@axiom-lattice/core 3.0.7 → 3.1.1

package/dist/index.d.mts

@@ -5494,12 +5494,13 @@ interface EvalRunService {
      *   executes in. Comes from the CALLER's runConfig, NOT from the eval
      *   project's targetServerConfig. Test cases are environment-agnostic;
      *   the run environment is decided at run time.
+     * @param taskId - Optional training task (round) this run belongs to.
      * @returns The newly created run ID
      */
     startRun(tenantId: string, projectId: string, suiteIds?: string[], caseIds?: string[], runConfig?: {
         workspaceId?: string;
         projectId?: string;
-    }): Promise<string>;
+    }, taskId?: string): Promise<string>;
     /**
      * Abort a running evaluation.
      *
@@ -5695,6 +5696,17 @@ declare class LatticeEvalSuite {
  * A judge that cannot discriminate these should not be trusted with real cases.
  */
 declare const DEFAULT_CALIBRATION_PROBES: CalibrationProbe[];
+/**
+ * Resolve the judge model for an eval project: an explicit judgeModelKey wins;
+ * otherwise the model registered under the "default" key (the convention apps
+ * use — see examples/deep_research), then the first registered model as a last
+ * resort. Throws only when no model is registered at all.
+ */
+declare function resolveJudgeModelKey(explicitKey: string | undefined, manager: {
+    getAllLattices: () => Array<{
+        key: string;
+    }>;
+}): string;
 /**
  * Manages a project with multiple evaluation suites.
  * Registers a per-project judge agent (keyed by project name and tenant)
@@ -8202,4 +8214,4 @@ declare class IdRemapper {
     remapSkillIds(graphDefinition: Record<string, unknown>, skillRemap: Record<string, string>): Record<string, unknown>;
 }
 
-export { AGENT_TASK_EVENT, type AddMessageParams, Agent, type AgentClient, type AgentExecutor, AgentInstanceManager, type AgentLattice, AgentLatticeManager, type AgentLifecycleEventName, AgentManager, type AgentStreamExecutor, type AgentThreadInterface, BUILTIN_PLUGINS, BUILTIN_SKILLS, type BackendFactory, type BackendProtocol, type BufferStats, type CacheEntry, type CalibrationProbe, type CaseRunResult, type Chunk, ChunkBuffer, ChunkBufferLatticeManager, CollectionLatticeManager, type ColumnInfo, CompositeBackend, type ConflictItem, type ConnectAllChannelsOptions, ConnectionRegistry, ConsoleLoggerClient, type CreateSandboxProviderConfig, type CronFields, CustomMetricsClient, type CustomMiddlewareFactory, CustomMiddlewareRegistry, DEFAULT_CALIBRATION_PROBES, type DatabaseConfig, type DatabaseType, DaytonaInstance, DaytonaProvider, type DaytonaProviderConfig, DefaultScheduleClient, type DeleteResult, DependencyResolver, E2BInstance, E2BProvider, type E2BProviderConfig, EMPTY_CONTENT_WARNING, type EditResult, type EmbeddingsInfo, type EmbeddingsLatticeInterface, EmbeddingsLatticeManager, type EnsureMicrosandboxInput, type EvalRunService, type ExportBundle, type ExportJob, type ExportableEntity, type ExportableEntityDefinition, ExportableEntityRegistry, type ExportableTypeInfo, type FileData, type FileInfo, FileSystemSkillStore, type FileSystemSkillStoreOptions, FilesystemBackend, type FsEntry, type GrepMatch, type IMessageQueueStore, type IMetricsServerClient, type ISqlDatabase, IdRemapper, type ImportApplyResult, type ImportEntityResult, type ImportEntityStatus, type ImportPreviewResult, InMemoryA2AApiKeyStore, InMemoryAssistantStore, InMemoryBindingStore, InMemoryChannelInstallationStore, InMemoryChunkBuffer, InMemoryCollectionStore, InMemoryDatabaseConfigStore, InMemoryMailboxStore, InMemoryMenuStore, InMemoryTaskListStore, InMemoryTaskStore, InMemoryTenantStore, InMemoryThreadMessageQueueStore, InMemoryThreadStore, InMemoryUserStore, InMemoryUserTenantLinkStore, type InsertionItem, type JudgeVerdict, LINE_NUMBER_WIDTH, type LangGraphStateChecker, type LatticeAgentStepConfig, LatticeEval, type LatticeEvalBatchReport, type LatticeEvalCase, type LatticeEvalCaseType, type LatticeEvalCaseWithTemplate, type LatticeEvalConfig, type LatticeEvalLogEvent, type LatticeEvalLogLevel, LatticeEvalProject, type LatticeEvalProjectType, type LatticeEvalResult, type LatticeEvalRubric, LatticeEvalSuite, type LatticeEvalSuiteType, type LatticeEvalTemplate, LocalSandboxInstance, LocalSandboxProvider, type LocalSandboxProviderConfig, type LoggerLattice, LoggerLatticeManager, MAX_LINE_LENGTH, type MailboxMessage, type MailboxStore, type McpLatticeInterface, McpLatticeManager, type McpServerInfo, MemoryBackend, MemoryLatticeManager, MemoryQueueClient, MemoryScheduleStorage, type MessageCompletedEvent, type MessageFailedEvent, type MessageStartedEvent, MessageType, MetricsServerManager, MicrosandboxRemoteInstance, MicrosandboxRemoteProvider, type MicrosandboxRemoteProviderClient, type MicrosandboxRemoteProviderConfig, MicrosandboxServiceClient, type MicrosandboxServiceClientConfig, type MicrosandboxShellExecInput, type ModelConfig, type ModelLatticeInterface, ModelLatticeManager, MysqlDatabase, type OutputType, type PendingMessage, PersonalAssistantConfig, PinoLoggerClient, PluginRegistry, PostgresDatabase, type PreviewError, PrometheusClient, type QueryResult, type QueueLattice, QueueLatticeManager, QueueMode, type QueuePendingEvent, RemoteSandboxInstance, RemoteSandboxProvider, type RemoteSandboxProviderConfig, type Resolution, type ResolutionAction, type ResolveAgentFn, type ResolvedConfig, type RunSandboxConfig, type RuntimeModelConfig, type STTModelInfo, STTModelLattice, type STTModelLatticeInterface, STTModelLatticeManager, type SandboxFileInfo, type SandboxFileService, SandboxFilesystem, type SandboxInstance, type SandboxIsolationLevel, SandboxLatticeManager, type SandboxManagerProtocol, type SandboxProvider, type SandboxProviderFactory, type SandboxProviderType, type SandboxShellService, SandboxSkillStore, type SandboxSkillStoreOptions, type SandboxVolumeDefinition, type ScheduleLattice, ScheduleLatticeManager, type SchedulerMiddlewareOptions, SemanticMetricsClient, type SharePayload, SimpleMemoryVectorStore, type SkillLattice, SkillLatticeManager, type SkillMeta, type SkillResource, SqlDatabaseManager, type StateAndStore, StateBackend, StoreBackend, type StoreLattice, StoreLatticeManager, type StoreType, type StoreTypeMap, TOOL_RESULT_TOKEN_LIMIT, TRUNCATION_GUIDANCE, type TableInfo, type TableSchema, type TaskEvent, type TaskListStore, type TaskSpec, TaskStatus, type TaskUpdatable, TeamAgentGraphBuilder, type TeamConfig, type TeamMiddlewareOptions, type TeamTask, type TeammateSpec, type TeammateToolsOptions, type ThreadBuffer, type ThreadBufferConfig, type ThreadBusyEvent, type ThreadIdleEvent, type ThreadInfo, type ThreadQueueConfig, type