@tangle-network/agent-eval 0.142.1 → 0.143.0

This diff represents the content of publicly available package versions that have been released to one of the supported registries. The information contained in this diff is provided for informational purposes only and reflects changes between package versions as they appear in their respective public registries.
Files changed (59) hide show
  1. package/CHANGELOG.md +24 -0
  2. package/README.md +21 -0
  3. package/dist/analyst/index.d.ts +5 -5
  4. package/dist/analyst/index.js +2 -2
  5. package/dist/{benchmark-command-DTyVFMUV.js → benchmark-command-CSddvlq7.js} +5 -5
  6. package/dist/{benchmark-command-DTyVFMUV.js.map → benchmark-command-CSddvlq7.js.map} +1 -1
  7. package/dist/benchmarks/index.d.ts +1 -1
  8. package/dist/benchmarks/index.js +1 -1
  9. package/dist/{benchmarks-Ce8GdJbc.js → benchmarks-Dg0B56tK.js} +4 -4
  10. package/dist/{benchmarks-Ce8GdJbc.js.map → benchmarks-Dg0B56tK.js.map} +1 -1
  11. package/dist/campaign/index.d.ts +3 -3
  12. package/dist/campaign/index.js +3 -3
  13. package/dist/{campaign-COFTVy3Q.js → campaign-C7ZGiZYG.js} +7 -5
  14. package/dist/campaign-C7ZGiZYG.js.map +1 -0
  15. package/dist/cli.js +1 -1
  16. package/dist/{client-BoqGxEqx.d.ts → client-Cgl6KasJ.d.ts} +2 -2
  17. package/dist/{client-BoqGxEqx.d.ts.map → client-Cgl6KasJ.d.ts.map} +1 -1
  18. package/dist/contract/index.d.ts +5 -5
  19. package/dist/contract/index.js +3 -3
  20. package/dist/{default-registry-Ci7wAAR8.d.ts → default-registry-D3uqKbo6.d.ts} +2 -2
  21. package/dist/{default-registry-Ci7wAAR8.d.ts.map → default-registry-D3uqKbo6.d.ts.map} +1 -1
  22. package/dist/{dspy-rlm-engine-Bw8Zz9CX.js → dspy-rlm-engine-CJ4sN8Wx.js} +2 -2
  23. package/dist/{dspy-rlm-engine-Bw8Zz9CX.js.map → dspy-rlm-engine-CJ4sN8Wx.js.map} +1 -1
  24. package/dist/hosted/index.d.ts +2 -2
  25. package/dist/{index-CKI1CXTL.d.ts → index-CaOjqCli.d.ts} +4 -4
  26. package/dist/{index-CKI1CXTL.d.ts.map → index-CaOjqCli.d.ts.map} +1 -1
  27. package/dist/{index-DCP4I2Qx.d.ts → index-DY83M1WL.d.ts} +7 -5
  28. package/dist/{index-DCP4I2Qx.d.ts.map → index-DY83M1WL.d.ts.map} +1 -1
  29. package/dist/index.d.ts +8 -8
  30. package/dist/index.js +4 -4
  31. package/dist/multishot/index.d.ts +1 -1
  32. package/dist/openapi.json +1 -1
  33. package/dist/{replay-BqTgoioO.d.ts → replay-BuJM6kLh.d.ts} +2 -2
  34. package/dist/{replay-BqTgoioO.d.ts.map → replay-BuJM6kLh.d.ts.map} +1 -1
  35. package/dist/{reward-hacking-GyN0kMd8.js → reward-hacking-DjTi9HLb.js} +34 -14
  36. package/dist/reward-hacking-DjTi9HLb.js.map +1 -0
  37. package/dist/rl.d.ts +263 -2
  38. package/dist/rl.d.ts.map +1 -1
  39. package/dist/rl.js +332 -3
  40. package/dist/rl.js.map +1 -1
  41. package/dist/{single-run-lock-BcvPQY57.js → single-run-lock-Do-8vBIJ.js} +2 -2
  42. package/dist/{single-run-lock-BcvPQY57.js.map → single-run-lock-Do-8vBIJ.js.map} +1 -1
  43. package/dist/{skill-usage-CNZTk_LT.d.ts → skill-usage-CRqaHi0R.d.ts} +3 -3
  44. package/dist/{skill-usage-CNZTk_LT.d.ts.map → skill-usage-CRqaHi0R.d.ts.map} +1 -1
  45. package/dist/{skillopt-optimization-method-D1dqGzzH.d.ts → skillopt-optimization-method-CwRvi-rL.d.ts} +21 -5
  46. package/dist/skillopt-optimization-method-CwRvi-rL.d.ts.map +1 -0
  47. package/dist/{skillopt-optimization-method-DV6c4ooy.js → skillopt-optimization-method-Di8Xkalt.js} +153 -50
  48. package/dist/skillopt-optimization-method-Di8Xkalt.js.map +1 -0
  49. package/dist/{tool-groups-DVQTy9lq.d.ts → tool-groups-Bqy4A3QB.d.ts} +2 -2
  50. package/dist/{tool-groups-DVQTy9lq.d.ts.map → tool-groups-Bqy4A3QB.d.ts.map} +1 -1
  51. package/dist/traces.d.ts +2 -2
  52. package/dist/{types-Dk7PB7vh.d.ts → types-KEqL1pZc.d.ts} +6 -4
  53. package/dist/{types-Dk7PB7vh.d.ts.map → types-KEqL1pZc.d.ts.map} +1 -1
  54. package/docs/verified-labels-flywheel.md +46 -0
  55. package/package.json +3 -3
  56. package/dist/campaign-COFTVy3Q.js.map +0 -1
  57. package/dist/reward-hacking-GyN0kMd8.js.map +0 -1
  58. package/dist/skillopt-optimization-method-D1dqGzzH.d.ts.map +0 -1
  59. package/dist/skillopt-optimization-method-DV6c4ooy.js.map +0 -1
package/dist/cli.js CHANGED
@@ -1,6 +1,6 @@
1
1
  #!/usr/bin/env node
2
2
  import { o as runRolloutReleaseCli } from "./hf-dataset-XggBupCr.js";
3
- import { n as runAnalystBenchmarkCommand } from "./benchmark-command-DTyVFMUV.js";
3
+ import { n as runAnalystBenchmarkCommand } from "./benchmark-command-CSddvlq7.js";
4
4
  import { a as runRpcBatch, o as runRpcOnce, p as handleVersion, r as startServerAsync, s as buildOpenApi } from "./server-Cu4M3NSO.js";
5
5
  import { writeFileSync } from "node:fs";
6
6
  //#region src/cli-config.ts
@@ -1,5 +1,5 @@
1
1
  import { l as RunTerminalOutcome } from "./run-record-Tb3TTtUn.js";
2
- import { _ as GateDecision, j as MutableSurface } from "./types-Dk7PB7vh.js";
2
+ import { _ as GateDecision, j as MutableSurface } from "./types-KEqL1pZc.js";
3
3
  import { o as InsightReport } from "./insight-report-Bh_8ksel.js";
4
4
  //#region src/hosted/types.d.ts
5
5
  declare const HOSTED_WIRE_VERSION: '2026-07-24.v1';
@@ -199,4 +199,4 @@ declare function hostedClientFromEnv(overrides?: Partial<HostedTenant> & {
199
199
  }): HostedClient | undefined;
200
200
  //#endregion
201
201
  export { UnixNanoTimestamp as _, hostedTenantFromEnv as a, EvalRunGenerationSnapshot as c, HostedIngestHeaders as d, HostedWireVersion as f, TraceSpanEvent as g, IngestTracesRequest as h, hostedClientFromEnv as i, EvalRunStatus as l, IngestResponse as m, HostedTenant as n, EvalRunCellScore as o, IngestEvalRunsRequest as p, createHostedClient as r, EvalRunEvent as s, HostedClient as t, HOSTED_WIRE_VERSION as u };
202
- //# sourceMappingURL=client-BoqGxEqx.d.ts.map
202
+ //# sourceMappingURL=client-Cgl6KasJ.d.ts.map
@@ -1 +1 @@
1
- {"version":3,"file":"client-BoqGxEqx.d.ts","names":[],"sources":["../src/hosted/types.ts","../src/hosted/client.ts"],"mappings":";;;;cAgCa;KACD,2BAA2B;;UAKtB;;EAEf;;EAEA;;EAEA,yBAAyB;;EAEzB;;;KAMU;UAQK;;EAEf;;EAEA;;EAEA;;EAEA,YAAY,eAAe;;EAE3B,iBAAiB;;EAEjB;;EAEA;;UAGe;;EAEf;;;EAGA;;;EAGA,UAAU;;EAEV,OAAO;;EAEP;;EAEA;;EAEA;;;;;;;UAQe;;EAEf;;EAEA;;EAEA;;EAEA,QAAQ;;EAER,QAAQ;;EAER,WAAW;;EAEX,aAAa;;EAEb,eAAe;;EAEf;;EAEA;;EAEA;;EAEA;;;;;;EAMA,gBAAgB;;;;;;KASN;;;;;;UAOK;EACf;EACA;EACA;EACA;EACA,mBAAmB;EACnB,iBAAiB;EACjB,YAAY;EACZ,SAAS;IACP,cAAc;IACd;IACA,aAAa;;EAEf;IAAW;IAAgC;;;EAE3C;;EAEA;;EAEA;;EAEA;;UAKe;EACf,aAAa;EACb,QAAQ;;UAGO;EACf,aAAa;EACb,OAAO;;UAGQ;;EAEf;;EAEA,UAAU;IAAQ;IAAe;;;;;UC1JlB;;EAEf;;EAEA;;EAEA;;EAEA,mBAAmB;;EAEnB;;EAEA;;UAGe;EACf,cAAc,OAAO,cAAc,0BAA0B,QAAQ;EACrE,eAAe,QAAQ,gBAAgB,0BAA0B,QAAQ;EACzE,aAAa,OAAO,kBAAkB,0BAA0B,QAAQ;WAC/D,QAAQ;WACR,aAAa;;iBAyGR,mBAAmB,QAAQ,eAAe;;;;;;;;;;;;;;;;;;;;;;;;;;;iBAsE1C,oBACd,YAAW,QAAQ;EAAkB,MAAM;IAC1C;iBAiBa,oBACd,YAAW,QAAQ;EAAkB,MAAM;IAC1C"}
1
+ {"version":3,"file":"client-Cgl6KasJ.d.ts","names":[],"sources":["../src/hosted/types.ts","../src/hosted/client.ts"],"mappings":";;;;cAgCa;KACD,2BAA2B;;UAKtB;;EAEf;;EAEA;;EAEA,yBAAyB;;EAEzB;;;KAMU;UAQK;;EAEf;;EAEA;;EAEA;;EAEA,YAAY,eAAe;;EAE3B,iBAAiB;;EAEjB;;EAEA;;UAGe;;EAEf;;;EAGA;;;EAGA,UAAU;;EAEV,OAAO;;EAEP;;EAEA;;EAEA;;;;;;;UAQe;;EAEf;;EAEA;;EAEA;;EAEA,QAAQ;;EAER,QAAQ;;EAER,WAAW;;EAEX,aAAa;;EAEb,eAAe;;EAEf;;EAEA;;EAEA;;EAEA;;;;;;EAMA,gBAAgB;;;;;;KASN;;;;;;UAOK;EACf;EACA;EACA;EACA;EACA,mBAAmB;EACnB,iBAAiB;EACjB,YAAY;EACZ,SAAS;IACP,cAAc;IACd;IACA,aAAa;;EAEf;IAAW;IAAgC;;;EAE3C;;EAEA;;EAEA;;EAEA;;UAKe;EACf,aAAa;EACb,QAAQ;;UAGO;EACf,aAAa;EACb,OAAO;;UAGQ;;EAEf;;EAEA,UAAU;IAAQ;IAAe;;;;;UC1JlB;;EAEf;;EAEA;;EAEA;;EAEA,mBAAmB;;EAEnB;;EAEA;;UAGe;EACf,cAAc,OAAO,cAAc,0BAA0B,QAAQ;EACrE,eAAe,QAAQ,gBAAgB,0BAA0B,QAAQ;EACzE,aAAa,OAAO,kBAAkB,0BAA0B,QAAQ;WAC/D,QAAQ;WACR,aAAa;;iBAyGR,mBAAmB,QAAQ,eAAe;;;;;;;;;;;;;;;;;;;;;;;;;;;iBAsE1C,oBACd,YAAW,QAAQ;EAAkB,MAAM;IAC1C;iBAiBa,oBACd,YAAW,QAAQ;EAAkB,MAAM;IAC1C"}
@@ -2,12 +2,12 @@ import { c as CostLedgerHandle, f as CostLedgerSummary, m as CostReceipt } from
2
2
  import { a as RunRecord, n as RunCostProvenance, s as RunSplitTag } from "../run-record-Tb3TTtUn.js";
3
3
  import { A as ChatClient, F as CreateChatClientOpts, V as createChatClient } from "../types-BjMFz88h.js";
4
4
  import { h as ProposalFindingOrigin, i as AnalystFinding, m as ProposalFinding, v as makeProposalFinding } from "../types-D3jh6F98.js";
5
- import { n as buildDefaultAnalystRegistry, t as DefaultAnalystRegistryOptions } from "../default-registry-Ci7wAAR8.js";
6
- 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, k as LabeledScenarioStore, 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-Dk7PB7vh.js";
5
+ import { n as buildDefaultAnalystRegistry, t as DefaultAnalystRegistryOptions } from "../default-registry-D3uqKbo6.js";
6
+ 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, k as LabeledScenarioStore, 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-KEqL1pZc.js";
7
7
  import { a as FailureClusterInsight, c as JudgeInsight, d as Recommendation, f as ReleaseSummary, i as FailureClassTally, l as LiftInsight, m as TokenUsageInsight, n as ExecutionErrorOutcomeCell, o as InsightReport, p as ScalarDistribution, r as ExecutionInsight, s as InterRaterInsight, t as CostProvenanceSummary, u as OutcomeCorrelationInsight } from "../insight-report-Bh_8ksel.js";
8
- import { A as summarizeExecution, D as ExecutionReport, E as AnalyzeRunsOptions, O as SummarizeExecutionOptions, k as analyzeRuns } from "../tool-groups-DVQTy9lq.js";
9
