@tangle-network/agent-eval 0.128.1 → 0.129.0

package/dist/contract/index.d.ts

@@ -216,9 +216,8 @@ type CostLedgerHandle = Pick<CostLedger, Exclude<keyof CostLedger, 'listPending'
  *     { baseUrl: 'https://router.tangle.tools/v1', apiKey: process.env.KEY },
  *   )
  *
- * This is THE llm-calling seam for agent-eval primitives that need structured
- * output (semantic concept judge, reviewer directives, critic scores). Primitives
- * that need free-form text use `callLlm` and parse output themselves.
+ * `createChatClient` wraps this implementation for provider-neutral package
+ * entry points. Direct callers can use `callLlm` or `callLlmJson`.
  */
 
 interface LlmMessage {
@@ -652,7 +651,7 @@ interface JudgeConfig<TArtifact, TScenario extends Scenario$1 = Scenario$1> {
 /** The canonical judge verdict shape — one declaration, shared by campaign
  *  judges and the multishot judge runner (which re-exports this type).
  *
- *  Scale is PRODUCER-DEFINED: campaign convention is [0,1]; the legacy
+ *  Scale is PRODUCER-DEFINED: campaign convention is [0,1]; the
  *  multishot runner emits 0-10. Cross-scale comparison must go through
  *  `detectScale` (src/campaign/gates/statistical-heldout.ts, used by
  *  promotion-policy) — never renormalize a producer's values in place, as
@@ -872,9 +871,6 @@ interface SurfaceProposer<TFindings = unknown> {
         reason?: string;
     };
 }
-/** Optional vocabulary alias. The loop is the optimizer; this object is the
- * proposer inside that loop. */
-type OptimizationProposer<TFindings = unknown> = SurfaceProposer<TFindings>;
 interface OptimizerConfigBase {
     populationSize: number;
     maxGenerations: number;
@@ -1152,8 +1148,6 @@ interface CampaignAggregates {
     byScenario: Record<string, ScenarioAggregate>;
     /** Canonical campaign accounting, including worker and judge calls. */
     cost: CostLedgerSummary;
-    /** Compatibility alias of `cost.totalCostUsd`. */
-    totalCostUsd: number;
     /** Cells whose dispatch completed, including cells whose later judge failed. */
     cellsExecuted: number;
     cellsSkipped: number;
@@ -1164,7 +1158,7 @@ interface CampaignAggregates {
     cellsDispatchFailed?: number;
     /** Present on results that record failure stages. */
     cellsJudgeFailed?: number;
-    /** Legacy failures whose stage was not recorded. */
+    /** Failures whose stage could not be classified. */
     cellsUnclassifiedFailed?: number;
 }
 interface CampaignResult<TArtifact = unknown, TScenario extends Scenario$1 = Scenario$1> {
@@ -1222,7 +1216,7 @@ interface CampaignStorage {
     write(path: string, content: string | Uint8Array): void;
     /** Append only when the current UTF-8 byte length matches `expectedBytes`.
      * Returns the new length, or undefined when another writer won. */
-    append?(path: string, content: string, expectedBytes: number): number | undefined;
+    append(path: string, content: string, expectedBytes: number): number | undefined;
 }
 /** Node-filesystem storage — the default. Lazily requires `node:fs` so the
  *  module imports cleanly in non-Node runtimes (where the caller passes
@@ -1358,9 +1352,9 @@ interface RunCampaignOptions<TScenario extends Scenario$1, TArtifact> {
     }) => string | undefined;
 }
 /** Durable `<cell>/failure-receipt.json` written before a failed cell can
- * trigger campaign-wide cancellation. The cell keeps its dispatch-only usage
- * fields for compatibility; `cost` covers every settled agent and judge call
- * attributed to this exact run attempt. */
+ * trigger campaign-wide cancellation. The cell records dispatch measurements;
+ * `cost` covers every settled agent and judge call attributed to this exact run
+ * attempt. */
 interface CampaignCellFailureReceipt<TArtifact = unknown> {
     schemaVersion: 1;
     runAttemptId: string;
@@ -1632,42 +1626,27 @@ interface RunImprovementLoopResult<TArtifact, TScenario extends Scenario$1> exte
 declare function runImprovementLoop<TScenario extends Scenario$1, TArtifact>(opts: RunImprovementLoopOptions<TScenario, TArtifact>): Promise<RunImprovementLoopResult<TArtifact, TScenario>>;
 
