@workbench-ai/workbench-core 0.0.66 → 0.0.68

package/dist/index.d.ts

@@ -1,220 +1,104 @@
-import type { AuthoredWorkbenchSourceDocument, CandidateCaseExecutionRef, CandidateCaseReview, CandidateFilePreview, CandidateFileSummary, CandidateRecord, CandidateSummary, EvalCaseResult, EngineResolveBinding, EvaluationRecord, EvaluationScorecard, RemoteWorkbenchEnvironment, RemoteWorkbenchEnvironmentVersion, RemoteWorkbenchFileInput, RemoteWorkbenchJob, Json, RunSummary, SurfaceSnapshotFile, UsageSummary, WorkbenchRuntimeBundle, WorkbenchRuntimeBundleStats, WorkbenchProjectSourceResources, WorkbenchProjectStateSource, WorkbenchCandidatePatch, WorkbenchExecutionCapability, WorkbenchExecutionSpec, WorkbenchResult } from "@workbench-ai/workbench-contract";
-import { type WorkbenchAdapterOperationExecutor, type WorkbenchAdapterOperationResult, type WorkbenchAdapterManifest } from "@workbench-ai/workbench-protocol";
-import { type GenericEngineCaseSpec, type GenericRunSpec, type WorkbenchCaseSelector, type WorkbenchEngineCase } from "./generic-spec.ts";
+import type { Json, RemoteWorkbenchJob, SurfaceSnapshotFile, WorkbenchArtifact, WorkbenchAutomationReadiness, WorkbenchSkillPatch, WorkbenchComparison, WorkbenchEvalSnapshot, WorkbenchExecutionSpec, WorkbenchInspectionSnapshot, WorkbenchJob, WorkbenchObjectPack, WorkbenchProjectState, WorkbenchRemote, WorkbenchRun, WorkbenchRunKind, WorkbenchAgent, WorkbenchSkillBundleSnapshot, WorkbenchSkillSource, WorkbenchStatus, WorkbenchTrace, WorkbenchVersion, UsageSummary } from "@workbench-ai/workbench-contract";
+import { type WorkbenchAdapterOperationExecutor } from "@workbench-ai/workbench-protocol";
+import type { WorkbenchAdapterAuthBundle, WorkbenchAdapterAuthTarget } from "./adapter-auth.ts";
+import { type SandboxExecutionFileStore, type SandboxPlane } from "./sandbox-plane.ts";
 import type { WorkbenchExecutionRuntimeInput } from "./execution-runtime-types.ts";
-import { createWorkbenchExecutionCapability, type SandboxExecutionFileStore, type SandboxPlane } from "./sandbox-plane.ts";
-import { type WorkbenchAdapterAuthBundle } from "./adapter-auth.ts";
-export { BENCHMARK_SPEC_FILE, CANDIDATE_SPEC_FILE, DEFAULT_EXECUTION_RESOURCES, engineCasePrivateFiles, engineCaseFilesForRuntimeInput, engineCasePublicFiles, engineResolveInvocationForSpec, engineResolveBindingForSpec, engineResolveBindingForSourceYaml, isWorkbenchCandidateManifestPath, parseWorkbenchSourceFiles, resolveEngineCaseExecutionConfig, resolveWorkbenchResolvedSourceYaml, resolveWorkbenchSourceFiles, runtimeNetwork, runtimeResources, serializeWorkbenchResolvedSourceYaml, validateWorkbenchResolvedSourceYaml, type AuthoredBenchmarkSpec, type WorkbenchCandidateImproveSpec, type WorkbenchCandidateManifestSpec, type WorkbenchCandidateRunSpec, type WorkbenchCaseSelector, type WorkbenchSelectionSpec, type GenericRunSpec, type GenericEngineCaseSpec, type WorkbenchEngineCase, type WorkbenchResolvedSource, } from "./generic-spec.ts";
-export { composeRuntimeDockerfileWithAdapterInstallers, type WorkbenchRuntimeAdapterInstaller, type WorkbenchRuntimeAdapterInstallerFile, } from "./runtime-dockerfile.ts";
-export { adapterCommandName, cloneWorkbenchAdapterManifest, collectWorkbenchAdapterAuthRequirements, collectWorkbenchAdapterInvocations, parseWorkbenchAdapterManifest, workbenchAdapterManifestRequiresAuth, workbenchAdapterManifestSupportsOperation, workbenchAdapterOperationCommand, workbenchAdapterOperationExecutor, withDefaultWorkbenchAdapterAuth, withDefaultWorkbenchAdapterAuthProfiles, type WorkbenchPrimitiveAdapterOperation, type WorkbenchAdapterOperation, type WorkbenchAdapterOperationExecutor, type WorkbenchAdapterOperationManifest, type WorkbenchAdapterSlotManifest, type WorkbenchAdapterAuthRequirement, type WorkbenchAdapterAuthManifest, type WorkbenchAdapterAuthMethodManifest, type WorkbenchAdapterInvocationLike, type WorkbenchAdapterManifest, } from "@workbench-ai/workbench-protocol";
-export { adapterAuthEnv, createWorkbenchAdapterAuthBundle, defaultWorkbenchAdapterAuthStoreRoot, localWorkbenchAdapterAuthStore, normalizeWorkbenchAdapterAuthTarget, parseWorkbenchAdapterAuthTarget, sanitizeWorkbenchAdapterAuthBundle, type WorkbenchAdapterAuthBundle, type WorkbenchAdapterAuthEnvVar, type WorkbenchAdapterAuthFile, type WorkbenchAdapterAuthStatus, type WorkbenchAdapterAuthStatusRecord, type WorkbenchAdapterAuthStore, type WorkbenchAdapterAuthTarget, } from "./adapter-auth.ts";
+export type { Json, RemoteWorkbenchJob, SurfaceSnapshotFile, WorkbenchArtifact, WorkbenchAutomationReadiness, WorkbenchComparison, WorkbenchComparisonCell, WorkbenchEvalSnapshot, WorkbenchExecutionResult, WorkbenchExecutionSpec, WorkbenchFileSurface, WorkbenchInspectionFileContent, WorkbenchInspectionFileOwnerKind, WorkbenchInspectionSnapshot, WorkbenchJob, WorkbenchLineageEdge, WorkbenchObjectPack, WorkbenchProjectState, WorkbenchRefs, WorkbenchRemote, WorkbenchRun, WorkbenchAgent, WorkbenchSkillBundleSnapshot, WorkbenchSkillInclude, WorkbenchSkillSource, WorkbenchStatus, WorkbenchTrace, WorkbenchVersion, } from "@workbench-ai/workbench-contract";
