@tangle-network/agent-eval 0.205.2 → 0.207.0

package/dist/benchmarks/index.d.ts

@@ -1,7 +1,7 @@
 import { c as RunSplitTag } from "../run-record-DF-fuu31.js";
 import { a as CampaignResult } from "../types-ZUy5HGY5.js";
-import "../index-BRdZ0oJo.js";
-import { Yt as CampaignStorage } from "../search-kernel-DNpsHb8P.js";
+import "../index-CojsA_EE.js";
+import { Xt as CampaignStorage } from "../search-kernel-CZ8KJQRw.js";
 import { a as BenchmarkFamily, c as BenchmarkSource, i as BenchmarkEvaluation, l as BenchmarkTaskKind, n as BenchmarkAdapter, o as BenchmarkResponder, r as BenchmarkDatasetItem, s as BenchmarkScenario, t as BENCHMARK_SPLIT_SEED, u as deterministicSplit } from "../types-DZQpUJse.js";
 //#region src/benchmarks/calibration.d.ts
 interface BenchmarkMetricCalibrationOptions<TPayload = unknown, TArtifact = string> {

package/dist/campaign/index.d.ts

@@ -4,7 +4,7 @@ import { t as SeriesDistribution } from "../descriptive-B2iPaT9J.js";
 import { $ as paretoSignificanceGate, B as PowerPreflight, G as BuildEvidenceVectorOptions, H as powerPreflight, J as ParetoSignificanceGateOptions, K as EvidenceVector, Q as paretoPolicy, U as AxisEvidence, V as PowerPreflightOptions, W as AxisVerdict, X as PromotionPolicy, Y as PromotionObjective, Z as buildEvidenceVector, _ as sequentialPairedGate, a as detectScale, c as pairHoldout, d as SequentialDecision, f as SequentialObservation, g as sequentialDecide, h as SequentialStreamState, i as PairedHoldout, l as SequentialDecideFn, m as SequentialPairedGateOptions, n as HeldoutSignificance, o as dimensionRegressions, p as SequentialPairedGate, q as ObjectiveSource, r as HeldoutSignificanceOptions, s as heldoutSignificance, t as DimensionRegression, u as SequentialDecideOptions } from "../statistical-heldout-COfgpWah.js";
 import { $ as SearchTokenAccounting, A as SearchLedgerHash, B as SearchOperationStartedEvent, C as SearchEdgeOperator, D as SearchFailureReason, E as SearchExecutionIdentity, F as SearchNodeRegisteredEvent, G as SearchSplit, H as SearchProposerKind, I as SearchNodeStatus, J as SearchSurfaceEffect, K as SearchSplitTasks, L as SearchOpenedEvent, M as SearchNodeDecidedEvent, N as SearchNodeDecision, O as SearchLedgerEntry, P as SearchNodeRef, Q as SearchTaskOutcome, R as SearchOperationKind, S as SearchEdgeAttribution, T as SearchEstimateMethod, U as SearchReservation, V as SearchProposer, W as SearchSourceRef, X as SearchSurfaceKind, Y as SearchSurfaceEvidence, Z as SearchTask, _ as SearchClaimPower, a as SearchAttemptAccounting, b as SearchClosedEvent, c as SearchCancelReason, d as SearchCellCancelledEvent, et as SearchTraceRef, f as SearchCellSettledEvent, g as SearchClaimFinalist, h as SearchClaimEstimator, i as SearchArtifactRef, j as SearchModelIdentity, k as SearchLedgerEvent, l as SearchCandidateSurface, m as SearchClaim, n as SEARCH_LEDGER_SCHEMA, o as SearchAudit, p as SearchCellStage, q as SearchSplits, r as SearchArtifactKind, s as SearchBudget, t as NodeEstimate, tt as SearchUnknown, u as SearchCellAllocatedEvent, v as SearchClaimTest, w as SearchEdgeRecordedEvent, x as SearchCostAccounting, y as SearchCloseReason, z as SearchOperationRecordedEvent } from "../search-ledger-types-p2xKfiDJ.js";
 import { $ as SearchNode, A as LabeledScenarioWrite, B as ScoredSurfaceOutcome, C as JudgeDimension, D as LabeledScenarioSampleArgs, E as LabeledScenarioRecord, F as ProposeContext, G as labelTrustRank, H as SurfaceProposer, I as ProposedCandidate, J as renderSearchSummary, K as SearchProposerView, L as RedactionStatus, M as OptimizerConfig, N as ParetoParent, O as LabeledScenarioSource, P as ProposalTrackContext, Q as SearchDecisionRecord, R as Scenario, S as JudgeConfig, T as LabelTrust, U as TraceSpan, V as SessionScript, W as isProposedCandidate, X as SearchCell, Y as searchProposerView, Z as SearchCompletion, _ as GateDecision, a as CampaignResult, at as SearchUnitScore, b as GenerationRecord, c as CampaignTraceWriter, ct as searchTaskSetDigest, d as DispatchContext, et as SearchOperation, f as DispatchFn, g as GateContribution, h as GateContext, i as CampaignCostMeter, it as SearchStateView, j as MutableSurface, k as LabeledScenarioStore, l as CodeSurface, lt as searchUnitScores, m as GateCheckStatus, n as CampaignArtifactWriter, nt as SearchSpend, o as CampaignScenarioIdentity, ot as searchCellId, p as Gate, q as SearchSummaryOptions, r as CampaignCellResult, rt as SearchState, s as CampaignTokenUsage, st as searchNodeId, t as CampaignAggregates, tt as SearchScoredCell, u as ComponentSurface, v as GateResult, w as JudgeScore, x as JudgeAggregate, y as GenerationCandidate, z as ScenarioAggregate } from "../types-ZUy5HGY5.js";
-import { $ as ScoreboardRow, $i as RunImprovementLoopOptions, $n as TraceAnalystScenario, $r as GepaOptimizationRecipe, $t as importExternalEvaluations, A as ScenarioSignal, Aa as CampaignCoverage, Ai as GepaCandidatePopulationSummary, An as CrossSurfaceComponentEvidence, B as FinalizedProfileMatrixResult, Bn as CrossSurfaceInteractionTask, Bt as FsLabeledScenarioStoreOptions, C as SingleRunLockOptions, Ca as CacheRead, Ci as compareOptimizationMethods, Cn as CrossSurfaceBootstrapPolicy, Cr as campaignMeanComposite, D as searchProposalExecution, Da as cellCachePath, Di as costFromLedgerSummary, Dn as CrossSurfaceCandidateOutcome, Dr as DefaultProductionGateOptions, E as searchModelIdentity, Ea as buildCellSchedule, Ei as combineComparisonCosts, En as CrossSurfaceCandidateEvidence, Er as DefaultProductionGateCheck, F as RetainedDispatchOptions, Fa as campaignSplitDigest, Fi as ExternalOptimizerObservationSummary, Fn as CrossSurfaceIneligibilityReason, G as RunProfileMatrixSegmentOptions, Gn as CrossSurfacePairwiseEntry, Gr as SkillOptOptimizationMethodConfig, Gt as ScoredRollout, H as ProfileMatrixPlan, Hn as CrossSurfacePairCompatibility, Ht as RolloutArgumentDiff, I as RetainedDispatchOutcome, Ia as campaignSplitDigestFromIdentities, Ii as ExternalOptimizerSubmittedCandidate, In as CrossSurfaceInteractionAwareSelection, It as scopedOptimizationMethod, J as runProfileMatrixSegment, Ji as llmJudge, Jn as CrossSurfaceSelectionPolicy, Jr as skillOptOptimizationMethod, Jt as rolloutArgumentDiff, K as createProfileMatrixPlan, Ki as LlmJudgeDimension, Kn as CrossSurfaceRankedSingle, Kr as SkillOptRunnerCommand, Kt as UngroundedLiteralReport, L as createRetainedDispatch, Li as readExternalOptimizerObservationArtifact, Ln as CrossSurfaceInteractionEffect, Lt as sequentialOptimizationMethod, M as selectDiscriminative, Ma as assertCampaignSplitIdentity, Mi as readGepaCandidatePopulationArtifact, Mn as CrossSurfaceDistribution, N as resolveRunDir, Na as campaignCoverage, Ni as ExternalOptimizerExecutionSummary, Nn as CrossSurfaceEligibility, O as searchReceiptAccounting, Oa as FinalEvidencePolicy, Oi as GepaCandidatePopulationArtifact, On as CrossSurfaceCandidateSummary, Or as DefaultProductionRewardHackingOptions, P as tangleTracesRoot, Pa as campaignScenarioIdentity, Pi as ExternalOptimizerObservationArtifact, Pn as CrossSurfaceEvidenceBreakdown, Q as ScoreboardRenderOptions, Qi as transientDispatchFailure, Qn as TraceAnalystArtifact, Qr as GepaOptimizationMethodConfig, Qt as externalSurface, R as CreateProfileMatrixPlanOptions, Rn as CrossSurfaceInteractionPath, Rt as neutralizeText, S as SingleRunLock, Sa as CacheIssueReason, Si as OptimizationTokenUsage, Sn as CrossSurfaceBestSingleSelection, Sr as campaignBreakdown, T as campaignCellSearchResult, Ta as CellScheduleSlot, Ti as ComparisonCost, Tn as CrossSurfaceCandidateComparison, Tr as heldOutGate, U as ProfileMatrixRow, Ui as ReferenceEquivalenceScenario, Un as CrossSurfacePairEvidence, Ut as RolloutArgumentDiffOptions, V as ProfileMatrixCoverage, Vi as ReferenceEquivalenceJudgeOptions, Vn as CrossSurfaceNaiveStackSelection, Vt as LabeledScenarioStoreError, W as ProfileMatrixSegmentResult, Wi as createReferenceEquivalenceJudge, Wn as CrossSurfacePairIncompatibilityReason, Wr as composeGate, Wt as RolloutCall, X as PlaybackDriver, Xi as isTransientTransportFailure, Xn as CrossSurfaceTaskRow, Xr as GepaEngineOptions, Xt as GepaEvaluationImport, Y as PlaybackContext, Yi as TransientFailureOptions, Yn as CrossSurfaceSelections, Yr as GepaAdaptiveEngineRun, Yt as ExternalSearchOptimizer, Z as PlaybackStep, Zi as quotaExhaustedUntil, Zn as BuildTraceAnalystSurfaceDispatchOptions, Zr as GepaEngineRun, Zt as GepaPopulationImport, _ as componentSurfaceIdentityMaterial, _a as runCampaign, _i as OptimizationMethodProvenance, _n as analyzeCrossSurfaceInteractions, _r as provenanceRecordPath, a as WorktreeAdapterError, aa as runOptimization, ai as ExternalTextOptimizationMethodConfig, an as EvalFixture, ar as LoopProvenanceArgsFromResult, at as scoreUserStory, b as surfaceDispatchRef, ba as PlanCampaignRunOptions, bi as OptimizationMethodScore, bn as CrossSurfaceAdditionRejectionReason, br as CampaignBreakdown, c as verifyCodeSurface, ca as SearchLedgerIntegrityError, ci as ExternalOptimizationExample, cn as EvalFixtureRunPlan, cr as LoopProvenanceEvidence, ct as ProfileDispatchFn, d as PhoenixEvaluationResultLike, da as openAutoPr, di as CompareOptimizationMethodsOptions, dn as LoadEvalFixtureScenariosOptions, dr as buildLoopProvenanceRecord, dt as RunProfileMatrixOptions, ea as RunImprovementLoopResult, ei as GepaRunnerCommand, en as importGepaPopulation, er as buildTraceAnalystSurfaceDispatch, et as ScoreboardSummary, f as PhoenixEvaluatorLike, fa as RunEvalOptions, fi as OptimizationMethod, fn as PlanEvalFixtureRunOptions, fr as campaignMeasurementDigest, ft as RunProfileMatrixResult, g as codeSurfaceIdentityMaterial, ga as RunCampaignOptions, gi as OptimizationMethodPairwise, gn as planEvalFixtureRun, gr as loopProvenanceSpans, h as assertCodeSurfaceIdentity, ha as CampaignCellRetryPolicy, hi as OptimizationMethodInput, hn as loadEvalFixtureScenarios, hr as loopProvenanceArgsFromResult, i as WorktreeAdapter, ia as RunOptimizationResult, ii as externalTextOptimizationMethod, in as neutralizationGate, ir as EmitLoopProvenanceResult, it as renderScoreboardMarkdown, j as scoreDiscrimination, ja as assertCampaignDesign, ji as GepaCandidateSelectionScore, jn as CrossSurfaceCompositionStep, k as DiscriminationScore, ka as FinalEvidenceUse, ki as GepaCandidatePopulationCandidate, kn as CrossSurfaceComponent, kr as defaultProductionGate, l as AutoevalsScoreLike, la as OpenAutoPrOptions, li as ExternalTextEvaluationResponse, ln as EvalFixtureScenario, lr as LoopProvenanceOptimizationMethod, lt as ProfileMatrixError, m as phoenixEvaluatorJudge, ma as CampaignCellFailureReceipt, mi as OptimizationMethodComposition, mn as loadEvalFixture, mr as emitLoopProvenance, mt as runProfileMatrix, n as GitWorktreeAdapterOptions, na as PremeasuredOptimizationBaseline, ni as OpenAICompatibleOptimizerModel, nn as recordSkillOptSearch, nr as BuildLoopProvenanceArgs, nt as UserStoryVerdict, o as gitWorktreeAdapter, oa as SearchLedgerConflictError, oi as ExternalTextOptimizerContext, on as EvalFixtureFile, or as LoopProvenanceBackend, ot as scoreboardSummary, p as autoevalsScorerJudge, pa as runEval, pi as OptimizationMethodComparison, pn as discoverEvalFixtures, pr as canonicalDigest, pt as ScenarioRollup, q as finalizeProfileMatrix, qi as LlmJudgeOptions, qn as CrossSurfaceRelativeCost, qr as SkillOptTrainerConfig, qt as classifyUngroundedLiterals, r as Worktree, ra as RunOptimizationOptions, ri as OptimizerModelBudget, rn as NeutralizationGateOptions, rr as EmitLoopProvenanceArgs, rt as makePlaybackDispatch, s as resolveWorktreePath, sa as SearchLedgerError, si as ExternalTextOptimizerResult, sn as EvalFixtureLoadOptions, sr as LoopProvenanceCandidate, st as userStoryScoreboard, t as CodeSurfaceVerification, ta as runImprovementLoop, ti as gepaOptimizationMethod, tn as recordGepaSearch, tr as traceAnalystQualityJudge, tt as UserStory, u as AutoevalsScorerLike, ua as OpenAutoPrResult, ui as decodeExternalTextCandidate, un as EvalFixtureValidationMode, ur as LoopProvenanceRecord, ut as ProfileSummary, v as renderSurfaceDiff, va as CampaignRunPlan, vi as OptimizationMethodResult, vn as AnalyzeCrossSurfaceInteractionsInput, vr as provenanceSpansPath, w as acquireSingleRunLock, wa as readCachedCell, wi as optimizationTokenUsageFromSummary, wn as CrossSurfaceCandidate, wr as HeldOutGateOptions, x as surfaceHash, xa as planCampaignRun, xi as OptimizationPackageSource, xn as CrossSurfaceAttemptCompleteness, y as surfaceContentHash, ya as CampaignRunPlanCell, yi as OptimizationMethodRunOptions, yn as CrossSurfaceAdditionDecision, yr as verifyLoopProvenanceRecord, z as FinalizeProfileMatrixOptions, zn as CrossSurfaceInteractionReport, zt as FsLabeledScenarioStore } from "../index-BRdZ0oJo.js";
-import { A as SearchCellPlan, At as SearchHistoryCoverageRow, B as aide, Bt as FileSearchLedger, C as decideSearchClaim, Ct as searchEdgeId, D as AshaOptions, Dt as SearchHistoryAdmissionOptions, E as verifySearchClaim, Et as CreateSearchHistoryReceiptInput, F as AideOptions, Ft as assertSearchHistoryMatchesState, G as EstimateNodeCellsInput, Gt as openSearchLedger, H as crowdedFrontierParent, Ht as SearchLedger, I as SearchExpansion, It as createSearchHistoryReceipt, J as SearchPosterior, Jt as validateSearchLedgerEvent, K as NodePosterior, Kt as parseSearchLedgerLine, L as SearchPolicy, Lt as searchHistoryCoverageRow, M as SearchUnitExtension, Mt as SearchHistoryReceipt, N as asha, Nt as SearchHistoryRequiredError, O as SearchAllocationView, Ot as SearchHistoryAuditSummary, P as uniform, Pt as assertCompleteSearchHistory, Q as searchPosterior, Qt as inMemoryCampaignStorage, R as SearchPolicyNode, Rt as verifySearchHistoryArtifact, S as SearchClaimVerification, St as developmentClaim, T as searchClaimReserveUsd, Tt as surfaceNode, U as incumbent, Ut as SearchLedgerAppendResult, V as beam, Vt as OpenSearchLedgerOptions, W as incumbentWithOperatorBandit, Wt as SearchLedgerTrustedHeadMode, X as estimateNodeFromCells, Xt as createRunCostLedger, Y as estimateNode, Yt as CampaignStorage, Z as searchCellSetDigest, Zt as fsCampaignStorage, _ as searchPolicyView, _t as SearchOpening, a as SearchCellWork, b as SearchClaimDecision, bt as SearchRunIdentity, c as SearchProposalBlob, d as SearchProposedChild, f as SearchProposerPort, g as searchExpansionIndex, gt as SearchLedgerBinding, h as searchDivergence, ht as RegisterSearchNodeInput, i as SearchCellResult, j as SearchRungDecision, jt as SearchHistoryPolicy, k as SearchAllocator, kt as SearchHistoryCoverage, l as SearchProposalRequest, m as runSearch, mt as RecordSearchEdgeInput, n as SEARCH_KERNEL_SOURCE, o as SearchExecutor, p as SearchRunResult, pt as AllocateSearchCellInput, q as SEARCH_ESTIMATOR, qt as replaySearchLedgerText, r as SearchArtifactCodec, s as SearchLane, t as RunSearchOptions, u as SearchProposalResult, v as PlanSearchClaimInput, vt as SearchRecorder, w as planSearchClaim, wt as surfaceDiff, x as SearchClaimPlan, xt as SettleSearchCellInput, y as SEARCH_CLAIM_RULE, yt as SearchRecorderOptions, z as SearchPolicyView, zt as verifySearchHistoryReceipt } from "../search-kernel-DNpsHb8P.js";
+import { $ as ScoreboardRow, $i as RunImprovementLoopOptions, $n as TraceAnalystScenario, $r as GepaOptimizationRecipe, $t as importExternalEvaluations, A as ScenarioSignal, Aa as CampaignCoverage, Ai as GepaCandidatePopulationSummary, An as CrossSurfaceComponentEvidence, B as FinalizedProfileMatrixResult, Bn as CrossSurfaceInteractionTask, Bt as FsLabeledScenarioStoreOptions, C as SingleRunLockOptions, Ca as CacheRead, Ci as compareOptimizationMethods, Cn as CrossSurfaceBootstrapPolicy, Cr as campaignMeanComposite, D as searchProposalExecution, Da as cellCachePath, Di as costFromLedgerSummary, Dn as CrossSurfaceCandidateOutcome, Dr as DefaultProductionGateOptions, E as searchModelIdentity, Ea as buildCellSchedule, Ei as combineComparisonCosts, En as CrossSurfaceCandidateEvidence, Er as DefaultProductionGateCheck, F as RetainedDispatchOptions, Fa as campaignSplitDigest, Fi as ExternalOptimizerObservationSummary, Fn as CrossSurfaceIneligibilityReason, G as RunProfileMatrixSegmentOptions, Gn as CrossSurfacePairwiseEntry, Gr as SkillOptOptimizationMethodConfig, Gt as ScoredRollout, H as ProfileMatrixPlan, Hn as CrossSurfacePairCompatibility, Ht as RolloutArgumentDiff, I as RetainedDispatchOutcome, Ia as campaignSplitDigestFromIdentities, Ii as ExternalOptimizerSubmittedCandidate, In as CrossSurfaceInteractionAwareSelection, It as scopedOptimizationMethod, J as runProfileMatrixSegment, Ji as llmJudge, Jn as CrossSurfaceSelectionPolicy, Jr as skillOptOptimizationMethod, Jt as rolloutArgumentDiff, K as createProfileMatrixPlan, Ki as LlmJudgeDimension, Kn as CrossSurfaceRankedSingle, Kr as SkillOptRunnerCommand, Kt as UngroundedLiteralReport, L as createRetainedDispatch, Li as readExternalOptimizerObservationArtifact, Ln as CrossSurfaceInteractionEffect, Lt as sequentialOptimizationMethod, M as selectDiscriminative, Ma as assertCampaignSplitIdentity, Mi as readGepaCandidatePopulationArtifact, Mn as CrossSurfaceDistribution, N as resolveRunDir, Na as campaignCoverage, Ni as ExternalOptimizerExecutionSummary, Nn as CrossSurfaceEligibility, O as searchReceiptAccounting, Oa as FinalEvidencePolicy, Oi as GepaCandidatePopulationArtifact, On as CrossSurfaceCandidateSummary, Or as DefaultProductionRewardHackingOptions, P as tangleTracesRoot, Pa as campaignScenarioIdentity, Pi as ExternalOptimizerObservationArtifact, Pn as CrossSurfaceEvidenceBreakdown, Q as ScoreboardRenderOptions, Qi as transientDispatchFailure, Qn as TraceAnalystArtifact, Qr as GepaOptimizationMethodConfig, Qt as externalSurface, R as CreateProfileMatrixPlanOptions, Rn as CrossSurfaceInteractionPath, Rt as neutralizeText, S as SingleRunLock, Sa as CacheIssueReason, Si as OptimizationTokenUsage, Sn as CrossSurfaceBestSingleSelection, Sr as campaignBreakdown, T as campaignCellSearchResult, Ta as CellScheduleSlot, Ti as ComparisonCost, Tn as CrossSurfaceCandidateComparison, Tr as heldOutGate, U as ProfileMatrixRow, Ui as ReferenceEquivalenceScenario, Un as CrossSurfacePairEvidence, Ut as RolloutArgumentDiffOptions, V as ProfileMatrixCoverage, Vi as ReferenceEquivalenceJudgeOptions, Vn as CrossSurfaceNaiveStackSelection, Vt as LabeledScenarioStoreError, W as ProfileMatrixSegmentResult, Wi as createReferenceEquivalenceJudge, Wn as CrossSurfacePairIncompatibilityReason, Wr as composeGate, Wt as RolloutCall, X as PlaybackDriver, Xi as isTransientTransportFailure, Xn as CrossSurfaceTaskRow, Xr as GepaEngineOptions, Xt as GepaEvaluationImport, Y as PlaybackContext, Yi as TransientFailureOptions, Yn as CrossSurfaceSelections, Yr as GepaAdaptiveEngineRun, Yt as ExternalSearchOptimizer, Z as PlaybackStep, Zi as quotaExhaustedUntil, Zn as BuildTraceAnalystSurfaceDispatchOptions, Zr as GepaEngineRun, Zt as GepaPopulationImport, _ as componentSurfaceIdentityMaterial, _a as runCampaign, _i as OptimizationMethodProvenance, _n as analyzeCrossSurfaceInteractions, _r as provenanceRecordPath, a as WorktreeAdapterError, aa as runOptimization, ai as ExternalTextOptimizationMethodConfig, an as EvalFixture, ar as LoopProvenanceArgsFromResult, at as scoreUserStory, b as surfaceDispatchRef, ba as PlanCampaignRunOptions, bi as OptimizationMethodScore, bn as CrossSurfaceAdditionRejectionReason, br as CampaignBreakdown, c as verifyCodeSurface, ca as SearchLedgerIntegrityError, ci as ExternalOptimizationExample, cn as EvalFixtureRunPlan, cr as LoopProvenanceEvidence, ct as ProfileDispatchFn, d as PhoenixEvaluationResultLike, da as openAutoPr, di as CompareOptimizationMethodsOptions, dn as LoadEvalFixtureScenariosOptions, dr as buildLoopProvenanceRecord, dt as RunProfileMatrixOptions, ea as RunImprovementLoopResult, ei as GepaRunnerCommand, en as importGepaPopulation, er as buildTraceAnalystSurfaceDispatch, et as ScoreboardSummary, f as PhoenixEvaluatorLike, fa as RunEvalOptions, fi as OptimizationMethod, fn as PlanEvalFixtureRunOptions, fr as campaignMeasurementDigest, ft as RunProfileMatrixResult, g as codeSurfaceIdentityMaterial, ga as RunCampaignOptions, gi as OptimizationMethodPairwise, gn as planEvalFixtureRun, gr as loopProvenanceSpans, h as assertCodeSurfaceIdentity, ha as CampaignCellRetryPolicy, hi as OptimizationMethodInput, hn as loadEvalFixtureScenarios, hr as loopProvenanceArgsFromResult, i as WorktreeAdapter, ia as RunOptimizationResult, ii as externalTextOptimizationMethod, in as neutralizationGate, ir as EmitLoopProvenanceResult, it as renderScoreboardMarkdown, j as scoreDiscrimination, ja as assertCampaignDesign, ji as GepaCandidateSelectionScore, jn as CrossSurfaceCompositionStep, k as DiscriminationScore, ka as FinalEvidenceUse, ki as GepaCandidatePopulationCandidate, kn as CrossSurfaceComponent, kr as defaultProductionGate, l as AutoevalsScoreLike, la as OpenAutoPrOptions, li as ExternalTextEvaluationResponse, ln as EvalFixtureScenario, lr as LoopProvenanceOptimizationMethod, lt as ProfileMatrixError, m as phoenixEvaluatorJudge, ma as CampaignCellFailureReceipt, mi as OptimizationMethodComposition, mn as loadEvalFixture, mr as emitLoopProvenance, mt as runProfileMatrix, n as GitWorktreeAdapterOptions, na as PremeasuredOptimizationBaseline, ni as OpenAICompatibleOptimizerModel, nn as recordSkillOptSearch, nr as BuildLoopProvenanceArgs, nt as UserStoryVerdict, o as gitWorktreeAdapter, oa as SearchLedgerConflictError, oi as ExternalTextOptimizerContext, on as EvalFixtureFile, or as LoopProvenanceBackend, ot as scoreboardSummary, p as autoevalsScorerJudge, pa as runEval, pi as OptimizationMethodComparison, pn as discoverEvalFixtures, pr as canonicalDigest, pt as ScenarioRollup, q as finalizeProfileMatrix, qi as LlmJudgeOptions, qn as CrossSurfaceRelativeCost, qr as SkillOptTrainerConfig, qt as classifyUngroundedLiterals, r as Worktree, ra as RunOptimizationOptions, ri as OptimizerModelBudget, rn as NeutralizationGateOptions, rr as EmitLoopProvenanceArgs, rt as makePlaybackDispatch, s as resolveWorktreePath, sa as SearchLedgerError, si as ExternalTextOptimizerResult, sn as EvalFixtureLoadOptions, sr as LoopProvenanceCandidate, st as userStoryScoreboard, t as CodeSurfaceVerification, ta as runImprovementLoop, ti as gepaOptimizationMethod, tn as recordGepaSearch, tr as traceAnalystQualityJudge, tt as UserStory, u as AutoevalsScorerLike, ua as OpenAutoPrResult, ui as decodeExternalTextCandidate, un as EvalFixtureValidationMode, ur as LoopProvenanceRecord, ut as ProfileSummary, v as renderSurfaceDiff, va as CampaignRunPlan, vi as OptimizationMethodResult, vn as AnalyzeCrossSurfaceInteractionsInput, vr as provenanceSpansPath, w as acquireSingleRunLock, wa as readCachedCell, wi as optimizationTokenUsageFromSummary, wn as CrossSurfaceCandidate, wr as HeldOutGateOptions, x as surfaceHash, xa as planCampaignRun, xi as OptimizationPackageSource, xn as CrossSurfaceAttemptCompleteness, y as surfaceContentHash, ya as CampaignRunPlanCell, yi as OptimizationMethodRunOptions, yn as CrossSurfaceAdditionDecision, yr as verifyLoopProvenanceRecord, z as FinalizeProfileMatrixOptions, zn as CrossSurfaceInteractionReport, zt as FsLabeledScenarioStore } from "../index-CojsA_EE.js";
+import { $ as searchPosterior, $t as inMemoryCampaignStorage, A as SearchCellPlan, At as SearchHistoryCoverage, B as aide, Bt as verifySearchHistoryReceipt, C as decideSearchClaim, Ct as developmentClaim, D as AshaOptions, Dt as CreateSearchHistoryReceiptInput, E as verifySearchClaim, Et as surfaceNode, F as AideOptions, Ft as assertCompleteSearchHistory, G as EstimateNodeCellsInput, Gt as SearchLedgerTrustedHeadMode, H as crowdedFrontierParent, Ht as OpenSearchLedgerOptions, I as SearchExpansion, It as assertSearchHistoryMatchesState, J as SearchPosterior, Jt as replaySearchLedgerText, K as NodePosterior, Kt as openSearchLedger, L as SearchPolicy, Lt as createSearchHistoryReceipt, M as SearchUnitExtension, Mt as SearchHistoryPolicy, N as asha, Nt as SearchHistoryReceipt, O as SearchAllocationView, Ot as SearchHistoryAdmissionOptions, P as uniform, Pt as SearchHistoryRequiredError, Q as searchEstimateMethod, Qt as fsCampaignStorage, R as SearchPolicyNode, Rt as searchHistoryCoverageRow, S as SearchClaimVerification, St as SettleSearchCellInput, T as searchClaimReserveUsd, Tt as surfaceDiff, U as incumbent, Ut as SearchLedger, V as beam, Vt as FileSearchLedger, W as incumbentWithOperatorBandit, Wt as SearchLedgerAppendResult, X as estimateNodeFromCells, Xt as CampaignStorage, Y as estimateNode, Yt as validateSearchLedgerEvent, Z as searchCellSetDigest, Zt as createRunCostLedger, _ as searchPolicyView, _t as SearchLedgerBinding, a as SearchCellWork, b as SearchClaimDecision, bt as SearchRecorderOptions, c as SearchProposalBlob, d as SearchProposedChild, f as SearchProposerPort, g as searchExpansionIndex, gt as RegisterSearchNodeInput, h as searchDivergence, ht as RecordSearchEdgeInput, i as SearchCellResult, j as SearchRungDecision, jt as SearchHistoryCoverageRow, k as SearchAllocator, kt as SearchHistoryAuditSummary, l as SearchProposalRequest, m as runSearch, mt as AllocateSearchCellInput, n as SEARCH_KERNEL_SOURCE, o as SearchExecutor, p as SearchRunResult, q as SEARCH_ESTIMATOR, qt as parseSearchLedgerLine, r as SearchArtifactCodec, s as SearchLane, t as RunSearchOptions, u as SearchProposalResult, v as PlanSearchClaimInput, vt as SearchOpening, w as planSearchClaim, wt as searchEdgeId, x as SearchClaimPlan, xt as SearchRunIdentity, y as SEARCH_CLAIM_RULE, yt as SearchRecorder, z as SearchPolicyView, zt as verifySearchHistoryArtifact } from "../search-kernel-CZ8KJQRw.js";
 import { B as ExternalOptimizerModelBudget, F as ExternalOptimizerCallbackLimits, G as ExternalOptimizerProcessLimits, H as ExternalOptimizerModelCallRequest, I as ExternalOptimizerChatRequest, L as ExternalOptimizerEndpointFormat, N as DEFAULT_EXTERNAL_OPTIMIZER_CALLBACK_LIMITS, P as DEFAULT_EXTERNAL_OPTIMIZER_PROCESS_LIMITS, Q as resolveExternalOptimizerProcessLimits, R as ExternalOptimizerEvaluationObservation, U as ExternalOptimizerModelCallResult, V as ExternalOptimizerModelCall, W as ExternalOptimizerModelExecutionObservation, Y as ExternalTextCandidate, Z as resolveExternalOptimizerCallbackLimits, q as ExternalOptimizerRunnerCommand, z as ExternalOptimizerEvaluationRefusalReason } from "../agent-profile-EpO0HURy.js";
-export { type AideOptions, type AllocateSearchCellInput, type AnalyzeCrossSurfaceInteractionsInput, type AshaOptions, type AutoevalsScoreLike, type AutoevalsScorerLike, type AxisEvidence, type AxisVerdict, type BuildEvidenceVectorOptions, type BuildLoopProvenanceArgs, type BuildTraceAnalystSurfaceDispatchOptions, type CacheIssueReason, type CacheRead, type CampaignAggregates, type CampaignArtifactWriter, type CampaignBreakdown, type CampaignCellFailureReceipt, type CampaignCellResult, type CampaignCellRetryPolicy, type CampaignCostMeter, type CampaignCoverage, type CampaignResult, type CampaignRunPlan, type CampaignRunPlanCell, type CampaignScenarioIdentity, type CampaignStorage, type CampaignTokenUsage, type CampaignTraceWriter, type CellScheduleSlot, type CodeSurface, type CodeSurfaceVerification, type CompareOptimizationMethodsOptions, type ComparisonCost, type ComponentSurface, type CostLedgerHandle, type CreateProfileMatrixPlanOptions, type CreateSearchHistoryReceiptInput, type CrossSurfaceAdditionDecision, type CrossSurfaceAdditionRejectionReason, type CrossSurfaceAttemptCompleteness, type CrossSurfaceBestSingleSelection, type CrossSurfaceBootstrapPolicy, type CrossSurfaceCandidate, type CrossSurfaceCandidateComparison, type CrossSurfaceCandidateEvidence, type CrossSurfaceCandidateOutcome, type CrossSurfaceCandidateSummary, type CrossSurfaceComponent, type CrossSurfaceComponentEvidence, type CrossSurfaceCompositionStep, type CrossSurfaceDistribution, type CrossSurfaceEligibility, type CrossSurfaceEvidenceBreakdown, type CrossSurfaceIneligibilityReason, type CrossSurfaceInteractionAwareSelection, type CrossSurfaceInteractionEffect, type CrossSurfaceInteractionPath, type CrossSurfaceInteractionReport, type CrossSurfaceInteractionTask, type CrossSurfaceNaiveStackSelection, type CrossSurfacePairCompatibility, type CrossSurfacePairEvidence, type CrossSurfacePairIncompatibilityReason, type CrossSurfacePairwiseEntry, type CrossSurfaceRankedSingle, type CrossSurfaceRelativeCost, type CrossSurfaceSelectionPolicy, type CrossSurfaceSelections, type CrossSurfaceTaskRow, DEFAULT_EXTERNAL_OPTIMIZER_CALLBACK_LIMITS, DEFAULT_EXTERNAL_OPTIMIZER_PROCESS_LIMITS, type DefaultProductionGateCheck, type DefaultProductionGateOptions, type DefaultProductionRewardHackingOptions, type DimensionRegression, type DiscriminationScore, type DispatchContext, type DispatchFn, type EmitLoopProvenanceArgs, type EmitLoopProvenanceResult, type EstimateNodeCellsInput, type EvalFixture, type EvalFixtureFile, type EvalFixtureLoadOptions, type EvalFixtureRunPlan, type EvalFixtureScenario, type EvalFixtureValidationMode, type EvidenceVector, type ExternalOptimizationExample, type ExternalOptimizerCallbackLimits, type ExternalOptimizerChatRequest, type ExternalOptimizerEndpointFormat, type ExternalOptimizerEvaluationObservation, type ExternalOptimizerEvaluationRefusalReason, type ExternalOptimizerExecutionSummary, type ExternalOptimizerModelBudget, type ExternalOptimizerModelCall, type ExternalOptimizerModelCallRequest, type ExternalOptimizerModelCallResult, type ExternalOptimizerModelExecutionObservation, type ExternalOptimizerObservationArtifact, type ExternalOptimizerObservationSummary, type ExternalOptimizerProcessLimits, type ExternalOptimizerRunnerCommand, type ExternalOptimizerSubmittedCandidate, type ExternalSearchOptimizer, type ExternalTextCandidate, type ExternalTextEvaluationResponse, type ExternalTextOptimizationMethodConfig, type ExternalTextOptimizerContext, type ExternalTextOptimizerResult, FileSearchLedger, type FinalEvidencePolicy, type FinalEvidenceUse, type FinalizeProfileMatrixOptions, type FinalizedProfileMatrixResult, FsLabeledScenarioStore, type FsLabeledScenarioStoreOptions, type Gate, type GateCheckStatus, type GateContext, type GateContribution, type GateDecision, type GateResult, type GenerationCandidate, type GenerationRecord, type GepaAdaptiveEngineRun, type GepaCandidatePopulationArtifact, type GepaCandidatePopulationCandidate, type GepaCandidatePopulationSummary, type GepaCandidateSelectionScore, type GepaEngineOptions, type GepaEngineRun, type GepaEvaluationImport, type GepaOptimizationMethodConfig, type GepaOptimizationRecipe, type GepaPopulationImport, type GepaRunnerCommand, type GitWorktreeAdapterOptions, type HeldOutGateOptions, type HeldoutSignificance, type HeldoutSignificanceOptions, type JudgeAggregate, type JudgeConfig, type JudgeDimension, type JudgeScore, type LabelTrust, type LabeledScenarioRecord, type LabeledScenarioSampleArgs, type LabeledScenarioSource, type LabeledScenarioStore, LabeledScenarioStoreError, type LabeledScenarioWrite, type LlmJudgeDimension, type LlmJudgeOptions, type LoadEvalFixtureScenariosOptions, type LoopProvenanceArgsFromResult, type LoopProvenanceBackend, type LoopProvenanceCandidate, type LoopProvenanceEvidence, type LoopProvenanceOptimizationMethod, type LoopProvenanceRecord, type MutableSurface, type NeutralizationGateOptions, type NodeEstimate, type NodePosterior, type ObjectiveSource, type OpenAICompatibleOptimizerModel, type OpenAutoPrOptions, type OpenAutoPrResult, type OpenSearchLedgerOptions, type OptimizationMethod, type OptimizationMethodComparison, type OptimizationMethodComposition, type OptimizationMethodInput, type OptimizationMethodPairwise, type OptimizationMethodProvenance, type OptimizationMethodResult, type OptimizationMethodRunOptions, type OptimizationMethodScore, type OptimizationPackageSource, type OptimizationTokenUsage, type OptimizerConfig, type OptimizerModelBudget, type PairedHoldout, type ParetoParent, type ParetoSignificanceGateOptions, type PendingCostCallView, type PhoenixEvaluationResultLike, type PhoenixEvaluatorLike, type PlanCampaignRunOptions, type PlanEvalFixtureRunOptions, type PlanSearchClaimInput, type PlaybackContext, type PlaybackDriver, type PlaybackStep, type PowerPreflight, type PowerPreflightOptions, type PremeasuredOptimizationBaseline, type ProfileDispatchFn, type ProfileMatrixCoverage, ProfileMatrixError, type ProfileMatrixPlan, type ProfileMatrixRow, type ProfileMatrixSegmentResult, type ProfileSummary, type PromotionObjective, type PromotionPolicy, type ProposalFinding, type ProposalFindingOrigin, type ProposalTrackContext, type ProposeContext, type ProposedCandidate, type RecordSearchEdgeInput, type RedactionStatus, type ReferenceEquivalenceJudgeOptions, type ReferenceEquivalenceScenario, type RegisterSearchNodeInput, type RetainedDispatchOptions, type RetainedDispatchOutcome, type RolloutArgumentDiff, type RolloutArgumentDiffOptions, type RolloutCall, type RunCampaignOptions, type RunEvalOptions, type RunImprovementLoopOptions, type RunImprovementLoopResult, type RunOptimizationOptions, type RunOptimizationResult, type RunProfileMatrixOptions, type RunProfileMatrixResult, type RunProfileMatrixSegmentOptions, type RunSearchOptions, SEARCH_CLAIM_RULE, SEARCH_ESTIMATOR, SEARCH_KERNEL_SOURCE, SEARCH_LEDGER_SCHEMA, type Scenario, type ScenarioAggregate, type ScenarioRollup, type ScenarioSignal, type ScoreboardRenderOptions, type ScoreboardRow, type ScoreboardSummary, type ScoredRollout, type ScoredSurfaceOutcome, type SearchAllocationView, type SearchAllocator, type SearchArtifactCodec, type SearchArtifactKind, type SearchArtifactRef, type SearchAttemptAccounting, type SearchAudit, type SearchBudget, type SearchCancelReason, type SearchCandidateSurface, type SearchCell, type SearchCellAllocatedEvent, type SearchCellCancelledEvent, type SearchCellPlan, type SearchCellResult, type SearchCellSettledEvent, type SearchCellStage, type SearchCellWork, type SearchClaim, type SearchClaimDecision, type SearchClaimEstimator, type SearchClaimFinalist, type SearchClaimPlan, type SearchClaimPower, type SearchClaimTest, type SearchClaimVerification, type SearchCloseReason, type SearchClosedEvent, type SearchCompletion, type SearchCostAccounting, type SearchDecisionRecord, type SearchEdgeAttribution, type SearchEdgeOperator, type SearchEdgeRecordedEvent, type SearchEstimateMethod, type SearchExecutionIdentity, type SearchExecutor, type SearchExpansion, type SearchFailureReason, type SearchHistoryAdmissionOptions, type SearchHistoryAuditSummary, type SearchHistoryCoverage, type SearchHistoryCoverageRow, type SearchHistoryPolicy, type SearchHistoryReceipt, SearchHistoryRequiredError, type SearchLane, type SearchLedger, type SearchLedgerAppendResult, type SearchLedgerBinding, SearchLedgerConflictError, type SearchLedgerEntry, SearchLedgerError, type SearchLedgerEvent, type SearchLedgerHash, SearchLedgerIntegrityError, type SearchLedgerTrustedHeadMode, type SearchModelIdentity, type SearchNode, type SearchNodeDecidedEvent, type SearchNodeDecision, type SearchNodeRef, type SearchNodeRegisteredEvent, type SearchNodeStatus, type SearchOpenedEvent, type SearchOpening, type SearchOperation, type SearchOperationKind, type SearchOperationRecordedEvent, type SearchOperationStartedEvent, type SearchPolicy, type SearchPolicyNode, type SearchPolicyView, type SearchPosterior, type SearchProposalBlob, type SearchProposalRequest, type SearchProposalResult, type SearchProposedChild, type SearchProposer, type SearchProposerKind, type SearchProposerPort, type SearchProposerView, SearchRecorder, type SearchRecorderOptions, type SearchReservation, type SearchRunIdentity, type SearchRunResult, type SearchRungDecision, type SearchScoredCell, type SearchSourceRef, type SearchSpend, type SearchSplit, type SearchSplitTasks, type SearchSplits, SearchState, SearchStateView, type SearchSummaryOptions, type SearchSurfaceEffect, type SearchSurfaceEvidence, type SearchSurfaceKind, type SearchTask, type SearchTaskOutcome, type SearchTokenAccounting, type SearchTraceRef, type SearchUnitExtension, type SearchUnitScore, type SearchUnknown, type SequentialDecideFn, type SequentialDecideOptions, type SequentialDecision, type SequentialObservation, type SequentialPairedGate, type SequentialPairedGateOptions, type SequentialStreamState, type SeriesDistribution, type SessionScript, type SettleSearchCellInput, type SingleRunLock, type SingleRunLockOptions, type SkillOptOptimizationMethodConfig, type SkillOptRunnerCommand, type SkillOptTrainerConfig, type SurfaceProposer, type TraceAnalystArtifact, type TraceAnalystScenario, type TraceSpan, type TransientFailureOptions, type UngroundedLiteralReport, type UserStory, type UserStoryVerdict, type Worktree, type WorktreeAdapter, WorktreeAdapterError, acquireSingleRunLock, aide, analyzeCrossSurfaceInteractions, asha, assertCampaignDesign, assertCampaignSplitIdentity, assertCodeSurfaceIdentity, assertCompleteSearchHistory, assertSearchHistoryMatchesState, autoevalsScorerJudge, beam, buildCellSchedule, buildEvidenceVector, buildLoopProvenanceRecord, buildTraceAnalystSurfaceDispatch, campaignBreakdown, campaignCellSearchResult, campaignCoverage, campaignMeanComposite, campaignMeasurementDigest, campaignScenarioIdentity, campaignSplitDigest, campaignSplitDigestFromIdentities, canonicalDigest, cellCachePath, classifyUngroundedLiterals, codeSurfaceIdentityMaterial, combineComparisonCosts, compareOptimizationMethods, componentSurfaceIdentityMaterial, composeGate, costFromLedgerSummary, createProfileMatrixPlan, createReferenceEquivalenceJudge, createRetainedDispatch, createRunCostLedger, createSearchHistoryReceipt, crowdedFrontierParent, decideSearchClaim, decodeExternalTextCandidate, defaultProductionGate, detectScale, developmentClaim, dimensionRegressions, discoverEvalFixtures, emitLoopProvenance, estimateNode, estimateNodeFromCells, externalSurface, externalTextOptimizationMethod, finalizeProfileMatrix, fsCampaignStorage, gepaOptimizationMethod, gitWorktreeAdapter, heldOutGate, heldoutSignificance, importExternalEvaluations, importGepaPopulation, inMemoryCampaignStorage, incumbent, incumbentWithOperatorBandit, isProposedCandidate, isTransientTransportFailure, labelTrustRank, llmJudge, loadEvalFixture, loadEvalFixtureScenarios, loopProvenanceArgsFromResult, loopProvenanceSpans, makePlaybackDispatch, makeProposalFinding, neutralizationGate, neutralizeText, openAutoPr, openSearchLedger, optimizationTokenUsageFromSummary, pairHoldout, paretoPolicy, paretoSignificanceGate, parseSearchLedgerLine, phoenixEvaluatorJudge, planCampaignRun, planEvalFixtureRun, planSearchClaim, powerPreflight, provenanceRecordPath, provenanceSpansPath, quotaExhaustedUntil, readCachedCell, readExternalOptimizerObservationArtifact, readGepaCandidatePopulationArtifact, recordGepaSearch, recordSkillOptSearch, renderScoreboardMarkdown, renderSearchSummary, renderSurfaceDiff, replaySearchLedgerText, resolveExternalOptimizerCallbackLimits, resolveExternalOptimizerProcessLimits, resolveRunDir, resolveWorktreePath, rolloutArgumentDiff, runCampaign, runEval, runImprovementLoop, runOptimization, runProfileMatrix, runProfileMatrixSegment, runSearch, scopedOptimizationMethod, scoreDiscrimination, scoreUserStory, scoreboardSummary, searchCellId, searchCellSetDigest, searchClaimReserveUsd, searchDivergence, searchEdgeId, searchExpansionIndex, searchHistoryCoverageRow, searchModelIdentity, searchNodeId, searchPolicyView, searchPosterior, searchProposalExecution, searchProposerView, searchReceiptAccounting, searchTaskSetDigest, searchUnitScores, selectDiscriminative, sequentialDecide, sequentialOptimizationMethod, sequentialPairedGate, skillOptOptimizationMethod, surfaceContentHash, surfaceDiff, surfaceDispatchRef, surfaceHash, surfaceNode, tangleTracesRoot, traceAnalystQualityJudge, transientDispatchFailure, uniform, userStoryScoreboard, validateSearchLedgerEvent, verifyCodeSurface, verifyLoopProvenanceRecord, verifySearchClaim, verifySearchHistoryArtifact, verifySearchHistoryReceipt };
+export { type AideOptions, type AllocateSearchCellInput, type AnalyzeCrossSurfaceInteractionsInput, type AshaOptions, type AutoevalsScoreLike, type AutoevalsScorerLike, type AxisEvidence, type AxisVerdict, type BuildEvidenceVectorOptions, type BuildLoopProvenanceArgs, type BuildTraceAnalystSurfaceDispatchOptions, type CacheIssueReason, type CacheRead, type CampaignAggregates, type CampaignArtifactWriter, type CampaignBreakdown, type CampaignCellFailureReceipt, type CampaignCellResult, type CampaignCellRetryPolicy, type CampaignCostMeter, type CampaignCoverage, type CampaignResult, type CampaignRunPlan, type CampaignRunPlanCell, type CampaignScenarioIdentity, type CampaignStorage, type CampaignTokenUsage, type CampaignTraceWriter, type CellScheduleSlot, type CodeSurface, type CodeSurfaceVerification, type CompareOptimizationMethodsOptions, type ComparisonCost, type ComponentSurface, type CostLedgerHandle, type CreateProfileMatrixPlanOptions, type CreateSearchHistoryReceiptInput, type CrossSurfaceAdditionDecision, type CrossSurfaceAdditionRejectionReason, type CrossSurfaceAttemptCompleteness, type CrossSurfaceBestSingleSelection, type CrossSurfaceBootstrapPolicy, type CrossSurfaceCandidate, type CrossSurfaceCandidateComparison, type CrossSurfaceCandidateEvidence, type CrossSurfaceCandidateOutcome, type CrossSurfaceCandidateSummary, type CrossSurfaceComponent, type CrossSurfaceComponentEvidence, type CrossSurfaceCompositionStep, type CrossSurfaceDistribution, type CrossSurfaceEligibility, type CrossSurfaceEvidenceBreakdown, type CrossSurfaceIneligibilityReason, type CrossSurfaceInteractionAwareSelection, type CrossSurfaceInteractionEffect, type CrossSurfaceInteractionPath, type CrossSurfaceInteractionReport, type CrossSurfaceInteractionTask, type CrossSurfaceNaiveStackSelection, type CrossSurfacePairCompatibility, type CrossSurfacePairEvidence, type CrossSurfacePairIncompatibilityReason, type CrossSurfacePairwiseEntry, type CrossSurfaceRankedSingle, type CrossSurfaceRelativeCost, type CrossSurfaceSelectionPolicy, type CrossSurfaceSelections, type CrossSurfaceTaskRow, DEFAULT_EXTERNAL_OPTIMIZER_CALLBACK_LIMITS, DEFAULT_EXTERNAL_OPTIMIZER_PROCESS_LIMITS, type DefaultProductionGateCheck, type DefaultProductionGateOptions, type DefaultProductionRewardHackingOptions, type DimensionRegression, type DiscriminationScore, type DispatchContext, type DispatchFn, type EmitLoopProvenanceArgs, type EmitLoopProvenanceResult, type EstimateNodeCellsInput, type EvalFixture, type EvalFixtureFile, type EvalFixtureLoadOptions, type EvalFixtureRunPlan, type EvalFixtureScenario, type EvalFixtureValidationMode, type EvidenceVector, type ExternalOptimizationExample, type ExternalOptimizerCallbackLimits, type ExternalOptimizerChatRequest, type ExternalOptimizerEndpointFormat, type ExternalOptimizerEvaluationObservation, type ExternalOptimizerEvaluationRefusalReason, type ExternalOptimizerExecutionSummary, type ExternalOptimizerModelBudget, type ExternalOptimizerModelCall, type ExternalOptimizerModelCallRequest, type ExternalOptimizerModelCallResult, type ExternalOptimizerModelExecutionObservation, type ExternalOptimizerObservationArtifact, type ExternalOptimizerObservationSummary, type ExternalOptimizerProcessLimits, type ExternalOptimizerRunnerCommand, type ExternalOptimizerSubmittedCandidate, type ExternalSearchOptimizer, type ExternalTextCandidate, type ExternalTextEvaluationResponse, type ExternalTextOptimizationMethodConfig, type ExternalTextOptimizerContext, type ExternalTextOptimizerResult, FileSearchLedger, type FinalEvidencePolicy, type FinalEvidenceUse, type FinalizeProfileMatrixOptions, type FinalizedProfileMatrixResult, FsLabeledScenarioStore, type FsLabeledScenarioStoreOptions, type Gate, type GateCheckStatus, type GateContext, type GateContribution, type GateDecision, type GateResult, type GenerationCandidate, type GenerationRecord, type GepaAdaptiveEngineRun, type GepaCandidatePopulationArtifact, type GepaCandidatePopulationCandidate, type GepaCandidatePopulationSummary, type GepaCandidateSelectionScore, type GepaEngineOptions, type GepaEngineRun, type GepaEvaluationImport, type GepaOptimizationMethodConfig, type GepaOptimizationRecipe, type GepaPopulationImport, type GepaRunnerCommand, type GitWorktreeAdapterOptions, type HeldOutGateOptions, type HeldoutSignificance, type HeldoutSignificanceOptions, type JudgeAggregate, type JudgeConfig, type JudgeDimension, type JudgeScore, type LabelTrust, type LabeledScenarioRecord, type LabeledScenarioSampleArgs, type LabeledScenarioSource, type LabeledScenarioStore, LabeledScenarioStoreError, type LabeledScenarioWrite, type LlmJudgeDimension, type LlmJudgeOptions, type LoadEvalFixtureScenariosOptions, type LoopProvenanceArgsFromResult, type LoopProvenanceBackend, type LoopProvenanceCandidate, type LoopProvenanceEvidence, type LoopProvenanceOptimizationMethod, type LoopProvenanceRecord, type MutableSurface, type NeutralizationGateOptions, type NodeEstimate, type NodePosterior, type ObjectiveSource, type OpenAICompatibleOptimizerModel, type OpenAutoPrOptions, type OpenAutoPrResult, type OpenSearchLedgerOptions, type OptimizationMethod, type OptimizationMethodComparison, type OptimizationMethodComposition, type OptimizationMethodInput, type OptimizationMethodPairwise, type OptimizationMethodProvenance, type OptimizationMethodResult, type OptimizationMethodRunOptions, type OptimizationMethodScore, type OptimizationPackageSource, type OptimizationTokenUsage, type OptimizerConfig, type OptimizerModelBudget, type PairedHoldout, type ParetoParent, type ParetoSignificanceGateOptions, type PendingCostCallView, type PhoenixEvaluationResultLike, type PhoenixEvaluatorLike, type PlanCampaignRunOptions, type PlanEvalFixtureRunOptions, type PlanSearchClaimInput, type PlaybackContext, type PlaybackDriver, type PlaybackStep, type PowerPreflight, type PowerPreflightOptions, type PremeasuredOptimizationBaseline, type ProfileDispatchFn, type ProfileMatrixCoverage, ProfileMatrixError, type ProfileMatrixPlan, type ProfileMatrixRow, type ProfileMatrixSegmentResult, type ProfileSummary, type PromotionObjective, type PromotionPolicy, type ProposalFinding, type ProposalFindingOrigin, type ProposalTrackContext, type ProposeContext, type ProposedCandidate, type RecordSearchEdgeInput, type RedactionStatus, type ReferenceEquivalenceJudgeOptions, type ReferenceEquivalenceScenario, type RegisterSearchNodeInput, type RetainedDispatchOptions, type RetainedDispatchOutcome, type RolloutArgumentDiff, type RolloutArgumentDiffOptions, type RolloutCall, type RunCampaignOptions, type RunEvalOptions, type RunImprovementLoopOptions, type RunImprovementLoopResult, type RunOptimizationOptions, type RunOptimizationResult, type RunProfileMatrixOptions, type RunProfileMatrixResult, type RunProfileMatrixSegmentOptions, type RunSearchOptions, SEARCH_CLAIM_RULE, SEARCH_ESTIMATOR, SEARCH_KERNEL_SOURCE, SEARCH_LEDGER_SCHEMA, type Scenario, type ScenarioAggregate, type ScenarioRollup, type ScenarioSignal, type ScoreboardRenderOptions, type ScoreboardRow, type ScoreboardSummary, type ScoredRollout, type ScoredSurfaceOutcome, type SearchAllocationView, type SearchAllocator, type SearchArtifactCodec, type SearchArtifactKind, type SearchArtifactRef, type SearchAttemptAccounting, type SearchAudit, type SearchBudget, type SearchCancelReason, type SearchCandidateSurface, type SearchCell, type SearchCellAllocatedEvent, type SearchCellCancelledEvent, type SearchCellPlan, type SearchCellResult, type SearchCellSettledEvent, type SearchCellStage, type SearchCellWork, type SearchClaim, type SearchClaimDecision, type SearchClaimEstimator, type SearchClaimFinalist, type SearchClaimPlan, type SearchClaimPower, type SearchClaimTest, type SearchClaimVerification, type SearchCloseReason, type SearchClosedEvent, type SearchCompletion, type SearchCostAccounting, type SearchDecisionRecord, type SearchEdgeAttribution, type SearchEdgeOperator, type SearchEdgeRecordedEvent, type SearchEstimateMethod, type SearchExecutionIdentity, type SearchExecutor, type SearchExpansion, type SearchFailureReason, type SearchHistoryAdmissionOptions, type SearchHistoryAuditSummary, type SearchHistoryCoverage, type SearchHistoryCoverageRow, type SearchHistoryPolicy, type SearchHistoryReceipt, SearchHistoryRequiredError, type SearchLane, type SearchLedger, type SearchLedgerAppendResult, type SearchLedgerBinding, SearchLedgerConflictError, type SearchLedgerEntry, SearchLedgerError, type SearchLedgerEvent, type SearchLedgerHash, SearchLedgerIntegrityError, type SearchLedgerTrustedHeadMode, type SearchModelIdentity, type SearchNode, type SearchNodeDecidedEvent, type SearchNodeDecision, type SearchNodeRef, type SearchNodeRegisteredEvent, type SearchNodeStatus, type SearchOpenedEvent, type SearchOpening, type SearchOperation, type SearchOperationKind, type SearchOperationRecordedEvent, type SearchOperationStartedEvent, type SearchPolicy, type SearchPolicyNode, type SearchPolicyView, type SearchPosterior, type SearchProposalBlob, type SearchProposalRequest, type SearchProposalResult, type SearchProposedChild, type SearchProposer, type SearchProposerKind, type SearchProposerPort, type SearchProposerView, SearchRecorder, type SearchRecorderOptions, type SearchReservation, type SearchRunIdentity, type SearchRunResult, type SearchRungDecision, type SearchScoredCell, type SearchSourceRef, type SearchSpend, type SearchSplit, type SearchSplitTasks, type SearchSplits, SearchState, SearchStateView, type SearchSummaryOptions, type SearchSurfaceEffect, type SearchSurfaceEvidence, type SearchSurfaceKind, type SearchTask, type SearchTaskOutcome, type SearchTokenAccounting, type SearchTraceRef, type SearchUnitExtension, type SearchUnitScore, type SearchUnknown, type SequentialDecideFn, type SequentialDecideOptions, type SequentialDecision, type SequentialObservation, type SequentialPairedGate, type SequentialPairedGateOptions, type SequentialStreamState, type SeriesDistribution, type SessionScript, type SettleSearchCellInput, type SingleRunLock, type SingleRunLockOptions, type SkillOptOptimizationMethodConfig, type SkillOptRunnerCommand, type SkillOptTrainerConfig, type SurfaceProposer, type TraceAnalystArtifact, type TraceAnalystScenario, type TraceSpan, type TransientFailureOptions, type UngroundedLiteralReport, type UserStory, type UserStoryVerdict, type Worktree, type WorktreeAdapter, WorktreeAdapterError, acquireSingleRunLock, aide, analyzeCrossSurfaceInteractions, asha, assertCampaignDesign, assertCampaignSplitIdentity, assertCodeSurfaceIdentity, assertCompleteSearchHistory, assertSearchHistoryMatchesState, autoevalsScorerJudge, beam, buildCellSchedule, buildEvidenceVector, buildLoopProvenanceRecord, buildTraceAnalystSurfaceDispatch, campaignBreakdown, campaignCellSearchResult, campaignCoverage, campaignMeanComposite, campaignMeasurementDigest, campaignScenarioIdentity, campaignSplitDigest, campaignSplitDigestFromIdentities, canonicalDigest, cellCachePath, classifyUngroundedLiterals, codeSurfaceIdentityMaterial, combineComparisonCosts, compareOptimizationMethods, componentSurfaceIdentityMaterial, composeGate, costFromLedgerSummary, createProfileMatrixPlan, createReferenceEquivalenceJudge, createRetainedDispatch, createRunCostLedger, createSearchHistoryReceipt, crowdedFrontierParent, decideSearchClaim, decodeExternalTextCandidate, defaultProductionGate, detectScale, developmentClaim, dimensionRegressions, discoverEvalFixtures, emitLoopProvenance, estimateNode, estimateNodeFromCells, externalSurface, externalTextOptimizationMethod, finalizeProfileMatrix, fsCampaignStorage, gepaOptimizationMethod, gitWorktreeAdapter, heldOutGate, heldoutSignificance, importExternalEvaluations, importGepaPopulation, inMemoryCampaignStorage, incumbent, incumbentWithOperatorBandit, isProposedCandidate, isTransientTransportFailure, labelTrustRank, llmJudge, loadEvalFixture, loadEvalFixtureScenarios, loopProvenanceArgsFromResult, loopProvenanceSpans, makePlaybackDispatch, makeProposalFinding, neutralizationGate, neutralizeText, openAutoPr, openSearchLedger, optimizationTokenUsageFromSummary, pairHoldout, paretoPolicy, paretoSignificanceGate, parseSearchLedgerLine, phoenixEvaluatorJudge, planCampaignRun, planEvalFixtureRun, planSearchClaim, powerPreflight, provenanceRecordPath, provenanceSpansPath, quotaExhaustedUntil, readCachedCell, readExternalOptimizerObservationArtifact, readGepaCandidatePopulationArtifact, recordGepaSearch, recordSkillOptSearch, renderScoreboardMarkdown, renderSearchSummary, renderSurfaceDiff, replaySearchLedgerText, resolveExternalOptimizerCallbackLimits, resolveExternalOptimizerProcessLimits, resolveRunDir, resolveWorktreePath, rolloutArgumentDiff, runCampaign, runEval, runImprovementLoop, runOptimization, runProfileMatrix, runProfileMatrixSegment, runSearch, scopedOptimizationMethod, scoreDiscrimination, scoreUserStory, scoreboardSummary, searchCellId, searchCellSetDigest, searchClaimReserveUsd, searchDivergence, searchEdgeId, searchEstimateMethod, searchExpansionIndex, searchHistoryCoverageRow, searchModelIdentity, searchNodeId, searchPolicyView, searchPosterior, searchProposalExecution, searchProposerView, searchReceiptAccounting, searchTaskSetDigest, searchUnitScores, selectDiscriminative, sequentialDecide, sequentialOptimizationMethod, sequentialPairedGate, skillOptOptimizationMethod, surfaceContentHash, surfaceDiff, surfaceDispatchRef, surfaceHash, surfaceNode, tangleTracesRoot, traceAnalystQualityJudge, transientDispatchFailure, uniform, userStoryScoreboard, validateSearchLedgerEvent, verifyCodeSurface, verifyLoopProvenanceRecord, verifySearchClaim, verifySearchHistoryArtifact, verifySearchHistoryReceipt };