 /**
- * ChatClient — the single LLM abstraction analysts call.
- *
- * agent-eval already ships an `LlmClient` (OpenAI-compatible, retry,
- * graceful JSON-schema degrade) and judges that talk to `TCloud`. Two
- * mixed patterns force every analyst author to pick a transport, which
- * couples analyst code to runtime concerns (cli-bridge vs router vs
- * sandbox-sdk) it shouldn't know about.
- *
- * `ChatClient` is one interface every analyst takes via `AnalystContext.chat`.
- * The operator decides at the registry boundary which transport binds
- * to it. Analyst code stays transport-agnostic; swapping production
- * (sandbox-sdk) for local dev (cli-bridge) or tests (mock) is a one-
- * line factory call.
- *
- * Designed to coexist: existing `LlmClient` callers and existing
- * `TCloud`-based judges keep working untouched. New analyst code uses
- * `ChatClient`. When old call sites migrate, they pick up budgeting,
- * cancellation, and unified telemetry for free.
+ * Provider-neutral chat contract for every model call made by agent-eval.
+ *
+ * Callers choose the transport at the package boundary with `createChatClient`.
+ * Evaluation code receives canonical requests and results without importing a
+ * provider SDK.
  */
 
 /**
- * Unified chat interface. Mirrors LlmCallRequest/Result so the OpenAI-
- * compatible mental model stays. Two methods: a one-shot `chat()` and
- * an `streamChat()` for future agentic loops (not yet exposed).
+ * Unified chat interface using the package's canonical LLM request and result.
  */
 interface ChatClient {
-    /** Display name of the bound transport — included in telemetry. */
+    /** Display name of the bound transport, included in telemetry. */
     readonly transport: ChatTransport;
-    /** Default model when caller omits — operators bind this per environment. */
+    /** Default model when the caller omits one. */
     readonly defaultModel?: string;
     /** Total provider attempts this transport can make for one chat call. */
     readonly maximumAttempts?: number;
     /** Implementations must enforce `req.maxTokens` when it is present. */
     chat(req: ChatRequest, opts?: ChatCallOpts): Promise<ChatResponse>;
 }
-type ChatTransport = 'router' | 'sandbox-sdk' | 'cli-bridge' | 'direct-provider' | 'mock';
+type ChatTransport = 'router' | 'sandbox-sdk' | 'cli-bridge' | 'direct-provider' | 'custom' | 'mock';
 interface ChatRequest extends Omit<LlmCallRequest, 'model'> {
     /** Optional — falls back to ChatClient.defaultModel. */
     model?: string;
@@ -1683,7 +1662,7 @@ interface ChatCallOpts {
     /** Stable provider idempotency key for retries/redrives of one paid call. */
     idempotencyKey?: string;
 }
-type CreateChatClientOpts = RouterTransportOpts | CliBridgeTransportOpts | DirectProviderTransportOpts | SandboxSdkTransportOpts | MockTransportOpts;
+type CreateChatClientOpts = RouterTransportOpts | CliBridgeTransportOpts | DirectProviderTransportOpts | SandboxSdkTransportOpts | CustomTransportOpts | MockTransportOpts;
 interface BaseTransportOpts {
     defaultModel?: string;
     /** Total provider attempts. Required for opaque transports used in capped runs. */
@@ -1705,15 +1684,18 @@ interface DirectProviderTransportOpts extends BaseTransportOpts {
     apiKey: string;
 }
 /**
- * Sandbox-SDK transport. Provided as a thin pass-through: the caller
- * supplies a callable that mimics LlmClient.chat() against an already-
- * configured Sandbox handle. We don't import the SDK here to keep
- * agent-eval dep-free of @tangle-network/sandbox.
+ * Sandbox-SDK transport. The caller supplies a canonical chat function for an
+ * already-configured Sandbox handle, so agent-eval does not import the SDK.
  */
 interface SandboxSdkTransportOpts extends BaseTransportOpts {
     transport: 'sandbox-sdk';
     chat: (req: ChatRequest, opts?: ChatCallOpts) => Promise<ChatResponse>;
 }
+/** Caller-adapted SDK or transport returning the canonical ChatResponse shape. */
+interface CustomTransportOpts extends BaseTransportOpts {
+    transport: 'custom';
+    chat: (req: ChatRequest, opts?: ChatCallOpts) => Promise<ChatResponse>;
+}
 /**
  * Mock transport for tests. The handler receives the request and returns
  * whatever the test wants. No retries, no JSON-schema degrade.
@@ -2550,6 +2532,18 @@ interface VerifiableRewardExtractionOptions {
      * doesn't report one. Default `0.7`.
      */
     judgeConfidenceFloor?: number;
+    /**
+     * Whether the anti-Goodhart realness gate applies. Default `true`, and the
+     * default is the one every training path must keep.
+     *
+     * Set `false` ONLY for detection and analysis. `rl/reward-hacking.ts` does,
+     * for the same reason it reads `observedScore` for its proxy: it measures the
+     * DIVERGENCE between the judge signal and the deterministic one, and a
+     * deterministic reward that another gate already forced to 0 manufactures
+     * exactly that divergence on exactly the gamed population. The detector would
+     * then be re-reporting a verdict it was supposed to reach independently.
+     */
+    applyRealnessGate?: boolean;
 }
 
