@workbench-ai/workbench-contract 0.0.66 → 0.0.68

package/dist/index.d.ts

@@ -1,683 +1,317 @@
 export type Json = null | boolean | number | string | Json[] | {
     [key: string]: Json;
 };
-export declare function isReservedWorkbenchAdapterAuthEnvName(name: string): boolean;
-export declare function assertWorkbenchAdapterAuthEnvNameAllowed(name: string): void;
-export interface RemoteWorkbenchProject {
-    id: string;
-    ownerUserId: string;
-    ownerUsername: string;
-    visibility: "private" | "public";
-    createdAt: string;
-    updatedAt: string;
-    activeEnvironmentVersionId: string;
-    currentSpecVersionId: string;
-    activeCandidateId?: string | null;
-    sourceFingerprint?: string;
-    starCount: number;
-}
-export interface RemoteWorkbenchProjectSummary {
-    id: string;
-    ownerUsername: string;
-    name: string;
-    description: string;
-    visibility: "private" | "public";
-    updatedAt: string;
-    currentSpecVersionId: string;
-    activeEnvironmentVersionId: string;
-    activeCandidateId?: string | null;
-    candidateCount: number;
-    evaluationCount: number;
-    runCount: number;
-    starCount: number;
-    viewerHasStarred?: boolean;
-}
-export interface WorkbenchSpecValidation {
-    ok: boolean;
-    errors: string[];
-    warnings: string[];
-}
-export interface RemoteWorkbenchSpecVersion {
-    id: string;
-    projectId: string;
-    ordinal: number;
-    sourceYaml: string;
-    createdAt: string;
-    updatedAt: string;
-    validation: WorkbenchSpecValidation;
-}
-export interface RemoteWorkbenchEnvironment {
-    id: string;
-    name: string;
-    description: string;
-    currentVersionId: string;
-    builtIn: boolean;
-    createdAt: string;
-    updatedAt: string;
-}
-export interface RemoteWorkbenchEnvironmentVersion {
-    id: string;
-    environmentId: string;
-    name: string;
-    spec: RemoteWorkbenchEnvironmentSpec;
-    imageRef: string;
-    sourceHash: string;
-    sourceType: "builtin" | "dockerfile";
-    build?: {
-        dockerfileRef?: BlobObjectRef;
-        logRef?: BlobObjectRef;
-        error?: string;
-        startedAt?: string;
-        finishedAt?: string;
-    };
-    status: "ready" | "building" | "failed";
-    createdAt: string;
-    updatedAt: string;
-}
-export interface RemoteWorkbenchEnvironmentSpec {
-    base: string;
-    resources: {
-        cpu: number;
-        memoryGb: number;
-        diskGb: number;
-        timeoutMinutes: number;
-    };
-    network: "off" | "on";
-}
-export interface BlobObjectRef {
-    bucket: string;
-    key: string;
-    byteLength: number;
-    sha256: string;
-}
-export type RemoteWorkbenchSnapshotKind = "candidate" | "engineResolve" | "adapters" | "runtime";
-export type WorkspaceWriteEncoding = "utf8" | "base64";
 export interface SurfaceSnapshotFile {
     path: string;
-    kind: "text" | "binary";
-    encoding: WorkspaceWriteEncoding;
+    kind?: "text" | "binary";
+    encoding?: "utf8" | "base64";
     content: string;
-    executable: boolean;
-    contentRedacted?: boolean;
-}
-export interface WorkbenchEngineCaseFiles {
-    public?: SurfaceSnapshotFile[];
-    private?: SurfaceSnapshotFile[];
-    source?: SurfaceSnapshotFile[];
-}
-export interface SurfaceSnapshot {
-    files: SurfaceSnapshotFile[];
+    executable?: boolean;
 }
-export interface RemoteWorkbenchFileInput {
+export type WorkbenchInspectionFileOwnerKind = "version" | "trace" | "artifact";
+export interface WorkbenchInspectionFileContent {
     path: string;
-    content: string;
-    encoding?: WorkspaceWriteEncoding;
+    kind?: SurfaceSnapshotFile["kind"];
+    encoding?: SurfaceSnapshotFile["encoding"];
     executable?: boolean;
+    content?: string;
+    unavailableReason?: string;
 }
-export interface EngineResolveBinding {
-    engine: string;
-    resolver: {
-        use: string;
-        withFingerprint: string;
-    };
+export declare function workbenchInspectionFileContentUnavailableReason(file: Pick<SurfaceSnapshotFile, "kind" | "encoding">): string | null;
+export declare function workbenchInspectionFileContent(file: SurfaceSnapshotFile): WorkbenchInspectionFileContent;
+export declare function workbenchInspectionFileManifest(file: SurfaceSnapshotFile): SurfaceSnapshotFile;
+export interface WorkbenchAgent {
+    name: string;
+    adapter: string;
+    model?: string;
+    config: Record<string, Json>;
 }
-export interface RemoteWorkbenchSnapshotBase {
+export type WorkbenchSkillSourceKind = "local" | "remote";
+export interface WorkbenchSkillInclude {
+    name: string;
+    kind: WorkbenchSkillSourceKind;
+    path?: string;
+    from?: string;
+    ref?: string;
+    resolvedRef?: string;
