@oro-ai/sdk 1.0.100 → 1.0.102

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 CHANGED
@@ -520,6 +520,14 @@ type AgentVersionProblemsResponse = {
520
520
  * Problem progress list
521
521
  */
522
522
  problems: Array<ProblemProgressEntry>;
523
+ /**
524
+ * Generated task results, using the same phase visibility as legacy problems
525
+ */
526
+ generated_results?: Array<GeneratedEvaluationResultRead>;
527
+ /**
528
+ * Safe task labels for generated qualifying evaluations
529
+ */
530
+ qualifying_tasks?: Array<QualifyingTaskRead>;
523
531
  };
524
532
  type AgentVersionPublic = {
525
533
  /**
@@ -926,6 +934,11 @@ type ChallengeResponse = {
926
934
  */
927
935
  expires_at: number;
928
936
  };
937
+ type CheckSummary = {
938
+ passed?: number;
939
+ failed?: number;
940
+ unknown?: number;
941
+ };
929
942
  type ChutesAuthStatusResponse = {
930
943
  /**
931
944
  * Whether the miner has a valid Chutes token stored
@@ -1219,6 +1232,43 @@ type EvalRunNotFoundError = {
1219
1232
  */
1220
1233
  error_code?: "EVAL_RUN_NOT_FOUND";
1221
1234
  };
1235
+ type EvaluationExecutionRead = {
1236
+ eval_run_id: string;
1237
+ agent_version_id: string;
1238
+ suite_id: number;
1239
+ validator_hotkey: string;
1240
+ status: string;
1241
+ completed_at: (string | null);
1242
+ score: (number | null);
1243
+ execution_kind: 'legacy_shoppingbench' | 'envpack' | 'unknown';
1244
+ execution_contract_id?: (string | null);
1245
+ environment?: (SafeEnvironmentMetadata | null);
1246
+ detail_released?: boolean;
1247
+ };
1248
+ type execution_kind = 'legacy_shoppingbench' | 'envpack' | 'unknown';
1249
+ type EvaluationItemRead = {
1250
+ item_id: string;
1251
+ item_kind: 'legacy_problem' | 'envpack_episode';
1252
+ family?: (string | null);
1253
+ status: string;
1254
+ score?: (number | null);
1255
+ verdict_status?: ('passed' | 'failed' | null);
1256
+ paid_reward?: (number | null);
1257
+ check_summary?: (CheckSummary | null);
1258
+ reward_summary?: (RewardSummary | null);
1259
+ step_count?: (number | null);
1260
+ eval_run_id: string;
1261
+ validator_hotkey: string;
1262
+ };
1263
+ type item_kind = 'legacy_problem' | 'envpack_episode';
1264
+ type EvaluationItemsRead = {
1265
+ eval_run_id: string;
1266
+ execution_kind: 'legacy_shoppingbench' | 'envpack' | 'unknown';
1267
+ detail_released: boolean;
1268
+ expected_item_count?: (number | null);
1269
+ persisted_item_count: number;
1270
+ items: Array<EvaluationItemRead>;
1271
+ };
1222
1272
  /**
1223
1273
  * Which phase of the qualifying+race cycle an evaluation belongs to.
1224
1274
  */
@@ -1288,6 +1338,9 @@ type EvaluationRunDetail = {
1288
1338
  score_components_summary?: ({
1289
1339
  [key: string]: unknown;
1290
1340
  } | null);
1341
+ execution_kind?: 'legacy_shoppingbench' | 'envpack' | 'unknown';
1342
+ execution_contract_id?: (string | null);
1343
+ environment?: (SafeEnvironmentMetadata | null);
1291
1344
  };
1292
1345
  type EvaluationRunPublic = {
1293
1346
  /**
@@ -1502,6 +1555,13 @@ type FileTooLargeError = {
1502
1555
  */
1503
1556
  error_code?: "FILE_TOO_LARGE";
1504
1557
  };
1558
+ /**
1559
+ * Generated run identity and task results returned by existing agent reads.
1560
+ */
1561
+ type GeneratedEvaluationResultRead = {
1562
+ identity: EvaluationExecutionRead;
1563
+ items: EvaluationItemsRead;
1564
+ };
1505
1565
  type HeartbeatRequest = {
1506
1566
  /**
1507
1567
  * Docker image digests for validator stack services
@@ -2108,6 +2168,11 @@ type ProgressUpdateResponse = {
2108
2168
  */
2109
2169
  accepted?: boolean;
2110
2170
  };
2171
+ type QualifyingTaskRead = {
2172
+ task_id: string;
2173
+ family: string;
2174
+ goal_text: string;
2175
+ };
2111
2176
  type RaceCurrentResponse = {
2112
2177
  /**
2113
2178
  * Current race, if any
@@ -2569,6 +2634,9 @@ type ReinstateRequest = {
2569
2634
  */
2570
2635
  reason?: (string | null);
2571
2636
  };
2637
+ type RewardSummary = {
2638
+ component_count?: number;
2639
+ };
2572
2640
  type RunAlreadyCompleteError = {
2573
2641
  /**
2574
2642
  * Error message describing what went wrong
@@ -2631,6 +2699,23 @@ type RunProblemsResponse = {
2631
2699
  */
2632
2700
  problems: Array<ProblemPublic>;
2633
2701
  };
2702
+ type SafeEnvironmentMetadata = {
2703
+ display_name: string;
2704
+ /**
2705
+ * Short audit revision; never a storage key
2706
+ */
2707
+ environment_revision: string;
2708
+ catalog_epoch: string;
2709
+ task_count: number;
2710
+ family_counts: {
2711
+ [key: string]: (number);
2712
+ };
2713
+ contract_version: string;
2714
+ runtime_version: string;
2715
+ tool_contract_version: string;
2716
+ verifier_version: string;
2717
+ result_schema_version: string;
2718
+ };
2634
2719
  /**
2635
2720
  * 400 - Candidate score does not exceed the challenge threshold.
2636
2721
  */
@@ -5172,4 +5257,4 @@ declare class SessionAuthManager {
5172
5257
  */
5173
5258
  declare function configureSessionAuth(baseUrl: string, config: SessionAuthConfig): SessionAuthManager;
5174
5259
 
5175
- export { type ActivateSuiteData, type ActivateSuiteError, type ActivateSuiteResponse, type ActivateSuiteResponse2, type AdminAgentCodeResponse, type AdminAgentVersionEntry, type AdminAgentVersionsResponse, type AdminEvaluationRunEntry, type AdminEvaluationRunsResponse, type AdminMinerEntry, type AdminMinersResponse, type AdminValidatorEntry, type AdminValidatorsResponse, type AdmissionReason, type AdmissionStatus, type AgentLatestVersion, type AgentNotFoundError, type AgentPublic, type AgentVersionHistoryEntry, type AgentVersionNotFoundError, type AgentVersionProblemsResponse, type AgentVersionPublic, type AgentVersionScoreEntry, type AgentVersionState, type AgentVersionStatus, type AgentVersionVariance, type AgentVersionVarianceResponse, type AlreadyInvalidatedError, type ApiProblemsData, type ApiProblemsError, type ApiProblemsResponse, type ApiRunData, type ApiRunError, type ApiRunResponse, type ArtifactDownloadRequest, type ArtifactDownloadResponse, type ArtifactNotFoundError, type ArtifactNotReleasedError, type ArtifactReleaseState, type ArtifactType, type AtCapacityError, type AuditEventEntry, type AuditEventsResponse, type BanMinerData, type BanMinerError, type BanMinerResponse, type BanRequest, type BanResponse, type BanValidatorData, type BanValidatorError, type BanValidatorResponse, type BittensorAuthConfig, type Body_submit_agent, type CachedSession, type CancelAgentVersionData, type CancelAgentVersionError, type CancelAgentVersionResponse, type CancelRequest, type CancelResponse, type ChallengeRequest, type ChallengeResponse, type ChutesAuthStatusResponse, type ClaimWorkData, type ClaimWorkError, type ClaimWorkResponse, type ClaimWorkResponse2, type ClearCooldownResponse, type ClearMinerCooldownData, type ClearMinerCooldownError, type ClearMinerCooldownResponse, type ClearRaceSelectionError, type ClearRaceSelectionResponse, type CloseQualifyingError, type CloseQualifyingResponse, type CloseQualifyingResponse2, type CodeAnalysisError, type CompleteRunData, type CompleteRunError, type CompleteRunRequest, type CompleteRunResponse, type CompleteRunResponse2, type CooldownActiveError, type CreateSessionEndpointData, type CreateSessionEndpointError, type CreateSessionEndpointResponse, type CreateSuiteData, type CreateSuiteError, type CreateSuiteRequest, type CreateSuiteResponse, type CreateSuiteResponse2, type CurrentRacesResponse, type DeepHealthCheckError, type DeepHealthCheckResponse, type DeleteInferenceCredentialData, type DeleteInferenceCredentialError, type DeleteInferenceCredentialResponse, type DiscardAgentVersionData, type DiscardAgentVersionError, type DiscardAgentVersionResponse, type DiscardRequest, type DiscardResponse, type EliminateAgentVersionData, type EliminateAgentVersionError, type EliminateAgentVersionResponse, type EliminateRequest, type EliminateResponse, type EpochStandings, type ErrorCategory, type EvalRunNotFoundError, type EvaluationPhase, type EvaluationRunDetail, type EvaluationRunPublic, type EvaluationRunStatus, type EvaluationRunStatusPublic, type ExchangeChutesCodeData, type ExchangeChutesCodeError, type ExchangeChutesCodeRequest, type ExchangeChutesCodeResponse, type ExchangeChutesCodeResponse2, type FileTooLargeError, type GetAgentVersionCodeData, type GetAgentVersionCodeError, type GetAgentVersionCodeResponse, type GetAgentVersionData, type GetAgentVersionError, type GetAgentVersionProblemsData, type GetAgentVersionProblemsError, type GetAgentVersionProblemsResponse, type GetAgentVersionResponse, type GetAgentVersionRunsData, type GetAgentVersionRunsError, type GetAgentVersionRunsResponse, type GetAgentVersionStatusData, type