 /**
@@ -2835,8 +2829,6 @@ interface JudgeInput {
     costPhase?: string;
     costTags?: Record<string, string>;
     signal?: AbortSignal;
-    /** Exact maximum provider attempts configured on the supplied TCloud client. */
-    tcloudMaximumAttempts?: number;
 }
 
 interface PairedBootstrapResult {
@@ -4747,13 +4739,8 @@ interface AnalystRunSummary {
     reason?: string;
     findings_count: number;
     latency_ms: number;
-    cost_usd: number;
-    /**
-     * Additive receipt for model usage. Registry-produced summaries populate it
-     * even when the analyst emits no findings. `cost_usd` remains the legacy
-     * numeric field; inspect `usage.cost` before treating zero as observed.
-     */
-    usage?: AnalystUsageReceipt;
+    /** Additive model usage and cost provenance for this analyst. */
+    usage: AnalystUsageReceipt;
     /** When `status='failed'`: the error class + message, never the full stack. */
     error?: {
         class: string;
@@ -4815,12 +4802,9 @@ type AnalystRunEvent = {
 /**
  * Typed Ax output for analyst findings.
  *
- * Replaces the legacy `findings:string[]` pattern (where every bullet
- * became a flat-severity `AnalystFinding`) with a structured object
- * array. Ax binds the field as `findings:json[]` so the provider emits
- * native structured output; at the kind-factory boundary we Zod-validate
- * each emitted finding so malformed rows fail loud instead of being
- * silently lifted with default severity.
+ * Ax binds the field as `findings:json[]` so the provider emits native
+ * structured output. At the kind-factory boundary every row is validated
+ * before it becomes an `AnalystFinding`.
  *
  * Why not `f.object().array()` directly in the signature? The Ax
  * signature string `question:string -> findings:json[]` already lets
@@ -4829,31 +4813,7 @@ type AnalystRunEvent = {
  * validation surface independent of which Ax version is installed.
  */
 
-/** Original public schema retained for stored rows and callback contracts. */
 declare const RawAnalystFindingSchema: z.ZodObject<{
-    evidence_uri: z.ZodString;
-    evidence_excerpt: z.ZodOptional<z.ZodString>;
-    severity: z.ZodEnum<{
-        low: "low";
-        high: "high";
-        medium: "medium";
-        critical: "critical";
-        info: "info";
-    }>;
-    claim: z.ZodString;
-    subject: z.ZodOptional<z.ZodString>;
-    confidence: z.ZodNumber;
-    rationale: z.ZodOptional<z.ZodString>;
-    recommended_action: z.ZodOptional<z.ZodString>;
-}, z.core.$strict>;
-type RawAnalystFinding = z.infer<typeof RawAnalystFindingSchema>;
-/**
- * Canonical plural-evidence contract. The preprocessor accepts the original
- * `evidence_uri` / `evidence_excerpt` pair and normalizes it into one evidence
- * item so persisted rows and older model fixtures remain readable. New output
- * always receives the plural shape.
- */
-declare const CanonicalRawAnalystFindingSchema: z.ZodPreprocess<z.ZodObject<{
     evidence: z.ZodArray<z.ZodObject<{
         uri: z.ZodString;
         excerpt: z.ZodOptional<z.ZodString>;
@@ -4870,8 +4830,8 @@ declare const CanonicalRawAnalystFindingSchema: z.ZodPreprocess<z.ZodObject<{
     confidence: z.ZodNumber;
     rationale: z.ZodOptional<z.ZodString>;
     recommended_action: z.ZodOptional<z.ZodString>;
-}, z.core.$strict>>;
-type CanonicalRawAnalystFinding = z.infer<typeof CanonicalRawAnalystFindingSchema>;
+}, z.core.$strict>;
+type RawAnalystFinding = z.infer<typeof RawAnalystFindingSchema>;
 