ThreadState, ThreadStatus, type ThreadStatusChangedEvent, TokenCache, type ToolDefinition, type ToolLattice, ToolLatticeManager, type UnknownToolHandlerConfig, type VectorStoreLatticeInterface, VectorStoreLatticeManager, type VectorStoreProviderLattice, VectorStoreProviderManager, VolumeFilesystem, type VolumeFsClient, type WorkflowValidationError, type WriteResult, abortWorkflowRun, agentInstanceManager, agentLatticeManager, buildGrepResultsDict, buildInput, buildNamedVolumeName, buildSandboxMetadataEnv, buildSkillFile, buildStateAnnotation, buildTableName, checkEmptyContent, clearEncryptionKeyCache, clearEvalRunService, collectionLatticeManager, compileInternal, compileWorkflow, computeSandboxName, configureStores, connectAllChannels, createAgentNode, createAgentTeam, createExecuteSqlQueryTool, createFileData, createInfoSqlTool, createListMetricsDataSourcesTool, createListMetricsServersTool, createListTablesSqlTool, createMapNode, createModelSelectorMiddleware, createNodeHandler, createQueryCheckerSqlTool, createQueryMetricDefinitionTool, createQueryMetricsListTool, createQuerySemanticMetricDataTool, createQuerySqlTool, createQueryTableDefinitionTool, createQueryTablesListTool, createResourceAddress, createSandboxProvider, createSchedulerMiddleware, createSharePayload, createTaskMiddleware, createTeamMiddleware, createTeammateTools, createUnknownToolHandlerMiddleware, createWidgetMiddleware, decrypt, describeCronExpression, documentLearningPlugin, documentParserPlugin, embeddingsLatticeManager, encrypt, ensureBuiltinAgentsForTenant, evaluateLatticeCaseWithLogs, eventBus, eventBus as eventBusDefault, extractFetcherError, extractOutput, fileDataToString, formatContentWithLineNumbers, formatGrepMatches, formatGrepResults, formatReadResponse, generateToken, getAgentClient, getAgentConfig, getAllAgentConfigs, getAllBuiltInSkillMetas, getAllToolDefinitions, getBindingRegistry, getBuiltInSkillContent, getBuiltInSkillMeta, getBuiltInSkillNames, getCheckpointSaver, getChunkBuffer, getCollectionEntryCount, getEmbeddingsClient, getEmbeddingsLattice, getEncryptionKey, getEvalRunService, getLoggerLattice, getMenuRegistry, getModelLattice, getNextCronTime, getOrCreateCollectionVectorStore, getQueueLattice, getSTTClient, getSTTClientWithTenant, getSTTModelLattice, getSandBoxManager, getScheduleLattice, getStoreLattice, getToolClient, getToolDefinition, getToolLattice, getVectorStoreClient, getVectorStoreLattice, getVectorStoreProvider, getWorkflowSignal, globSearchFiles, grepMatchesFromFiles, grepSearchFiles, hasChunkBuffer, invokeWithRetry, isBuiltInSkill, isUsingDefaultKey, isValidCronExpression, isValidSandboxName, isValidSkillName, listCollectionEntries, listSandboxProviderTypes, loggerLatticeManager, mcpManager, metricsServerManager, modelLatticeManager, normalizeSandboxName, parallelLimit, parseCronExpression, parseJudgeVerdict, parseSkillFrontmatter, parseYaml, performStringReplacement, queueLatticeManager, registerAgentLattice, registerAgentLatticeWithTenant, registerAgentLattices, registerBuiltinSkill, registerCheckpointSaver, registerChunkBuffer, registerEmbeddingsLattice, registerExistingTool, registerLoggerLattice, registerModelLattice, registerQueueLattice, registerSTTModelLattice, registerSandboxProviderType, registerScheduleLattice, registerStoreLattice, registerTeammateAgent, registerToolLattice, registerVectorStoreLattice, registerVectorStoreProvider, registerWorkflowRun, removeCollectionVectorStore, renderTemplate, resolvePath, sandboxLatticeManager, sanitizeToolCallId, scheduleLatticeManager, serializePluginMeta, setBindingRegistry, setEvalRunService, setMenuRegistry, skillLatticeManager, sqlDatabaseManager, storeLatticeManager, sttModelLatticeManager, toJsonSchema, toSafeStateExpr, toolLatticeManager, truncateIfTooLong, unregisterTeammateAgent, unregisterWorkflowRun, updateFileData, validateAgentInput, validateDSL, validateEncryptionKey, validatePath, validateSkillName, validateToolInput, vectorStoreLatticeManager, vectorStoreProviderManager };
+export { AGENT_TASK_EVENT, type AddMessageParams, Agent, type AgentClient, type AgentExecutor, AgentInstanceManager, type AgentLattice, AgentLatticeManager, type AgentLifecycleEventName, AgentManager, type AgentStreamExecutor, type AgentThreadInterface, BUILTIN_PLUGINS, BUILTIN_SKILLS, type BackendFactory, type BackendProtocol, type BufferStats, type CacheEntry, type CalibrationProbe, type CaseRunResult, type Chunk, ChunkBuffer, ChunkBufferLatticeManager, CollectionLatticeManager, type ColumnInfo, CompositeBackend, type ConflictItem, type ConnectAllChannelsOptions, ConnectionRegistry, ConsoleLoggerClient, type CreateSandboxProviderConfig, type CronFields, CustomMetricsClient, type CustomMiddlewareFactory, CustomMiddlewareRegistry, DEFAULT_CALIBRATION_PROBES, type DatabaseConfig, type DatabaseType, DaytonaInstance, DaytonaProvider, type DaytonaProviderConfig, DefaultScheduleClient, type DeleteResult, DependencyResolver, E2BInstance, E2BProvider, type E2BProviderConfig, EMPTY_CONTENT_WARNING, type EditResult, type EmbeddingsInfo, type EmbeddingsLatticeInterface, EmbeddingsLatticeManager, type EnsureMicrosandboxInput, type EvalRunService, type ExportBundle, type ExportJob, type ExportableEntity, type ExportableEntityDefinition, ExportableEntityRegistry, type ExportableTypeInfo, type FileData, type FileInfo, FileSystemSkillStore, type FileSystemSkillStoreOptions, FilesystemBackend, type FsEntry, type GrepMatch, type IMessageQueueStore, type IMetricsServerClient, type ISqlDatabase, IdRemapper, type ImportApplyResult, type ImportEntityResult, type ImportEntityStatus, type ImportPreviewResult, InMemoryA2AApiKeyStore, InMemoryAssistantStore, InMemoryBindingStore, InMemoryChannelInstallationStore, InMemoryChunkBuffer, InMemoryCollectionStore, InMemoryDatabaseConfigStore, InMemoryMailboxStore, InMemoryMenuStore, InMemoryTaskListStore, InMemoryTaskStore, InMemoryTenantStore, InMemoryThreadMessageQueueStore, InMemoryThreadStore, InMemoryUserStore, InMemoryUserTenantLinkStore, type InsertionItem, type JudgeVerdict, LINE_NUMBER_WIDTH, type LangGraphStateChecker, type LatticeAgentStepConfig, LatticeEval, type LatticeEvalBatchReport, type LatticeEvalCase, type LatticeEvalCaseType, type LatticeEvalCaseWithTemplate, type LatticeEvalConfig, type LatticeEvalLogEvent, type LatticeEvalLogLevel, LatticeEvalProject, type LatticeEvalProjectType, type LatticeEvalResult, type LatticeEvalRubric, LatticeEvalSuite, type LatticeEvalSuiteType, type LatticeEvalTemplate, LocalSandboxInstance, LocalSandboxProvider, type LocalSandboxProviderConfig, type LoggerLattice, LoggerLatticeManager, MAX_LINE_LENGTH, type