+export { workbenchInspectionFileContent, workbenchInspectionFileContentUnavailableReason, workbenchInspectionFileManifest, } from "@workbench-ai/workbench-contract";
 export type { WorkbenchExecutionRuntimeInput, WorkbenchWorkloadStepCommand, } from "./execution-runtime-types.ts";
-export { asRuntimeRecord, importNodeModule, jsonRecord, nodeBuiltin, normalizeRelativePath, normalizeWorkbenchWorkerId, normalizeRuntimeRegistry, quoteShellArg, readSurfaceFiles, resolveDockerRuntimeImageRef, resolveWorkbenchWorkerId, writeSurfaceFiles, } from "./runtime-utils.ts";
-export { assignUsageRole, extractExecutionUsageFromTrace, mergeUsageSummaries, } from "./execution-usage.ts";
-export { createWorkbenchProgressStdoutParser, publishWorkbenchProgressStdoutEnvelope, } from "./execution-events.ts";
-export { persistWorkbenchAdapterAuthUpdates, } from "./adapter-auth-updates.ts";
-export { resolveSandboxTemplateImage, } from "./sandbox-backends/template-images.ts";
-export { readOutputTraceFiles, workbenchTraceExecutionDirectory, workbenchTraceRunDirectory, workbenchTraceRunDirectoryName, } from "./trace-files.ts";
-export { assertWorkbenchAdapterOperationSupport, assertWorkbenchAdapterOperationResultOk, collectWorkbenchAdapterOperationIssues, collectWorkbenchAdapterOperationRequirements, ensureWorkbenchAdapterOutputDir, WORKBENCH_ADAPTER_RESULT_FILE, normalizeWorkbenchAdapterOperationRequest, normalizeWorkbenchAdapterOperationResult, readWorkbenchAdapterOperationRequest, readWorkbenchAdapterOperationResult, workbenchAdapterOperationResultPath, writeWorkbenchAdapterOperationResult, type WorkbenchAdapterOperationRequest, type WorkbenchAdapterOperationResult, type WorkbenchAdapterOperationResultValue, type WorkbenchAdapterOperationRequirement, type WorkbenchEngineResolveResult, type WorkbenchEngineCaseSpec, } from "@workbench-ai/workbench-protocol";
-export { applyWorkbenchCandidatePatch, type ApplyWorkbenchCandidatePatchInput, } from "./candidate-patch.ts";
-export { createWorkbenchSandboxFileStore, createSandboxAdapterRequest, executionResultFromCompletedSandboxJob, isSurfaceSnapshotFile, materializeWorkbenchSandboxInput, readWorkbenchExecutionSpec, sanitizeWorkbenchExecutionJobForSandbox, } from "./sandbox-inputs.ts";
-export { compileWorkbenchExecutionGraph, type CompileExecutionGraphInput, type WorkbenchExecutionGraph, type WorkbenchExecutionGraphNode, } from "./execution-graph.ts";
-export { createBaselineCandidateExecution, createBaselineCandidateJob, createWorkbenchExecutionJob, expectedWorkbenchRunJobCount, engineCaseForCase, engineCaseIds, attemptJobCountForRunSpec, workbenchExecutionJobPurpose, MAX_WORKBENCH_RUN_BUDGET, planWorkbenchExecutionJobsForPurpose, validateWorkbenchRunEnvelope, workbenchExecutionJobId, type WorkbenchRunWorkflow, } from "./execution-jobs.ts";
+export { attachSandboxMetadataToJob, createSandboxAdapterRequest, createWorkbenchSandboxFileStore, executionResultFromCompletedSandboxJob, isSurfaceSnapshotFile, readWorkbenchExecutionSpec, } from "./sandbox-inputs.ts";
+export { createWorkbenchExecutionCapability, createWorkbenchSandboxAllocation, collectExecutionCapabilityScopeIssues, collectSandboxAllocationScopeIssues, collectSandboxHandleScopeIssues, assertSandboxBackendSupportsNetworkPolicy, executeValidatedSandboxExecution, type SandboxBackendCapabilities, type SandboxBackendDescriptor, type SandboxCreateRequest, type SandboxEnvironmentImage, type SandboxExecRequest, type SandboxExecutionFileStore, type SandboxHandle, type SandboxMaterializedInput, type SandboxPlane, } from "./sandbox-plane.ts";
+export { createDockerSandboxBackendDescriptor, createDockerSandboxPlane, DOCKER_SANDBOX_BACKEND, } from "./sandbox-backends/index.ts";
 export { addCapacity, capacityFits, runWorkbenchExecutionDag, subtractCapacity, workbenchJobDependencies, workbenchJobHostCost, workbenchJobResources, type WorkbenchExecutionDagCapacity, type WorkbenchExecutionDagResult, type WorkbenchExecutionDagRunInput, } from "./execution-scheduler.ts";
-export { assertWorkbenchExecutionIsolation, collectWorkbenchExecutionIsolationIssues, validateWorkbenchExecutionOutputPayloads, type WorkbenchExecutionOutputPayloads, } from "./execution-outputs.ts";
-export { collectSandboxAllocationScopeIssues, collectExecutionCapabilityScopeIssues, collectSandboxHandleScopeIssues, createWorkbenchSandboxAllocation, createWorkbenchSandboxExecutionMetadata, createWorkbenchExecutionCapability, executeValidatedSandboxExecution, type SandboxExecutionFileStore, type SandboxExecutionOptions, type SandboxBackendCapabilities, type SandboxBackendDescriptor, type SandboxCreateRequest, type SandboxEnvironmentImage, type SandboxExecRequest, type SandboxHandle, type SandboxMaterializedInput, type SandboxPlane, type ValidatedSandboxExecutionResult, } from "./sandbox-plane.ts";