GetAgentVersionStatusError, type GetAgentVersionStatusResponse, type GetAgentVersionVarianceData, type GetAgentVersionVarianceError, type GetAgentVersionVarianceResponse, type GetArtifactDownloadUrlData, type GetArtifactDownloadUrlError, type GetArtifactDownloadUrlResponse, type GetAuditEventsData, type GetAuditEventsError, type GetAuditEventsResponse, type GetChutesAuthStatusError, type GetChutesAuthStatusResponse, type GetCurrentRaceError, type GetCurrentRaceResponse, type GetCurrentRacesError, type GetCurrentRacesResponse, type GetCurrentSuiteError, type GetCurrentSuiteResponse, type GetEvaluationRunData, type GetEvaluationRunError, type GetEvaluationRunResponse, type GetInferenceAuthStatusOneData, type GetInferenceAuthStatusOneError, type GetInferenceAuthStatusOneResponse, type GetInferenceModelsData, type GetInferenceModelsError, type GetInferenceModelsResponse, type GetLeaderboardData, type GetLeaderboardError, type GetLeaderboardResponse, type GetOwnedAgentVersionStatusData, type GetOwnedAgentVersionStatusError, type GetOwnedAgentVersionStatusResponse, type GetPendingEvaluationsData, type GetPendingEvaluationsError, type GetPendingEvaluationsResponse, type GetRaceDetailData, type GetRaceDetailError, type GetRaceDetailResponse, type GetRaceDiagnosticsData, type GetRaceDiagnosticsError, type GetRaceDiagnosticsResponse, type GetRaceHistoryData, type GetRaceHistoryError, type GetRaceHistoryResponse, type GetRaceValidatorVarianceData, type GetRaceValidatorVarianceError, type GetRaceValidatorVarianceResponse, type GetReaperStatsError, type GetReaperStatsResponse, type GetRunProblemsData, type GetRunProblemsError, type GetRunProblemsResponse, type GetRunningEvaluationsError, type GetRunningEvaluationsResponse, type GetSuiteProblemsData, type GetSuiteProblemsError, type GetSuiteProblemsResponse, type GetTopAgentError, type GetTopAgentResponse, type GetTopHistoryData, type GetTopHistoryError, type GetTopHistoryResponse, type GetTopMinerPayoutError, type GetTopMinerPayoutResponse, type GetTrajectoryHistoryError, type GetTrajectoryHistoryResponse, type GetValidatorFailuresData, type GetValidatorFailuresError, type GetValidatorFailuresResponse, type GetValidatorPauseError, type GetValidatorPauseResponse, type GetValidatorResourceSamplesData, type GetValidatorResourceSamplesError, type GetValidatorResourceSamplesResponse, type GetValidatorScoresData, type GetValidatorScoresError, type GetValidatorScoresResponse, type GetValidatorsError, type GetValidatorsResponse, type GetWeightSaltError, type GetWeightSaltResponse, type HTTPValidationError, type HealthCheckError, type HealthCheckResponse, type HeartbeatData, type HeartbeatError, type HeartbeatRequest, type HeartbeatResponse, type HeartbeatResponse2, type InferenceAuthListResponse, type InferenceAuthStatusResponse, type InferenceModelsResponse, type InferenceTokenGrant, type InvalidAgentNameError, type InvalidArtifactTypeError, type InvalidFileError, type InvalidProblemIdError, type InvalidateEvaluationRunData, type InvalidateEvaluationRunError, type InvalidateEvaluationRunResponse, type InvalidateRunRequest, type JoinWaitlistData, type JoinWaitlistError, type JoinWaitlistResponse, type LeaderboardEntry, type LeaderboardResponse, type LeaseExpiredError, type ListAgentVersions1Data, type ListAgentVersions1Error, type ListAgentVersions1Response, type ListAgentVersionsData, type ListAgentVersionsError, type ListAgentVersionsResponse, type ListEvaluationRunsData, type ListEvaluationRunsError, type ListEvaluationRunsResponse, type ListInferenceAuthError, type ListInferenceAuthResponse, type ListMinerAgentsError, type ListMinerAgentsResponse, type ListMinersData, type ListMinersError, type ListMinersResponse, type ListSuitesError, type ListSuitesResponse, type ListValidatorsData, type ListValidatorsError, type ListValidatorsResponse, type LogoutData, type LogoutError, type LogoutResponse, type LogoutResponse2, type MinerAgentsResponse, type MinerNotFoundError, type MinerRaceSelectionRequest, type MissingParameterError, type MissingScoreError, type NoActiveSuiteError, type NoSelectionError, type NotRunOwnerError, type NotVersionOwnerError, type OroErrorCode, type PendingEvaluation, type PendingEvaluationSummary, type PendingEvaluationsResponse, type PinnedFinisher, type PooledWindowRace, type PostValidatorPauseData, type PostValidatorPauseError, type PostValidatorPauseResponse, type PostValidatorResumeData, type PostValidatorResumeError, type PostValidatorResumeResponse, type PresignUploadData, type PresignUploadError, type PresignUploadRequest, type PresignUploadResponse, type PresignUploadResponse2, type ProblemNotFoundError, type ProblemProgressEntry, type ProblemProgressUpdate, type ProblemPublic, type ProblemStatus, type ProgressUpdateRequest, type ProgressUpdateResponse, type RaceCurrentResponse, type RaceDetailResponse, type RaceDiagnosticsResponse, type RaceHistoryResponse, type RaceInFlightError, type RaceLockedError, type RaceNotFoundError, type RacePublic, type RaceQualifierEntry, type RaceQualifierPublic, type RaceSummary, type RaceValidatorVarianceResponse, type RaceWorkItemEntry, type RankedInferenceModel, type RankedInferenceModelsResponse, type RateLimitExceededError, type ReaperStatsResponse, type ReevaluateAgentVersionData, type ReevaluateAgentVersionError, type ReevaluateAgentVersionResponse, type ReevaluateRequest, type ReevaluateResponse, type ReinstateAgentVersionData, type ReinstateAgentVersionError, type ReinstateAgentVersionResponse, type ReinstateEliminationData, type ReinstateEliminationError, type ReinstateEliminationRequest, type ReinstateEliminationResponse, type ReinstateEliminationResponse2, type ReinstateRequest, type RequestChallengeData, type RequestChallengeError, type RequestChallengeResponse, type RetryConfig, type RetryContext, type RunAlreadyCompleteError, type RunProblemsResponse, type RunningEvaluation, type ScoreBelowThresholdError, type SessionAuthConfig, SessionAuthManager, type SessionInfo, type SessionRequest, type SessionResponse, type SetDefaultInferenceProviderData, type SetDefaultInferenceProviderError, type SetDefaultInferenceProviderResponse, type SetDefaultProviderRequest, type SetRaceSelectionData, type SetRaceSelectionError, type SetRaceSelectionResponse, type SetTopAgentData, type SetTopAgentError, type SetTopAgentResponse, type SetTopRequest, type SetTopResponse, type SimilarityCheckUnavailableError, type SpreadBucket, type StoreChutesTokenData, type StoreChutesTokenError, type StoreChutesTokenRequest, type StoreChutesTokenResponse, type StoreInferenceCredentialData, type StoreInferenceCredentialError, type StoreInferenceCredentialRequest, type StoreInferenceCredentialResponse, type SubmitAgentData, type SubmitAgentError, type SubmitAgentResponse, type SubmitAgentResponse2, type SuiteNotFoundError, type SuitePublic, type SuiteWithProblemsResponse, type TerminalStatus, type TopAgentResponse, type TopHistoryEntry, type TopHistoryResponse, type TopMinerPayoutResponse, type TrajectoryDay, type TrajectoryHistoryResponse, type UnbanMinerData, type UnbanMinerError, type UnbanMinerResponse, type UnbanValidatorData, type UnbanValidatorError, type UnbanValidatorResponse, type UpdateProgressData, type UpdateProgressError, type UpdateProgressResponse, type UpdateValidatorData, type UpdateValidatorError, type UpdateValidatorRequest, type UpdateValidatorResponse, type ValidationError, type ValidationErrorError, type ValidatorCurrentAgent, type ValidatorFailureEntry, type ValidatorFailuresResponse, type ValidatorNotFoundError, type ValidatorPauseRequest, type ValidatorPauseStatus, type ValidatorProblemResult, type ValidatorPublic, type ValidatorResourceSampleEntry, type ValidatorResourceSamplesResponse, type ValidatorResumeRequest, type ValidatorScoreSummary, type ValidatorScoresResponse, type ValidatorStatus, type ValidatorVarianceEntry, type WaitlistSignupRequest, type WaitlistSignupResponse, type WeightSaltResponse, type WorkItemStatus, activateSuite, apiProblems, apiRun, banMiner, banValidator, cancelAgentVersion, claimWork, classifyError, classifyStatus, clearMinerCooldown, clearRaceSelection, client, closeQualifying, completeRun, computeDelay, configureBittensorAuth, configurePublicClient, configureSessionAuth, createRetryFetch, createSessionEndpoint, createSuite, deepHealthCheck, deleteInferenceCredential, discardAgentVersion, eliminateAgentVersion, exchangeChutesCode, generateAuthHeaders, getAgentVersion, getAgentVersionCode, getAgentVersionProblems, getAgentVersionRuns, getAgentVersionStatus, getAgentVersionVariance, getArtifactDownloadUrl, getAuditEvents, getChutesAuthStatus, getCurrentRace, getCurrentRaces, getCurrentSuite, getErrorCode, getErrorDetail, getEvaluationRun, getInferenceAuthStatusOne, getInferenceModels, getLeaderboard, getOwnedAgentVersionStatus, getPendingEvaluations, getRaceDetail, getRaceDiagnostics, getRaceHistory, getRaceValidatorVariance, getReaperStats, getRunProblems, getRunningEvaluations, getSuiteProblems, getTopAgent, getTopHistory, getTopMinerPayout, getTrajectoryHistory, getValidatorFailures, getValidatorPause, getValidatorResourceSamples, getValidatorScores, getValidators, getWeightSalt, hasDetail, hasErrorCode, healthCheck, heartbeat, invalidateEvaluationRun, isTransient, isTransientError, joinWaitlist, listAgentVersions, listAgentVersions1, listEvaluationRuns, listInferenceAuth, listMinerAgents, listMiners, listSuites, listValidators, logout, parseRetryAfter, postValidatorPause, postValidatorResume, presignUpload, type provider, reevaluateAgentVersion, reinstateAgentVersion, reinstateElimination, requestChallenge, setDefaultInferenceProvider, setRaceSelection, setTopAgent, storeChutesToken, storeInferenceCredential, submitAgent, unbanMiner, unbanValidator, updateProgress, updateValidator };