MailboxMessage, type MailboxStore, type McpLatticeInterface, McpLatticeManager, type McpServerInfo, MemoryBackend, MemoryLatticeManager, MemoryQueueClient, MemoryScheduleStorage, type MessageCompletedEvent, type MessageFailedEvent, type MessageStartedEvent, MessageType, MetricsServerManager, MicrosandboxRemoteInstance, MicrosandboxRemoteProvider, type MicrosandboxRemoteProviderClient, type MicrosandboxRemoteProviderConfig, MicrosandboxServiceClient, type MicrosandboxServiceClientConfig, type MicrosandboxShellExecInput, type ModelConfig, type ModelLatticeInterface, ModelLatticeManager, MysqlDatabase, type OutputType, type PendingMessage, PersonalAssistantConfig, PinoLoggerClient, PluginRegistry, PostgresDatabase, type PreviewError, PrometheusClient, type QueryResult, type QueueLattice, QueueLatticeManager, QueueMode, type QueuePendingEvent, RemoteSandboxInstance, RemoteSandboxProvider, type RemoteSandboxProviderConfig, type Resolution, type ResolutionAction, type ResolveAgentFn, type ResolvedConfig, type RunSandboxConfig, type RuntimeModelConfig, type STTModelInfo, STTModelLattice, type STTModelLatticeInterface, STTModelLatticeManager, type SandboxFileInfo, type SandboxFileService, SandboxFilesystem, type SandboxInstance, type SandboxIsolationLevel, SandboxLatticeManager, type SandboxManagerProtocol, type SandboxProvider, type SandboxProviderFactory, type SandboxProviderType, type SandboxShellService, SandboxSkillStore, type SandboxSkillStoreOptions, type SandboxVolumeDefinition, type ScheduleLattice, ScheduleLatticeManager, type SchedulerMiddlewareOptions, SemanticMetricsClient, type SharePayload, SimpleMemoryVectorStore, type SkillLattice, SkillLatticeManager, type SkillMeta, type SkillResource, SqlDatabaseManager, type StateAndStore, StateBackend, StoreBackend, type StoreLattice, StoreLatticeManager, type StoreType, type StoreTypeMap, TOOL_RESULT_TOKEN_LIMIT, TRUNCATION_GUIDANCE, type TableInfo, type TableSchema, type TaskEvent, type TaskListStore, type TaskSpec, TaskStatus, type TaskUpdatable, TeamAgentGraphBuilder, type TeamConfig, type TeamMiddlewareOptions, type TeamTask, type TeammateSpec, type TeammateToolsOptions, type ThreadBuffer, type ThreadBufferConfig, type ThreadBusyEvent, type ThreadIdleEvent, type ThreadInfo, type ThreadQueueConfig, type ThreadState, ThreadStatus, type ThreadStatusChangedEvent, TokenCache, type ToolDefinition, type ToolLattice, ToolLatticeManager, type UnknownToolHandlerConfig, type VectorStoreLatticeInterface, VectorStoreLatticeManager, type VectorStoreProviderLattice, VectorStoreProviderManager, VolumeFilesystem, type VolumeFsClient, type WorkflowValidationError, type WriteResult, abortWorkflowRun, agentInstanceManager, agentLatticeManager, buildGrepResultsDict, buildInput, buildNamedVolumeName, buildSandboxMetadataEnv, buildSkillFile, buildStateAnnotation, buildTableName, checkEmptyContent, clearEncryptionKeyCache, clearEvalRunService, collectionLatticeManager, compileInternal, compileWorkflow, computeSandboxName, configureStores, connectAllChannels, createAgentNode, createAgentTeam, createExecuteSqlQueryTool, createFileData, createInfoSqlTool, createListMetricsDataSourcesTool, createListMetricsServersTool, createListTablesSqlTool, createMapNode, createModelSelectorMiddleware, createNodeHandler, createQueryCheckerSqlTool, createQueryMetricDefinitionTool, createQueryMetricsListTool, createQuerySemanticMetricDataTool, createQuerySqlTool, createQueryTableDefinitionTool, createQueryTablesListTool, createResourceAddress, createSandboxProvider, createSchedulerMiddleware, createSharePayload, createTaskMiddleware, createTeamMiddleware, createTeammateTools, createUnknownToolHandlerMiddleware, createWidgetMiddleware, decrypt, describeCronExpression, documentLearningPlugin, documentParserPlugin, embeddingsLatticeManager, encrypt, ensureBuiltinAgentsForTenant, evaluateLatticeCaseWithLogs, eventBus, eventBus as eventBusDefault, extractFetcherError, extractOutput, fileDataToString, formatContentWithLineNumbers, formatGrepMatches, formatGrepResults, formatReadResponse, generateToken, getAgentClient, getAgentConfig, getAllAgentConfigs, getAllBuiltInSkillMetas, getAllToolDefinitions, getBindingRegistry, getBuiltInSkillContent, getBuiltInSkillMeta, getBuiltInSkillNames, getCheckpointSaver, getChunkBuffer, getCollectionEntryCount, getEmbeddingsClient, getEmbeddingsLattice, getEncryptionKey, getEvalRunService, getLoggerLattice, getMenuRegistry, getModelLattice, getNextCronTime, getOrCreateCollectionVectorStore, getQueueLattice, getSTTClient, getSTTClientWithTenant, getSTTModelLattice, getSandBoxManager, getScheduleLattice, getStoreLattice, getToolClient, getToolDefinition, getToolLattice, getVectorStoreClient, getVectorStoreLattice, getVectorStoreProvider, getWorkflowSignal, globSearchFiles, grepMatchesFromFiles, grepSearchFiles, hasChunkBuffer, invokeWithRetry, isBuiltInSkill, isUsingDefaultKey, isValidCronExpression, isValidSandboxName, isValidSkillName, listCollectionEntries, listSandboxProviderTypes, loggerLatticeManager, mcpManager, metricsServerManager, modelLatticeManager, normalizeSandboxName, parallelLimit, parseCronExpression, parseJudgeVerdict, parseSkillFrontmatter, parseYaml, performStringReplacement, queueLatticeManager, registerAgentLattice, registerAgentLatticeWithTenant, registerAgentLattices, registerBuiltinSkill, registerCheckpointSaver, registerChunkBuffer, registerEmbeddingsLattice, registerExistingTool, registerLoggerLattice, registerModelLattice, registerQueueLattice, registerSTTModelLattice, registerSandboxProviderType, registerScheduleLattice, registerStoreLattice, registerTeammateAgent, registerToolLattice, registerVectorStoreLattice, registerVectorStoreProvider, registerWorkflowRun, removeCollectionVectorStore, renderTemplate, resolveJudgeModelKey, resolvePath, sandboxLatticeManager, sanitizeToolCallId, scheduleLatticeManager, serializePluginMeta, setBindingRegistry, setEvalRunService, setMenuRegistry, skillLatticeManager, sqlDatabaseManager, storeLatticeManager, sttModelLatticeManager, toJsonSchema, toSafeStateExpr, toolLatticeManager, truncateIfTooLong, unregisterTeammateAgent, unregisterWorkflowRun, updateFileData, validateAgentInput, validateDSL, validateEncryptionKey, validatePath, validateSkillName, validateToolInput, vectorStoreLatticeManager, vectorStoreProviderManager };

package/dist/index.d.ts

@@ -5494,12 +5494,13 @@ interface EvalRunService {
      *   executes in. Comes from the CALLER's runConfig, NOT from the eval
      *   project's targetServerConfig. Test cases are environment-agnostic;
      *   the run environment is decided at run time.