+    hash?: string;
+    files?: SurfaceSnapshotFile[];
+}
+export interface WorkbenchSkillSource {
+    name: string;
+    kind: WorkbenchSkillSourceKind;
+    path?: string;
+    from?: string;
+    ref?: string;
+    resolvedRef?: string;
+    hash?: string;
+    includes?: WorkbenchSkillInclude[];
+}
+export interface WorkbenchSkillBundleSnapshot {
+    hash: string;
+    skillName: string;
+    entryName: string;
+    source: WorkbenchSkillSource;
     files: SurfaceSnapshotFile[];
-    updatedAt: string;
-}
-export interface RemoteWorkbenchEngineResolveSnapshot extends RemoteWorkbenchSnapshotBase {
-    kind: "engineResolve";
-    engineResolveBinding: EngineResolveBinding;
-}
-export interface RemoteWorkbenchStandardSnapshot extends RemoteWorkbenchSnapshotBase {
-    kind: Exclude<RemoteWorkbenchSnapshotKind, "engineResolve">;
-}
-export type RemoteWorkbenchSnapshot = RemoteWorkbenchEngineResolveSnapshot | RemoteWorkbenchStandardSnapshot;
-export type CandidateStatus = "running" | "evaluated" | "repair_exhausted" | "eval_error" | "agent_error";
-export interface MetricStats {
-    count: number;
-    mean: number;
-    variance: number;
-    stddev: number;
-    min: number;
-    max: number;
-}
-export type EvalCaseStatus = "completed" | "error";
-export type EvalCaseSource = Record<string, Json>;
-export interface CandidateCaseCriterionScore {
-    criterion_id: string;
-    label: string;
-    score: number;
-    pass: boolean;
-    errors?: string[];
-    rationale?: string;
+    includedSkills: WorkbenchSkillInclude[];
+    createdAt: string;
 }
-export interface EvalCaseResult {
+export interface WorkbenchVersion {
     id: string;
-    label?: string;
-    split?: string;
-    status?: EvalCaseStatus;
-    durationMs?: number;
-    metrics: Record<string, number>;
-    source?: EvalCaseSource;
-    feedback?: Json;
-    criteria?: CandidateCaseCriterionScore[];
-}
-export type ExecutionRole = "improver" | "runner" | "engine";
-export type ExecutionUsageCostSource = "provider" | "estimated" | "mixed";
-export interface ExecutionUsage {
-    provider?: string;
-    model?: string;
-    inputTokens?: number;
-    uncachedInputTokens?: number;
-    cachedInputTokens?: number;
-    cacheCreationInputTokens?: number;
-    cacheReadInputTokens?: number;
-    outputTokens?: number;
-    reasoningOutputTokens?: number;
-    totalTokens?: number;
-    costUsd?: number;
-    costSource?: ExecutionUsageCostSource;
-    pricingSource?: string;
-}
-export interface UsageSummary {
-    total?: ExecutionUsage;
-    improver?: ExecutionUsage;
-    runner?: ExecutionUsage;
-    engine?: ExecutionUsage;
+    hash: string;
+    message: string;
+    parentIds: string[];
+    createdAt: string;
+    files: SurfaceSnapshotFile[];
 }
-export interface EvaluationCandidateSummary {
-    id: string;
-    kind: "candidate";
-    label?: string;
+export interface WorkbenchEvalSnapshot {
+    hash: string;
+    files: SurfaceSnapshotFile[];
+    caseCount: number;
 }
-export type EvaluationSampleStatus = "planned" | "running" | "completed" | "error";
-export type EvaluationStatus = EvaluationSampleStatus | "partial";
-export interface EvaluationSampleRecord {
+export type WorkbenchRunKind = "eval" | "improve" | "compare" | "retry";
+export type WorkbenchRunStatus = "running" | "succeeded" | "failed" | "canceled";
+export type WorkbenchJobStatus = "queued" | "running" | "succeeded" | "failed" | "canceled";
+export type WorkbenchArtifactKind = "file" | "directory" | "log" | "scorecard";
+export interface WorkbenchRun {
     id: string;
-    index: number;
-    candidate: EvaluationCandidateSummary;
-    status: EvaluationSampleStatus;
-    startedAt?: string;
+    kind: WorkbenchRunKind;
+    versionId: string;
+    skillName: string;
+    skillBundleHash: string;
+    evalHash: string;
+    agentName: string;
+    agentHash: string;
+    status: WorkbenchRunStatus;
+    score?: number;
+    costUsd?: number;
+    latencyMs?: number;
+    jobIds?: string[];
+    traceIds: string[];
+    createdAt: string;
     finishedAt?: string;
-    durationMs?: number;
-    metrics?: Record<string, number>;
-    usage?: UsageSummary;
+    parentRunId?: string;
+    outputVersionId?: string;
     error?: string;
-    cases?: EvalCaseResult[];
-    feedback?: Json;
-}
-export interface EvaluationCaseStats {
-    id: string;
-    label?: string;
-    split?: string;
-    status?: EvalCaseStatus;
-    sampleCount: number;
-    metrics: Record<string, MetricStats>;
-    durationMs?: MetricStats;
 }
-export interface EvaluationUsageStats {
-    total?: ExecutionUsageStats;
-    improver?: ExecutionUsageStats;
-    runner?: ExecutionUsageStats;
-    engine?: ExecutionUsageStats;