5260
+ export { type ActivateSuiteData, type ActivateSuiteError, type ActivateSuiteResponse, type ActivateSuiteResponse2, type AdminAgentCodeResponse, type AdminAgentVersionEntry, type AdminAgentVersionsResponse, type AdminEvaluationRunEntry, type AdminEvaluationRunsResponse, type AdminMinerEntry, type AdminMinersResponse, type AdminValidatorEntry, type AdminValidatorsResponse, type AdmissionReason, type AdmissionStatus, type AgentLatestVersion, type AgentNotFoundError, type AgentPublic, type AgentVersionHistoryEntry, type AgentVersionNotFoundError, type AgentVersionProblemsResponse, type AgentVersionPublic, type AgentVersionScoreEntry, type AgentVersionState, type AgentVersionStatus, type AgentVersionVariance, type AgentVersionVarianceResponse, type AlreadyInvalidatedError, type ApiProblemsData, type ApiProblemsError, type ApiProblemsResponse, type ApiRunData, type ApiRunError, type ApiRunResponse, type ArtifactDownloadRequest, type ArtifactDownloadResponse, type ArtifactNotFoundError, type ArtifactNotReleasedError, type ArtifactReleaseState, type ArtifactType, type AtCapacityError, type AuditEventEntry, type AuditEventsResponse, type BanMinerData, type BanMinerError, type BanMinerResponse, type BanRequest, type BanResponse, type BanValidatorData, type BanValidatorError, type BanValidatorResponse, type BittensorAuthConfig, type Body_submit_agent, type CachedSession, type CancelAgentVersionData, type CancelAgentVersionError, type CancelAgentVersionResponse, type CancelRequest, type CancelResponse, type ChallengeRequest, type ChallengeResponse, type CheckSummary, type ChutesAuthStatusResponse, type ClaimWorkData, type ClaimWorkError, type ClaimWorkResponse, type ClaimWorkResponse2, type ClearCooldownResponse, type ClearMinerCooldownData, type ClearMinerCooldownError, type ClearMinerCooldownResponse, type ClearRaceSelectionError, type ClearRaceSelectionResponse, type CloseQualifyingError, type CloseQualifyingResponse, type CloseQualifyingResponse2, type CodeAnalysisError, type CompleteRunData, type CompleteRunError, type CompleteRunRequest, type CompleteRunResponse, type CompleteRunResponse2, type CooldownActiveError, type CreateSessionEndpointData, type CreateSessionEndpointError, type CreateSessionEndpointResponse, type CreateSuiteData, type CreateSuiteError, type CreateSuiteRequest, type CreateSuiteResponse, type CreateSuiteResponse2, type CurrentRacesResponse, type DeepHealthCheckError, type DeepHealthCheckResponse, type DeleteInferenceCredentialData, type DeleteInferenceCredentialError, type DeleteInferenceCredentialResponse, type DiscardAgentVersionData, type DiscardAgentVersionError, type DiscardAgentVersionResponse, type DiscardRequest, type DiscardResponse, type EliminateAgentVersionData, type EliminateAgentVersionError, type EliminateAgentVersionResponse, type EliminateRequest, type EliminateResponse, type EpochStandings, type ErrorCategory, type EvalRunNotFoundError, type EvaluationExecutionRead, type EvaluationItemRead, type EvaluationItemsRead, type EvaluationPhase, type EvaluationRunDetail, type EvaluationRunPublic, type EvaluationRunStatus, type EvaluationRunStatusPublic, type ExchangeChutesCodeData, type ExchangeChutesCodeError, type ExchangeChutesCodeRequest, type ExchangeChutesCodeResponse, type ExchangeChutesCodeResponse2, type FileTooLargeError, type GeneratedEvaluationResultRead, type GetAgentVersionCodeData, type GetAgentVersionCodeError, type GetAgentVersionCodeResponse, type GetAgentVersionData, type GetAgentVersionError, type GetAgentVersionProblemsData, type GetAgentVersionProblemsError, type GetAgentVersionProblemsResponse, type GetAgentVersionResponse, type GetAgentVersionRunsData, type GetAgentVersionRunsError, type GetAgentVersionRunsResponse, type GetAgentVersionStatusData, type GetAgentVersionStatusError, type GetAgentVersionStatusResponse, type GetAgentVersionVarianceData, type GetAgentVersionVarianceError, type GetAgentVersionVarianceResponse, type GetArtifactDownloadUrlData, type GetArtifactDownloadUrlError, type GetArtifactDownloadUrlResponse, type GetAuditEventsData, type GetAuditEventsError, type GetAuditEventsResponse, type GetChutesAuthStatusError, type GetChutesAuthStatusResponse, type GetCurrentRaceError, type GetCurrentRaceResponse, type GetCurrentRacesError, type GetCurrentRacesResponse, type GetCurrentSuiteError, type GetCurrentSuiteResponse, type GetEvaluationRunData, type GetEvaluationRunError, type GetEvaluationRunResponse, type GetInferenceAuthStatusOneData, type GetInferenceAuthStatusOneError, type GetInferenceAuthStatusOneResponse, type GetInferenceModelsData, type GetInferenceModelsError, type GetInferenceModelsResponse, type GetLeaderboardData, type GetLeaderboardError, type GetLeaderboardResponse, type GetOwnedAgentVersionStatusData, type GetOwnedAgentVersionStatusError, type GetOwnedAgentVersionStatusResponse, type GetPendingEvaluationsData, type GetPendingEvaluationsError, type GetPendingEvaluationsResponse, type GetRaceDetailData, type GetRaceDetailError, type GetRaceDetailResponse, type GetRaceDiagnosticsData, type GetRaceDiagnosticsError, type GetRaceDiagnosticsResponse, type GetRaceHistoryData, type GetRaceHistoryError, type GetRaceHistoryResponse, type GetRaceValidatorVarianceData, type GetRaceValidatorVarianceError, type GetRaceValidatorVarianceResponse, type GetReaperStatsError, type GetReaperStatsResponse, type GetRunProblemsData, type GetRunProblemsError, type GetRunProblemsResponse, type GetRunningEvaluationsError, type GetRunningEvaluationsResponse, type GetSuiteProblemsData, type GetSuiteProblemsError, type GetSuiteProblemsResponse, type GetTopAgentError, type GetTopAgentResponse, type GetTopHistoryData, type GetTopHistoryError, type GetTopHistoryResponse, type GetTopMinerPayoutError, type GetTopMinerPayoutResponse, type GetTrajectoryHistoryError, type GetTrajectoryHistoryResponse, type GetValidatorFailuresData, type GetValidatorFailuresError, type GetValidatorFailuresResponse, type GetValidatorPauseError, type GetValidatorPauseResponse, type GetValidatorResourceSamplesData, type GetValidatorResourceSamplesError, type GetValidatorResourceSamplesResponse, type GetValidatorScoresData, type GetValidatorScoresError, type GetValidatorScoresResponse, type GetValidatorsError, type GetValidatorsResponse, type GetWeightSaltError, type GetWeightSaltResponse, type HTTPValidationError, type HealthCheckError, type HealthCheckResponse, type HeartbeatData, type HeartbeatError, type HeartbeatRequest, type HeartbeatResponse, type HeartbeatResponse2, type InferenceAuthListResponse, type InferenceAuthStatusResponse, type InferenceModelsResponse, type InferenceTokenGrant, type InvalidAgentNameError, type InvalidArtifactTypeError, type InvalidFileError, type InvalidProblemIdError, type InvalidateEvaluationRunData, type InvalidateEvaluationRunError, type InvalidateEvaluationRunResponse, type InvalidateRunRequest, type JoinWaitlistData, type JoinWaitlistError, type JoinWaitlistResponse, type LeaderboardEntry, type LeaderboardResponse, type LeaseExpiredError, type ListAgentVersions1Data, type ListAgentVersions1Error, type ListAgentVersions1Response, type ListAgentVersionsData, type ListAgentVersionsError, type ListAgentVersionsResponse, type ListEvaluationRunsData, type ListEvaluationRunsError, type ListEvaluationRunsResponse, type ListInferenceAuthError, type ListInferenceAuthResponse, type ListMinerAgentsError, type ListMinerAgentsResponse, type ListMinersData, type ListMinersError, type ListMinersResponse, type ListSuitesError, type ListSuitesResponse, type