+     * @param taskId - Optional training task (round) this run belongs to.
      * @returns The newly created run ID
      */
     startRun(tenantId: string, projectId: string, suiteIds?: string[], caseIds?: string[], runConfig?: {
         workspaceId?: string;
         projectId?: string;
-    }): Promise<string>;
+    }, taskId?: string): Promise<string>;
     /**
      * Abort a running evaluation.
      *
@@ -5695,6 +5696,17 @@ declare class LatticeEvalSuite {
  * A judge that cannot discriminate these should not be trusted with real cases.
  */
 declare const DEFAULT_CALIBRATION_PROBES: CalibrationProbe[];
+/**
+ * Resolve the judge model for an eval project: an explicit judgeModelKey wins;
+ * otherwise the model registered under the "default" key (the convention apps
+ * use — see examples/deep_research), then the first registered model as a last
+ * resort. Throws only when no model is registered at all.
+ */
+declare function resolveJudgeModelKey(explicitKey: string | undefined, manager: {
+    getAllLattices: () => Array<{
+        key: string;
+    }>;
+}): string;
 /**
  * Manages a project with multiple evaluation suites.
  * Registers a per-project judge agent (keyed by project name and tenant)
@@ -8202,4 +8214,4 @@ declare class IdRemapper {
     remapSkillIds(graphDefinition: Record<string, unknown>, skillRemap: Record<string, string>): Record<string, unknown>;
 }
 
-export { AGENT_TASK_EVENT, type AddMessageParams, Agent, type AgentClient, type AgentExecutor, AgentInstanceManager, type AgentLattice, AgentLatticeManager, type AgentLifecycleEventName, AgentManager, type AgentStreamExecutor, type AgentThreadInterface, BUILTIN_PLUGINS, BUILTIN_SKILLS, type BackendFactory, type BackendProtocol, type BufferStats, type CacheEntry, type CalibrationProbe, type CaseRunResult, type Chunk, ChunkBuffer, ChunkBufferLatticeManager, CollectionLatticeManager, type ColumnInfo, CompositeBackend, type ConflictItem, type ConnectAllChannelsOptions, ConnectionRegistry, ConsoleLoggerClient, type CreateSandboxProviderConfig, type CronFields, CustomMetricsClient, type CustomMiddlewareFactory, CustomMiddlewareRegistry, DEFAULT_CALIBRATION_PROBES, type DatabaseConfig, type DatabaseType, DaytonaInstance, DaytonaProvider, type DaytonaProviderConfig, DefaultScheduleClient, type DeleteResult, DependencyResolver, E2BInstance, E2BProvider, type E2BProviderConfig, EMPTY_CONTENT_WARNING, type EditResult, type EmbeddingsInfo, type EmbeddingsLatticeInterface, EmbeddingsLatticeManager, type EnsureMicrosandboxInput, type EvalRunService, type ExportBundle, type ExportJob, type ExportableEntity, type ExportableEntityDefinition, ExportableEntityRegistry, type ExportableTypeInfo, type FileData, type FileInfo, FileSystemSkillStore, type FileSystemSkillStoreOptions, FilesystemBackend, type FsEntry, type GrepMatch, type IMessageQueueStore, type IMetricsServerClient, type ISqlDatabase, IdRemapper, type ImportApplyResult, type ImportEntityResult, type ImportEntityStatus, type ImportPreviewResult, InMemoryA2AApiKeyStore, InMemoryAssistantStore, InMemoryBindingStore, InMemoryChannelInstallationStore, InMemoryChunkBuffer, InMemoryCollectionStore, InMemoryDatabaseConfigStore, InMemoryMailboxStore, InMemoryMenuStore, InMemoryTaskListStore, InMemoryTaskStore, InMemoryTenantStore, InMemoryThreadMessageQueueStore, InMemoryThreadStore, InMemoryUserStore, InMemoryUserTenantLinkStore, type InsertionItem, type JudgeVerdict, LINE_NUMBER_WIDTH, type LangGraphStateChecker, type LatticeAgentStepConfig, LatticeEval, type LatticeEvalBatchReport, type LatticeEvalCase, type LatticeEvalCaseType, type LatticeEvalCaseWithTemplate, type LatticeEvalConfig, type LatticeEvalLogEvent, type LatticeEvalLogLevel, LatticeEvalProject, type LatticeEvalProjectType, type LatticeEvalResult, type LatticeEvalRubric, LatticeEvalSuite, type LatticeEvalSuiteType, type LatticeEvalTemplate, LocalSandboxInstance, LocalSandboxProvider, type LocalSandboxProviderConfig, type LoggerLattice, LoggerLatticeManager, MAX_LINE_LENGTH, type MailboxMessage, type MailboxStore, type McpLatticeInterface, McpLatticeManager, type McpServerInfo, MemoryBackend, MemoryLatticeManager, MemoryQueueClient, MemoryScheduleStorage, type MessageCompletedEvent, type MessageFailedEvent, type MessageStartedEvent, MessageType, MetricsServerManager, MicrosandboxRemoteInstance, MicrosandboxRemoteProvider, type MicrosandboxRemoteProviderClient, type MicrosandboxRemoteProviderConfig, MicrosandboxServiceClient, type MicrosandboxServiceClientConfig, type MicrosandboxShellExecInput, type ModelConfig, type ModelLatticeInterface, ModelLatticeManager, MysqlDatabase, type OutputType, type PendingMessage, PersonalAssistantConfig, PinoLoggerClient, PluginRegistry, PostgresDatabase, type PreviewError, PrometheusClient, type QueryResult, type QueueLattice, QueueLatticeManager, QueueMode, type QueuePendingEvent, RemoteSandboxInstance, RemoteSandboxProvider, type RemoteSandboxProviderConfig, type Resolution, type ResolutionAction, type ResolveAgentFn, type ResolvedConfig, type RunSandboxConfig, type RuntimeModelConfig, type STTModelInfo, STTModelLattice, type STTModelLatticeInterface, STTModelLatticeManager, type SandboxFileInfo, type SandboxFileService, SandboxFilesystem, type SandboxInstance, type SandboxIsolationLevel, SandboxLatticeManager, type SandboxManagerProtocol, type SandboxProvider, type SandboxProviderFactory, type SandboxProviderType, type SandboxShellService, SandboxSkillStore, type SandboxSkillStoreOptions, type SandboxVolumeDefinition, type ScheduleLattice, ScheduleLatticeManager, type SchedulerMiddlewareOptions, SemanticMetricsClient, type SharePayload, SimpleMemoryVectorStore, type SkillLattice, SkillLatticeManager, type SkillMeta, type SkillResource, SqlDatabaseManager, type StateAndStore, StateBackend, StoreBackend, type StoreLattice, StoreLatticeManager, type StoreType, type StoreTypeMap, TOOL_RESULT_TOKEN_LIMIT, TRUNCATION_GUIDANCE, type TableInfo, type TableSchema, type TaskEvent, type TaskListStore, type TaskSpec, TaskStatus, type TaskUpdatable, TeamAgentGraphBuilder, type TeamConfig, type