- import { A as RunEvalOptions, B as OptimizerModelBudget, Bt as ComparisonCost, C as RunImprovementLoopOptions, Cn as ReferenceEquivalenceJudgeResult, D as RunOptimizationOptions, Dn as LlmJudgeDimension, E as PremeasuredOptimizationBaseline, En as runReferenceEquivalenceJudge, F as GepaOptimizationMethodConfig, Ft as ExternalTextOptimizerContext, G as ObjectiveSource, Gt as OptimizationMethodProvenance, H as AxisVerdict, Ht as OptimizationMethodComparison, I as GepaOptimizationRecipe, It as ExternalTextOptimizerResult, J as PromotionPolicy, K as ParetoSignificanceGateOptions, Kt as OptimizationMethodResult, L as GepaRunnerCommand, Lt as ExternalOptimizationExample, M as GepaAdaptiveEngineRun, Mt as composeGate, N as GepaEngineOptions, Nt as externalTextOptimizationMethod, On as LlmJudgeOptions, P as GepaEngineRun, Pt as ExternalTextOptimizationMethodConfig, R as gepaOptimizationMethod, Rt as ExternalTextEvaluationResponse, Sn as ReferenceEquivalenceJudgeOptions, T as runImprovementLoop, Tn as createReferenceEquivalenceJudge, U as BuildEvidenceVectorOptions, Ut as OptimizationMethodInput, V as AxisEvidence, Vt as OptimizationMethod, W as EvidenceVector, X as paretoPolicy, Xt as OptimizationTokenUsage, Y as buildEvidenceVector, Yt as OptimizationPackageSource, Z as paretoSignificanceGate, Zt as compareOptimizationMethods, bn as REFERENCE_EQUIVALENCE_JUDGE_VERSION, dt as DefaultProductionGateCheck, en as CampaignCellFailureReceipt, ft as DefaultProductionGateOptions, i as skillOptOptimizationMethod, in as RunCampaignOptions, j as runEval, kn as llmJudge, ln as fsCampaignStorage, lt as HeldOutGateOptions, mn as campaignSplitDigest, mt as defaultProductionGate, n as SkillOptRunnerCommand, on as runCampaign, ot as PowerPreflight, p as LoopProvenanceRecord, pt as DefaultProductionRewardHackingOptions, q as PromotionObjective, r as SkillOptTrainerConfig, sn as CampaignStorage, t as SkillOptOptimizationMethodConfig, un as inMemoryCampaignStorage, ut as heldOutGate, w as RunImprovementLoopResult, wn as ReferenceEquivalenceScenario, xn as ReferenceEquivalenceJudgeInput, yn as REFERENCE_EQUIVALENCE_INPUT_LIMITS, z as OpenAICompatibleOptimizerModel, zt as CompareOptimizationMethodsOptions } from "../skillopt-optimization-method-D1dqGzzH.js";
10
- import { c as EvalRunGenerationSnapshot, g as TraceSpanEvent, n as HostedTenant, o as EvalRunCellScore, s as EvalRunEvent } from "../client-BoqGxEqx.js";
8
+ import { A as summarizeExecution, D as ExecutionReport, E as AnalyzeRunsOptions, O as SummarizeExecutionOptions, k as analyzeRuns } from "../tool-groups-Bqy4A3QB.js";
9
+ import { A as RunEvalOptions, B as OptimizerModelBudget, Bt as ComparisonCost, C as RunImprovementLoopOptions, Cn as ReferenceEquivalenceJudgeResult, D as RunOptimizationOptions, Dn as LlmJudgeDimension, E as PremeasuredOptimizationBaseline, En as runReferenceEquivalenceJudge, F as GepaOptimizationMethodConfig, Ft as ExternalTextOptimizerContext, G as ObjectiveSource, Gt as OptimizationMethodProvenance, H as AxisVerdict, Ht as OptimizationMethodComparison, I as GepaOptimizationRecipe, It as ExternalTextOptimizerResult, J as PromotionPolicy, K as ParetoSignificanceGateOptions, Kt as OptimizationMethodResult, L as GepaRunnerCommand, Lt as ExternalOptimizationExample, M as GepaAdaptiveEngineRun, Mt as composeGate, N as GepaEngineOptions, Nt as externalTextOptimizationMethod, On as LlmJudgeOptions, P as GepaEngineRun, Pt as ExternalTextOptimizationMethodConfig, R as gepaOptimizationMethod, Rt as ExternalTextEvaluationResponse, Sn as ReferenceEquivalenceJudgeOptions, T as runImprovementLoop, Tn as createReferenceEquivalenceJudge, U as BuildEvidenceVectorOptions, Ut as OptimizationMethodInput, V as AxisEvidence, Vt as OptimizationMethod, W as EvidenceVector, X as paretoPolicy, Xt as OptimizationTokenUsage, Y as buildEvidenceVector, Yt as OptimizationPackageSource, Z as paretoSignificanceGate, Zt as compareOptimizationMethods, bn as REFERENCE_EQUIVALENCE_JUDGE_VERSION, dt as DefaultProductionGateCheck, en as CampaignCellFailureReceipt, ft as DefaultProductionGateOptions, i as skillOptOptimizationMethod, in as RunCampaignOptions, j as runEval, kn as llmJudge, ln as fsCampaignStorage, lt as HeldOutGateOptions, mn as campaignSplitDigest, mt as defaultProductionGate, n as SkillOptRunnerCommand, on as runCampaign, ot as PowerPreflight, p as LoopProvenanceRecord, pt as DefaultProductionRewardHackingOptions, q as PromotionObjective, r as SkillOptTrainerConfig, sn as CampaignStorage, t as SkillOptOptimizationMethodConfig, un as inMemoryCampaignStorage, ut as heldOutGate, w as RunImprovementLoopResult, wn as ReferenceEquivalenceScenario, xn as ReferenceEquivalenceJudgeInput, yn as REFERENCE_EQUIVALENCE_INPUT_LIMITS, z as OpenAICompatibleOptimizerModel, zt as CompareOptimizationMethodsOptions } from "../skillopt-optimization-method-CwRvi-rL.js";
10
+ import { c as EvalRunGenerationSnapshot, g as TraceSpanEvent, n as HostedTenant, o as EvalRunCellScore, s as EvalRunEvent } from "../client-Cgl6KasJ.js";
11
11
  import { i as InMemoryOutcomeStore, n as FileSystemOutcomeStore, o as OutcomeStore, r as FileSystemOutcomeStoreOptions, t as DeploymentOutcome } from "../outcome-store-BYHIuO0e.js";
12
12
  import { AgentCandidateBenchmarkCellRef, AgentCandidateBenchmarkSuiteInputs, AgentCandidateBenchmarkTask, AgentCandidateBenchmarkTaskMaterial, AgentCandidateBundle, AgentCandidateEvaluationPolicy, AgentCandidateExperiment, AgentCandidateExperimentMaterial, AgentCandidateExperimentMeasurement, AgentImprovementCost, AgentImprovementMeasuredComparison, AgentProfileImprovementExperiment, AgentProfileImprovementExperimentMaterial, AgentProfileImprovementMeasuredComparison, AgentProfileImprovementMeasurement, AgentProfileImprovementRunCell, AgentProfileImprovementRunReceipt, AgentProfileImprovementSuiteInputs, AgentProfileImprovementTask, AgentProfileImprovementTaskMaterial, CandidateExecutionEvidence, Sha256Digest } from "@tangle-network/agent-interface";
13
13
  //#region src/contract/self-improve.d.ts
@@ -5,14 +5,14 @@ import { C as createChatClient, t as buildDefaultAnalystRegistry } from "../defa
5
5
  import { LLM_MODEL_ATTR_KEYS, SPAN_KIND_ATTR_KEYS } from "../trace-attributes.js";
6
6
  import { nt as classifyOtlpSpanRole, rt as isOtlpModelCall } from "../kind-factory-Bvwe3pup.js";
7
7
  import { s as makeProposalFinding } from "../usage-receipt-CgxMEBZq.js";
8
- import { i as inMemoryCampaignStorage, n as createRunCostLedger, r as fsCampaignStorage } from "../single-run-lock-BcvPQY57.js";
8
+ import { i as inMemoryCampaignStorage, n as createRunCostLedger, r as fsCampaignStorage } from "../single-run-lock-Do-8vBIJ.js";
9
9
  import { m as mapConcurrentRange } from "../ledger-core-Dxz0Rkwa.js";
10
- import { A as assertOptimizationResult, B as surfaceHash, Dt as decidePairedPromotion, J as runCampaign, O as composeGate, S as defaultProductionGate, T as heldoutSignificance, Y as resolveRunDir, _ as buildEvidenceVector, a as emitLoopProvenance, at as createReferenceEquivalenceJudge, b as powerPreflight, d as runImprovementLoop, et as campaignSplitDigest, g as gepaOptimizationMethod, ht as llmJudge, it as REFERENCE_EQUIVALENCE_JUDGE_VERSION, j as compareOptimizationMethods, k as externalTextOptimizationMethod, o as loopProvenanceArgsFromResult, ot as runReferenceEquivalenceJudge, p as runEval, rt as REFERENCE_EQUIVALENCE_INPUT_LIMITS, t as skillOptOptimizationMethod, v as paretoPolicy, x as heldOutGate, y as paretoSignificanceGate, z as surfaceContentHash } from "../skillopt-optimization-method-DV6c4ooy.js";
10
+ import { A as assertOptimizationResult, B as surfaceHash, Dt as decidePairedPromotion, J as runCampaign, O as composeGate, S as defaultProductionGate, T as heldoutSignificance, Y as resolveRunDir, _ as buildEvidenceVector, a as emitLoopProvenance, at as createReferenceEquivalenceJudge, b as powerPreflight, d as runImprovementLoop, et as campaignSplitDigest, g as gepaOptimizationMethod, ht as llmJudge, it as REFERENCE_EQUIVALENCE_JUDGE_VERSION, j as compareOptimizationMethods, k as externalTextOptimizationMethod, o as loopProvenanceArgsFromResult, ot as runReferenceEquivalenceJudge, p as runEval, rt as REFERENCE_EQUIVALENCE_INPUT_LIMITS, t as skillOptOptimizationMethod, v as paretoPolicy, x as heldOutGate, y as paretoSignificanceGate, z as surfaceContentHash } from "../skillopt-optimization-method-Di8Xkalt.js";
11
11
  import { E as pairedBootstrap } from "../statistics-ByxzSiOM.js";
12
12
  import { i as parseRunRecordSafe, r as modelHasSnapshot } from "../run-record-vRgqWmJw.js";
13
13
  import { a as recordAggregateMeasurements, i as readTaskFailureLabels, o as summarizeExecutionMeasurements, s as summarizeTraceErrors, t as extractUsage } from "../extract-usage-7l1Xq5ti.js";
14
14
  import { n as summarizeExecution, t as analyzeRuns } from "../analyze-runs-BScZvqMV.js";
15
- import { a as campaignCellExecutionEvidence, c as campaignCellToRunRecord, o as campaignCellJudgeDimensions, s as campaignCellTaskScore } from "../reward-hacking-GyN0kMd8.js";
15
+ import { c as campaignCellTaskScore, l as campaignCellToRunRecord, o as campaignCellExecutionEvidence, s as campaignCellJudgeDimensions } from "../reward-hacking-DjTi9HLb.js";
16
16
  import { n as InMemoryOutcomeStore, t as FileSystemOutcomeStore } from "../outcome-store-ChBKlTd_.js";
17
17
  import { t as createHostedClient } from "../client-LIuo-KPv.js";
18
18
  import { dirname, join } from "node:path";
@@ -1,7 +1,7 @@
1
1
  import { N as TraceAnalystSpan, i as AnalystFinding, n as AnalystContext, x as TraceAnalysisStore } from "./types-D3jh6F98.js";
2
2
  import { l as ExactCapableAnalyst } from "./exact-types-B0lJV3tu.js";
3
3
  import { D as AnalystRegistryOptions, E as AnalystRegistry } from "./completion-verifier-D15NHYSk.js";
4
- import { i as TraceAnalysisEngine, o as TraceAnalysisEngineResult, p as RawAnalystFinding, s as TraceAnalystLimits, t as TraceToolGroupName } from "./tool-groups-DVQTy9lq.js";
4
+ import { i as TraceAnalysisEngine, o as TraceAnalysisEngineResult, p as RawAnalystFinding, s as TraceAnalystLimits, t as TraceToolGroupName } from "./tool-groups-Bqy4A3QB.js";
5
5
  //#region src/trace-analyst/behavioral-metrics.d.ts
6
6
  type SuboptimalCode = 'monotonic-input-growth' | 'output-length-decay' | 'single-tool-dependency' | 'no-self-verification';
7
7
  interface SuboptimalSignal {
@@ -115,4 +115,4 @@ interface DefaultAnalystRegistryOptions {
115
115
  declare function buildDefaultAnalystRegistry(options?: DefaultAnalystRegistryOptions): AnalystRegistry;
116
116
  //#endregion
117
117
  export { createTraceAnalyst as a, runTraceAnalyst as c, deriveEfficiencyFindings as d, BehavioralMetrics as f, computeTraceMetrics as g, SuboptimalSignal as h, TraceAnalystDefinition as i, BehavioralAnalystOptions as l, SuboptimalCode as m, buildDefaultAnalystRegistry as n, renderPriorFindings as o, BehavioralTokenSequence as p, CreateTraceAnalystOptions as r, renderUpstreamFindings as s, DefaultAnalystRegistryOptions as t, behavioralAnalyst as u };
118
- //# sourceMappingURL=default-registry-Ci7wAAR8.d.ts.map
118
+ //# sourceMappingURL=default-registry-D3uqKbo6.d.ts.map
@@ -1 +1 @@
1