package/dist/campaign/index.js

@@ -3,7 +3,7 @@ import { A as assertCompleteSearchHistory, B as searchReceiptAccounting, C as in
 import { D as powerPreflight, F as surfaceContentHash, G as campaignCoverage, H as assertCampaignDesign, I as surfaceDispatchRef, J as campaignSplitDigestFromIdentities, K as campaignScenarioIdentity, L as surfaceHash, M as codeSurfaceIdentityMaterial, N as componentSurfaceIdentityMaterial, P as renderSurfaceDiff, U as assertCampaignSplitIdentity, _ as provenanceRecordPath, b as campaignBreakdown, d as buildLoopProvenanceRecord, f as campaignMeasurementDigest, g as loopProvenanceSpans, h as loopProvenanceArgsFromResult, j as assertCodeSurfaceIdentity, m as emitLoopProvenance, p as canonicalDigest, q as campaignSplitDigest, v as provenanceSpansPath, x as campaignMeanComposite, y as verifyLoopProvenanceRecord } from "../campaign-evidence-DW3Krxov.js";
 import { a as SearchLedgerError, i as SearchLedgerConflictError, o as SearchLedgerIntegrityError } from "../search-ledger-file-DY0AXX3W.js";
 import { C as inMemoryCampaignStorage, S as fsCampaignStorage, l as DEFAULT_EXTERNAL_OPTIMIZER_CALLBACK_LIMITS, u as DEFAULT_EXTERNAL_OPTIMIZER_PROCESS_LIMITS, v as resolveExternalOptimizerCallbackLimits, x as createRunCostLedger, y as resolveExternalOptimizerProcessLimits } from "../external-optimizer-subprocess-BmKtH3T6.js";
-import { S as searchCellSetDigest, a as searchPolicyView, b as estimateNode, c as planSearchClaim, i as searchExpansionIndex, l as searchClaimReserveUsd, n as runSearch, o as SEARCH_CLAIM_RULE, r as searchDivergence, s as decideSearchClaim, t as SEARCH_KERNEL_SOURCE, u as verifySearchClaim, w as searchPosterior, x as estimateNodeFromCells, y as SEARCH_ESTIMATOR } from "../search-kernel-mzeNuzr8.js";
+import { C as searchEstimateMethod, S as searchCellSetDigest, a as searchPolicyView, b as estimateNode, c as planSearchClaim, i as searchExpansionIndex, l as searchClaimReserveUsd, n as runSearch, o as SEARCH_CLAIM_RULE, r as searchDivergence, s as decideSearchClaim, t as SEARCH_KERNEL_SOURCE, u as verifySearchClaim, w as searchPosterior, x as estimateNodeFromCells, y as SEARCH_ESTIMATOR } from "../search-kernel-mzeNuzr8.js";
 import { a as validateSearchLedgerEvent, c as SearchStateView, d as searchTaskSetDigest, f as searchUnitScores, i as replaySearchLedgerText, l as searchCellId, n as openSearchLedger, o as SEARCH_LEDGER_SCHEMA, r as parseSearchLedgerLine, s as SearchState, t as FileSearchLedger, u as searchNodeId } from "../search-ledger-CXQPwdWz.js";
 import { t as acquireSingleRunLock } from "../single-run-lock-fVu3MZWH.js";
 import { n as searchProposerView, t as renderSearchSummary } from "../search-summary-CDURlgZR.js";
@@ -12,4 +12,4 @@ import { a as importGepaPopulation, c as heldOutGate, d as decodeExternalTextCan
 import { A as loadEvalFixtureScenarios, C as FsLabeledScenarioStore, D as neutralizationGate, E as rolloutArgumentDiff, M as analyzeCrossSurfaceInteractions, N as buildTraceAnalystSurfaceDispatch, O as discoverEvalFixtures, P as traceAnalystQualityJudge, S as neutralizeText, T as classifyUngroundedLiterals, _ as scoreUserStory, a as autoevalsScorerJudge, b as scopedOptimizationMethod, c as selectDiscriminative, d as finalizeProfileMatrix, f as runProfileMatrixSegment, g as renderScoreboardMarkdown, h as makePlaybackDispatch, i as verifyCodeSurface, j as planEvalFixtureRun, k as loadEvalFixture, l as createRetainedDispatch, m as runProfileMatrix, n as gitWorktreeAdapter, o as phoenixEvaluatorJudge, p as ProfileMatrixError, r as resolveWorktreePath, s as scoreDiscrimination, t as WorktreeAdapterError, u as createProfileMatrixPlan, v as scoreboardSummary, w as LabeledScenarioStoreError, x as sequentialOptimizationMethod, y as userStoryScoreboard } from "../campaign-BL591dWw.js";
 import { n as paretoPolicy, r as paretoSignificanceGate, t as buildEvidenceVector } from "../promotion-policy-CCmyEH_c.js";
 import { n as sequentialPairedGate, t as sequentialDecide } from "../sequential-DMqr3sni.js";
-export { DEFAULT_EXTERNAL_OPTIMIZER_CALLBACK_LIMITS, DEFAULT_EXTERNAL_OPTIMIZER_PROCESS_LIMITS, FileSearchLedger, FsLabeledScenarioStore, LabeledScenarioStoreError, ProfileMatrixError, SEARCH_CLAIM_RULE, SEARCH_ESTIMATOR, SEARCH_KERNEL_SOURCE, SEARCH_LEDGER_SCHEMA, SearchHistoryRequiredError, SearchLedgerConflictError, SearchLedgerError, SearchLedgerIntegrityError, SearchRecorder, SearchState, SearchStateView, WorktreeAdapterError, acquireSingleRunLock, aide, analyzeCrossSurfaceInteractions, asha, assertCampaignDesign, assertCampaignSplitIdentity, assertCodeSurfaceIdentity, assertCompleteSearchHistory, assertSearchHistoryMatchesState, autoevalsScorerJudge, beam, buildCellSchedule, buildEvidenceVector, buildLoopProvenanceRecord, buildTraceAnalystSurfaceDispatch, campaignBreakdown, campaignCellSearchResult, campaignCoverage, campaignMeanComposite, campaignMeasurementDigest, campaignScenarioIdentity, campaignSplitDigest, campaignSplitDigestFromIdentities, canonicalDigest, cellCachePath, classifyUngroundedLiterals, codeSurfaceIdentityMaterial, combineComparisonCosts, compareOptimizationMethods, componentSurfaceIdentityMaterial, composeGate, costFromLedgerSummary, createProfileMatrixPlan, createReferenceEquivalenceJudge, createRetainedDispatch, createRunCostLedger, createSearchHistoryReceipt, crowdedFrontierParent, decideSearchClaim, decodeExternalTextCandidate, defaultProductionGate, detectScale, developmentClaim, dimensionRegressions, discoverEvalFixtures, emitLoopProvenance, estimateNode, estimateNodeFromCells, externalSurface, externalTextOptimizationMethod, finalizeProfileMatrix, fsCampaignStorage, gepaOptimizationMethod, gitWorktreeAdapter, heldOutGate, heldoutSignificance, importExternalEvaluations, importGepaPopulation, inMemoryCampaignStorage, incumbent, incumbentWithOperatorBandit, isProposedCandidate, isTransientTransportFailure, labelTrustRank, llmJudge, loadEvalFixture, loadEvalFixtureScenarios, loopProvenanceArgsFromResult, loopProvenanceSpans, makePlaybackDispatch, makeProposalFinding, neutralizationGate, neutralizeText, openAutoPr, openSearchLedger, optimizationTokenUsageFromSummary, pairHoldout, paretoPolicy, paretoSignificanceGate, parseSearchLedgerLine, phoenixEvaluatorJudge, planCampaignRun, planEvalFixtureRun, planSearchClaim, powerPreflight, provenanceRecordPath, provenanceSpansPath, quotaExhaustedUntil, readCachedCell, readExternalOptimizerObservationArtifact, readGepaCandidatePopulationArtifact, recordGepaSearch, recordSkillOptSearch, renderScoreboardMarkdown, renderSearchSummary, renderSurfaceDiff, replaySearchLedgerText, resolveExternalOptimizerCallbackLimits, resolveExternalOptimizerProcessLimits, resolveRunDir, resolveWorktreePath, rolloutArgumentDiff, runCampaign, runEval, runImprovementLoop, runOptimization, runProfileMatrix, runProfileMatrixSegment, runSearch, scopedOptimizationMethod, scoreDiscrimination, scoreUserStory, scoreboardSummary, searchCellId, searchCellSetDigest, searchClaimReserveUsd, searchDivergence, searchEdgeId, searchExpansionIndex, searchHistoryCoverageRow, searchModelIdentity, searchNodeId, searchPolicyView, searchPosterior, searchProposalExecution, searchProposerView, searchReceiptAccounting, searchTaskSetDigest, searchUnitScores, selectDiscriminative, sequentialDecide, sequentialOptimizationMethod, sequentialPairedGate, skillOptOptimizationMethod, surfaceContentHash, surfaceDiff, surfaceDispatchRef, surfaceHash, surfaceNode, tangleTracesRoot, traceAnalystQualityJudge, transientDispatchFailure, uniform, userStoryScoreboard, validateSearchLedgerEvent, verifyCodeSurface, verifyLoopProvenanceRecord, verifySearchClaim, verifySearchHistoryArtifact, verifySearchHistoryReceipt };
+export { DEFAULT_EXTERNAL_OPTIMIZER_CALLBACK_LIMITS, DEFAULT_EXTERNAL_OPTIMIZER_PROCESS_LIMITS, FileSearchLedger, FsLabeledScenarioStore, LabeledScenarioStoreError, ProfileMatrixError, SEARCH_CLAIM_RULE, SEARCH_ESTIMATOR, SEARCH_KERNEL_SOURCE, SEARCH_LEDGER_SCHEMA, SearchHistoryRequiredError, SearchLedgerConflictError, SearchLedgerError, SearchLedgerIntegrityError, SearchRecorder, SearchState, SearchStateView, WorktreeAdapterError, acquireSingleRunLock, aide, analyzeCrossSurfaceInteractions, asha, assertCampaignDesign, assertCampaignSplitIdentity, assertCodeSurfaceIdentity, assertCompleteSearchHistory, assertSearchHistoryMatchesState, autoevalsScorerJudge, beam, buildCellSchedule, buildEvidenceVector, buildLoopProvenanceRecord, buildTraceAnalystSurfaceDispatch, campaignBreakdown, campaignCellSearchResult, campaignCoverage, campaignMeanComposite, campaignMeasurementDigest, campaignScenarioIdentity, campaignSplitDigest, campaignSplitDigestFromIdentities, canonicalDigest, cellCachePath, classifyUngroundedLiterals, codeSurfaceIdentityMaterial, combineComparisonCosts, compareOptimizationMethods, componentSurfaceIdentityMaterial, composeGate, costFromLedgerSummary, createProfileMatrixPlan, createReferenceEquivalenceJudge, createRetainedDispatch, createRunCostLedger, createSearchHistoryReceipt, crowdedFrontierParent, decideSearchClaim, decodeExternalTextCandidate, defaultProductionGate, detectScale, developmentClaim, dimensionRegressions, discoverEvalFixtures, emitLoopProvenance, estimateNode, estimateNodeFromCells, externalSurface, externalTextOptimizationMethod, finalizeProfileMatrix, fsCampaignStorage, gepaOptimizationMethod, gitWorktreeAdapter, heldOutGate, heldoutSignificance, importExternalEvaluations, importGepaPopulation, inMemoryCampaignStorage, incumbent, incumbentWithOperatorBandit, isProposedCandidate, isTransientTransportFailure, labelTrustRank, llmJudge, loadEvalFixture, loadEvalFixtureScenarios, loopProvenanceArgsFromResult, loopProvenanceSpans, makePlaybackDispatch, makeProposalFinding, neutralizationGate, neutralizeText, openAutoPr, openSearchLedger, optimizationTokenUsageFromSummary, pairHoldout, paretoPolicy, paretoSignificanceGate, parseSearchLedgerLine, phoenixEvaluatorJudge, planCampaignRun, planEvalFixtureRun, planSearchClaim, powerPreflight, provenanceRecordPath, provenanceSpansPath, quotaExhaustedUntil, readCachedCell, readExternalOptimizerObservationArtifact, readGepaCandidatePopulationArtifact, recordGepaSearch, recordSkillOptSearch, renderScoreboardMarkdown, renderSearchSummary, renderSurfaceDiff, replaySearchLedgerText, resolveExternalOptimizerCallbackLimits, resolveExternalOptimizerProcessLimits, resolveRunDir, resolveWorktreePath, rolloutArgumentDiff, runCampaign, runEval, runImprovementLoop, runOptimization, runProfileMatrix, runProfileMatrixSegment, runSearch, scopedOptimizationMethod, scoreDiscrimination, scoreUserStory, scoreboardSummary, searchCellId, searchCellSetDigest, searchClaimReserveUsd, searchDivergence, searchEdgeId, searchEstimateMethod, searchExpansionIndex, searchHistoryCoverageRow, searchModelIdentity, searchNodeId, searchPolicyView, searchPosterior, searchProposalExecution, searchProposerView, searchReceiptAccounting, searchTaskSetDigest, searchUnitScores, selectDiscriminative, sequentialDecide, sequentialOptimizationMethod, sequentialPairedGate, skillOptOptimizationMethod, surfaceContentHash, surfaceDiff, surfaceDispatchRef, surfaceHash, surfaceNode, tangleTracesRoot, traceAnalystQualityJudge, transientDispatchFailure, uniform, userStoryScoreboard, validateSearchLedgerEvent, verifyCodeSurface, verifyLoopProvenanceRecord, verifySearchClaim, verifySearchHistoryArtifact, verifySearchHistoryReceipt };

package/dist/cli.js

@@ -5,9 +5,9 @@ import { t as renderSearchSummary } from "./search-summary-CDURlgZR.js";
 import { c as hostedTenantFromEnv, g as SearchRunKindSchema, r as shipSearchLedger, t as SearchShipConflictError, u as SEARCH_LEDGER_BATCH_MAX_BYTES } from "./search-shipper-DhhofkDX.js";
 import { i as runRolloutReleaseCli } from "./hf-dataset-BR2yhuNn.js";
 import { t as createChatClient } from "./chat-client-5LLl4Ctl.js";
-import { t as runAnalystBenchmarkCommand } from "./benchmark-command-DuDTN5ID.js";
+import { t as runAnalystBenchmarkCommand } from "./benchmark-command-DtP94cOB.js";
 import { a as runRpcBatch, o as runRpcOnce, p as handleVersion, r as startServerAsync, s as buildOpenApi } from "./server-CoBZO1fj.js";
-import { A as editCreditText, E as front, f as renderMetaSearchText, h as landscape, k as editCredit, l as metaSearch, m as formatLandscape, n as taskMatrix, o as skillManifold, r as formatSkillManifold, t as tree, x as surfaceTextEdits } from "./tree-By7LtLUu.js";
+import { A as editCreditText, E as front, f as renderMetaSearchText, h as landscape, k as editCredit, l as metaSearch, m as formatLandscape, n as taskMatrix, o as skillManifold, r as formatSkillManifold, t as tree, x as surfaceTextEdits } from "./tree-CN8mlzUO.js";
 import { t as runSupervisorRunCommand } from "./report-command-JrjiWReS.js";
 import { createHash } from "node:crypto";
 import { existsSync, readFileSync, writeFileSync } from "node:fs";
@@ -391,6 +391,7 @@ function renderFrontText(data) {
 }
 function renderTaskMatrixText(data) {
 	const lines = [`task matrix (${data.split} split, ${data.direction}): ${data.nodeIds.length} node(s) × ${data.unitIds.length} unit(s), ${data.nodeClusters.length} node cluster(s), ${data.unitClusters.length} unit cluster(s)`];
+	if (data.sampling) lines.push(`  clustered ${data.sampling.nodesClustered} of ${data.sampling.nodesTotal} nodes (cap ${data.sampling.cap}), evenly sampled in registration order; unit clusters and specialist gains use the same nodes`);
 	for (const row of data.specialistGain) {
 		const gain = row.gain === null ? `insufficient (${row.insufficient})` : `${row.gain.toFixed(4)} (best ${row.bestNodeId}, ${row.nodesContributing} nodes on ${row.unitIds.length} units)`;
 		lines.push(`  cluster ${row.clusterId} [${row.unitIds.join(", ")}]: specialist gain ${gain}`);