ListValidatorsData, type ListValidatorsError, type ListValidatorsResponse, type LogoutData, type LogoutError, type LogoutResponse, type LogoutResponse2, type MinerAgentsResponse, type MinerNotFoundError, type MinerRaceSelectionRequest, type MissingParameterError, type MissingScoreError, type NoActiveSuiteError, type NoSelectionError, type NotRunOwnerError, type NotVersionOwnerError, type OroErrorCode, type PendingEvaluation, type PendingEvaluationSummary, type PendingEvaluationsResponse, type PinnedFinisher, type PooledWindowRace, type PostValidatorPauseData, type PostValidatorPauseError, type PostValidatorPauseResponse, type PostValidatorResumeData, type PostValidatorResumeError, type PostValidatorResumeResponse, type PresignUploadData, type PresignUploadError, type PresignUploadRequest, type PresignUploadResponse, type PresignUploadResponse2, type ProblemNotFoundError, type ProblemProgressEntry, type ProblemProgressUpdate, type ProblemPublic, type ProblemStatus, type ProgressUpdateRequest, type ProgressUpdateResponse, type QualifyingTaskRead, type RaceCurrentResponse, type RaceDetailResponse, type RaceDiagnosticsResponse, type RaceHistoryResponse, type RaceInFlightError, type RaceLockedError, type RaceNotFoundError, type RacePublic, type RaceQualifierEntry, type RaceQualifierPublic, type RaceSummary, type RaceValidatorVarianceResponse, type RaceWorkItemEntry, type RankedInferenceModel, type RankedInferenceModelsResponse, type RateLimitExceededError, type ReaperStatsResponse, type ReevaluateAgentVersionData, type ReevaluateAgentVersionError, type ReevaluateAgentVersionResponse, type ReevaluateRequest, type ReevaluateResponse, type ReinstateAgentVersionData, type ReinstateAgentVersionError, type ReinstateAgentVersionResponse, type ReinstateEliminationData, type ReinstateEliminationError, type ReinstateEliminationRequest, type ReinstateEliminationResponse, type ReinstateEliminationResponse2, type ReinstateRequest, type RequestChallengeData, type RequestChallengeError, type RequestChallengeResponse, type RetryConfig, type RetryContext, type RewardSummary, type RunAlreadyCompleteError, type RunProblemsResponse, type RunningEvaluation, type SafeEnvironmentMetadata, type ScoreBelowThresholdError, type SessionAuthConfig, SessionAuthManager, type SessionInfo, type SessionRequest, type SessionResponse, type SetDefaultInferenceProviderData, type SetDefaultInferenceProviderError, type SetDefaultInferenceProviderResponse, type SetDefaultProviderRequest, type SetRaceSelectionData, type SetRaceSelectionError, type SetRaceSelectionResponse, type SetTopAgentData, type SetTopAgentError, type SetTopAgentResponse, type SetTopRequest, type SetTopResponse, type SimilarityCheckUnavailableError, type SpreadBucket, type StoreChutesTokenData, type StoreChutesTokenError, type StoreChutesTokenRequest, type StoreChutesTokenResponse, type StoreInferenceCredentialData, type StoreInferenceCredentialError, type StoreInferenceCredentialRequest, type StoreInferenceCredentialResponse, type SubmitAgentData, type SubmitAgentError, type SubmitAgentResponse, type SubmitAgentResponse2, type SuiteNotFoundError, type SuitePublic, type SuiteWithProblemsResponse, type TerminalStatus, type TopAgentResponse, type TopHistoryEntry, type TopHistoryResponse, type TopMinerPayoutResponse, type TrajectoryDay, type TrajectoryHistoryResponse, type UnbanMinerData, type UnbanMinerError, type UnbanMinerResponse, type UnbanValidatorData, type UnbanValidatorError, type UnbanValidatorResponse, type UpdateProgressData, type UpdateProgressError, type UpdateProgressResponse, type UpdateValidatorData, type UpdateValidatorError, type UpdateValidatorRequest, type UpdateValidatorResponse, type ValidationError, type ValidationErrorError, type ValidatorCurrentAgent, type ValidatorFailureEntry, type ValidatorFailuresResponse, type ValidatorNotFoundError, type ValidatorPauseRequest, type ValidatorPauseStatus, type ValidatorProblemResult, type ValidatorPublic, type ValidatorResourceSampleEntry, type ValidatorResourceSamplesResponse, type ValidatorResumeRequest, type ValidatorScoreSummary, type ValidatorScoresResponse, type ValidatorStatus, type ValidatorVarianceEntry, type WaitlistSignupRequest, type WaitlistSignupResponse, type WeightSaltResponse, type WorkItemStatus, activateSuite, apiProblems, apiRun, banMiner, banValidator, cancelAgentVersion, claimWork, classifyError, classifyStatus, clearMinerCooldown, clearRaceSelection, client, closeQualifying, completeRun, computeDelay, configureBittensorAuth, configurePublicClient, configureSessionAuth, createRetryFetch, createSessionEndpoint, createSuite, deepHealthCheck, deleteInferenceCredential, discardAgentVersion, eliminateAgentVersion, exchangeChutesCode, type execution_kind, generateAuthHeaders, getAgentVersion, getAgentVersionCode, getAgentVersionProblems, getAgentVersionRuns, getAgentVersionStatus, getAgentVersionVariance, getArtifactDownloadUrl, getAuditEvents, getChutesAuthStatus, getCurrentRace, getCurrentRaces, getCurrentSuite, getErrorCode, getErrorDetail, getEvaluationRun, getInferenceAuthStatusOne, getInferenceModels, getLeaderboard, getOwnedAgentVersionStatus, getPendingEvaluations, getRaceDetail, getRaceDiagnostics, getRaceHistory, getRaceValidatorVariance, getReaperStats, getRunProblems, getRunningEvaluations, getSuiteProblems, getTopAgent, getTopHistory, getTopMinerPayout, getTrajectoryHistory, getValidatorFailures, getValidatorPause, getValidatorResourceSamples, getValidatorScores, getValidators, getWeightSalt, hasDetail, hasErrorCode, healthCheck, heartbeat, invalidateEvaluationRun, isTransient, isTransientError, type item_kind, joinWaitlist, listAgentVersions, listAgentVersions1, listEvaluationRuns, listInferenceAuth, listMinerAgents, listMiners, listSuites, listValidators, logout, parseRetryAfter, postValidatorPause, postValidatorResume, presignUpload, type provider, reevaluateAgentVersion, reinstateAgentVersion, reinstateElimination, requestChallenge, setDefaultInferenceProvider, setRaceSelection, setTopAgent, storeChutesToken, storeInferenceCredential, submitAgent, unbanMiner, unbanValidator, updateProgress, updateValidator };
package/dist/index.d.ts CHANGED
@@ -520,6 +520,14 @@ type AgentVersionProblemsResponse = {
520
520
  * Problem progress list
521
521
  */
522
522
  problems: Array<ProblemProgressEntry>;
523
+ /**
524
+ * Generated task results, using the same phase visibility as legacy problems
525
+ */
526
+ generated_results?: Array<GeneratedEvaluationResultRead>;
527
+ /**
528
+ * Safe task labels for generated qualifying evaluations
529
+ */
530
+ qualifying_tasks?: Array<QualifyingTaskRead>;
523
531
  };
524
532
  type AgentVersionPublic = {
525
533
  /**
@@ -926,6 +934,11 @@ type ChallengeResponse = {
926
934
  */
927
935
  expires_at: number;
928
936
  };
937
+ type CheckSummary = {
938
+ passed?: number;
939
+ failed?: number;
940
+ unknown?: number;
941
+ };
929
942
  type ChutesAuthStatusResponse = {
930
943
  /**
931
944
  * Whether the miner has a valid Chutes token stored
@@ -1219,6 +1232,43 @@ type EvalRunNotFoundError = {
1219
1232
  */
1220
1233
  error_code?: "EVAL_RUN_NOT_FOUND";
1221
1234
  };
1235
+ type EvaluationExecutionRead = {
1236
+ eval_run_id: string;
1237
+ agent_version_id: string;
1238
+ suite_id: number;
1239
+ validator_hotkey: string;
1240
+ status: string;
1241
+ completed_at: (string | null);
1242
+ score: (number | null);
1243
+ execution_kind: 'legacy_shoppingbench' | 'envpack' | 'unknown';
1244
+ execution_contract_id?: (string | null);
1245
+ environment?: (SafeEnvironmentMetadata | null);
1246
+ detail_released?: boolean;
1247
+ };
1248
+ type execution_kind = 'legacy_shoppingbench' | 'envpack' | 'unknown';
1249
+ type EvaluationItemRead = {
1250
+ item_id: string;
1251
+ item_kind: 'legacy_problem' | 'envpack_episode';
1252
+ family?: (string | null);
1253
+ status: string;
1254
+ score?: (number | null);
1255
+ verdict_status?: ('passed' | 'failed' | null);
1256
+ paid_reward?: (number | null);
1257
+ check_summary?: (CheckSummary | null);
1258
+ reward_summary?: (RewardSummary | null);
1259
+ step_count?: (number | null);
1260
+ eval_run_id: string;
1261
+ validator_hotkey: string;
1262
+ };
1263
+ type item_kind = 'legacy_problem' | 'envpack_episode';
1264
+ type EvaluationItemsRead = {
1265
+ eval_run_id: string;
1266
+ execution_kind: 'legacy_shoppingbench' | 'envpack' | 'unknown';
1267
+ detail_released: boolean;
1268
+ expected_item_count?: (number | null);
1269
+ persisted_item_count: number;
1270
+ items: Array<EvaluationItemRead>;
1271
+ };
1222
1272
  /**
1223
1273
  * Which phase of the qualifying+race cycle an evaluation belongs to.