- {"version":3,"file":"default-registry-Ci7wAAR8.d.ts","names":[],"sources":["../src/trace-analyst/behavioral-metrics.ts","../src/analyst/behavioral-analyst.ts","../src/analyst/kind-factory.ts","../src/analyst/default-registry.ts"],"mappings":";;;;;KAoBY;UAMK;EACf,MAAM;EACN;;EAEA;;EAEA,UAAU;;UAGK;;EAEf;EACA;;EAEA,gBAAgB;;EAEhB;EACA;EACA,eAAe;EACf;EACA;;EAEA;EACA;EACA,SAAS;;UAGM;EACf;EACA;EACA,sBAAsB;EACtB,uBAAuB;;;;;;iBAoFT,oBAAoB,gBAAgB,qBAAqB;;;UC9GxD;;EAEf;;EAEA;;;;;;iBA+Dc,yBACd,SAAS,mBACT;EAAQ;EAAoB;IAC3B;;iBAkCa,kBACd,UAAS,2BACR,oBAAoB;;;;UCrHN;EACf;EACA;EACA;EACA;;EAEA,sBAAsB,SAAS;;EAE/B;;EAEA,WAAW;;EAEX,kBACE,OAAO,oBACP,SAAS,wCACe;EAC1B,SAAS,QAAQ;EACjB,eAAe,KAAK,mBAAmB,SAAS,mBAAmB;EACnE;EACA;;UAGe;EACf,QAAQ;;EAER;EACA;;;iBAIoB,gBAAgB;EACpC,YAAY;EACZ,QAAQ;EACR,OAAO;EACP,SAAS;EACT;IACE,QAAQ;;iBAgFI,mBACd,YAAY,wBACZ,SAAS,4BACR,oBAAoB;iBA+QP,oBAAoB,OAAO;iBAe3B,uBAAuB,UAAU;;;UCnahC;;;;;;EAMf,SAAS;EACT,uBAAuB;;EAEvB;EACA,aAAa;EACb,WAAW;;iBAGG,4BACd,UAAS,gCACR"}
1
+ {"version":3,"file":"default-registry-D3uqKbo6.d.ts","names":[],"sources":["../src/trace-analyst/behavioral-metrics.ts","../src/analyst/behavioral-analyst.ts","../src/analyst/kind-factory.ts","../src/analyst/default-registry.ts"],"mappings":";;;;;KAoBY;UAMK;EACf,MAAM;EACN;;EAEA;;EAEA,UAAU;;UAGK;;EAEf;EACA;;EAEA,gBAAgB;;EAEhB;EACA;EACA,eAAe;EACf;EACA;;EAEA;EACA;EACA,SAAS;;UAGM;EACf;EACA;EACA,sBAAsB;EACtB,uBAAuB;;;;;;iBAoFT,oBAAoB,gBAAgB,qBAAqB;;;UC9GxD;;EAEf;;EAEA;;;;;;iBA+Dc,yBACd,SAAS,mBACT;EAAQ;EAAoB;IAC3B;;iBAkCa,kBACd,UAAS,2BACR,oBAAoB;;;;UCrHN;EACf;EACA;EACA;EACA;;EAEA,sBAAsB,SAAS;;EAE/B;;EAEA,WAAW;;EAEX,kBACE,OAAO,oBACP,SAAS,wCACe;EAC1B,SAAS,QAAQ;EACjB,eAAe,KAAK,mBAAmB,SAAS,mBAAmB;EACnE;EACA;;UAGe;EACf,QAAQ;;EAER;EACA;;;iBAIoB,gBAAgB;EACpC,YAAY;EACZ,QAAQ;EACR,OAAO;EACP,SAAS;EACT;IACE,QAAQ;;iBAgFI,mBACd,YAAY,wBACZ,SAAS,4BACR,oBAAoB;iBA+QP,oBAAoB,OAAO;iBAe3B,uBAAuB,UAAU;;;UCnahC;;;;;;EAMf,SAAS;EACT,uBAAuB;;EAEvB;EACA,aAAa;EACb,WAAW;;iBAGG,4BACd,UAAS,gCACR"}
@@ -1,6 +1,6 @@
1
1
  import { s as resolveModelPricing } from "./metrics-C9YY1OcL.js";
2
2
  import { D as RawAnalystFindingSchema } from "./kind-factory-Bvwe3pup.js";
3
- import { S as removeCredentialEnvironment, d as runWithCleanup, f as startExternalOptimizerModelProxy, h as sendJson, l as runExternalOptimizerProcess, m as listenLocal, p as closeServer } from "./single-run-lock-BcvPQY57.js";
3
+ import { S as removeCredentialEnvironment, d as runWithCleanup, f as startExternalOptimizerModelProxy, h as sendJson, l as runExternalOptimizerProcess, m as listenLocal, p as closeServer } from "./single-run-lock-Do-8vBIJ.js";
4
4
  import { randomBytes } from "node:crypto";
5
5
  import { createServer } from "node:http";
6
6
  //#region src/analyst/trace-tool-callback.ts
@@ -351,4 +351,4 @@ function isRecord(value) {
351
351
  //#endregion
352
352
  export { createDspyRlmTraceEngine as t };
353
353
 
354
- //# sourceMappingURL=dspy-rlm-engine-Bw8Zz9CX.js.map
354
+ //# sourceMappingURL=dspy-rlm-engine-CJ4sN8Wx.js.map
@@ -1 +1 @@
1
- {"version":3,"file":"dspy-rlm-engine-Bw8Zz9CX.js","names":["isRecord"],"sources":["../src/analyst/trace-tool-callback.ts","../src/analyst/dspy-rlm-engine.ts"],"sourcesContent":["import { randomBytes } from 'node:crypto'\nimport { createServer, type IncomingMessage, type ServerResponse } from 'node:http'\nimport { closeServer, listenLocal, sendJson } from '../campaign/external-optimizer-http'\nimport type { TraceAnalysisToolDescriptor } from '../trace-analyst/tools'\n\nconst MAX_REQUEST_BYTES = 1_000_000\nconst MAX_RESPONSE_BYTES = 4_000_000\n\nexport interface TraceToolCallback {\n url: string\n token: string\n calls: () => number\n close: () => Promise<void>\n}\n\n/** Expose one bounded trace-tool set only on an authenticated loopback socket. */\nexport async function startTraceToolCallback(args: {\n tools: readonly TraceAnalysisToolDescriptor[]\n maxCalls: number\n signal?: AbortSignal\n}): Promise<TraceToolCallback> {\n if (!Number.isSafeInteger(args.maxCalls) || args.maxCalls <= 0) {\n throw new TypeError('trace tool callback maxCalls must be a positive safe integer')\n }\n args.signal?.throwIfAborted()\n const byName = new Map(args.tools.map((tool) => [tool.name, tool]))\n if (byName.size !== args.tools.length) {\n throw new Error('trace tool callback received duplicate tool names')\n }\n\n const token = randomBytes(32).toString('hex')\n let calls = 0\n let accepting = true\n let closePromise: Promise<void> | undefined\n const activeControllers = new Set<AbortController>()\n const activeHandlers = new Set<Promise<void>>()\n const server = createServer((request, response) => {\n if (!accepting) {\n sendJsonIfOpen(response, 503, { error: 'trace tool callback is closing' })\n return\n }\n const controller = new AbortController()\n const abortRequest = (): void => {\n request.destroy()\n response.destroy()\n }\n activeControllers.add(controller)\n controller.signal.addEventListener('abort', abortRequest, { once: true })\n\n let handler!: Promise<void>\n handler = handleRequest(request, response, controller.signal).finally(() => {\n controller.signal.removeEventListener('abort', abortRequest)\n activeControllers.delete(controller)\n activeHandlers.delete(handler)\n })\n activeHandlers.add(handler)\n void handler.catch(() => undefined)\n })\n const port = await listenLocal(server)\n const close = (): Promise<void> => {\n closePromise ??= closeCallback()\n return closePromise\n }\n const onAbort = (): void => {\n void close().catch(() => undefined)\n }\n args.signal?.addEventListener('abort', onAbort, { once: true })\n if (args.signal?.aborted) onAbort()\n\n return {\n url: `http://127.0.0.1:${port}/call`,\n token,\n calls: () => calls,\n close,\n }\n\n async function handleRequest(\n request: IncomingMessage,\n response: ServerResponse,\n signal: AbortSignal,\n ): Promise<void> {\n try {\n if (request.method !== 'POST' || request.url !== '/call') {\n sendJsonIfOpen(response, 404, { error: 'not found' })\n return\n }\n if (request.headers.authorization !== `Bearer ${token}`) {\n sendJsonIfOpen(response, 401, { error: 'unauthorized' })\n return\n }\n if (calls >= args.maxCalls) {\n sendJsonIfOpen(response, 429, { error: 'trace tool call limit reached' })\n return\n }\n const body = await readJson(request)\n if (!isRecord(body) || typeof body.name !== 'string' || !('args' in body)) {\n sendJsonIfOpen(response, 400, { error: 'name and args are required' })\n return\n }\n const tool = byName.get(body.name)\n if (!tool) {\n sendJsonIfOpen(response, 404, { error: `unknown trace tool '${body.name}'` })\n return\n }\n calls += 1\n const result = await tool.handler(body.args, { signal })\n const encoded = JSON.stringify({ result })\n if (Buffer.byteLength(encoded) > MAX_RESPONSE_BYTES) {\n sendJsonIfOpen(response, 413, { error: 'trace tool response too large' })\n return\n }\n response.writeHead(200, {\n 'content-type': 'application/json; charset=utf-8',\n 'content-length': String(Buffer.byteLength(encoded)),\n })\n response.end(encoded)\n } catch (error) {\n sendJsonIfOpen(response, signal.aborted ? 499 : 400, {\n error: error instanceof Error ? error.message : String(error),\n })\n }\n }\n\n async function closeCallback(): Promise<void> {\n args.signal?.removeEventListener('abort', onAbort)\n accepting = false\n const closingServer = closeServer(server)\n server.closeIdleConnections?.()\n for (const controller of activeControllers) controller.abort()\n const [serverResult] = await Promise.allSettled([\n closingServer,\n waitForActiveHandlers(activeHandlers),\n ])\n if (activeControllers.size !== 0 || activeHandlers.size !== 0) {\n throw new Error('trace tool callback closed with active requests')\n }\n if (serverResult?.status === 'rejected') throw serverResult.reason\n }\n}\n\nasync function waitForActiveHandlers(activeHandlers: Set<Promise<void>>): Promise<void> {\n while (activeHandlers.size > 0) {\n await Promise.allSettled([...activeHandlers])\n }\n}\n\nfunction readJson(request: IncomingMessage): Promise<unknown> {\n return new Promise((resolve, reject) => {\n let size = 0\n const chunks: Buffer[] = []\n request.on('data', (chunk: Buffer) => {\n size += chunk.length\n if (size > MAX_REQUEST_BYTES) {\n reject(new Error('trace tool request too large'))\n request.destroy()\n return\n }\n chunks.push(chunk)\n })\n request.on('error', reject)\n request.on('end', () => {\n try {\n resolve(JSON.parse(Buffer.concat(chunks).toString('utf8')))\n } catch (error) {\n reject(error)\n }\n })\n })\n}\n\nfunction sendJsonIfOpen(response: ServerResponse, status: number, body: unknown): void {\n if (response.destroyed || response.writableEnded) return\n sendJson(response, status, body)\n}\n\nfunction isRecord(value: unknown): value is Record<string, unknown> {\n return typeof value === 'object' && value !== null && !Array.isArray(value)\n}\n","import {\n type ExternalOptimizerModelProxy,\n type ExternalOptimizerRunnerCommand,\n removeCredentialEnvironment,\n} from '../campaign/external-optimizer-contracts'\nimport { startExternalOptimizerModelProxy } from '../campaign/external-optimizer-model-proxy'\nimport { runWithCleanup } from '../campaign/external-optimizer-resources'\nimport { runExternalOptimizerProcess } from '../campaign/external-optimizer-subprocess'\nimport type { CustomTokenPricing } from '../cost-ledger'\nimport { resolveModelPricing } from '../metrics'\nimport type { TraceAnalysisEngine, TraceAnalysisEngineResult } from './engine'\nimport { type RawAnalystFinding, RawAnalystFindingSchema } from './finding-signature'\nimport { startTraceToolCallback } from './trace-tool-callback'\n\nconst DEFAULT_TIMEOUT_MS = 10 * 60_000\n// 4096 is below what current coding models emit for a full findings array:\n// glm-5.2 through an OpenAI-compatible gateway returns 8192 and the request is\n// rejected outright (`502 — provider reported 8192 completion tokens,\n// exceeding requested limit 4096`), so the default failed 2/2 smoke cases\n// before any analysis ran. The cap exists to bound spend, and `maxCostUsd`\n// already does that directly, so it starts above what a real completion needs.\nconst DEFAULT_MODEL_OUTPUT_TOKENS = 16_384\nconst DEFAULT_MAX_COST_USD = 1\nconst MAX_MODEL_REQUEST_BYTES = 16 * 1024 * 1024\nconst MAX_MODEL_RESPONSE_BYTES = 4 * 1024 * 1024\nconst BRIDGE_MODULE = 'agent_eval_rpc.dspy_rlm_bridge'\n/** Bumped whenever this engine's execution behavior changes. */\nconst DSPY_RLM_ENGINE_VERSION = '1.0.0'\n\nexport interface DspyRlmTraceEngineOptions {\n baseUrl: string\n apiKey: string\n model: string\n /** Exact provider rates. Required when the model is absent from the pricing table. */\n pricing?: CustomTokenPricing\n /** Maximum provider spend for one investigation. Default: 1 USD. */\n maxCostUsd?: number\n /** Controller response cap. Default: 16384. */\n maxOutputTokens?: number\n /**\n * Thinking tokens one controller turn may bill on top of its completion.