@tangle-network/agent-eval 0.92.0 → 0.94.0

package/dist/contract/index.d.ts

@@ -1,34 +1,34 @@
 import { S as Scenario, M as MutableSurface, b as DispatchContext, a as JudgeConfig, I as ImprovementDriver, G as Gate, L as LabeledScenarioStore, c as GateDecision } from '../types-BU-7W85F.js';
 export { C as CampaignAggregates, d as CampaignArtifactWriter, e as CampaignCellResult, f as CampaignCostMeter, g as CampaignResult, h as CampaignTraceWriter, i as CodeSurface, D as Dispatch, j as GateContext, k as GateResult, l as GenerationCandidate, m as GenerationRecord, n as JudgeDimension, J as JudgeScore, o as Mutator, O as OptimizerConfig, p as SessionScript } from '../types-BU-7W85F.js';
-import { L as LoopProvenanceRecord } from '../provenance-LnqRT0sS.js';
-export { A as AxisEvidence, a as AxisVerdict, B as BuildEvidenceVectorOptions, D as DefaultProductionGateOptions, E as EvidenceVector, b as EvolutionaryDriverOptions, H as HeldOutGateOptions, O as ObjectiveSource, P as ParetoSignificanceGateOptions, c as PromotionObjective, d as PromotionPolicy, R as RunEvalOptions, e as buildEvidenceVector, f as composeGate, g as defaultProductionGate, h as evolutionaryDriver, i as heldOutGate, p as paretoPolicy, j as paretoSignificanceGate, r as runEval } from '../provenance-LnqRT0sS.js';
-import { C as CampaignStorage, R as RunOptimizationOptions, a as RunImprovementLoopResult } from '../run-improvement-loop-5z_l5zDz.js';
-export { G as GepaDriverOptions, b as RunCampaignOptions, c as RunImprovementLoopOptions, f as fsCampaignStorage, g as gepaDriver, i as inMemoryCampaignStorage, r as runCampaign, d as runImprovementLoop } from '../run-improvement-loop-5z_l5zDz.js';
+import { L as LoopProvenanceRecord } from '../provenance-P-bCL2Fo.js';
+export { A as AxisEvidence, a as AxisVerdict, B as BuildEvidenceVectorOptions, D as DefaultProductionGateOptions, E as EvidenceVector, b as EvolutionaryDriverOptions, H as HeldOutGateOptions, O as ObjectiveSource, P as ParetoSignificanceGateOptions, c as PromotionObjective, d as PromotionPolicy, R as RunEvalOptions, e as buildEvidenceVector, f as composeGate, g as defaultProductionGate, h as evolutionaryDriver, i as heldOutGate, p as paretoPolicy, j as paretoSignificanceGate, r as runEval } from '../provenance-P-bCL2Fo.js';
+import { C as CampaignStorage, R as RunOptimizationOptions, a as RunImprovementLoopResult } from '../run-improvement-loop-DBahB8Ax.js';
+export { G as GepaDriverOptions, b as RunCampaignOptions, c as RunImprovementLoopOptions, f as fsCampaignStorage, g as gepaDriver, i as inMemoryCampaignStorage, r as runCampaign, d as runImprovementLoop } from '../run-improvement-loop-DBahB8Ax.js';
 export { D as DeploymentOutcome, F as FileSystemOutcomeStore, a as FileSystemOutcomeStoreOptions, I as InMemoryOutcomeStore, b as OutcomeStore } from '../outcome-store-rnXLEqSn.js';
 import { HostedTenant, EvalRunCellScore, EvalRunGenerationSnapshot, EvalRunEvent, TraceSpanEvent } from '../hosted/index.js';
 import { R as RunRecord, a as RunSplitTag } from '../run-record-e7vj1uZQ.js';
-import { I as InsightReport } from '../insight-report-BBwvOh6x.js';
-export { F as FailureClusterInsight, a as InterRaterInsight, J as JudgeInsight, L as LiftInsight, O as OutcomeCorrelationInsight, R as Recommendation, b as ReleaseSummary, S as ScalarDistribution } from '../insight-report-BBwvOh6x.js';
-export { D as DefaultAnalystRegistryOptions, b as buildDefaultAnalystRegistry } from '../default-registry-zoGHUQEH.js';
-export { c as AnalystFinding } from '../types-2VVIL04s.js';
-export { A as AnalyzeRunsOptions, a as analyzeRuns } from '../analyze-runs-DwCEkpO_.js';
+import { I as InsightReport } from '../insight-report-DWl3z9tl.js';
+export { F as FailureClusterInsight, a as InterRaterInsight, J as JudgeInsight, L as LiftInsight, O as OutcomeCorrelationInsight, R as Recommendation, b as ReleaseSummary, S as ScalarDistribution } from '../insight-report-DWl3z9tl.js';
+export { D as DefaultAnalystRegistryOptions, b as buildDefaultAnalystRegistry } from '../default-registry-6dhErQbs.js';
+export { c as AnalystFinding } from '../types-Ce17tDlG.js';
+export { A as AnalyzeRunsOptions, a as analyzeRuns } from '../analyze-runs-B6Ljo_dI.js';
 export { C as CodeAgentSessionDiagnostic, a as CodeAgentSessionIntakeOptions, b as CodeAgentSessionIntakeResult, c as CodeAgentSessionMetrics, d as CodeAgentSessionSource, P as ParsedCodeAgentJsonl, f as fromClaudeCodeSession, e as fromCodexSession, g as fromKimiCodeSession, h as fromOpenCodeSession, i as fromPiSession, j as fromPigraphSession, p as parseCodeAgentJsonl } from '../code-agent-session-BO8nCnv3.js';
-import '../red-team-BXHil6c8.js';
+import '../red-team-BWdoyleI.js';
 import '../dataset-BbGkaN2I.js';
 import '../errors-CzMUYo7b.js';
-import '../store-CKUAgsJz.js';
+import '../store-BcFXE6LG.js';
 import '../schema-m0gsnbt3.js';
 import '../pareto-E-pembql.js';
-import '../statistics-C7PozGrZ.js';
+import '../statistics-CCJpTGOS.js';
 import '../judge-calibration-0p2QcWNE.js';
-import '../types-Croy5h7V.js';
+import '../types-C7DGg5ex.js';
 import '@tangle-network/tcloud';
-import '../llm-client-BeEcAokY.js';
+import '../llm-client-Bj7g0rqu.js';
 import '../raw-provider-sink-C46HDghv.js';
-import '../summary-report-DGmUucwQ.js';
-import '../failure-cluster-CL7IVgkJ.js';
+import '../summary-report-BDOFevaT.js';
+import '../failure-cluster-DH9Flgcf.js';
 import '@ax-llm/ax';
-import '../kind-factory-5b7xXXOr.js';
+import '../kind-factory-0BhLSI27.js';
 import '../store-C1YxJDEK.js';
 import 'zod';
 