1224
1274
  */
@@ -1288,6 +1338,9 @@ type EvaluationRunDetail = {
1288
1338
  score_components_summary?: ({
1289
1339
  [key: string]: unknown;
1290
1340
  } | null);
1341
+ execution_kind?: 'legacy_shoppingbench' | 'envpack' | 'unknown';
1342
+ execution_contract_id?: (string | null);
1343
+ environment?: (SafeEnvironmentMetadata | null);
1291
1344
  };
1292
1345
  type EvaluationRunPublic = {
1293
1346
  /**
@@ -1502,6 +1555,13 @@ type FileTooLargeError = {
1502
1555
  */
1503
1556
  error_code?: "FILE_TOO_LARGE";
1504
1557
  };
1558
+ /**
1559
+ * Generated run identity and task results returned by existing agent reads.
1560
+ */
1561
+ type GeneratedEvaluationResultRead = {
1562
+ identity: EvaluationExecutionRead;
1563
+ items: EvaluationItemsRead;
1564
+ };
1505
1565
  type HeartbeatRequest = {
1506
1566
  /**
1507
1567
  * Docker image digests for validator stack services
@@ -2108,6 +2168,11 @@ type ProgressUpdateResponse = {
2108
2168
  */
2109
2169
  accepted?: boolean;
2110
2170
  };
2171
+ type QualifyingTaskRead = {
2172
+ task_id: string;
2173
+ family: string;
2174
+ goal_text: string;
2175
+ };
2111
2176
  type RaceCurrentResponse = {
2112
2177
  /**
2113
2178
  * Current race, if any
@@ -2569,6 +2634,9 @@ type ReinstateRequest = {
2569
2634
  */
2570
2635
  reason?: (string | null);
2571
2636
  };
2637
+ type RewardSummary = {
2638
+ component_count?: number;
2639
+ };
2572
2640
  type RunAlreadyCompleteError = {
2573
2641
  /**
2574
2642
  * Error message describing what went wrong
@@ -2631,6 +2699,23 @@ type RunProblemsResponse = {
2631
2699
  */
2632
2700
  problems: Array<ProblemPublic>;
2633
2701
  };
2702
+ type SafeEnvironmentMetadata = {
2703
+ display_name: string;
2704
+ /**
2705
+ * Short audit revision; never a storage key
2706
+ */
2707
+ environment_revision: string;
2708
+ catalog_epoch: string;
2709
+ task_count: number;
2710
+ family_counts: {
2711
+ [key: string]: (number);
2712
+ };
2713
+ contract_version: string;
2714
+ runtime_version: string;
2715
+ tool_contract_version: string;
2716
+ verifier_version: string;
2717
+ result_schema_version: string;
2718
+ };
2634
2719
  /**
2635
2720
  * 400 - Candidate score does not exceed the challenge threshold.
2636
2721
  */
@@ -5172,4 +5257,4 @@ declare class SessionAuthManager {
5172
5257
  */
5173
5258
  declare function configureSessionAuth(baseUrl: string, config: SessionAuthConfig): SessionAuthManager;
5174
5259
 
5175
- export { type ActivateSuiteData, type ActivateSuiteError, type ActivateSuiteResponse, type ActivateSuiteResponse2, type AdminAgentCodeResponse, type AdminAgentVersionEntry, type AdminAgentVersionsResponse, type AdminEvaluationRunEntry, type AdminEvaluationRunsResponse, type AdminMinerEntry, type AdminMinersResponse, type AdminValidatorEntry, type AdminValidatorsResponse, type AdmissionReason, type AdmissionStatus, type AgentLatestVersion, type AgentNotFoundError, type AgentPublic, type AgentVersionHistoryEntry, type AgentVersionNotFoundError, type AgentVersionProblemsResponse, type AgentVersionPublic, type AgentVersionScoreEntry, type AgentVersionState, type AgentVersionStatus, type AgentVersionVariance, type AgentVersionVarianceResponse, type AlreadyInvalidatedError, type ApiProblemsData, type ApiProblemsError, type ApiProblemsResponse, type ApiRunData, type ApiRunError, type ApiRunResponse, type ArtifactDownloadRequest, type ArtifactDownloadResponse, type ArtifactNotFoundError, type ArtifactNotReleasedError, type ArtifactReleaseState, type ArtifactType, type AtCapacityError, type AuditEventEntry, type AuditEventsResponse, type BanMinerData, type BanMinerError, type BanMinerResponse, type BanRequest, type BanResponse, type BanValidatorData, type BanValidatorError, type BanValidatorResponse, type BittensorAuthConfig, type Body_submit_agent, type CachedSession, type CancelAgentVersionData, type CancelAgentVersionError, type CancelAgentVersionResponse, type CancelRequest, type CancelResponse, type ChallengeRequest, type ChallengeResponse, type ChutesAuthStatusResponse, type ClaimWorkData, type ClaimWorkError, type ClaimWorkResponse, type ClaimWorkResponse2, type ClearCooldownResponse, type ClearMinerCooldownData, type ClearMinerCooldownError, type ClearMinerCooldownResponse, type ClearRaceSelectionError, type ClearRaceSelectionResponse, type CloseQualifyingError, type CloseQualifyingResponse, type CloseQualifyingResponse2, type CodeAnalysisError, type CompleteRunData, type CompleteRunError, type CompleteRunRequest, type CompleteRunResponse, type CompleteRunResponse2, type CooldownActiveError, type CreateSessionEndpointData, type CreateSessionEndpointError, type CreateSessionEndpointResponse, type CreateSuiteData, type CreateSuiteError, type CreateSuiteRequest, type CreateSuiteResponse, type CreateSuiteResponse2, type CurrentRacesResponse, type DeepHealthCheckError, type DeepHealthCheckResponse, type DeleteInferenceCredentialData, type DeleteInferenceCredentialError, type DeleteInferenceCredentialResponse, type DiscardAgentVersionData, type DiscardAgentVersionError, type DiscardAgentVersionResponse, type DiscardRequest, type DiscardResponse, type EliminateAgentVersionData, type EliminateAgentVersionError, type EliminateAgentVersionResponse, type EliminateRequest, type EliminateResponse, type EpochStandings, type ErrorCategory, type EvalRunNotFoundError, type EvaluationPhase, type EvaluationRunDetail, type EvaluationRunPublic, type EvaluationRunStatus, type EvaluationRunStatusPublic, type ExchangeChutesCodeData, type ExchangeChutesCodeError, type ExchangeChutesCodeRequest, type ExchangeChutesCodeResponse, type ExchangeChutesCodeResponse2, type FileTooLargeError, type GetAgentVersionCodeData, type GetAgentVersionCodeError, type GetAgentVersionCodeResponse, type GetAgentVersionData, type GetAgentVersionError, type GetAgentVersionProblemsData, type GetAgentVersionProblemsError, type GetAgentVersionProblemsResponse, type GetAgentVersionResponse, type GetAgentVersionRunsData, type GetAgentVersionRunsError, type GetAgentVersionRunsResponse, type GetAgentVersionStatusData, type GetAgentVersionStatusError, type GetAgentVersionStatusResponse, type GetAgentVersionVarianceData, type GetAgentVersionVarianceError, type GetAgentVersionVarianceResponse, type GetArtifactDownloadUrlData, type GetArtifactDownloadUrlError, type GetArtifactDownloadUrlResponse, type GetAuditEventsData, type GetAuditEventsError, type GetAuditEventsResponse, type GetChutesAuthStatusError, type GetChutesAuthStatusResponse, type GetCurrentRaceError, type GetCurrentRaceResponse, type GetCurrentRacesError, type GetCurrentRacesResponse, type GetCurrentSuiteError, type GetCurrentSuiteResponse, type GetEvaluationRunData, type GetEvaluationRunError, type GetEvaluationRunResponse, type GetInferenceAuthStatusOneData, type GetInferenceAuthStatusOneError, type GetInferenceAuthStatusOneResponse, type GetInferenceModelsData, type GetInferenceModelsError, type GetInferenceModelsResponse, type GetLeaderboardData, type GetLeaderboardError, type GetLeaderboardResponse, type GetOwnedAgentVersionStatusData, type GetOwnedAgentVersionStatusError, type GetOwnedAgentVersionStatusResponse, type GetPendingEvaluationsData, type GetPendingEvaluationsError, type GetPendingEvaluationsResponse, type GetRaceDetailData, type GetRaceDetailError, type GetRaceDetailResponse, type GetRaceDiagnosticsData, type GetRaceDiagnosticsError, type GetRaceDiagnosticsResponse, type GetRaceHistoryData, type GetRaceHistoryError, type GetRaceHistoryResponse, type GetRaceValidatorVarianceData, type GetRaceValidatorVarianceError, type GetRaceValidatorVarianceResponse, type GetReaperStatsError, type GetReaperStatsResponse, type GetRunProblemsData, type GetRunProblemsError, type GetRunProblemsResponse, type GetRunningEvaluationsError, type GetRunningEvaluationsResponse, type GetSuiteProblemsData, type GetSuiteProblemsError, type GetSuiteProblemsResponse, type GetTopAgentError, type GetTopAgentResponse, type GetTopHistoryData, type GetTopHistoryError, type GetTopHistoryResponse, type GetTopMinerPayoutError, type GetTopMinerPayoutResponse, type GetTrajectoryHistoryError, type GetTrajectoryHistoryResponse, type GetValidatorFailuresData, type GetValidatorFailuresError, type GetValidatorFailuresResponse, type GetValidatorPauseError, type GetValidatorPauseResponse, type GetValidatorResourceSamplesData, type GetValidatorResourceSamplesError, type GetValidatorResourceSamplesResponse, type GetValidatorScoresData, type GetValidatorScoresError, type GetValidatorScoresResponse, type GetValidatorsError, type GetValidatorsResponse, type GetWeightSaltError, type GetWeightSaltResponse, type HTTPValidationError, type HealthCheckError, type