@oro-ai/sdk 1.0.20 → 1.0.22

package/dist/index.d.mts

@@ -581,6 +581,10 @@ type AgentVersionStatus = {
      * Final score if eligible
      */
     final_score?: (number | null);
+    /**
+     * Average reasoning coefficient from included successful runs
+     */
+    reasoning_coefficient?: (number | null);
     /**
      * Whether artifacts are released
      */
@@ -1155,6 +1159,12 @@ type EvaluationRunDetail = {
     sandbox_metadata?: ({
         [key: string]: unknown;
     } | null);
+    /**
+     * Score breakdown including reasoning details
+     */
+    score_components_summary?: ({
+        [key: string]: unknown;
+    } | null);
 };
 /**
  * Public representation of an evaluation run.
@@ -1967,6 +1977,15 @@ type RunningEvaluation = {
      */
     progress_percent?: number;
 };
+/**
+ * Problems for an evaluation run.
+ */
+type RunProblemsResponse = {
+    /**
+     * Problems to evaluate against
+     */
+    problems: Array<ProblemPublic>;
+};
 /**
  * 400 - Candidate score does not exceed the challenge threshold.
  */
@@ -2764,6 +2783,16 @@ type CompleteRunData = {
 };
 type CompleteRunResponse2 = (CompleteRunResponse);
 type CompleteRunError = ((MissingScoreError | ValidationError) | EvalRunNotFoundError | (NotRunOwnerError | RunAlreadyCompleteError) | HTTPValidationError);
+type GetRunProblemsData = {
+    path: {
+        /**
+         * Evaluation run ID
+         */
+        eval_run_id: string;
+    };
+};
+type GetRunProblemsResponse = (RunProblemsResponse);
+type GetRunProblemsError = (EvalRunNotFoundError | HTTPValidationError);
 type BanMinerData = {
     body: BanRequest;
     path: {
@@ -3285,6 +3314,11 @@ declare const presignUpload: <ThrowOnError extends boolean = false>(options: Opt
  * Finalize an evaluation run with terminal status and score.
  */
 declare const completeRun: <ThrowOnError extends boolean = false>(options: OptionsLegacyParser<CompleteRunData, ThrowOnError>) => _hey_api_client_fetch.RequestResult<CompleteRunResponse, CompleteRunError, ThrowOnError>;
+/**
+ * Get problems for an evaluation run
+ * Returns the problems that should be evaluated for a given run.
+ */
+declare const getRunProblems: <ThrowOnError extends boolean = false>(options: OptionsLegacyParser<GetRunProblemsData, ThrowOnError>) => _hey_api_client_fetch.RequestResult<RunProblemsResponse, GetRunProblemsError, ThrowOnError>;
 /**
  * Ban a miner
  * Ban a miner from submitting agents.
@@ -3775,4 +3809,4 @@ declare class SessionAuthManager {
  */
 declare function configureSessionAuth(baseUrl: string, config: SessionAuthConfig): SessionAuthManager;
 