-}
-export interface ExecutionUsageStats {
-    inputTokens?: MetricStats;
-    uncachedInputTokens?: MetricStats;
-    cachedInputTokens?: MetricStats;
-    cacheCreationInputTokens?: MetricStats;
-    cacheReadInputTokens?: MetricStats;
-    outputTokens?: MetricStats;
-    reasoningOutputTokens?: MetricStats;
-    totalTokens?: MetricStats;
-    costUsd?: MetricStats;
+export type WorkbenchAutomationReadinessLevel = "insufficient" | "assist" | "review" | "automate";
+export interface WorkbenchAutomationReadiness {
+    level: WorkbenchAutomationReadinessLevel;
+    label: string;
+    reason: string;
+    runId?: string;
+    score?: number;
+    caseCount?: number;
+    jobCount?: number;
 }
-export interface EvaluationRecord {
-    candidate: EvaluationCandidateSummary;
-    status: EvaluationStatus;
-    sampleCount: number;
-    completedSampleCount: number;
-    errorSampleCount: number;
+export interface WorkbenchJob {
+    id: string;
+    runId: string;
+    kind: WorkbenchRunKind;
+    versionId: string;
+    skillName: string;
+    skillBundleHash: string;
+    evalHash: string;
+    agentName: string;
+    agentHash: string;
+    caseId: string;
+    sample: number;
+    status: WorkbenchJobStatus;
+    score?: number;
+    command?: string;
+    dockerImage?: string;
+    exitCode?: number;
+    artifactIds: string[];
+    traceIds: string[];
+    createdAt: string;
     startedAt?: string;
     finishedAt?: string;
-    metrics?: Record<string, MetricStats>;
-    durationMs?: MetricStats;
-    usage?: EvaluationUsageStats;
-    cases?: EvaluationCaseStats[];
-    samples: EvaluationSampleRecord[];
+    durationMs?: number;
     error?: string;
 }
-export interface EvaluationSummary {
+export interface WorkbenchArtifact {
     id: string;
     runId: string;
-    benchmarkFingerprint: string;
-    candidateFingerprint: string;
-    candidateId: string;
-    candidateName?: string;
-    candidateVersion: number;
-    candidateRunId?: string;
-    candidateRunName?: string;
+    jobId: string;
+    kind: WorkbenchArtifactKind;
+    path: string;
     createdAt: string;
-    updatedAt: string;
-    status: EvaluationStatus;
-    sampleCount: number;
-    completedSampleCount: number;
-    errorSampleCount: number;
-    metrics?: Record<string, MetricStats>;
-    selectionMetric?: string;
-    selectionLabel?: string;
-    selectionScore?: MetricStats;
-    durationMs?: MetricStats;
-    usage?: EvaluationUsageStats;
-    error?: string;
-}
-export interface EvaluationScorecard extends EvaluationSummary {
-    evaluation: EvaluationRecord;
+    files: SurfaceSnapshotFile[];
 }
-export interface WorkbenchEvaluationMetricDescriptor {
+export interface WorkbenchTrace {
     id: string;
-    label: string;
-    direction: "higher" | "lower";
-    kind: "number" | "duration_ms" | "currency_usd";
-    group: "metric" | "execution" | "usage" | "other";
-    primary: boolean;
-    semanticRole?: "performance" | "speed" | "cost";
-}
-export interface WorkbenchEvaluationComparisonRow {
-    evaluationId: string;
     runId: string;
-    candidateId: string;
-    candidateLabel: string;
-    configurationLabel: string;
-    status: EvaluationSummary["status"];
-    score: number | null;
-    metrics: Record<string, number>;
+    jobId?: string;
+    versionId: string;
+    skillName: string;
+    skillBundleHash: string;
+    agentName: string;
     createdAt: string;
-    updatedAt: string;
-    error?: string;
+    request: Json;
+    result: Json;
+    files: SurfaceSnapshotFile[];
 }
-export interface WorkbenchCandidateEvaluationRollup {
-    candidateId: string;
-    candidateLabel: string;
-    evaluationCount: number;
-    completeEvaluationCount: number;
-    scoredEvaluationCount: number;
-    bestEvaluationId: string | null;
-    bestScore: number | null;
-    meanScore: number | null;
-}
-export interface WorkbenchEvaluationComparison {
-    evaluations: EvaluationSummary[];
-    rows: WorkbenchEvaluationComparisonRow[];
-    candidates: WorkbenchCandidateEvaluationRollup[];
-    metrics: WorkbenchEvaluationMetricDescriptor[];
-}
-export interface CandidateSummary {
-    id: string;
-    name?: string;
-    version: number;
-    ordinal: number;
-    benchmarkFingerprint: string;
-    candidateFingerprint: string;
-    ownerUserId?: string;
-    ownerUsername?: string;
-    visibility?: "private" | "public";
+export interface WorkbenchLineageEdge {
+    parentId: string;
+    childId: string;
+    runId?: string;
+    reason: "version" | "improve" | "switch" | "publish";
     createdAt: string;
-    baseId?: string;
-    referenceIds: string[];
-    status: CandidateStatus;
-    fileChanges: string[];