package/dist/contract/index.js

@@ -18,40 +18,40 @@ import {
   paretoPolicy,
   paretoSignificanceGate,
   runEval
-} from "../chunk-6SOJM3VR.js";
+} from "../chunk-SD2YFWQQ.js";
 import {
   analyzeRuns
-} from "../chunk-Y47J2LJ3.js";
+} from "../chunk-2KNZHH3P.js";
 import {
   emitLoopProvenance,
   gepaDriver,
   heldOutGate,
   runImprovementLoop,
   surfaceContentHash
-} from "../chunk-ZFIBGEOL.js";
+} from "../chunk-EGPMSBEZ.js";
 import {
   fsCampaignStorage,
   inMemoryCampaignStorage,
   runCampaign
-} from "../chunk-TWS7AZEY.js";
+} from "../chunk-2K6UUZ7P.js";
 import "../chunk-VI2UW6B6.js";
 import {
   buildDefaultAnalystRegistry
-} from "../chunk-UMMZHCPB.js";
-import "../chunk-D3V5B42D.js";
+} from "../chunk-MIFZUPEK.js";
+import "../chunk-Q5LIB7BC.js";
 import "../chunk-5HRORJQY.js";
 import {
   FileSystemOutcomeStore,
   InMemoryOutcomeStore
 } from "../chunk-3RF76KTD.js";
-import "../chunk-CVVHBFGN.js";
-import "../chunk-CY6U5S3X.js";
+import "../chunk-CWNP4DV4.js";
+import "../chunk-URWVKRCS.js";
 import "../chunk-45EEMHTC.js";
-import "../chunk-LMZQ2Z4U.js";
+import "../chunk-56GC6VYO.js";
 import "../chunk-GGE4NNQT.js";
-import "../chunk-47X6LRCE.js";
+import "../chunk-UFSG7ACU.js";
 import "../chunk-5BKGXME7.js";
-import "../chunk-QAY5UIJO.js";
+import "../chunk-QMUEXQJS.js";
 import "../chunk-PC4UYEBM.js";
 import "../chunk-VSMTAMNK.js";
 import "../chunk-3BFEG2F6.js";