 /**
  * Analyst-kind factory — the typed way to define trace analysts.
@@ -4880,7 +4840,7 @@ type CanonicalRawAnalystFinding = z.infer<typeof CanonicalRawAnalystFindingSchem
  * and bounded Ax subqueries target one failure-mode lens (failure-mode
  * classification, knowledge gap discovery, knowledge poisoning,
  * self-improvement, ...). Kinds emit findings in the typed
- * `CanonicalRawAnalystFinding` shape via a JSON-array Ax output; the factory
+ * `RawAnalystFinding` shape via a JSON-array Ax output; the factory
  * validates each row with Zod and lifts it into `AnalystFinding[]`.
  *
  * Composition rules:
@@ -4943,7 +4903,7 @@ interface TraceAnalystKindSpec {
  */
 interface TraceAnalystGolden {
     question: string;
-    expected: ReadonlyArray<Omit<CanonicalRawAnalystFinding, 'confidence'>>;
+    expected: ReadonlyArray<Omit<RawAnalystFinding, 'confidence'>>;
 }
 
 /**
@@ -5806,4 +5766,4 @@ interface FromOtelSpansOptions {
 }
 declare function fromOtelSpans(opts: FromOtelSpansOptions): RunRecord[];
 
-export { type AgentEvalAgent, type AgentEvalEvaluateOptions, type AgentEvalImproveOptions, 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 CampaignCostMeter, type CampaignResult, type CampaignStorage, type CampaignTraceWriter, type CandidateExperimentExecutionInput, type ChatClient, 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 CodeSurface, 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 EvalCellScoreDelta, type EvalDimensionDelta, type EvalGenerationDiff, type EvalReportingSuiteInput, type EvalReportingSuiteOptions, type EvalReportingSuiteResult, type EvalRunDiff, 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 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 OptimizationProposer, 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, REFERENCE_EQUIVALENCE_INPUT_LIMITS, REFERENCE_EQUIVALENCE_JUDGE_VERSION, type Recommendation, type ReferenceEquivalenceJudgeInput, type ReferenceEquivalenceJudgeOptions, type ReferenceEquivalenceJudgeResult, type ReferenceEquivalenceScenario, type ReleaseSummary, type RunCampaignOptions, type RunCandidateExperimentOptions, type RunEvalOptions, type RunImprovementLoopOptions, type RunImprovementLoopResult, type RunRecordRejection, type ScalarDistribution, type Scenario$1 as Scenario, type SealCandidateBenchmarkSuiteOptions, type SelfImproveBudget, type SelfImproveOptions, type SelfImproveProgressEvent, 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, diffGenerations, diffRunBaselineToWinner, diffRuns, evalReportingSuite, evaluatePairedMeasurements, externalTextOptimizationMethod, fromClaudeCodeSession, fromCodexSession, fromFeedbackTable, fromKimiCodeSession, fromOpenCodeSession, fromOtelSpans, fromPiSession, fromPigraphSession, fromRunRecordDir, fsCampaignStorage, gepaOptimizationMethod, heldOutGate, inMemoryCampaignStorage, llmJudge, measuredComparisonFromCandidateExperiment, observeCodeAgentSession, paretoPolicy, paretoSignificanceGate, parseAgentTrace, parseCodeAgentJsonl, partitionRunsByAuthoringModel, runCampaign, runCandidateExperiment, runEval, runImprovementLoop, runReferenceEquivalenceJudge, sealCandidateBenchmarkSuite, sealCandidateBenchmarkTask, sealCandidateExperiment, selfImprove, skillOptOptimizationMethod, summarizeExecution, verifyCandidateBenchmarkSuite, verifyCandidateBenchmarkSuiteInputs, verifyCandidateBenchmarkTask, verifyCandidateExperiment, verifyCandidateExperimentComparison };
+export { type AgentEvalAgent, type AgentEvalEvaluateOptions, type AgentEvalImproveOptions, 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 CampaignCostMeter, type CampaignResult, type CampaignStorage, type CampaignTraceWriter, type CandidateExperimentExecutionInput, type ChatClient, 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 CodeSurface, 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 EvalCellScoreDelta, type EvalDimensionDelta, type EvalGenerationDiff, type EvalReportingSuiteInput, type EvalReportingSuiteOptions, type EvalReportingSuiteResult, type EvalRunDiff, 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 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, REFERENCE_EQUIVALENCE_INPUT_LIMITS, REFERENCE_EQUIVALENCE_JUDGE_VERSION, type Recommendation, type ReferenceEquivalenceJudgeInput, type ReferenceEquivalenceJudgeOptions, type ReferenceEquivalenceJudgeResult, type ReferenceEquivalenceScenario, type ReleaseSummary, type RunCampaignOptions, type RunCandidateExperimentOptions, type RunEvalOptions, type RunImprovementLoopOptions, type RunImprovementLoopResult, type RunRecordRejection, type ScalarDistribution, type Scenario$1 as Scenario, type SealCandidateBenchmarkSuiteOptions, type SelfImproveBudget, type SelfImproveOptions, type SelfImproveProgressEvent, 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, diffGenerations, diffRunBaselineToWinner, diffRuns, evalReportingSuite, evaluatePairedMeasurements, externalTextOptimizationMethod, fromClaudeCodeSession, fromCodexSession, fromFeedbackTable, fromKimiCodeSession, fromOpenCodeSession, fromOtelSpans, fromPiSession, fromPigraphSession, fromRunRecordDir, fsCampaignStorage, gepaOptimizationMethod, heldOutGate, inMemoryCampaignStorage, llmJudge, measuredComparisonFromCandidateExperiment, observeCodeAgentSession, paretoPolicy, paretoSignificanceGate, parseAgentTrace, parseCodeAgentJsonl, partitionRunsByAuthoringModel, runCampaign, runCandidateExperiment, runEval, runImprovementLoop, runReferenceEquivalenceJudge, sealCandidateBenchmarkSuite, sealCandidateBenchmarkTask, sealCandidateExperiment, selfImprove, skillOptOptimizationMethod, summarizeExecution, verifyCandidateBenchmarkSuite, verifyCandidateBenchmarkSuiteInputs, verifyCandidateBenchmarkTask, verifyCandidateExperiment, verifyCandidateExperimentComparison };