package/dist/cli.js.map

@@ -1 +1 @@
-{"version":3,"file":"cli.js","names":[],"sources":["../src/cli-config.ts","../src/search-command.ts","../src/cli.ts"],"sourcesContent":["/**\n * Provider configuration for the `agent-eval` binary.\n *\n * The binary reads endpoint and credential values from its environment.\n * TypeScript library callers bind a transport explicitly. The maintained\n * HTTP transport accepts their credential and sends provider requests.\n */\n\nimport { type ChatClient, createChatClient } from './analyst/chat-client'\n\nexport interface CliLlmConfig {\n  /** Judge transport for the wire handlers. Absent when no provider is configured. */\n  chat?: ChatClient\n  model?: string\n}\n\n/** The endpoint the binary resolved from its environment. */\nexport interface CliProviderRoute {\n  baseUrl: string\n  apiKey: string\n  model?: string\n}\n\n/**\n * Environment precedence for the binary's provider route, with no client built.\n *\n * Both halves are required: a base URL without a key, or a key without an\n * endpoint, is a half-configured server. `/v1/judge` then refuses with\n * `llm_not_configured` instead of calling an unintended endpoint.\n */\nexport function resolveCliProviderRoute(\n  env: NodeJS.ProcessEnv = process.env,\n): CliProviderRoute | undefined {\n  const explicitBaseUrl = nonEmpty(env.AGENT_EVAL_LLM_BASE_URL)\n  const explicitApiKey = nonEmpty(env.AGENT_EVAL_LLM_API_KEY)\n  const openAiApiKey = nonEmpty(env.OPENAI_API_KEY)\n  const tangleApiKey = nonEmpty(env.TANGLE_API_KEY)\n  const baseUrl =\n    explicitBaseUrl ??\n    nonEmpty(env.OPENAI_BASE_URL) ??\n    nonEmpty(env.TANGLE_ROUTER_URL) ??\n    (openAiApiKey ? 'https://api.openai.com/v1' : undefined) ??\n    (tangleApiKey ? 'https://router.tangle.tools/v1' : undefined)\n  const apiKey = explicitApiKey ?? openAiApiKey ?? tangleApiKey\n  const model =\n    nonEmpty(env.AGENT_EVAL_LLM_MODEL) ?? nonEmpty(env.OPENAI_MODEL) ?? nonEmpty(env.TANGLE_MODEL)\n\n  if (!baseUrl || !apiKey) return undefined\n  return { baseUrl, apiKey, ...(model ? { model } : {}) }\n}\n\nexport function resolveCliLlmConfig(env: NodeJS.ProcessEnv = process.env): CliLlmConfig {\n  const route = resolveCliProviderRoute(env)\n  if (!route) return {}\n  return {\n    chat: cliChatClient({ baseUrl: route.baseUrl, apiKey: route.apiKey }, route.model),\n    ...(route.model ? { model: route.model } : {}),\n  }\n}\n\n/**\n * The binary's transport IS the published one — `createChatClient({ transport:\n * 'openai-compatible' })` — so the server and a library consumer cannot drift\n * onto two different clients for the same endpoint.\n *\n * The one thing kept local is the error text. A generic \"no model on the\n * request\" tells a server operator nothing; this names the variable to set.\n */\nfunction cliChatClient(\n  route: { baseUrl: string; apiKey: string },\n  defaultModel: string | undefined,\n): ChatClient {\n  const client = createChatClient({\n    transport: 'openai-compatible',\n    baseUrl: route.baseUrl,\n    apiKey: route.apiKey,\n    ...(defaultModel ? { defaultModel } : {}),\n  })\n  return {\n    ...client,\n    chat: (req, callOpts) => {\n      if (!req.model && !defaultModel) {\n        throw new Error(\n          'agent-eval: no model on the request and no AGENT_EVAL_LLM_MODEL configured',\n        )\n      }\n      return client.chat(req, callOpts)\n    },\n  }\n}\n\nfunction nonEmpty(value: string | undefined): string | undefined {\n  const trimmed = value?.trim()\n  return trimmed ? trimmed : undefined\n}\n","/**\n * `agent-eval search <subcommand>`: work with a search ledger from a terminal.\n *\n *   agent-eval search ship <search-ledger.jsonl> --run-kind optimization|eval [--content full|digests]\n *   agent-eval search show <search-ledger.jsonl> [--tree] [--operator-yield] [--front] [--task-matrix]\n *                          [--edit-credit] [--landscape] [--skill-manifold] [--json]\n *   agent-eval search show <search-ledger.jsonl> [<search-ledger.jsonl> ...] --meta [--objective <key>] [--json]\n *\n * `ship` sends the ledger to the hosted store named by `TANGLE_INGEST_URL`,\n * `TANGLE_INGEST_API_KEY` and `TANGLE_TENANT_ID`, starting from the store's\n * head, so it finishes or resumes a ship that a loop could not complete.\n *\n * `show` verifies the ledger and prints `renderSearchSummary`: the same\n * compact text a proposer reads, on the search's own ranking split (the\n * selection split when the search declares one, else train). There is no\n * local HTML renderer; the hosted store's page is the visual view (§5 of the\n * search-tree design). A lens flag adds that lens's text form after the\n * summary, so an agent reads what a person sees; `--json` prints the lens\n * results as JSON instead, the shape a view renders.\n *\n * `show --meta` reads every ledger it is given and prints the `metaSearch`\n * lens over them instead of one search's summary: each search as one node\n * scored by its held-out lift per known dollar, the configurations best first,\n * and the searches arranged by derivation and containment.\n */\n\nimport { createHash } from 'node:crypto'\nimport { existsSync, readFileSync } from 'node:fs'\nimport { open, readFile } from 'node:fs/promises'\nimport { dirname, join } from 'node:path'\nimport { fileURLToPath } from 'node:url'\nimport { replaySearchLedgerText } from './campaign/search-ledger'\nimport type { SearchArtifactRef } from './campaign/search-ledger-types'\nimport type { SearchStateView } from './campaign/search-state'\nimport { renderSearchSummary } from './campaign/search-summary'\nimport { hostedTenantFromEnv } from './hosted/client'\nimport { SEARCH_LEDGER_BATCH_MAX_BYTES, SearchRunKindSchema } from './hosted/search-ledger-wire'\nimport { SearchShipConflictError, shipSearchLedger } from './hosted/search-shipper'\nimport { editCredit, editCreditText } from './search/lenses/edit-credit'\nimport { type FrontData, front } from './search/lenses/front'\nimport { formatLandscape, landscape, surfaceTextEdits } from './search/lenses/landscape'\nimport { metaSearch, renderMetaSearchText } from './search/lenses/meta-search'\nimport { type OperatorYieldData, operatorYield } from './search/lenses/operator-yield'\nimport { INSUFFICIENT_FROM } from './search/lenses/shared'\nimport { formatSkillManifold, skillManifold } from './search/lenses/skill-manifold'\nimport { type TaskMatrixData, taskMatrix } from './search/lenses/task-matrix'\nimport { type TreeData, type TreeNode, tree } from './search/lenses/tree'\n\nconst USAGE = `usage: agent-eval search <subcommand> ...\n\n  ship <search-ledger.jsonl> --run-kind optimization|eval [--content full|digests]\n        Ships the ledger's entries, and the blobs they name, to the hosted store in\n        TANGLE_INGEST_URL (or TANGLE_ORCHESTRATOR_URL) as tenant TANGLE_TENANT_ID with\n        TANGLE_INGEST_API_KEY (or TANGLE_API_KEY). It starts from the store's head, so\n        running it again sends only what the store lacks. Prints the result as JSON.\n        Exits 1 when the store holds a different chain for the search.\n\n  show <search-ledger.jsonl> [--tree] [--operator-yield] [--front] [--task-matrix]\n                              [--edit-credit] [--landscape] [--skill-manifold] [--json]\n        Verifies the ledger and prints its search summary: the leading nodes\n        against the root, the most recently discarded nodes and why, and a log\n        of recent proposals. The same text a proposer reads as context, on the\n        search's own ranking split. Each lens flag adds that lens's text form\n        (search-tree-design §12) below the summary — the same JSON\n        Intelligence, discovery lab, VerticalBench and agent-runtime read.\n        --tree           the tidy tree of nodes and edges.\n        --operator-yield outcome counts and yield per known dollar, by edge operator.\n        --front          the Pareto front over score and known cost per cell.\n        --task-matrix    nodes × units, clustered, with specialist gain per cluster.\n        --edit-credit    every edit as a gene followed down the lineage, its\n                         credit, interacting pairs, and skill candidates. It\n                         reads node artifacts from the blobs the ledger names,\n                         verified by digest.\n        --landscape      nodes placed by line edits between their profile\n                         surfaces (read from the blobs, verified by digest), the\n                         kriged score surface, its basins, and the plateau score\n                         a policy drafts on.\n        --skill-manifold the node × unit score matrix factored into skill axes,\n                         its cross-validated intrinsic dimension, and the unit\n                         that best separates the leaders.\n        --json           prints the requested lenses as JSON instead of text.\n\n  show <search-ledger.jsonl> [<search-ledger.jsonl> ...] --meta [--objective <key>] [--json]\n        Verifies every ledger and prints the meta-search lens over them: each\n        search as one node, scored by its claim's held-out lift per known\n        dollar (unscored searches say why), configurations best first with\n        their interval, method and n, and the searches arranged by derivation\n        and containment. --objective names the objective the best\n        configuration is chosen within when the searches span several.`\n\nexport async function runSearchCommand(argv: string[]): Promise<number> {\n  const [subcommand, ...rest] = argv\n  if (subcommand === undefined || subcommand === '--help' || subcommand === '-h') {\n    process.stdout.write(`${USAGE}\\n`)\n    return 0\n  }\n  if (subcommand === 'show') return await runShowCommand(rest)\n  if (subcommand !== 'ship') {\n    process.stderr.write(`unknown search subcommand: ${subcommand}\\n${USAGE}\\n`)\n    return 1\n  }\n  if (rest.some((arg) => arg === undefined || arg === '--help' || arg === '-h')) {\n    process.stdout.write(`${USAGE}\\n`)\n    return 0\n  }\n  const { path, flags } = parseShipArgs(rest)\n  const runKind = SearchRunKindSchema.safeParse(flags['run-kind'])\n  if (!runKind.success) throw new Error(`--run-kind must be optimization or eval\\n${USAGE}`)\n  const content = flags.content ?? 'full'\n  if (content !== 'full' && content !== 'digests') {\n    throw new Error(`--content must be full or digests\\n${USAGE}`)\n  }\n  const tenant = hostedTenantFromEnv()\n  if (!tenant) {\n    throw new Error(\n      'set TANGLE_INGEST_URL, TANGLE_INGEST_API_KEY and TANGLE_TENANT_ID to name the store',\n    )\n  }\n  const searchId = await firstLineSearchId(path)\n  try {\n    const shipped = await shipSearchLedger({\n      tenant,\n      ledger: { path, searchId },\n      runKind: runKind.data,\n      content,\n    })\n    process.stdout.write(`${JSON.stringify(shipped, null, 2)}\\n`)\n    return shipped.head.nextSequence === shipped.localLines ? 0 : 1\n  } catch (error) {\n    if (!(error instanceof SearchShipConflictError)) throw error\n    process.stderr.write(`${error.message}\\n`)\n    return 1\n  }\n}\n\nasync function runShowCommand(argv: string[]): Promise<number> {\n  if (argv.includes('--help') || argv.includes('-h')) {\n    process.stdout.write(`${USAGE}\\n`)\n    return 0\n  }\n  if (argv.includes('--meta')) return await runShowMetaCommand(argv)\n  const positional = argv.filter((arg) => !arg.startsWith('--'))\n  const flags = new Set(argv.filter((arg) => arg.startsWith('--')))\n  const unknown = [...flags].filter((flag) => !SHOW_FLAGS.has(flag))\n  if (positional.length !== 1 || unknown.length > 0) {\n    if (unknown.length > 0) process.stderr.write(`unknown flag ${unknown.join(', ')}\\n`)\n    process.stdout.write(`${USAGE}\\n`)\n    return 1\n  }\n  const path = positional[0]!\n  const searchId = await firstLineSearchId(path)\n  const text = await readFile(path, 'utf8')\n  const state = replaySearchLedgerText(text, searchId, path)\n  const split =\n    state.header && state.header.splits.selection.tasks.length > 0 ? 'selection' : 'train'\n  const lenses = showLenses(state, path, flags)\n  if (flags.has('--json')) {\n    if (lenses.length === 0) {\n      process.stderr.write('--json needs a lens flag\\n')\n      return 1\n    }\n    const json = Object.fromEntries(lenses.map((lens) => [lens.name, lens.result]))\n    process.stdout.write(`${JSON.stringify(json, null, 2)}\\n`)\n    return 0\n  }\n  const sections = [renderSearchSummary(state, { split }), ...lenses.map((lens) => lens.text)]\n  process.stdout.write(`${sections.join('\\n\\n')}\\n`)\n  return 0\n}\n\nasync function runShowMetaCommand(argv: string[]): Promise<number> {\n  const paths: string[] = []\n  let objective: string | undefined\n  let json = false\n  for (let index = 0; index < argv.length; index++) {\n    const token = argv[index]!\n    if (token === '--meta') continue\n    if (token === '--json') {\n      json = true\n      continue\n    }\n    if (token === '--objective') {\n      objective = argv[++index]\n      if (objective === undefined) throw new Error(`--objective needs a value\\n${USAGE}`)\n      continue\n    }\n    if (token.startsWith('--')) throw new Error(`unknown flag ${token}\\n${USAGE}`)\n    paths.push(token)\n  }\n  if (paths.length === 0) throw new Error(`show --meta needs at least one ledger\\n${USAGE}`)\n  const states = []\n  for (const path of paths) {\n    const searchId = await firstLineSearchId(path)\n    states.push(replaySearchLedgerText(await readFile(path, 'utf8'), searchId, path))\n  }\n  const lens = metaSearch(states, objective === undefined ? {} : { objective })\n  process.stdout.write(\n    json ? `${JSON.stringify({ metaSearch: lens }, null, 2)}\\n` : `${renderMetaSearchText(lens)}\\n`,\n  )\n  return 0\n}\n\nconst SHOW_FLAGS = new Set([\n  '--edit-credit',\n  '--json',\n  '--tree',\n  '--operator-yield',\n  '--front',\n  '--task-matrix',\n  '--landscape',\n  '--skill-manifold',\n])\n\nfunction showLenses(\n  state: SearchStateView,\n  ledgerPath: string,\n  flags: ReadonlySet<string>,\n): Array<{ name: string; result: unknown; text: string }> {\n  const lenses: Array<{ name: string; result: unknown; text: string }> = []\n  if (flags.has('--tree')) {\n    const result = tree(state)\n    lenses.push({ name: 'tree', result, text: renderTreeText(result.data) })\n  }\n  if (flags.has('--operator-yield')) {\n    const result = operatorYield(state)\n    lenses.push({ name: 'operatorYield', result, text: renderOperatorYieldText(result.data) })\n  }\n  if (flags.has('--front')) {\n    const result = front(state)\n    lenses.push({ name: 'front', result, text: renderFrontText(result.data) })\n  }\n  if (flags.has('--task-matrix')) {\n    const result = taskMatrix(state)\n    lenses.push({ name: 'taskMatrix', result, text: renderTaskMatrixText(result.data) })\n  }\n  if (flags.has('--edit-credit')) {\n    const result = editCredit(state, { readArtifact: ledgerBlobReader(ledgerPath) })\n    lenses.push({ name: 'editCredit', result, text: editCreditText(result) })\n  }\n  if (flags.has('--landscape')) {\n    const read = ledgerBlobReader(ledgerPath)\n    const result = landscape(\n      state,\n      surfaceTextEdits((surface) => read(surface.artifact) ?? null),\n    )\n    lenses.push({ name: 'landscape', result, text: formatLandscape(result) })\n  }\n  if (flags.has('--skill-manifold')) {\n    const result = skillManifold(state)\n    lenses.push({ name: 'skillManifold', result, text: formatSkillManifold(result) })\n  }\n  return lenses\n}\n\n/**\n * Reads a blob a ledger names: at its `file:` URI, else under `blobs/` beside\n * the ledger, where `SearchRecorder` writes them. Bytes that do not hash to the\n * reference's digest and length read as unavailable, never as content.\n */\nfunction ledgerBlobReader(ledgerPath: string): (ref: SearchArtifactRef) => unknown | undefined {\n  const cache = new Map<string, unknown>()\n  return (ref) => {\n    if (cache.has(ref.sha256)) return cache.get(ref.sha256)\n    const hex = ref.sha256.slice('sha256:'.length)\n    const candidates = [join(dirname(ledgerPath), 'blobs', `${hex}.json`)]\n    if (ref.uri.startsWith('file:')) candidates.unshift(fileURLToPath(ref.uri))\n    let value: unknown\n    for (const candidate of candidates) {\n      if (!existsSync(candidate)) continue\n      const bytes = readFileSync(candidate)\n      const digest = createHash('sha256').update(bytes).digest('hex')\n      if (digest !== hex || bytes.byteLength !== ref.byteLength) continue\n      try {\n        value = JSON.parse(bytes.toString('utf8'))\n      } catch {\n        // Verified bytes that are not JSON are not content a lens can read.\n      }\n      break\n    }\n    cache.set(ref.sha256, value)\n    return value\n  }\n}\n\nfunction renderTreeText(data: TreeData): string {\n  const lines = [\n    `tree: ${data.nodeCount} node${data.nodeCount === 1 ? '' : 's'}, ${data.edgeCount} edge${data.edgeCount === 1 ? '' : 's'}`,\n  ]\n  const walk = (node: TreeNode, prefix: string, isLast: boolean): void => {\n    const branch = prefix === '' ? '' : isLast ? '└─ ' : '├─ '\n    const op = node.operator ? `${node.operator} → ` : ''\n    const rung = node.rung === null ? '' : ` rung ${node.rung}`\n    const depth = node.depth === null ? ' depth unknown' : ''\n    const cost =\n      node.unknownCostCells === 0\n        ? `$${node.knownCostUsd.toFixed(2)}`\n        : `$${node.knownCostUsd.toFixed(2)} known + at least $${node.floorUsd.toFixed(2)} over ${node.unknownCostCells} unknown-cost cell${node.unknownCostCells === 1 ? '' : 's'}`\n    lines.push(\n      `${prefix}${branch}${op}${node.nodeId} (${node.status ?? 'undecided'}${rung}${depth}) ${cost}`,\n    )\n    const childPrefix = prefix + (prefix === '' ? '' : isLast ? '   ' : '│  ')\n    node.children.forEach((child, index) => {\n      walk(child, childPrefix, index === node.children.length - 1)\n    })\n  }\n  data.roots.forEach((root, index) => {\n    walk(root, '', index === data.roots.length - 1)\n  })\n  return lines.join('\\n')\n}\n\nfunction renderOperatorYieldText(data: OperatorYieldData): string {\n  const lines = [`operator yield (${data.split} split, gain per known $ of proposal + screen):`]\n  if (data.rows.length === 0) lines.push('  no edges yet')\n  for (const row of data.rows) {\n    const y = row.yield\n    const outcomes = Object.entries(row.outcomes)\n      .filter(([, count]) => count > 0)\n      .map(([status, count]) => `${status}=${count}`)\n      .join(' ')\n    const yieldText =\n      y.method === 'none'\n        ? y.n === 0\n          ? 'yield: no measured children'\n          : `yield unknown (${y.n} of ${INSUFFICIENT_FROM}, none)`\n        : y.method === 'insufficient'\n          ? `yield≈${y.mean!.toFixed(4)}/$ (${y.n} of ${INSUFFICIENT_FROM}, insufficient)`\n          : `yield≈${y.mean!.toFixed(4)}/$ [${y.interval![0].toFixed(4)}, ${y.interval![1].toFixed(4)}] (${y.method}, ${y.intervalMethod}, n=${y.n})`\n    const nodes = Object.values(row.outcomes).reduce((sum, count) => sum + count, 0)\n    const reproposals = row.reproposals > 0 ? `, ${row.reproposals} re-proposals` : ''\n    lines.push(\n      `  ${row.operator}: ${row.proposals} proposals${reproposals} → ${nodes} node${nodes === 1 ? '' : 's'} (${outcomes || 'no decisions yet'}); ${yieldText}; ${row.excluded} excluded`,\n    )\n  }\n  return lines.join('\\n')\n}\n\nfunction renderFrontText(data: FrontData): string {\n  const onFront = data.rows.filter((row) => row.onFront)\n  const reasons = Object.entries(data.exclusions)\n    .filter(([, count]) => count > 0)\n    .map(([reason, count]) => `${reason} ${count}`)\n    .join(', ')\n  const lines = [\n    `front (${data.split} split, axes: ${data.axes.join(', ')}): ${data.frontierSize} of ${data.rows.length} node(s) on the frontier, ${data.excludedCount} excluded${reasons ? ` (${reasons})` : ''}`,\n  ]\n  for (const row of onFront) {\n    lines.push(\n      `  ${row.nodeId}: score=${row.score!.toFixed(4)} on ${row.units} units (${row.method}) costPerCell=$${row.costPerCellUsd!.toFixed(4)}`,\n    )\n  }\n  return lines.join('\\n')\n}\n\nfunction renderTaskMatrixText(data: TaskMatrixData): string {\n  const lines = [\n    `task matrix (${data.split} split, ${data.direction}): ${data.nodeIds.length} node(s) × ${data.unitIds.length} unit(s), ${data.nodeClusters.length} node cluster(s), ${data.unitClusters.length} unit cluster(s)`,\n  ]\n  for (const row of data.specialistGain) {\n    const gain =\n      row.gain === null\n        ? `insufficient (${row.insufficient})`\n        : `${row.gain.toFixed(4)} (best ${row.bestNodeId}, ${row.nodesContributing} nodes on ${row.unitIds.length} units)`\n    lines.push(`  cluster ${row.clusterId} [${row.unitIds.join(', ')}]: specialist gain ${gain}`)\n  }\n  return lines.join('\\n')\n}\n\nfunction parseShipArgs(argv: string[]): { path: string; flags: Record<string, string> } {\n  const flags: Record<string, string> = {}\n  const positional: string[] = []\n  for (let index = 0; index < argv.length; index++) {\n    const token = argv[index]!\n    if (!token.startsWith('--')) {\n      positional.push(token)\n      continue\n    }\n    const [name, inline] = token.slice(2).split('=', 2) as [string, string | undefined]\n    if (name !== 'run-kind' && name !== 'content') throw new Error(`unknown flag --${name}`)\n    const value = inline ?? argv[++index]\n    if (value === undefined) throw new Error(`--${name} needs a value`)\n    flags[name] = value\n  }\n  if (positional.length !== 1) throw new Error(USAGE)\n  return { path: positional[0]!, flags }\n}\n\n/** The ledger names its search on every line; read the first. */\nasync function firstLineSearchId(path: string): Promise<string> {\n  const handle = await open(path, 'r')\n  try {\n    // A line longer than one request's limit cannot ship anyway.\n    const bytes = Buffer.alloc(SEARCH_LEDGER_BATCH_MAX_BYTES + 1)\n    const { bytesRead } = await handle.read(bytes, 0, bytes.byteLength, 0)\n    const newline = bytes.subarray(0, bytesRead).indexOf(0x0a)\n    if (newline < 0) throw new Error(`${path} holds no complete ledger line`)\n    const first = JSON.parse(bytes.toString('utf8', 0, newline)) as { searchId?: unknown }\n    if (typeof first.searchId !== 'string') throw new Error(`${path} line 1 names no searchId`)\n    return first.searchId\n  } finally {\n    await handle.close()\n  }\n}\n","#!/usr/bin/env node\n/**\n * agent-eval CLI.