-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 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 ClearMinerCooldownData, type ClearMinerCooldownError, type ClearMinerCooldownResponse, 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 DiscardAgentVersionData, type DiscardAgentVersionError, type DiscardAgentVersionResponse, type DiscardRequest, type DiscardResponse, type ErrorCategory, type EvalRunNotFoundError, type EvaluationRunDetail, type EvaluationRunPublic, type EvaluationRunStatus, type EvaluationRunStatusPublic, 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 GetCurrentSuiteError, type GetCurrentSuiteResponse, type GetEvaluationRunData, type GetEvaluationRunError, type GetEvaluationRunResponse, type GetLeaderboardData, type GetLeaderboardError, type GetLeaderboardResponse, type GetOwnedAgentVersionStatusData, type GetOwnedAgentVersionStatusError, type GetOwnedAgentVersionStatusResponse, type GetPendingEvaluationsData, type GetPendingEvaluationsError, type GetPendingEvaluationsResponse, type GetReaperStatsError, type GetReaperStatsResponse, type GetRunningEvaluationsError, type GetRunningEvaluationsResponse, type GetSuiteProblemsData, type GetSuiteProblemsError, type GetSuiteProblemsResponse, type GetTopAgentError, type GetTopAgentResponse, type GetTopHistoryData, type GetTopHistoryError, type GetTopHistoryResponse, type GetValidatorScoresData, type GetValidatorScoresError, type GetValidatorScoresResponse, type GetValidatorsError, type GetValidatorsResponse, type HTTPValidationError, type HealthCheckError, type HealthCheckResponse, type HeartbeatData, type HeartbeatError, type HeartbeatRequest, type HeartbeatResponse, type HeartbeatResponse2, 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 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 MissingParameterError, type MissingScoreError, type NoActiveSuiteError, type NotRunOwnerError, type OroErrorCode, type PendingEvaluation, type PendingEvaluationSummary, type PendingEvaluationsResponse, 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 RateLimitExceededError, type ReaperStatsResponse, type ReevaluateAgentVersionData, type ReevaluateAgentVersionError, type ReevaluateAgentVersionResponse, type ReevaluateRequest, type ReevaluateResponse, type ReinstateAgentVersionData, type ReinstateAgentVersionError, type ReinstateAgentVersionResponse, type ReinstateRequest, type RequestChallengeData, type RequestChallengeError, type RequestChallengeResponse, type RetryConfig, type RetryContext, type RunAlreadyCompleteError, type RunningEvaluation, type ScoreBelowThresholdError, type SessionAuthConfig, SessionAuthManager, type SessionInfo, type SessionRequest, type SessionResponse, type SetTopAgentData, type SetTopAgentError, type SetTopAgentResponse, type SetTopRequest, type SetTopResponse, type StoreChutesTokenData, type StoreChutesTokenError, type StoreChutesTokenRequest, type StoreChutesTokenResponse, type SubmitAgentData, type SubmitAgentError, type SubmitAgentResponse, type SubmitAgentResponse2, type SuiteNotFoundError, type SuitePublic, type SuiteWithProblemsResponse, type TerminalStatus, type TopAgentResponse, type TopHistoryEntry, type TopHistoryResponse, type UnbanMinerData, type UnbanMinerError, type UnbanMinerResponse, type UnbanValidatorData, type UnbanValidatorError, type UnbanValidatorResponse, type UpdateProgressData, type UpdateProgressError, type UpdateProgressResponse, type ValidationError, type ValidatorCurrentAgent, type ValidatorNotFoundError, type ValidatorProblemResult, type ValidatorPublic, type ValidatorScoreSummary, type ValidatorScoresResponse, type ValidatorStatus, type WaitlistSignupRequest, type WaitlistSignupResponse, type WorkItemStatus, activateSuite, banMiner, banValidator, cancelAgentVersion, claimWork, classifyError, classifyStatus, clearMinerCooldown, client, completeRun, computeDelay, configureBittensorAuth, configurePublicClient, configureSessionAuth, createRetryFetch, createSessionEndpoint, createSuite, discardAgentVersion, generateAuthHeaders, getAgentVersion, getAgentVersionCode, getAgentVersionProblems, getAgentVersionRuns, getAgentVersionStatus, getAgentVersionVariance, getArtifactDownloadUrl, getAuditEvents, getChutesAuthStatus, getCurrentSuite, getErrorCode, getErrorDetail, getEvaluationRun, getLeaderboard, getOwnedAgentVersionStatus, getPendingEvaluations, getReaperStats, getRunningEvaluations, getSuiteProblems, getTopAgent, getTopHistory, getValidatorScores, getValidators, hasDetail, hasErrorCode, healthCheck, heartbeat, invalidateEvaluationRun, isTransient, isTransientError, joinWaitlist, listAgentVersions, listAgentVersions1, listEvaluationRuns, listMinerAgents, listMiners, listSuites, listValidators, logout, parseRetryAfter, presignUpload, reevaluateAgentVersion, reinstateAgentVersion, requestChallenge, setTopAgent, storeChutesToken, submitAgent, unbanMiner, unbanValidator, updateProgress };
+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 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 ClearMinerCooldownData, type ClearMinerCooldownError, type ClearMinerCooldownResponse, 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 DiscardAgentVersionData, type DiscardAgentVersionError, type DiscardAgentVersionResponse, type DiscardRequest, type DiscardResponse, type ErrorCategory, type EvalRunNotFoundError, type EvaluationRunDetail, type EvaluationRunPublic, type EvaluationRunStatus, type EvaluationRunStatusPublic, 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 GetCurrentSuiteError, type GetCurrentSuiteResponse, type GetEvaluationRunData, type GetEvaluationRunError, type GetEvaluationRunResponse, type GetLeaderboardData, type GetLeaderboardError, type GetLeaderboardResponse, type GetOwnedAgentVersionStatusData, type GetOwnedAgentVersionStatusError, type GetOwnedAgentVersionStatusResponse, type GetPendingEvaluationsData, type GetPendingEvaluationsError, type GetPendingEvaluationsResponse, 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 GetValidatorScoresData, type GetValidatorScoresError, type GetValidatorScoresResponse, type GetValidatorsError, type GetValidatorsResponse, type HTTPValidationError, type HealthCheckError, type HealthCheckResponse, type HeartbeatData, type HeartbeatError, type HeartbeatRequest, type HeartbeatResponse, type HeartbeatResponse2, 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 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 MissingParameterError, type MissingScoreError, type NoActiveSuiteError, type NotRunOwnerError, type OroErrorCode, type PendingEvaluation, type PendingEvaluationSummary, type PendingEvaluationsResponse, 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 RateLimitExceededError, type ReaperStatsResponse, type ReevaluateAgentVersionData, type ReevaluateAgentVersionError, type ReevaluateAgentVersionResponse, type ReevaluateRequest, type ReevaluateResponse, type ReinstateAgentVersionData, type ReinstateAgentVersionError, type ReinstateAgentVersionResponse, 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 SetTopAgentData, type SetTopAgentError, type SetTopAgentResponse, type SetTopRequest, type SetTopResponse, type StoreChutesTokenData, type StoreChutesTokenError, type StoreChutesTokenRequest, type StoreChutesTokenResponse, type SubmitAgentData, type SubmitAgentError, type SubmitAgentResponse, type SubmitAgentResponse2, type SuiteNotFoundError, type SuitePublic, type SuiteWithProblemsResponse, type TerminalStatus, type TopAgentResponse, type TopHistoryEntry, type TopHistoryResponse, type UnbanMinerData, type UnbanMinerError, type UnbanMinerResponse, type UnbanValidatorData, type UnbanValidatorError, type UnbanValidatorResponse, type UpdateProgressData, type UpdateProgressError, type UpdateProgressResponse, type ValidationError, type ValidatorCurrentAgent, type ValidatorNotFoundError, type ValidatorProblemResult, type ValidatorPublic, type ValidatorScoreSummary, type ValidatorScoresResponse, type ValidatorStatus, type WaitlistSignupRequest, type WaitlistSignupResponse, type WorkItemStatus, activateSuite, banMiner, banValidator, cancelAgentVersion, claimWork, classifyError, classifyStatus, clearMinerCooldown, client, completeRun, computeDelay, configureBittensorAuth, configurePublicClient, configureSessionAuth, createRetryFetch, createSessionEndpoint, createSuite, discardAgentVersion, generateAuthHeaders, getAgentVersion, getAgentVersionCode, getAgentVersionProblems, getAgentVersionRuns, getAgentVersionStatus, getAgentVersionVariance, getArtifactDownloadUrl, getAuditEvents, getChutesAuthStatus, getCurrentSuite, getErrorCode, getErrorDetail, getEvaluationRun, getLeaderboard, getOwnedAgentVersionStatus, getPendingEvaluations, getReaperStats, getRunProblems, getRunningEvaluations, getSuiteProblems, getTopAgent, getTopHistory, getValidatorScores, getValidators, hasDetail, hasErrorCode, healthCheck, heartbeat, invalidateEvaluationRun, isTransient, isTransientError, joinWaitlist, listAgentVersions, listAgentVersions1, listEvaluationRuns, listMinerAgents, listMiners, listSuites, listValidators, logout, parseRetryAfter, presignUpload, reevaluateAgentVersion, reinstateAgentVersion, requestChallenge, setTopAgent, storeChutesToken, submitAgent, unbanMiner, unbanValidator, updateProgress };