\n * A reasoning model bills these beyond `maxOutputTokens`, so the cost\n * reservation must cover them. Default: four times the completion cap.\n */\n maxReasoningTokens?: number\n /**\n * How the controller's reasoning and code fields are obtained.\n *\n * `tolerant` parses marker output strictly first, then recovers the fields\n * deterministically from prose plus a fenced code block — the shape coding\n * models naturally emit — at no extra model cost. `two-step` extracts with a\n * second call per turn. `chat` accepts marker output only. Default:\n * `tolerant`.\n */\n controlAdapter?: 'chat' | 'two-step' | 'tolerant'\n /** Python command used to load agent-eval-rpc[dspy]. Default: python. */\n runner?: ExternalOptimizerRunnerCommand\n /** Whole investigation deadline. Default: 10 minutes. */\n timeoutMs?: number\n}\n\n/** Use the official DSPy RLM as a bounded recursive trace-analysis engine. */\nexport function createDspyRlmTraceEngine(options: DspyRlmTraceEngineOptions): TraceAnalysisEngine {\n assertOptions(options)\n const maxOutputTokens = options.maxOutputTokens ?? DEFAULT_MODEL_OUTPUT_TOKENS\n if (!Number.isSafeInteger(maxOutputTokens) || maxOutputTokens <= 0) {\n throw new TypeError('DSPy RLM maxOutputTokens must be a positive safe integer')\n }\n const controlAdapter = options.controlAdapter ?? 'tolerant'\n const maxReasoningTokens = options.maxReasoningTokens ?? maxOutputTokens * 4\n if (!Number.isSafeInteger(maxReasoningTokens) || maxReasoningTokens < 0) {\n throw new TypeError('DSPy RLM maxReasoningTokens must be a non-negative safe integer')\n }\n const timeoutMs = options.timeoutMs ?? DEFAULT_TIMEOUT_MS\n if (!Number.isSafeInteger(timeoutMs) || timeoutMs <= 0) {\n throw new TypeError('DSPy RLM timeoutMs must be a positive safe integer')\n }\n const maxCostUsd = options.maxCostUsd ?? DEFAULT_MAX_COST_USD\n if (!Number.isFinite(maxCostUsd) || maxCostUsd <= 0) {\n throw new TypeError('DSPy RLM maxCostUsd must be positive and finite')\n }\n const pricing = options.pricing ?? pricingForModel(options.model)\n\n const runner = sanitizedRunner(options.runner)\n return {\n id: 'dspy-rlm',\n description: 'Official DSPy RLM with bounded trace tools and metered model calls.',\n model: options.model,\n version: DSPY_RLM_ENGINE_VERSION,\n executionConfig: {\n bridge_module: BRIDGE_MODULE,\n base_url: options.baseUrl,\n model: options.model,\n api_key_provided: true,\n pricing: { ...pricing },\n max_cost_usd: maxCostUsd,\n max_output_tokens: maxOutputTokens,\n max_reasoning_tokens: maxReasoningTokens,\n control_adapter: controlAdapter,\n timeout_ms: timeoutMs,\n max_request_bytes: MAX_MODEL_REQUEST_BYTES,\n max_response_bytes: MAX_MODEL_RESPONSE_BYTES,\n runner: runner ? 'caller-supplied' : 'default',\n runner_command: runner?.command ?? null,\n },\n async analyze(request) {\n const callback = await startTraceToolCallback({\n tools: request.tools,\n maxCalls: request.limits.maxToolCalls,\n ...(request.signal ? { signal: request.signal } : {}),\n })\n let modelProxy: ExternalOptimizerModelProxy | undefined\n const result = await runWithCleanup({\n label: 'DSPy RLM trace-analysis resources',\n run: async () => {\n modelProxy = await startExternalOptimizerModelProxy({\n upstreamBaseUrl: options.baseUrl,\n upstreamApiKey: options.apiKey,\n model: options.model,\n budget: {\n maxCostUsd,\n maxRequests: request.limits.maxIterations + request.limits.maxLlmCalls + 1,\n maxRequestBytes: MAX_MODEL_REQUEST_BYTES,\n maxResponseBytes: MAX_MODEL_RESPONSE_BYTES,\n maxOutputTokensPerRequest: maxOutputTokens,\n maxReasoningTokensPerRequest: maxReasoningTokens,\n requestTimeoutMs: timeoutMs,\n pricing,\n },\n costLedger: request.costLedger,\n channel: 'analyst',\n phase: request.costPhase,\n actor: request.analystId,\n ...(request.costTags ? { tags: request.costTags } : {}),\n ...(request.signal ? { signal: request.signal } : {}),\n })\n request.log?.('trace analyst engine started', {\n engine: 'dspy-rlm',\n model: options.model,\n tools: request.tools.map((tool) => tool.name),\n limits: request.limits,\n })\n const raw = await runExternalOptimizerProcess<unknown>({\n label: 'DSPy RLM trace analysis',\n tempPrefix: 'agent-eval-dspy-rlm-',\n module: BRIDGE_MODULE,\n input: {\n operation: 'analyze',\n question: request.question,\n instructions: request.instructions,\n modelProxy: {\n baseUrl: modelProxy.baseUrl,\n apiKey: modelProxy.apiKey,\n model: options.model,\n maxOutputTokens,\n },\n toolCallback: {\n url: callback.url,\n token: callback.token,\n },\n toolSpecs: request.tools.map(({ name, description, parameters }) => ({\n name,\n description,\n parameters,\n })),\n controlAdapter,\n limits: {\n maxIterations: request.limits.maxIterations,\n maxLlmCalls: request.limits.maxLlmCalls,\n maxOutputChars: request.limits.maxOutputChars,\n },\n },\n ...(runner ? { runner } : {}),\n timeoutMs,\n ...(request.signal ? { signal: request.signal } : {}),\n })\n const parsed = parseBridgeOutput(raw, (index, reason) => {\n request.log?.('finding rejected: bridge row failed schema validation', {\n engine: 'dspy-rlm',\n index,\n reason,\n })\n })\n const successfulCompletions = modelProxy.successfulCompletions()\n const requestAttempts = modelProxy.requestAttempts()\n const rateLimitRetries = modelProxy.rateLimitRetries()\n if (parsed.modelCalls !== successfulCompletions) {\n throw new Error(\n `DSPy RLM reported ${parsed.modelCalls} model calls, but the provider proxy recorded ${successfulCompletions}`,\n )\n }\n return {\n ...parsed,\n toolCalls: callback.calls(),\n runtime: {\n ...parsed.runtime,\n modelRequestAttempts: requestAttempts,\n modelSuccessfulCompletions: successfulCompletions,\n modelRateLimitRetries: rateLimitRetries,\n },\n } satisfies TraceAnalysisEngineResult\n },\n cleanup: async () => {\n const results = await Promise.allSettled([\n ...(modelProxy ? [modelProxy.close()] : []),\n callback.close(),\n ])\n const errors = results.flatMap((entry) =>\n entry.status === 'rejected' ? [entry.reason] : [],\n )\n if (errors.length > 0) {\n throw new AggregateError(errors, 'DSPy RLM resource cleanup failed')\n }\n },\n })\n request.log?.('trace analyst engine completed', {\n engine: 'dspy-rlm',\n model_calls: result.modelCalls,\n model_request_attempts: result.runtime.modelRequestAttempts,\n model_rate_limit_retries: result.runtime.modelRateLimitRetries,\n tool_calls: result.toolCalls,\n findings: result.findings.length,\n })\n return result\n },\n }\n}\n\nfunction parseBridgeOutput(\n value: unknown,\n onRejectedFinding: (index: number, reason: string) => void,\n): Omit<TraceAnalysisEngineResult, 'toolCalls'> {\n if (!isRecord(value)) throw new Error('DSPy RLM bridge output must be an object')\n if (typeof value.answer !== 'string' || !value.answer.trim()) {\n throw new Error('DSPy RLM bridge returned no answer')\n }\n if (!Array.isArray(value.findings)) {\n throw new Error('DSPy RLM bridge findings must be an array')\n }\n // Findings are model output: one malformed row is model noise, not a bridge\n // fault, and the rest of the paid investigation must survive it. Rejected\n // rows are logged per row and counted in runtime.rejectedFindings.\n let rejectedFindings = 0\n const findings: RawAnalystFinding[] = []\n value.findings.forEach((finding, index) => {\n const parsed = RawAnalystFindingSchema.safeParse(finding)\n if (!parsed.success) {\n rejectedFindings += 1\n onRejectedFinding(\n index,\n parsed.error.issues.map((issue) => `${issue.path.join('.')}: ${issue.message}`).join('; '),\n )\n return\n }\n findings.push(parsed.data)\n })\n if (!Array.isArray(value.trajectory)) {\n throw new Error('DSPy RLM bridge trajectory must be an array')\n }\n if (!Number.isSafeInteger(value.modelCalls) || (value.modelCalls as number) <= 0) {\n throw new Error('DSPy RLM bridge modelCalls must be a positive safe integer')\n }\n if (!isRecord(value.runtime)) {\n throw new Error('DSPy RLM bridge runtime must be an object')\n }\n return {\n answer: value.answer,\n findings,\n trajectory: value.trajectory,\n modelCalls: value.modelCalls as number,\n runtime: { ...value.runtime, rejectedFindings },\n }\n}\n\nfunction assertOptions(options: DspyRlmTraceEngineOptions): void {\n for (const [name, value] of [\n ['baseUrl', options.baseUrl],\n ['apiKey', options.apiKey],\n ['model', options.model],\n ] as const) {\n if (typeof value !== 'string' || !value.trim()) {\n throw new TypeError(`DSPy RLM ${name} must be a non-empty string`)\n }\n }\n}\n\nfunction pricingForModel(model: string): CustomTokenPricing {\n const pricing = resolveModelPricing(model)\n if (!pricing) {\n throw new Error(\n `no pricing is configured for '${model}'; provide DspyRlmTraceEngineOptions.pricing`,\n )\n }\n return {\n inputUsdPerMillion: pricing.input * 1_000,\n outputUsdPerMillion: pricing.output * 1_000,\n }\n}\n\nfunction sanitizedRunner(\n runner: ExternalOptimizerRunnerCommand | undefined,\n): ExternalOptimizerRunnerCommand | undefined {\n if (!runner) return undefined\n return {\n ...(runner.command ? { command: runner.command } : {}),\n ...(runner.args ? { args: runner.args } : {}),\n ...(runner.env ? { env: removeCredentialEnvironment(runner.env) } : {}),\n }\n}\n\nfunction isRecord(value: unknown): value is Record<string, unknown> {\n return typeof value === 'object' && value !