-export { buildCandidateCaseExecutionRefs, buildWorkbenchExecutionEvidence, isWorkbenchExecutionActive, readWorkbenchExecutionId, readWorkbenchExecutionMetadataNumber, readWorkbenchExecutionMetadataString, readWorkbenchExecutionPurpose, resolveWorkbenchJobGroupStatus, } from "./execution-evidence.ts";
-export { buildCandidateLineage, buildWorkbenchEvaluationComparison, buildWorkbenchEvaluationMetricDescriptors, formatEvaluationConfigurationLabel, isCompleteEvaluationSummary, readEvaluationScore, type WorkbenchCandidateEvaluationRollup, type WorkbenchEvaluationComparison, type WorkbenchEvaluationComparisonRow, type WorkbenchEvaluationMetricDescriptor, } from "@workbench-ai/workbench-contract";
-export { WorkbenchInspectionError, createWorkbenchInspection, type WorkbenchFailureDetail, type WorkbenchFailureDiagnosis, type WorkbenchFailureKind, type WorkbenchInspection, type WorkbenchInspectionBackend, type WorkbenchInspectionCandidateInput, type WorkbenchInspectionCandidatePreviewInput, type WorkbenchInspectionCaseReviewInput, type WorkbenchInspectionErrorOptions, type WorkbenchInspectionEvaluationInput, type WorkbenchInspectionExecutionInput, type WorkbenchInspectionExecutionPreviewInput, type WorkbenchInspectionFileListInput, type WorkbenchInspectionFilePreviewInput, type WorkbenchInspectionPreviewInput, type WorkbenchInspectionRunDetail, type WorkbenchInspectionRunInput, } from "./inspection.ts";
-export { buildWorkbenchTraceSessionsFromFiles, combineWorkbenchTraceSessions, finalizeWorkbenchExecutionTraceForJob, mergeWorkbenchExecutionTracesByJob, readWorkbenchExecutionTraceFiles, traceSessionLabel, type WorkbenchTraceMergeJob, } from "./execution-traces.ts";
-export { DOCKER_SANDBOX_BACKEND, assertSandboxHostHealthForBackend, createDockerSandboxBackendDescriptor, createDockerSandboxPlane, resolveWorkbenchSandboxBackendName, sandboxBackendAdmissionForResources, sandboxHostHealthExpectationForBackend, type SandboxBackendAdmission, type SandboxBackendHostCost, type SandboxBackendRequestedResources, type SandboxHostHealthExpectation, type WorkbenchSandboxBackendName, } from "./sandbox-backends/index.ts";
-export type { WorkbenchExecutionEventPublisher, WorkbenchExecutionProgressTarget, } from "./execution-events.ts";
-export type { CandidateCaseReview, CandidateRecord, EngineResolveBinding, EvaluationScorecard, RemoteWorkbenchJob, Json, RunSummary, RuntimeEvent, SurfaceSnapshotFile, WorkbenchRuntimeBundle, WorkbenchRuntimeBundleStats, WorkbenchRuntimeRun, WorkbenchRuntimeCandidateFiles, WorkbenchRuntimeExecutionFiles, WorkbenchRuntimeImportResult, WorkbenchProjectSourceResources, WorkbenchProjectState, WorkbenchProjectStateBase, WorkbenchProjectStateImportResult, WorkbenchProjectStateRemote, WorkbenchProjectStateSource, WorkbenchRemoteRunRequest, WorkbenchExecutionCapability, WorkbenchExecutionTrace, WorkbenchTraceSession, WorkbenchSandboxHandle, WorkbenchSandboxExecutionMetadata, } from "@workbench-ai/workbench-contract";
-export declare function sanitizeWorkbenchRuntimeJobForExchange(job: RemoteWorkbenchJob): RemoteWorkbenchJob;
-export declare function compactWorkbenchRuntimeJobForExchange(job: RemoteWorkbenchJob): RemoteWorkbenchJob;
-export declare function sanitizeWorkbenchRuntimeCandidateForExchange(candidate: CandidateRecord): CandidateRecord;
-export declare function workbenchRuntimeCandidateIdentityForExchange(candidate: CandidateRecord): {
-    id: string;
-    candidateFingerprint: string;
-    baseId: string | null;
-    referenceIds: string[];
-};
-export declare function mergeWorkbenchRuntimeCandidateForExchange(left: CandidateRecord, right: CandidateRecord): CandidateRecord;
-export interface WorkbenchBenchmarkContentFingerprintInput {
-    sourceYaml: string;
-    engineResolveFiles: readonly SurfaceSnapshotFile[];
-    engineResolveBinding: EngineResolveBinding;
-    adapterFiles?: readonly SurfaceSnapshotFile[];
-    adapterManifests?: readonly WorkbenchAdapterManifest[];
-    runtimeFiles?: readonly SurfaceSnapshotFile[];
-    resources?: WorkbenchProjectSourceResources | null;
-    network?: WorkbenchProjectStateSource["network"] | null;
-}
-export interface WorkbenchCandidateContentFingerprintInput {
-    sourceYaml: string;
-    candidateFiles: readonly SurfaceSnapshotFile[];
-    adapterFiles?: readonly SurfaceSnapshotFile[];
-    adapterManifests?: readonly WorkbenchAdapterManifest[];
-}
-export declare function workbenchBenchmarkContentFingerprint(input: WorkbenchBenchmarkContentFingerprintInput): string;
-export declare function workbenchCandidateContentFingerprint(input: WorkbenchCandidateContentFingerprintInput): string;
-export declare function workbenchProjectSourceFingerprint(input: Omit<WorkbenchProjectStateSource, "files" | "revisionId" | "fingerprint">): string;
-export declare function workbenchRuntimeBundleFingerprint(bundle: WorkbenchRuntimeBundle): string;