package/dist/index.d.ts

@@ -581,6 +581,10 @@ type AgentVersionStatus = {
      * Final score if eligible
      */
     final_score?: (number | null);
+    /**
+     * Average reasoning coefficient from included successful runs
+     */
+    reasoning_coefficient?: (number | null);
     /**
      * Whether artifacts are released
      */
@@ -1155,6 +1159,12 @@ type EvaluationRunDetail = {
     sandbox_metadata?: ({
         [key: string]: unknown;
     } | null);
+    /**
+     * Score breakdown including reasoning details
+     */
+    score_components_summary?: ({
+        [key: string]: unknown;
+    } | null);
 };
 /**
  * Public representation of an evaluation run.
@@ -1967,6 +1977,15 @@ type RunningEvaluation = {
      */
     progress_percent?: number;
 };
+/**
+ * Problems for an evaluation run.
+ */
+type RunProblemsResponse = {
+    /**
+     * Problems to evaluate against
+     */
+    problems: Array<ProblemPublic>;
+};
 /**
  * 400 - Candidate score does not exceed the challenge threshold.
  */
@@ -2764,6 +2783,16 @@ type CompleteRunData = {
 };
 type CompleteRunResponse2 = (CompleteRunResponse);
 type CompleteRunError = ((MissingScoreError | ValidationError) | EvalRunNotFoundError | (NotRunOwnerError | RunAlreadyCompleteError) | HTTPValidationError);
+type GetRunProblemsData = {
+    path: {
+        /**
+         * Evaluation run ID
+         */
+        eval_run_id: string;
+    };
+};
+type GetRunProblemsResponse = (RunProblemsResponse);
+type GetRunProblemsError = (EvalRunNotFoundError | HTTPValidationError);
 type BanMinerData = {
     body: BanRequest;
     path: {
@@ -3285,6 +3314,11 @@ declare const presignUpload: <ThrowOnError extends boolean = false>(options: Opt
  * Finalize an evaluation run with terminal status and score.
  */
 declare const completeRun: <ThrowOnError extends boolean = false>(options: OptionsLegacyParser<CompleteRunData, ThrowOnError>) => _hey_api_client_fetch.RequestResult<CompleteRunResponse, CompleteRunError, ThrowOnError>;
+/**
+ * Get problems for an evaluation run
+ * Returns the problems that should be evaluated for a given run.
+ */
+declare const getRunProblems: <ThrowOnError extends boolean = false>(options: OptionsLegacyParser<GetRunProblemsData, ThrowOnError>) => _hey_api_client_fetch.RequestResult<RunProblemsResponse, GetRunProblemsError, ThrowOnError>;
 /**
  * Ban a miner
  * Ban a miner from submitting agents.
@@ -3775,4 +3809,4 @@ declare class SessionAuthManager {
  */
 declare function configureSessionAuth(baseUrl: string, config: SessionAuthConfig): SessionAuthManager;
 