\n *\n *   agent-eval serve [--port 5005] [--host 127.0.0.1]\n *   agent-eval rpc <method>          # one request from stdin → one response on stdout\n *   agent-eval rpc-batch <method>    # JSONL stdin → JSONL stdout\n *   agent-eval analyst-benchmark ... # benchmark a real-model trace analyst\n *   agent-eval supervisor-run report <runDir> [--format headline|markdown|json]\n *   agent-eval search ship <search-ledger.jsonl> --run-kind optimization|eval\n *   agent-eval search show <search-ledger.jsonl>\n *   agent-eval openapi [--out path]  # write OpenAPI spec\n *   agent-eval version\n *\n * <method> is one of: judge, listRubrics, version. When omitted, the\n * stdin payload must be a full {method, params} envelope.\n */\nimport { writeFileSync } from 'node:fs'\nimport { runAnalystBenchmarkCommand } from './analyst/benchmark-command'\nimport { resolveCliLlmConfig } from './cli-config'\nimport { runRolloutReleaseCli } from './rollout/release/hf-dataset'\nimport { runSearchCommand } from './search-command'\nimport { runSupervisorRunCommand } from './supervisor-run/report-command'\nimport { handleVersion } from './wire/handlers'\nimport { buildOpenApi } from './wire/openapi'\nimport { runRpcBatch, runRpcOnce } from './wire/rpc'\nimport { startServerAsync } from './wire/server'\n\ninterface Args {\n  command: string\n  positional: string[]\n  flags: Record<string, string>\n}\n\nfunction parseArgs(argv: string[]): Args {\n  const [command, ...rest] = argv\n  const positional: string[] = []\n  const flags: Record<string, string> = {}\n  for (let i = 0; i < rest.length; i++) {\n    const tok = rest[i]!\n    if (tok.startsWith('--')) {\n      const raw = tok.slice(2)\n      const equalsAt = raw.indexOf('=')\n      if (equalsAt >= 0) {\n        flags[raw.slice(0, equalsAt)] = raw.slice(equalsAt + 1)\n        continue\n      }\n      const key = raw\n      if (key === 'help') {\n        flags[key] = 'true'\n        continue\n      }\n      const next = rest[i + 1]\n      if (next != null && !next.startsWith('--')) {\n        flags[key] = next\n        i++\n      } else {\n        flags[key] = 'true'\n      }\n    } else if (tok === '-h') {\n      flags.help = 'true'\n    } else {\n      positional.push(tok)\n    }\n  }\n  return { command: command ?? 'help', positional, flags }\n}\n\nconst HELP = `agent-eval: evaluation RPC and HTTP server.\n\nCommands:\n  serve [--port 5005] [--host 127.0.0.1]\n        Start the HTTP server. POST /v1/judge, GET /v1/rubrics, GET /v1/version, GET /openapi.json.\n  rpc <method>\n        Read one JSON object from stdin (the params for <method>), write one\n        JSON object to stdout. Methods: judge, listRubrics, version.\n  rpc-batch <method>\n        Like 'rpc' but JSONL in / JSONL out.\n  openapi [--out openapi.json]\n        Write the OpenAPI 3.1 spec.\n  rollout-release <ledger.jsonl...> --out <dir> [--formats sft,verifiers,rft,raw] [--include-proposers] [--push <org/name>]\n        Build a HuggingFace-ready dataset dir from tangle.rollout.v1 ledgers:\n        validate, fail-closed split filter, scrub, export formats + card.\n  analyst-benchmark --dataset <agentrx|codetracebench> --labels <path> --trace-dir <path> [--artifact-dir <path>] --out <dir> ...\n        Compare an empty baseline with a real-model trace analyst on public labels.\n        Run with --help for all required inputs and controls.\n  supervisor-run report <runDir> [--format headline|markdown|json]\n        Report one Runtime or loops supervisor run directory. The status comes\n        from Runtime's own result.json / failure.json; every missing measurement\n        stays named. Exits 1 when the directory cannot be read.\n  search ship <search-ledger.jsonl> --run-kind optimization|eval [--content full|digests]\n        Ship a search ledger and its blobs to the hosted store in TANGLE_INGEST_URL,\n        starting from the store's head. Run with --help for the environment it reads.\n  search show <search-ledger.jsonl>\n        Verify a search ledger and print its summary: leading nodes, recently\n        discarded nodes and recent proposals, the same text a proposer reads.\n  version\n        Print server + wire-protocol version JSON.\n\nJudge provider:\n  Set AGENT_EVAL_LLM_BASE_URL, AGENT_EVAL_LLM_API_KEY, and AGENT_EVAL_LLM_MODEL.\n  OPENAI_* and TANGLE_* equivalents are also accepted. This binary is the only\n  place that reads a provider credential; the library never does. Without both\n  a base URL and a key, /v1/judge refuses with llm_not_configured.\n\nWithout arguments, prints this help.`\n\nasync function main(): Promise<number> {\n  const { command, positional, flags } = parseArgs(process.argv.slice(2))\n  assertKnownFlags(command, flags)\n\n  if (command === 'analyst-benchmark' && flags.help === 'true') {\n    return await runAnalystBenchmarkCommand(process.argv.slice(3))\n  }\n  if (command === 'search') return await runSearchCommand(process.argv.slice(3))\n  if (flags.help === 'true') {\n    process.stdout.write(`${HELP}\\n`)\n    return 0\n  }\n\n  switch (command) {\n    case 'serve': {\n      const port = parsePort(flags.port ?? '5005')\n      const host = flags.host ?? '127.0.0.1'\n      const llm = resolveCliLlmConfig()\n      const { server } = await startServerAsync({\n        port,\n        host,\n        ...(llm.chat ? { chat: llm.chat } : {}),\n        ...(llm.model ? { judgeModel: llm.model } : {}),\n      })\n      // Keep process alive on SIGINT/SIGTERM\n      const shutdown = (sig: string) => {\n        // eslint-disable-next-line no-console\n        console.log(`[agent-eval] received ${sig}, shutting down`)\n        server.close(() => process.exit(0))\n        // Force exit after 5s if close hangs\n        setTimeout(() => process.exit(1), 5000).unref()\n      }\n      process.on('SIGINT', () => shutdown('SIGINT'))\n      process.on('SIGTERM', () => shutdown('SIGTERM'))\n      // Block forever\n      await new Promise(() => {})\n      return 0\n    }\n    case 'rpc': {\n      const [method] = positional\n      const llm = resolveCliLlmConfig()\n      return await runRpcOnce(method, {\n        ...(llm.chat ? { chat: llm.chat } : {}),\n        ...(llm.model ? { judgeModel: llm.model } : {}),\n      })\n    }\n    case 'rpc-batch': {\n      const [method] = positional\n      const llm = resolveCliLlmConfig()\n      return await runRpcBatch(method, {\n        ...(llm.chat ? { chat: llm.chat } : {}),\n        ...(llm.model ? { judgeModel: llm.model } : {}),\n      })\n    }\n    case 'openapi': {\n      const out = flags.out ?? 'openapi.json'\n      const spec = buildOpenApi(handleVersion().version)\n      writeFileSync(out, `${JSON.stringify(spec, null, 2)}\\n`, 'utf-8')\n      // eslint-disable-next-line no-console\n      console.log(`[agent-eval] wrote OpenAPI 3.1 spec to ${out}`)\n      return 0\n    }\n    case 'rollout-release': {\n      // The subcommand owns its own flag grammar (multi-value --formats,\n      // boolean --include-proposers) — pass raw argv through untouched.\n      return await runRolloutReleaseCli(process.argv.slice(3))\n    }\n    case 'analyst-benchmark': {\n      return await runAnalystBenchmarkCommand(process.argv.slice(3))\n    }\n    case 'supervisor-run': {\n      // The subcommand owns its own flag grammar; pass raw argv through.\n      return await runSupervisorRunCommand(process.argv.slice(3))\n    }\n    case 'version': {\n      process.stdout.write(`${JSON.stringify(handleVersion(), null, 2)}\\n`)\n      return 0\n    }\n    case '--version': {\n      process.stdout.write(`${handleVersion().version}\\n`)\n      return 0\n    }\n    case 'help':\n    case '--help':\n    case '-h':\n    case '':\n      process.stdout.write(`${HELP}\\n`)\n      return 0\n    default:\n      process.stderr.write(`unknown command: ${command}\\n${HELP}\\n`)\n      return 1\n  }\n}\n\nconst FLAGS_BY_COMMAND: Record<string, ReadonlySet<string>> = {\n  serve: new Set(['help', 'host', 'port']),\n  rpc: new Set(['help']),\n  'rpc-batch': new Set(['help']),\n  openapi: new Set(['help', 'out']),\n  version: new Set(['help']),\n  help: new Set(),\n  '--help': new Set(),\n  '-h': new Set(),\n  '--version': new Set(),\n  '': new Set(),\n}\n\nfunction assertKnownFlags(command: string, flags: Record<string, string>): void {\n  // These subcommands parse their own argv.\n  if (\n    command === 'rollout-release' ||\n    command === 'analyst-benchmark' ||\n    command === 'supervisor-run' ||\n    command === 'search'\n  ) {\n    return\n  }\n  const allowed = FLAGS_BY_COMMAND[command]\n  if (!allowed) return\n  const unknown = Object.keys(flags).filter((flag) => !allowed.has(flag))\n  if (unknown.length > 0) {\n    throw new Error(`unknown flag for ${command || 'help'}: --${unknown[0]}`)\n  }\n}\n\nfunction parsePort(raw: string): number {\n  const port = Number(raw)\n  if (!Number.isInteger(port) || port < 0 || port > 65_535) {\n    throw new Error(`--port must be an integer from 0 to 65535; received ${JSON.stringify(raw)}`)\n  }\n  return port\n}\n\nmain()\n  .then((code) => {\n    process.exitCode = code\n  })\n  .catch((err) => {\n    // eslint-disable-next-line no-console\n    console.error('[agent-eval] cli error:', err)\n    process.exitCode = 1\n  })\n"],"mappings":";;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;AA8BA,SAAgB,wBACd,MAAyB,QAAQ,KACH;CAC9B,MAAM,kBAAkB,SAAS,IAAI,uBAAuB;CAC5D,MAAM,iBAAiB,SAAS,IAAI,sBAAsB;CAC1D,MAAM,eAAe,SAAS,IAAI,cAAc;CAChD,MAAM,eAAe,SAAS,IAAI,cAAc;CAChD,MAAM,UACJ,mBACA,SAAS,IAAI,eAAe,KAC5B,SAAS,IAAI,iBAAiB,MAC7B,eAAe,8BAA8B,KAAA,OAC7C,eAAe,mCAAmC,KAAA;CACrD,MAAM,SAAS,kBAAkB,gBAAgB;CACjD,MAAM,QACJ,SAAS,IAAI,oBAAoB,KAAK,SAAS,IAAI,YAAY,KAAK,SAAS,IAAI,YAAY;CAE/F,IAAI,CAAC,WAAW,CAAC,QAAQ,OAAO,KAAA;CAChC,OAAO;EAAE;EAAS;EAAQ,GAAI,QAAQ,EAAE,MAAM,IAAI,CAAC;CAAG;AACxD;AAEA,SAAgB,oBAAoB,MAAyB,QAAQ,KAAmB;CACtF,MAAM,QAAQ,wBAAwB,GAAG;CACzC,IAAI,CAAC,OAAO,OAAO,CAAC;CACpB,OAAO;EACL,MAAM,cAAc;GAAE,SAAS,MAAM;GAAS,QAAQ,MAAM;EAAO,GAAG,MAAM,KAAK;EACjF,GAAI,MAAM,QAAQ,EAAE,OAAO,MAAM,MAAM,IAAI,CAAC;CAC9C;AACF;;;;;;;;;AAUA,SAAS,cACP,OACA,cACY;CACZ,MAAM,SAAS,iBAAiB;EAC9B,WAAW;EACX,SAAS,MAAM;EACf,QAAQ,MAAM;EACd,GAAI,eAAe,EAAE,aAAa,IAAI,CAAC;CACzC,CAAC;CACD,OAAO;EACL,GAAG;EACH,OAAO,KAAK,aAAa;GACvB,IAAI,CAAC,IAAI,SAAS,CAAC,cACjB,MAAM,IAAI,MACR,4EACF;GAEF,OAAO,OAAO,KAAK,KAAK,QAAQ;EAClC;CACF;AACF;AAEA,SAAS,SAAS,OAA+C;CAC/D,MAAM,UAAU,OAAO,KAAK;CAC5B,OAAO,UAAU,UAAU,KAAA;AAC7B;;;;;;;;;;;;;;;;;;;;;;;;;;;;AC9CA,MAAM,QAAQ;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;AA0Cd,eAAsB,iBAAiB,MAAiC;CACtE,MAAM,CAAC,YAAY,GAAG,QAAQ;CAC9B,IAAI,eAAe,KAAA,KAAa,eAAe,YAAY,eAAe,MAAM;EAC9E,QAAQ,OAAO,MAAM,GAAG,MAAM,GAAG;EACjC,OAAO;CACT;CACA,IAAI,eAAe,QAAQ,OAAO,MAAM,eAAe,IAAI;CAC3D,IAAI,eAAe,QAAQ;EACzB,QAAQ,OAAO,MAAM,8BAA8B,WAAW,IAAI,MAAM,GAAG;EAC3E,OAAO;CACT;CACA,IAAI,KAAK,MAAM,QAAQ,QAAQ,KAAA,KAAa,QAAQ,YAAY,QAAQ,IAAI,GAAG;EAC7E,QAAQ,OAAO,MAAM,GAAG,MAAM,GAAG;EACjC,OAAO;CACT;CACA,MAAM,EAAE,MAAM,UAAU,cAAc,IAAI;CAC1C,MAAM,UAAU,oBAAoB,UAAU,MAAM,WAAW;CAC/D,IAAI,CAAC,QAAQ,SAAS,MAAM,IAAI,MAAM,4CAA4C,OAAO;CACzF,MAAM,UAAU,MAAM,WAAW;CACjC,IAAI,YAAY,UAAU,YAAY,WACpC,MAAM,IAAI,MAAM,sCAAsC,OAAO;CAE/D,MAAM,SAAS,oBAAoB;CACnC,IAAI,CAAC,QACH,MAAM,IAAI,MACR,qFACF;CAEF,MAAM,WAAW,MAAM,kBAAkB,IAAI;CAC7C,IAAI;EACF,MAAM,UAAU,MAAM,iBAAiB;GACrC;GACA,QAAQ;IAAE;IAAM;GAAS;GACzB,SAAS,QAAQ;GACjB;EACF,CAAC;EACD,QAAQ,OAAO,MAAM,GAAG,KAAK,UAAU,SAAS,MAAM,CAAC,EAAE,GAAG;EAC5D,OAAO,QAAQ,KAAK,iBAAiB,QAAQ,aAAa,IAAI;CAChE,SAAS,OAAO;EACd,IAAI,EAAE,iBAAiB,0BAA0B,MAAM;EACvD,QAAQ,OAAO,MAAM,GAAG,MAAM,QAAQ,GAAG;EACzC,OAAO;CACT;AACF;AAEA,eAAe,eAAe,MAAiC;CAC7D,IAAI,KAAK,SAAS,QAAQ,KAAK,KAAK,SAAS,IAAI,GAAG;EAClD,QAAQ,OAAO,MAAM,GAAG,MAAM,GAAG;EACjC,OAAO;CACT;CACA,IAAI,KAAK,SAAS,QAAQ,GAAG,OAAO,MAAM,mBAAmB,IAAI;CACjE,MAAM,aAAa,KAAK,QAAQ,QAAQ,CAAC,IAAI,WAAW,IAAI,CAAC;CAC7D,MAAM,QAAQ,IAAI,IAAI,KAAK,QAAQ,QAAQ,IAAI,WAAW,IAAI,CAAC,CAAC;CAChE,MAAM,UAAU,CAAC,GAAG,KAAK,CAAC,CAAC,QAAQ,SAAS,CAAC,WAAW,IAAI,IAAI,CAAC;CACjE,IAAI,WAAW,WAAW,KAAK,QAAQ,SAAS,GAAG;EACjD,IAAI,QAAQ,SAAS,GAAG,QAAQ,OAAO,MAAM,gBAAgB,QAAQ,KAAK,IAAI,EAAE,GAAG;EACnF,QAAQ,OAAO,MAAM,GAAG,MAAM,GAAG;EACjC,OAAO;CACT;CACA,MAAM,OAAO,WAAW;CACxB,MAAM,WAAW,MAAM,kBAAkB,IAAI;CAC7C,MAAM,OAAO,MAAM,SAAS,MAAM,MAAM;CACxC,MAAM,QAAQ,uBAAuB,MAAM,UAAU,IAAI;CACzD,MAAM,QACJ,MAAM,UAAU,MAAM,OAAO,OAAO,UAAU,MAAM,SAAS,IAAI,cAAc;CACjF,MAAM,SAAS,WAAW,OAAO,MAAM,KAAK;CAC5C,IAAI,MAAM,IAAI,QAAQ,GAAG;EACvB,IAAI,OAAO,WAAW,GAAG;GACvB,QAAQ,OAAO,MAAM,4BAA4B;GACjD,OAAO;EACT;EACA,MAAM,OAAO,OAAO,YAAY,OAAO,KAAK,SAAS,CAAC,KAAK,MAAM,KAAK,MAAM,CAAC,CAAC;EAC9E,QAAQ,OAAO,MAAM,GAAG,KAAK,UAAU,MAAM,MAAM,CAAC,EAAE,GAAG;EACzD,OAAO;CACT;CACA,MAAM,WAAW,CAAC,oBAAoB,OAAO,EAAE,MAAM,CAAC,GAAG,GAAG,OAAO,KAAK,SAAS,KAAK,IAAI,CAAC;CAC3F,QAAQ,OAAO,MAAM,GAAG,SAAS,KAAK,MAAM,EAAE,GAAG;CACjD,OAAO;AACT;AAEA,eAAe,mBAAmB,MAAiC;CACjE,MAAM,QAAkB,CAAC;CACzB,IAAI;CACJ,IAAI,OAAO;CACX,KAAK,IAAI,QAAQ,GAAG,QAAQ,KAAK,QAAQ,SAAS;EAChD,MAAM,QAAQ,KAAK;EACnB,IAAI,UAAU,UAAU;EACxB,IAAI,UAAU,UAAU;GACtB,OAAO;GACP;EACF;EACA,IAAI,UAAU,eAAe;GAC3B,YAAY,KAAK,EAAE;GACnB,IAAI,cAAc,KAAA,GAAW,MAAM,IAAI,MAAM,8BAA8B,OAAO;GAClF;EACF;EACA,IAAI,MAAM,WAAW,IAAI,GAAG,MAAM,IAAI,MAAM,gBAAgB,MAAM,IAAI,OAAO;EAC7E,MAAM,KAAK,KAAK;CAClB;CACA,IAAI,MAAM,WAAW,GAAG,MAAM,IAAI,MAAM,0CAA0C,OAAO;CACzF,MAAM,SAAS,CAAC;CAChB,KAAK,MAAM,QAAQ,OAAO;EACxB,MAAM,WAAW,MAAM,kBAAkB,IAAI;EAC7C,OAAO,KAAK,uBAAuB,MAAM,SAAS,MAAM,MAAM,GAAG,UAAU,IAAI,CAAC;CAClF;CACA,MAAM,OAAO,WAAW,QAAQ,cAAc,KAAA,IAAY,CAAC,IAAI,EAAE,UAAU,CAAC;CAC5E,QAAQ,OAAO,MACb,OAAO,GAAG,KAAK,UAAU,EAAE,YAAY,KAAK,GAAG,MAAM,CAAC,EAAE,MAAM,GAAG,qBAAqB,IAAI,EAAE,GAC9F;CACA,OAAO;AACT;AAEA,MAAM,6BAAa,IAAI,IAAI;CACzB;CACA;CACA;CACA;CACA;CACA;CACA;CACA;AACF,CAAC;AAED,SAAS,WACP,OACA,YACA,OACwD;CACxD,MAAM,SAAiE,CAAC;CACxE,IAAI,MAAM,IAAI,QAAQ,GAAG;EACvB,MAAM,SAAS,KAAK,KAAK;EACzB,OAAO,KAAK;GAAE,MAAM;GAAQ;GAAQ,MAAM,eAAe,OAAO,IAAI;EAAE,CAAC;CACzE;CACA,IAAI,MAAM,IAAI,kBAAkB,GAAG;EACjC,MAAM,SAAS,cAAc,KAAK;EAClC,OAAO,KAAK;GAAE,MAAM;GAAiB;GAAQ,MAAM,wBAAwB,OAAO,IAAI;EAAE,CAAC;CAC3F;CACA,IAAI,MAAM,IAAI,SAAS,GAAG;EACxB,MAAM,SAAS,MAAM,KAAK;EAC1B,OAAO,KAAK;GAAE,MAAM;GAAS;GAAQ,MAAM,gBAAgB,OAAO,IAAI;EAAE,CAAC;CAC3E;CACA,IAAI,MAAM,IAAI,eAAe,GAAG;EAC9B,MAAM,SAAS,WAAW,KAAK;EAC/B,OAAO,KAAK;GAAE,MAAM;GAAc;GAAQ,MAAM,qBAAqB,OAAO,IAAI;EAAE,CAAC;CACrF;CACA,IAAI,MAAM,IAAI,eAAe,GAAG;EAC9B,MAAM,SAAS,WAAW,OAAO,EAAE,cAAc,iBAAiB,UAAU,EAAE,CAAC;EAC/E,OAAO,KAAK;GAAE,MAAM;GAAc;GAAQ,MAAM,eAAe,MAAM;EAAE,CAAC;CAC1E;CACA,IAAI,MAAM,IAAI,aAAa,GAAG;EAC5B,MAAM,OAAO,iBAAiB,UAAU;EACxC,MAAM,SAAS,UACb,OACA,kBAAkB,YAAY,KAAK,QAAQ,QAAQ,KAAK,IAAI,CAC9D;EACA,OAAO,KAAK;GAAE,MAAM;GAAa;GAAQ,MAAM,gBAAgB,MAAM;EAAE,CAAC;CAC1E;CACA,IAAI,MAAM,IAAI,kBAAkB,GAAG;EACjC,MAAM,SAAS,cAAc,KAAK;EAClC,OAAO,KAAK;GAAE,MAAM;GAAiB;GAAQ,MAAM,oBAAoB,MAAM;EAAE,CAAC;CAClF;CACA,OAAO;AACT;;;;;;AAOA,SAAS,iBAAiB,YAAqE;CAC7F,MAAM,wBAAQ,IAAI,IAAqB;CACvC,QAAQ,QAAQ;EACd,IAAI,MAAM,IAAI,IAAI,MAAM,GAAG,OAAO,MAAM,IAAI,IAAI,MAAM;EACtD,MAAM,MAAM,IAAI,OAAO,MAAM,CAAgB;EAC7C,MAAM,aAAa,CAAC,KAAK,QAAQ,UAAU,GAAG,SAAS,GAAG,IAAI,MAAM,CAAC;EACrE,IAAI,IAAI,IAAI,WAAW,OAAO,GAAG,WAAW,QAAQ,cAAc,IAAI,GAAG,CAAC;EAC1E,IAAI;EACJ,KAAK,MAAM,aAAa,YAAY;GAClC,IAAI,CAAC,WAAW,SAAS,GAAG;GAC5B,MAAM,QAAQ,aAAa,SAAS;GAEpC,IADe,WAAW,QAAQ,CAAC,CAAC,OAAO,KAAK,CAAC,CAAC,OAAO,KAChD,MAAM,OAAO,MAAM,eAAe,IAAI,YAAY;GAC3D,IAAI;IACF,QAAQ,KAAK,MAAM,MAAM,SAAS,MAAM,CAAC;GAC3C,QAAQ,CAER;GACA;EACF;EACA,MAAM,IAAI,IAAI,QAAQ,KAAK;EAC3B,OAAO;CACT;AACF;AAEA,SAAS,eAAe,MAAwB;CAC9C,MAAM,QAAQ,CACZ,SAAS,KAAK,UAAU,OAAO,KAAK,cAAc,IAAI,KAAK,IAAI,IAAI,KAAK,UAAU,OAAO,KAAK,cAAc,IAAI,KAAK,KACvH;CACA,MAAM,QAAQ,MAAgB,QAAgB,WAA0B;EACtE,MAAM,SAAS,WAAW,KAAK,KAAK,SAAS,QAAQ;EACrD,MAAM,KAAK,KAAK,WAAW,GAAG,KAAK,SAAS,OAAO;EACnD,MAAM,OAAO,KAAK,SAAS,OAAO,KAAK,SAAS,KAAK;EACrD,MAAM,QAAQ,KAAK,UAAU,OAAO,mBAAmB;EACvD,MAAM,OACJ,KAAK,qBAAqB,IACtB,IAAI,KAAK,aAAa,QAAQ,CAAC,MAC/B,IAAI,KAAK,aAAa,QAAQ,CAAC,EAAE,qBAAqB,KAAK,SAAS,QAAQ,CAAC,EAAE,QAAQ,KAAK,iBAAiB,oBAAoB,KAAK,qBAAqB,IAAI,KAAK;EAC1K,MAAM,KACJ,GAAG,SAAS,SAAS,KAAK,KAAK,OAAO,IAAI,KAAK,UAAU,cAAc,OAAO,MAAM,IAAI,MAC1F;EACA,MAAM,cAAc,UAAU,WAAW,KAAK,KAAK,SAAS,QAAQ;EACpE,KAAK,SAAS,SAAS,OAAO,UAAU;GACtC,KAAK,OAAO,aAAa,UAAU,KAAK,SAAS,SAAS,CAAC;EAC7D,CAAC;CACH;CACA,KAAK,MAAM,SAAS,MAAM,UAAU;EAClC,KAAK,MAAM,IAAI,UAAU,KAAK,MAAM,SAAS,CAAC;CAChD,CAAC;CACD,OAAO,MAAM,KAAK,IAAI;AACxB;AAEA,SAAS,wBAAwB,MAAiC;CAChE,MAAM,QAAQ,CAAC,mBAAmB,KAAK,MAAM,gDAAgD;CAC7F,IAAI,KAAK,KAAK,WAAW,GAAG,MAAM,KAAK,gBAAgB;CACvD,KAAK,MAAM,OAAO,KAAK,MAAM;EAC3B,MAAM,IAAI,IAAI;EACd,MAAM,WAAW,OAAO,QAAQ,IAAI,QAAQ,CAAC,CAC1C,QAAQ,GAAG,WAAW,QAAQ,CAAC,CAAC,CAChC,KAAK,CAAC,QAAQ,WAAW,GAAG,OAAO,GAAG,OAAO,CAAC,CAC9C,KAAK,GAAG;EACX,MAAM,YACJ,EAAE,WAAW,SACT,EAAE,MAAM,IACN,gCACA,kBAAkB,EAAE,EAAE,MAAM,kBAAkB,WAChD,EAAE,WAAW,iBACX,SAAS,EAAE,KAAM,QAAQ,CAAC,EAAE,MAAM,EAAE,EAAE,MAAM,kBAAkB,mBAC9D,SAAS,EAAE,KAAM,QAAQ,CAAC,EAAE,MAAM,EAAE,SAAU,EAAE,CAAC,QAAQ,CAAC,EAAE,IAAI,EAAE,SAAU,EAAE,CAAC,QAAQ,CAAC,EAAE,KAAK,EAAE,OAAO,IAAI,EAAE,eAAe,MAAM,EAAE,EAAE;EAC/I,MAAM,QAAQ,OAAO,OAAO,IAAI,QAAQ,CAAC,CAAC,QAAQ,KAAK,UAAU,MAAM,OAAO,CAAC;EAC/E,MAAM,cAAc,IAAI,cAAc,IAAI,KAAK,IAAI,YAAY,iBAAiB;EAChF,MAAM,KACJ,KAAK,IAAI,SAAS,IAAI,IAAI,UAAU,YAAY,YAAY,KAAK,MAAM,OAAO,UAAU,IAAI,KAAK,IAAI,IAAI,YAAY,mBAAmB,KAAK,UAAU,IAAI,IAAI,SAAS,UAC1K;CACF;CACA,OAAO,MAAM,KAAK,IAAI;AACxB;AAEA,SAAS,gBAAgB,MAAyB;CAChD,MAAM,UAAU,KAAK,KAAK,QAAQ,QAAQ,IAAI,OAAO;CACrD,MAAM,UAAU,OAAO,QAAQ,KAAK,UAAU,CAAC,CAC5C,QAAQ,GAAG,WAAW,QAAQ,CAAC,CAAC,CAChC,KAAK,CAAC,QAAQ,WAAW,GAAG,OAAO,GAAG,OAAO,CAAC,CAC9C,KAAK,IAAI;CACZ,MAAM,QAAQ,CACZ,UAAU,KAAK,MAAM,gBAAgB,KAAK,KAAK,KAAK,IAAI,EAAE,KAAK,KAAK,aAAa,MAAM,KAAK,KAAK,OAAO,4BAA4B,KAAK,cAAc,WAAW,UAAU,KAAK,QAAQ,KAAK,IAChM;CACA,KAAK,MAAM,OAAO,SAChB,MAAM,KACJ,KAAK,IAAI,OAAO,UAAU,IAAI,MAAO,QAAQ,CAAC,EAAE,MAAM,IAAI,MAAM,UAAU,IAAI,OAAO,iBAAiB,IAAI,eAAgB,QAAQ,CAAC,GACrI;CAEF,OAAO,MAAM,KAAK,IAAI;AACxB;AAEA,SAAS,qBAAqB,MAA8B;CAC1D,MAAM,QAAQ,CACZ,gBAAgB,KAAK,MAAM,UAAU,KAAK,UAAU,KAAK,KAAK,QAAQ,OAAO,aAAa,KAAK,QAAQ,OAAO,YAAY,KAAK,aAAa,OAAO,oBAAoB,KAAK,aAAa,OAAO,iBAClM;CACA,KAAK,MAAM,OAAO,KAAK,gBAAgB;EACrC,MAAM,OACJ,IAAI,SAAS,OACT,iBAAiB,IAAI,aAAa,KAClC,GAAG,IAAI,KAAK,QAAQ,CAAC,EAAE,SAAS,IAAI,WAAW,IAAI,IAAI,kBAAkB,YAAY,IAAI,QAAQ,OAAO;EAC9G,MAAM,KAAK,aAAa,IAAI,UAAU,IAAI,IAAI,QAAQ,KAAK,IAAI,EAAE,qBAAqB,MAAM;CAC9F;CACA,OAAO,MAAM,KAAK,IAAI;AACxB;AAEA,SAAS,cAAc,MAAiE;CACtF,MAAM,QAAgC,CAAC;CACvC,MAAM,aAAuB,CAAC;CAC9B,KAAK,IAAI,QAAQ,GAAG,QAAQ,KAAK,QAAQ,SAAS;EAChD,MAAM,QAAQ,KAAK;EACnB,IAAI,CAAC,MAAM,WAAW,IAAI,GAAG;GAC3B,WAAW,KAAK,KAAK;GACrB;EACF;EACA,MAAM,CAAC,MAAM,UAAU,MAAM,MAAM,CAAC,CAAC,CAAC,MAAM,KAAK,CAAC;EAClD,IAAI,SAAS,cAAc,SAAS,WAAW,MAAM,IAAI,MAAM,kBAAkB,MAAM;EACvF,MAAM,QAAQ,UAAU,KAAK,EAAE;EAC/B,IAAI,UAAU,KAAA,GAAW,MAAM,IAAI,MAAM,KAAK,KAAK,eAAe;EAClE,MAAM,QAAQ;CAChB;CACA,IAAI,WAAW,WAAW,GAAG,MAAM,IAAI,MAAM,KAAK;CAClD,OAAO;EAAE,MAAM,WAAW;EAAK;CAAM;AACvC;;AAGA,eAAe,kBAAkB,MAA+B;CAC9D,MAAM,SAAS,MAAM,KAAK,MAAM,GAAG;CACnC,IAAI;EAEF,MAAM,QAAQ,OAAO,MAAM,gCAAgC,CAAC;EAC5D,MAAM,EAAE,cAAc,MAAM,OAAO,KAAK,OAAO,GAAG,MAAM,YAAY,CAAC;EACrE,MAAM,UAAU,MAAM,SAAS,GAAG,SAAS,CAAC,CAAC,QAAQ,EAAI;EACzD,IAAI,UAAU,GAAG,MAAM,IAAI,MAAM,GAAG,KAAK,+BAA+B;EACxE,MAAM,QAAQ,KAAK,MAAM,MAAM,SAAS,QAAQ,GAAG,OAAO,CAAC;EAC3D,IAAI,OAAO,MAAM,aAAa,UAAU,MAAM,IAAI,MAAM,GAAG,KAAK,0BAA0B;EAC1F,OAAO,MAAM;CACf,UAAU;EACR,MAAM,OAAO,MAAM;CACrB;AACF;;;;;;;;;;;;;;;;;;;AChXA,SAAS,UAAU,MAAsB;CACvC,MAAM,CAAC,SAAS,GAAG,QAAQ;CAC3B,MAAM,aAAuB,CAAC;CAC9B,MAAM,QAAgC,CAAC;CACvC,KAAK,IAAI,IAAI,GAAG,IAAI,KAAK,QAAQ,KAAK;EACpC,MAAM,MAAM,KAAK;EACjB,IAAI,IAAI,WAAW,IAAI,GAAG;GACxB,MAAM,MAAM,IAAI,MAAM,CAAC;GACvB,MAAM,WAAW,IAAI,QAAQ,GAAG;GAChC,IAAI,YAAY,GAAG;IACjB,MAAM,IAAI,MAAM,GAAG,QAAQ,KAAK,IAAI,MAAM,WAAW,CAAC;IACtD;GACF;GACA,MAAM,MAAM;GACZ,IAAI,QAAQ,QAAQ;IAClB,MAAM,OAAO;IACb;GACF;GACA,MAAM,OAAO,KAAK,IAAI;GACtB,IAAI,QAAQ,QAAQ,CAAC,KAAK,WAAW,IAAI,GAAG;IAC1C,MAAM,OAAO;IACb;GACF,OACE,MAAM,OAAO;EAEjB,OAAO,IAAI,QAAQ,MACjB,MAAM,OAAO;OAEb,WAAW,KAAK,GAAG;CAEvB;CACA,OAAO;EAAE,SAAS,WAAW;EAAQ;EAAY;CAAM;AACzD;AAEA,MAAM,OAAO;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;AAuCb,eAAe,OAAwB;CACrC,MAAM,EAAE,SAAS,YAAY,UAAU,UAAU,QAAQ,KAAK,MAAM,CAAC,CAAC;CACtE,iBAAiB,SAAS,KAAK;CAE/B,IAAI,YAAY,uBAAuB,MAAM,SAAS,QACpD,OAAO,MAAM,2BAA2B,QAAQ,KAAK,MAAM,CAAC,CAAC;CAE/D,IAAI,YAAY,UAAU,OAAO,MAAM,iBAAiB,QAAQ,KAAK,MAAM,CAAC,CAAC;CAC7E,IAAI,MAAM,SAAS,QAAQ;EACzB,QAAQ,OAAO,MAAM,GAAG,KAAK,GAAG;EAChC,OAAO;CACT;CAEA,QAAQ,SAAR;EACE,KAAK,SAAS;GACZ,MAAM,OAAO,UAAU,MAAM,QAAQ,MAAM;GAC3C,MAAM,OAAO,MAAM,QAAQ;GAC3B,MAAM,MAAM,oBAAoB;GAChC,MAAM,EAAE,WAAW,MAAM,iBAAiB;IACxC;IACA;IACA,GAAI,IAAI,OAAO,EAAE,MAAM,IAAI,KAAK,IAAI,CAAC;IACrC,GAAI,IAAI,QAAQ,EAAE,YAAY,IAAI,MAAM,IAAI,CAAC;GAC/C,CAAC;GAED,MAAM,YAAY,QAAgB;IAEhC,QAAQ,IAAI,yBAAyB,IAAI,gBAAgB;IACzD,OAAO,YAAY,QAAQ,KAAK,CAAC,CAAC;IAElC,iBAAiB,QAAQ,KAAK,CAAC,GAAG,GAAI,CAAC,CAAC,MAAM;GAChD;GACA,QAAQ,GAAG,gBAAgB,SAAS,QAAQ,CAAC;GAC7C,QAAQ,GAAG,iBAAiB,SAAS,SAAS,CAAC;GAE/C,MAAM,IAAI,cAAc,CAAC,CAAC;GAC1B,OAAO;EACT;EACA,KAAK,OAAO;GACV,MAAM,CAAC,UAAU;GACjB,MAAM,MAAM,oBAAoB;GAChC,OAAO,MAAM,WAAW,QAAQ;IAC9B,GAAI,IAAI,OAAO,EAAE,MAAM,IAAI,KAAK,IAAI,CAAC;IACrC,GAAI,IAAI,QAAQ,EAAE,YAAY,IAAI,MAAM,IAAI,CAAC;GAC/C,CAAC;EACH;EACA,KAAK,aAAa;GAChB,MAAM,CAAC,UAAU;GACjB,MAAM,MAAM,oBAAoB;GAChC,OAAO,MAAM,YAAY,QAAQ;IAC/B,GAAI,IAAI,OAAO,EAAE,MAAM,IAAI,KAAK,IAAI,CAAC;IACrC,GAAI,IAAI,QAAQ,EAAE,YAAY,IAAI,MAAM,IAAI,CAAC;GAC/C,CAAC;EACH;EACA,KAAK,WAAW;GACd,MAAM,MAAM,MAAM,OAAO;GACzB,MAAM,OAAO,aAAa,cAAc,CAAC,CAAC,OAAO;GACjD,cAAc,KAAK,GAAG,KAAK,UAAU,MAAM,MAAM,CAAC,EAAE,KAAK,OAAO;GAEhE,QAAQ,IAAI,0CAA0C,KAAK;GAC3D,OAAO;EACT;EACA,KAAK,mBAGH,OAAO,MAAM,qBAAqB,QAAQ,KAAK,MAAM,CAAC,CAAC;EAEzD,KAAK,qBACH,OAAO,MAAM,2BAA2B,QAAQ,KAAK,MAAM,CAAC,CAAC;EAE/D,KAAK,kBAEH,OAAO,MAAM,wBAAwB,QAAQ,KAAK,MAAM,CAAC,CAAC;EAE5D,KAAK;GACH,QAAQ,OAAO,MAAM,GAAG,KAAK,UAAU,cAAc,GAAG,MAAM,CAAC,EAAE,GAAG;GACpE,OAAO;EAET,KAAK;GACH,QAAQ,OAAO,MAAM,GAAG,cAAc,CAAC,CAAC,QAAQ,GAAG;GACnD,OAAO;EAET,KAAK;EACL,KAAK;EACL,KAAK;EACL,KAAK;GACH,QAAQ,OAAO,MAAM,GAAG,KAAK,GAAG;GAChC,OAAO;EACT;GACE,QAAQ,OAAO,MAAM,oBAAoB,QAAQ,IAAI,KAAK,GAAG;GAC7D,OAAO;CACX;AACF;AAEA,MAAM,mBAAwD;CAC5D,uBAAO,IAAI,IAAI;EAAC;EAAQ;EAAQ;CAAM,CAAC;CACvC,qBAAK,IAAI,IAAI,CAAC,MAAM,CAAC;CACrB,6BAAa,IAAI,IAAI,CAAC,MAAM,CAAC;CAC7B,yBAAS,IAAI,IAAI,CAAC,QAAQ,KAAK,CAAC;CAChC,yBAAS,IAAI,IAAI,CAAC,MAAM,CAAC;CACzB,sBAAM,IAAI,IAAI;CACd,0BAAU,IAAI,IAAI;CAClB,sBAAM,IAAI,IAAI;CACd,6BAAa,IAAI,IAAI;CACrB,oBAAI,IAAI,IAAI;AACd;AAEA,SAAS,iBAAiB,SAAiB,OAAqC;CAE9E,IACE,YAAY,qBACZ,YAAY,uBACZ,YAAY,oBACZ,YAAY,UAEZ;CAEF,MAAM,UAAU,iBAAiB;CACjC,IAAI,CAAC,SAAS;CACd,MAAM,UAAU,OAAO,KAAK,KAAK,CAAC,CAAC,QAAQ,SAAS,CAAC,QAAQ,IAAI,IAAI,CAAC;CACtE,IAAI,QAAQ,SAAS,GACnB,MAAM,IAAI,MAAM,oBAAoB,WAAW,OAAO,MAAM,QAAQ,IAAI;AAE5E;AAEA,SAAS,UAAU,KAAqB;CACtC,MAAM,OAAO,OAAO,GAAG;CACvB,IAAI,CAAC,OAAO,UAAU,IAAI,KAAK,OAAO,KAAK,OAAO,OAChD,MAAM,IAAI,MAAM,uDAAuD,KAAK,UAAU,GAAG,GAAG;CAE9F,OAAO;AACT;AAEA,KAAK,CAAC,CACH,MAAM,SAAS;CACd,QAAQ,WAAW;AACrB,CAAC,CAAC,CACD,OAAO,QAAQ;CAEd,QAAQ,MAAM,2BAA2B,GAAG;CAC5C,QAAQ,WAAW;AACrB,CAAC"}