HealthCheckResponse, type HeartbeatData, type HeartbeatError, type HeartbeatRequest, type HeartbeatResponse, type HeartbeatResponse2, type InferenceAuthListResponse, type InferenceAuthStatusResponse, type InferenceModelsResponse, type InferenceTokenGrant, type InvalidAgentNameError, type InvalidArtifactTypeError, type InvalidFileError, type InvalidProblemIdError, type InvalidateEvaluationRunData, type InvalidateEvaluationRunError, type InvalidateEvaluationRunResponse, type InvalidateRunRequest, type JoinWaitlistData, type JoinWaitlistError, type JoinWaitlistResponse, type LeaderboardEntry, type LeaderboardResponse, type LeaseExpiredError, type ListAgentVersions1Data, type ListAgentVersions1Error, type ListAgentVersions1Response, type ListAgentVersionsData, type ListAgentVersionsError, type ListAgentVersionsResponse, type ListEvaluationRunsData, type ListEvaluationRunsError, type ListEvaluationRunsResponse, type ListInferenceAuthError, type ListInferenceAuthResponse, type ListMinerAgentsError, type ListMinerAgentsResponse, type ListMinersData, type ListMinersError, type ListMinersResponse, type ListSuitesError, type ListSuitesResponse, type ListValidatorsData, type ListValidatorsError, type ListValidatorsResponse, type LogoutData, type LogoutError, type LogoutResponse, type LogoutResponse2, type MinerAgentsResponse, type MinerNotFoundError, type MinerRaceSelectionRequest, type MissingParameterError, type MissingScoreError, type NoActiveSuiteError, type NoSelectionError, type NotRunOwnerError, type NotVersionOwnerError, type OroErrorCode, type PendingEvaluation, type PendingEvaluationSummary, type PendingEvaluationsResponse, type PinnedFinisher, type PooledWindowRace, type PostValidatorPauseData, type PostValidatorPauseError, type PostValidatorPauseResponse, type PostValidatorResumeData, type PostValidatorResumeError, type PostValidatorResumeResponse, type PresignUploadData, type PresignUploadError, type PresignUploadRequest, type PresignUploadResponse, type PresignUploadResponse2, type ProblemNotFoundError, type ProblemProgressEntry, type ProblemProgressUpdate, type ProblemPublic, type ProblemStatus, type ProgressUpdateRequest, type ProgressUpdateResponse, type RaceCurrentResponse, type RaceDetailResponse, type RaceDiagnosticsResponse, type RaceHistoryResponse, type RaceInFlightError, type RaceLockedError, type RaceNotFoundError, type RacePublic, type RaceQualifierEntry, type RaceQualifierPublic, type RaceSummary, type RaceValidatorVarianceResponse, type RaceWorkItemEntry, type RankedInferenceModel, type RankedInferenceModelsResponse, type RateLimitExceededError, type ReaperStatsResponse, type ReevaluateAgentVersionData, type ReevaluateAgentVersionError, type ReevaluateAgentVersionResponse, type ReevaluateRequest, type ReevaluateResponse, type ReinstateAgentVersionData, type ReinstateAgentVersionError, type ReinstateAgentVersionResponse, type ReinstateEliminationData, type ReinstateEliminationError, type ReinstateEliminationRequest, type ReinstateEliminationResponse, type ReinstateEliminationResponse2, type ReinstateRequest, type RequestChallengeData, type RequestChallengeError, type RequestChallengeResponse, type RetryConfig, type RetryContext, type RunAlreadyCompleteError, type RunProblemsResponse, type RunningEvaluation, type ScoreBelowThresholdError, type SessionAuthConfig, SessionAuthManager, type SessionInfo, type SessionRequest, type SessionResponse, type SetDefaultInferenceProviderData, type SetDefaultInferenceProviderError, type SetDefaultInferenceProviderResponse, type SetDefaultProviderRequest, type SetRaceSelectionData, type SetRaceSelectionError, type SetRaceSelectionResponse, type SetTopAgentData, type SetTopAgentError, type SetTopAgentResponse, type SetTopRequest, type SetTopResponse, type SimilarityCheckUnavailableError, type SpreadBucket, type StoreChutesTokenData, type StoreChutesTokenError, type StoreChutesTokenRequest, type StoreChutesTokenResponse, type StoreInferenceCredentialData, type StoreInferenceCredentialError, type StoreInferenceCredentialRequest, type StoreInferenceCredentialResponse, type SubmitAgentData, type SubmitAgentError, type SubmitAgentResponse, type SubmitAgentResponse2, type SuiteNotFoundError, type SuitePublic, type SuiteWithProblemsResponse, type TerminalStatus, type TopAgentResponse, type TopHistoryEntry, type TopHistoryResponse, type TopMinerPayoutResponse, type TrajectoryDay, type TrajectoryHistoryResponse, type UnbanMinerData, type UnbanMinerError, type UnbanMinerResponse, type UnbanValidatorData, type UnbanValidatorError, type UnbanValidatorResponse, type UpdateProgressData, type UpdateProgressError, type UpdateProgressResponse, type UpdateValidatorData, type UpdateValidatorError, type UpdateValidatorRequest, type UpdateValidatorResponse, type ValidationError, type ValidationErrorError, type ValidatorCurrentAgent, type ValidatorFailureEntry, type ValidatorFailuresResponse, type ValidatorNotFoundError, type ValidatorPauseRequest, type ValidatorPauseStatus, type ValidatorProblemResult, type ValidatorPublic, type ValidatorResourceSampleEntry, type ValidatorResourceSamplesResponse, type ValidatorResumeRequest, type ValidatorScoreSummary, type ValidatorScoresResponse, type ValidatorStatus, type ValidatorVarianceEntry, type WaitlistSignupRequest, type WaitlistSignupResponse, type WeightSaltResponse, type WorkItemStatus, activateSuite, apiProblems, apiRun, banMiner, banValidator, cancelAgentVersion, claimWork, classifyError, classifyStatus, clearMinerCooldown, clearRaceSelection, client, closeQualifying, completeRun, computeDelay, configureBittensorAuth, configurePublicClient, configureSessionAuth, createRetryFetch, createSessionEndpoint, createSuite, deepHealthCheck, deleteInferenceCredential, discardAgentVersion, eliminateAgentVersion, exchangeChutesCode, generateAuthHeaders, getAgentVersion, getAgentVersionCode, getAgentVersionProblems, getAgentVersionRuns, getAgentVersionStatus, getAgentVersionVariance, getArtifactDownloadUrl, getAuditEvents, getChutesAuthStatus, getCurrentRace, getCurrentRaces, getCurrentSuite, getErrorCode, getErrorDetail, getEvaluationRun, getInferenceAuthStatusOne, getInferenceModels, getLeaderboard, getOwnedAgentVersionStatus, getPendingEvaluations, getRaceDetail, getRaceDiagnostics, getRaceHistory, getRaceValidatorVariance, getReaperStats, getRunProblems, getRunningEvaluations, getSuiteProblems, getTopAgent, getTopHistory, getTopMinerPayout, getTrajectoryHistory, getValidatorFailures, getValidatorPause, getValidatorResourceSamples, getValidatorScores, getValidators, getWeightSalt, hasDetail, hasErrorCode, healthCheck, heartbeat, invalidateEvaluationRun, isTransient, isTransientError, joinWaitlist, listAgentVersions, listAgentVersions1, listEvaluationRuns, listInferenceAuth, listMinerAgents, listMiners, listSuites, listValidators, logout, parseRetryAfter, postValidatorPause, postValidatorResume, presignUpload, type provider, reevaluateAgentVersion, reinstateAgentVersion, reinstateElimination, requestChallenge, setDefaultInferenceProvider, setRaceSelection, setTopAgent, storeChutesToken, storeInferenceCredential, submitAgent, unbanMiner, unbanValidator, updateProgress, updateValidator };
5260
+ export { type ActivateSuiteData, type ActivateSuiteError, type ActivateSuiteResponse, type ActivateSuiteResponse2, type AdminAgentCodeResponse, type AdminAgentVersionEntry, type AdminAgentVersionsResponse, type AdminEvaluationRunEntry, type AdminEvaluationRunsResponse, type AdminMinerEntry, type AdminMinersResponse, type AdminValidatorEntry, type AdminValidatorsResponse, type AdmissionReason, type AdmissionStatus, type AgentLatestVersion, type AgentNotFoundError, type AgentPublic, type AgentVersionHistoryEntry, type AgentVersionNotFoundError, type AgentVersionProblemsResponse, type AgentVersionPublic, type AgentVersionScoreEntry, type AgentVersionState, type AgentVersionStatus, type AgentVersionVariance, type AgentVersionVarianceResponse, type AlreadyInvalidatedError, type ApiProblemsData, type ApiProblemsError, type ApiProblemsResponse, type ApiRunData, type ApiRunError, type ApiRunResponse, type ArtifactDownloadRequest, type ArtifactDownloadResponse, type ArtifactNotFoundError, type ArtifactNotReleasedError, type ArtifactReleaseState, type ArtifactType, type AtCapacityError, type AuditEventEntry, type AuditEventsResponse, type BanMinerData, type BanMinerError, type BanMinerResponse, type BanRequest, type BanResponse, type BanValidatorData, type BanValidatorError, type BanValidatorResponse, type BittensorAuthConfig, type Body_submit_agent, type CachedSession, type CancelAgentVersionData, type CancelAgentVersionError, type CancelAgentVersionResponse, type CancelRequest, type