== null && !Array.isArray(value)\n}\n"],"mappings":";;;;;;AAKA,MAAM,oBAAoB;AAC1B,MAAM,qBAAqB;;AAU3B,eAAsB,uBAAuB,MAId;CAC7B,IAAI,CAAC,OAAO,cAAc,KAAK,QAAQ,KAAK,KAAK,YAAY,GAC3D,MAAM,IAAI,UAAU,8DAA8D;CAEpF,KAAK,QAAQ,eAAe;CAC5B,MAAM,SAAS,IAAI,IAAI,KAAK,MAAM,KAAK,SAAS,CAAC,KAAK,MAAM,IAAI,CAAC,CAAC;CAClE,IAAI,OAAO,SAAS,KAAK,MAAM,QAC7B,MAAM,IAAI,MAAM,mDAAmD;CAGrE,MAAM,QAAQ,YAAY,EAAE,CAAC,CAAC,SAAS,KAAK;CAC5C,IAAI,QAAQ;CACZ,IAAI,YAAY;CAChB,IAAI;CACJ,MAAM,oCAAoB,IAAI,IAAqB;CACnD,MAAM,iCAAiB,IAAI,IAAmB;CAC9C,MAAM,SAAS,cAAc,SAAS,aAAa;EACjD,IAAI,CAAC,WAAW;GACd,eAAe,UAAU,KAAK,EAAE,OAAO,iCAAiC,CAAC;GACzE;EACF;EACA,MAAM,aAAa,IAAI,gBAAgB;EACvC,MAAM,qBAA2B;GAC/B,QAAQ,QAAQ;GAChB,SAAS,QAAQ;EACnB;EACA,kBAAkB,IAAI,UAAU;EAChC,WAAW,OAAO,iBAAiB,SAAS,cAAc,EAAE,MAAM,KAAK,CAAC;EAExE,IAAI;EACJ,UAAU,cAAc,SAAS,UAAU,WAAW,MAAM,CAAC,CAAC,cAAc;GAC1E,WAAW,OAAO,oBAAoB,SAAS,YAAY;GAC3D,kBAAkB,OAAO,UAAU;GACnC,eAAe,OAAO,OAAO;EAC/B,CAAC;EACD,eAAe,IAAI,OAAO;EAC1B,QAAa,YAAY,KAAA,CAAS;CACpC,CAAC;CACD,MAAM,OAAO,MAAM,YAAY,MAAM;CACrC,MAAM,cAA6B;EACjC,iBAAiB,cAAc;EAC/B,OAAO;CACT;CACA,MAAM,gBAAsB;EAC1B,MAAW,CAAC,CAAC,YAAY,KAAA,CAAS;CACpC;CACA,KAAK,QAAQ,iBAAiB,SAAS,SAAS,EAAE,MAAM,KAAK,CAAC;CAC9D,IAAI,KAAK,QAAQ,SAAS,QAAQ;CAElC,OAAO;EACL,KAAK,oBAAoB,KAAK;EAC9B;EACA,aAAa;EACb;CACF;CAEA,eAAe,cACb,SACA,UACA,QACe;EACf,IAAI;GACF,IAAI,QAAQ,WAAW,UAAU,QAAQ,QAAQ,SAAS;IACxD,eAAe,UAAU,KAAK,EAAE,OAAO,YAAY,CAAC;IACpD;GACF;GACA,IAAI,QAAQ,QAAQ,kBAAkB,UAAU,SAAS;IACvD,eAAe,UAAU,KAAK,EAAE,OAAO,eAAe,CAAC;IACvD;GACF;GACA,IAAI,SAAS,KAAK,UAAU;IAC1B,eAAe,UAAU,KAAK,EAAE,OAAO,gCAAgC,CAAC;IACxE;GACF;GACA,MAAM,OAAO,MAAM,SAAS,OAAO;GACnC,IAAI,CAACA,WAAS,IAAI,KAAK,OAAO,KAAK,SAAS,YAAY,EAAE,UAAU,OAAO;IACzE,eAAe,UAAU,KAAK,EAAE,OAAO,6BAA6B,CAAC;IACrE;GACF;GACA,MAAM,OAAO,OAAO,IAAI,KAAK,IAAI;GACjC,IAAI,CAAC,MAAM;IACT,eAAe,UAAU,KAAK,EAAE,OAAO,uBAAuB,KAAK,KAAK,GAAG,CAAC;IAC5E;GACF;GACA,SAAS;GACT,MAAM,SAAS,MAAM,KAAK,QAAQ,KAAK,MAAM,EAAE,OAAO,CAAC;GACvD,MAAM,UAAU,KAAK,UAAU,EAAE,OAAO,CAAC;GACzC,IAAI,OAAO,WAAW,OAAO,IAAI,oBAAoB;IACnD,eAAe,UAAU,KAAK,EAAE,OAAO,gCAAgC,CAAC;IACxE;GACF;GACA,SAAS,UAAU,KAAK;IACtB,gBAAgB;IAChB,kBAAkB,OAAO,OAAO,WAAW,OAAO,CAAC;GACrD,CAAC;GACD,SAAS,IAAI,OAAO;EACtB,SAAS,OAAO;GACd,eAAe,UAAU,OAAO,UAAU,MAAM,KAAK,EACnD,OAAO,iBAAiB,QAAQ,MAAM,UAAU,OAAO,KAAK,EAC9D,CAAC;EACH;CACF;CAEA,eAAe,gBAA+B;EAC5C,KAAK,QAAQ,oBAAoB,SAAS,OAAO;EACjD,YAAY;EACZ,MAAM,gBAAgB,YAAY,MAAM;EACxC,OAAO,uBAAuB;EAC9B,KAAK,MAAM,cAAc,mBAAmB,WAAW,MAAM;EAC7D,MAAM,CAAC,gBAAgB,MAAM,QAAQ,WAAW,CAC9C,eACA,sBAAsB,cAAc,CACtC,CAAC;EACD,IAAI,kBAAkB,SAAS,KAAK,eAAe,SAAS,GAC1D,MAAM,IAAI,MAAM,iDAAiD;EAEnE,IAAI,cAAc,WAAW,YAAY,MAAM,aAAa;CAC9D;AACF;AAEA,eAAe,sBAAsB,gBAAmD;CACtF,OAAO,eAAe,OAAO,GAC3B,MAAM,QAAQ,WAAW,CAAC,GAAG,cAAc,CAAC;AAEhD;AAEA,SAAS,SAAS,SAA4C;CAC5D,OAAO,IAAI,SAAS,SAAS,WAAW;EACtC,IAAI,OAAO;EACX,MAAM,SAAmB,CAAC;EAC1B,QAAQ,GAAG,SAAS,UAAkB;GACpC,QAAQ,MAAM;GACd,IAAI,OAAO,mBAAmB;IAC5B,uBAAO,IAAI,MAAM,8BAA8B,CAAC;IAChD,QAAQ,QAAQ;IAChB;GACF;GACA,OAAO,KAAK,KAAK;EACnB,CAAC;EACD,QAAQ,GAAG,SAAS,MAAM;EAC1B,QAAQ,GAAG,aAAa;GACtB,IAAI;IACF,QAAQ,KAAK,MAAM,OAAO,OAAO,MAAM,CAAC,CAAC,SAAS,MAAM,CAAC,CAAC;GAC5D,SAAS,OAAO;IACd,OAAO,KAAK;GACd;EACF,CAAC;CACH,CAAC;AACH;AAEA,SAAS,eAAe,UAA0B,QAAgB,MAAqB;CACrF,IAAI,SAAS,aAAa,SAAS,eAAe;CAClD,SAAS,UAAU,QAAQ,IAAI;AACjC;AAEA,SAASA,WAAS,OAAkD;CAClE,OAAO,OAAO,UAAU,YAAY,UAAU,QAAQ,CAAC,MAAM,QAAQ,KAAK;AAC5E;;;ACnKA,MAAM,qBAAqB,KAAK;AAOhC,MAAM,8BAA8B;AACpC,MAAM,uBAAuB;AAC7B,MAAM,0BAA0B,KAAK,OAAO;AAC5C,MAAM,2BAA2B,IAAI,OAAO;AAC5C,MAAM,gBAAgB;;AAEtB,MAAM,0BAA0B;;AAmChC,SAAgB,yBAAyB,SAAyD;CAChG,cAAc,OAAO;CACrB,MAAM,kBAAkB,QAAQ,mBAAmB;CACnD,IAAI,CAAC,OAAO,cAAc,eAAe,KAAK,mBAAmB,GAC/D,MAAM,IAAI,UAAU,0DAA0D;CAEhF,MAAM,iBAAiB,QAAQ,kBAAkB;CACjD,MAAM,qBAAqB,QAAQ,sBAAsB,kBAAkB;CAC3E,IAAI,CAAC,OAAO,cAAc,kBAAkB,KAAK,qBAAqB,GACpE,MAAM,IAAI,UAAU,iEAAiE;CAEvF,MAAM,YAAY,QAAQ,aAAa;CACvC,IAAI,CAAC,OAAO,cAAc,SAAS,KAAK,aAAa,GACnD,MAAM,IAAI,UAAU,oDAAoD;CAE1E,MAAM,aAAa,QAAQ,cAAc;CACzC,IAAI,CAAC,OAAO,SAAS,UAAU,KAAK,cAAc,GAChD,MAAM,IAAI,UAAU,iDAAiD;CAEvE,MAAM,UAAU,QAAQ,WAAW,gBAAgB,QAAQ,KAAK;CAEhE,MAAM,SAAS,gBAAgB,QAAQ,MAAM;CAC7C,OAAO;EACL,IAAI;EACJ,aAAa;EACb,OAAO,QAAQ;EACf,SAAS;EACT,iBAAiB;GACf,eAAe;GACf,UAAU,QAAQ;GAClB,OAAO,QAAQ;GACf,kBAAkB;GAClB,SAAS,EAAE,GAAG,QAAQ;GACtB,cAAc;GACd,mBAAmB;GACnB,sBAAsB;GACtB,iBAAiB;GACjB,YAAY;GACZ,mBAAmB;GACnB,oBAAoB;GACpB,QAAQ,SAAS,oBAAoB;GACrC,gBAAgB,QAAQ,WAAW;EACrC;EACA,MAAM,QAAQ,SAAS;GACrB,MAAM,WAAW,MAAM,uBAAuB;IAC5C,OAAO,QAAQ;IACf,UAAU,QAAQ,OAAO;IACzB,GAAI,QAAQ,SAAS,EAAE,QAAQ,QAAQ,OAAO,IAAI,CAAC;GACrD,CAAC;GACD,IAAI;GACJ,MAAM,SAAS,MAAM,eAAe;IAClC,OAAO;IACP,KAAK,YAAY;KACf,aAAa,MAAM,iCAAiC;MAClD,iBAAiB,QAAQ;MACzB,gBAAgB,QAAQ;MACxB,OAAO,QAAQ;MACf,QAAQ;OACN;OACA,aAAa,QAAQ,OAAO,gBAAgB,QAAQ,OAAO,cAAc;OACzE,iBAAiB;OACjB,kBAAkB;OAClB,2BAA2B;OAC3B,8BAA8B;OAC9B,kBAAkB;OAClB;MACF;MACA,YAAY,QAAQ;MACpB,SAAS;MACT,OAAO,QAAQ;MACf,OAAO,QAAQ;MACf,GAAI,QAAQ,WAAW,EAAE,MAAM,QAAQ,SAAS,IAAI,CAAC;MACrD,GAAI,QAAQ,SAAS,EAAE,QAAQ,QAAQ,OAAO,IAAI,CAAC;KACrD,CAAC;KACD,QAAQ,MAAM,gCAAgC;MAC5C,QAAQ;MACR,OAAO,QAAQ;MACf,OAAO,QAAQ,MAAM,KAAK,SAAS,KAAK,IAAI;MAC5C,QAAQ,QAAQ;KAClB,CAAC;KAmCD,MAAM,SAAS,kBAAkB,MAlCf,4BAAqC;MACrD,OAAO;MACP,YAAY;MACZ,QAAQ;MACR,OAAO;OACL,WAAW;OACX,UAAU,QAAQ;OAClB,cAAc,QAAQ;OACtB,YAAY;QACV,SAAS,WAAW;QACpB,QAAQ,WAAW;QACnB,OAAO,QAAQ;QACf;OACF;OACA,cAAc;QACZ,KAAK,SAAS;QACd,OAAO,SAAS;OAClB;OACA,WAAW,QAAQ,MAAM,KAAK,EAAE,MAAM,aAAa,kBAAkB;QACnE;QACA;QACA;OACF,EAAE;OACF;OACA,QAAQ;QACN,eAAe,QAAQ,OAAO;QAC9B,aAAa,QAAQ,OAAO;QAC5B,gBAAgB,QAAQ,OAAO;OACjC;MACF;MACA,GAAI,SAAS,EAAE,OAAO,IAAI,CAAC;MAC3B;MACA,GAAI,QAAQ,SAAS,EAAE,QAAQ,QAAQ,OAAO,IAAI,CAAC;KACrD,CAAC,IACsC,OAAO,WAAW;MACvD,QAAQ,MAAM,yDAAyD;OACrE,QAAQ;OACR;OACA;MACF,CAAC;KACH,CAAC;KACD,MAAM,wBAAwB,WAAW,sBAAsB;KAC/D,MAAM,kBAAkB,WAAW,gBAAgB;KACnD,MAAM,mBAAmB,WAAW,iBAAiB;KACrD,IAAI,OAAO,eAAe,uBACxB,MAAM,IAAI,MACR,qBAAqB,OAAO,WAAW,gDAAgD,uBACzF;KAEF,OAAO;MACL,GAAG;MACH,WAAW,SAAS,MAAM;MAC1B,SAAS;OACP,GAAG,OAAO;OACV,sBAAsB;OACtB,4BAA4B;OAC5B,uBAAuB;MACzB;KACF;IACF;IACA,SAAS,YAAY;KAKnB,MAAM,UAAS,MAJO,QAAQ,WAAW,CACvC,GAAI,aAAa,CAAC,WAAW,MAAM,CAAC,IAAI,CAAC,GACzC,SAAS,MAAM,CACjB,CAAC,EAAA,CACsB,SAAS,UAC9B,MAAM,WAAW,aAAa,CAAC,MAAM,MAAM,IAAI,CAAC,CAClD;KACA,IAAI,OAAO,SAAS,GAClB,MAAM,IAAI,eAAe,QAAQ,kCAAkC;IAEvE;GACF,CAAC;GACD,QAAQ,MAAM,kCAAkC;IAC9C,QAAQ;IACR,aAAa,OAAO;IACpB,wBAAwB,OAAO,QAAQ;IACvC,0BAA0B,OAAO,QAAQ;IACzC,YAAY,OAAO;IACnB,UAAU,OAAO,SAAS;GAC5B,CAAC;GACD,OAAO;EACT;CACF;AACF;AAEA,SAAS,kBACP,OACA,mBAC8C;CAC9C,IAAI,CAAC,SAAS,KAAK,GAAG,MAAM,IAAI,MAAM,0CAA0C;CAChF,IAAI,OAAO,MAAM,WAAW,YAAY,CAAC,MAAM,OAAO,KAAK,GACzD,MAAM,IAAI,MAAM,oCAAoC;CAEtD,IAAI,CAAC,MAAM,QAAQ,MAAM,QAAQ,GAC/B,MAAM,IAAI,MAAM,2CAA2C;CAK7D,IAAI,mBAAmB;CACvB,MAAM,WAAgC,CAAC;CACvC,MAAM,SAAS,SAAS,SAAS,UAAU;EACzC,MAAM,SAAS,wBAAwB,UAAU,OAAO;EACxD,IAAI,CAAC,OAAO,SAAS;GACnB,oBAAoB;GACpB,kBACE,OACA,OAAO,MAAM,OAAO,KAAK,UAAU,GAAG,MAAM,KAAK,KAAK,GAAG,EAAE,IAAI,MAAM,SAAS,CAAC,CAAC,KAAK,IAAI,CAC3F;GACA;EACF;EACA,SAAS,KAAK,OAAO,IAAI;CAC3B,CAAC;CACD,IAAI,CAAC,MAAM,QAAQ,MAAM,UAAU,GACjC,MAAM,IAAI,MAAM,6CAA6C;CAE/D,IAAI,CAAC,OAAO,cAAc,MAAM,UAAU,KAAM,MAAM,cAAyB,GAC7E,MAAM,IAAI,MAAM,4DAA4D;CAE9E,IAAI,CAAC,SAAS,MAAM,OAAO,GACzB,MAAM,IAAI,MAAM,2CAA2C;CAE7D,OAAO;EACL,QAAQ,MAAM;EACd;EACA,YAAY,MAAM;EAClB,YAAY,MAAM;EAClB,SAAS;GAAE,GAAG,MAAM;GAAS;EAAiB;CAChD;AACF;AAEA,SAAS,cAAc,SAA0C;CAC/D,KAAK,MAAM,CAAC,MAAM,UAAU;EAC1B,CAAC,WAAW,QAAQ,OAAO;EAC3B,CAAC,UAAU,QAAQ,MAAM;EACzB,CAAC,SAAS,QAAQ,KAAK;CACzB,GACE,IAAI,OAAO,UAAU,YAAY,CAAC,MAAM,KAAK,GAC3C,MAAM,IAAI,UAAU,YAAY,KAAK,4BAA4B;AAGvE;AAEA,SAAS,gBAAgB,OAAmC;CAC1D,MAAM,UAAU,oBAAoB,KAAK;CACzC,IAAI,CAAC,SACH,MAAM,IAAI,MACR,iCAAiC,MAAM,6CACzC;CAEF,OAAO;EACL,oBAAoB,QAAQ,QAAQ;EACpC,qBAAqB,QAAQ,SAAS;CACxC;AACF;AAEA,SAAS,gBACP,QAC4C;CAC5C,IAAI,CAAC,QAAQ,OAAO,KAAA;CACpB,OAAO;EACL,GAAI,OAAO,UAAU,EAAE,SAAS,OAAO,QAAQ,IAAI,CAAC;EACpD,GAAI,OAAO,OAAO,EAAE,MAAM,OAAO,KAAK,IAAI,CAAC;EAC3C,GAAI,OAAO,MAAM,EAAE,KAAK,4BAA4B,OAAO,GAAG,EAAE,IAAI,CAAC;CACvE;AACF;AAEA,SAAS,SAAS,OAAkD;CAClE,OAAO,OAAO,UAAU,YAAY,UAAU,QAAQ,CAAC,MAAM,QAAQ,KAAK;AAC5E"}
1
+ {"version":3,"file":"dspy-rlm-engine-CJ4sN8Wx.js","names":["isRecord"],"sources":["../src/analyst/trace-tool-callback.ts","../src/analyst/dspy-rlm-engine.ts"],"sourcesContent":["import { randomBytes } from 'node:crypto'\nimport { createServer, type IncomingMessage, type ServerResponse } from 'node:http'\nimport { closeServer, listenLocal, sendJson } from '../campaign/external-optimizer-http'\nimport type { TraceAnalysisToolDescriptor } from '../trace-analyst/tools'\n\nconst MAX_REQUEST_BYTES = 1_000_000\nconst MAX_RESPONSE_BYTES = 4_000_000\n\nexport interface TraceToolCallback {\n url: string\n token: string\n calls: () => number\n close: () => Promise<void>\n}\n\n/** Expose one bounded trace-tool set only on an authenticated loopback socket. */\nexport async function startTraceToolCallback(args: {\n tools: readonly TraceAnalysisToolDescriptor[]\n maxCalls: number\n signal?: AbortSignal\n}): Promise<TraceToolCallback> {\n if (!Number.isSafeInteger(args.maxCalls) || args.maxCalls <= 0) {\n throw new TypeError('trace tool callback maxCalls must be a positive safe integer')\n }\n args.signal?.throwIfAborted()\n const byName = new Map(args.tools.map((tool) => [tool.name, tool]))\n if (byName.size !== args.tools.length) {\n throw new Error('trace tool callback received duplicate tool names')\n }\n\n const token = randomBytes(32).toString('hex')\n let calls = 0\n let accepting = true\n let closePromise: Promise<void> | undefined\n const activeControllers = new Set<AbortController>()\n const activeHandlers = new Set<Promise<void>>()\n const server = createServer((request, response) => {\n if (!accepting) {\n sendJsonIfOpen(response, 503, { error: 'trace tool callback is closing' })\n return\n }\n const controller = new AbortController()\n const abortRequest = (): void => {\n request.destroy()\n response.destroy()\n }\n activeControllers.add(controller)\n controller.signal.addEventListener('abort', abortRequest, { once: true })\n\n let handler!: Promise<void>\n handler = handleRequest(request, response, controller.signal).finally(() => {\n controller.signal.removeEventListener('abort', abortRequest)\n activeControllers.delete(controller)\n activeHandlers.delete(handler)\n })\n activeHandlers.add(handler)\n void handler.catch(() => undefined)\n })\n const port = await listenLocal(server)\n const close = (): Promise<void> => {\n closePromise ??= closeCallback()\n return closePromise\n }\n const onAbort = (): void => {\n void close().catch(() => undefined)\n }\n args.signal?.addEventListener('abort', onAbort, { once: true })\n if (args.signal?.aborted) onAbort()\n\n return {\n url: `http://127.0.0.1:${port}/call`,\n token,\n calls: () => calls,\n close,\n }\n\n async function handleRequest(\n request: IncomingMessage,\n response: ServerResponse,\n signal: AbortSignal,\n ): Promise<void> {\n try {\n if (request.method !== 'POST' || request.url !== '/call') {\n sendJsonIfOpen(response, 404, { error: 'not found' })\n return\n }\n if (request.headers.authorization !== `Bearer ${token}`) {\n sendJsonIfOpen(response, 401, { error: 'unauthorized' })\n return\n }\n if (calls >= args.maxCalls) {\n sendJsonIfOpen(response, 429, { error: 'trace tool call limit reached' })\n return\n }\n const body = await readJson(request)\n if (!isRecord(body) || typeof body.name !== 'string' || !('args' in body)) {\n sendJsonIfOpen(response, 400, { error: 'name and args are required' })\n return\n }\n const tool = byName.get(body.name)\n if (!tool) {\n sendJsonIfOpen(response, 404, { error: `unknown trace tool '${body.name}'` })\n return\n }\n calls += 1\n const result = await tool.handler(body.args, { signal })\n const encoded = JSON.stringify({ result })\n if (Buffer.byteLength(encoded) > MAX_RESPONSE_BYTES) {\n sendJsonIfOpen(response, 413, { error: 'trace tool response too large' })\n return\n }\n response.writeHead(200, {\n 'content-type': 'application/json; charset=utf-8',\n 'content-length': String(Buffer.byteLength(encoded)),\n })\n response.end(encoded)\n } catch (error) {\n sendJsonIfOpen(response, signal.aborted ? 