+{"version":3,"file":"cli.js","names":[],"sources":["../src/cli-config.ts","../src/search-command.ts","../src/cli.ts"],"sourcesContent":["/**\n * Provider configuration for the `agent-eval` binary.\n *\n * The binary reads endpoint and credential values from its environment.\n * TypeScript library callers bind a transport explicitly. The maintained\n * HTTP transport accepts their credential and sends provider requests.\n */\n\nimport { type ChatClient, createChatClient } from './analyst/chat-client'\n\nexport interface CliLlmConfig {\n  /** Judge transport for the wire handlers. Absent when no provider is configured. */\n  chat?: ChatClient\n  model?: string\n}\n\n/** The endpoint the binary resolved from its environment. */\nexport interface CliProviderRoute {\n  baseUrl: string\n  apiKey: string\n  model?: string\n}\n\n/**\n * Environment precedence for the binary's provider route, with no client built.\n *\n * Both halves are required: a base URL without a key, or a key without an\n * endpoint, is a half-configured server. `/v1/judge` then refuses with\n * `llm_not_configured` instead of calling an unintended endpoint.\n */\nexport function resolveCliProviderRoute(\n  env: NodeJS.ProcessEnv = process.env,\n): CliProviderRoute | undefined {\n  const explicitBaseUrl = nonEmpty(env.AGENT_EVAL_LLM_BASE_URL)\n  const explicitApiKey = nonEmpty(env.AGENT_EVAL_LLM_API_KEY)\n  const openAiApiKey = nonEmpty(env.OPENAI_API_KEY)\n  const tangleApiKey = nonEmpty(env.TANGLE_API_KEY)\n  const baseUrl =\n    explicitBaseUrl ??\n    nonEmpty(env.OPENAI_BASE_URL) ??\n    nonEmpty(env.TANGLE_ROUTER_URL) ??\n    (openAiApiKey ? 'https://api.openai.com/v1' : undefined) ??\n    (tangleApiKey ? 'https://router.tangle.tools/v1' : undefined)\n  const apiKey = explicitApiKey ?? openAiApiKey ?? tangleApiKey\n  const model =\n    nonEmpty(env.AGENT_EVAL_LLM_MODEL) ?? nonEmpty(env.OPENAI_MODEL) ?? nonEmpty(env.TANGLE_MODEL)\n\n  if (!baseUrl || !apiKey) return undefined\n  return { baseUrl, apiKey, ...(model ? { model } : {}) }\n}\n\nexport function resolveCliLlmConfig(env: NodeJS.ProcessEnv = process.env): CliLlmConfig {\n  const route = resolveCliProviderRoute(env)\n  if (!route) return {}\n  return {\n    chat: cliChatClient({ baseUrl: route.baseUrl, apiKey: route.apiKey }, route.model),\n    ...(route.model ? { model: route.model } : {}),\n  }\n}\n\n/**\n * The binary's transport IS the published one — `createChatClient({ transport:\n * 'openai-compatible' })` — so the server and a library consumer cannot drift\n * onto two different clients for the same endpoint.\n *\n * The one thing kept local is the error text. A generic \"no model on the\n * request\" tells a server operator nothing; this names the variable to set.\n */\nfunction cliChatClient(\n  route: { baseUrl: string; apiKey: string },\n  defaultModel: string | undefined,\n): ChatClient {\n  const client = createChatClient({\n    transport: 'openai-compatible',\n    baseUrl: route.baseUrl,\n    apiKey: route.apiKey,\n    ...(defaultModel ? { defaultModel } : {}),\n  })\n  return {\n    ...client,\n    chat: (req, callOpts) => {\n      if (!req.model && !defaultModel) {\n        throw new Error(\n          'agent-eval: no model on the request and no AGENT_EVAL_LLM_MODEL configured',\n        )\n      }\n      return client.chat(req, callOpts)\n    },\n  }\n}\n\nfunction nonEmpty(value: string | undefined): string | undefined {\n  const trimmed = value?.trim()\n  return trimmed ? trimmed : undefined\n}\n","/**\n * `agent-eval search <subcommand>`: work with a search ledger from a terminal.\n *\n *   agent-eval search ship <search-ledger.jsonl> --run-kind optimization|eval [--content full|digests]\n *   agent-eval search show <search-ledger.jsonl> [--tree] [--operator-yield] [--front] [--task-matrix]\n *                          [--edit-credit] [--landscape] [--skill-manifold] [--json]\n *   agent-eval search show <search-ledger.jsonl> [<search-ledger.jsonl> ...] --meta [--objective <key>] [--json]\n *\n * `ship` sends the ledger to the hosted store named by `TANGLE_INGEST_URL`,\n * `TANGLE_INGEST_API_KEY` and `TANGLE_TENANT_ID`, starting from the store's\n * head, so it finishes or resumes a ship that a loop could not complete.\n *\n * `show` verifies the ledger and prints `renderSearchSummary`: the same\n * compact text a proposer reads, on the search's own ranking split (the\n * selection split when the search declares one, else train). There is no\n * local HTML renderer; the hosted store's page is the visual view (§5 of the\n * search-tree design). A lens flag adds that lens's text form after the\n * summary, so an agent reads what a person sees; `--json` prints the lens\n * results as JSON instead, the shape a view renders.\n *\n * `show --meta` reads every ledger it is given and prints the `metaSearch`\n * lens over them instead of one search's summary: each search as one node\n * scored by its held-out lift per known dollar, the configurations best first,\n * and the searches arranged by derivation and containment.\n */\n\nimport { createHash } from 'node:crypto'\nimport { existsSync, readFileSync } from 'node:fs'\nimport { open, readFile } from 'node:fs/promises'\nimport { dirname, join } from 'node:path'\nimport { fileURLToPath } from 'node:url'\nimport { replaySearchLedgerText } from './campaign/search-ledger'\nimport type { SearchArtifactRef } from './campaign/search-ledger-types'\nimport type { SearchStateView } from './campaign/search-state'\nimport { renderSearchSummary } from './campaign/search-summary'\nimport { hostedTenantFromEnv } from './hosted/client'\nimport { SEARCH_LEDGER_BATCH_MAX_BYTES, SearchRunKindSchema } from './hosted/search-ledger-wire'\nimport { SearchShipConflictError, shipSearchLedger } from './hosted/search-shipper'\nimport { editCredit, editCreditText } from './search/lenses/edit-credit'\nimport { type FrontData, front } from './search/lenses/front'\nimport { formatLandscape, landscape, surfaceTextEdits } from './search/lenses/landscape'\nimport { metaSearch, renderMetaSearchText } from './search/lenses/meta-search'\nimport { type OperatorYieldData, operatorYield } from './search/lenses/operator-yield'\nimport { INSUFFICIENT_FROM } from './search/lenses/shared'\nimport { formatSkillManifold, skillManifold } from './search/lenses/skill-manifold'\nimport { type TaskMatrixData, taskMatrix } from './search/lenses/task-matrix'\nimport { type TreeData, type TreeNode, tree } from './search/lenses/tree'\n\nconst USAGE = `usage: agent-eval search <subcommand> ...\n\n  ship <search-ledger.jsonl> --run-kind optimization|eval [--content full|digests]\n        Ships the ledger's entries, and the blobs they name, to the hosted store in\n        TANGLE_INGEST_URL (or TANGLE_ORCHESTRATOR_URL) as tenant TANGLE_TENANT_ID with\n        TANGLE_INGEST_API_KEY (or TANGLE_API_KEY). It starts from the store's head, so\n        running it again sends only what the store lacks. Prints the result as JSON.\n        Exits 1 when the store holds a different chain for the search.\n\n  show <search-ledger.jsonl> [--tree] [--operator-yield] [--front] [--task-matrix]\n                              [--edit-credit] [--landscape] [--skill-manifold] [--json]\n        Verifies the ledger and prints its search summary: the leading nodes\n        against the root, the most recently discarded nodes and why, and a log\n        of recent proposals. The same text a proposer reads as context, on the\n        search's own ranking split. Each lens flag adds that lens's text form\n        (search-tree-design §12) below the summary — the same JSON\n        Intelligence, discovery lab, VerticalBench and agent-runtime read.\n        --tree           the tidy tree of nodes and edges.\n        --operator-yield outcome counts and yield per known dollar, by edge operator.\n        --front          the Pareto front over score and known cost per cell.\n        --task-matrix    nodes × units, clustered, with specialist gain per cluster.\n        --edit-credit    every edit as a gene followed down the lineage, its\n                         credit, interacting pairs, and skill candidates. It\n                         reads node artifacts from the blobs the ledger names,\n                         verified by digest.\n        --landscape      nodes placed by line edits between their profile\n                         surfaces (read from the blobs, verified by digest), the\n                         kriged score surface, its basins, and the plateau score\n                         a policy drafts on.\n        --skill-manifold the node × unit score matrix factored into skill axes,\n                         its cross-validated intrinsic dimension, and the unit\n                         that best separates the leaders.\n        --json           prints the requested lenses as JSON instead of text.\n\n  show <search-ledger.jsonl> [<search-ledger.jsonl> ...] --meta [--objective <key>] [--json]\n        Verifies every ledger and prints the meta-search lens over them: each\n        search as one node, scored by its claim's held-out lift per known\n        dollar (unscored searches say why), configurations best first with\n        their interval, method and n, and the searches arranged by derivation\n        and containment. --objective names the objective the best\n        configuration is chosen within when the searches span several.`\n\nexport async function runSearchCommand(argv: string[]): Promise<number> {\n  const [subcommand, ...rest] = argv\n  if (subcommand === undefined || subcommand === '--help' || subcommand === '-h') {\n    process.stdout.write(`${USAGE}\\n`)\n    return 0\n  }\n  if (subcommand === 'show') return await runShowCommand(rest)\n  if (subcommand !== 'ship') {\n    process.stderr.write(`unknown search subcommand: ${subcommand}\\n${USAGE}\\n`)\n    return 1\n  }\n  if (rest.some((arg) => arg === undefined || arg === '--help' || arg === '-h')) {\n    process.stdout.write(`${USAGE}\\n`)\n    return 0\n  }\n  const { path, flags } = parseShipArgs(rest)\n  const runKind = SearchRunKindSchema.safeParse(flags['run-kind'])\n  if (!runKind.success) throw new Error(`--run-kind must be optimization or eval\\n${USAGE}`)\n  const content = flags.content ?? 'full'\n  if (content !== 'full' && content !== 'digests') {\n    throw new Error(`--content must be full or digests\\n${USAGE}`)\n  }\n  const tenant = hostedTenantFromEnv()\n  if (!tenant) {\n    throw new Error(\n      'set TANGLE_INGEST_URL, TANGLE_INGEST_API_KEY and TANGLE_TENANT_ID to name the store',\n    )\n  }\n  const searchId = await firstLineSearchId(path)\n  try {\n    const shipped = await shipSearchLedger({\n      tenant,\n      ledger: { path, searchId },\n      runKind: runKind.data,\n      content,\n    })\n    process.stdout.write(`${JSON.stringify(shipped, null, 2)}\\n`)\n    return shipped.head.nextSequence === shipped.localLines ? 0 : 1\n  } catch (error) {\n    if (!(error instanceof SearchShipConflictError)) throw error\n    process.stderr.write(`${error.message}\\n`)\n    return 1\n  }\n}\n\nasync function runShowCommand(argv: string[]): Promise<number> {\n  if (argv.includes('--help') || argv.includes('-h')) {\n    process.stdout.write(`${USAGE}\\n`)\n    return 0\n  }\n  if (argv.includes('--meta')) return await runShowMetaCommand(argv)\n  const positional = argv.filter((arg) => !arg.startsWith('--'))\n  const flags = new Set(argv.filter((arg) => arg.startsWith('--')))\n  const unknown = [...flags].filter((flag) => !SHOW_FLAGS.has(flag))\n  if (positional.length !== 1 || unknown.length > 0) {\n    if (unknown.length > 0) process.stderr.write(`unknown flag ${unknown.join(', ')}\\n`)\n    process.stdout.write(`${USAGE}\\n`)\n    return 1\n  }\n  const path = positional[0]!\n  const searchId = await firstLineSearchId(path)\n  const text = await readFile(path, 'utf8')\n  const state = replaySearchLedgerText(text, searchId, path)\n  const split =\n    state.header && state.header.splits.selection.tasks.length > 0 ? 'selection' : 'train'\n  const lenses = showLenses(state, path, flags)\n  if (flags.has('--json')) {\n    if (lenses.length === 0) {\n      process.stderr.write('--json needs a lens flag\\n')\n      return 1\n    }\n    const json = Object.fromEntries(lenses.map((lens) => [lens.name, lens.result]))\n    process.stdout.write(`${JSON.stringify(json, null, 2)}\\n`)\n    return 0\n  }\n  const sections = [renderSearchSummary(state, { split }), ...lenses.map((lens) => lens.text)]\n  process.stdout.write(`${sections.join('\\n\\n')}\\n`)\n  return 0\n}\n\nasync function runShowMetaCommand(argv: string[]): Promise<number> {\n  const paths: string[] = []\n  let objective: string | undefined\n  let json = false\n  for (let index = 0; index < argv.length; index++) {\n    const token = argv[index]!\n    if (token === '--meta') continue\n    if (token === '--json') {\n      json = true\n      continue\n    }\n    if (token === '--objective') {\n      objective = argv[++index]\n      if (objective === undefined) throw new Error(`--objective needs a value\\n${USAGE}`)\n      continue\n    }\n    if (token.startsWith('--')) throw new Error(`unknown flag ${token}\\n${USAGE}`)\n    paths.push(token)\n  }\n  if (paths.length === 0) throw new Error(`show --meta needs at least one ledger\\n${USAGE}`)\n  const states = []\n  for (const path of paths) {\n    const searchId = await firstLineSearchId(path)\n    states.push(replaySearchLedgerText(await readFile(path, 'utf8'), searchId, path))\n  }\n  const lens = metaSearch(states, objective === undefined ? {} : { objective })\n  process.stdout.write(\n    json ? `${JSON.stringify({ metaSearch: lens }, null, 2)}\\n` : `${renderMetaSearchText(lens)}\\n`,\n  )\n  return 0\n}\n\nconst SHOW_FLAGS = new Set([\n  '--edit-credit',\n  '--json',\n  '--tree',\n  '--operator-yield',\n  '--front',\n  '--task-matrix',\n  '--landscape',\n  '--skill-manifold',\n])\n\nfunction showLenses(\n  state: SearchStateView,\n  ledgerPath: string,\n  flags: ReadonlySet<string>,\n): Array<{ name: string; result: unknown; text: string }> {\n  const lenses: Array<{ name: string; result: unknown; text: string }> = []\n  if (flags.has('--tree')) {\n    const result = tree(state)\n    lenses.push({ name: 'tree', result, text: renderTreeText(result.data) })\n  }\n  if (flags.has('--operator-yield')) {\n    const result = operatorYield(state)\n    lenses.push({ name: 'operatorYield', result, text: renderOperatorYieldText(result.data) })\n  }\n  if (flags.has('--front')) {\n    const result = front(state)\n    lenses.push({ name: 'front', result, text: renderFrontText(result.data) })\n  }\n  if (flags.has('--task-matrix')) {\n    const result = taskMatrix(state)\n    lenses.push({ name: 'taskMatrix', result, text: renderTaskMatrixText(result.data) })\n  }\n  if (flags.has('--edit-credit')) {\n    const result = editCredit(state, { readArtifact: ledgerBlobReader(ledgerPath) })\n    lenses.push({ name: 'editCredit', result, text: editCreditText(result) })\n  }\n  if (flags.has('--landscape')) {\n    const read = ledgerBlobReader(ledgerPath)\n    const result = landscape(\n      state,\n      surfaceTextEdits((surface) => read(surface.artifact) ?? null),\n    )\n    lenses.push({ name: 'landscape', result, text: formatLandscape(result) })\n  }\n  if (flags.has('--skill-manifold')) {\n    const result = skillManifold(state)\n    lenses.push({ name: 'skillManifold', result, text: formatSkillManifold(result) })\n  }\n  return lenses\n}\n\n/**\n * Reads a blob a ledger names: at its `file:` URI, else under `blobs/` beside\n * the ledger, where `SearchRecorder` writes them. Bytes that do not hash to the\n * reference's digest and length read as unavailable, never as content.\n */\nfunction ledgerBlobReader(ledgerPath: string): (ref: SearchArtifactRef) => unknown | undefined {\n  const cache = new Map<string, unknown>()\n  return (ref) => {\n    if (cache.has(ref.sha256)) return cache.get(ref.sha256)\n    const hex = ref.sha256.slice('sha256:'.length)\n    const candidates = [join(dirname(ledgerPath), 'blobs', `${hex}.json`)]\n    if (ref.uri.startsWith('file:')) candidates.unshift(fileURLToPath(ref.uri))\n    let value: unknown\n    for (const candidate of candidates) {\n      if (!existsSync(candidate)) continue\n      const bytes = readFileSync(candidate)\n      const digest = createHash('sha256').update(bytes).digest('hex')\n      if (digest !== hex || bytes.byteLength !== ref.byteLength) continue\n      try {\n        value = JSON.parse(bytes.toString('utf8'))\n      } catch {\n        // Verified bytes that are not JSON are not content a lens can read.\n      }\n      break\n    }\n    cache.set(ref.sha256, value)\n    return value\n  }\n}\n\nfunction renderTreeText(data: TreeData): string {\n  const lines = [\n    `tree: ${data.nodeCount} node${data.nodeCount === 1 ? '' : 's'}, ${data.edgeCount} edge${data.edgeCount === 1 ? '' : 's'}`,\n  ]\n  const walk = (node: TreeNode, prefix: string, isLast: boolean): void => {\n    const branch = prefix === '' ? '' : isLast ? '└─ ' : '├─ '\n    const op = node.operator ? `${node.operator} → ` : ''\n    const rung = node.rung === null ? '' : ` rung ${node.rung}`\n    const depth = node.depth === null ? ' depth unknown' : ''\n    const cost =\n      node.unknownCostCells === 0\n        ? `$${node.knownCostUsd.toFixed(2)}`\n        : `$${node.knownCostUsd.toFixed(2)} known + at least $${node.floorUsd.toFixed(2)} over ${node.unknownCostCells} unknown-cost cell${node.unknownCostCells === 1 ? '' : 's'}`\n    lines.push(\n      `${prefix}${branch}${op}${node.nodeId} (${node.status ?? 'undecided'}${rung}${depth}) ${cost}`,\n    )\n    const childPrefix = prefix + (prefix === '' ? '' : isLast ? '   ' : '│  ')\n    node.children.forEach((child, index) => {\n      walk(child, childPrefix, index === node.children.length - 1)\n    })\n  }\n  data.roots.forEach((root, index) => {\n    walk(root, '', index === data.roots.length - 1)\n  })\n  return lines.join('\\n')\n}\n\nfunction renderOperatorYieldText(data: OperatorYieldData): string {\n  const lines = [`operator yield (${data.split} split, gain per known $ of proposal + screen):`]\n  if (data.rows.length === 0) lines.push('  no edges yet')\n  for (const row of data.rows) {\n    const y = row.yield\n    const outcomes = Object.entries(row.outcomes)\n      .filter(([, count]) => count > 0)\n      .map(([status, count]) => `${status}=${count}`)\n      .join(' ')\n    const yieldText =\n      y.method === 'none'\n        ? y.n === 0\n          ? 'yield: no measured children'\n          : `yield unknown (${y.n} of ${INSUFFICIENT_FROM}, none)`\n        : y.method === 'insufficient'\n          ? `yield≈${y.mean!.toFixed(4)}/$ (${y.n} of ${INSUFFICIENT_FROM}, insufficient)`\n          : `yield≈${y.mean!.toFixed(4)}/$ [${y.interval![0].toFixed(4)}, ${y.interval![1].toFixed(4)}] (${y.method}, ${y.intervalMethod}, n=${y.n})`\n    const nodes = Object.values(row.outcomes).reduce((sum, count) => sum + count, 0)\n    const reproposals = row.reproposals > 0 ? `, ${row.reproposals} re-proposals` : ''\n    lines.push(\n      `  ${row.operator}: ${row.proposals} proposals${reproposals} → ${nodes} node${nodes === 1 ? '' : 's'} (${outcomes || 'no decisions yet'}); ${yieldText}; ${row.excluded} excluded`,\n    )\n  }\n  return lines.join('\\n')\n}\n\nfunction renderFrontText(data: FrontData): string {\n  const onFront = data.rows.filter((row) => row.onFront)\n  const reasons = Object.entries(data.exclusions)\n    .filter(([, count]) => count > 0)\n    .map(([reason, count]) => `${reason} ${count}`)\n    .join(', ')\n  const lines = [\n    `front (${data.split} split, axes: ${data.axes.join(', ')}): ${data.frontierSize} of ${data.rows.length} node(s) on the frontier, ${data.excludedCount} excluded${reasons ? ` (${reasons})` : ''}`,\n  ]\n  for (const row of onFront) {\n    lines.push(\n      `  ${row.nodeId}: score=${row.score!.toFixed(4)} on ${row.units} units (${row.method}) costPerCell=$${row.costPerCellUsd!.toFixed(4)}`,\n    )\n  }\n  return lines.join('\\n')\n}\n\nfunction renderTaskMatrixText(data: TaskMatrixData): string {\n  const lines = [\n    `task matrix (${data.split} split, ${data.direction}): ${data.nodeIds.length} node(s) × ${data.unitIds.length} unit(s), ${data.nodeClusters.length} node cluster(s), ${data.unitClusters.length} unit cluster(s)`,\n  ]\n  if (data.sampling) {\n    lines.push(\n      `  clustered ${data.sampling.nodesClustered} of ${data.sampling.nodesTotal} nodes (cap ${data.sampling.cap}), evenly sampled in registration order; unit clusters and specialist gains use the same nodes`,\n    )\n  }\n  for (const row of data.specialistGain) {\n    const gain =\n      row.gain === null\n        ? `insufficient (${row.insufficient})`\n        : `${row.gain.toFixed(4)} (best ${row.bestNodeId}, ${row.nodesContributing} nodes on ${row.unitIds.length} units)`\n    lines.push(`  cluster ${row.clusterId} [${row.unitIds.join(', ')}]: specialist gain ${gain}`)\n  }\n  return lines.join('\\n')\n}\n\nfunction parseShipArgs(argv: string[]): { path: string; flags: Record<string, string> } {\n  const flags: Record<string, string> = {}\n  const positional: string[] = []\n  for (let index = 0; index < argv.length; index++) {\n    const token = argv[index]!\n    if (!token.startsWith('--')) {\n      positional.push(token)\n      continue\n    }\n    const [name, inline] = token.slice(2).split('=', 2) as [string, string | undefined]\n    if (name !== 'run-kind' && name !== 'content') throw new Error(`unknown flag --${name}`)\n    const value = inline ?? argv[++index]\n    if (value === undefined) throw new Error(`--${name} needs a value`)\n    flags[name] = value\n  }\n  if (positional.length !== 1) throw new Error(USAGE)\n  return { path: positional[0]!, flags }\n}\n\n/** The ledger names its search on every line; read the first. */\nasync function firstLineSearchId(path: string): Promise<string> {\n  const handle = await open(path, 'r')\n  try {\n    // A line longer than one request's limit cannot ship anyway.\n    const bytes = Buffer.alloc(SEARCH_LEDGER_BATCH_MAX_BYTES + 1)\n    const { bytesRead } = await handle.read(bytes, 0, bytes.byteLength, 0)\n    const newline = bytes.subarray(0, bytesRead).indexOf(0x0a)\n    if (newline < 0) throw new Error(`${path} holds no complete ledger line`)\n    const first = JSON.parse(bytes.toString('utf8', 0, newline)) as { searchId?: unknown }\n    if (typeof first.searchId !== 'string') throw new Error(`${path} line 1 names no searchId`)\n    return first.searchId\n  } finally {\n    await handle.close()\n  }\n}\n","#!/usr/bin/env node\n/**\n * agent-eval CLI.