package/dist/{control-_Qb7skHX.d.ts → control-D6qwHXIR.d.ts}

@@ -1,8 +1,8 @@
-import { b as FeedbackLabel, p as ProposedSideEffect } from './feedback-trajectory-D9OVLrg9.js';
-import { C as ControlEvalResult, a as ControlRunResult, h as ControlRuntimeConfig } from './control-runtime-DuFBYg7A.js';
-import { T as TraceEmitter } from './emitter-DEZwY14K.js';
+import { b as FeedbackLabel, p as ProposedSideEffect } from './feedback-trajectory-BxY0cKfs.js';
+import { C as ControlEvalResult, a as ControlRunResult, h as ControlRuntimeConfig } from './control-runtime-Acf9CGhw.js';
+import { T as TraceEmitter } from './emitter-C2rqGH_l.js';
 import { F as FailureClass } from './schema-m0gsnbt3.js';
-import { T as TraceStore } from './store-CKUAgsJz.js';
+import { T as TraceStore } from './store-BcFXE6LG.js';
 import { a as RunSplitTag, b as RunTokenUsage, R as RunRecord } from './run-record-e7vj1uZQ.js';
 
 interface ActionExecutionPolicy {

package/dist/{control-runtime-DuFBYg7A.d.ts → control-runtime-Acf9CGhw.d.ts}

@@ -1,6 +1,6 @@
-import { T as TraceEmitter } from './emitter-DEZwY14K.js';
+import { T as TraceEmitter } from './emitter-C2rqGH_l.js';
 import { F as FailureClass } from './schema-m0gsnbt3.js';
-import { T as TraceStore } from './store-CKUAgsJz.js';
+import { T as TraceStore } from './store-BcFXE6LG.js';
 
 /**
  * Policy-based agent control runtime.

package/dist/control.d.ts

@@ -1,9 +1,9 @@
-export { A as ActionExecutionPolicy, a as ActionPolicyDecision, C as ControlRunToRunRecordOptions, e as ProposeReviewConfig, f as ProposeReviewControlAction, g as ProposeReviewControlConfig, h as ProposeReviewControlResult, i as ProposeReviewControlState, j as ProposeReviewReport, p as RunEvidenceMetadata, s as controlRunToRunRecord, u as evaluateActionPolicy, x as runProposeReview, y as runProposeReviewAsControlLoop, z as scoreFromEvals } from './control-_Qb7skHX.js';
-export { c as ControlActionFailureMode, d as ControlActionOutcome, e as ControlBudget, f as ControlContext, g as ControlDecision, C as ControlEvalResult, a as ControlRunResult, h as ControlRuntimeConfig, i as ControlRuntimeError, j as ControlSeverity, b as ControlStep, k as ControlStopPolicies, S as StopDecision, l as allCriticalPassed, o as objectiveEval, r as runAgentControlLoop, s as stopOnNoProgress, m as stopOnRepeatedAction, n as subjectiveEval } from './control-runtime-DuFBYg7A.js';
-import './feedback-trajectory-D9OVLrg9.js';
+export { A as ActionExecutionPolicy, a as ActionPolicyDecision, C as ControlRunToRunRecordOptions, e as ProposeReviewConfig, f as ProposeReviewControlAction, g as ProposeReviewControlConfig, h as ProposeReviewControlResult, i as ProposeReviewControlState, j as ProposeReviewReport, p as RunEvidenceMetadata, s as controlRunToRunRecord, u as evaluateActionPolicy, x as runProposeReview, y as runProposeReviewAsControlLoop, z as scoreFromEvals } from './control-D6qwHXIR.js';
+export { c as ControlActionFailureMode, d as ControlActionOutcome, e as ControlBudget, f as ControlContext, g as ControlDecision, C as ControlEvalResult, a as ControlRunResult, h as ControlRuntimeConfig, i as ControlRuntimeError, j as ControlSeverity, b as ControlStep, k as ControlStopPolicies, S as StopDecision, l as allCriticalPassed, o as objectiveEval, r as runAgentControlLoop, s as stopOnNoProgress, m as stopOnRepeatedAction, n as subjectiveEval } from './control-runtime-Acf9CGhw.js';
+import './feedback-trajectory-BxY0cKfs.js';
 import './dataset-BbGkaN2I.js';
 import './errors-CzMUYo7b.js';
-import './emitter-DEZwY14K.js';
+import './emitter-C2rqGH_l.js';
 import './schema-m0gsnbt3.js';
-import './store-CKUAgsJz.js';
+import './store-BcFXE6LG.js';
 import './run-record-e7vj1uZQ.js';