-export declare function workbenchSurfaceFilesEqualForExchange(left: readonly SurfaceSnapshotFile[], right: readonly SurfaceSnapshotFile[]): boolean;
-export declare function workbenchRuntimeBundleStats(bundle: WorkbenchRuntimeBundle): WorkbenchRuntimeBundleStats;
-export declare function workbenchRuntimeExplicitActiveId(args: {
-    candidates: readonly CandidateRecord[];
-    runs: readonly RunSummary[];
-    preferredActiveId?: string | null;
-    benchmarkFingerprint: string;
-}): string | null;
-export interface WorkbenchRunMaterialization {
-    candidates: CandidateRecord[];
-    candidateFiles: Record<string, SurfaceSnapshotFile[]>;
-    evaluations: EvaluationScorecard[];
-    activeCandidateId: string | null;
-    selectedCandidate: CandidateRecord | null;
-    completedJobCount: number;
-    failedJobCount: number;
-}
-export interface WorkbenchRunWorkload {
-    job: RemoteWorkbenchJob;
-    spec: GenericRunSpec;
-    candidateId: string;
-    attemptIndex: number;
-    sampleIndex: number;
+export { createWorkbenchExecutionJob, planWorkbenchExecutionJobsForPurpose, workbenchExecutionJobId, } from "./execution-jobs.ts";
+export { applyWorkbenchSkillPatch, } from "./skill-patch.ts";
+export { asRuntimeRecord, importNodeModule, jsonRecord, nodeBuiltin, normalizeRelativePath, normalizeRuntimeRegistry, normalizeWorkbenchWorkerId, quoteShellArg, readSurfaceFiles, resolveDockerRuntimeImageRef, resolveWorkbenchWorkerId, writeSurfaceFiles, } from "./runtime-utils.ts";
+export { assignUsageRole, extractExecutionUsageFromTrace, mergeUsageSummaries, normalizeUsageSummary, } from "./execution-usage.ts";
+export { composeRuntimeDockerfileWithAdapterInstallers, type WorkbenchRuntimeAdapterInstaller, } from "./runtime-dockerfile.ts";
+export { readOutputTraceFiles, workbenchTraceExecutionDirectory, workbenchTraceRunDirectory, workbenchTraceRunDirectoryName, } from "./trace-files.ts";
+export { createWorkbenchProgressStdoutParser, publishWorkbenchProgressStdoutEnvelope, type WorkbenchExecutionEventPublisher, type WorkbenchExecutionProgressTarget, } from "./execution-events.ts";
+export { persistWorkbenchAdapterAuthUpdates, } from "./adapter-auth-updates.ts";
+export { adapterAuthEnv, createWorkbenchAdapterAuthBundle, defaultWorkbenchAdapterAuthStoreRoot, localWorkbenchAdapterAuthStore, normalizeWorkbenchAdapterAuthTarget, parseWorkbenchAdapterAuthTarget, sanitizeWorkbenchAdapterAuthBundle, type WorkbenchAdapterAuthBundle, type WorkbenchAdapterAuthEnvVar, type WorkbenchAdapterAuthFile, type WorkbenchAdapterAuthStatus, type WorkbenchAdapterAuthStatusRecord, type WorkbenchAdapterAuthStore, type WorkbenchAdapterAuthTarget, } from "./adapter-auth.ts";
+export { mergeWorkbenchExecutionTracesByJob, } from "./execution-traces.ts";
+export declare class WorkbenchUserError extends Error {
+    constructor(message: string);
+}
+export interface WorkbenchCommandOptions {
+    dir?: string;
+    authToken?: string;
+}
+export interface WorkbenchEvalOptions extends WorkbenchCommandOptions {
+    version?: string;
+    skill?: string;
+    agent?: string;
+    samples?: number;
+    kind?: WorkbenchRunKind;
+    parentRunId?: string;
+    caseIds?: readonly string[];
+    selectedSamples?: readonly WorkbenchCaseSampleSelection[];
+    rerun?: boolean;
+}
+export interface WorkbenchStateEvalOptions {
+    authToken?: string;
+    version?: string;
+    evalHash?: string;
+    skill?: string;
+    agent?: string;
+    samples?: number;
+    kind?: WorkbenchRunKind;
+    parentRunId?: string;
+    caseIds?: readonly string[];
+    selectedSamples?: readonly WorkbenchCaseSampleSelection[];
+    rerun?: boolean;
+}
+export interface WorkbenchStateImproveOptions {
+    authToken?: string;
+    version?: string;
+    evalHash?: string;
+    skill?: string;
+    agent?: string;
+    budget?: number;
+    samples?: number;
+    parentRunId?: string;
+    evidenceTraceIds?: readonly string[];
+}
+export interface WorkbenchCaseSampleSelection {
     caseId: string;
-    candidateFiles: SurfaceSnapshotFile[];
-    engineResolveFiles: SurfaceSnapshotFile[];
-    traceFiles: SurfaceSnapshotFile[];
-    engineCase?: WorkbenchEngineCase;
-    engineCaseSpec?: GenericEngineCaseSpec;
-    prompt: string;
-    changedPaths: string[];
-    baseId: string | null;
-}
-export interface RuntimeWorkloadResult {
-    files: SurfaceSnapshotFile[];
-    fileChanges: string[];
-    operationResults?: WorkbenchAdapterOperationResult[];
-    workspaceFiles?: SurfaceSnapshotFile[];
-    candidatePatch?: WorkbenchCandidatePatch;
-    result?: WorkbenchResult;
-    metrics?: Record<string, number>;
-    cases?: EvalCaseResult[];
-    usage?: UsageSummary;
-    summary?: string;
-    feedback?: Json;
-    exitCode?: number;
-    error?: string;
-    startedAt?: string;
-    finishedAt?: string;
-    durationMs?: number;
-}
-export declare const DEFAULT_ENVIRONMENT_VERSIONS: RemoteWorkbenchEnvironmentVersion[];