\n *\n *   agent-eval serve [--port 5005] [--host 127.0.0.1]\n *   agent-eval rpc <method>          # one request from stdin → one response on stdout\n *   agent-eval rpc-batch <method>    # JSONL stdin → JSONL stdout\n *   agent-eval analyst-benchmark ... # benchmark a real-model trace analyst\n *   agent-eval supervisor-run report <runDir> [--format headline|markdown|json]\n *   agent-eval search ship <search-ledger.jsonl> --run-kind optimization|eval\n *   agent-eval search show <search-ledger.jsonl>\n *   agent-eval openapi [--out path]  # write OpenAPI spec\n *   agent-eval version\n *\n * <method> is one of: judge, listRubrics, version. When omitted, the\n * stdin payload must be a full {method, params} envelope.\n */\nimport { writeFileSync } from 'node:fs'\nimport { runAnalystBenchmarkCommand } from './analyst/benchmark-command'\nimport { resolveCliLlmConfig } from './cli-config'\nimport { runRolloutReleaseCli } from './rollout/release/hf-dataset'\nimport { runSearchCommand } from './search-command'\nimport { runSupervisorRunCommand } from './supervisor-run/report-command'\nimport { handleVersion } from './wire/handlers'\nimport { buildOpenApi } from './wire/openapi'\nimport { runRpcBatch, runRpcOnce } from './wire/rpc'\nimport { startServerAsync } from './wire/server'\n\ninterface Args {\n  command: string\n  positional: string[]\n  flags: Record<string, string>\n}\n\nfunction parseArgs(argv: string[]): Args {\n  const [command, ...rest] = argv\n  const positional: string[] = []\n  const flags: Record<string, string> = {}\n  for (let i = 0; i < rest.length; i++) {\n    const tok = rest[i]!\n    if (tok.startsWith('--')) {\n      const raw = tok.slice(2)\n      const equalsAt = raw.indexOf('=')\n      if (equalsAt >= 0) {\n        flags[raw.slice(0, equalsAt)] = raw.slice(equalsAt + 1)\n        continue\n      }\n      const key = raw\n      if (key === 'help') {\n        flags[key] = 'true'\n        continue\n      }\n      const next = rest[i + 1]\n      if (next != null && !next.startsWith('--')) {\n        flags[key] = next\n        i++\n      } else {\n        flags[key] = 'true'\n      }\n    } else if (tok === '-h') {\n      flags.help = 'true'\n    } else {\n      positional.push(tok)\n    }\n  }\n  return { command: command ?? 'help', positional, flags }\n}\n\nconst HELP = `agent-eval: evaluation RPC and HTTP server.\n\nCommands:\n  serve [--port 5005] [--host 127.0.0.1]\n        Start the HTTP server. POST /v1/judge, GET /v1/rubrics, GET /v1/version, GET /openapi.json.\n  rpc <method>\n        Read one JSON object from stdin (the params for <method>), write one\n        JSON object to stdout. Methods: judge, listRubrics, version.\n  rpc-batch <method>\n        Like 'rpc' but JSONL in / JSONL out.\n  openapi [--out openapi.json]\n        Write the OpenAPI 3.1 spec.\n  rollout-release <ledger.jsonl...> --out <dir> [--formats sft,verifiers,rft,raw] [--include-proposers] [--push <org/name>]\n        Build a HuggingFace-ready dataset dir from tangle.rollout.v1 ledgers:\n        validate, fail-closed split filter, scrub, export formats + card.\n  analyst-benchmark --dataset <agentrx|codetracebench> --labels <path> --trace-dir <path> [--artifact-dir <path>] --out <dir> ...\n        Compare an empty baseline with a real-model trace analyst on public labels.\n        Run with --help for all required inputs and controls.\n  supervisor-run report <runDir> [--format headline|markdown|json]\n        Report one Runtime or loops supervisor run directory. The status comes\n        from Runtime's own result.json / failure.json; every missing measurement\n        stays named. Exits 1 when the directory cannot be read.\n  search ship <search-ledger.jsonl> --run-kind optimization|eval [--content full|digests]\n        Ship a search ledger and its blobs to the hosted store in TANGLE_INGEST_URL,\n        starting from the store's head. Run with --help for the environment it reads.\n  search show <search-ledger.jsonl>\n        Verify a search ledger and print its summary: leading nodes, recently\n        discarded nodes and recent proposals, the same text a proposer reads.\n  version\n        Print server + wire-protocol version JSON.\n\nJudge provider:\n  Set AGENT_EVAL_LLM_BASE_URL, AGENT_EVAL_LLM_API_KEY, and AGENT_EVAL_LLM_MODEL.\n  OPENAI_* and TANGLE_* equivalents are also accepted. This binary is the only\n  place that reads a provider credential; the library never does. Without both\n  a base URL and a key, /v1/judge refuses with llm_not_configured.\n\nWithout arguments, prints this help.`\n\nasync function main(): Promise<number> {\n  const { command, positional, flags } = parseArgs(process.argv.slice(2))\n  assertKnownFlags(command, flags)\n\n  if (command === 'analyst-benchmark' && flags.help === 'true') {\n    return await runAnalystBenchmarkCommand(process.argv.slice(3))\n  }\n  if (command === 'search') return await runSearchCommand(process.argv.slice(3))\n  if (flags.help === 'true') {\n    process.stdout.write(`${HELP}\\n`)\n    return 0\n  }\n\n  switch (command) {\n    case 'serve': {\n      const port = parsePort(flags.port ?? '5005')\n      const host = flags.host ?? '127.0.0.1'\n      const llm = resolveCliLlmConfig()\n      const { server } = await startServerAsync({\n        port,\n        host,\n        ...(llm.chat ? { chat: llm.chat } : {}),\n        ...(llm.model ? { judgeModel: llm.model } : {}),\n      })\n      // Keep process alive on SIGINT/SIGTERM\n      const shutdown = (sig: string) => {\n        // eslint-disable-next-line no-console\n        console.log(`[agent-eval] received ${sig}, shutting down`)\n        server.close(() => process.exit(0))\n        // Force exit after 5s if close hangs\n        setTimeout(() => process.exit(1), 5000).unref()\n      }\n      process.on('SIGINT', () => shutdown('SIGINT'))\n      process.on('SIGTERM', () => shutdown('SIGTERM'))\n      // Block forever\n      await new Promise(() => {})\n      return 0\n    }\n    case 'rpc': {\n      const [method] = positional\n      const llm = resolveCliLlmConfig()\n      return await runRpcOnce(method, {\n        ...(llm.chat ? { chat: llm.chat } : {}),\n        ...(llm.model ? { judgeModel: llm.model } : {}),\n      })\n    }\n    case 'rpc-batch': {\n      const [method] = positional\n      const llm = resolveCliLlmConfig()\n      return await runRpcBatch(method, {\n        ...(llm.chat ? { chat: llm.chat } : {}),\n        ...(llm.model ? { judgeModel: llm.model } : {}),\n      })\n    }\n    case 'openapi': {\n      const out = flags.out ?? 'openapi.json'\n      const spec = buildOpenApi(handleVersion().version)\n      writeFileSync(out, `${JSON.stringify(spec, null, 2)}\\n`, 'utf-8')\n      // eslint-disable-next-line no-console\n      console.log(`[agent-eval] wrote OpenAPI 3.1 spec to ${out}`)\n      return 0\n    }\n    case 'rollout-release': {\n      // The subcommand owns its own flag grammar (multi-value --formats,\n      // boolean --include-proposers) — pass raw argv through untouched.\n      return await runRolloutReleaseCli(process.argv.slice(3))\n    }\n    case 'analyst-benchmark': {\n      return await runAnalystBenchmarkCommand(process.argv.slice(3))\n    }\n    case 'supervisor-run': {\n      // The subcommand owns its own flag grammar; pass raw argv through.\n      return await runSupervisorRunCommand(process.argv.slice(3))\n    }\n    case 'version': {\n      process.stdout.write(`${JSON.stringify(handleVersion(), null, 2)}\\n`)\n      return 0\n    }\n    case '--version': {\n      process.stdout.write(`${handleVersion().version}\\n`)\n      return 0\n    }\n    case 'help':\n    case '--help':\n    case '-h':\n    case '':\n      process.stdout.write(`${HELP}\\n`)\n      return 0\n    default:\n      process.stderr.write(`unknown command: ${command}\\n${HELP}\\n`)\n      return 1\n  }\n}\n\nconst FLAGS_BY_COMMAND: Record<string, ReadonlySet<string>> = {\n  serve: new Set(['help', 'host', 'port']),\n  rpc: new Set(['help']),\n  'rpc-batch': new Set(['help']),\n  openapi: new Set(['help', 'out']),\n  version: new Set(['help']),\n  help: new Set(),\n  '--help': new Set(),\n  '-h': new Set(),\n  '--version': new Set(),\n  '': new Set(),\n}\n\nfunction assertKnownFlags(command: string, flags: Record<string, string>): void {\n  // These subcommands parse their own argv.\n  if (\n    command === 'rollout-release' ||\n    command === 'analyst-benchmark' ||\n    command === 'supervisor-run' ||\n    command === 'search'\n  ) {\n    return\n  }\n  const allowed = FLAGS_BY_COMMAND[command]\n  if (!allowed) return\n  const unknown = Object.keys(flags).filter((flag) => !allowed.has(flag))\n  if (unknown.length > 0) {\n    throw new Error(`unknown flag for ${command || 'help'}: --${unknown[0]}`)\n  }\n}\n\nfunction parsePort(raw: string): number {\n  const port = Number(raw)\n  if (!Number.isInteger(port) || port < 0 || port > 65_535) {\n    throw new Error(`--port must be an integer from 0 to 65535; received ${JSON.stringify(raw)}`)\n  }\n  return port\n}\n\nmain()\n  .then((code) => {\n    process.exitCode = code\n  })\n  .catch((err) => {\n    // eslint-disable-next-line no-console\n    console.error('[agent-eval] cli error:', err)\n    process.exitCode = 1\n  })\n"],"mappings":";;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;AA8BA,SAAgB,wBACd,MAAyB,QAAQ,KACH;CAC9B,MAAM,kBAAkB,SAAS,IAAI,uBAAuB;CAC5D,MAAM,iBAAiB,SAAS,IAAI,sBAAsB;CAC1D,MAAM,eAAe,SAAS,IAAI,cAAc;CAChD,MAAM,eAAe,SAAS,IAAI,cAAc;CAChD,MAAM,UACJ,mBACA,SAAS,IAAI,eAAe,KAC5B,SAAS,IAAI,iBAAiB,MAC7B,eAAe,8BAA8B,KAAA,OAC7C,eAAe,mCAAmC,KAAA;CACrD,MAAM,SAAS,kBAAkB,gBAAgB;CACjD,MAAM,QACJ,SAAS,IAAI,oBAAoB,KAAK,SAAS,IAAI,YAAY,KAAK,SAAS,IAAI,YAAY;CAE/F,IAAI,CAAC,WAAW,CAAC,QAAQ,OAAO,KAAA;CAChC,OAAO;EAAE;EAAS;EAAQ,GAAI,QAAQ,EAAE,MAAM,IAAI,CAAC;CAAG;AACxD;AAEA,SAAgB,oBAAoB,MAAyB,QAAQ,KAAmB;CACtF,MAAM,QAAQ,wBAAwB,GAAG;CACzC,IAAI,CAAC,OAAO,OAAO,CAAC;CACpB,OAAO;EACL,MAAM,cAAc;GAAE,SAAS,MAAM;GAAS,QAAQ,MAAM;EAAO,GAAG,MAAM,KAAK;EACjF,GAAI,MAAM,QAAQ,EAAE,OAAO,MAAM,MAAM,IAAI,CAAC;CAC9C;AACF;;;;;;;;;AAUA,SAAS,cACP,OACA,cACY;CACZ,MAAM,SAAS,iBAAiB;EAC9B,WAAW;EACX,SAAS,MAAM;EACf,QAAQ,MAAM;EACd,GAAI,eAAe,EAAE,aAAa,IAAI,CAAC;CACzC,CAAC;CACD,OAAO;EACL,GAAG;EACH,OAAO,KAAK,aAAa;GACvB,IAAI,CAAC,IAAI,SAAS,CAAC,cACjB,MAAM,IAAI,MACR,4EACF;GAEF,OAAO,OAAO,KAAK,KAAK,QAAQ;EAClC;CACF;AACF;AAEA,SAAS,SAAS,OAA+C;CAC/D,MAAM,UAAU,OAAO,KAAK;CAC5B,OAAO,UAAU,UAAU,KAAA;AAC7B;;;;;;;;;;;;;;;;;;;;;;;;;;;;AC9CA,MAAM,QAAQ;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;AA0Cd,eAAsB,iBAAiB,MAAiC;CACtE,MAAM,CAAC,YAAY,GAAG,QAAQ;CAC9B,IAAI,eAAe,KAAA,KAAa,eAAe,YAAY,eAAe,MAAM;EAC9E,QAAQ,OAAO,MAAM,GAAG,MAAM,GAAG;EACjC,OAAO;CACT;CACA,IAAI,eAAe,QAAQ,OAAO,MAAM,eAAe,IAAI;CAC3D,IAAI,eAAe,QAAQ;EACzB,QAAQ,OAAO,MAAM,8BAA8B,WAAW,IAAI,MAAM,GAAG;EAC3E,OAAO;CACT;CACA,IAAI,KAAK,MAAM,QAAQ,QAAQ,KAAA,KAAa,QAAQ,YAAY,QAAQ,IAAI,GAAG;EAC7E,QAAQ,OAAO,MAAM,GAAG,MAAM,GAAG;EACjC,OAAO;CACT;CACA,MAAM,EAAE,MAAM,UAAU,cAAc,IAAI;CAC1C,MAAM,UAAU,oBAAoB,UAAU,MAAM,WAAW;CAC/D,IAAI,CAAC,QAAQ,SAAS,MAAM,IAAI,MAAM,4CAA4C,OAAO;CACzF,MAAM,UAAU,MAAM,WAAW;CACjC,IAAI,YAAY,UAAU,YAAY,WACpC,MAAM,IAAI,MAAM,sCAAsC,OAAO;CAE/D,MAAM,SAAS,oBAAoB;CACnC,IAAI,CAAC,QACH,MAAM,IAAI,MACR,qFACF;CAEF,MAAM,WAAW,MAAM,kBAAkB,IAAI;CAC7C,IAAI;EACF,MAAM,UAAU,MAAM,iBAAiB;GACrC;GACA,QAAQ;IAAE;IAAM;GAAS;GACzB,SAAS,QAAQ;GACjB;EACF,CAAC;EACD,QAAQ,OAAO,MAAM,GAAG,KAAK,UAAU,SAAS,MAAM,CAAC,EAAE,GAAG;EAC5D,OAAO,QAAQ,KAAK,iBAAiB,QAAQ,aAAa,IAAI;CAChE,SAAS,OAAO;EACd,IAAI,EAAE,iBAAiB,0BAA0B,MAAM;EACvD,QAAQ,OAAO,MAAM,GAAG,MAAM,QAAQ,GAAG;EACzC,OAAO;CACT;AACF;AAEA,eAAe,eAAe,MAAiC;CAC7D,IAAI,KAAK,SAAS,QAAQ,KAAK,KAAK,SAAS,IAAI,GAAG;EAClD,QAAQ,OAAO,MAAM,GAAG,MAAM,GAAG;EACjC,OAAO;CACT;CACA,IAAI,KAAK,SAAS,QAAQ,GAAG,OAAO,MAAM,mBAAmB,IAAI;CACjE,MAAM,aAAa,KAAK,QAAQ,QAAQ,CAAC,IAAI,WAAW,IAAI,CAAC;CAC7D,MAAM,QAAQ,IAAI,IAAI,KAAK,QAAQ,QAAQ,IAAI,WAAW,IAAI,CAAC,CAAC;CAChE,MAAM,UAAU,CAAC,GAAG,KAAK,CAAC,CAAC,QAAQ,SAAS,CAAC,WAAW,IAAI,IAAI,CAAC;CACjE,IAAI,WAAW,WAAW,KAAK,QAAQ,SAAS,GAAG;EACjD,IAAI,QAAQ,SAAS,GAAG,QAAQ,OAAO,MAAM,gBAAgB,QAAQ,KAAK,IAAI,EAAE,GAAG;EACnF,QAAQ,OAAO,MAAM,GAAG,MAAM,GAAG;EACjC,OAAO;CACT;CACA,MAAM,OAAO,WAAW;CACxB,MAAM,WAAW,MAAM,kBAAkB,IAAI;CAC7C,MAAM,OAAO,MAAM,SAAS,MAAM,MAAM;CACxC,MAAM,QAAQ,uBAAuB,MAAM,UAAU,IAAI;CACzD,MAAM,QACJ,MAAM,UAAU,MAAM,OAAO,OAAO,UAAU,MAAM,SAAS,IAAI,cAAc;CACjF,MAAM,SAAS,WAAW,OAAO,MAAM,KAAK;CAC5C,IAAI,MAAM,IAAI,QAAQ,GAAG;EACvB,IAAI,OAAO,WAAW,GAAG;GACvB,QAAQ,OAAO,MAAM,4BAA4B;GACjD,OAAO;EACT;EACA,MAAM,OAAO,OAAO,YAAY,OAAO,KAAK,SAAS,CAAC,KAAK,MAAM,KAAK,MAAM,CAAC,CAAC;EAC9E,QAAQ,OAAO,MAAM,GAAG,KAAK,UAAU,MAAM,MAAM,CAAC,EAAE,GAAG;EACzD,OAAO;CACT;CACA,MAAM,WAAW,CAAC,oBAAoB,OAAO,EAAE,MAAM,CAAC,GAAG,GAAG,OAAO,KAAK,SAAS,KAAK,IAAI,CAAC;CAC3F,QAAQ,OAAO,MAAM,GAAG,SAAS,KAAK,MAAM,EAAE,GAAG;CACjD,OAAO;AACT;AAEA,eAAe,mBAAmB,MAAiC;CACjE,MAAM,QAAkB,CAAC;CACzB,IAAI;CACJ,IAAI,OAAO;CACX,KAAK,IAAI,QAAQ,GAAG,QAAQ,KAAK,QAAQ,SAAS;EAChD,MAAM,QAAQ,KAAK;EACnB,IAAI,UAAU,UAAU;EACxB,IAAI,UAAU,UAAU;GACtB,OAAO;GACP;EACF;EACA,IAAI,UAAU,eAAe;GAC3B,YAAY,KAAK,EAAE;GACnB,IAAI,cAAc,KAAA,GAAW,MAAM,IAAI,MAAM,8BAA8B,OAAO;GAClF;EACF;EACA,IAAI,MAAM,WAAW,IAAI,GAAG,MAAM,IAAI,MAAM,gBAAgB,MAAM,IAAI,OAAO;EAC7E,MAAM,KAAK,KAAK;CAClB;CACA,IAAI,MAAM,WAAW,GAAG,MAAM,IAAI,MAAM,0CAA0C,OAAO;CACzF,MAAM,SAAS,CAAC;CAChB,KAAK,MAAM,QAAQ,OAAO;EACxB,MAAM,WAAW,MAAM,kBAAkB,IAAI;EAC7C,OAAO,KAAK,uBAAuB,MAAM,SAAS,MAAM,MAAM,GAAG,UAAU,IAAI,CAAC;CAClF;CACA,MAAM,OAAO,WAAW,QAAQ,cAAc,KAAA,IAAY,CAAC,IAAI,EAAE,UAAU,CAAC;CAC5E,QAAQ,OAAO,MACb,OAAO,GAAG,KAAK,UAAU,EAAE,YAAY,KAAK,GAAG,MAAM,CAAC,EAAE,MAAM,GAAG,qBAAqB,IAAI,EAAE,GAC9F;CACA,OAAO;AACT;AAEA,MAAM,6BAAa,IAAI,IAAI;CACzB;CACA;CACA;CACA;CACA;CACA;CACA;CACA;AACF,CAAC;AAED,SAAS,WACP,OACA,YACA,OACwD;CACxD,MAAM,SAAiE,CAAC;CACxE,IAAI,MAAM,IAAI,QAAQ,GAAG;EACvB,MAAM,SAAS,KAAK,KAAK;EACzB,OAAO,KAAK;GAAE,MAAM;GAAQ;GAAQ,MAAM,eAAe,OAAO,IAAI;EAAE,CAAC;CACzE;CACA,IAAI,MAAM,IAAI,kBAAkB,GAAG;EACjC,MAAM,SAAS,cAAc,KAAK;EAClC,OAAO,KAAK;GAAE,MAAM;GAAiB;GAAQ,MAAM,wBAAwB,OAAO,IAAI;EAAE,CAAC;CAC3F;CACA,IAAI,MAAM,IAAI,SAAS,GAAG;EACxB,MAAM,SAAS,MAAM,KAAK;EAC1B,OAAO,KAAK;GAAE,MAAM;GAAS;GAAQ,MAAM,gBAAgB,OAAO,IAAI;EAAE,CAAC;CAC3E;CACA,IAAI,MAAM,IAAI,eAAe,GAAG;EAC9B,MAAM,SAAS,WAAW,KAAK;EAC/B,OAAO,KAAK;GAAE,MAAM;GAAc;GAAQ,MAAM,qBAAqB,OAAO,IAAI;EAAE,CAAC;CACrF;CACA,IAAI,MAAM,IAAI,eAAe,GAAG;EAC9B,MAAM,SAAS,WAAW,OAAO,EAAE,cAAc,iBAAiB,UAAU,EAAE,CAAC;EAC/E,OAAO,KAAK;GAAE,MAAM;GAAc;GAAQ,MAAM,eAAe,MAAM;EAAE,CAAC;CAC1E;CACA,IAAI,MAAM,IAAI,aAAa,GAAG;EAC5B,MAAM,OAAO,iBAAiB,UAAU;EACxC,MAAM,SAAS,UACb,OACA,kBAAkB,YAAY,KAAK,QAAQ,QAAQ,KAAK,IAAI,CAC9D;EACA,OAAO,KAAK;GAAE,MAAM;GAAa;GAAQ,MAAM,gBAAgB,MAAM;EAAE,CAAC;CAC1E;CACA,IAAI,MAAM,IAAI,kBAAkB,GAAG;EACjC,MAAM,SAAS,cAAc,KAAK;EAClC,OAAO,KAAK;GAAE,MAAM;GAAiB;GAAQ,MAAM,oBAAoB,MAAM;EAAE,CAAC;CAClF;CACA,OAAO;AACT;;;;;;AAOA,SAAS,iBAAiB,YAAqE;CAC7F,MAAM,wBAAQ,IAAI,IAAqB;CACvC,QAAQ,QAAQ;EACd,IAAI,MAAM,IAAI,IAAI,MAAM,GAAG,OAAO,MAAM,IAAI,IAAI,MAAM;EACtD,MAAM,MAAM,IAAI,OAAO,MAAM,CAAgB;EAC7C,MAAM,aAAa,CAAC,KAAK,QAAQ,UAAU,GAAG,SAAS,GAAG,IAAI,MAAM,CAAC;EACrE,IAAI,IAAI,IAAI,WAAW,OAAO,GAAG,WAAW,QAAQ,cAAc,IAAI,GAAG,CAAC;EAC1E,IAAI;EACJ,KAAK,MAAM,aAAa,YAAY;GAClC,IAAI,CAAC,WAAW,SAAS,GAAG;GAC5B,MAAM,QAAQ,aAAa,SAAS;GAEpC,IADe,WAAW,QAAQ,CAAC,CAAC,OAAO,KAAK,CAAC,CAAC,OAAO,KAChD,MAAM,OAAO,MAAM,eAAe,IAAI,YAAY;GAC3D,IAAI;IACF,QAAQ,KAAK,MAAM,MAAM,SAAS,MAAM,CAAC;GAC3C,QAAQ,CAER;GACA;EACF;EACA,MAAM,IAAI,IAAI,QAAQ,KAAK;EAC3B,OAAO;CACT;AACF;AAEA,SAAS,eAAe,MAAwB;CAC9C,MAAM,QAAQ,CACZ,SAAS,KAAK,UAAU,OAAO,KAAK,cAAc,IAAI,KAAK,IAAI,IAAI,KAAK,UAAU,OAAO,KAAK,cAAc,IAAI,KAAK,KACvH;CACA,MAAM,QAAQ,MAAgB,QAAgB,WAA0B;EACtE,MAAM,SAAS,WAAW,KAAK,KAAK,SAAS,QAAQ;EACrD,MAAM,KAAK,KAAK,WAAW,GAAG,KAAK,SAAS,OAAO;EACnD,MAAM,OAAO,KAAK,SAAS,OAAO,KAAK,SAAS,KAAK;EACrD,MAAM,QAAQ,KAAK,UAAU,OAAO,mBAAmB;EACvD,MAAM,OACJ,KAAK,qBAAqB,IACtB,IAAI,KAAK,aAAa,QAAQ,CAAC,MAC/B,IAAI,KAAK,aAAa,QAAQ,CAAC,EAAE,qBAAqB,KAAK,SAAS,QAAQ,CAAC,EAAE,QAAQ,KAAK,iBAAiB,oBAAoB,KAAK,qBAAqB,IAAI,KAAK;EAC1K,MAAM,KACJ,GAAG,SAAS,SAAS,KAAK,KAAK,OAAO,IAAI,KAAK,UAAU,cAAc,OAAO,MAAM,IAAI,MAC1F;EACA,MAAM,cAAc,UAAU,WAAW,KAAK,KAAK,SAAS,QAAQ;EACpE,KAAK,SAAS,SAAS,OAAO,UAAU;GACtC,KAAK,OAAO,aAAa,UAAU,KAAK,SAAS,SAAS,CAAC;EAC7D,CAAC;CACH;CACA,KAAK,MAAM,SAAS,MAAM,UAAU;EAClC,KAAK,MAAM,IAAI,UAAU,KAAK,MAAM,SAAS,CAAC;CAChD,CAAC;CACD,OAAO,MAAM,KAAK,IAAI;AACxB;AAEA,SAAS,wBAAwB,MAAiC;CAChE,MAAM,QAAQ,CAAC,mBAAmB,KAAK,MAAM,gDAAgD;CAC7F,IAAI,KAAK,KAAK,WAAW,GAAG,MAAM,KAAK,gBAAgB;CACvD,KAAK,MAAM,OAAO,KAAK,MAAM;EAC3B,MAAM,IAAI,IAAI;EACd,MAAM,WAAW,OAAO,QAAQ,IAAI,QAAQ,CAAC,CAC1C,QAAQ,GAAG,WAAW,QAAQ,CAAC,CAAC,CAChC,KAAK,CAAC,QAAQ,WAAW,GAAG,OAAO,GAAG,OAAO,CAAC,CAC9C,KAAK,GAAG;EACX,MAAM,YACJ,EAAE,WAAW,SACT,EAAE,MAAM,IACN,gCACA,kBAAkB,EAAE,EAAE,MAAM,kBAAkB,WAChD,EAAE,WAAW,iBACX,SAAS,EAAE,KAAM,QAAQ,CAAC,EAAE,MAAM,EAAE,EAAE,MAAM,kBAAkB,mBAC9D,SAAS,EAAE,KAAM,QAAQ,CAAC,EAAE,MAAM,EAAE,SAAU,EAAE,CAAC,QAAQ,CAAC,EAAE,IAAI,EAAE,SAAU,EAAE,CAAC,QAAQ,CAAC,EAAE,KAAK,EAAE,OAAO,IAAI,EAAE,eAAe,MAAM,EAAE,EAAE;EAC/I,MAAM,QAAQ,OAAO,OAAO,IAAI,QAAQ,CAAC,CAAC,QAAQ,KAAK,UAAU,MAAM,OAAO,CAAC;EAC/E,MAAM,cAAc,IAAI,cAAc,IAAI,KAAK,IAAI,YAAY,iBAAiB;EAChF,MAAM,KACJ,KAAK,IAAI,SAAS,IAAI,IAAI,UAAU,YAAY,YAAY,KAAK,MAAM,OAAO,UAAU,IAAI,KAAK,IAAI,IAAI,YAAY,mBAAmB,KAAK,UAAU,IAAI,IAAI,SAAS,UAC1K;CACF;CACA,OAAO,MAAM,KAAK,IAAI;AACxB;AAEA,SAAS,gBAAgB,MAAyB;CAChD,MAAM,UAAU,KAAK,KAAK,QAAQ,QAAQ,IAAI,OAAO;CACrD,MAAM,UAAU,OAAO,QAAQ,KAAK,UAAU,CAAC,CAC5C,QAAQ,GAAG,WAAW,QAAQ,CAAC,CAAC,CAChC,KAAK,CAAC,QAAQ,WAAW,GAAG,OAAO,GAAG,OAAO,CAAC,CAC9C,KAAK,IAAI;CACZ,MAAM,QAAQ,CACZ,UAAU,KAAK,MAAM,gBAAgB,KAAK,KAAK,KAAK,IAAI,EAAE,KAAK,KAAK,aAAa,MAAM,KAAK,KAAK,OAAO,4BAA4B,KAAK,cAAc,WAAW,UAAU,KAAK,QAAQ,KAAK,IAChM;CACA,KAAK,MAAM,OAAO,SAChB,MAAM,KACJ,KAAK,IAAI,OAAO,UAAU,IAAI,MAAO,QAAQ,CAAC,EAAE,MAAM,IAAI,MAAM,UAAU,IAAI,OAAO,iBAAiB,IAAI,eAAgB,QAAQ,CAAC,GACrI;CAEF,OAAO,MAAM,KAAK,IAAI;AACxB;AAEA,SAAS,qBAAqB,MAA8B;CAC1D,MAAM,QAAQ,CACZ,gBAAgB,KAAK,MAAM,UAAU,KAAK,UAAU,KAAK,KAAK,QAAQ,OAAO,aAAa,KAAK,QAAQ,OAAO,YAAY,KAAK,aAAa,OAAO,oBAAoB,KAAK,aAAa,OAAO,iBAClM;CACA,IAAI,KAAK,UACP,MAAM,KACJ,eAAe,KAAK,SAAS,eAAe,MAAM,KAAK,SAAS,WAAW,cAAc,KAAK,SAAS,IAAI,+FAC7G;CAEF,KAAK,MAAM,OAAO,KAAK,gBAAgB;EACrC,MAAM,OACJ,IAAI,SAAS,OACT,iBAAiB,IAAI,aAAa,KAClC,GAAG,IAAI,KAAK,QAAQ,CAAC,EAAE,SAAS,IAAI,WAAW,IAAI,IAAI,kBAAkB,YAAY,IAAI,QAAQ,OAAO;EAC9G,MAAM,KAAK,aAAa,IAAI,UAAU,IAAI,IAAI,QAAQ,KAAK,IAAI,EAAE,qBAAqB,MAAM;CAC9F;CACA,OAAO,MAAM,KAAK,IAAI;AACxB;AAEA,SAAS,cAAc,MAAiE;CACtF,MAAM,QAAgC,CAAC;CACvC,MAAM,aAAuB,CAAC;CAC9B,KAAK,IAAI,QAAQ,GAAG,QAAQ,KAAK,QAAQ,SAAS;EAChD,MAAM,QAAQ,KAAK;EACnB,IAAI,CAAC,MAAM,WAAW,IAAI,GAAG;GAC3B,WAAW,KAAK,KAAK;GACrB;EACF;EACA,MAAM,CAAC,MAAM,UAAU,MAAM,MAAM,CAAC,CAAC,CAAC,MAAM,KAAK,CAAC;EAClD,IAAI,SAAS,cAAc,SAAS,WAAW,MAAM,IAAI,MAAM,kBAAkB,MAAM;EACvF,MAAM,QAAQ,UAAU,KAAK,EAAE;EAC/B,IAAI,UAAU,KAAA,GAAW,MAAM,IAAI,MAAM,KAAK,KAAK,eAAe;EAClE,MAAM,QAAQ;CAChB;CACA,IAAI,WAAW,WAAW,GAAG,MAAM,IAAI,MAAM,KAAK;CAClD,OAAO;EAAE,MAAM,WAAW;EAAK;CAAM;AACvC;;AAGA,eAAe,kBAAkB,MAA+B;CAC9D,MAAM,SAAS,MAAM,KAAK,MAAM,GAAG;CACnC,IAAI;EAEF,MAAM,QAAQ,OAAO,MAAM,gCAAgC,CAAC;EAC5D,MAAM,EAAE,cAAc,MAAM,OAAO,KAAK,OAAO,GAAG,MAAM,YAAY,CAAC;EACrE,MAAM,UAAU,MAAM,SAAS,GAAG,SAAS,CAAC,CAAC,QAAQ,EAAI;EACzD,IAAI,UAAU,GAAG,MAAM,IAAI,MAAM,GAAG,KAAK,+BAA+B;EACxE,MAAM,QAAQ,KAAK,MAAM,MAAM,SAAS,QAAQ,GAAG,OAAO,CAAC;EAC3D,IAAI,OAAO,MAAM,aAAa,UAAU,MAAM,IAAI,MAAM,GAAG,KAAK,0BAA0B;EAC1F,OAAO,MAAM;CACf,UAAU;EACR,MAAM,OAAO,MAAM;CACrB;AACF;;;;;;;;;;;;;;;;;;;ACrXA,SAAS,UAAU,MAAsB;CACvC,MAAM,CAAC,SAAS,GAAG,QAAQ;CAC3B,MAAM,aAAuB,CAAC;CAC9B,MAAM,QAAgC,CAAC;CACvC,KAAK,IAAI,IAAI,GAAG,IAAI,KAAK,QAAQ,KAAK;EACpC,MAAM,MAAM,KAAK;EACjB,IAAI,IAAI,WAAW,IAAI,GAAG;GACxB,MAAM,MAAM,IAAI,MAAM,CAAC;GACvB,MAAM,WAAW,IAAI,QAAQ,GAAG;GAChC,IAAI,YAAY,GAAG;IACjB,MAAM,IAAI,MAAM,GAAG,QAAQ,KAAK,IAAI,MAAM,WAAW,CAAC;IACtD;GACF;GACA,MAAM,MAAM;GACZ,IAAI,QAAQ,QAAQ;IAClB,MAAM,OAAO;IACb;GACF;GACA,MAAM,OAAO,KAAK,IAAI;GACtB,IAAI,QAAQ,QAAQ,CAAC,KAAK,WAAW,IAAI,GAAG;IAC1C,MAAM,OAAO;IACb;GACF,OACE,MAAM,OAAO;EAEjB,OAAO,IAAI,QAAQ,MACjB,MAAM,OAAO;OAEb,WAAW,KAAK,GAAG;CAEvB;CACA,OAAO;EAAE,SAAS,WAAW;EAAQ;EAAY;CAAM;AACzD;AAEA,MAAM,OAAO;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;AAuCb,eAAe,OAAwB;CACrC,MAAM,EAAE,SAAS,YAAY,UAAU,UAAU,QAAQ,KAAK,MAAM,CAAC,CAAC;CACtE,iBAAiB,SAAS,KAAK;CAE/B,IAAI,YAAY,uBAAuB,MAAM,SAAS,QACpD,OAAO,MAAM,2BAA2B,QAAQ,KAAK,MAAM,CAAC,CAAC;CAE/D,IAAI,YAAY,UAAU,OAAO,MAAM,iBAAiB,QAAQ,KAAK,MAAM,CAAC,CAAC;CAC7E,IAAI,MAAM,SAAS,QAAQ;EACzB,QAAQ,OAAO,MAAM,GAAG,KAAK,GAAG;EAChC,OAAO;CACT;CAEA,QAAQ,SAAR;EACE,KAAK,SAAS;GACZ,MAAM,OAAO,UAAU,MAAM,QAAQ,MAAM;GAC3C,MAAM,OAAO,MAAM,QAAQ;GAC3B,MAAM,MAAM,oBAAoB;GAChC,MAAM,EAAE,WAAW,MAAM,iBAAiB;IACxC;IACA;IACA,GAAI,IAAI,OAAO,EAAE,MAAM,IAAI,KAAK,IAAI,CAAC;IACrC,GAAI,IAAI,QAAQ,EAAE,YAAY,IAAI,MAAM,IAAI,CAAC;GAC/C,CAAC;GAED,MAAM,YAAY,QAAgB;IAEhC,QAAQ,IAAI,yBAAyB,IAAI,gBAAgB;IACzD,OAAO,YAAY,QAAQ,KAAK,CAAC,CAAC;IAElC,iBAAiB,QAAQ,KAAK,CAAC,GAAG,GAAI,CAAC,CAAC,MAAM;GAChD;GACA,QAAQ,GAAG,gBAAgB,SAAS,QAAQ,CAAC;GAC7C,QAAQ,GAAG,iBAAiB,SAAS,SAAS,CAAC;GAE/C,MAAM,IAAI,cAAc,CAAC,CAAC;GAC1B,OAAO;EACT;EACA,KAAK,OAAO;GACV,MAAM,CAAC,UAAU;GACjB,MAAM,MAAM,oBAAoB;GAChC,OAAO,MAAM,WAAW,QAAQ;IAC9B,GAAI,IAAI,OAAO,EAAE,MAAM,IAAI,KAAK,IAAI,CAAC;IACrC,GAAI,IAAI,QAAQ,EAAE,YAAY,IAAI,MAAM,IAAI,CAAC;GAC/C,CAAC;EACH;EACA,KAAK,aAAa;GAChB,MAAM,CAAC,UAAU;GACjB,MAAM,MAAM,oBAAoB;GAChC,OAAO,MAAM,YAAY,QAAQ;IAC/B,GAAI,IAAI,OAAO,EAAE,MAAM,IAAI,KAAK,IAAI,CAAC;IACrC,GAAI,IAAI,QAAQ,EAAE,YAAY,IAAI,MAAM,IAAI,CAAC;GAC/C,CAAC;EACH;EACA,KAAK,WAAW;GACd,MAAM,MAAM,MAAM,OAAO;GACzB,MAAM,OAAO,aAAa,cAAc,CAAC,CAAC,OAAO;GACjD,cAAc,KAAK,GAAG,KAAK,UAAU,MAAM,MAAM,CAAC,EAAE,KAAK,OAAO;GAEhE,QAAQ,IAAI,0CAA0C,KAAK;GAC3D,OAAO;EACT;EACA,KAAK,mBAGH,OAAO,MAAM,qBAAqB,QAAQ,KAAK,MAAM,CAAC,CAAC;EAEzD,KAAK,qBACH,OAAO,MAAM,2BAA2B,QAAQ,KAAK,MAAM,CAAC,CAAC;EAE/D,KAAK,kBAEH,OAAO,MAAM,wBAAwB,QAAQ,KAAK,MAAM,CAAC,CAAC;EAE5D,KAAK;GACH,QAAQ,OAAO,MAAM,GAAG,KAAK,UAAU,cAAc,GAAG,MAAM,CAAC,EAAE,GAAG;GACpE,OAAO;EAET,KAAK;GACH,QAAQ,OAAO,MAAM,GAAG,cAAc,CAAC,CAAC,QAAQ,GAAG;GACnD,OAAO;EAET,KAAK;EACL,KAAK;EACL,KAAK;EACL,KAAK;GACH,QAAQ,OAAO,MAAM,GAAG,KAAK,GAAG;GAChC,OAAO;EACT;GACE,QAAQ,OAAO,MAAM,oBAAoB,QAAQ,IAAI,KAAK,GAAG;GAC7D,OAAO;CACX;AACF;AAEA,MAAM,mBAAwD;CAC5D,uBAAO,IAAI,IAAI;EAAC;EAAQ;EAAQ;CAAM,CAAC;CACvC,qBAAK,IAAI,IAAI,CAAC,MAAM,CAAC;CACrB,6BAAa,IAAI,IAAI,CAAC,MAAM,CAAC;CAC7B,yBAAS,IAAI,IAAI,CAAC,QAAQ,KAAK,CAAC;CAChC,yBAAS,IAAI,IAAI,CAAC,MAAM,CAAC;CACzB,sBAAM,IAAI,IAAI;CACd,0BAAU,IAAI,IAAI;CAClB,sBAAM,IAAI,IAAI;CACd,6BAAa,IAAI,IAAI;CACrB,oBAAI,IAAI,IAAI;AACd;AAEA,SAAS,iBAAiB,SAAiB,OAAqC;CAE9E,IACE,YAAY,qBACZ,YAAY,uBACZ,YAAY,oBACZ,YAAY,UAEZ;CAEF,MAAM,UAAU,iBAAiB;CACjC,IAAI,CAAC,SAAS;CACd,MAAM,UAAU,OAAO,KAAK,KAAK,CAAC,CAAC,QAAQ,SAAS,CAAC,QAAQ,IAAI,IAAI,CAAC;CACtE,IAAI,QAAQ,SAAS,GACnB,MAAM,IAAI,MAAM,oBAAoB,WAAW,OAAO,MAAM,QAAQ,IAAI;AAE5E;AAEA,SAAS,UAAU,KAAqB;CACtC,MAAM,OAAO,OAAO,GAAG;CACvB,IAAI,CAAC,OAAO,UAAU,IAAI,KAAK,OAAO,KAAK,OAAO,OAChD,MAAM,IAAI,MAAM,uDAAuD,KAAK,UAAU,GAAG,GAAG;CAE9F,OAAO;AACT;AAEA,KAAK,CAAC,CACH,MAAM,SAAS;CACd,QAAQ,WAAW;AACrB,CAAC,CAAC,CACD,OAAO,QAAQ;CAEd,QAAQ,MAAM,2BAA2B,GAAG;CAC5C,QAAQ,WAAW;AACrB,CAAC"}