-    usage?: UsageSummary;
-}
-export interface CandidateRecord extends CandidateSummary {
-    eval?: EvaluationRecord;
-    prompt?: string;
-    meta?: Json;
+    message?: string;
 }
-export interface CandidateLineageNode {
-    id: string;
-    active: boolean;
-    summary: CandidateSummary;
+export interface WorkbenchRemote {
+    name: string;
+    url: string;
+    type: "workbench";
 }
-export interface CandidateLineageEdge {
-    id: string;
-    kind: "anchor";
-    sourceId: string;
-    targetId: string;
-}
-export interface CandidateLineageGraph {
-    activeId: string | null;
-    nodes: CandidateLineageNode[];
-    edges: CandidateLineageEdge[];
-}
-export declare function buildCandidateLineage(args: {
-    summaries: readonly CandidateSummary[];
-    activeId: string | null;
-}): CandidateLineageGraph;
-export declare function buildWorkbenchEvaluationComparison(evaluations: readonly EvaluationSummary[]): WorkbenchEvaluationComparison;
-export declare function buildWorkbenchEvaluationMetricDescriptors(evaluations: readonly EvaluationSummary[]): WorkbenchEvaluationMetricDescriptor[];
-export declare function readEvaluationScore(evaluation: EvaluationSummary): number | null;
-export declare function isCompleteEvaluationSummary(evaluation: Pick<EvaluationSummary, "status" | "sampleCount" | "completedSampleCount" | "errorSampleCount">): boolean;
-export declare function formatEvaluationConfigurationLabel(evaluation: Pick<EvaluationSummary, "candidateRunName" | "candidateRunId">): string;
-export type CandidatePreviewMode = "diff" | "raw" | "rendered";
-export type CandidatePreviewKind = "text" | "markdown" | "table" | "spreadsheet" | "image" | "pdf" | "unsupported";
-export type CandidatePreviewSourceEncoding = "utf8" | "base64";
-export type CandidateFileStatus = "added" | "modified" | "unchanged";
-export interface CandidateFileSummary {
-    path: string;
-    old_path: string | null;
-    status: CandidateFileStatus;
-    mime_type: string | null;
-    preview_kind: CandidatePreviewKind;
-    additions: number;
-    deletions: number;
-}
-export interface CandidateFilePreviewSource {
-    content: string;
-    encoding: CandidatePreviewSourceEncoding;
+export interface WorkbenchRefs {
+    current?: string;
+    [name: string]: string | undefined;
 }
-export interface CandidateFilePreview {
+export interface WorkbenchProjectState {
+    schema: "workbench.skill.state.v1";
+    root: string;
+    currentVersionId?: string;
+    refs: WorkbenchRefs;
+    remotes: Record<string, WorkbenchRemote>;
+    defaultSkill?: string;
+    defaultAgent?: string;
+    versions: WorkbenchVersion[];
+    skillSources: WorkbenchSkillSource[];
+    skillBundles: WorkbenchSkillBundleSnapshot[];
+    evals: WorkbenchEvalSnapshot[];
+    agents: WorkbenchAgent[];
+    runs: WorkbenchRun[];
+    jobs: WorkbenchJob[];
+    traces: WorkbenchTrace[];
+    artifacts: WorkbenchArtifact[];
+    lineage: WorkbenchLineageEdge[];
+}
+export interface WorkbenchStatus {
+    root: string;
+    initialized: boolean;
+    currentSkillHash?: string;
+    currentVersionId?: string;
+    hasUnversionedChanges: boolean;
+    defaultSkill?: string;
+    defaultAgent?: string;
+    versionCount: number;
+    skillCount: number;
+    agentCount: number;
+    runCount: number;
+    remoteCount: number;
+    pendingSyncCount?: number;
+    lastScore?: number;
+    automationReadiness?: WorkbenchAutomationReadiness;
+}
+export interface WorkbenchComparisonCell {
+    versionId: string;
+    skillName: string;
+    skillBundleHash: string;
+    evalHash: string;
+    agentName: string;
+    runId?: string;
+    score?: number;
+    costUsd?: number;
+    latencyMs?: number;
+    automationReadiness?: WorkbenchAutomationReadiness;
+}
+export interface WorkbenchComparison {
+    evalHash?: string;
+    versions: WorkbenchVersion[];
+    skills: WorkbenchSkillBundleSnapshot[];
+    agents: WorkbenchAgent[];
+    cells: WorkbenchComparisonCell[];
+}
+export interface WorkbenchInspectionSnapshot {
+    root: string;
+    status: WorkbenchStatus;
+    versions: WorkbenchVersion[];
+    skillSources: WorkbenchSkillSource[];
+    skillBundles: WorkbenchSkillBundleSnapshot[];
+    agents: WorkbenchAgent[];
+    runs: WorkbenchRun[];
+    jobs: WorkbenchJob[];
+    traces: WorkbenchTrace[];
+    artifacts: WorkbenchArtifact[];
+    lineage: WorkbenchLineageEdge[];
+    remotes: WorkbenchRemote[];
+    refs: WorkbenchRefs;
+    publication?: WorkbenchPublication;
+}
+export interface WorkbenchPublication {
+    versionId: string;
+    installUrl: string;
+    pinnedInstallUrl: string;
+}
+export interface WorkbenchObjectPack {
+    schema: "workbench.object-pack.v1";