-export declare const DEFAULT_ENVIRONMENTS: RemoteWorkbenchEnvironment[];
-export declare function loadAuthoredWorkbenchSourceDocument(args: {
-    sourceYaml: string;
-    path?: string;
-    sourceFiles?: readonly SurfaceSnapshotFile[];
-    cases?: RemoteWorkbenchFileInput[];
-}): AuthoredWorkbenchSourceDocument;
-export declare function materializeWorkbenchRunResult(args: {
-    runId: string;
-    benchmarkFingerprint: string;
-    sourceYaml?: string;
-    benchmarkSourceFiles?: readonly SurfaceSnapshotFile[];
-    candidateFingerprint?: string;
-    candidateSourceFiles?: readonly SurfaceSnapshotFile[];
-    startedAt: string;
-    spec: GenericRunSpec;
-    jobs: readonly RemoteWorkbenchJob[];
-    previousCandidate?: CandidateRecord | null;
-    existingCandidateCount: number;
-    selection?: {
-        metric: string;
-        caseIds?: readonly string[];
-        label?: string;
-    };
-}): WorkbenchRunMaterialization;
-export declare function evaluationScorecardId(runId: string, candidateId: string): string;
-export declare function selectExecutionOutputFilesForInspection(args: {
-    purpose: string | null | undefined;
-    files: readonly SurfaceSnapshotFile[];
-    output?: Record<string, unknown> | null | undefined;
-}): SurfaceSnapshotFile[];
-export declare function isWorkbenchInternalOutputPath(filePath: string): boolean;
-export declare function isGeneratedExecutionOutputPath(filePath: string): boolean;
-export declare function createOptimizerTraceInputFiles(args: {
-    jobs: readonly RemoteWorkbenchJob[];
-}): SurfaceSnapshotFile[];
-export interface WorkbenchSelectionPolicy {
-    metric: string;
-    selector: WorkbenchCaseSelector;
-}
-export declare function workbenchImproveOptimizeSelector(spec: GenericRunSpec): WorkbenchCaseSelector;
-export declare function workbenchImproveSelectionPolicy(spec: GenericRunSpec): WorkbenchSelectionPolicy;
-export declare function workbenchEngineCaseIdsForSelector(engineCases: readonly WorkbenchEngineCase[], selector: WorkbenchCaseSelector): string[];
-export declare function workbenchEngineCaseIdsForImproveEvaluation(args: {
-    spec: GenericRunSpec;
-    engineCases: readonly WorkbenchEngineCase[];
-}): string[];
-export declare function filterOptimizerTraceJobsForCaseIds(jobs: readonly RemoteWorkbenchJob[], caseIds: readonly string[]): RemoteWorkbenchJob[];
-export declare function formatWorkbenchCaseSelector(selector: WorkbenchCaseSelector): string;
-export declare function formatWorkbenchSelectionPolicy(policy: WorkbenchSelectionPolicy): string;
-export declare function workbenchCaseSelectorUsesAllCases(selector: WorkbenchCaseSelector): boolean;
-export declare function evaluationMeanMetrics(evaluation: Pick<EvaluationRecord, "metrics"> | null | undefined): Record<string, number> | undefined;
-export declare function candidateRecordWithoutDerivedFields(candidate: CandidateRecord): CandidateRecord;
-export declare function candidateSummaryFromRecord(candidate: CandidateRecord): CandidateSummary;
-export declare function workbenchRunExecutionFingerprint(args: {
-    sourceYaml?: string | null;
-    adapterFiles?: readonly SurfaceSnapshotFile[];
-    specVersionId?: string | null;
-    environmentVersionId?: string | null;
-}): string;
-export interface WorkbenchProjectSourceFilesInput {
-    specSource?: string;
-    specFiles?: readonly SurfaceSnapshotFile[];
-    candidateFilesPath: string;
-    candidateFiles: readonly SurfaceSnapshotFile[];
-    engineResolveFilesPath: string;
-    engineResolveFiles: readonly SurfaceSnapshotFile[];
-    adapterFiles?: readonly SurfaceSnapshotFile[];
-    dockerfilePath?: string;
-    dockerfile?: string | null;
-    dockerfiles?: readonly SurfaceSnapshotFile[];
-}
-export declare function buildWorkbenchProjectSourceFiles(input: WorkbenchProjectSourceFilesInput): SurfaceSnapshotFile[];
-export declare function readWorkbenchSpecDockerfilePath(spec: {
-    environment: {
-        dockerfile?: unknown;
+    sample: number;
+}
+export interface WorkbenchCheckResult {
+    ok: true;
+    status: WorkbenchStatus;
+    cases: number;
+    skills: number;
+    agents: number;
+    plan: {
+        source: {
+            skillFiles: number;
+            evalFiles: number;
+            caseCount: number;
+            smokeCaseCount: number;
+        };
+        skills: Array<{
+            name: string;
+            bundleHash: string;
+            includedSkillCount: number;
+            fileCount: number;
+        }>;
+        agents: Array<{
+            name: string;
+            adapter: string;
+            model?: string;
+            providerBacked: boolean;
+            executionMode: "local-command" | "provider-backed";
+            network: WorkbenchExecutionSpec["policy"]["network"];
+            resources: WorkbenchExecutionSpec["policy"]["resources"];
+            image: string;
+            auth?: "local-adapter-auth";