TeamMiddlewareOptions, type TeamTask, type TeammateSpec, type TeammateToolsOptions, type ThreadBuffer, type ThreadBufferConfig, type ThreadBusyEvent, type ThreadIdleEvent, type ThreadInfo, type ThreadQueueConfig, type ThreadState, ThreadStatus, type ThreadStatusChangedEvent, TokenCache, type ToolDefinition, type ToolLattice, ToolLatticeManager, type UnknownToolHandlerConfig, type VectorStoreLatticeInterface, VectorStoreLatticeManager, type VectorStoreProviderLattice, VectorStoreProviderManager, VolumeFilesystem, type VolumeFsClient, type WorkflowValidationError, type WriteResult, abortWorkflowRun, agentInstanceManager, agentLatticeManager, buildGrepResultsDict, buildInput, buildNamedVolumeName, buildSandboxMetadataEnv, buildSkillFile, buildStateAnnotation, buildTableName, checkEmptyContent, clearEncryptionKeyCache, clearEvalRunService, collectionLatticeManager, compileInternal, compileWorkflow, computeSandboxName, configureStores, connectAllChannels, createAgentNode, createAgentTeam, createExecuteSqlQueryTool, createFileData, createInfoSqlTool, createListMetricsDataSourcesTool, createListMetricsServersTool, createListTablesSqlTool, createMapNode, createModelSelectorMiddleware, createNodeHandler, createQueryCheckerSqlTool, createQueryMetricDefinitionTool, createQueryMetricsListTool, createQuerySemanticMetricDataTool, createQuerySqlTool, createQueryTableDefinitionTool, createQueryTablesListTool, createResourceAddress, createSandboxProvider, createSchedulerMiddleware, createSharePayload, createTaskMiddleware, createTeamMiddleware, createTeammateTools, createUnknownToolHandlerMiddleware, createWidgetMiddleware, decrypt, describeCronExpression, documentLearningPlugin, documentParserPlugin, embeddingsLatticeManager, encrypt, ensureBuiltinAgentsForTenant, evaluateLatticeCaseWithLogs, eventBus, eventBus as eventBusDefault, extractFetcherError, extractOutput, fileDataToString, formatContentWithLineNumbers, formatGrepMatches, formatGrepResults, formatReadResponse, generateToken, getAgentClient, getAgentConfig, getAllAgentConfigs, getAllBuiltInSkillMetas, getAllToolDefinitions, getBindingRegistry, getBuiltInSkillContent, getBuiltInSkillMeta, getBuiltInSkillNames, getCheckpointSaver, getChunkBuffer, getCollectionEntryCount, getEmbeddingsClient, getEmbeddingsLattice, getEncryptionKey, getEvalRunService, getLoggerLattice, getMenuRegistry, getModelLattice, getNextCronTime, getOrCreateCollectionVectorStore, getQueueLattice, getSTTClient, getSTTClientWithTenant, getSTTModelLattice, getSandBoxManager, getScheduleLattice, getStoreLattice, getToolClient, getToolDefinition, getToolLattice, getVectorStoreClient, getVectorStoreLattice, getVectorStoreProvider, getWorkflowSignal, globSearchFiles, grepMatchesFromFiles, grepSearchFiles, hasChunkBuffer, invokeWithRetry, isBuiltInSkill, isUsingDefaultKey, isValidCronExpression, isValidSandboxName, isValidSkillName, listCollectionEntries, listSandboxProviderTypes, loggerLatticeManager, mcpManager, metricsServerManager, modelLatticeManager, normalizeSandboxName, parallelLimit, parseCronExpression, parseJudgeVerdict, parseSkillFrontmatter, parseYaml, performStringReplacement, queueLatticeManager, registerAgentLattice, registerAgentLatticeWithTenant, registerAgentLattices, registerBuiltinSkill, registerCheckpointSaver, registerChunkBuffer, registerEmbeddingsLattice, registerExistingTool, registerLoggerLattice, registerModelLattice, registerQueueLattice, registerSTTModelLattice, registerSandboxProviderType, registerScheduleLattice, registerStoreLattice, registerTeammateAgent, registerToolLattice, registerVectorStoreLattice, registerVectorStoreProvider, registerWorkflowRun, removeCollectionVectorStore, renderTemplate, resolvePath, sandboxLatticeManager, sanitizeToolCallId, scheduleLatticeManager, serializePluginMeta, setBindingRegistry, setEvalRunService, setMenuRegistry, skillLatticeManager, sqlDatabaseManager, storeLatticeManager, sttModelLatticeManager, toJsonSchema, toSafeStateExpr, toolLatticeManager, truncateIfTooLong, unregisterTeammateAgent, unregisterWorkflowRun, updateFileData, validateAgentInput, validateDSL, validateEncryptionKey, validatePath, validateSkillName, validateToolInput, vectorStoreLatticeManager, vectorStoreProviderManager };
+export { AGENT_TASK_EVENT, type AddMessageParams, Agent, type AgentClient, type AgentExecutor, AgentInstanceManager, type AgentLattice, AgentLatticeManager, type AgentLifecycleEventName, AgentManager, type AgentStreamExecutor, type AgentThreadInterface, BUILTIN_PLUGINS, BUILTIN_SKILLS, type BackendFactory, type BackendProtocol, type BufferStats, type CacheEntry, type CalibrationProbe, type CaseRunResult, type Chunk, ChunkBuffer, ChunkBufferLatticeManager, CollectionLatticeManager, type ColumnInfo, CompositeBackend, type ConflictItem, type ConnectAllChannelsOptions, ConnectionRegistry, ConsoleLoggerClient, type CreateSandboxProviderConfig, type CronFields, CustomMetricsClient, type CustomMiddlewareFactory, CustomMiddlewareRegistry, DEFAULT_CALIBRATION_PROBES, type DatabaseConfig, type DatabaseType, DaytonaInstance, DaytonaProvider, type DaytonaProviderConfig, DefaultScheduleClient, type DeleteResult, DependencyResolver, E2BInstance, E2BProvider, type E2BProviderConfig, EMPTY_CONTENT_WARNING, type EditResult, type EmbeddingsInfo, type EmbeddingsLatticeInterface, EmbeddingsLatticeManager, type EnsureMicrosandboxInput, type EvalRunService, type ExportBundle, type ExportJob, type ExportableEntity, type ExportableEntityDefinition, ExportableEntityRegistry, type ExportableTypeInfo, type FileData, type FileInfo, FileSystemSkillStore, type FileSystemSkillStoreOptions, FilesystemBackend, type FsEntry, type GrepMatch, type IMessageQueueStore, type IMetricsServerClient, type ISqlDatabase, IdRemapper, type ImportApplyResult, type ImportEntityResult, type ImportEntityStatus, type ImportPreviewResult, InMemoryA2AApiKeyStore, InMemoryAssistantStore, InMemoryBindingStore, InMemoryChannelInstallationStore, InMemoryChunkBuffer, InMemoryCollectionStore, InMemoryDatabaseConfigStore, InMemoryMailboxStore, InMemoryMenuStore, InMemoryTaskListStore, InMemoryTaskStore, InMemoryTenantStore, InMemoryThreadMessageQueueStore, InMemoryThreadStore, InMemoryUserStore, InMemoryUserTenantLinkStore, type