+    createdAt: string;
+    refs: WorkbenchRefs;
+    defaultSkill?: string;
+    defaultAgent?: string;
+    versions: WorkbenchVersion[];
+    skillSources: WorkbenchSkillSource[];
+    skillBundles: WorkbenchSkillBundleSnapshot[];
+    evals: WorkbenchEvalSnapshot[];
+    agents: WorkbenchAgent[];
+    runs: WorkbenchRun[];
+    jobs: WorkbenchJob[];
+    traces: WorkbenchTrace[];
+    artifacts: WorkbenchArtifact[];
+    lineage: WorkbenchLineageEdge[];
+}
+export interface WorkbenchFilePreview {
     path: string;
-    view: CandidatePreviewMode;
-    mime_type: string | null;
-    preview_kind: CandidatePreviewKind;
-    diff: string | null;
-    source: CandidateFilePreviewSource | null;
-    rendered_html: string | null;
-}
-export interface CandidateCaseCriterionResult {
-    criterion_id: string;
-    pass: boolean;
-    score: number;
-    errors: string[];
-    rationale?: string;
+    source?: SurfaceSnapshotFile;
+    renderedText?: string;
+    diff?: string;
 }
-export interface CandidateCaseExecutionRef {
-    runId: string;
-    kind: string;
-    role: WorkbenchExecutionEventRole;
-    status: RemoteWorkbenchJobStatus;
-    jobIds: string[];
-    executionIds: string[];
-    createdAt?: string;
-    startedAt?: string;
-    finishedAt?: string;
-    durationMs?: number;
-    caseId?: string;
-    sampleIndex?: number;
-    attemptIndex?: number;
-}
-export interface CandidateCaseReview {
-    candidateId: string;
-    caseId: string;
-    caseLabel: string;
-    sampleId?: string;
-    sampleIndex?: number;
-    status?: EvalCaseStatus | RemoteWorkbenchJobStatus;
-    metrics: Record<string, number>;
-    durationMs?: number;
-    source?: EvalCaseSource;
-    feedback?: Json;
-    executions: CandidateCaseExecutionRef[];
-    criteria_results: CandidateCaseCriterionResult[];
-}
-export type RunStatus = "queued" | "running" | "finished";
-export type RunOutcome = "ok" | "error" | "cancelled";
-export type RemoteRunWorkflow = "eval" | "improve";
-export interface RunSummary {
-    id: string;
-    workflow: RemoteRunWorkflow;
-    benchmarkFingerprint: string;
-    status: RunStatus;
-    candidateId?: string | null;
-    candidateRunId?: string;
-    candidateRunName?: string;
-    startedAt: string;
-    finishedAt?: string;
-    durationMs?: number;
-    improver: string;
-    engineRun: string;
-    strategy: string;
-    optimizeOn?: string;
-    selectBy?: string;
-    budget: number;
-    repairBudget: number;
-    attemptsRequested: number;
-    attemptsExecuted: number;
-    samples: number;
-    executionFingerprint?: string;
-    stoppedReason?: "budget_exhausted" | "completed" | "dry_run" | "cancelled";
-    outcome?: RunOutcome;
-    error?: string;
-    activeCandidateId?: string | null;
-    outputCandidateId?: string | null;
-}
-export interface WorkbenchRuntimeRun extends RunSummary {
-    jobCount?: number;
-    completedJobCount?: number;
-    failedJobCount?: number;
-}
-export interface RuntimeEvent {
-    id: string;
-    at: string;
-    type: "run_started" | "job_queued" | "job_started" | "job_progress" | "sandbox_allocated" | "sandbox_stopped" | "candidate_created" | "candidate_evaluated" | "active_changed" | "run_finished";
-    runId?: string;
-    jobId?: string;
-    candidateId?: string;
-    baseId?: string;
-    activeId?: string;
-    status?: CandidateStatus | RemoteWorkbenchJobStatus;
-    metrics?: Record<string, number>;
-    detail?: Record<string, Json>;
-}
-export interface RuntimeSnapshot {
-    workspaceRoot: string;
-    activeId: string | null;
-    currentBenchmarkFingerprint: string | null;
-    summaries: CandidateSummary[];
-    evaluations: EvaluationSummary[];
-    runs: RunSummary[];
-}
-export interface WorkbenchRuntimeCandidateFiles {
-    candidateId: string;
+export interface WorkbenchFileSurface {
     files: SurfaceSnapshotFile[];
+    preview: WorkbenchFilePreview | null;
 }
-export interface WorkbenchRuntimeExecutionFiles {
-    jobId: string;
-    files: SurfaceSnapshotFile[];
+export interface WorkbenchSpecValidation {
+    ok: boolean;
+    errors: string[];
+    warnings: string[];
 }
-export interface WorkbenchRuntimeBundle {
-    schema: "workbench.runtime.bundle.v1";
-    activeId: string | null;
-    candidates: CandidateRecord[];
-    candidateFiles: WorkbenchRuntimeCandidateFiles[];
-    evaluations: EvaluationScorecard[];
-    runs: WorkbenchRuntimeRun[];
-    jobs: RemoteWorkbenchJob[];
-    executionFiles: WorkbenchRuntimeExecutionFiles[];
-    events: RuntimeEvent[];
-}
-export interface WorkbenchRuntimeBundleStats {
-    candidates: number;
-    candidateFiles: number;
-    evaluations: number;
-    runs: number;
-    jobs: number;