+        }>;
+        readiness: WorkbenchAutomationReadiness;
     };
-}): string;
-export declare function isCandidateSourceFilePath(filePath: string): boolean;
-export declare function filterCandidateSourceFiles(files: readonly SurfaceSnapshotFile[]): SurfaceSnapshotFile[];
-export declare function normalizeSurfaceFiles(files: RemoteWorkbenchFileInput[]): SurfaceSnapshotFile[];
-export declare function filterSurfaceFilesByInclude<T extends {
-    path: string;
-}>(files: readonly T[], include: readonly string[] | undefined): T[];
-export declare function summarizeCandidateFiles(files: readonly SurfaceSnapshotFile[], changedPaths?: readonly string[]): CandidateFileSummary[];
-export declare function createCandidateFilePreview(args: {
-    files: readonly SurfaceSnapshotFile[];
-    path: string;
-    view: "diff" | "raw" | "rendered";
-}): CandidateFilePreview;
-export declare function createCaseReview(args: {
-    candidate: CandidateRecord;
-    caseId: string;
-    executions?: CandidateCaseExecutionRef[];
-}): CandidateCaseReview;
-export declare function createWorkbenchRunWorkload(args: {
-    job: RemoteWorkbenchJob;
-    spec: GenericRunSpec;
-    baseFiles: readonly SurfaceSnapshotFile[];
-    engineResolveFiles: readonly SurfaceSnapshotFile[];
-    engineCases: readonly WorkbenchEngineCase[];
-    traceFiles?: readonly SurfaceSnapshotFile[];
-}): WorkbenchRunWorkload;
+}
 export interface WorkbenchExecutionJobOptions {
     sandboxBackend: string;
     loadLocalAdapterAuthProfiles?: boolean;
@@ -222,13 +106,234 @@ export interface WorkbenchExecutionJobOptions {
     createSandboxPlaneForBackend?: (backend: string, args: WorkbenchExecutionRuntimeInput, startedAt: string, fileStore: SandboxExecutionFileStore) => SandboxPlane;
 }
 export declare function executeWorkbenchExecutionJob(args: WorkbenchExecutionRuntimeInput, options: WorkbenchExecutionJobOptions): Promise<RemoteWorkbenchJob>;
-export declare function workbenchExecutionExecutorForRuntimeInput(args: Pick<WorkbenchExecutionRuntimeInput, "job" | "adapterManifests" | "runtimeControlOperation">): WorkbenchAdapterOperationExecutor;
+export declare function requiredWorkbenchAdapterAuthTargetsForRuntimeInput(args: Pick<WorkbenchExecutionRuntimeInput, "job" | "adapterManifests" | "runtimeControlOperation" | "spec">): WorkbenchAdapterAuthTarget[];
 export declare function workbenchExecutionPurpose(job: RemoteWorkbenchJob): WorkbenchExecutionSpec["purpose"] | null;
-export declare function executeAdapterInCurrentRuntime(args: WorkbenchExecutionRuntimeInput, execution: WorkbenchExecutionSpec, startedAt: string, capability: ReturnType<typeof createWorkbenchExecutionCapability>): Promise<RemoteWorkbenchJob>;
-export declare function executeRuntimeControlOperationSequenceInCurrentRuntime(args: WorkbenchExecutionRuntimeInput, execution: WorkbenchExecutionSpec, startedAt: string, capability?: WorkbenchExecutionCapability): Promise<RemoteWorkbenchJob>;
-export declare function stageWorkbenchRunWorkload(root: string, workload: WorkbenchRunWorkload): Promise<void>;
-export declare function workloadTimeoutMs(spec: GenericRunSpec): number;
-export declare function findEnvironmentVersionForImage(image: string, versions: readonly RemoteWorkbenchEnvironmentVersion[]): RemoteWorkbenchEnvironmentVersion | null;
-export declare function normalizeDockerImageRef(image: string): string;
-export declare function environmentVersionTimeoutMs(version: Pick<RemoteWorkbenchEnvironmentVersion, "spec"> | null | undefined): number;
+export declare function workbenchExecutionExecutorForRuntimeInput(args: Pick<WorkbenchExecutionRuntimeInput, "job" | "adapterManifests" | "runtimeControlOperation">): WorkbenchAdapterOperationExecutor;
+export declare function executeAdapterInCurrentRuntime(args: WorkbenchExecutionRuntimeInput, execution: WorkbenchExecutionSpec, startedAt: string): Promise<RemoteWorkbenchJob>;
+export declare function executeRuntimeControlOperationSequenceInCurrentRuntime(args: WorkbenchExecutionRuntimeInput, execution: WorkbenchExecutionSpec, startedAt: string): Promise<RemoteWorkbenchJob>;
+export declare function isWorkbenchInternalOutputPath(filePath: string): boolean;
+export interface WorkbenchImproveOptions extends WorkbenchCommandOptions {
+    version?: string;
+    skill?: string;
+    agent?: string;
+    budget?: number;
+    samples?: number;
+    parentRunId?: string;
+    evidenceTraceIds?: readonly string[];
+}
+export interface WorkbenchImproveResult {
+    run: WorkbenchRun;
+    version: WorkbenchVersion;
+    switched: boolean;
+    promoted: boolean;
+    promotionReason: string;
+    incumbentRunId?: string;
+    incumbentScore?: number;
+    outputScore?: number;
+}
+export interface WorkbenchCompareOptions extends WorkbenchCommandOptions {
+    versions?: string;
+    skills?: string;
+    agents?: string;
+}
+export interface WorkbenchRemoteOptions extends WorkbenchCommandOptions {