CancelResponse, type ChallengeRequest, type ChallengeResponse, type CheckSummary, type ChutesAuthStatusResponse, type ClaimWorkData, type ClaimWorkError, type ClaimWorkResponse, type ClaimWorkResponse2, type ClearCooldownResponse, type ClearMinerCooldownData, type ClearMinerCooldownError, type ClearMinerCooldownResponse, type ClearRaceSelectionError, type ClearRaceSelectionResponse, type CloseQualifyingError, type CloseQualifyingResponse, type CloseQualifyingResponse2, type CodeAnalysisError, type CompleteRunData, type CompleteRunError, type CompleteRunRequest, type CompleteRunResponse, type CompleteRunResponse2, type CooldownActiveError, type CreateSessionEndpointData, type CreateSessionEndpointError, type CreateSessionEndpointResponse, type CreateSuiteData, type CreateSuiteError, type CreateSuiteRequest, type CreateSuiteResponse, type CreateSuiteResponse2, type CurrentRacesResponse, type DeepHealthCheckError, type DeepHealthCheckResponse, type DeleteInferenceCredentialData, type DeleteInferenceCredentialError, type DeleteInferenceCredentialResponse, type DiscardAgentVersionData, type DiscardAgentVersionError, type DiscardAgentVersionResponse, type DiscardRequest, type DiscardResponse, type EliminateAgentVersionData, type EliminateAgentVersionError, type EliminateAgentVersionResponse, type EliminateRequest, type EliminateResponse, type EpochStandings, type ErrorCategory, type EvalRunNotFoundError, type EvaluationExecutionRead, type EvaluationItemRead, type EvaluationItemsRead, type EvaluationPhase, type EvaluationRunDetail, type EvaluationRunPublic, type EvaluationRunStatus, type EvaluationRunStatusPublic, type ExchangeChutesCodeData, type ExchangeChutesCodeError, type ExchangeChutesCodeRequest, type ExchangeChutesCodeResponse, type ExchangeChutesCodeResponse2, type FileTooLargeError, type GeneratedEvaluationResultRead, type GetAgentVersionCodeData, type GetAgentVersionCodeError, type GetAgentVersionCodeResponse, type GetAgentVersionData, type GetAgentVersionError, type GetAgentVersionProblemsData, type GetAgentVersionProblemsError, type GetAgentVersionProblemsResponse, type GetAgentVersionResponse, type GetAgentVersionRunsData, type GetAgentVersionRunsError, type GetAgentVersionRunsResponse, type GetAgentVersionStatusData, type GetAgentVersionStatusError, type GetAgentVersionStatusResponse, type GetAgentVersionVarianceData, type GetAgentVersionVarianceError, type GetAgentVersionVarianceResponse, type GetArtifactDownloadUrlData, type GetArtifactDownloadUrlError, type GetArtifactDownloadUrlResponse, type GetAuditEventsData, type GetAuditEventsError, type GetAuditEventsResponse, type GetChutesAuthStatusError, type GetChutesAuthStatusResponse, type GetCurrentRaceError, type GetCurrentRaceResponse, type GetCurrentRacesError, type GetCurrentRacesResponse, type GetCurrentSuiteError, type GetCurrentSuiteResponse, type GetEvaluationRunData, type GetEvaluationRunError, type GetEvaluationRunResponse, type GetInferenceAuthStatusOneData, type GetInferenceAuthStatusOneError, type GetInferenceAuthStatusOneResponse, type GetInferenceModelsData, type GetInferenceModelsError, type GetInferenceModelsResponse, type GetLeaderboardData, type GetLeaderboardError, type GetLeaderboardResponse, type GetOwnedAgentVersionStatusData, type GetOwnedAgentVersionStatusError, type GetOwnedAgentVersionStatusResponse, type GetPendingEvaluationsData, type GetPendingEvaluationsError, type GetPendingEvaluationsResponse, type GetRaceDetailData, type GetRaceDetailError, type GetRaceDetailResponse, type GetRaceDiagnosticsData, type GetRaceDiagnosticsError, type GetRaceDiagnosticsResponse, type GetRaceHistoryData, type GetRaceHistoryError, type GetRaceHistoryResponse, type GetRaceValidatorVarianceData, type GetRaceValidatorVarianceError, type GetRaceValidatorVarianceResponse, type GetReaperStatsError, type GetReaperStatsResponse, type GetRunProblemsData, type GetRunProblemsError, type GetRunProblemsResponse, type GetRunningEvaluationsError, type GetRunningEvaluationsResponse, type GetSuiteProblemsData, type GetSuiteProblemsError, type GetSuiteProblemsResponse, type GetTopAgentError, type GetTopAgentResponse, type GetTopHistoryData, type GetTopHistoryError, type GetTopHistoryResponse, type GetTopMinerPayoutError, type GetTopMinerPayoutResponse, type GetTrajectoryHistoryError, type GetTrajectoryHistoryResponse, type GetValidatorFailuresData, type GetValidatorFailuresError, type GetValidatorFailuresResponse, type GetValidatorPauseError, type GetValidatorPauseResponse, type GetValidatorResourceSamplesData, type GetValidatorResourceSamplesError, type GetValidatorResourceSamplesResponse, type GetValidatorScoresData, type GetValidatorScoresError, type GetValidatorScoresResponse, type GetValidatorsError, type GetValidatorsResponse, type GetWeightSaltError, type GetWeightSaltResponse, type HTTPValidationError, type HealthCheckError, type HealthCheckResponse, type HeartbeatData, type HeartbeatError, type HeartbeatRequest, type HeartbeatResponse, type HeartbeatResponse2, type InferenceAuthListResponse, type InferenceAuthStatusResponse, type InferenceModelsResponse, type InferenceTokenGrant, type InvalidAgentNameError, type InvalidArtifactTypeError, type InvalidFileError, type InvalidProblemIdError, type InvalidateEvaluationRunData, type InvalidateEvaluationRunError, type InvalidateEvaluationRunResponse, type InvalidateRunRequest, type JoinWaitlistData, type JoinWaitlistError, type JoinWaitlistResponse, type LeaderboardEntry, type LeaderboardResponse, type LeaseExpiredError, type ListAgentVersions1Data, type ListAgentVersions1Error, type ListAgentVersions1Response, type ListAgentVersionsData, type ListAgentVersionsError, type ListAgentVersionsResponse, type ListEvaluationRunsData, type ListEvaluationRunsError, type ListEvaluationRunsResponse, type ListInferenceAuthError, type ListInferenceAuthResponse, type ListMinerAgentsError, type ListMinerAgentsResponse, type ListMinersData, type ListMinersError, type ListMinersResponse, type ListSuitesError, type ListSuitesResponse, type ListValidatorsData, type ListValidatorsError, type ListValidatorsResponse, type LogoutData, type LogoutError, type LogoutResponse, type LogoutResponse2, type MinerAgentsResponse, type MinerNotFoundError, type MinerRaceSelectionRequest, type MissingParameterError, type MissingScoreError, type NoActiveSuiteError, type NoSelectionError, type NotRunOwnerError, type NotVersionOwnerError, type OroErrorCode, type PendingEvaluation, type PendingEvaluationSummary, type PendingEvaluationsResponse, type PinnedFinisher, type PooledWindowRace, type PostValidatorPauseData, type PostValidatorPauseError, type PostValidatorPauseResponse, type PostValidatorResumeData, type PostValidatorResumeError, type PostValidatorResumeResponse, type PresignUploadData, type PresignUploadError, type PresignUploadRequest, type PresignUploadResponse, type PresignUploadResponse2, type ProblemNotFoundError, type ProblemProgressEntry, type ProblemProgressUpdate, type ProblemPublic, type ProblemStatus, type ProgressUpdateRequest, type ProgressUpdateResponse, type QualifyingTaskRead, type RaceCurrentResponse, type RaceDetailResponse, type RaceDiagnosticsResponse, type RaceHistoryResponse, type RaceInFlightError, type RaceLockedError, type RaceNotFoundError, type RacePublic, type RaceQualifierEntry, type RaceQualifierPublic, type RaceSummary, type RaceValidatorVarianceResponse, type RaceWorkItemEntry, type RankedInferenceModel, type RankedInferenceModelsResponse, type RateLimitExceededError, type ReaperStatsResponse, type ReevaluateAgentVersionData, type ReevaluateAgentVersionError, type ReevaluateAgentVersionResponse, type ReevaluateRequest, type ReevaluateResponse, type ReinstateAgentVersionData, type ReinstateAgentVersionError, type ReinstateAgentVersionResponse, type ReinstateEliminationData, type ReinstateEliminationError, type ReinstateEliminationRequest, type ReinstateEliminationResponse, type ReinstateEliminationResponse2, type ReinstateRequest, type RequestChallengeData, type RequestChallengeError, type RequestChallengeResponse, type RetryConfig, type RetryContext, type RewardSummary, type RunAlreadyCompleteError, type RunProblemsResponse, type RunningEvaluation, type SafeEnvironmentMetadata, type ScoreBelowThresholdError, type SessionAuthConfig, SessionAuthManager, type SessionInfo, type SessionRequest, type SessionResponse, type SetDefaultInferenceProviderData, type SetDefaultInferenceProviderError, type SetDefaultInferenceProviderResponse, type SetDefaultProviderRequest, type SetRaceSelectionData, type SetRaceSelectionError, type SetRaceSelectionResponse, type SetTopAgentData, type SetTopAgentError, type SetTopAgentResponse, type SetTopRequest, type