package/dist/contract/index.d.ts

@@ -3,12 +3,12 @@ import { _ as ProposalFinding, i as AnalystFinding, v as ProposalFindingOrigin,
 import { A as createChatClient, T as CreateChatClientOpts, x as ChatClient } from "../types-DLxe96rp.js";
 import { $ as paretoSignificanceGate, G as BuildEvidenceVectorOptions, J as ParetoSignificanceGateOptions, K as EvidenceVector, Q as paretoPolicy, U as AxisEvidence, W as AxisVerdict, X as PromotionPolicy, Y as PromotionObjective, Z as buildEvidenceVector, q as ObjectiveSource } from "../statistical-heldout-COfgpWah.js";
 import { C as JudgeDimension, H as SurfaceProposer, M as OptimizerConfig, R as Scenario, S as JudgeConfig, V as SessionScript, _ as GateDecision, a as CampaignResult, b as GenerationRecord, c as CampaignTraceWriter, d as DispatchContext, f as DispatchFn, g as GateContribution, h as GateContext, i as CampaignCostMeter, j as MutableSurface, l as CodeSurface, m as GateCheckStatus, n as CampaignArtifactWriter, p as Gate, r as CampaignCellResult, t as CampaignAggregates, v as GateResult, w as JudgeScore, y as GenerationCandidate } from "../types-ZUy5HGY5.js";
-import { $i as RunImprovementLoopOptions, $r as GepaOptimizationRecipe, Bi as ReferenceEquivalenceJudgeInput, Ci as compareOptimizationMethods, Dr as DefaultProductionGateOptions, Er as DefaultProductionGateCheck, Fa as campaignSplitDigest, Gi as runReferenceEquivalenceJudge, Gr as SkillOptOptimizationMethodConfig, Hi as ReferenceEquivalenceJudgeResult, Ji as llmJudge, Jr as skillOptOptimizationMethod, Ki as LlmJudgeDimension, Kr as SkillOptRunnerCommand, Oa as FinalEvidencePolicy, Or as DefaultProductionRewardHackingOptions, Qi as transientDispatchFailure, Qr as GepaOptimizationMethodConfig, Ri as REFERENCE_EQUIVALENCE_INPUT_LIMITS, Si as OptimizationTokenUsage, Ti as ComparisonCost, Tr as heldOutGate, Ui as ReferenceEquivalenceScenario, Vi as ReferenceEquivalenceJudgeOptions, Wi as createReferenceEquivalenceJudge, Wr as composeGate, Xr as GepaEngineOptions, Yr as GepaAdaptiveEngineRun, Zr as GepaEngineRun, _a as runCampaign, _i as OptimizationMethodProvenance, ai as ExternalTextOptimizationMethodConfig, ci as ExternalOptimizationExample, di as CompareOptimizationMethodsOptions, ea as RunImprovementLoopResult, ei as GepaRunnerCommand, fa as RunEvalOptions, fi as OptimizationMethod, ga as RunCampaignOptions, ha as CampaignCellRetryPolicy, hi as OptimizationMethodInput, ii as externalTextOptimizationMethod, ka as FinalEvidenceUse, kr as defaultProductionGate, li as ExternalTextEvaluationResponse, ma as CampaignCellFailureReceipt, ni as OpenAICompatibleOptimizerModel, oi as ExternalTextOptimizerContext, pa as runEval, pi as OptimizationMethodComparison, qi as LlmJudgeOptions, qr as SkillOptTrainerConfig, ri as OptimizerModelBudget, si as ExternalTextOptimizerResult, ta as runImprovementLoop, ti as gepaOptimizationMethod, vi as OptimizationMethodResult, wr as HeldOutGateOptions, xi as OptimizationPackageSource, zi as REFERENCE_EQUIVALENCE_JUDGE_VERSION } from "../index-BRdZ0oJo.js";
-import { Qt as inMemoryCampaignStorage, Yt as CampaignStorage, Zt as fsCampaignStorage } from "../search-kernel-DNpsHb8P.js";
+import { $i as RunImprovementLoopOptions, $r as GepaOptimizationRecipe, Bi as ReferenceEquivalenceJudgeInput, Ci as compareOptimizationMethods, Dr as DefaultProductionGateOptions, Er as DefaultProductionGateCheck, Fa as campaignSplitDigest, Gi as runReferenceEquivalenceJudge, Gr as SkillOptOptimizationMethodConfig, Hi as ReferenceEquivalenceJudgeResult, Ji as llmJudge, Jr as skillOptOptimizationMethod, Ki as LlmJudgeDimension, Kr as SkillOptRunnerCommand, Oa as FinalEvidencePolicy, Or as DefaultProductionRewardHackingOptions, Qi as transientDispatchFailure, Qr as GepaOptimizationMethodConfig, Ri as REFERENCE_EQUIVALENCE_INPUT_LIMITS, Si as OptimizationTokenUsage, Ti as ComparisonCost, Tr as heldOutGate, Ui as ReferenceEquivalenceScenario, Vi as ReferenceEquivalenceJudgeOptions, Wi as createReferenceEquivalenceJudge, Wr as composeGate, Xr as GepaEngineOptions, Yr as GepaAdaptiveEngineRun, Zr as GepaEngineRun, _a as runCampaign, _i as OptimizationMethodProvenance, ai as ExternalTextOptimizationMethodConfig, ci as ExternalOptimizationExample, di as CompareOptimizationMethodsOptions, ea as RunImprovementLoopResult, ei as GepaRunnerCommand, fa as RunEvalOptions, fi as OptimizationMethod, ga as RunCampaignOptions, ha as CampaignCellRetryPolicy, hi as OptimizationMethodInput, ii as externalTextOptimizationMethod, ka as FinalEvidenceUse, kr as defaultProductionGate, li as ExternalTextEvaluationResponse, ma as CampaignCellFailureReceipt, ni as OpenAICompatibleOptimizerModel, oi as ExternalTextOptimizerContext, pa as runEval, pi as OptimizationMethodComparison, qi as LlmJudgeOptions, qr as SkillOptTrainerConfig, ri as OptimizerModelBudget, si as ExternalTextOptimizerResult, ta as runImprovementLoop, ti as gepaOptimizationMethod, vi as OptimizationMethodResult, wr as HeldOutGateOptions, xi as OptimizationPackageSource, zi as REFERENCE_EQUIVALENCE_JUDGE_VERSION } from "../index-CojsA_EE.js";
+import { $t as inMemoryCampaignStorage, Qt as fsCampaignStorage, Xt as CampaignStorage } from "../search-kernel-CZ8KJQRw.js";
 import { i as InMemoryOutcomeStore, n as FileSystemOutcomeStore, o as OutcomeStore, r as FileSystemOutcomeStoreOptions, t as DeploymentOutcome } from "../outcome-store-CNt4iZ67.js";
 import { r as HostedTenant } from "../client-BzL9niqi.js";
 import { _ as ReleaseSummary, a as summarizeExecution, c as ExecutionInsight, d as InsightReport, f as InterRaterInsight, g as Recommendation, h as OutcomeCorrelationInsight, i as analyzeRuns, l as FailureClassTally, m as LiftInsight, n as ExecutionReport, o as CostProvenanceSummary, p as JudgeInsight, r as SummarizeExecutionOptions, s as ExecutionErrorOutcomeCell, t as AnalyzeRunsOptions, u as FailureClusterInsight, v as ScalarDistribution, y as TokenUsageInsight } from "../analyze-runs-Q74o4Hfj.js";
-import { _ as SelfImproveMethodResult, a as DefinedAgentEval, c as SelfImproveMethodOptions, d as SelfImproveProposerOptions, f as SelfImproveProposerResult, g as SelfImproveMethodProvenance, h as selfImprove, i as DefineAgentEvalOptions, l as SelfImproveOptions, m as SelfImproveRunError, n as AgentEvalEvaluateOptions, o as defineAgentEval, p as SelfImproveResult, r as AgentEvalImproveOptions, s as SelfImproveBudget, t as AgentEvalAgent, u as SelfImproveProgressEvent } from "../define-agent-eval-CKB6wyML.js";
-import { $ as FromRunRecordDirResult, A as CodeAgentSessionActionStatus, B as AgentTraceFile, C as fromOpenCodeSession, Ct as measuredComparisonFromCandidateExperiment, D as streamCodeAgentJsonlFile, Dt as sealCandidateExperiment, E as parseCodeAgentJsonlFile, Et as sealCandidateBenchmarkTask, F as CodeAgentSessionTerminalStatus, G as PartitionByAuthoringModelResult, H as AgentTraceRange, I as observeCodeAgentSession, J as EvalReportingSuiteInput, K as parseAgentTrace, L as AgentTraceContributor, M as CodeAgentSessionExecutionReceipt, N as CodeAgentSessionObservation, O as CodeAgentSessionAction, Ot as verifyCandidateExperiment, P as CodeAgentSessionSource, Q as FromRunRecordDirOptions, R as AgentTraceContributorType, S as fromKimiCodeSession, St as evaluatePairedMeasurements, T as parseCodeAgentJsonl, Tt as sealCandidateBenchmarkSuite, U as AgentTraceRecord, V as AgentTraceIndex, W as AuthoringProvenance, X as EvalReportingSuiteResult, Y as EvalReportingSuiteOptions, Z as evalReportingSuite, _ as CodeAgentSessionIntakeResult, _t as PairedMeasurement, a as FromFeedbackTableOptions, at as RunAgentProfileImprovementExperimentOptions, b as fromClaudeCodeSession, bt as RunCandidateExperimentOptions, c as CodeAgentStoreObservation, ct as runAgentProfileImprovementExperiment, d as CodeAgentStoreScope, dt as sealAgentProfileImprovementTask, et as RunRecordRejection, f as CodeAgentStoreSession, ft as verifyAgentProfileImprovementExperimentComparison, g as CodeAgentSessionIntakeOptions, gt as EvaluatePairedMeasurementsOptions, h as CodeAgentSessionDiagnostic, ht as CompareCandidateExperimentOptions, i as FeedbackTableRow, it as CompareAgentProfileImprovementExperimentOptions, j as CodeAgentSessionActionSurface, k as CodeAgentSessionActionKind, kt as verifyCandidateExperimentComparison, l as CodeAgentStoreRejection, lt as sealAgentProfileImprovementExperiment, m as CodeAgentJsonlLine, mt as CandidateExperimentRun, n as fromOtelSpans, nt as AgentProfileImprovementExperimentExecutionInput, o as FromFeedbackTableResult, ot as SealAgentProfileImprovementSuiteOptions, p as observeCodeAgentStore, pt as CandidateExperimentExecutionInput, q as partitionRunsByAuthoringModel, r as FeedbackTableMeta, rt as AgentProfileImprovementExperimentRun, s as fromFeedbackTable, st as measuredComparisonFromAgentProfileImprovementExperiment, t as FromOtelSpansOptions, tt as fromRunRecordDir, u as CodeAgentStoreRejections, ut as sealAgentProfileImprovementSuite, v as CodeAgentSessionMetrics, vt as PairedMeasurementAdapter, w as fromPiSession, wt as runCandidateExperiment, x as fromCodexSession, xt as SealCandidateBenchmarkSuiteOptions, y as ParsedCodeAgentJsonl, yt as PairedMeasurementEvaluation, z as AgentTraceConversation } from "../index-R8Xhod8e.js";
+import { _ as SelfImproveMethodResult, a as DefinedAgentEval, c as SelfImproveMethodOptions, d as SelfImproveProposerOptions, f as SelfImproveProposerResult, g as SelfImproveMethodProvenance, h as selfImprove, i as DefineAgentEvalOptions, l as SelfImproveOptions, m as SelfImproveRunError, n as AgentEvalEvaluateOptions, o as defineAgentEval, p as SelfImproveResult, r as AgentEvalImproveOptions, s as SelfImproveBudget, t as AgentEvalAgent, u as SelfImproveProgressEvent } from "../define-agent-eval-B2YV-RDE.js";
+import { $ as FromRunRecordDirResult, A as CodeAgentSessionActionStatus, B as AgentTraceFile, C as fromOpenCodeSession, Ct as measuredComparisonFromCandidateExperiment, D as streamCodeAgentJsonlFile, Dt as sealCandidateExperiment, E as parseCodeAgentJsonlFile, Et as sealCandidateBenchmarkTask, F as CodeAgentSessionTerminalStatus, G as PartitionByAuthoringModelResult, H as AgentTraceRange, I as observeCodeAgentSession, J as EvalReportingSuiteInput, K as parseAgentTrace, L as AgentTraceContributor, M as CodeAgentSessionExecutionReceipt, N as CodeAgentSessionObservation, O as CodeAgentSessionAction, Ot as verifyCandidateExperiment, P as CodeAgentSessionSource, Q as FromRunRecordDirOptions, R as AgentTraceContributorType, S as fromKimiCodeSession, St as evaluatePairedMeasurements, T as parseCodeAgentJsonl, Tt as sealCandidateBenchmarkSuite, U as AgentTraceRecord, V as AgentTraceIndex, W as AuthoringProvenance, X as EvalReportingSuiteResult, Y as EvalReportingSuiteOptions, Z as evalReportingSuite, _ as CodeAgentSessionIntakeResult, _t as PairedMeasurement, a as FromFeedbackTableOptions, at as RunAgentProfileImprovementExperimentOptions, b as fromClaudeCodeSession, bt as RunCandidateExperimentOptions, c as CodeAgentStoreObservation, ct as runAgentProfileImprovementExperiment, d as CodeAgentStoreScope, dt as sealAgentProfileImprovementTask, et as RunRecordRejection, f as CodeAgentStoreSession, ft as verifyAgentProfileImprovementExperimentComparison, g as CodeAgentSessionIntakeOptions, gt as EvaluatePairedMeasurementsOptions, h as CodeAgentSessionDiagnostic, ht as CompareCandidateExperimentOptions, i as FeedbackTableRow, it as CompareAgentProfileImprovementExperimentOptions, j as CodeAgentSessionActionSurface, k as CodeAgentSessionActionKind, kt as verifyCandidateExperimentComparison, l as CodeAgentStoreRejection, lt as sealAgentProfileImprovementExperiment, m as CodeAgentJsonlLine, mt as CandidateExperimentRun, n as fromOtelSpans, nt as AgentProfileImprovementExperimentExecutionInput, o as FromFeedbackTableResult, ot as SealAgentProfileImprovementSuiteOptions, p as observeCodeAgentStore, pt as CandidateExperimentExecutionInput, q as partitionRunsByAuthoringModel, r as FeedbackTableMeta, rt as AgentProfileImprovementExperimentRun, s as fromFeedbackTable, st as measuredComparisonFromAgentProfileImprovementExperiment, t as FromOtelSpansOptions, tt as fromRunRecordDir, u as CodeAgentStoreRejections, ut as sealAgentProfileImprovementSuite, v as CodeAgentSessionMetrics, vt as PairedMeasurementAdapter, w as fromPiSession, wt as runCandidateExperiment, x as fromCodexSession, xt as SealCandidateBenchmarkSuiteOptions, y as ParsedCodeAgentJsonl, yt as PairedMeasurementEvaluation, z as AgentTraceConversation } from "../index-DVDiJIhi.js";
 import { n as buildDefaultAnalystRegistry, t as DefaultAnalystRegistryOptions } from "../default-registry-BHEfZHQd.js";
 export { type AgentEvalAgent, type AgentEvalEvaluateOptions, type AgentEvalImproveOptions, type AgentProfileImprovementExperimentExecutionInput, type AgentProfileImprovementExperimentRun, type AgentTraceContributor, type AgentTraceContributorType, type AgentTraceConversation, type AgentTraceFile, type AgentTraceIndex, type AgentTraceRange, type AgentTraceRecord, type AnalystFinding, type AnalyzeRunsOptions, type AuthoringProvenance, type AxisEvidence, type AxisVerdict, type BuildEvidenceVectorOptions, type CampaignAggregates, type CampaignArtifactWriter, type CampaignCellFailureReceipt, type CampaignCellResult, type CampaignCellRetryPolicy, type CampaignCostMeter, type CampaignResult, type CampaignStorage, type CampaignTraceWriter, type CandidateExperimentExecutionInput, type CandidateExperimentRun, type ChatClient, type CodeAgentJsonlLine, type CodeAgentSessionAction, type CodeAgentSessionActionKind, type CodeAgentSessionActionStatus, type CodeAgentSessionActionSurface, type CodeAgentSessionDiagnostic, type CodeAgentSessionExecutionReceipt, type CodeAgentSessionIntakeOptions, type CodeAgentSessionIntakeResult, type CodeAgentSessionMetrics, type CodeAgentSessionObservation, type CodeAgentSessionSource, type CodeAgentSessionTerminalStatus, type CodeAgentStoreObservation, type CodeAgentStoreRejection, type CodeAgentStoreRejections, type CodeAgentStoreScope, type CodeAgentStoreSession, type CodeSurface, type CompareAgentProfileImprovementExperimentOptions, type CompareCandidateExperimentOptions, type CompareOptimizationMethodsOptions, type ComparisonCost, type CostLedgerHandle, type CostProvenanceSummary, type CreateChatClientOpts, type DefaultAnalystRegistryOptions, type DefaultProductionGateCheck, type DefaultProductionGateOptions, type DefaultProductionRewardHackingOptions, type DefineAgentEvalOptions, type DefinedAgentEval, type DeploymentOutcome, type DispatchFn as Dispatch, type DispatchContext, type EvalReportingSuiteInput, type EvalReportingSuiteOptions, type EvalReportingSuiteResult, type EvaluatePairedMeasurementsOptions, type EvidenceVector, type ExecutionErrorOutcomeCell, type ExecutionInsight, type ExecutionReport, type ExternalOptimizationExample, type ExternalTextEvaluationResponse, type ExternalTextOptimizationMethodConfig, type ExternalTextOptimizerContext, type ExternalTextOptimizerResult, type FailureClassTally, type FailureClusterInsight, type FeedbackTableMeta, type FeedbackTableRow, FileSystemOutcomeStore, type FileSystemOutcomeStoreOptions, type FinalEvidencePolicy, type FinalEvidenceUse, type FromFeedbackTableOptions, type FromFeedbackTableResult, type FromOtelSpansOptions, type FromRunRecordDirOptions, type FromRunRecordDirResult, type Gate, type GateCheckStatus, type GateContext, type GateContribution, type GateDecision, type GateResult, type GenerationCandidate, type GenerationRecord, type GepaAdaptiveEngineRun, type GepaEngineOptions, type GepaEngineRun, type GepaOptimizationMethodConfig, type GepaOptimizationRecipe, type GepaRunnerCommand, type HeldOutGateOptions, type HostedTenant, InMemoryOutcomeStore, type InsightReport, type InterRaterInsight, type JudgeConfig, type JudgeDimension, type JudgeInsight, type JudgeScore, type LiftInsight, type LlmJudgeDimension, type LlmJudgeOptions, type MutableSurface, type ObjectiveSource, type OpenAICompatibleOptimizerModel, type OptimizationMethod, type OptimizationMethodComparison, type OptimizationMethodInput, type OptimizationMethodProvenance, type OptimizationMethodResult, type OptimizationPackageSource, type OptimizationTokenUsage, type OptimizerConfig, type OptimizerModelBudget, type OutcomeCorrelationInsight, type OutcomeStore, type PairedMeasurement, type PairedMeasurementAdapter, type PairedMeasurementEvaluation, type ParetoSignificanceGateOptions, type ParsedCodeAgentJsonl, type PartitionByAuthoringModelResult, type PromotionObjective, type PromotionPolicy, type ProposalFinding, type ProposalFindingOrigin, REFERENCE_EQUIVALENCE_INPUT_LIMITS, REFERENCE_EQUIVALENCE_JUDGE_VERSION, type Recommendation, type ReferenceEquivalenceJudgeInput, type ReferenceEquivalenceJudgeOptions, type ReferenceEquivalenceJudgeResult, type ReferenceEquivalenceScenario, type ReleaseSummary, type RunAgentProfileImprovementExperimentOptions, type RunCampaignOptions, type RunCandidateExperimentOptions, type RunEvalOptions, type RunImprovementLoopOptions, type RunImprovementLoopResult, type RunRecordRejection, type ScalarDistribution, type Scenario, type SealAgentProfileImprovementSuiteOptions, type SealCandidateBenchmarkSuiteOptions, type SelfImproveBudget, type SelfImproveMethodOptions, type SelfImproveMethodProvenance, type SelfImproveMethodResult, type SelfImproveOptions, type SelfImproveProgressEvent, type SelfImproveProposerOptions, type SelfImproveProposerResult, type SelfImproveResult, SelfImproveRunError, type SessionScript, type SkillOptOptimizationMethodConfig, type SkillOptRunnerCommand, type SkillOptTrainerConfig, type SummarizeExecutionOptions, type SurfaceProposer, type TokenUsageInsight, analyzeRuns, buildDefaultAnalystRegistry, buildEvidenceVector, campaignSplitDigest, compareOptimizationMethods, composeGate, createChatClient, createReferenceEquivalenceJudge, defaultProductionGate, defineAgentEval, evalReportingSuite, evaluatePairedMeasurements, externalTextOptimizationMethod, fromClaudeCodeSession, fromCodexSession, fromFeedbackTable, fromKimiCodeSession, fromOpenCodeSession, fromOtelSpans, fromPiSession, fromRunRecordDir, fsCampaignStorage, gepaOptimizationMethod, heldOutGate, inMemoryCampaignStorage, llmJudge, makeProposalFinding, measuredComparisonFromAgentProfileImprovementExperiment, measuredComparisonFromCandidateExperiment, observeCodeAgentSession, observeCodeAgentStore, paretoPolicy, paretoSignificanceGate, parseAgentTrace, parseCodeAgentJsonl, parseCodeAgentJsonlFile, partitionRunsByAuthoringModel, runAgentProfileImprovementExperiment, runCampaign, runCandidateExperiment, runEval, runImprovementLoop, runReferenceEquivalenceJudge, sealAgentProfileImprovementExperiment, sealAgentProfileImprovementSuite, sealAgentProfileImprovementTask, sealCandidateBenchmarkSuite, sealCandidateBenchmarkTask, sealCandidateExperiment, selfImprove, skillOptOptimizationMethod, streamCodeAgentJsonlFile, summarizeExecution, transientDispatchFailure, verifyAgentProfileImprovementExperimentComparison, verifyCandidateExperiment, verifyCandidateExperimentComparison };

package/dist/{define-agent-eval-CKB6wyML.d.ts → define-agent-eval-B2YV-RDE.d.ts}

@@ -3,8 +3,8 @@ import { _ as ProposalFinding } from "./types-6VEehmEg.js";
 import { B as PowerPreflight, ft as EvidenceReceipt, it as CampaignEvidenceContext } from "./statistical-heldout-COfgpWah.js";
 import { nt as EvaluationClaim } from "./search-ledger-types-p2xKfiDJ.js";
 import { H as SurfaceProposer, R as Scenario, S as JudgeConfig, a as CampaignResult, d as DispatchContext, j as MutableSurface, k as LabeledScenarioStore, p as Gate, v as GateResult } from "./types-ZUy5HGY5.js";
-import { Oa as FinalEvidencePolicy, Ti as ComparisonCost, _i as OptimizationMethodProvenance, da as openAutoPr, ea as RunImprovementLoopResult, fa as RunEvalOptions, fi as OptimizationMethod, ga as RunCampaignOptions, ha as CampaignCellRetryPolicy, ka as FinalEvidenceUse, na as PremeasuredOptimizationBaseline, ra as RunOptimizationOptions, ur as LoopProvenanceRecord, vi as OptimizationMethodResult, xr as CampaignComparisonUnits } from "./index-BRdZ0oJo.js";
-import { At as SearchHistoryCoverageRow, Dt as SearchHistoryAdmissionOptions, Mt as SearchHistoryReceipt, Yt as CampaignStorage } from "./search-kernel-DNpsHb8P.js";
+import { Oa as FinalEvidencePolicy, Ti as ComparisonCost, _i as OptimizationMethodProvenance, da as openAutoPr, ea as RunImprovementLoopResult, fa as RunEvalOptions, fi as OptimizationMethod, ga as RunCampaignOptions, ha as CampaignCellRetryPolicy, ka as FinalEvidenceUse, na as PremeasuredOptimizationBaseline, ra as RunOptimizationOptions, ur as LoopProvenanceRecord, vi as OptimizationMethodResult, xr as CampaignComparisonUnits } from "./index-CojsA_EE.js";
+import { Nt as SearchHistoryReceipt, Ot as SearchHistoryAdmissionOptions, Xt as CampaignStorage, jt as SearchHistoryCoverageRow } from "./search-kernel-CZ8KJQRw.js";
 import { r as HostedTenant } from "./client-BzL9niqi.js";
 import { d as InsightReport } from "./analyze-runs-Q74o4Hfj.js";
 //#region src/campaign/presets/run-final-comparison.d.ts
@@ -512,4 +512,4 @@ interface DefinedAgentEval<TScenario extends Scenario, TArtifact> {
 declare function defineAgentEval<TScenario extends Scenario, TArtifact>(defaults: DefineAgentEvalOptions<TScenario, TArtifact>): DefinedAgentEval<TScenario, TArtifact>;
 //#endregion
 export { SelfImproveMethodResult as _, DefinedAgentEval as a, SelfImproveMethodOptions as c, SelfImproveProposerOptions as d, SelfImproveProposerResult as f, SelfImproveMethodProvenance as g, selfImprove as h, DefineAgentEvalOptions as i, SelfImproveOptions as l, SelfImproveRunError as m, AgentEvalEvaluateOptions as n, defineAgentEval as o, SelfImproveResult as p, AgentEvalImproveOptions as r, SelfImproveBudget as s, AgentEvalAgent as t, SelfImproveProgressEvent as u };
-//# sourceMappingURL=define-agent-eval-CKB6wyML.d.ts.map
+//# sourceMappingURL=define-agent-eval-B2YV-RDE.d.ts.map

package/dist/{define-agent-eval-CKB6wyML.d.ts.map → define-agent-eval-B2YV-RDE.d.ts.map}

