter key list.
130
+ */
131
+ candidates?: Array<components.EvalCandidateObject> | undefined;
128
132
  taskIds: Array<string> | null;
129
133
  judgeModel: string | null;
130
134
  embeddingModel: string | null;
@@ -593,6 +597,7 @@ export const RerunFailedEvalSamplesResponseBody$inboundSchema: z.ZodType<
593
597
  team_id: z.string(),
594
598
  status: components.EvalRunStatus$inboundSchema,
595
599
  models: z.array(z.string()),
600
+ candidates: z.array(components.EvalCandidateObject$inboundSchema).optional(),
596
601
  task_ids: z.nullable(z.array(z.string())),
597
602
  judge_model: z.nullable(z.string()),
598
603
  embedding_model: z.nullable(z.string()),
@@ -648,6 +653,7 @@ export type RerunFailedEvalSamplesResponseBody$Outbound = {
648
653
  team_id: string;
649
654
  status: string;
650
655
  models: Array<string>;
656
+ candidates?: Array<components.EvalCandidateObject$Outbound> | undefined;
651
657
  task_ids: Array<string> | null;
652
658
  judge_model: string | null;
653
659
  embedding_model: string | null;
@@ -684,6 +690,7 @@ export const RerunFailedEvalSamplesResponseBody$outboundSchema: z.ZodType<
684
690
  teamId: z.string(),
685
691
  status: components.EvalRunStatus$outboundSchema,
686
692
  models: z.array(z.string()),
693
+ candidates: z.array(components.EvalCandidateObject$outboundSchema).optional(),
687
694
  taskIds: z.nullable(z.array(z.string())),
688
695
  judgeModel: z.nullable(z.string()),
689
696
  embeddingModel: z.nullable(z.string()),
package/src/models/operations/retryfailedevalrun.ts CHANGED
@@ -125,6 +125,10 @@ export type RetryFailedEvalRunResponseBody = {
125
125
  teamId: string;
126
126
  status: components.EvalRunStatus;
127
127
  models: Array<string>;
128
+ /**
129
+ * Typed candidates, including pinned saved-agent versions. models remains the legacy storage/filter key list.
130
+ */
131
+ candidates?: Array<components.EvalCandidateObject> | undefined;
128
132
  taskIds: Array<string> | null;
129
133
  judgeModel: string | null;
130
134
  embeddingModel: string | null;
@@ -587,6 +591,7 @@ export const RetryFailedEvalRunResponseBody$inboundSchema: z.ZodType<
587
591
  team_id: z.string(),
588
592
  status: components.EvalRunStatus$inboundSchema,
589
593
  models: z.array(z.string()),
594
+ candidates: z.array(components.EvalCandidateObject$inboundSchema).optional(),
590
595
  task_ids: z.nullable(z.array(z.string())),
591
596
  judge_model: z.nullable(z.string()),
592
597
  embedding_model: z.nullable(z.string()),
@@ -640,6 +645,7 @@ export type RetryFailedEvalRunResponseBody$Outbound = {
640
645
  team_id: string;
641
646
  status: string;
642
647
  models: Array<string>;
648
+ candidates?: Array<components.EvalCandidateObject$Outbound> | undefined;
643
649
  task_ids: Array<string> | null;
644
650
  judge_model: string | null;
645
651
  embedding_model: string | null;
@@ -676,6 +682,7 @@ export const RetryFailedEvalRunResponseBody$outboundSchema: z.ZodType<
676
682
  teamId: z.string(),
677
683
  status: components.EvalRunStatus$outboundSchema,
678
684
  models: z.array(z.string()),
685
+ candidates: z.array(components.EvalCandidateObject$outboundSchema).optional(),
679
686
  taskIds: z.nullable(z.array(z.string())),
680
687
  judgeModel: z.nullable(z.string()),
681
688
  embeddingModel: z.nullable(z.string()),
package/src/react-query/llmEvalsCreateRun.ts CHANGED
@@ -48,7 +48,7 @@ export type LlmEvalsCreateRunMutationError =
48
48
  * Create an eval run
49
49
  *
50
50
  * @remarks
51
- * Starts a durable eval run over the selected suite version, tasks, and models.
51
+ * Starts a durable eval run over the selected suite version and tasks, using model or pinned saved-agent candidates. Supply exactly one of candidates or the legacy models field.
52
52
  */
53
53
  export function useLlmEvalsCreateRunMutation(
54
54
  options?: MutationHookOptions<
package/src/sdk/evals.ts CHANGED
@@ -269,7 +269,7 @@ export class Evals extends ClientSDK {
269
269
  * Create an eval run
270
270
  *
271
271
  * @remarks
272
- * Starts a durable eval run over the selected suite version, tasks, and models.
272
+ * Starts a durable eval run over the selected suite version and tasks, using model or pinned saved-agent candidates. Supply exactly one of candidates or the legacy models field.
273
273
  */
274
274
  async createRun(
275
275
  request: operations.CreateEvalRunRequest,