+    remote?: string;
+    authToken?: string;
+}
+export interface WorkbenchPublishOptions extends WorkbenchCommandOptions {
+    version?: string;
+    remote?: string;
+    visibility?: "private" | "public";
+    authToken?: string;
+}
+export interface WorkbenchPublishResult {
+    remote: WorkbenchRemote;
+    version: WorkbenchVersion;
+    visibility: "private" | "public";
+    installUrl: string;
+    pinnedInstallUrl: string;
+}
+export interface WorkbenchDiffEntry {
+    path: string;
+    status: "added" | "removed" | "modified";
+    before?: string;
+    after?: string;
+}
+export interface WorkbenchCaseRecord {
+    id: string;
+    path: string;
+    content: string;
+}
+export interface WorkbenchQueuedSkillEvalJobInput {
+    kind: "workbench.skill.eval.job.v1";
+    ownerUserId?: string;
+    skillId?: string;
+    runId: string;
+    jobId: string;
+    artifactId?: string;
+    traceId?: string;
+    versionId: string;
+    evalHash: string;
+    agentName: string;
+    caseId: string;
+    sample: number;
+    state: WorkbenchProjectState;
+    adapterAuthProfiles?: readonly WorkbenchAdapterAuthBundle[];
+}
+export interface WorkbenchEvalCaseRuntime extends WorkbenchCaseRecord {
+    files: SurfaceSnapshotFile[];
+    command?: string;
+    smoke?: boolean;
+}
+export interface WorkbenchSkillEvalRuntimeInputArgs {
+    ownerUserId: string;
+    projectId: string;
+    runId: string;
+    jobId: string;
+    versionId: string;
+    evalHash: string;
+    evalSnapshot: WorkbenchEvalSnapshot;
+    skillName?: string;
+    skillBundleHash?: string;
+    agent: WorkbenchAgent;
+    versionFiles: readonly SurfaceSnapshotFile[];
+    runtimeCase: WorkbenchEvalCaseRuntime;
+    sample: number;
+    createdAt?: string;
+    attempt?: number;
+    environmentDockerfile?: string;
+    environmentImageRef?: string;
+}
+export interface WorkbenchSkillImproveRuntimeInputArgs {
+    ownerUserId: string;
+    projectId: string;
+    runId: string;
+    jobId: string;
+    baseVersionId: string;
+    evalHash: string;
+    skillName?: string;
+    skillBundleHash?: string;
+    agent: WorkbenchAgent;
+    baseFiles: readonly SurfaceSnapshotFile[];
+    traces: readonly WorkbenchTrace[];
+    createdAt?: string;
+    attempt?: number;
+    environmentDockerfile?: string;
+    environmentImageRef?: string;
+}
+export interface WorkbenchSkillImprovementPatchApplication {
+    state: WorkbenchProjectState;
+    version: WorkbenchVersion;
+    created: boolean;
+}
+export declare function initWorkbenchSkill(options?: WorkbenchCommandOptions): Promise<WorkbenchStatus>;
+export declare function workbenchStatus(options?: WorkbenchCommandOptions): Promise<WorkbenchStatus>;
+export declare function checkWorkbenchSkill(options?: WorkbenchCommandOptions): Promise<WorkbenchCheckResult>;
+export declare function listWorkbenchVersions(options?: WorkbenchCommandOptions): Promise<WorkbenchVersion[]>;
+export declare function evalWorkbenchSkill(options?: WorkbenchEvalOptions): Promise<WorkbenchRun[]>;
+export declare function evalWorkbenchProjectState(state: WorkbenchProjectState, options?: WorkbenchStateEvalOptions): Promise<{
+    state: WorkbenchProjectState;
+    runs: WorkbenchRun[];
+}>;
+export declare function improveWorkbenchProjectState(state: WorkbenchProjectState, options?: WorkbenchStateImproveOptions): Promise<{
+    state: WorkbenchProjectState;
+    runs: WorkbenchRun[];
+}>;
+export declare function listWorkbenchProjectStateEvalCases(state: WorkbenchProjectState): Promise<WorkbenchCaseRecord[]>;
+export declare function listWorkbenchProjectStateEvalRuntimeCases(state: WorkbenchProjectState, evalHash?: string): Promise<WorkbenchEvalCaseRuntime[]>;
+export declare function createWorkbenchSkillEvalRuntimeInput(args: WorkbenchSkillEvalRuntimeInputArgs): WorkbenchExecutionRuntimeInput;
+export declare function createWorkbenchSkillImproveRuntimeInput(args: WorkbenchSkillImproveRuntimeInputArgs): WorkbenchExecutionRuntimeInput;
+export declare function workbenchImprovementEvidenceFromTraces(traces: readonly WorkbenchTrace[]): string[];
+export declare function workbenchImprovementEvidenceTraces(traces: readonly WorkbenchTrace[]): WorkbenchTrace[];
+export declare function workbenchImprovementEvidenceTracesForVersion(state: WorkbenchProjectState, options: {
+    versionId: string;
+    skillName: string;
+    agent: WorkbenchAgent;
+    traceIds?: readonly string[];
+}): WorkbenchTrace[];
+export declare function workbenchSkillImproveCanUseQueuedAdapter(agent: WorkbenchAgent): boolean;
+export declare function workbenchSkillImproveAdapterRequirementMessage(agent: WorkbenchAgent): string;
+export declare function readWorkbenchSkillImprovementPatchFromRemoteJob(remoteJob: RemoteWorkbenchJob): WorkbenchSkillPatch | null;
+export declare function applyWorkbenchSkillImprovementPatch(state: WorkbenchProjectState, args: {