@@ -1 +1 @@
-{"version":3,"file":"define-agent-eval-CKB6wyML.d.ts","names":["_$__0.FinalEvidenceUse","_$__0.GateResult"],"sources":["../src/campaign/presets/run-final-comparison.ts","../src/contract/self-improve-method.ts","../src/contract/self-improve.ts","../src/contract/define-agent-eval.ts"],"mappings":";;;;;;;;;;UAaiB,uBAAuB,kBAAkB,UAAU,mBAC1D,KAAK,mBAAmB,WAAW;EAC3C,iBAAiB;EACjB,eAAe;EACf,sBACE,SAAS,gBACT,UAAU,WACV,KAAK,WAAW,mBAAmB,WAAW,+BAC3C,QAAQ;EACb,MAAM,KAAK,WAAW;EACtB,QAAQ;EACR,gBAAgB;EAChB;EACA;EACA,cAAc,QAAQ,gBAAgB,UAAU,mBAAmB;;;iBAI/C,mBAAmB,kBAAkB,UAAU,WACnE,MAAM,uBAAuB,WAAW,aAAU;;;;;;;;;;;;kBAA5CA;;;;;;;;;;;UCAS;EACf;EACA;EACA;EACA;EACA;EACA;EACA;EACA;EACA,oBAAoB,YAAY,0BAA0B;EAC1D,QAAQ;EACR;IACE;IACA;IACA;IACA;IACA;IACA;IACA;IACA,sBAAsB;;EAExB,MAAM,QAAQ,kBAAkB;EAChC;EACA;EACA;EACA;EACA,MAAM;EACN;EACA;;UAGe,wBAAwB,kBAAkB,UAAU,mBAC3D,KACN,0BAA0B,WAAW;EAWvC;IAAa,UAAU;IAAiB,QAAQ;;EAChD,uBAAuB;EACvB;;EAEA,UAAU,0BAA0B,WAAW;EAC/C,QAAQ,KAAK,0BAA0B,WAAW;IAChD;;EAEF,YAAY;EACZ,cAAc,YAAY,0BAA0B,WAAW;;EAE/D,MAAM;;EAEN,YAAY;EACZ,UAAU,0BAA0B,WAAW;EAC/C,KAAK,QAAQ,kBAAkB,mBAAmB,WAAW;IAC3D;IACA,iBAAiB,0BAA0B,WAAW;IACtD,eAAe,0BAA0B,WAAW;IACpD;IACA,QAAQ;IACR,MAAM;IACN,WAAW,kBAAkB;;;;;UCpBhB;;;EAGf;;;;EAIA;;EAEA;;EAEA;;;EAGA;;;EAGA;;;;EAIA;;EAEA,mBAAmB;;;;;;;;;EASnB;;EAEA;;;;;EAKA;;KAGU;EACN;EAA0B;;EAC1B;EAA4B;EAAuB;;EACnD;EAA4B;EAAe;;EAC3C;EAA8B;EAAe;EAAuB;;EAGpE;EAAsB;EAAkB;;EACxC;EAAyB;EAAW;EAAY;EAAa;;UAElD,mBAAmB,kBAAkB,UAAU,mBACtD;;;;;;;;;;;;;;EAcR,QAAQ,SAAS,gBAAgB,UAAU,WAAW,KAAK,oBAAoB,QAAQ;;;;;;;EAQvF;;EAEA;;;EAIA,WAAW;;EAGX,OAAO,YAAY,WAAW;;;EAI9B,iBAAiB;;EAGjB,SAAS;;;;;;;;;;;;;EAcT,sBAAsB,gCAAgC,WAAW;;;;;EAMjE,WAAW,gBAAgB;;;;;;EAO3B,SAAS,mBAAmB,WAAW;;;EAIvC,qBAAqB;;;EAIrB,OAAO,KAAK,WAAW;;;;;;;;EASvB,cAAc,eAAe,gBAAgB,iBAAiB,mBAAmB;;;;EAKjF,UAAU;;;;EAKV;;;EAIA,gBAAgB,QAAQ,uBAAuB;;;;;EAM/C,iBAAiB;IACf,UAAU;IACV;IACA;;;;EAKF;;;;;;;EAQA,YAAY;;;EAIZ,cAAc,OAAO;;;;EAKrB;EACA;EACA;;;;;;;;;;;;;EAcA,eAAe;;;;EAKf,eAAe;;EAGf;;;;;;;;EASA;;;;;;;;;EAUA,oBAAoB,uBAAuB,WAAW;;;EAItD,WAAW;;;;;EAMX,SAAS,uBAAuB,WAAW;;;EAI3C,eAAe,uBAAuB,WAAW;;EAEjD,WAAW;;EAEX,QAAQ;;EAER,gBAAgB;;UAGD,0BAA0B,kBAAkB,UAAU;EACrE;EACA,QAAQ;EACR,gBAAgB;;;;EAIhB;IACE;IACA,aAAa;;;;;EAKf;IACE;IACA,aAAa;IACb,SAAS;;;IAGT;;;;IAIA;;;;;;EAMF;;;EAGA;;;;EAIA,YAAY;;EAEZ;;;EAGA;;EAEA;;EAEA;;EAEA,MAAM;;;EAGN,UAAU;;;EAGV,gBAAgB;;EAEhB;IACE;IACA,MAAM;IACN;IACA,aAAa;;;;;;;;EAQf,SAAS;;;;EAIT,QAAQ;;;;;;EAMR,KAAK,yBAAyB,WAAW;;KAG/B,kBAAkB,kBAAkB,UAAU,aACtD,0BAA0B,WAAW,aACrC,wBAAwB,WAAW;KAE3B,yBAAyB,kBAAkB,UAAU,aAAa,KAC5E,mBAAmB,WAAW;EAG9B,QAAQ,mBAAmB,WAAW;EACtC;EACA,gBAAgB,QAAQ;;KAGd,2BAA2B,kBAAkB,UAAU,aAAa,KAC9E,mBAAmB,WAAW;EAG9B;EACA,gBAAgB,QAAQ;;;cAIb,4BAA4B;WAC9B,MAAM;WACN,UAAU;EAEnB,YAAY,gBAAgB,QAAQ;;;;;;;;;;;;;;;;;;;;;;;;;;;;iBA0NtB,YAAY,kBAAkB,UAAU,WACtD,MAAM,yBAAyB,WAAW,aACzC,QAAQ,wBAAwB,WAAW;iBAC9B,YAAY,kBAAkB,UAAU,WACtD,MAAM,2BAA2B,WAAW,aAC3C,QAAQ,0BAA0B,WAAW;iBAChC,YAAY,kBAAkB,UAAU,WACtD,MAAM,mBAAmB,WAAW,aACnC,QAAQ,kBAAkB,WAAW;;;KC9nB5B,eAAe,kBAAkB,UAAU,cACrD,SAAS,gBACT,UAAU,WACV,KAAK,oBACF,QAAQ;KAED,uBAAuB,kBAAkB,UAAU,aAAa,mBAC1E,WACA;UAGe,yBAAyB,kBAAkB,UAAU,mBAC5D,KACN,eAAe,WAAW;;EAI5B,YAAY;;EAEZ,UAAU;;EAEV,QAAQ,eAAe,WAAW;;EAElC,QAAQ,YAAY,WAAW;;EAE/B,SAAS,YAAY,WAAW;;EAEhC;;KAGU,wBAAwB,kBAAkB,UAAU,aAAa,KAC3E,QAAQ,mBAAmB,WAAW;EAGtC,SAAS,QAAQ;EACjB,eAAe,QAAQ;;UAGR,iBAAiB,kBAAkB,UAAU;;WAEnD,oBAAoB;;WAEpB,iBAAiB;;;;;EAK1B,SACE,OAAO,yBAAyB,WAAW,aAC1C,QAAQ,eAAe,WAAW;;;;;;;EAOrC,QACE,OAAO,wBAAwB,WAAW,aACzC,QAAQ,kBAAkB,WAAW;;;;;;;;;iBAU1B,gBAAgB,kBAAkB,UAAU,WAC1D,UAAU,uBAAuB,WAAW,aAC3C,iBAAiB,WAAW"}
+{"version":3,"file":"define-agent-eval-B2YV-RDE.d.ts","names":["_$__0.FinalEvidenceUse","_$__0.GateResult"],"sources":["../src/campaign/presets/run-final-comparison.ts","../src/contract/self-improve-method.ts","../src/contract/self-improve.ts","../src/contract/define-agent-eval.ts"],"mappings":";;;;;;;;;;UAaiB,uBAAuB,kBAAkB,UAAU,mBAC1D,KAAK,mBAAmB,WAAW;EAC3C,iBAAiB;EACjB,eAAe;EACf,sBACE,SAAS,gBACT,UAAU,WACV,KAAK,WAAW,mBAAmB,WAAW,+BAC3C,QAAQ;EACb,MAAM,KAAK,WAAW;EACtB,QAAQ;EACR,gBAAgB;EAChB;EACA;EACA,cAAc,QAAQ,gBAAgB,UAAU,mBAAmB;;;iBAI/C,mBAAmB,kBAAkB,UAAU,WACnE,MAAM,uBAAuB,WAAW,aAAU;;;;;;;;;;;;kBAA5CA;;;;;;;;;;;UCAS;EACf;EACA;EACA;EACA;EACA;EACA;EACA;EACA;EACA,oBAAoB,YAAY,0BAA0B;EAC1D,QAAQ;EACR;IACE;IACA;IACA;IACA;IACA;IACA;IACA;IACA,sBAAsB;;EAExB,MAAM,QAAQ,kBAAkB;EAChC;EACA;EACA;EACA;EACA,MAAM;EACN;EACA;;UAGe,wBAAwB,kBAAkB,UAAU,mBAC3D,KACN,0BAA0B,WAAW;EAWvC;IAAa,UAAU;IAAiB,QAAQ;;EAChD,uBAAuB;EACvB;;EAEA,UAAU,0BAA0B,WAAW;EAC/C,QAAQ,KAAK,0BAA0B,WAAW;IAChD;;EAEF,YAAY;EACZ,cAAc,YAAY,0BAA0B,WAAW;;EAE/D,MAAM;;EAEN,YAAY;EACZ,UAAU,0BAA0B,WAAW;EAC/C,KAAK,QAAQ,kBAAkB,mBAAmB,WAAW;IAC3D;IACA,iBAAiB,0BAA0B,WAAW;IACtD,eAAe,0BAA0B,WAAW;IACpD;IACA,QAAQ;IACR,MAAM;IACN,WAAW,kBAAkB;;;;;UCpBhB;;;EAGf;;;;EAIA;;EAEA;;EAEA;;;EAGA;;;EAGA;;;;EAIA;;EAEA,mBAAmB;;;;;;;;;EASnB;;EAEA;;;;;EAKA;;KAGU;EACN;EAA0B;;EAC1B;EAA4B;EAAuB;;EACnD;EAA4B;EAAe;;EAC3C;EAA8B;EAAe;EAAuB;;EAGpE;EAAsB;EAAkB;;EACxC;EAAyB;EAAW;EAAY;EAAa;;UAElD,mBAAmB,kBAAkB,UAAU,mBACtD;;;;;;;;;;;;;;EAcR,QAAQ,SAAS,gBAAgB,UAAU,WAAW,KAAK,oBAAoB,QAAQ;;;;;;;EAQvF;;EAEA;;;EAIA,WAAW;;EAGX,OAAO,YAAY,WAAW;;;EAI9B,iBAAiB;;EAGjB,SAAS;;;;;;;;;;;;;EAcT,sBAAsB,gCAAgC,WAAW;;;;;EAMjE,WAAW,gBAAgB;;;;;;EAO3B,SAAS,mBAAmB,WAAW;;;EAIvC,qBAAqB;;;EAIrB,OAAO,KAAK,WAAW;;;;;;;;EASvB,cAAc,eAAe,gBAAgB,iBAAiB,mBAAmB;;;;EAKjF,UAAU;;;;EAKV;;;EAIA,gBAAgB,QAAQ,uBAAuB;;;;;EAM/C,iBAAiB;IACf,UAAU;IACV;IACA;;;;EAKF;;;;;;;EAQA,YAAY;;;EAIZ,cAAc,OAAO;;;;EAKrB;EACA;EACA;;;;;;;;;;;;;EAcA,eAAe;;;;EAKf,eAAe;;EAGf;;;;;;;;EASA;;;;;;;;;EAUA,oBAAoB,uBAAuB,WAAW;;;EAItD,WAAW;;;;;EAMX,SAAS,uBAAuB,WAAW;;;EAI3C,eAAe,uBAAuB,WAAW;;EAEjD,WAAW;;EAEX,QAAQ;;EAER,gBAAgB;;UAGD,0BAA0B,kBAAkB,UAAU;EACrE;EACA,QAAQ;EACR,gBAAgB;;;;EAIhB;IACE;IACA,aAAa;;;;;EAKf;IACE;IACA,aAAa;IACb,SAAS;;;IAGT;;;;IAIA;;;;;;EAMF;;;EAGA;;;;EAIA,YAAY;;EAEZ;;;EAGA;;EAEA;;EAEA;;EAEA,MAAM;;;EAGN,UAAU;;;EAGV,gBAAgB;;EAEhB;IACE;IACA,MAAM;IACN;IACA,aAAa;;;;;;;;EAQf,SAAS;;;;EAIT,QAAQ;;;;;;EAMR,KAAK,yBAAyB,WAAW;;KAG/B,kBAAkB,kBAAkB,UAAU,aACtD,0BAA0B,WAAW,aACrC,wBAAwB,WAAW;KAE3B,yBAAyB,kBAAkB,UAAU,aAAa,KAC5E,mBAAmB,WAAW;EAG9B,QAAQ,mBAAmB,WAAW;EACtC;EACA,gBAAgB,QAAQ;;KAGd,2BAA2B,kBAAkB,UAAU,aAAa,KAC9E,mBAAmB,WAAW;EAG9B;EACA,gBAAgB,QAAQ;;;cAIb,4BAA4B;WAC9B,MAAM;WACN,UAAU;EAEnB,YAAY,gBAAgB,QAAQ;;;;;;;;;;;;;;;;;;;;;;;;;;;;iBA0NtB,YAAY,kBAAkB,UAAU,WACtD,MAAM,yBAAyB,WAAW,aACzC,QAAQ,wBAAwB,WAAW;iBAC9B,YAAY,kBAAkB,UAAU,WACtD,MAAM,2BAA2B,WAAW,aAC3C,QAAQ,0BAA0B,WAAW;iBAChC,YAAY,kBAAkB,UAAU,WACtD,MAAM,mBAAmB,WAAW,aACnC,QAAQ,kBAAkB,WAAW;;;KC9nB5B,eAAe,kBAAkB,UAAU,cACrD,SAAS,gBACT,UAAU,WACV,KAAK,oBACF,QAAQ;KAED,uBAAuB,kBAAkB,UAAU,aAAa,mBAC1E,WACA;UAGe,yBAAyB,kBAAkB,UAAU,mBAC5D,KACN,eAAe,WAAW;;EAI5B,YAAY;;EAEZ,UAAU;;EAEV,QAAQ,eAAe,WAAW;;EAElC,QAAQ,YAAY,WAAW;;EAE/B,SAAS,YAAY,WAAW;;EAEhC;;KAGU,wBAAwB,kBAAkB,UAAU,aAAa,KAC3E,QAAQ,mBAAmB,WAAW;EAGtC,SAAS,QAAQ;EACjB,eAAe,QAAQ;;UAGR,iBAAiB,kBAAkB,UAAU;;WAEnD,oBAAoB;;WAEpB,iBAAiB;;;;;EAK1B,SACE,OAAO,yBAAyB,WAAW,aAC1C,QAAQ,eAAe,WAAW;;;;;;;EAOrC,QACE,OAAO,wBAAwB,WAAW,aACzC,QAAQ,kBAAkB,WAAW;;;;;;;;;iBAU1B,gBAAgB,kBAAkB,UAAU,WAC1D,UAAU,uBAAuB,WAAW,aAC3C,iBAAiB,WAAW"}

package/dist/{index-BRdZ0oJo.d.ts → index-CojsA_EE.d.ts}

@@ -8,7 +8,7 @@ import { a as PairedPromotionDecision, d as PairedBootstrapResult } from "./pair
 import { E as SearchExecutionIdentity, G as SearchSplit, V as SearchProposer, a as SearchAttemptAccounting, i as SearchArtifactRef, j as SearchModelIdentity, nt as EvaluationClaim, rt as EvaluationUnitSummary, tt as SearchUnknown, z as SearchOperationRecordedEvent } from "./search-ledger-types-p2xKfiDJ.js";
 import { r as LedgerHash } from "./canonical-CFpojCN5.js";
 import { A as LabeledScenarioWrite, C as JudgeDimension, D as LabeledScenarioSampleArgs, E as LabeledScenarioRecord, H as SurfaceProposer, N as ParetoParent, O as LabeledScenarioSource, R as Scenario, S as JudgeConfig, T as LabelTrust, _ as GateDecision, a as CampaignResult, b as GenerationRecord, c as CampaignTraceWriter, d as DispatchContext, f as DispatchFn, g as GateContribution, i as CampaignCostMeter, j as MutableSurface, k as LabeledScenarioStore, l as CodeSurface, o as CampaignScenarioIdentity, p as Gate, r as CampaignCellResult, u as ComponentSurface, v as GateResult, y as GenerationCandidate } from "./types-ZUy5HGY5.js";
-import { Dt as SearchHistoryAdmissionOptions, L as SearchPolicy, Mt as SearchHistoryReceipt, Yt as CampaignStorage, bt as SearchRunIdentity, gt as SearchLedgerBinding, ht as RegisterSearchNodeInput, i as SearchCellResult, kt as SearchHistoryCoverage, vt as SearchRecorder } from "./search-kernel-DNpsHb8P.js";
+import { At as SearchHistoryCoverage, L as SearchPolicy, Nt as SearchHistoryReceipt, Ot as SearchHistoryAdmissionOptions, Xt as CampaignStorage, _t as SearchLedgerBinding, gt as RegisterSearchNodeInput, i as SearchCellResult, xt as SearchRunIdentity, yt as SearchRecorder } from "./search-kernel-CZ8KJQRw.js";
 import { B as ExternalOptimizerModelBudget, E as AnalystIssueExpectation, F as ExternalOptimizerCallbackLimits, J as ExternalOptimizerWireCounts, K as ExternalOptimizerResumeMode, R as ExternalOptimizerEvaluationObservation, V as ExternalOptimizerModelCall, X as ExternalTextEvaluationRequest, Y as ExternalTextCandidate, m as AnalystBenchmarkLabelState, q as ExternalOptimizerRunnerCommand, t as AgentProfile$1, u as AnalystBenchmarkCase } from "./agent-profile-EpO0HURy.js";
 import { r as DatasetScenario, t as Dataset } from "./dataset-DQqhOCPt.js";
 import { a as CheckerIdentity, n as VerdictCertification, t as DefaultVerdict, x as VerificationStrategySource } from "./verdict-E4eRNf7-.js";
@@ -3390,4 +3390,4 @@ declare function verifyCodeSurface(surface: CodeSurface, worktreeDir?: string):
 declare function resolveWorktreePath(surface: CodeSurface, worktreeDir?: string): string;
 //#endregion
 export { ScoreboardRow as $, RunImprovementLoopOptions as $i, TraceAnalystScenario as $n, GepaOptimizationRecipe as $r, importExternalEvaluations as $t, ScenarioSignal as A, CampaignCoverage as Aa, GepaCandidatePopulationSummary as Ai, CrossSurfaceComponentEvidence as An, DEFAULT_RED_TEAM_CORPUS as Ar, RequirementCheck as At, FinalizedProfileMatrixResult as B, ReferenceEquivalenceJudgeInput as Bi, CrossSurfaceInteractionTask as Bn, CanaryKind as Br, FsLabeledScenarioStoreOptions as Bt, SingleRunLockOptions as C, CacheRead as Ca, compareOptimizationMethods as Ci, CrossSurfaceBootstrapPolicy as Cn, campaignMeanComposite as Cr, extractProducedState as Ct, searchProposalExecution as D, cellCachePath as Da, costFromLedgerSummary as Di, CrossSurfaceCandidateOutcome as Dn, DefaultProductionGateOptions as Dr, LlmCorrectnessCheckerOpts as Dt, searchModelIdentity as E, buildCellSchedule as Ea, combineComparisonCosts as Ei, CrossSurfaceCandidateEvidence as En, DefaultProductionGateCheck as Er, CorrectnessChecker as Et, RetainedDispatchOptions as F, campaignSplitDigest as Fa, ExternalOptimizerObservationSummary as Fi, CrossSurfaceIneligibilityReason as Fn, redTeamDataset as Fr, verifyCompletion as Ft, RunProfileMatrixSegmentOptions as G, runReferenceEquivalenceJudge as Gi, CrossSurfacePairwiseEntry as Gn, SkillOptOptimizationMethodConfig as Gr, ScoredRollout as Gt, ProfileMatrixPlan as H, ReferenceEquivalenceJudgeResult as Hi, CrossSurfacePairCompatibility as Hn, CanaryReport as Hr, RolloutArgumentDiff as Ht, RetainedDispatchOutcome as I, campaignSplitDigestFromIdentities as Ia, ExternalOptimizerSubmittedCandidate as Ii, CrossSurfaceInteractionAwareSelection as In, redTeamReport as Ir, scopedOptimizationMethod as It, runProfileMatrixSegment as J, llmJudge as Ji, CrossSurfaceSelectionPolicy as Jn, skillOptOptimizationMethod as Jr, rolloutArgumentDiff as Jt, createProfileMatrixPlan as K, LlmJudgeDimension as Ki, CrossSurfaceRankedSingle as Kn, SkillOptRunnerCommand as Kr, UngroundedLiteralReport as Kt, createRetainedDispatch as L, readExternalOptimizerObservationArtifact as Li, CrossSurfaceInteractionEffect as Ln, scoreRedTeamOutput as Lr, sequentialOptimizationMethod as Lt, selectDiscriminative as M, assertCampaignSplitIdentity as Ma, readGepaCandidatePopulationArtifact as Mi, CrossSurfaceDistribution as Mn, RedTeamCategory as Mr, TaskGold as Mt, resolveRunDir as N, campaignCoverage as Na, ExternalOptimizerExecutionSummary as Ni, CrossSurfaceEligibility as Nn, RedTeamFinding as Nr, completionVerdict as Nt, searchReceiptAccounting as O, FinalEvidencePolicy as Oa, GepaCandidatePopulationArtifact as Oi, CrossSurfaceCandidateSummary as On, DefaultProductionRewardHackingOptions as Or, ProducedProposal as Ot, tangleTracesRoot as P, campaignScenarioIdentity as Pa, ExternalOptimizerObservationArtifact as Pi, CrossSurfaceEvidenceBreakdown as Pn, RedTeamReport as Pr, createLlmCorrectnessChecker as Pt, ScoreboardRenderOptions as Q, transientDispatchFailure as Qi, TraceAnalystArtifact as Qn, GepaOptimizationMethodConfig as Qr, externalSurface as Qt, CreateProfileMatrixPlanOptions as R, REFERENCE_EQUIVALENCE_INPUT_LIMITS as Ri, CrossSurfaceInteractionPath as Rn, CanaryAlert as Rr, neutralizeText as Rt, SingleRunLock as S, CacheIssueReason as Sa, OptimizationTokenUsage as Si, CrossSurfaceBestSingleSelection as Sn, campaignBreakdown as Sr, ToolCallEventLike as St, campaignCellSearchResult as T, CellScheduleSlot as Ta, ComparisonCost as Ti, CrossSurfaceCandidateComparison as Tn, heldOutGate as Tr, CompletionVerdict as Tt, ProfileMatrixRow as U, ReferenceEquivalenceScenario as Ui, CrossSurfacePairEvidence as Un, runCanaries as Ur, RolloutArgumentDiffOptions as Ut, ProfileMatrixCoverage as V, ReferenceEquivalenceJudgeOptions as Vi, CrossSurfaceNaiveStackSelection as Vn, CanaryOptions as Vr, LabeledScenarioStoreError as Vt, ProfileMatrixSegmentResult as W, createReferenceEquivalenceJudge as Wi, CrossSurfacePairIncompatibilityReason as Wn, composeGate as Wr, RolloutCall as Wt, PlaybackDriver as X, isTransientTransportFailure as Xi, CrossSurfaceTaskRow as Xn, GepaEngineOptions as Xr, GepaEvaluationImport as Xt, PlaybackContext as Y, TransientFailureOptions as Yi, CrossSurfaceSelections as Yn, GepaAdaptiveEngineRun as Yr, ExternalSearchOptimizer as Yt, PlaybackStep as Z, quotaExhaustedUntil as Zi, BuildTraceAnalystSurfaceDispatchOptions as Zn, GepaEngineRun as Zr, GepaPopulationImport as Zt, componentSurfaceIdentityMaterial as _, runCampaign as _a, OptimizationMethodProvenance as _i, analyzeCrossSurfaceInteractions as _n, provenanceRecordPath as _r, assertRealBackend as _t, WorktreeAdapterError as a, runOptimization as aa, ExternalTextOptimizationMethodConfig as ai, EvalFixture as an, LoopProvenanceArgsFromResult as ar, scoreUserStory as at, surfaceDispatchRef as b, PlanCampaignRunOptions as ba, OptimizationMethodScore as bi, CrossSurfaceAdditionRejectionReason as bn, CampaignBreakdown as br, ProposalEventLike as bt, verifyCodeSurface as c, SearchLedgerIntegrityError as ca, ExternalOptimizationExample as ci, EvalFixtureRunPlan as cn, LoopProvenanceEvidence as cr, ProfileDispatchFn as ct, PhoenixEvaluationResultLike as d, openAutoPr as da, CompareOptimizationMethodsOptions as di, LoadEvalFixtureScenariosOptions as dn, buildLoopProvenanceRecord as dr, RunProfileMatrixOptions as dt, RunImprovementLoopResult as ea, GepaRunnerCommand as ei, importGepaPopulation as en, buildTraceAnalystSurfaceDispatch as er, ScoreboardSummary as et, PhoenixEvaluatorLike as f, RunEvalOptions as fa, OptimizationMethod as fi, PlanEvalFixtureRunOptions as fn, campaignMeasurementDigest as fr, RunProfileMatrixResult as ft, codeSurfaceIdentityMaterial as g, RunCampaignOptions as ga, OptimizationMethodPairwise as gi, planEvalFixtureRun as gn, loopProvenanceSpans as gr, BackendIntegrityReport as gt, assertCodeSurfaceIdentity as h, CampaignCellRetryPolicy as ha, OptimizationMethodInput as hi, loadEvalFixtureScenarios as hn, loopProvenanceArgsFromResult as hr, BackendIntegrityError as ht, WorktreeAdapter as i, RunOptimizationResult as ia, externalTextOptimizationMethod as ii, neutralizationGate as in, EmitLoopProvenanceResult as ir, renderScoreboardMarkdown as it, scoreDiscrimination as j, assertCampaignDesign as ja, GepaCandidateSelectionScore as ji, CrossSurfaceCompositionStep as jn, RedTeamCase as jr, SatisfiedBy as jt, DiscriminationScore as k, FinalEvidenceUse as ka, GepaCandidatePopulationCandidate as ki, CrossSurfaceComponent as kn, defaultProductionGate as kr, ProducedState as kt, AutoevalsScoreLike as l, OpenAutoPrOptions as la, ExternalTextEvaluationResponse as li, EvalFixtureScenario as ln, LoopProvenanceOptimizationMethod as lr, ProfileMatrixError as lt, phoenixEvaluatorJudge as m, CampaignCellFailureReceipt as ma, OptimizationMethodComposition as mi, loadEvalFixture as mn, emitLoopProvenance as mr, runProfileMatrix as mt, GitWorktreeAdapterOptions as n, PremeasuredOptimizationBaseline as na, OpenAICompatibleOptimizerModel as ni, recordSkillOptSearch as nn, BuildLoopProvenanceArgs as nr, UserStoryVerdict as nt, gitWorktreeAdapter as o, SearchLedgerConflictError as oa, ExternalTextOptimizerContext as oi, EvalFixtureFile as on, LoopProvenanceBackend as or, scoreboardSummary as ot, autoevalsScorerJudge as p, runEval as pa, OptimizationMethodComparison as pi, discoverEvalFixtures as pn, canonicalDigest as pr, ScenarioRollup as pt, finalizeProfileMatrix as q, LlmJudgeOptions as qi, CrossSurfaceRelativeCost as qn, SkillOptTrainerConfig as qr, classifyUngroundedLiterals as qt, Worktree as r, RunOptimizationOptions as ra, OptimizerModelBudget as ri, NeutralizationGateOptions as rn, EmitLoopProvenanceArgs as rr, makePlaybackDispatch as rt, resolveWorktreePath as s, SearchLedgerError as sa, ExternalTextOptimizerResult as si, EvalFixtureLoadOptions as sn, LoopProvenanceCandidate as sr, userStoryScoreboard as st, CodeSurfaceVerification as t, runImprovementLoop as ta, gepaOptimizationMethod as ti, recordGepaSearch as tn, traceAnalystQualityJudge as tr, UserStory as tt, AutoevalsScorerLike as u, OpenAutoPrResult as ua, decodeExternalTextCandidate as ui, EvalFixtureValidationMode as un, LoopProvenanceRecord as ur, ProfileSummary as ut, renderSurfaceDiff as v, CampaignRunPlan as va, OptimizationMethodResult as vi, AnalyzeCrossSurfaceInteractionsInput as vn, provenanceSpansPath as vr, summarizeBackendIntegrity as vt, acquireSingleRunLock as w, readCachedCell as wa, optimizationTokenUsageFromSummary as wi, CrossSurfaceCandidate as wn, HeldOutGateOptions as wr, CompletionRequirement as wt, surfaceHash as x, planCampaignRun as xa, OptimizationPackageSource as xi, CrossSurfaceAttemptCompleteness as xn, CampaignComparisonUnits as xr, RuntimeEventLike as xt, surfaceContentHash as y, CampaignRunPlanCell as ya, OptimizationMethodRunOptions as yi, CrossSurfaceAdditionDecision as yn, verifyLoopProvenanceRecord as yr, ArtifactEventLike as yt, FinalizeProfileMatrixOptions as z, REFERENCE_EQUIVALENCE_JUDGE_VERSION as zi, CrossSurfaceInteractionReport as zn, CanaryEvaluation as zr, FsLabeledScenarioStore as zt };
-//# sourceMappingURL=index-BRdZ0oJo.d.ts.map
+//# sourceMappingURL=index-CojsA_EE.d.ts.map