-    executionFiles: number;
-    events: number;
-    activeId: string | null;
-}
-export interface WorkbenchRuntimeImportResult {
-    changed: boolean;
-    stats: WorkbenchRuntimeBundleStats;
-}
-export interface WorkbenchProjectSourceResources {
-    cpu?: number;
-    memoryGb?: number;
-    diskGb?: number;
-    timeoutMinutes?: number;
-}
-export interface WorkbenchProjectStateSource {
-    source: string;
-    files: SurfaceSnapshotFile[];
-    candidateFiles: SurfaceSnapshotFile[];
-    engineResolveFiles: SurfaceSnapshotFile[];
-    engineResolveBinding: EngineResolveBinding;
-    adapterFiles: SurfaceSnapshotFile[];
-    dockerfile: string;
-    runtimeDockerfile: string;
-    runtimeFiles: SurfaceSnapshotFile[];
+export interface RemoteWorkbenchEnvironmentSpec {
+    base: string;
+    resources: {
+        cpu: number;
+        memoryGb: number;
+        diskGb: number;
+        timeoutMinutes: number;
+    };
     network: "off" | "on";
-    resources: WorkbenchProjectSourceResources;
-    revisionId?: string;
-    fingerprint?: string;
-}
-export interface WorkbenchProjectStateBase {
-    sourceRevisionId?: string;
-    sourceFingerprint?: string;
-    runtimeFingerprint?: string;
 }
-export interface WorkbenchProjectStateRemote {
+export interface RemoteWorkbenchEnvironmentVersion {
     id: string;
-    remote: string;
-    ownerUsername: string;
+    environmentId: string;
     name: string;
-    visibility: "private" | "public";
-}
-export interface WorkbenchProjectState {
-    schema: "workbench.project.state.v1";
-    project: WorkbenchProjectStateRemote;
-    base: WorkbenchProjectStateBase;
-    source: WorkbenchProjectStateSource;
-    runtime: WorkbenchRuntimeBundle;
-}
-export interface WorkbenchProjectStateImportResult {
-    changed: boolean;
-    source: {
-        changed: boolean;
-        revisionId?: string;
-        fingerprint?: string;
-    };
-    runtime: WorkbenchRuntimeImportResult;
-    state: WorkbenchProjectState;
-}
-export type WorkbenchRemoteContractSchema = "workbench.remote.capabilities.v1" | "workbench.remote.run.request.v1" | "workbench.remote.job.claim_request.v1" | "workbench.remote.job.claim.v1" | "workbench.remote.job.renewal.v1" | "workbench.remote.job.renewal_result.v1" | "workbench.remote.job.progress.v1" | "workbench.remote.job.completion.v1" | "workbench.remote.job.retry.v1";
-export type WorkbenchRemoteProductionSandbox = "firecracker";
-export type WorkbenchRemoteLocalSandbox = "docker";
-export type WorkbenchRemoteNetworkPolicy = "open" | "none";
-export interface WorkbenchRemoteCapabilities {
-    schema: "workbench.remote.capabilities.v1";
-    contractVersion: 1;
-    projectState: {
-        schema: WorkbenchProjectState["schema"];
-        guardedSourceWrites: true;
-        immutableRuntimeFacts: true;
-    };
-    execution: {
-        fencedJobLeases: true;
-        idempotentCompletion: true;
-        progressIsBestEffort: true;
-        maxJobsPerRun: number;
-    };
-    sandbox: {
-        production: WorkbenchRemoteProductionSandbox;
-        local: WorkbenchRemoteLocalSandbox;
-        networkPolicies: WorkbenchRemoteNetworkPolicy[];
-    };
-    blobs: {
-        contentAddressed: boolean;
-        maxUploadBytes: number;
+    spec: RemoteWorkbenchEnvironmentSpec;
+    imageRef: string;
+    sourceHash: string;
+    sourceType: "builtin" | "dockerfile";
+    build?: {
+        dockerfileRef?: BlobObjectRef;
+        logRef?: BlobObjectRef;
+        error?: string;
+        startedAt?: string;
+        finishedAt?: string;
     };
+    status: "ready" | "building" | "failed";
+    createdAt: string;
+    updatedAt: string;
 }
-export interface WorkbenchRemoteRunRequest {
-    schema: "workbench.remote.run.request.v1";
-    workflow: "eval" | "improve";
-    budget?: number;
-    samples: number;
-    candidateId?: string;
-    sourceYaml?: string;
-    candidateFiles?: RemoteWorkbenchFileInput[];
-    adapterFiles?: RemoteWorkbenchFileInput[];
-    selectedSamples?: Array<{
-        caseId: string;
-        sampleIndex: number;
-    }>;
-    preserveActive?: boolean;
-    rerun?: boolean;
-}
-export interface AuthoredWorkbenchCandidateRunSpec extends WorkbenchAuthoredAdapterSpec {
-    name: string;
-}
-export interface WorkbenchCaseSelector {
-    all?: true;
-    split?: string;
-}
-export interface WorkbenchSelectionSpec {
-    metric: string;
-    cases?: WorkbenchCaseSelector;
-}
-export interface AuthoredWorkbenchCandidateImproveSpec extends WorkbenchAuthoredAdapterSpec {
-    edits: string[];
-    optimizeOn?: WorkbenchCaseSelector;
-    selectBy?: WorkbenchSelectionSpec;
-}
-export interface AuthoredWorkbenchCandidateSpec {