499 : 400, {\n error: error instanceof Error ? error.message : String(error),\n })\n }\n }\n\n async function closeCallback(): Promise<void> {\n args.signal?.removeEventListener('abort', onAbort)\n accepting = false\n const closingServer = closeServer(server)\n server.closeIdleConnections?.()\n for (const controller of activeControllers) controller.abort()\n const [serverResult] = await Promise.allSettled([\n closingServer,\n waitForActiveHandlers(activeHandlers),\n ])\n if (activeControllers.size !== 0 || activeHandlers.size !== 0) {\n throw new Error('trace tool callback closed with active requests')\n }\n if (serverResult?.status === 'rejected') throw serverResult.reason\n }\n}\n\nasync function waitForActiveHandlers(activeHandlers: Set<Promise<void>>): Promise<void> {\n while (activeHandlers.size > 0) {\n await Promise.allSettled([...activeHandlers])\n }\n}\n\nfunction readJson(request: IncomingMessage): Promise<unknown> {\n return new Promise((resolve, reject) => {\n let size = 0\n const chunks: Buffer[] = []\n request.on('data', (chunk: Buffer) => {\n size += chunk.length\n if (size > MAX_REQUEST_BYTES) {\n reject(new Error('trace tool request too large'))\n request.destroy()\n return\n }\n chunks.push(chunk)\n })\n request.on('error', reject)\n request.on('end', () => {\n try {\n resolve(JSON.parse(Buffer.concat(chunks).toString('utf8')))\n } catch (error) {\n reject(error)\n }\n })\n })\n}\n\nfunction sendJsonIfOpen(response: ServerResponse, status: number, body: unknown): void {\n if (response.destroyed || response.writableEnded) return\n sendJson(response, status, body)\n}\n\nfunction isRecord(value: unknown): value is Record<string, unknown> {\n return typeof value === 'object' && value !== null && !Array.isArray(value)\n}\n","import {\n type ExternalOptimizerModelProxy,\n type ExternalOptimizerRunnerCommand,\n removeCredentialEnvironment,\n} from '../campaign/external-optimizer-contracts'\nimport { startExternalOptimizerModelProxy } from '../campaign/external-optimizer-model-proxy'\nimport { runWithCleanup } from '../campaign/external-optimizer-resources'\nimport { runExternalOptimizerProcess } from '../campaign/external-optimizer-subprocess'\nimport type { CustomTokenPricing } from '../cost-ledger'\nimport { resolveModelPricing } from '../metrics'\nimport type { TraceAnalysisEngine, TraceAnalysisEngineResult } from './engine'\nimport { type RawAnalystFinding, RawAnalystFindingSchema } from './finding-signature'\nimport { startTraceToolCallback } from './trace-tool-callback'\n\nconst DEFAULT_TIMEOUT_MS = 10 * 60_000\n// 4096 is below what current coding models emit for a full findings array:\n// glm-5.2 through an OpenAI-compatible gateway returns 8192 and the request is\n// rejected outright (`502 — provider reported 8192 completion tokens,\n// exceeding requested limit 4096`), so the default failed 2/2 smoke cases\n// before any analysis ran. The cap exists to bound spend, and `maxCostUsd`\n// already does that directly, so it starts above what a real completion needs.\nconst DEFAULT_MODEL_OUTPUT_TOKENS = 16_384\nconst DEFAULT_MAX_COST_USD = 1\nconst MAX_MODEL_REQUEST_BYTES = 16 * 1024 * 1024\nconst MAX_MODEL_RESPONSE_BYTES = 4 * 1024 * 1024\nconst BRIDGE_MODULE = 'agent_eval_rpc.dspy_rlm_bridge'\n/** Bumped whenever this engine's execution behavior changes. */\nconst DSPY_RLM_ENGINE_VERSION = '1.0.0'\n\nexport interface DspyRlmTraceEngineOptions {\n baseUrl: string\n apiKey: string\n model: string\n /** Exact provider rates. Required when the model is absent from the pricing table. */\n pricing?: CustomTokenPricing\n /** Maximum provider spend for one investigation. Default: 1 USD. */\n maxCostUsd?: number\n /** Controller response cap. Default: 16384. */\n maxOutputTokens?: number\n /**\n * Thinking tokens one controller turn may bill on top of its completion.\n * A reasoning model bills these beyond `maxOutputTokens`, so the cost\n * reservation must cover them. Default: four times the completion cap.\n */\n maxReasoningTokens?: number\n /**\n * How the controller's reasoning and code fields are obtained.\n *\n * `tolerant` parses marker output strictly first, then recovers the fields\n * deterministically from prose plus a fenced code block — the shape coding\n * models naturally emit — at no extra model cost. `two-step` extracts with a\n * second call per turn. `chat` accepts marker output only. Default:\n * `tolerant`.\n */\n controlAdapter?: 'chat' | 'two-step' | 'tolerant'\n /** Python command used to load agent-eval-rpc[dspy]. Default: python. */\n runner?: ExternalOptimizerRunnerCommand\n /** Whole investigation deadline. Default: 10 minutes. */\n timeoutMs?: number\n}\n\n/** Use the official DSPy RLM as a bounded recursive trace-analysis engine. */\nexport function createDspyRlmTraceEngine(options: DspyRlmTraceEngineOptions): TraceAnalysisEngine {\n assertOptions(options)\n const maxOutputTokens = options.maxOutputTokens ?? DEFAULT_MODEL_OUTPUT_TOKENS\n if (!Number.isSafeInteger(maxOutputTokens) || maxOutputTokens <= 0) {\n throw new TypeError('DSPy RLM maxOutputTokens must be a positive safe integer')\n }\n const controlAdapter = options.controlAdapter ?? 'tolerant'\n const maxReasoningTokens = options.maxReasoningTokens ?? maxOutputTokens * 4\n if (!Number.isSafeInteger(maxReasoningTokens) || maxReasoningTokens < 0) {\n throw new TypeError('DSPy RLM maxReasoningTokens must be a non-negative safe integer')\n }\n const timeoutMs = options.timeoutMs ?? DEFAULT_TIMEOUT_MS\n if (!Number.isSafeInteger(timeoutMs) || timeoutMs <= 0) {\n throw new TypeError('DSPy RLM timeoutMs must be a positive safe integer')\n }\n const maxCostUsd = options.maxCostUsd ?? DEFAULT_MAX_COST_USD\n if (!Number.isFinite(maxCostUsd) || maxCostUsd <= 0) {\n throw new TypeError('DSPy RLM maxCostUsd must be positive and finite')\n }\n const pricing = options.pricing ?? pricingForModel(options.model)\n\n const runner = sanitizedRunner(options.runner)\n return {\n id: 'dspy-rlm',\n description: 'Official DSPy RLM with bounded trace tools and metered model calls.',\n model: options.model,\n version: DSPY_RLM_ENGINE_VERSION,\n executionConfig: {\n bridge_module: BRIDGE_MODULE,\n base_url: options.baseUrl,\n model: options.model,\n api_key_provided: true,\n pricing: { ...pricing },\n max_cost_usd: maxCostUsd,\n max_output_tokens: maxOutputTokens,\n max_reasoning_tokens: maxReasoningTokens,\n control_adapter: controlAdapter,\n timeout_ms: timeoutMs,\n max_request_bytes: MAX_MODEL_REQUEST_BYTES,\n max_response_bytes: MAX_MODEL_RESPONSE_BYTES,\n runner: runner ? 'caller-supplied' : 'default',\n runner_command: runner?.command ?? null,\n },\n async analyze(request) {\n const callback = await startTraceToolCallback({\n tools: request.tools,\n maxCalls: request.limits.maxToolCalls,\n ...(request.signal ? { signal: request.signal } : {}),\n })\n let modelProxy: ExternalOptimizerModelProxy | undefined\n const result = await runWithCleanup({\n label: 'DSPy RLM trace-analysis resources',\n run: async () => {\n modelProxy = await startExternalOptimizerModelProxy({\n upstreamBaseUrl: options.baseUrl,\n upstreamApiKey: options.apiKey,\n model: options.model,\n budget: {\n maxCostUsd,\n maxRequests: request.limits.maxIterations + request.limits.maxLlmCalls + 1,\n maxRequestBytes: MAX_MODEL_REQUEST_BYTES,\n maxResponseBytes: MAX_MODEL_RESPONSE_BYTES,\n maxOutputTokensPerRequest: maxOutputTokens,\n maxReasoningTokensPerRequest: maxReasoningTokens,\n requestTimeoutMs: timeoutMs,\n pricing,\n },\n costLedger: request.costLedger,\n channel: 'analyst',\n phase: request.costPhase,\n actor: request.analystId,\n ...(request.costTags ? { tags: request.costTags } : {}),\n ...(request.signal ? { signal: request.signal } : {}),\n })\n request.log?.('trace analyst engine started', {\n engine: 'dspy-rlm',\n model: options.model,\n tools: request.tools.map((tool) => tool.name),\n limits: request.limits,\n })\n const raw = await runExternalOptimizerProcess<unknown>({\n label: 'DSPy RLM trace analysis',\n tempPrefix: 'agent-eval-dspy-rlm-',\n module: BRIDGE_MODULE,\n input: {\n operation: 'analyze',\n question: request.question,\n instructions: request.instructions,\n modelProxy: {\n baseUrl: modelProxy.baseUrl,\n apiKey: modelProxy.apiKey,\n model: options.model,\n maxOutputTokens,\n },\n toolCallback: {\n url: callback.url,\n token: callback.token,\n },\n toolSpecs: request.tools.map(({ name, description, parameters }) => ({\n name,\n description,\n parameters,\n })),\n controlAdapter,\n limits: {\n maxIterations: request.limits.maxIterations,\n maxLlmCalls: request.limits.maxLlmCalls,\n maxOutputChars: request.limits.maxOutputChars,\n },\n },\n ...(runner ? { runner } : {}),\n timeoutMs,\n ...(request.signal ? { signal: request.signal } : {}),\n })\n const parsed = parseBridgeOutput(raw, (index, reason) => {\n request.log?.('finding rejected: bridge row failed schema validation', {\n engine: 'dspy-rlm',\n index,\n reason,\n })\n })\n const successfulCompletions = modelProxy.successfulCompletions()\n const requestAttempts = modelProxy.requestAttempts()\n const rateLimitRetries = modelProxy.rateLimitRetries()\n if (parsed.modelCalls !== successfulCompletions) {\n throw new Error(\n `DSPy RLM reported ${parsed.modelCalls} model calls, but the provider proxy recorded ${successfulCompletions}`,\n )\n }\n return {\n ...parsed,\n toolCalls: callback.calls(),\n runtime: {\n ...parsed.runtime,\n modelRequestAttempts: requestAttempts,\n modelSuccessfulCompletions: successfulCompletions,\n modelRateLimitRetries: rateLimitRetries,\n },\n } satisfies TraceAnalysisEngineResult\n },\n cleanup: async () => {\n const results = await Promise.allSettled([\n ...(modelProxy ? [modelProxy.close()] : []),\n callback.close(),\n ])\n const errors = results.flatMap((entry) =>\n entry.status === 'rejected' ? [entry.reason] : [],\n )\n if (errors.length > 0) {\n throw new AggregateError(errors, 'DSPy RLM resource cleanup failed')\n }\n },\n })\n request.log?.