package/dist/control.js

@@ -4,7 +4,7 @@ import {
   runProposeReview,
   runProposeReviewAsControlLoop,
   scoreFromEvals
-} from "./chunk-RTWFUK6A.js";
+} from "./chunk-E4GH6USR.js";
 import {
   allCriticalPassed,
   objectiveEval,
@@ -12,7 +12,7 @@ import {
   stopOnNoProgress,
   stopOnRepeatedAction,
   subjectiveEval
-} from "./chunk-NCRFYPS3.js";
+} from "./chunk-YEHAEDUD.js";
 import "./chunk-KWRRMR3J.js";
 import "./chunk-TVVP3ZZQ.js";
 import "./chunk-VSMTAMNK.js";

package/dist/{corpus-BoR-041R.d.ts → corpus-B8A4BDR3.d.ts}

@@ -1,6 +1,6 @@
 import { R as RunRecord, a as RunSplitTag } from './run-record-e7vj1uZQ.js';
 import { S as Span } from './schema-m0gsnbt3.js';
-import { T as TraceStore } from './store-CKUAgsJz.js';
+import { T as TraceStore } from './store-BcFXE6LG.js';
 
 /**
  * Preference dataset extraction — bridge from `RunRecord[]` to RL training.

package/dist/{counterfactual-Dwibr5IW.d.ts → counterfactual-DlOz8PBx.d.ts}

@@ -1,6 +1,6 @@
-import { T as TraceEmitter } from './emitter-DEZwY14K.js';
-import { T as TraceStore } from './store-CKUAgsJz.js';
-import { T as Trajectory, a as TrajectoryStep } from './trajectory-GEdXJCL5.js';
+import { T as TraceEmitter } from './emitter-C2rqGH_l.js';
+import { T as TraceStore } from './store-BcFXE6LG.js';
+import { T as Trajectory, a as TrajectoryStep } from './trajectory-2TkpSEVh.js';
 
 /**
  * Counterfactual replay — "what would have happened if we'd changed

package/dist/{default-registry-zoGHUQEH.d.ts → default-registry-6dhErQbs.d.ts}

@@ -1,6 +1,6 @@
 import { AxAIService } from '@ax-llm/ax';
-import { T as TraceAnalystKindSpec } from './kind-factory-5b7xXXOr.js';
-import { A as Analyst, a as AnalystContext, b as AnalystRunSummary, c as AnalystFinding, d as AnalystRunResult, C as ChatClient, e as AnalystRunInputs, f as AnalystRunEvent } from './types-2VVIL04s.js';
+import { T as TraceAnalystKindSpec } from './kind-factory-0BhLSI27.js';
+import { A as Analyst, a as AnalystContext, b as AnalystRunSummary, c as AnalystFinding, d as AnalystRunResult, C as ChatClient, e as AnalystRunInputs, f as AnalystRunEvent } from './types-Ce17tDlG.js';
 
 /**
  * AnalystRegistry — orchestrate N analysts against one run.

package/dist/diagnose.d.ts

@@ -1,16 +1,16 @@
-import { C as CounterfactualMutation, a as attributeCounterfactuals, b as CounterfactualRunner } from './counterfactual-Dwibr5IW.js';
-export { c as CounterfactualContext, d as CounterfactualResult } from './counterfactual-Dwibr5IW.js';
+import { C as CounterfactualMutation, a as attributeCounterfactuals, b as CounterfactualRunner } from './counterfactual-DlOz8PBx.js';
+export { c as CounterfactualContext, d as CounterfactualResult } from './counterfactual-DlOz8PBx.js';
 import { S as Span } from './schema-m0gsnbt3.js';
-import { T as TraceStore } from './store-CKUAgsJz.js';
-import { a as TrajectoryStep, T as Trajectory } from './trajectory-GEdXJCL5.js';
-import { h as AnalystSeverity, c as AnalystFinding } from './types-2VVIL04s.js';
-import { C as CorpusRecord } from './corpus-BoR-041R.js';
+import { T as TraceStore } from './store-BcFXE6LG.js';
+import { a as TrajectoryStep, T as Trajectory } from './trajectory-2TkpSEVh.js';
+import { h as AnalystSeverity, c as AnalystFinding } from './types-Ce17tDlG.js';
+import { C as CorpusRecord } from './corpus-B8A4BDR3.js';
 import { R as RunRecord } from './run-record-e7vj1uZQ.js';
-import './emitter-DEZwY14K.js';
+import './emitter-C2rqGH_l.js';
 import './store-C1YxJDEK.js';
-import './types-Croy5h7V.js';
+import './types-C7DGg5ex.js';
 import '@tangle-network/tcloud';
-import './llm-client-BeEcAokY.js';
+import './llm-client-Bj7g0rqu.js';
 import './errors-CzMUYo7b.js';
 import './raw-provider-sink-C46HDghv.js';
 