-    name: string;
-    description?: string;
-    files: WorkbenchPathRef;
-    prepare?: WorkbenchCandidatePrepareSpec;
-    defaultRun?: string;
-    selectedRunId?: string;
-    runs: Record<string, AuthoredWorkbenchCandidateRunSpec>;
-    improve?: AuthoredWorkbenchCandidateImproveSpec;
-}
-export interface WorkbenchCandidatePrepareSpec {
-    command: string;
-}
-export interface WorkbenchPathRef {
-    path: string;
-}
-export interface WorkbenchAuthoredAdapterSpec {
-    use: string;
-    auth?: string | Record<string, string>;
-    with?: Record<string, Json>;
-}
-export interface AuthoredWorkbenchRuntimeSpec {
-    dockerfile: string;
-    resources?: {
-        cpu?: number;
-        memoryGb?: number;
-        diskGb?: number;
-        timeoutMinutes?: number;
-    };
-    network?: {
-        egress?: "none" | "open";
+export interface EngineResolveBinding {
+    engine: string;
+    resolver: {
+        use: string;
+        withFingerprint: string;
     };
 }
-export type AuthoredWorkbenchImproveSpec = WorkbenchAuthoredAdapterSpec;
-export type AuthoredWorkbenchRunSpec = WorkbenchAuthoredAdapterSpec;
-export type AuthoredWorkbenchScoreSpec = WorkbenchAuthoredAdapterSpec;
-export interface AuthoredWorkbenchEngineConfig {
-    tasks?: WorkbenchAuthoredAdapterSpec;
-    environment: AuthoredWorkbenchRuntimeSpec;
-    score: AuthoredWorkbenchScoreSpec;
-}
-export interface AuthoredWorkbenchEngineSpec {
-    use: string;
-    auth?: string | Record<string, string>;
-    with?: AuthoredWorkbenchEngineConfig | Record<string, Json>;
-}
-export interface AuthoredWorkbenchBenchmarkSpec {
-    name: string;
-    description: string;
-    engine: AuthoredWorkbenchEngineSpec;
-}
-export interface AuthoredWorkbenchSourceSpec {
-    version: 4;
-    benchmark: AuthoredWorkbenchBenchmarkSpec;
-    candidate: AuthoredWorkbenchCandidateSpec;
-}
 export type WorkbenchExecutionPurpose = "improve" | "attempt";
 export type WorkbenchSandboxTemplateKind = "snapshot" | "oci";
 export interface WorkbenchAdapterInvocation {
@@ -704,11 +338,11 @@ export interface WorkbenchSandboxAllocation {
 }
 export interface WorkbenchExecutionCapability {
     executionId: string;
-    candidate: {
+    skill: {
         tenantId: string;
         projectId: string;
         runId: string;
-        candidateId?: string;
+        versionId?: string;
     };
     inputs: WorkbenchExecutionInputRef[];
     outputPrefix: string;
@@ -735,7 +369,7 @@ export interface WorkbenchExecutionInputRef {
     mountPath: string;
     writable: boolean;
 }
-export type WorkbenchExecutionOutputSchema = "workbench.candidate_patch.v1" | "workbench.result.v1" | string;
+export type WorkbenchExecutionOutputSchema = "workbench.skill_patch.v1" | "workbench.result.v1" | string;
 export interface WorkbenchExecutionOutputContract {
     name: string;
     schema: WorkbenchExecutionOutputSchema;
@@ -761,7 +395,7 @@ export interface WorkbenchExecutionSpec {
     id: string;
     projectId: string;
     runId: string;
-    candidateId?: string;
+    versionId?: string;
     purpose: WorkbenchExecutionPurpose;
     adapter: WorkbenchAdapterInvocation;
     sandbox: WorkbenchSandboxTemplate;
@@ -770,12 +404,87 @@ export interface WorkbenchExecutionSpec {
     policy: WorkbenchExecutionPolicy;
     metadata: Record<string, Json>;
 }
-export interface WorkbenchCandidatePatch {
+export interface BlobObjectRef {
+    bucket: string;
+    key: string;
+    byteLength: number;
+    sha256: string;
+}
+export interface WorkbenchSkillPatch {
     files: SurfaceSnapshotFile[];
     fileChanges: string[];
     summary?: string;
     feedback?: Json;
 }
+export interface WorkbenchCaseCriterionScore {
+    criterion_id: string;
+    label: string;
+    score: number;
+    pass: boolean;
+    errors?: string[];
+    rationale?: string;
+}
+export interface MetricStats {
+    count: number;
+    mean: number;
+    variance: number;
+    stddev: number;
+    min: number;
+    max: number;
+}
+export type EvalCaseStatus = "completed" | "error";
+export type EvalCaseSource = Record<string, Json>;
+export interface EvalCaseResult {
+    id: string;
+    label?: string;
+    split?: string;
+    status?: EvalCaseStatus;
+    durationMs?: number;
+    metrics: Record<string, number>;
+    source?: EvalCaseSource;
+    feedback?: Json;
+    criteria?: WorkbenchCaseCriterionScore[];
+}
+export type ExecutionRole = "improver" | "runner" | "engine";
+export type ExecutionUsageCostSource = "provider" | "estimated" | "mixed";
+export interface ExecutionUsage {
+    provider?: string;
+    model?: string;
+    inputTokens?: number;
+    uncachedInputTokens?: number;
+    cachedInputTokens?: number;