package/dist/contract/index.js

@@ -14,7 +14,7 @@ import {
 import {
   analyzeRuns,
   summarizeExecution
-} from "../chunk-NACAGYSY.js";
+} from "../chunk-M4YBQKIJ.js";
 import {
   REFERENCE_EQUIVALENCE_INPUT_LIMITS,
   REFERENCE_EQUIVALENCE_JUDGE_VERSION,
@@ -40,7 +40,7 @@ import {
   skillOptOptimizationMethod,
   surfaceContentHash,
   surfaceHash
-} from "../chunk-NKAGIDE2.js";
+} from "../chunk-2QU3YOPR.js";
 import {
   campaignSplitDigest,
   createRunCostLedger,
@@ -48,31 +48,31 @@ import {
   inMemoryCampaignStorage,
   resolveRunDir,
   runCampaign
-} from "../chunk-EZJEIH2R.js";
+} from "../chunk-C6LXANRU.js";
 import {
   buildDefaultAnalystRegistry,
   createChatClient
-} from "../chunk-DJKY2TSY.js";
+} from "../chunk-BSO5JDQH.js";
 import "../chunk-HHWE3POT.js";
 import "../chunk-WGXIEX7P.js";
 import {
   FileSystemOutcomeStore,
   InMemoryOutcomeStore
 } from "../chunk-3RF76KTD.js";
-import "../chunk-NYLOYM6N.js";
+import "../chunk-EG66UGL4.js";
 import {
   campaignCellExecutionEvidence,
   campaignCellJudgeDimensions,
   campaignCellTaskScore,
   campaignCellToRunRecord
-} from "../chunk-2MKQIFS4.js";
-import "../chunk-PBE2LOSS.js";
-import "../chunk-VLOATJQ2.js";
-import "../chunk-DPUHNQLN.js";
+} from "../chunk-E7QXT7SX.js";
+import "../chunk-SFLLL76A.js";
+import "../chunk-TJVT4QFF.js";
+import "../chunk-7FO3TNPI.js";
 import {
   pairedBootstrap
-} from "../chunk-MHELPNRP.js";
-import "../chunk-WS3NZZQQ.js";
+} from "../chunk-ZHTZ4EYI.js";
+import "../chunk-VCZ5FQYW.js";
 import "../chunk-VI2UW6B6.js";
 import {
   readTaskFailureLabels,
@@ -91,8 +91,9 @@ import "../chunk-PC4UYEBM.js";
 import {
   modelHasSnapshot,
   parseRunRecordSafe
-} from "../chunk-2JX3CFMB.js";
+} from "../chunk-56TAVBOK.js";
 import "../chunk-MA6HLL3S.js";
+import "../chunk-OIUOT4QD.js";
 import {
   ValidationError
 } from "../chunk-ONWEPEDO.js";
@@ -512,7 +513,7 @@ async function shipEvalRunToHosted(tenant, opts, summary, raw, runDir) {
       surface,
       cells,
       compositeMean,
-      costUsd: campaign.aggregates.totalCostUsd,
+      costUsd: campaign.aggregates.cost.totalCostUsd,
       durationMs
     };
   }