-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 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 ClearMinerCooldownData, type ClearMinerCooldownError, type ClearMinerCooldownResponse, 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 DiscardAgentVersionData, type DiscardAgentVersionError, type DiscardAgentVersionResponse, type DiscardRequest, type DiscardResponse, type ErrorCategory, type EvalRunNotFoundError, type EvaluationRunDetail, type EvaluationRunPublic, type EvaluationRunStatus, type EvaluationRunStatusPublic, 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 GetCurrentSuiteError, type GetCurrentSuiteResponse, type GetEvaluationRunData, type GetEvaluationRunError, type GetEvaluationRunResponse, type GetLeaderboardData, type GetLeaderboardError, type GetLeaderboardResponse, type GetOwnedAgentVersionStatusData, type GetOwnedAgentVersionStatusError, type GetOwnedAgentVersionStatusResponse, type GetPendingEvaluationsData, type GetPendingEvaluationsError, type GetPendingEvaluationsResponse, type GetReaperStatsError, type GetReaperStatsResponse, type GetRunningEvaluationsError, type GetRunningEvaluationsResponse, type GetSuiteProblemsData, type GetSuiteProblemsError, type GetSuiteProblemsResponse, type GetTopAgentError, type GetTopAgentResponse, type GetTopHistoryData, type GetTopHistoryError, type GetTopHistoryResponse, type GetValidatorScoresData, type GetValidatorScoresError, type GetValidatorScoresResponse, type GetValidatorsError, type GetValidatorsResponse, type HTTPValidationError, type HealthCheckError, type HealthCheckResponse, type HeartbeatData, type HeartbeatError, type HeartbeatRequest, type HeartbeatResponse, type HeartbeatResponse2, 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 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 MissingParameterError, type MissingScoreError, type NoActiveSuiteError, type NotRunOwnerError, type OroErrorCode, type PendingEvaluation, type PendingEvaluationSummary, type PendingEvaluationsResponse, 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 RateLimitExceededError, type ReaperStatsResponse, type ReevaluateAgentVersionData, type ReevaluateAgentVersionError, type ReevaluateAgentVersionResponse, type ReevaluateRequest, type ReevaluateResponse, type ReinstateAgentVersionData, type ReinstateAgentVersionError, type ReinstateAgentVersionResponse, type ReinstateRequest, type RequestChallengeData, type RequestChallengeError, type RequestChallengeResponse, type RetryConfig, type RetryContext, type RunAlreadyCompleteError, type RunningEvaluation, type ScoreBelowThresholdError, type SessionAuthConfig, SessionAuthManager, type SessionInfo, type SessionRequest, type SessionResponse, type SetTopAgentData, type SetTopAgentError, type SetTopAgentResponse, type SetTopRequest, type SetTopResponse, type StoreChutesTokenData, type StoreChutesTokenError, type StoreChutesTokenRequest, type StoreChutesTokenResponse, type SubmitAgentData, type SubmitAgentError, type SubmitAgentResponse, type SubmitAgentResponse2, type SuiteNotFoundError, type SuitePublic, type SuiteWithProblemsResponse, type TerminalStatus, type TopAgentResponse, type TopHistoryEntry, type TopHistoryResponse, type UnbanMinerData, type UnbanMinerError, type UnbanMinerResponse, type UnbanValidatorData, type UnbanValidatorError, type UnbanValidatorResponse, type UpdateProgressData, type UpdateProgressError, type UpdateProgressResponse, type ValidationError, type ValidatorCurrentAgent, type ValidatorNotFoundError, type ValidatorProblemResult, type ValidatorPublic, type ValidatorScoreSummary, type ValidatorScoresResponse, type ValidatorStatus, type WaitlistSignupRequest, type WaitlistSignupResponse, type WorkItemStatus, activateSuite, banMiner, banValidator, cancelAgentVersion, claimWork, classifyError, classifyStatus, clearMinerCooldown, client, completeRun, computeDelay, configureBittensorAuth, configurePublicClient, configureSessionAuth, createRetryFetch, createSessionEndpoint, createSuite, discardAgentVersion, generateAuthHeaders, getAgentVersion, getAgentVersionCode, getAgentVersionProblems, getAgentVersionRuns, getAgentVersionStatus, getAgentVersionVariance, getArtifactDownloadUrl, getAuditEvents, getChutesAuthStatus, getCurrentSuite, getErrorCode, getErrorDetail, getEvaluationRun, getLeaderboard, getOwnedAgentVersionStatus, getPendingEvaluations, getReaperStats, getRunningEvaluations, getSuiteProblems, getTopAgent, getTopHistory, getValidatorScores, getValidators, hasDetail, hasErrorCode, healthCheck, heartbeat, invalidateEvaluationRun, isTransient, isTransientError, joinWaitlist, listAgentVersions, listAgentVersions1, listEvaluationRuns, listMinerAgents, listMiners, listSuites, listValidators, logout, parseRetryAfter, presignUpload, reevaluateAgentVersion, reinstateAgentVersion, requestChallenge, setTopAgent, storeChutesToken, submitAgent, unbanMiner, unbanValidator, updateProgress };