('trace analyst engine completed', {\n engine: 'dspy-rlm',\n model_calls: result.modelCalls,\n model_request_attempts: result.runtime.modelRequestAttempts,\n model_rate_limit_retries: result.runtime.modelRateLimitRetries,\n tool_calls: result.toolCalls,\n findings: result.findings.length,\n })\n return result\n },\n }\n}\n\nfunction parseBridgeOutput(\n value: unknown,\n onRejectedFinding: (index: number, reason: string) => void,\n): Omit<TraceAnalysisEngineResult, 'toolCalls'> {\n if (!isRecord(value)) throw new Error('DSPy RLM bridge output must be an object')\n if (typeof value.answer !== 'string' || !value.answer.trim()) {\n throw new Error('DSPy RLM bridge returned no answer')\n }\n if (!Array.isArray(value.findings)) {\n throw new Error('DSPy RLM bridge findings must be an array')\n }\n // Findings are model output: one malformed row is model noise, not a bridge\n // fault, and the rest of the paid investigation must survive it. Rejected\n // rows are logged per row and counted in runtime.rejectedFindings.\n let rejectedFindings = 0\n const findings: RawAnalystFinding[] = []\n value.findings.forEach((finding, index) => {\n const parsed = RawAnalystFindingSchema.safeParse(finding)\n if (!parsed.success) {\n rejectedFindings += 1\n onRejectedFinding(\n index,\n parsed.error.issues.map((issue) => `${issue.path.join('.')}: ${issue.message}`).join('; '),\n )\n return\n }\n findings.push(parsed.data)\n })\n if (!Array.isArray(value.trajectory)) {\n throw new Error('DSPy RLM bridge trajectory must be an array')\n }\n if (!Number.isSafeInteger(value.modelCalls) || (value.modelCalls as number) <= 0) {\n throw new Error('DSPy RLM bridge modelCalls must be a positive safe integer')\n }\n if (!isRecord(value.runtime)) {\n throw new Error('DSPy RLM bridge runtime must be an object')\n }\n return {\n answer: value.answer,\n findings,\n trajectory: value.trajectory,\n modelCalls: value.modelCalls as number,\n runtime: { ...value.runtime, rejectedFindings },\n }\n}\n\nfunction assertOptions(options: DspyRlmTraceEngineOptions): void {\n for (const [name, value] of [\n ['baseUrl', options.baseUrl],\n ['apiKey', options.apiKey],\n ['model', options.model],\n ] as const) {\n if (typeof value !== 'string' || !value.trim()) {\n throw new TypeError(`DSPy RLM ${name} must be a non-empty string`)\n }\n }\n}\n\nfunction pricingForModel(model: string): CustomTokenPricing {\n const pricing = resolveModelPricing(model)\n if (!pricing) {\n throw new Error(\n `no pricing is configured for '${model}'; provide DspyRlmTraceEngineOptions.pricing`,\n )\n }\n return {\n inputUsdPerMillion: pricing.input * 1_000,\n outputUsdPerMillion: pricing.output * 1_000,\n }\n}\n\nfunction sanitizedRunner(\n runner: ExternalOptimizerRunnerCommand | undefined,\n): ExternalOptimizerRunnerCommand | undefined {\n if (!runner) return undefined\n return {\n ...(runner.command ? { command: runner.command } : {}),\n ...(runner.args ? { args: runner.args } : {}),\n ...(runner.env ? { env: removeCredentialEnvironment(runner.env) } : {}),\n }\n}\n\nfunction isRecord(value: unknown): value is Record<string, unknown> {\n return typeof value === 'object' && value !== null && !Array.isArray(value)\n}\n"],"mappings":";;;;;;AAKA,MAAM,oBAAoB;AAC1B,MAAM,qBAAqB;;AAU3B,eAAsB,uBAAuB,MAId;CAC7B,IAAI,CAAC,OAAO,cAAc,KAAK,QAAQ,KAAK,KAAK,YAAY,GAC3D,MAAM,IAAI,UAAU,8DAA8D;CAEpF,KAAK,QAAQ,eAAe;CAC5B,MAAM,SAAS,IAAI,IAAI,KAAK,MAAM,KAAK,SAAS,CAAC,KAAK,MAAM,IAAI,CAAC,CAAC;CAClE,IAAI,OAAO,SAAS,KAAK,MAAM,QAC7B,MAAM,IAAI,MAAM,mDAAmD;CAGrE,MAAM,QAAQ,YAAY,EAAE,CAAC,CAAC,SAAS,KAAK;CAC5C,IAAI,QAAQ;CACZ,IAAI,YAAY;CAChB,IAAI;CACJ,MAAM,oCAAoB,IAAI,IAAqB;CACnD,MAAM,iCAAiB,IAAI,IAAmB;CAC9C,MAAM,SAAS,cAAc,SAAS,aAAa;EACjD,IAAI,CAAC,WAAW;GACd,eAAe,UAAU,KAAK,EAAE,OAAO,iCAAiC,CAAC;GACzE;EACF;EACA,MAAM,aAAa,IAAI,gBAAgB;EACvC,MAAM,qBAA2B;GAC/B,QAAQ,QAAQ;GAChB,SAAS,QAAQ;EACnB;EACA,kBAAkB,IAAI,UAAU;EAChC,WAAW,OAAO,iBAAiB,SAAS,cAAc,EAAE,MAAM,KAAK,CAAC;EAExE,IAAI;EACJ,UAAU,cAAc,SAAS,UAAU,WAAW,MAAM,CAAC,CAAC,cAAc;GAC1E,WAAW,OAAO,oBAAoB,SAAS,YAAY;GAC3D,kBAAkB,OAAO,UAAU;GACnC,eAAe,OAAO,OAAO;EAC/B,CAAC;EACD,eAAe,IAAI,OAAO;EAC1B,QAAa,YAAY,KAAA,CAAS;CACpC,CAAC;CACD,MAAM,OAAO,MAAM,YAAY,MAAM;CACrC,MAAM,cAA6B;EACjC,iBAAiB,cAAc;EAC/B,OAAO;CACT;CACA,MAAM,gBAAsB;EAC1B,MAAW,CAAC,CAAC,YAAY,KAAA,CAAS;CACpC;CACA,KAAK,QAAQ,iBAAiB,SAAS,SAAS,EAAE,MAAM,KAAK,CAAC;CAC9D,IAAI,KAAK,QAAQ,SAAS,QAAQ;CAElC,OAAO;EACL,KAAK,oBAAoB,KAAK;EAC9B;EACA,aAAa;EACb;CACF;CAEA,eAAe,cACb,SACA,UACA,QACe;EACf,IAAI;GACF,IAAI,QAAQ,WAAW,UAAU,QAAQ,QAAQ,SAAS;IACxD,eAAe,UAAU,KAAK,EAAE,OAAO,YAAY,CAAC;IACpD;GACF;GACA,IAAI,QAAQ,QAAQ,kBAAkB,UAAU,SAAS;IACvD,eAAe,UAAU,KAAK,EAAE,OAAO,eAAe,CAAC;IACvD;GACF;GACA,IAAI,SAAS,KAAK,UAAU;IAC1B,eAAe,UAAU,KAAK,EAAE,OAAO,gCAAgC,CAAC;IACxE;GACF;GACA,MAAM,OAAO,MAAM,SAAS,OAAO;GACnC,IAAI,CAACA,WAAS,IAAI,KAAK,OAAO,KAAK,SAAS,YAAY,EAAE,UAAU,OAAO;IACzE,eAAe,UAAU,KAAK,EAAE,OAAO,6BAA6B,CAAC;IACrE;GACF;GACA,MAAM,OAAO,OAAO,IAAI,KAAK,IAAI;GACjC,IAAI,CAAC,MAAM;IACT,eAAe,UAAU,KAAK,EAAE,OAAO,uBAAuB,KAAK,KAAK,GAAG,CAAC;IAC5E;GACF;GACA,SAAS;GACT,MAAM,SAAS,MAAM,KAAK,QAAQ,KAAK,MAAM,EAAE,OAAO,CAAC;GACvD,MAAM,UAAU,KAAK,UAAU,EAAE,OAAO,CAAC;GACzC,IAAI,OAAO,WAAW,OAAO,IAAI,oBAAoB;IACnD,eAAe,UAAU,KAAK,EAAE,OAAO,gCAAgC,CAAC;IACxE;GACF;GACA,SAAS,UAAU,KAAK;IACtB,gBAAgB;IAChB,kBAAkB,OAAO,OAAO,WAAW,OAAO,CAAC;GACrD,CAAC;GACD,SAAS,IAAI,OAAO;EACtB,SAAS,OAAO;GACd,eAAe,UAAU,OAAO,UAAU,MAAM,KAAK,EACnD,OAAO,iBAAiB,QAAQ,MAAM,UAAU,OAAO,KAAK,EAC9D,CAAC;EACH;CACF;CAEA,eAAe,gBAA+B;EAC5C,KAAK,QAAQ,oBAAoB,SAAS,OAAO;EACjD,YAAY;EACZ,MAAM,gBAAgB,YAAY,MAAM;EACxC,OAAO,uBAAuB;EAC9B,KAAK,MAAM,cAAc,mBAAmB,WAAW,MAAM;EAC7D,MAAM,CAAC,gBAAgB,MAAM,QAAQ,WAAW,CAC9C,eACA,sBAAsB,cAAc,CACtC,CAAC;EACD,IAAI,kBAAkB,SAAS,KAAK,eAAe,SAAS,GAC1D,MAAM,IAAI,MAAM,iDAAiD;EAEnE,IAAI,cAAc,WAAW,YAAY,MAAM,aAAa;CAC9D;AACF;AAEA,eAAe,sBAAsB,gBAAmD;CACtF,OAAO,eAAe,OAAO,GAC3B,MAAM,QAAQ,WAAW,CAAC,GAAG,cAAc,CAAC;AAEhD;AAEA,SAAS,SAAS,SAA4C;CAC5D,OAAO,IAAI,SAAS,SAAS,WAAW;EACtC,IAAI,OAAO;EACX,MAAM,SAAmB,CAAC;EAC1B,QAAQ,GAAG,SAAS,UAAkB;GACpC,QAAQ,MAAM;GACd,IAAI,OAAO,mBAAmB;IAC5B,uBAAO,IAAI,MAAM,8BAA8B,CAAC;IAChD,QAAQ,QAAQ;IAChB;GACF;GACA,OAAO,KAAK,KAAK;EACnB,CAAC;EACD,QAAQ,GAAG,SAAS,MAAM;EAC1B,QAAQ,GAAG,aAAa;GACtB,IAAI;IACF,QAAQ,KAAK,MAAM,OAAO,OAAO,MAAM,CAAC,CAAC,SAAS,MAAM,CAAC,CAAC;GAC5D,SAAS,OAAO;IACd,OAAO,KAAK;GACd;EACF,CAAC;CACH,CAAC;AACH;AAEA,SAAS,eAAe,UAA0B,QAAgB,MAAqB;CACrF,IAAI,SAAS,aAAa,SAAS,eAAe;CAClD,SAAS,UAAU,QAAQ,IAAI;AACjC;AAEA,SAASA,WAAS,OAAkD;CAClE,OAAO,OAAO,UAAU,YAAY,UAAU,QAAQ,CAAC,MAAM,QAAQ,KAAK;AAC5E;;;ACnKA,MAAM,qBAAqB,KAAK;AAOhC,MAAM,8BAA8B;AACpC,MAAM,uBAAuB;AAC7B,MAAM,0BAA0B,KAAK,OAAO;AAC5C,MAAM,2BAA2B,IAAI,OAAO;AAC5C,MAAM,gBAAgB;;AAEtB,MAAM,0BAA0B;;AAmChC,SAAgB,yBAAyB,SAAyD;CAChG,cAAc,OAAO;CACrB,MAAM,kBAAkB,QAAQ,mBAAmB;CACnD,IAAI,CAAC,OAAO,cAAc,eAAe,KAAK,mBAAmB,GAC/D,MAAM,IAAI,UAAU,0DAA0D;CAEhF,MAAM,iBAAiB,QAAQ,kBAAkB;CACjD,MAAM,qBAAqB,QAAQ,sBAAsB,kBAAkB;CAC3E,IAAI,CAAC,OAAO,cAAc,kBAAkB,KAAK,qBAAqB,GACpE,MAAM,IAAI,UAAU,iEAAiE;CAEvF,MAAM,YAAY,QAAQ,aAAa;CACvC,IAAI,CAAC,OAAO,cAAc,SAAS,KAAK,aAAa,GACnD,MAAM,IAAI,UAAU,oDAAoD;CAE1E,MAAM,aAAa,QAAQ,cAAc;CACzC,IAAI,CAAC,OAAO,SAAS,UAAU,KAAK,cAAc,GAChD,MAAM,IAAI,UAAU,iDAAiD;CAEvE,MAAM,UAAU,QAAQ,WAAW,gBAAgB,QAAQ,KAAK;CAEhE,MAAM,SAAS,gBAAgB,QAAQ,MAAM;CAC7C,OAAO;EACL,IAAI;EACJ,aAAa;EACb,OAAO,QAAQ;EACf,SAAS;EACT,iBAAiB;GACf,eAAe;GACf,UAAU,QAAQ;GAClB,OAAO,QAAQ;GACf,kBAAkB;GAClB,SAAS,EAAE,GAAG,QAAQ;GACtB,cAAc;GACd,mBAAmB;GACnB,sBAAsB;GACtB,iBAAiB;GACjB,YAAY;GACZ,mBAAmB;GACnB,oBAAoB;GACpB,QAAQ,SAAS,oBAAoB;GACrC,gBAAgB,QAAQ,WAAW;EACrC;EACA,MAAM,QAAQ,SAAS;GACrB,MAAM,WAAW,MAAM,uBAAuB;IAC5C,OAAO,QAAQ;IACf,UAAU,QAAQ,OAAO;IACzB,GAAI,QAAQ,SAAS,EAAE,QAAQ,QAAQ,OAAO,IAAI,CAAC;GACrD,CAAC;GACD,IAAI;GACJ,MAAM,SAAS,MAAM,eAAe;IAClC,OAAO;IACP,KAAK,YAAY;KACf,aAAa,MAAM,iCAAiC;MAClD,iBAAiB,QAAQ;MACzB,gBAAgB,QAAQ;MACxB,OAAO,QAAQ;MACf,QAAQ;OACN;OACA,aAAa,QAAQ,OAAO,gBAAgB,QAAQ,OAAO,cAAc;OACzE,iBAAiB;OACjB,kBAAkB;OAClB,2BAA2B;OAC3B,8BAA8B;OAC9B,kBAAkB;OAClB;MACF;MACA,YAAY,QAAQ;MACpB,SAAS;MACT,OAAO,QAAQ;MACf,OAAO,QAAQ;MACf,GAAI,QAAQ,WAAW,EAAE,MAAM,QAAQ,SAAS,IAAI,CAAC;MACrD,GAAI,QAAQ,SAAS,EAAE,QAAQ,QAAQ,OAAO,IAAI,CAAC;KACrD,CAAC;KACD,QAAQ,MAAM,gCAAgC;MAC5C,QAAQ;MACR,OAAO,QAAQ;MACf,OAAO,QAAQ,MAAM,KAAK,SAAS,KAAK,IAAI;MAC5C,QAAQ,QAAQ;KAClB,CAAC;KAmCD,MAAM,SAAS,kBAAkB,MAlCf,4BAAqC;MACrD,OAAO;MACP,YAAY;MACZ,QAAQ;MACR,OAAO;OACL,WAAW;OACX,UAAU,QAAQ;OAClB,cAAc,QAAQ;OACtB,YAAY;QACV,SAAS,WAAW;QACpB,QAAQ,WAAW;QACnB,OAAO,QAAQ;QACf;OACF;OACA,cAAc;QACZ,KAAK,SAAS;QACd,OAAO,SAAS;OAClB;OACA,WAAW,QAAQ,MAAM,KAAK,EAAE,MAAM,aAAa,kBAAkB;QACnE;QACA;QACA;OACF,EAAE;OACF;OACA,QAAQ;QACN,eAAe,QAAQ,OAAO;QAC9B,aAAa,QAAQ,OAAO;QAC5B,gBAAgB,QAAQ,OAAO;OACjC;MACF;MACA,GAAI,SAAS,EAAE,OAAO,IAAI,CAAC;MAC3B;MACA,GAAI,QAAQ,SAAS,EAAE,QAAQ,QAAQ,OAAO,IAAI,CAAC;KACrD,CAAC,IACsC,OAAO,WAAW;MACvD,QAAQ,MAAM,yDAAyD;OACrE,QAAQ;OACR;OACA;MACF,CAAC;KACH,CAAC;KACD,MAAM,wBAAwB,WAAW,sBAAsB;KAC/D,MAAM,kBAAkB,WAAW,gBAAgB;KACnD,MAAM,mBAAmB,WAAW,iBAAiB;KACrD,IAAI,OAAO,eAAe,uBACxB,MAAM,IAAI,MACR,qBAAqB,OAAO,WAAW,gDAAgD,uBACzF;KAEF,OAAO;MACL,GAAG;MACH,WAAW,SAAS,MAAM;MAC1B,SAAS;OACP,GAAG,OAAO;OACV,sBAAsB;OACtB,4BAA4B;OAC5B,uBAAuB;MACzB;KACF;IACF;IACA,SAAS,YAAY;KAKnB,MAAM,UAAS,MAJO,QAAQ,WAAW,CACvC,GAAI,aAAa,CAAC,WAAW,MAAM,CAAC,IAAI,CAAC,GACzC,SAAS,MAAM,CACjB,CAAC,EAAA,CACsB,SAAS,UAC9B,MAAM,WAAW,aAAa,CAAC,MAAM,MAAM,IAAI,CAAC,CAClD;KACA,IAAI,OAAO,SAAS,GAClB,MAAM,IAAI,eAAe,QAAQ,kCAAkC;IAEvE;GACF,CAAC;GACD,QAAQ,MAAM,kCAAkC;IAC9C,QAAQ;IACR,aAAa,OAAO;IACpB,wBAAwB,OAAO,QAAQ;IACvC,0BAA0B,OAAO,QAAQ;IACzC,YAAY,OAAO;IACnB,UAAU,OAAO,SAAS;GAC5B,CAAC;GACD,OAAO;EACT;CACF;AACF;AAEA,SAAS,kBACP,OACA,mBAC8C;CAC9C,IAAI,CAAC,SAAS,KAAK,GAAG,MAAM,IAAI,MAAM,0CAA0C;CAChF,IAAI,OAAO,MAAM,WAAW,YAAY,CAAC,MAAM,OAAO,KAAK,GACzD,MAAM,IAAI,MAAM,oCAAoC;CAEtD,IAAI,CAAC,MAAM,QAAQ,MAAM,QAAQ,GAC/B,MAAM,IAAI,MAAM,2CAA2C;CAK7D,IAAI,mBAAmB;CACvB,MAAM,WAAgC,CAAC;CACvC,MAAM,SAAS,SAAS,SAAS,UAAU;EACzC,MAAM,SAAS,wBAAwB,UAAU,OAAO;EACxD,IAAI,CAAC,OAAO,SAAS;GACnB,oBAAoB;GACpB,kBACE,OACA,OAAO,MAAM,OAAO,KAAK,UAAU,GAAG,MAAM,KAAK,KAAK,GAAG,EAAE,IAAI,MAAM,SAAS,CAAC,CAAC,KAAK,IAAI,CAC3F;GACA;EACF;EACA,SAAS,KAAK,OAAO,IAAI;CAC3B,CAAC;CACD,IAAI,CAAC,MAAM,QAAQ,MAAM,UAAU,GACjC,MAAM,IAAI,MAAM,6CAA6C;CAE/D,IAAI,CAAC,OAAO,cAAc,MAAM,UAAU,KAAM,MAAM,cAAyB,GAC7E,MAAM,IAAI,MAAM,4DAA4D;CAE9E,IAAI,CAAC,SAAS,MAAM,OAAO,GACzB,MAAM,IAAI,MAAM,2CAA2C;CAE7D,OAAO;EACL,QAAQ,MAAM;EACd;EACA,YAAY,MAAM;EAClB,YAAY,MAAM;EAClB,SAAS;GAAE,GAAG,MAAM;GAAS;EAAiB;CAChD;AACF;AAEA,SAAS,cAAc,SAA0C;CAC/D,KAAK,MAAM,CAAC,MAAM,UAAU;EAC1B,CAAC,WAAW,QAAQ,OAAO;EAC3B,CAAC,UAAU,QAAQ,MAAM;EACzB,CAAC,SAAS,QAAQ,KAAK;CACzB,GACE,IAAI,OAAO,UAAU,YAAY,CAAC,MAAM,KAAK,GAC3C,MAAM,IAAI,UAAU,YAAY,KAAK,4BAA4B;AAGvE;AAEA,SAAS,gBAAgB,OAAmC;CAC1D,MAAM,UAAU,oBAAoB,KAAK;CACzC,IAAI,CAAC,SACH,MAAM,IAAI,MACR,iCAAiC,MAAM,6CACzC;CAEF,OAAO;EACL,oBAAoB,QAAQ,QAAQ;EACpC,qBAAqB,QAAQ,SAAS;CACxC;AACF;AAEA,SAAS,gBACP,QAC4C;CAC5C,IAAI,CAAC,QAAQ,OAAO,KAAA;CACpB,OAAO;EACL,GAAI,OAAO,UAAU,EAAE,SAAS,OAAO,QAAQ,IAAI,CAAC;EACpD,GAAI,OAAO,OAAO,EAAE,MAAM,OAAO,KAAK,IAAI,CAAC;EAC3C,GAAI,OAAO,MAAM,EAAE,KAAK,4BAA4B,OAAO,GAAG,EAAE,IAAI,CAAC;CACvE;AACF;AAEA,SAAS,SAAS,OAAkD;CAClE,OAAO,OAAO,UAAU,YAAY,UAAU,QAAQ,CAAC,MAAM,QAAQ,KAAK;AAC5E"}