InsertionItem, type JudgeVerdict, LINE_NUMBER_WIDTH, type LangGraphStateChecker, type LatticeAgentStepConfig, LatticeEval, type LatticeEvalBatchReport, type LatticeEvalCase, type LatticeEvalCaseType, type LatticeEvalCaseWithTemplate, type LatticeEvalConfig, type LatticeEvalLogEvent, type LatticeEvalLogLevel, LatticeEvalProject, type LatticeEvalProjectType, type LatticeEvalResult, type LatticeEvalRubric, LatticeEvalSuite, type LatticeEvalSuiteType, type LatticeEvalTemplate, LocalSandboxInstance, LocalSandboxProvider, type LocalSandboxProviderConfig, type LoggerLattice, LoggerLatticeManager, MAX_LINE_LENGTH, type MailboxMessage, type MailboxStore, type McpLatticeInterface, McpLatticeManager, type McpServerInfo, MemoryBackend, MemoryLatticeManager, MemoryQueueClient, MemoryScheduleStorage, type MessageCompletedEvent, type MessageFailedEvent, type MessageStartedEvent, MessageType, MetricsServerManager, MicrosandboxRemoteInstance, MicrosandboxRemoteProvider, type MicrosandboxRemoteProviderClient, type MicrosandboxRemoteProviderConfig, MicrosandboxServiceClient, type MicrosandboxServiceClientConfig, type MicrosandboxShellExecInput, type ModelConfig, type ModelLatticeInterface, ModelLatticeManager, MysqlDatabase, type OutputType, type PendingMessage, PersonalAssistantConfig, PinoLoggerClient, PluginRegistry, PostgresDatabase, type PreviewError, PrometheusClient, type QueryResult, type QueueLattice, QueueLatticeManager, QueueMode, type QueuePendingEvent, RemoteSandboxInstance, RemoteSandboxProvider, type RemoteSandboxProviderConfig, type Resolution, type ResolutionAction, type ResolveAgentFn, type ResolvedConfig, type RunSandboxConfig, type RuntimeModelConfig, type STTModelInfo, STTModelLattice, type STTModelLatticeInterface, STTModelLatticeManager, type SandboxFileInfo, type SandboxFileService, SandboxFilesystem, type SandboxInstance, type SandboxIsolationLevel, SandboxLatticeManager, type SandboxManagerProtocol, type SandboxProvider, type SandboxProviderFactory, type SandboxProviderType, type SandboxShellService, SandboxSkillStore, type SandboxSkillStoreOptions, type SandboxVolumeDefinition, type ScheduleLattice, ScheduleLatticeManager, type SchedulerMiddlewareOptions, SemanticMetricsClient, type SharePayload, SimpleMemoryVectorStore, type SkillLattice, SkillLatticeManager, type SkillMeta, type SkillResource, SqlDatabaseManager, type StateAndStore, StateBackend, StoreBackend, type StoreLattice, StoreLatticeManager, type StoreType, type StoreTypeMap, TOOL_RESULT_TOKEN_LIMIT, TRUNCATION_GUIDANCE, type TableInfo, type TableSchema, type TaskEvent, type TaskListStore, type TaskSpec, TaskStatus, type TaskUpdatable, TeamAgentGraphBuilder, type TeamConfig, type TeamMiddlewareOptions, type TeamTask, type TeammateSpec, type TeammateToolsOptions, type ThreadBuffer, type ThreadBufferConfig, type ThreadBusyEvent, type ThreadIdleEvent, type ThreadInfo, type ThreadQueueConfig, type ThreadState, ThreadStatus, type ThreadStatusChangedEvent, TokenCache, type ToolDefinition, type ToolLattice, ToolLatticeManager, type UnknownToolHandlerConfig, type VectorStoreLatticeInterface, VectorStoreLatticeManager, type VectorStoreProviderLattice, VectorStoreProviderManager, VolumeFilesystem, type VolumeFsClient, type WorkflowValidationError, type WriteResult, abortWorkflowRun, agentInstanceManager, agentLatticeManager, buildGrepResultsDict, buildInput, buildNamedVolumeName, buildSandboxMetadataEnv, buildSkillFile, buildStateAnnotation, buildTableName, checkEmptyContent, clearEncryptionKeyCache, clearEvalRunService, collectionLatticeManager, compileInternal, compileWorkflow, computeSandboxName, configureStores, connectAllChannels, createAgentNode, createAgentTeam, createExecuteSqlQueryTool, createFileData, createInfoSqlTool, createListMetricsDataSourcesTool, createListMetricsServersTool, createListTablesSqlTool, createMapNode, createModelSelectorMiddleware, createNodeHandler, createQueryCheckerSqlTool, createQueryMetricDefinitionTool, createQueryMetricsListTool, createQuerySemanticMetricDataTool, createQuerySqlTool, createQueryTableDefinitionTool, createQueryTablesListTool, createResourceAddress, createSandboxProvider, createSchedulerMiddleware, createSharePayload, createTaskMiddleware, createTeamMiddleware, createTeammateTools, createUnknownToolHandlerMiddleware, createWidgetMiddleware, decrypt, describeCronExpression, documentLearningPlugin, documentParserPlugin, embeddingsLatticeManager, encrypt, ensureBuiltinAgentsForTenant, evaluateLatticeCaseWithLogs, eventBus, eventBus as eventBusDefault, extractFetcherError, extractOutput, fileDataToString, formatContentWithLineNumbers, formatGrepMatches, formatGrepResults, formatReadResponse, generateToken, getAgentClient, getAgentConfig, getAllAgentConfigs, getAllBuiltInSkillMetas, getAllToolDefinitions, getBindingRegistry, getBuiltInSkillContent, getBuiltInSkillMeta, getBuiltInSkillNames, getCheckpointSaver, getChunkBuffer, getCollectionEntryCount, getEmbeddingsClient, getEmbeddingsLattice, getEncryptionKey, getEvalRunService, getLoggerLattice, getMenuRegistry, getModelLattice, getNextCronTime, getOrCreateCollectionVectorStore, getQueueLattice, getSTTClient, getSTTClientWithTenant, getSTTModelLattice, getSandBoxManager, getScheduleLattice, getStoreLattice, getToolClient, getToolDefinition, getToolLattice, getVectorStoreClient, getVectorStoreLattice, getVectorStoreProvider, getWorkflowSignal, globSearchFiles, grepMatchesFromFiles, grepSearchFiles, hasChunkBuffer, invokeWithRetry, isBuiltInSkill, isUsingDefaultKey, isValidCronExpression, isValidSandboxName, isValidSkillName, listCollectionEntries, listSandboxProviderTypes, loggerLatticeManager, mcpManager, metricsServerManager, modelLatticeManager, normalizeSandboxName, parallelLimit, parseCronExpression, parseJudgeVerdict, parseSkillFrontmatter, parseYaml, performStringReplacement, queueLatticeManager, registerAgentLattice, registerAgentLatticeWithTenant, registerAgentLattices, registerBuiltinSkill, registerCheckpointSaver, registerChunkBuffer, registerEmbeddingsLattice, registerExistingTool, registerLoggerLattice, registerModelLattice, registerQueueLattice, registerSTTModelLattice, registerSandboxProviderType, registerScheduleLattice, registerStoreLattice, registerTeammateAgent, registerToolLattice, registerVectorStoreLattice, registerVectorStoreProvider, registerWorkflowRun, removeCollectionVectorStore, renderTemplate, resolveJudgeModelKey, resolvePath, sandboxLatticeManager, sanitizeToolCallId, scheduleLatticeManager, serializePluginMeta, setBindingRegistry, setEvalRunService, setMenuRegistry, skillLatticeManager, sqlDatabaseManager, storeLatticeManager, sttModelLatticeManager, toJsonSchema, toSafeStateExpr, toolLatticeManager, truncateIfTooLong, unregisterTeammateAgent, unregisterWorkflowRun, updateFileData, validateAgentInput, validateDSL, validateEncryptionKey, validatePath, validateSkillName, validateToolInput, vectorStoreLatticeManager, vectorStoreProviderManager };

package/dist/index.js

@@ -1852,6 +1852,7 @@ __export(index_exports, {
   registerWorkflowRun: () => registerWorkflowRun,
   removeCollectionVectorStore: () => removeCollectionVectorStore,
   renderTemplate: () => renderTemplate,
+  resolveJudgeModelKey: () => resolveJudgeModelKey,
   resolvePath: () => resolvePath,
   sandboxLatticeManager: () => sandboxLatticeManager,
   sanitizeToolCallId: () => sanitizeToolCallId,
@@ -4501,6 +4502,7 @@ var InMemoryTaskStore = class {
       projectId: params.projectId,
       dueDate: params.dueDate,
       metadata: params.metadata,
+      files: params.files,
       parentId: params.parentId,
       sourceId: params.sourceId,
       context: params.context,
@@ -4558,6 +4560,7 @@ var InMemoryTaskStore = class {
     const updated = {
       ...existing,
       ...updates,
+      files: updates.files !== void 0 ? updates.files : existing.files,
       updatedAt: /* @__PURE__ */ new Date()
     };
     tenantTasks.set(id, updated);
@@ -16043,6 +16046,11 @@ var manageTaskSchema = import_zod43.z.object({
   dependencies: import_zod43.z.array(import_zod43.z.string()).optional().describe("List of task IDs that must be completed before this task can start"),
   result: import_zod43.z.string().optional().describe("Result summary when task is completed"),
   failureReason: import_zod43.z.string().optional().describe("Reason for failure (use when status='failed')"),
+  files: import_zod43.z.array(import_zod43.z.object({
+    uri: import_zod43.z.string().describe("Uniquely locates the resource: http(s):// URL, /s/:token share, or sandbox path"),
+    name: import_zod43.z.string().optional().describe("Display name"),
+    addedBy: import_zod43.z.enum(["user", "agent"]).optional().describe("Who attached the file")
+  })).optional().describe("File references attached to this task"),
   summary: import_zod43.z.string().optional().describe("Brief summary of the operation")
 });
 function buildReviewMarkdown(task) {
@@ -16101,7 +16109,8 @@ function createTaskMiddleware() {
           requireReview: input.requireReview,
           dependencies: input.dependencies,
           workspaceId,
-          projectId
+          projectId,
+          files: input.files
         });
         return JSON.stringify({ success: true, data: task });
       }
@@ -16180,7 +16189,8 @@ function createTaskMiddleware() {
           "result",
           "failureReason",
           "requireReview",
-          "dependencies"
+          "dependencies",
+          "files"
         ];
         for (const field of settableFields) {
           if (input[field] !== void 0) {
@@ -20012,11 +20022,11 @@ ${BASE_PROMPT}` : BASE_PROMPT;
       defaultInterruptOn: interruptOn,
       subagents,
       generalPurposeAgent: true
-    }),
-    // Enables Anthropic prompt caching for improved performance and reduced costs
-    (0, import_langchain54.anthropicPromptCachingMiddleware)({
-      unsupportedModelBehavior: "ignore"
     })
+    // Enables Anthropic prompt caching for improved performance and reduced costs
+    // anthropicPromptCachingMiddleware({
+    //   unsupportedModelBehavior: "ignore",
+    // })
   ];
   if (interruptOn) {
     middleware.push((0, import_langchain54.humanInTheLoopMiddleware)({ interruptOn }));
@@ -23629,7 +23639,7 @@ registerToolLattice(
   "update_agent",
   {
     name: "update_agent",
-    description: "Update an existing agent's configuration. Provide the agent ID and the fields to change. Returns the updated agent's ID and name. NOTE: For user approval/confirmation scenarios, add middleware with type: 'ask_user_to_clarify'.",
+    description: "Update an existing agent's configuration. Provide the agent ID and the fields to change. Returns the updated agent's ID and name. NOTE: For user approval/confirmation scenarios, add middleware with type: 'ask_user_to_clarify'. IMPORTANT: do NOT set 'modelKey' unless the user explicitly specified a model \u2014 leaving it unset makes the runtime use the 'default' model.",
     schema: updateAgentSchema
   },
   async (input, exeConfig) => {
@@ -23704,6 +23714,27 @@ registerToolLattice(
     }
   }
 );
+registerToolLattice(
+  "list_models",
+  {
+    name: "list_models",
+    description: "List all registered models. Returns each model's key (use this string as modelKey when creating eval projects via manage_eval create_project, or as an agent's modelKey in create_agent/update_agent) and its display name.",
+    schema: import_zod47.default.object({})
+  },
+  async (_input) => {
+    try {
+      const lattices = modelLatticeManager.getAllLattices();
+      return JSON.stringify({
+        models: lattices.map((l) => ({
+          key: l.key,
+          name: l.client.name ?? l.key
+        }))
+      });
+    } catch (error) {
+      return JSON.stringify({ error: `Failed to list models: ${error.message}` });
+    }
+  }
+);
 registerToolLattice(
   "invoke_agent",
   {
@@ -23835,6 +23866,24 @@ The skills document WHY and HOW; these gates are the unskippable
 minimum. If you cannot satisfy a gate (e.g. user says skip), record it
 and proceed only on the user's explicit instruction.