+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 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 ClearMinerCooldownData, type ClearMinerCooldownError, type ClearMinerCooldownResponse, 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 DiscardAgentVersionData, type DiscardAgentVersionError, type DiscardAgentVersionResponse, type DiscardRequest, type DiscardResponse, type ErrorCategory, type EvalRunNotFoundError, type EvaluationRunDetail, type EvaluationRunPublic, type EvaluationRunStatus, type EvaluationRunStatusPublic, 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 GetCurrentSuiteError, type GetCurrentSuiteResponse, type GetEvaluationRunData, type GetEvaluationRunError, type GetEvaluationRunResponse, type GetLeaderboardData, type GetLeaderboardError, type GetLeaderboardResponse, type GetOwnedAgentVersionStatusData, type GetOwnedAgentVersionStatusError, type GetOwnedAgentVersionStatusResponse, type GetPendingEvaluationsData, type GetPendingEvaluationsError, type GetPendingEvaluationsResponse, 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 GetValidatorScoresData, type GetValidatorScoresError, type GetValidatorScoresResponse, type GetValidatorsError, type GetValidatorsResponse, type HTTPValidationError, type HealthCheckError, type HealthCheckResponse, type HeartbeatData, type HeartbeatError, type HeartbeatRequest, type HeartbeatResponse, type HeartbeatResponse2, 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 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 MissingParameterError, type MissingScoreError, type NoActiveSuiteError, type NotRunOwnerError, type OroErrorCode, type PendingEvaluation, type PendingEvaluationSummary, type PendingEvaluationsResponse, 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 RateLimitExceededError, type ReaperStatsResponse, type ReevaluateAgentVersionData, type ReevaluateAgentVersionError, type ReevaluateAgentVersionResponse, type ReevaluateRequest, type ReevaluateResponse, type ReinstateAgentVersionData, type ReinstateAgentVersionError, type ReinstateAgentVersionResponse, 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 SetTopAgentData, type SetTopAgentError, type SetTopAgentResponse, type SetTopRequest, type SetTopResponse, type StoreChutesTokenData, type StoreChutesTokenError, type StoreChutesTokenRequest, type StoreChutesTokenResponse, type SubmitAgentData, type SubmitAgentError, type SubmitAgentResponse, type SubmitAgentResponse2, type SuiteNotFoundError, type SuitePublic, type SuiteWithProblemsResponse, type TerminalStatus, type TopAgentResponse, type TopHistoryEntry, type TopHistoryResponse, type UnbanMinerData, type UnbanMinerError, type UnbanMinerResponse, type UnbanValidatorData, type UnbanValidatorError, type UnbanValidatorResponse, type UpdateProgressData, type UpdateProgressError, type UpdateProgressResponse, type ValidationError, type ValidatorCurrentAgent, type ValidatorNotFoundError, type ValidatorProblemResult, type ValidatorPublic, type ValidatorScoreSummary, type ValidatorScoresResponse, type ValidatorStatus, type WaitlistSignupRequest, type WaitlistSignupResponse, type WorkItemStatus, activateSuite, banMiner, banValidator, cancelAgentVersion, claimWork, classifyError, classifyStatus, clearMinerCooldown, client, completeRun, computeDelay, configureBittensorAuth, configurePublicClient, configureSessionAuth, createRetryFetch, createSessionEndpoint, createSuite, discardAgentVersion, generateAuthHeaders, getAgentVersion, getAgentVersionCode, getAgentVersionProblems, getAgentVersionRuns, getAgentVersionStatus, getAgentVersionVariance, getArtifactDownloadUrl, getAuditEvents, getChutesAuthStatus, getCurrentSuite, getErrorCode, getErrorDetail, getEvaluationRun, getLeaderboard, getOwnedAgentVersionStatus, getPendingEvaluations, getReaperStats, getRunProblems, getRunningEvaluations, getSuiteProblems, getTopAgent, getTopHistory, getValidatorScores, getValidators, hasDetail, hasErrorCode, healthCheck, heartbeat, invalidateEvaluationRun, isTransient, isTransientError, joinWaitlist, listAgentVersions, listAgentVersions1, listEvaluationRuns, listMinerAgents, listMiners, listSuites, listValidators, logout, parseRetryAfter, presignUpload, reevaluateAgentVersion, reinstateAgentVersion, requestChallenge, setTopAgent, storeChutesToken, submitAgent, unbanMiner, unbanValidator, updateProgress };