+    baseVersionId: string;
+    agent: WorkbenchAgent;
+    patch: WorkbenchSkillPatch;
+    runId?: string;
+    createdAt?: string;
+}): WorkbenchSkillImprovementPatchApplication;
+export declare function decideWorkbenchImprovementPromotion(run: WorkbenchRun, incumbentRun: WorkbenchRun | undefined): {
+    promoted: boolean;
+    reason: string;
+};
+export declare function normalizeWorkbenchSkillEvalEnvironmentDockerfile(source: string | undefined): string | undefined;
+export interface WorkbenchVersionRuntimeSnapshot {
+    evalSnapshot: WorkbenchEvalSnapshot;
+    cases: WorkbenchEvalCaseRuntime[];
+    agents: WorkbenchAgent[];
+    selectedAgents: WorkbenchAgent[];
+    defaultAgent?: string;
+    skillSources: WorkbenchSkillSource[];
+    skillBundles: WorkbenchSkillBundleSnapshot[];
+    defaultSkill?: string;
+    environmentDockerfile?: string;
+}
+export declare function createWorkbenchVersionRuntimeSnapshot(version: WorkbenchVersion, options?: {
+    skill?: string;
+    agent?: string;
+    evalHash?: string;
+    authToken?: string;
+}): Promise<WorkbenchVersionRuntimeSnapshot>;
+export declare function executeQueuedWorkbenchSkillEvalJob(input: WorkbenchQueuedSkillEvalJobInput): Promise<{
+    run: WorkbenchRun;
+    job: WorkbenchJob;
+    artifact: WorkbenchArtifact;
+    trace: WorkbenchTrace;
+}>;
+export declare function improveWorkbenchSkill(options?: WorkbenchImproveOptions): Promise<WorkbenchImproveResult>;
+export declare function compareWorkbench(options?: WorkbenchCompareOptions): Promise<WorkbenchComparison>;
+export declare function switchWorkbenchVersion(versionRef: string, options?: WorkbenchCommandOptions): Promise<WorkbenchVersion>;
+export declare function diffWorkbenchVersions(range: string, options?: WorkbenchCommandOptions): Promise<WorkbenchDiffEntry[]>;
+export declare function showWorkbenchRef(ref: string, options?: WorkbenchCommandOptions): Promise<unknown>;
+export declare function filesForWorkbenchRef(ref: string, options?: WorkbenchCommandOptions): Promise<SurfaceSnapshotFile[]>;
+export declare function listWorkbenchCases(options?: WorkbenchCommandOptions): Promise<WorkbenchCaseRecord[]>;
+export declare function showWorkbenchCase(caseId: string, options?: WorkbenchCommandOptions): Promise<WorkbenchCaseRecord>;
+export declare function addWorkbenchCase(options?: WorkbenchCommandOptions & {
+    fromTraceId?: string;
+}): Promise<WorkbenchCaseRecord>;
+export declare function removeWorkbenchCase(caseId: string, options?: WorkbenchCommandOptions): Promise<{
+    removed: string;
+}>;
+export declare function listWorkbenchAgents(options?: WorkbenchCommandOptions): Promise<WorkbenchAgent[]>;
+export declare function addWorkbenchAgent(input: WorkbenchCommandOptions & {
+    name: string;
+    adapter: string;
+    model?: string;
+    config?: Record<string, Json>;
+}): Promise<WorkbenchAgent>;
+export declare function removeWorkbenchAgent(name: string, options?: WorkbenchCommandOptions): Promise<{
+    removed: string;
+}>;
+export declare function setDefaultWorkbenchAgent(name: string, options?: WorkbenchCommandOptions): Promise<WorkbenchAgent>;
+export declare function addWorkbenchRemote(name: string, url: string, options?: WorkbenchCommandOptions): Promise<WorkbenchRemote>;
+export declare function listWorkbenchRemotes(options?: WorkbenchCommandOptions): Promise<WorkbenchRemote[]>;
+export declare function syncWorkbenchRemote(options?: WorkbenchRemoteOptions): Promise<{
+    remote: WorkbenchRemote;
+    pushed: number;
+    pulled: number;
+}>;
+export declare function publishWorkbenchVersion(options?: WorkbenchPublishOptions): Promise<WorkbenchPublishResult>;
+export declare function createWorkbenchInspectionSnapshot(options?: WorkbenchCommandOptions): Promise<WorkbenchInspectionSnapshot>;
+export declare function exportObjectPack(state: WorkbenchProjectState): WorkbenchObjectPack;
+interface ImportObjectPackOptions {
+    refs?: "merge" | "none";
+}
+export declare function importObjectPack(state: WorkbenchProjectState, pack: WorkbenchObjectPack, options?: ImportObjectPackOptions): void;
+export declare function hashJson(value: unknown): string;
+export declare function hashFiles(files: readonly SurfaceSnapshotFile[]): string;
+export declare function readWorkbenchSkillRunOutputScore(output: unknown): number;
+export declare function readWorkbenchSkillRunOutputUsage(output: unknown): UsageSummary | undefined;
+export declare function readWorkbenchSkillTraceResultsCostUsd(results: readonly unknown[]): number | undefined;
+export declare function readWorkbenchSkillUsageCostUsd(usage: UsageSummary | undefined): number | undefined;
+export declare function isDefaultWorkbenchSkillEvalEnvironmentDockerfile(source: string): boolean;
+export declare function automationReadinessForRuns(runs: readonly WorkbenchRun[], jobs: readonly WorkbenchJob[]): WorkbenchAutomationReadiness;
 //# sourceMappingURL=index.d.ts.map