+    cacheCreationInputTokens?: number;
+    cacheReadInputTokens?: number;
+    outputTokens?: number;
+    reasoningOutputTokens?: number;
+    totalTokens?: number;
+    costUsd?: number;
+    costSource?: ExecutionUsageCostSource;
+    pricingSource?: string;
+}
+export interface UsageSummary {
+    total?: ExecutionUsage;
+    improver?: ExecutionUsage;
+    runner?: ExecutionUsage;
+    engine?: ExecutionUsage;
+}
+export interface EvaluationUsageStats {
+    total?: ExecutionUsageStats;
+    improver?: ExecutionUsageStats;
+    runner?: ExecutionUsageStats;
+    engine?: ExecutionUsageStats;
+}
+export interface ExecutionUsageStats {
+    inputTokens?: MetricStats;
+    uncachedInputTokens?: MetricStats;
+    cachedInputTokens?: MetricStats;
+    cacheCreationInputTokens?: MetricStats;
+    cacheReadInputTokens?: MetricStats;
+    reasoningOutputTokens?: MetricStats;
+    outputTokens?: MetricStats;
+    totalTokens?: MetricStats;
+    costUsd?: MetricStats;
+}
 export interface WorkbenchResult {
     score: number;
     metrics?: Record<string, number>;
@@ -888,6 +597,24 @@ export interface WorkbenchTraceSession {
     trace: WorkbenchExecutionTrace;
     metadata?: Record<string, Json>;
 }
+export type RemoteWorkbenchJobStatus = "queued" | "running" | "succeeded" | "failed" | "cancelled";
+export type RemoteWorkbenchJobKind = "execute";
+export interface RemoteWorkbenchJob {
+    id: string;
+    projectId: string;
+    runId: string;
+    versionId?: string;
+    kind: RemoteWorkbenchJobKind;
+    status: RemoteWorkbenchJobStatus;
+    attempt: number;
+    createdAt: string;
+    updatedAt: string;
+    startedAt?: string;
+    finishedAt?: string;
+    input: Json;
+    output?: Json;
+    error?: string;
+}
 export interface WorkbenchExecutionEvidence {
     id: string;
     kind: string;
@@ -896,7 +623,7 @@ export interface WorkbenchExecutionEvidence {
     status: RemoteWorkbenchJobStatus;
     jobIds: string[];
     executionIds: string[];
-    candidateId?: string;
+    versionId?: string;
     caseId?: string;
     sampleIndex?: number;
     attemptIndex?: number;
@@ -908,44 +635,6 @@ export interface WorkbenchExecutionTraceDetail {
     runId: string;
     executions: WorkbenchExecutionEvidence[];
 }
-export interface AuthoredWorkbenchCaseSummary {
-    id: string;
-    slug: string;
-    path: string;
-    name: string;
-    split?: string;
-    fileCount: number;
-}
-export interface AuthoredWorkbenchSourceFile {
-    path: string;
-    content: string;
-}
-export interface AuthoredWorkbenchSourceDocument {
-    path: string;
-    exists: boolean;
-    source_yaml: string;
-    source_files: AuthoredWorkbenchSourceFile[];
-    spec: AuthoredWorkbenchSourceSpec | null;
-    cases: AuthoredWorkbenchCaseSummary[];
-}
-export type RemoteWorkbenchJobStatus = "queued" | "running" | "succeeded" | "failed" | "cancelled";
-export type RemoteWorkbenchJobKind = "execute";
-export interface RemoteWorkbenchJob {
-    id: string;
-    projectId: string;
-    runId: string;
-    candidateId?: string;
-    kind: RemoteWorkbenchJobKind;
-    status: RemoteWorkbenchJobStatus;
-    attempt: number;
-    createdAt: string;
-    updatedAt: string;
-    startedAt?: string;
-    finishedAt?: string;
-    input: Json;
-    output?: Json;
-    error?: string;
-}
 export interface WorkbenchRemoteJobClaimRequest {
     schema: "workbench.remote.job.claim_request.v1";
     ownerUserId: string;
@@ -1015,32 +704,6 @@ export interface WorkbenchRemoteJobRetry {
     leaseToken: string;
     reason: string;
 }
-export interface RemoteWorkbenchRun extends WorkbenchRuntimeRun {
-    projectId: string;
-    environmentVersionId?: string;
-    specVersionId: string;
-    candidateId: string | null;
-    activeCandidateId?: string | null;
-    outputCandidateId?: string | null;
-    input: {
-        benchmarkFingerprint: string;
-        candidateFingerprint: string;
-        baseCandidateId: string | null;
-        payerUserId?: string;
-        candidateOwnerUserId?: string;
-        candidateOwnerUsername?: string;
-        preserveActiveCandidateId?: string | null;
-        selectedSamples?: Array<{
-            caseId: string;
-            sampleIndex: number;
-        }>;
-        sourceYaml?: string;
-        candidateSourceFiles?: SurfaceSnapshotFile[];
-        baseFiles: SurfaceSnapshotFile[];
-        engineResolveFiles: SurfaceSnapshotFile[];
-    };
-    jobCount: number;
-    completedJobCount: number;
-    failedJobCount: number;
-}
+export declare function isReservedWorkbenchAdapterAuthEnvName(name: string): boolean;
+export declare function assertWorkbenchAdapterAuthEnvNameAllowed(name: string): void;
 //# sourceMappingURL=index.d.ts.map