package/dist/index.js

@@ -67,6 +67,7 @@ __export(index_exports, {
   getOwnedAgentVersionStatus: () => getOwnedAgentVersionStatus,
   getPendingEvaluations: () => getPendingEvaluations,
   getReaperStats: () => getReaperStats,
+  getRunProblems: () => getRunProblems,
   getRunningEvaluations: () => getRunningEvaluations,
   getSuiteProblems: () => getSuiteProblems,
   getTopAgent: () => getTopAgent,
@@ -297,6 +298,12 @@ var completeRun = (options) => {
     url: "/v1/validator/evaluation-runs/{eval_run_id}/complete"
   });
 };
+var getRunProblems = (options) => {
+  return (options?.client ?? client).get({
+    ...options,
+    url: "/v1/validator/evaluation-runs/{eval_run_id}/problems"
+  });
+};
 var banMiner = (options) => {
   return (options?.client ?? client).post({
     ...options,
@@ -828,6 +835,7 @@ function configureSessionAuth(baseUrl, config) {
   getOwnedAgentVersionStatus,
   getPendingEvaluations,
   getReaperStats,
+  getRunProblems,
   getRunningEvaluations,
   getSuiteProblems,
   getTopAgent,

package/dist/index.mjs

@@ -192,6 +192,12 @@ var completeRun = (options) => {
     url: "/v1/validator/evaluation-runs/{eval_run_id}/complete"
   });
 };
+var getRunProblems = (options) => {
+  return (options?.client ?? client).get({
+    ...options,
+    url: "/v1/validator/evaluation-runs/{eval_run_id}/problems"
+  });
+};
 var banMiner = (options) => {
   return (options?.client ?? client).post({
     ...options,
@@ -722,6 +728,7 @@ export {
   getOwnedAgentVersionStatus,
   getPendingEvaluations,
   getReaperStats,
+  getRunProblems,
   getRunningEvaluations,
   getSuiteProblems,
   getTopAgent,

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "@oro-ai/sdk",
-  "version": "1.0.20",
+  "version": "1.0.22",
   "description": "Official TypeScript SDK for the ORO Bittensor Subnet API",
   "main": "dist/index.js",
   "module": "dist/index.mjs",