package/dist/diagnose.js

@@ -10,7 +10,7 @@ import {
 } from "./chunk-45EEMHTC.js";
 import {
   confidenceInterval
-} from "./chunk-LMZQ2Z4U.js";
+} from "./chunk-56GC6VYO.js";
 import {
   validateRunRecord
 } from "./chunk-KWRRMR3J.js";

package/dist/{emitter-DEZwY14K.d.ts → emitter-C2rqGH_l.d.ts}

@@ -1,5 +1,5 @@
 import { b as RunOutcome, R as Run, S as Span, c as SpanKind, L as LlmSpan, T as ToolSpan, d as RetrievalSpan, J as JudgeSpan, e as SandboxSpan, E as EventKind, a as TraceEvent, B as BudgetLedgerEntry, A as Artifact, M as Message } from './schema-m0gsnbt3.js';
-import { T as TraceStore } from './store-CKUAgsJz.js';
+import { T as TraceStore } from './store-BcFXE6LG.js';
 
 /**
  * TraceEmitter — hierarchical span builder that auto-parents using an

package/dist/{failure-cluster-CL7IVgkJ.d.ts → failure-cluster-DH9Flgcf.d.ts}

@@ -1,5 +1,5 @@
 import { R as Run, S as Span, a as TraceEvent, F as FailureClass } from './schema-m0gsnbt3.js';
-import { T as TraceStore } from './store-CKUAgsJz.js';
+import { T as TraceStore } from './store-BcFXE6LG.js';
 
 /**
  * Failure taxonomy — canonical classes + a default classifier.

package/dist/{feedback-trajectory-D9OVLrg9.d.ts → feedback-trajectory-BxY0cKfs.d.ts}

@@ -1,4 +1,4 @@
-import { C as ControlEvalResult, a as ControlRunResult, b as ControlStep } from './control-runtime-DuFBYg7A.js';
+import { C as ControlEvalResult, a as ControlRunResult, b as ControlStep } from './control-runtime-Acf9CGhw.js';
 import { D as DatasetSplit, a as DatasetScenario } from './dataset-BbGkaN2I.js';
 
 type FeedbackArtifactType = 'text' | 'code' | 'plan' | 'research' | 'action' | 'ui' | 'decision' | 'data' | 'other';

package/dist/governance/index.d.ts

@@ -1,8 +1,8 @@
 import { c as DatasetManifest } from '../dataset-BbGkaN2I.js';
 import { a as CalibrationResult } from '../judge-calibration-0p2QcWNE.js';
 import { b as OutcomeStore } from '../outcome-store-rnXLEqSn.js';
-import { d as RedTeamReport } from '../red-team-BXHil6c8.js';
-import { T as TraceStore } from '../store-CKUAgsJz.js';
+import { d as RedTeamReport } from '../red-team-BWdoyleI.js';
+import { T as TraceStore } from '../store-BcFXE6LG.js';
 import '../errors-CzMUYo7b.js';
 import '../schema-m0gsnbt3.js';
 

package/dist/{harness-optimizer-EnEnQPsr.d.ts → harness-optimizer-mOl9XX_O.d.ts}

@@ -1,5 +1,5 @@
 import { O as Objective, P as ParetoResult } from './pareto-E-pembql.js';
-import { R as RunScore, a as RunTrace, b as RunScoreWeights } from './run-critic-BAIjX99r.js';
+import { R as RunScore, a as RunTrace, b as RunScoreWeights } from './run-critic-CmMf05uV.js';
 
 interface SteeringRolePrompt {
     system?: string;

package/dist/hosted/index.d.ts

@@ -1,11 +1,11 @@
 import { M as MutableSurface, c as GateDecision } from '../types-BU-7W85F.js';
-import { I as InsightReport } from '../insight-report-BBwvOh6x.js';
+import { I as InsightReport } from '../insight-report-DWl3z9tl.js';
 import '../run-record-e7vj1uZQ.js';
 import '../errors-CzMUYo7b.js';
 import '../schema-m0gsnbt3.js';
-import '../summary-report-DGmUucwQ.js';
-import '../failure-cluster-CL7IVgkJ.js';
-import '../store-CKUAgsJz.js';
+import '../summary-report-BDOFevaT.js';
+import '../failure-cluster-DH9Flgcf.js';
+import '../store-BcFXE6LG.js';
 import '../judge-calibration-0p2QcWNE.js';
 
 /**