SetTopResponse, type SimilarityCheckUnavailableError, type SpreadBucket, type StoreChutesTokenData, type StoreChutesTokenError, type StoreChutesTokenRequest, type StoreChutesTokenResponse, type StoreInferenceCredentialData, type StoreInferenceCredentialError, type StoreInferenceCredentialRequest, type StoreInferenceCredentialResponse, type SubmitAgentData, type SubmitAgentError, type SubmitAgentResponse, type SubmitAgentResponse2, type SuiteNotFoundError, type SuitePublic, type SuiteWithProblemsResponse, type TerminalStatus, type TopAgentResponse, type TopHistoryEntry, type TopHistoryResponse, type TopMinerPayoutResponse, type TrajectoryDay, type TrajectoryHistoryResponse, type UnbanMinerData, type UnbanMinerError, type UnbanMinerResponse, type UnbanValidatorData, type UnbanValidatorError, type UnbanValidatorResponse, type UpdateProgressData, type UpdateProgressError, type UpdateProgressResponse, type UpdateValidatorData, type UpdateValidatorError, type UpdateValidatorRequest, type UpdateValidatorResponse, type ValidationError, type ValidationErrorError, type ValidatorCurrentAgent, type ValidatorFailureEntry, type ValidatorFailuresResponse, type ValidatorNotFoundError, type ValidatorPauseRequest, type ValidatorPauseStatus, type ValidatorProblemResult, type ValidatorPublic, type ValidatorResourceSampleEntry, type ValidatorResourceSamplesResponse, type ValidatorResumeRequest, type ValidatorScoreSummary, type ValidatorScoresResponse, type ValidatorStatus, type ValidatorVarianceEntry, type WaitlistSignupRequest, type WaitlistSignupResponse, type WeightSaltResponse, type WorkItemStatus, activateSuite, apiProblems, apiRun, banMiner, banValidator, cancelAgentVersion, claimWork, classifyError, classifyStatus, clearMinerCooldown, clearRaceSelection, client, closeQualifying, completeRun, computeDelay, configureBittensorAuth, configurePublicClient, configureSessionAuth, createRetryFetch, createSessionEndpoint, createSuite, deepHealthCheck, deleteInferenceCredential, discardAgentVersion, eliminateAgentVersion, exchangeChutesCode, type execution_kind, generateAuthHeaders, getAgentVersion, getAgentVersionCode, getAgentVersionProblems, getAgentVersionRuns, getAgentVersionStatus, getAgentVersionVariance, getArtifactDownloadUrl, getAuditEvents, getChutesAuthStatus, getCurrentRace, getCurrentRaces, getCurrentSuite, getErrorCode, getErrorDetail, getEvaluationRun, getInferenceAuthStatusOne, getInferenceModels, getLeaderboard, getOwnedAgentVersionStatus, getPendingEvaluations, getRaceDetail, getRaceDiagnostics, getRaceHistory, getRaceValidatorVariance, getReaperStats, getRunProblems, getRunningEvaluations, getSuiteProblems, getTopAgent, getTopHistory, getTopMinerPayout, getTrajectoryHistory, getValidatorFailures, getValidatorPause, getValidatorResourceSamples, getValidatorScores, getValidators, getWeightSalt, hasDetail, hasErrorCode, healthCheck, heartbeat, invalidateEvaluationRun, isTransient, isTransientError, type item_kind, joinWaitlist, listAgentVersions, listAgentVersions1, listEvaluationRuns, listInferenceAuth, listMinerAgents, listMiners, listSuites, listValidators, logout, parseRetryAfter, postValidatorPause, postValidatorResume, presignUpload, type provider, reevaluateAgentVersion, reinstateAgentVersion, reinstateElimination, requestChallenge, setDefaultInferenceProvider, setRaceSelection, setTopAgent, storeChutesToken, storeInferenceCredential, submitAgent, unbanMiner, unbanValidator, updateProgress, updateValidator };
package/package.json CHANGED
@@ -1,6 +1,6 @@
1
1
  {
2
2
  "name": "@oro-ai/sdk",
3
- "version": "1.0.100",
3
+ "version": "1.0.102",
4
4
  "description": "Official TypeScript SDK for the ORO Bittensor Subnet API",
5
5
  "main": "dist/index.js",
6
6
  "module": "dist/index.mjs",
@@ -536,6 +536,14 @@ export type AgentVersionProblemsResponse = {
536
536
  * Problem progress list
537
537
  */
538
538
  problems: Array<ProblemProgressEntry>;
539
+ /**
540
+ * Generated task results, using the same phase visibility as legacy problems
541
+ */
542
+ generated_results?: Array<GeneratedEvaluationResultRead>;
543
+ /**
544
+ * Safe task labels for generated qualifying evaluations
545
+ */
546
+ qualifying_tasks?: Array<QualifyingTaskRead>;
539
547
  };
540
548
 
541
549
  export type AgentVersionPublic = {
@@ -966,6 +974,12 @@ export type ChallengeResponse = {
966
974
  expires_at: number;
967
975
  };
968
976
 
977
+ export type CheckSummary = {
978
+ passed?: number;
979
+ failed?: number;
980
+ unknown?: number;
981
+ };
982
+
969
983
  export type ChutesAuthStatusResponse = {
970
984
  /**
971
985
  * Whether the miner has a valid Chutes token stored
@@ -1276,6 +1290,48 @@ export type EvalRunNotFoundError = {
1276
1290
  error_code?: "EVAL_RUN_NOT_FOUND";
1277
1291
  };
1278
1292
 
1293
+ export type EvaluationExecutionRead = {
1294
+ eval_run_id: string;
1295
+ agent_version_id: string;
1296
+ suite_id: number;
1297
+ validator_hotkey: string;
1298
+ status: string;
1299
+ completed_at: (string | null);
1300
+ score: (number | null);
1301
+ execution_kind: 'legacy_shoppingbench' | 'envpack' | 'unknown';
1302
+ execution_contract_id?: (string | null);
1303
+ environment?: (SafeEnvironmentMetadata | null);
1304
+ detail_released?: boolean;
1305
+ };
1306
+
1307
+ export type execution_kind = 'legacy_shoppingbench' | 'envpack' | 'unknown';
1308
+
1309
+ export type EvaluationItemRead = {
1310
+ item_id: string;
1311
+ item_kind: 'legacy_problem' | 'envpack_episode';
1312
+ family?: (string | null);
1313
+ status: string;
1314
+ score?: (number | null);
1315
+ verdict_status?: ('passed' | 'failed' | null);
1316
+ paid_reward?: (number | null);
1317
+ check_summary?: (CheckSummary | null);
1318
+ reward_summary?: (RewardSummary | null);
1319
+ step_count?: (number | null);
1320
+ eval_run_id: string;
1321
+ validator_hotkey: string;
1322
+ };
1323
+
1324
+ export type item_kind = 'legacy_problem' | 'envpack_episode';
1325
+
1326
+ export type EvaluationItemsRead = {
1327
+ eval_run_id: string;
1328
+ execution_kind: 'legacy_shoppingbench' | 'envpack' | 'unknown';
1329
+ detail_released: boolean;
1330
+ expected_item_count?: (number | null);
1331
+ persisted_item_count: number;
1332
+ items: Array<EvaluationItemRead>;
1333
+ };
1334
+
1279
1335
  /**
1280
1336
  * Which phase of the qualifying+race cycle an evaluation belongs to.
1281
1337
  */
@@ -1346,6 +1402,9 @@ export type EvaluationRunDetail = {
1346
1402
  score_components_summary?: ({
1347
1403
  [key: string]: unknown;
1348
1404
  } | null);
1405
+ execution_kind?: 'legacy_shoppingbench' | 'envpack' | 'unknown';
1406
+ execution_contract_id?: (string | null);
1407
+ environment?: (SafeEnvironmentMetadata | null);
1349
1408
  };
1350
1409
 
1351
1410
  export type EvaluationRunPublic = {
@@ -1567,6 +1626,14 @@ export type FileTooLargeError = {
1567
1626
  error_code?: "FILE_TOO_LARGE";
1568
1627
  };
1569
1628
 
1629
+ /**
1630
+ * Generated run identity and task results returned by existing agent reads.
1631
+ */
1632
+ export type GeneratedEvaluationResultRead = {
1633
+ identity: EvaluationExecutionRead;
1634
+ items: EvaluationItemsRead;
1635
+ };
1636
+
1570
1637
  export type HeartbeatRequest = {
1571
1638
  /**
1572
1639
  * Docker image digests for validator stack services
@@ -2213,6 +2280,12 @@ export type ProgressUpdateResponse = {
2213
2280
  accepted?: boolean;
2214
2281
  };
2215
2282
 
2283
+ export type QualifyingTaskRead = {
2284
+ task_id: string;
2285
+ family: string;
2286
+ goal_text: string;
2287
+ };
2288
+
2216
2289
  export type RaceCurrentResponse = {
2217
2290
  /**
2218
2291
  * Current race, if any
@@ -2696,6 +2769,10 @@ export type ReinstateRequest = {
2696
2769
  reason?: (string | null);
2697
2770
  };
2698
2771
 
2772
+ export type RewardSummary = {
2773
+ component_count?: number;
2774
+ };
2775
+
2699
2776
  export type RunAlreadyCompleteError = {
2700
2777
  /**
2701
2778
  * Error message describing what went wrong
@@ -2761,6 +2838,24 @@ export type RunProblemsResponse = {
2761
2838
  problems: Array<ProblemPublic>;
2762
2839
  };
2763
2840
 
2841
+ export type SafeEnvironmentMetadata = {
2842
+ display_name: string;
2843
+ /**
2844
+ * Short audit revision; never a storage key
2845
+ */
2846
+ environment_revision: string;
2847
+ catalog_epoch: string;
2848
+ task_count: number;
2849
+ family_counts: {
2850
+ [key: string]: (number);
2851
+ };
2852
+ contract_version: string;
2853
+ runtime_version: string;
2854
+ tool_contract_version: string;
2855
+ verifier_version: string;
2856
+ result_schema_version: string;
2857
+ };
2858
+
2764
2859
  /**
2765
2860
  * 400 - Candidate score does not exceed the challenge threshold.
2766
2861
  */