package/src/generated/sdk.gen.ts

@@ -1,7 +1,7 @@
 // This file is auto-generated by @hey-api/openapi-ts
 
 import { createClient, createConfig, type OptionsLegacyParser, formDataBodySerializer } from '@hey-api/client-fetch';
-import type { HealthCheckError, HealthCheckResponse, ListSuitesError, ListSuitesResponse, GetCurrentSuiteError, GetCurrentSuiteResponse, GetSuiteProblemsData, GetSuiteProblemsError, GetSuiteProblemsResponse, GetLeaderboardData, GetLeaderboardError, GetLeaderboardResponse, GetTopAgentError, GetTopAgentResponse, GetTopHistoryData, GetTopHistoryError, GetTopHistoryResponse, GetAgentVersionStatusData, GetAgentVersionStatusError, GetAgentVersionStatusResponse, GetAgentVersionRunsData, GetAgentVersionRunsError, GetAgentVersionRunsResponse, GetAgentVersionProblemsData, GetAgentVersionProblemsError, GetAgentVersionProblemsResponse, GetAgentVersionData, GetAgentVersionError, GetAgentVersionResponse, GetArtifactDownloadUrlData, GetArtifactDownloadUrlError, GetArtifactDownloadUrlResponse, GetEvaluationRunData, GetEvaluationRunError, GetEvaluationRunResponse, GetValidatorsError, GetValidatorsResponse, GetRunningEvaluationsError, GetRunningEvaluationsResponse, GetPendingEvaluationsData, GetPendingEvaluationsError, GetPendingEvaluationsResponse, JoinWaitlistData, JoinWaitlistError, JoinWaitlistResponse, RequestChallengeData, RequestChallengeError, RequestChallengeResponse, CreateSessionEndpointData, CreateSessionEndpointError, CreateSessionEndpointResponse, LogoutData, LogoutError, LogoutResponse2, SubmitAgentData, SubmitAgentError, SubmitAgentResponse2, StoreChutesTokenData, StoreChutesTokenError, StoreChutesTokenResponse, GetChutesAuthStatusError, GetChutesAuthStatusResponse, ListMinerAgentsError, ListMinerAgentsResponse, ListAgentVersionsData, ListAgentVersionsError, ListAgentVersionsResponse, GetOwnedAgentVersionStatusData, GetOwnedAgentVersionStatusError, GetOwnedAgentVersionStatusResponse, ClaimWorkData, ClaimWorkError, ClaimWorkResponse2, HeartbeatData, HeartbeatError, HeartbeatResponse2, UpdateProgressData, UpdateProgressError, UpdateProgressResponse, PresignUploadData, PresignUploadError, PresignUploadResponse2, CompleteRunData, CompleteRunError, CompleteRunResponse2, BanMinerData, BanMinerError, BanMinerResponse, UnbanMinerData, UnbanMinerError, UnbanMinerResponse, BanValidatorData, BanValidatorError, BanValidatorResponse, UnbanValidatorData, UnbanValidatorError, UnbanValidatorResponse, DiscardAgentVersionData, DiscardAgentVersionError, DiscardAgentVersionResponse, ReinstateAgentVersionData, ReinstateAgentVersionError, ReinstateAgentVersionResponse, SetTopAgentData, SetTopAgentError, SetTopAgentResponse, InvalidateEvaluationRunData, InvalidateEvaluationRunError, InvalidateEvaluationRunResponse, ReevaluateAgentVersionData, ReevaluateAgentVersionError, ReevaluateAgentVersionResponse, CancelAgentVersionData, CancelAgentVersionError, CancelAgentVersionResponse, CreateSuiteData, CreateSuiteError, CreateSuiteResponse2, ActivateSuiteData, ActivateSuiteError, ActivateSuiteResponse2, GetAuditEventsData, GetAuditEventsError, GetAuditEventsResponse, GetReaperStatsError, GetReaperStatsResponse, ClearMinerCooldownData, ClearMinerCooldownError, ClearMinerCooldownResponse, ListMinersData, ListMinersError, ListMinersResponse, ListValidatorsData, ListValidatorsError, ListValidatorsResponse, ListAgentVersions1Data, ListAgentVersions1Error, ListAgentVersions1Response, ListEvaluationRunsData, ListEvaluationRunsError, ListEvaluationRunsResponse, GetValidatorScoresData, GetValidatorScoresError, GetValidatorScoresResponse, GetAgentVersionVarianceData, GetAgentVersionVarianceError, GetAgentVersionVarianceResponse, GetAgentVersionCodeData, GetAgentVersionCodeError, GetAgentVersionCodeResponse } from './types.gen';
+import type { HealthCheckError, HealthCheckResponse, ListSuitesError, ListSuitesResponse, GetCurrentSuiteError, GetCurrentSuiteResponse, GetSuiteProblemsData, GetSuiteProblemsError, GetSuiteProblemsResponse, GetLeaderboardData, GetLeaderboardError, GetLeaderboardResponse, GetTopAgentError, GetTopAgentResponse, GetTopHistoryData, GetTopHistoryError, GetTopHistoryResponse, GetAgentVersionStatusData, GetAgentVersionStatusError, GetAgentVersionStatusResponse, GetAgentVersionRunsData, GetAgentVersionRunsError, GetAgentVersionRunsResponse, GetAgentVersionProblemsData, GetAgentVersionProblemsError, GetAgentVersionProblemsResponse, GetAgentVersionData, GetAgentVersionError, GetAgentVersionResponse, GetArtifactDownloadUrlData, GetArtifactDownloadUrlError, GetArtifactDownloadUrlResponse, GetEvaluationRunData, GetEvaluationRunError, GetEvaluationRunResponse, GetValidatorsError, GetValidatorsResponse, GetRunningEvaluationsError, GetRunningEvaluationsResponse, GetPendingEvaluationsData, GetPendingEvaluationsError, GetPendingEvaluationsResponse, JoinWaitlistData, JoinWaitlistError, JoinWaitlistResponse, RequestChallengeData, RequestChallengeError, RequestChallengeResponse, CreateSessionEndpointData, CreateSessionEndpointError, CreateSessionEndpointResponse, LogoutData, LogoutError, LogoutResponse2, SubmitAgentData, SubmitAgentError, SubmitAgentResponse2, StoreChutesTokenData, StoreChutesTokenError, StoreChutesTokenResponse, GetChutesAuthStatusError, GetChutesAuthStatusResponse, ListMinerAgentsError, ListMinerAgentsResponse, ListAgentVersionsData, ListAgentVersionsError, ListAgentVersionsResponse, GetOwnedAgentVersionStatusData, GetOwnedAgentVersionStatusError, GetOwnedAgentVersionStatusResponse, ClaimWorkData, ClaimWorkError, ClaimWorkResponse2, HeartbeatData, HeartbeatError, HeartbeatResponse2, UpdateProgressData, UpdateProgressError, UpdateProgressResponse, PresignUploadData, PresignUploadError, PresignUploadResponse2, CompleteRunData, CompleteRunError, CompleteRunResponse2, GetRunProblemsData, GetRunProblemsError, GetRunProblemsResponse, BanMinerData, BanMinerError, BanMinerResponse, UnbanMinerData, UnbanMinerError, UnbanMinerResponse, BanValidatorData, BanValidatorError, BanValidatorResponse, UnbanValidatorData, UnbanValidatorError, UnbanValidatorResponse, DiscardAgentVersionData, DiscardAgentVersionError, DiscardAgentVersionResponse, ReinstateAgentVersionData, ReinstateAgentVersionError, ReinstateAgentVersionResponse, SetTopAgentData, SetTopAgentError, SetTopAgentResponse, InvalidateEvaluationRunData, InvalidateEvaluationRunError, InvalidateEvaluationRunResponse, ReevaluateAgentVersionData, ReevaluateAgentVersionError, ReevaluateAgentVersionResponse, CancelAgentVersionData, CancelAgentVersionError, CancelAgentVersionResponse, CreateSuiteData, CreateSuiteError, CreateSuiteResponse2, ActivateSuiteData, ActivateSuiteError, ActivateSuiteResponse2, GetAuditEventsData, GetAuditEventsError, GetAuditEventsResponse, GetReaperStatsError, GetReaperStatsResponse, ClearMinerCooldownData, ClearMinerCooldownError, ClearMinerCooldownResponse, ListMinersData, ListMinersError, ListMinersResponse, ListValidatorsData, ListValidatorsError, ListValidatorsResponse, ListAgentVersions1Data, ListAgentVersions1Error, ListAgentVersions1Response, ListEvaluationRunsData, ListEvaluationRunsError, ListEvaluationRunsResponse, GetValidatorScoresData, GetValidatorScoresError, GetValidatorScoresResponse, GetAgentVersionVarianceData, GetAgentVersionVarianceError, GetAgentVersionVarianceResponse, GetAgentVersionCodeData, GetAgentVersionCodeError, GetAgentVersionCodeResponse } from './types.gen';
 