 
+LEARNING ROUND KICKOFF \u2014 a message that names an existing target agent
+id AND an existing tracking task id (a "learning round"). This protocol
+OVERRIDES the defaults above:
+- The target agent ALREADY EXISTS (an empty placeholder). Build and
+  refine it via update_agent on that exact id. NEVER call create_agent \u2014
+  a new agent would disconnect the round's tracking.
+- The parent task ALREADY EXISTS \u2014 your create-a-task-first duty is
+  satisfied by it. Create subtasks with manage_task under its id
+  (parentId); NEVER create a new parent task for the round.
+- The round is pre-approved \u2014 skip the DESIGN\u2192CONFIRM gates: show the
+  design in your reply, then build directly.
+- Do NOT set modelKey in update_agent unless the user explicitly named a
+  model \u2014 leaving it unset makes the runtime use the 'default' model.
+- Every update_agent call must be mirrored by a manage_task work item
+  whose summary names the changed keys (e.g. "update_agent: prompt,
+  modelKey") \u2014 the round feed highlights these so the user can see what
+  changed between iterations.
+
 Your sub-skills (accessible via the MOC or direct loading):
 - [[learn-capability]] \u2014 Learn from any source material \u2192 skills + agents
 - [[agent-build]] \u2014 Design and build single agents (REACT/DEEP_AGENT)
@@ -23949,6 +23998,7 @@ var agentArchitectConfig = {
   tools: [
     "list_agents",
     "list_tools",
+    "list_models",
     "list_middleware_types",
     "list_connections",
     "get_agent",
@@ -28423,14 +28473,22 @@ var DEFAULT_CALIBRATION_PROBES = [
     expectedPass: false
   }
 ];
+function resolveJudgeModelKey(explicitKey, manager) {
+  if (explicitKey) return explicitKey;
+  const models = manager.getAllLattices();
+  const defaultModel = models.find((m) => m.key === "default");
+  if (defaultModel) return defaultModel.key;
+  const first = models[0];
+  if (!first) {
+    throw new Error("No model registered \u2014 register a model or provide a judgeModelKey");
+  }
+  return first.key;
+}
 var LatticeEvalProject = class {
   constructor(project, onCaseComplete) {
     this.suites = /* @__PURE__ */ new Map();
     this.project = project;
-    const judgeModelKey = this.project.judge_agent_config.modelKey;
-    if (!judgeModelKey) {
-      throw new Error("judge_agent_config.modelKey is required \u2014 only pre-registered models are allowed");
-    }
+    const judgeModelKey = resolveJudgeModelKey(this.project.judge_agent_config.modelKey, modelLatticeManager);
     if (!modelLatticeManager.hasLattice(judgeModelKey)) {
       throw new Error(
         `Judge model "${judgeModelKey}" not found. Register it first or provide a valid modelKey.`
@@ -30832,6 +30890,7 @@ function createManageEvalTool() {
     description: import_zod64.z.string().optional(),
     judgeModelKey: import_zod64.z.string().optional(),
     concurrency: import_zod64.z.number().optional(),
+    targetAgentId: import_zod64.z.string().optional().describe("Optional for create_project \u2014 the agent this eval project verifies. Recorded in targetServerConfig so the agent's detail page can find its eval data without relying on project naming."),
     suiteId: import_zod64.z.string().optional(),
     caseId: import_zod64.z.string().optional(),
     inputMessage: import_zod64.z.string().optional(),
@@ -30860,10 +30919,16 @@ function createManageEvalTool() {
             data = await store.createProject(tid, (0, import_uuid12.v4)(), {
               name: input.name,
               description: input.description,
-              judgeModelConfig: { modelKey: input.judgeModelKey },
+              // No judge key given → default to the 'default' model key (the same
+              // convention the UI model selector uses); never an empty modelKey.
+              judgeModelConfig: { modelKey: input.judgeModelKey ?? "default" },
               targetServerConfig: {
                 workspace_id: ctx.workspaceId,
-                project_id: ctx.projectId
+                project_id: ctx.projectId,
+                // The project↔agent association lives here (not in the project
+                // name): an eval project is agent-agnostic by design, this
+                // pointer is merely the agent's designated verifier.
+                ...input.targetAgentId ? { targetAgentId: input.targetAgentId } : {}
               },
               concurrency: input.concurrency ?? 3
             });
@@ -30930,8 +30995,8 @@ function createManageEvalTool() {
       name: "manage_eval",
       description: `Create, update, delete evaluation projects, suites, and test cases.
  
-Project: create_project(name, description?, judgeModelKey?, concurrency?) | update_project | delete_project
-  judgeModelKey defaults to first available model. concurrency defaults to 3.
+Project: create_project(name, description?, judgeModelKey?, concurrency?, targetAgentId?) | update_project | delete_project
+  judgeModelKey defaults to the "default" model. concurrency defaults to 3.
   delete_project rejected if active runs exist.
   **When creating a project from within a workspace, the workspace/project context is
   automatically bound \u2014 eval runs will execute in the same workspace.**
@@ -30954,6 +31019,7 @@ function createRunEvalTool() {
     suiteIds: import_zod64.z.array(import_zod64.z.string()).optional().describe("Optional for start \u2014 only run these suites (e.g. dev set only). Omit to run all."),
     caseIds: import_zod64.z.array(import_zod64.z.string()).optional().describe("Optional for start \u2014 only run these cases across the selected suites. Omit to run all cases in those suites."),
     runId: import_zod64.z.string().optional().describe("Required for status, resume, abort"),
+    taskId: import_zod64.z.string().optional().describe("Optional for start \u2014 training task ID this run belongs to (round association)"),
     sleepMs: import_zod64.z.number().int().min(0).max(12e4).optional().describe("Optional for status \u2014 sleep this many ms BEFORE checking the run, to pace polling (e.g. 15000 \u2192 30000 \u2192 60000 \u2192 120000). Omit to check immediately."),
     wait: import_zod64.z.boolean().optional().describe("Optional for start \u2014 defaults to true: block synchronously (up to ~150s) and return final results in one call. Set false to return the runId immediately and poll.")
   });
@@ -30971,7 +31037,7 @@ function createRunEvalTool() {
           switch (input.action) {
             case "start": {
               const ctx = workspaceContext(exeConfig);
-              const runId = await svc.startRun(tid, input.projectId, input.suiteIds, input.caseIds, ctx);
+              const runId = await svc.startRun(tid, input.projectId, input.suiteIds, input.caseIds, ctx, input.taskId);
               if (input.wait === false) {
                 data = sanitize({ runId, message: "Run started. Poll with run_eval status (backoff: 15s\u219230s\u219260s\u2192max 120s)." });
                 break;
@@ -33216,6 +33282,7 @@ registerBuiltinPlugins();
   registerWorkflowRun,
   removeCollectionVectorStore,
   renderTemplate,
+  resolveJudgeModelKey,
   resolvePath,
   sandboxLatticeManager,
   sanitizeToolCallId,