 export const client = createClient(createConfig());
 
@@ -364,6 +364,17 @@ export const completeRun = <ThrowOnError extends boolean = false>(options: Optio
     });
 };
 
+/**
+ * Get problems for an evaluation run
+ * Returns the problems that should be evaluated for a given run.
+ */
+export const getRunProblems = <ThrowOnError extends boolean = false>(options: OptionsLegacyParser<GetRunProblemsData, ThrowOnError>) => {
+    return (options?.client ?? client).get<GetRunProblemsResponse, GetRunProblemsError, ThrowOnError>({
+        ...options,
+        url: '/v1/validator/evaluation-runs/{eval_run_id}/problems'
+    });
+};
+
 /**
  * Ban a miner
  * Ban a miner from submitting agents.

package/src/generated/types.gen.ts

@@ -601,6 +601,10 @@ export type AgentVersionStatus = {
      * Final score if eligible
      */
     final_score?: (number | null);
+    /**
+     * Average reasoning coefficient from included successful runs
+     */
+    reasoning_coefficient?: (number | null);
     /**
      * Whether artifacts are released
      */
@@ -1205,6 +1209,12 @@ export type EvaluationRunDetail = {
      */
     sandbox_metadata?: ({
     [key: string]: unknown;
+} | null);
+    /**
+     * Score breakdown including reasoning details
+     */
+    score_components_summary?: ({
+    [key: string]: unknown;
 } | null);
 };
 
@@ -2060,6 +2070,16 @@ export type RunningEvaluation = {
     progress_percent?: number;
 };
 
+/**
+ * Problems for an evaluation run.
+ */
+export type RunProblemsResponse = {
+    /**
+     * Problems to evaluate against
+     */
+    problems: Array<ProblemPublic>;
+};
+
 /**
  * 400 - Candidate score does not exceed the challenge threshold.
  */
@@ -2967,6 +2987,19 @@ export type CompleteRunResponse2 = (CompleteRunResponse);
 
 export type CompleteRunError = ((MissingScoreError | ValidationError) | EvalRunNotFoundError | (NotRunOwnerError | RunAlreadyCompleteError) | HTTPValidationError);
 
+export type GetRunProblemsData = {
+    path: {
+        /**
+         * Evaluation run ID
+         */
+        eval_run_id: string;
+    };
+};
+
+export type GetRunProblemsResponse = (RunProblemsResponse);
+
+export type GetRunProblemsError = (EvalRunNotFoundError | HTTPValidationError);
+
 export type BanMinerData = {
     body: BanRequest;
     path: {