@@ -1,6 +1,6 @@
1
- import { j as MutableSurface } from "../types-Dk7PB7vh.js";
1
+ import { j as MutableSurface } from "../types-KEqL1pZc.js";
2
2
  import { o as InsightReport } from "../insight-report-Bh_8ksel.js";
3
- import { _ as UnixNanoTimestamp, a as hostedTenantFromEnv, c as EvalRunGenerationSnapshot, d as HostedIngestHeaders, f as HostedWireVersion, g as TraceSpanEvent, h as IngestTracesRequest, i as hostedClientFromEnv, l as EvalRunStatus, m as IngestResponse, n as HostedTenant, o as EvalRunCellScore, p as IngestEvalRunsRequest, r as createHostedClient, s as EvalRunEvent, t as HostedClient, u as HOSTED_WIRE_VERSION } from "../client-BoqGxEqx.js";
3
+ import { _ as UnixNanoTimestamp, a as hostedTenantFromEnv, c as EvalRunGenerationSnapshot, d as HostedIngestHeaders, f as HostedWireVersion, g as TraceSpanEvent, h as IngestTracesRequest, i as hostedClientFromEnv, l as EvalRunStatus, m as IngestResponse, n as HostedTenant, o as EvalRunCellScore, p as IngestEvalRunsRequest, r as createHostedClient, s as EvalRunEvent, t as HostedClient, u as HOSTED_WIRE_VERSION } from "../client-Cgl6KasJ.js";
4
4
  import { z } from "zod";
5
5
  //#region src/hosted/schemas.d.ts
6
6
  declare const UnixNanoTimestampSchema: z.ZodType<UnixNanoTimestamp>;
@@ -1,7 +1,7 @@
1
1
  import { s as RunSplitTag } from "./run-record-Tb3TTtUn.js";
2
- import { R as Scenario, a as CampaignResult, d as DispatchContext } from "./types-Dk7PB7vh.js";
3
- import { sn as CampaignStorage } from "./skillopt-optimization-method-D1dqGzzH.js";
4
- import "./index-DCP4I2Qx.js";
2
+ import { R as Scenario, a as CampaignResult, d as DispatchContext } from "./types-KEqL1pZc.js";
3
+ import { sn as CampaignStorage } from "./skillopt-optimization-method-CwRvi-rL.js";
4
+ import "./index-DY83M1WL.js";
5
5
  //#region src/benchmarks/types.d.ts
6
6
  type BenchmarkTaskKind = 'retrieval' | 'rag-answer' | 'hallucination' | 'kb-improvement' | 'routing' | 'custom';
7
7
  type BenchmarkFamily = 'beir' | 'mteb-retrieval' | 'msmarco' | 'trec-dl' | 'miracl' | 'lotte' | 'bright' | 'crag' | 'hotpotqa' | 'kilt' | 'ragtruth' | 'faithbench' | 'first-party' | 'custom';
@@ -332,4 +332,4 @@ declare namespace index_d_exports {
332
332
  }
333
333
  //#endregion
334
334
  export { BenchmarkMetricCalibrationOptions as A, BenchmarkSource as B, BenchmarkRunOptions as C, runBenchmarkAdapter as D, renderBenchmarkReportMarkdown as E, BenchmarkDatasetItem as F, deterministicSplit as H, BenchmarkEvaluation as I, BenchmarkFamily as L, calibrateBenchmarkMetric as M, BENCHMARK_SPLIT_SEED as N, summarizeBenchmarkCampaign as O, BenchmarkAdapter as P, BenchmarkResponder as R, BenchmarkReport as S, BenchmarkSliceSummary as T, BenchmarkTaskKind as V, parseJsonlRows as _, StandardRetrievalDocument as a, retrievalMetricsAtCutoff as b, StandardRetrievalQrel as c, buildStandardRetrievalItems as d, createRetrievalIdBenchmarkAdapter as f, parseBeirQueriesJsonl as g, parseBeirCorpusJsonl as h, StandardRetrievalArtifact as i, BenchmarkMetricCalibrationResult as j, index_d_exports$1 as k, StandardRetrievalQuery as l, normalizeRetrievedDocumentIds as m, BuildStandardRetrievalItemsOptions as n, StandardRetrievalEvaluationOptions as o, evaluateStandardRetrieval as p, RetrievalIdAdapterOptions as r, StandardRetrievalPayload as s, index_d_exports as t, StandardRetrievalResult as u, parseQrels as v, BenchmarkRunResult as w, BenchmarkDistribution as x, parseTsvRows as y, BenchmarkScenario as z };
335
- //# sourceMappingURL=index-CKI1CXTL.d.ts.map
335
+ //# sourceMappingURL=index-CaOjqCli.d.ts.map
@@ -1 +1 @@
1
- {"version":3,"file":"index-CKI1CXTL.d.ts","names":[],"sources":["../src/benchmarks/types.ts","../src/benchmarks/calibration.ts","../src/benchmarks/routing/dataset.ts","../src/benchmarks/routing/index.ts","../src/benchmarks/runner.ts","../src/benchmarks/standard-formats.ts","../src/benchmarks/index.ts"],"mappings":";;;;;KASY;KAQA;UAgBK,qBAAqB;;;EAGpC;;EAEA,SAAS;;EAET,QAAQ;;EAER,SAAS;;EAET,WAAW;;EAEX;;EAEA,SAAS;EACT,WAAW;;UAGI;;;;EAIf;;EAEA;;EAEA,aAAa;;;EAGb,KAAK;EACL;;UAGe;EACf;EACA;EACA;EACA;EACA;;;UAQe,iBAAiB,kBAAkB,oBAAoB;;EAEtE;EACA,SAAS;EACT,WAAW;EACX;EACA,SAAS;EACT;;;;;EAKA,YAAY,OAAO,cAAc,QAAQ,qBAAqB;;EAE9D,SAAS,MAAM,qBAAqB,WAAW,UAAU,YAAY,QAAQ;;;;EAI7E,YAAY,iBAAiB;;UAGd,kBAAkB,4BAA4B;EAC7D;EACA;EACA,QAAQ;EACR,UAAU;EACV,UAAU;EACV,MAAM,qBAAqB;;KAGjB,mBAAmB,oBAAoB,uBAAuB;EACxE,UAAU,kBAAkB;EAC5B,MAAM,qBAAqB;EAC3B,SAAS;MACL,QAAQ,aAAa;;;cAoBd;;;;;;;;;iBAUG,mBACd,gBACA,gBACC;;;UCjJc,kCAAkC,oBAAoB;EACrE,SAAS,iBAAiB,qBAAqB,WAAW,UAAU;EACpE,MAAM,qBAAqB;EAC3B,cAAc;EACd,gBAAgB;EAChB;EACA;EACA;;UAGe;EACf;EACA,MAAM;EACN,QAAQ;EACR;EACA;EACA;EACA;;iBAGoB,yBAAyB,oBAAoB,oBACjE,SAAS,kCAAkC,UAAU,aACpD,QAAQ;;;;;;;;;;;;;;;;;;;;;;;;UCFM;EACf;EACA;EACA;;EAEA;;EAEA;;EAEA;;cAGW,iBAAiB;;;;KChBlB,iBAAiB;KACjB,qBAAqB,qBAAqB;cAEhD,0BAA0B,iBAAiB,oBAAoB;WAC1D;WACA;WACA;WACA;WACA;EAEH,YAAY,OAAO,cAAc,QAAQ;EAMzC,SAAS,MAAM,oBAAoB,mBAAmB,QAAQ;EAqBpE,YAAY,iBAAiB;;;;;;;;iBAef,mBAAmB;cAOtB,cAAW,OAjDG,gBAAc,QAAQ;cAkDpC,WAAQ,MA5CE,oBAAkB,qBAAqB,QAAQ;cA6CzD,cAAW,mBAxBO;;;UClCd,oBAAoB,oBAAoB;EACvD,SAAS,iBAAiB,qBAAqB,WAAW,UAAU;EACpE,SAAS,mBAAmB,UAAU;EACtC,kBAAkB;EAClB;EACA;EACA;EACA;EACA;EACA;EACA;EACA;EACA;EACA,UAAU;EACV,YAAY;;UAGG;EACf;EACA,QAAQ;EACR,UAAU;EACV,SAAS;EACT;EACA;EACA;EACA;EACA;EACA;EACA;EACA;EACA;EACA;EACA;EACA,QAAQ,eAAe;EACvB,MAAM,eAAe;EACrB,YAAY,eAAe;EAC3B,OAAO;EACP,SAAS;EACT,WAAW;;UAGI;EACf;EACA;EACA;EACA,OAAO;EACP,SAAS;EACT,WAAW;;UAGI;EACf;EACA;EACA;EACA;EACA;EACA;;UAGe,mBAAmB,oBAAoB;EACtD,WAAW,MAAM,kBAAkB;EACnC,UAAU,eAAe,WAAW,kBAAkB;EACtD,QAAQ;EACR;EACA;;iBAGoB,oBAAoB,oBAAoB,oBAC5D,SAAS,oBAAoB,UAAU,aACtC,QAAQ,mBAAmB,UAAU;iBAmGxB,2BAA2B,UAAU,WAAW;EAC9D,SAAS,iBAAiB,qBAAqB,WAAW,UAAU;EACpE,WAAW,MAAM,kBAAkB;EACnC,UAAU,eAAe,WAAW,kBAAkB;IACpD;iBA6CY,8BAA8B,QAAQ;;;UCpOrC;EACf;EACA;EACA;EACA,WAAW;;UAGI;EACf;EACA;EACA,WAAW;;UAGI;EACf;EACA;EACA;;UAGe;EACf;EACA;EACA;EACA,gBAAgB;EAChB,SAAS,eAAe;EACxB,WAAW;;UAGI;EACf;EACA,QAAQ;EACR,kBAAkB;EAClB,gBAAgB;EAChB,kBAAkB;EAClB;EACA,SAAS;EACT;EACA,WAAW,oBAAoB;;UAGhB,kCAAkC;EACjD,oBAAoB;EACpB;EACA;EACA;EACA;;UAGe;EACf;EACA;EACA;EACA;;KAGU,kEAGC;EAEP;EACA;EACA,mBAAmB;;UAGR;EACf,oBAAoB;EACpB;EACA;EACA;EACA;;iBAGc,eAAe,aAAa,eAAe;iBAc3C,aAAa;iBASb,WAAW,eAAe;iBAgB1B,qBAAqB,eAAe;iBAgBpC,sBAAsB,eAAe;iBAerC,4BACd,SAAS,qCACR,MAAM,qBAAqB;iBAuCd,kCACd,SAAS,4BACR,iBACD,qBAAqB,2BACrB,0BACA;iBAqBc,0BACd,SAAS,0BACT,UAAU,2BACV,UAAS;;;;;;;;;;iBAsCK,8BACd,UAAU,2BACV,oBAAmB;iBA0BL,yBAAyB;EACvC;EACA,gBAAgB;EAChB;IACE"}
1
+ {"version":3,"file":"index-CaOjqCli.d.ts","names":[],"sources":["../src/benchmarks/types.ts","../src/benchmarks/calibration.ts","../src/benchmarks/routing/dataset.ts","../src/benchmarks/routing/index.ts","../src/benchmarks/runner.ts","../src/benchmarks/standard-formats.ts","../src/benchmarks/index.ts"],"mappings":";;;;;KASY;KAQA;UAgBK,qBAAqB;;;EAGpC;;EAEA,SAAS;;EAET,QAAQ;;EAER,SAAS;;EAET,WAAW;;EAEX;;EAEA,SAAS;EACT,WAAW;;UAGI;;;;EAIf;;EAEA;;EAEA,aAAa;;;EAGb,KAAK;EACL;;UAGe;EACf;EACA;EACA;EACA;EACA;;;UAQe,iBAAiB,kBAAkB,oBAAoB;;EAEtE;EACA,SAAS;EACT,WAAW;EACX;EACA,SAAS;EACT;;;;;EAKA,YAAY,OAAO,cAAc,QAAQ,qBAAqB;;EAE9D,SAAS,MAAM,qBAAqB,WAAW,UAAU,YAAY,QAAQ;;;;EAI7E,YAAY,iBAAiB;;UAGd,kBAAkB,4BAA4B;EAC7D;EACA;EACA,QAAQ;EACR,UAAU;EACV,UAAU;EACV,MAAM,qBAAqB;;KAGjB,mBAAmB,oBAAoB,uBAAuB;EACxE,UAAU,kBAAkB;EAC5B,MAAM,qBAAqB;EAC3B,SAAS;MACL,QAAQ,aAAa;;;cAoBd;;;;;;;;;iBAUG,mBACd,gBACA,gBACC;;;UCjJc,kCAAkC,oBAAoB;EACrE,SAAS,iBAAiB,qBAAqB,WAAW,UAAU;EACpE,MAAM,qBAAqB;EAC3B,cAAc;EACd,gBAAgB;EAChB;EACA;EACA;;UAGe;EACf;EACA,MAAM;EACN,QAAQ;EACR;EACA;EACA;EACA;;iBAGoB,yBAAyB,oBAAoB,oBACjE,SAAS,kCAAkC,UAAU,aACpD,QAAQ;;;;;;;;;;;;;;;;;;;;;;;;UCFM;EACf;EACA;EACA;;EAEA;;EAEA;;EAEA;;cAGW,iBAAiB;;;;KChBlB,iBAAiB;KACjB,qBAAqB,qBAAqB;cAEhD,0BAA0B,iBAAiB,oBAAoB;WAC1D;WACA;WACA;WACA;WACA;EAEH,YAAY,OAAO,cAAc,QAAQ;EAMzC,SAAS,MAAM,oBAAoB,mBAAmB,QAAQ;EAqBpE,YAAY,iBAAiB;;;;;;;;iBAef,mBAAmB;cAOtB,cAAW,OAjDG,gBAAc,QAAQ;cAkDpC,WAAQ,MA5CE,oBAAkB,qBAAqB,QAAQ;cA6CzD,cAAW,mBAxBO;;;UClCd,oBAAoB,oBAAoB;EACvD,SAAS,iBAAiB,qBAAqB,WAAW,UAAU;EACpE,SAAS,mBAAmB,UAAU;EACtC,kBAAkB;EAClB;EACA;EACA;EACA;EACA;EACA;EACA;EACA;EACA;EACA,UAAU;EACV,YAAY;;UAGG;EACf;EACA,QAAQ;EACR,UAAU;EACV,SAAS;EACT;EACA;EACA;EACA;EACA;EACA;EACA;EACA;EACA;EACA;EACA;EACA,QAAQ,eAAe;EACvB,MAAM,eAAe;EACrB,YAAY,eAAe;EAC3B,OAAO;EACP,SAAS;EACT,WAAW;;UAGI;EACf;EACA;EACA;EACA,OAAO;EACP,SAAS;EACT,WAAW;;UAGI;EACf;EACA;EACA;EACA;EACA;EACA;;UAGe,mBAAmB,oBAAoB;EACtD,WAAW,MAAM,kBAAkB;EACnC,UAAU,eAAe,WAAW,kBAAkB;EACtD,QAAQ;EACR;EACA;;iBAGoB,oBAAoB,oBAAoB,oBAC5D,SAAS,oBAAoB,UAAU,aACtC,QAAQ,mBAAmB,UAAU;iBAmGxB,2BAA2B,UAAU,WAAW;EAC9D,SAAS,iBAAiB,qBAAqB,WAAW,UAAU;EACpE,WAAW,MAAM,kBAAkB;EACnC,UAAU,eAAe,WAAW,kBAAkB;IACpD;iBA6CY,8BAA8B,QAAQ;;;UCpOrC;EACf;EACA;EACA;EACA,WAAW;;UAGI;EACf;EACA;EACA,WAAW;;UAGI;EACf;EACA;EACA;;UAGe;EACf;EACA;EACA;EACA,gBAAgB;EAChB,SAAS,eAAe;EACxB,WAAW;;UAGI;EACf;EACA,QAAQ;EACR,kBAAkB;EAClB,gBAAgB;EAChB,kBAAkB;EAClB;EACA,SAAS;EACT;EACA,WAAW,oBAAoB;;UAGhB,kCAAkC;EACjD,oBAAoB;EACpB;EACA;EACA;EACA;;UAGe;EACf;EACA;EACA;EACA;;KAGU,kEAGC;EAEP;EACA;EACA,mBAAmB;;UAGR;EACf,oBAAoB;EACpB;EACA;EACA;EACA;;iBAGc,eAAe,aAAa,eAAe;iBAc3C,aAAa;iBASb,WAAW,eAAe;iBAgB1B,qBAAqB,eAAe;iBAgBpC,sBAAsB,eAAe;iBAerC,4BACd,SAAS,qCACR,MAAM,qBAAqB;iBAuCd,kCACd,SAAS,4BACR,iBACD,qBAAqB,2BACrB,0BACA;iBAqBc,0BACd,SAAS,0BACT,UAAU,2BACV,UAAS;;;;;;;;;;iBAsCK,8BACd,UAAU,2BACV,oBAAmB;iBA0BL,yBAAyB;EACvC;EACA,gBAAgB;EAChB;IACE"}
@@ -1,13 +1,13 @@
1
1
  import { l as ValidationError, t as AgentEvalError } from "./errors-DkfjIDvD.js";
2
- import { T as RunPaidCallInput, m as CostReceipt } from "./cost-ledger-FuQvHxPm.js";
2
+ import { T as RunPaidCallInput, m as CostReceipt, p as CostProvenance } from "./cost-ledger-FuQvHxPm.js";
3
3
  import { a as RunRecord, s as RunSplitTag } from "./run-record-Tb3TTtUn.js";
4
4
  import { f as AnalystUsageReceipt, i as AnalystFinding, l as AnalystRunResult, x as TraceAnalysisStore } from "./types-D3jh6F98.js";
5
5
  import { n as CompletionVerdict, o as ProducedState, r as CorrectnessChecker, t as CompletionRequirement } from "./completion-verifier-D15NHYSk.js";
6
6
  import { _ as AnalystIssueExpectation, a as AnalystBenchmarkLabelState, t as AnalystBenchmarkCase } from "./benchmark-DxaZfy0w.js";
7
- import { A as LabeledScenarioWrite, D as LabeledScenarioSampleArgs, E as LabeledScenarioRecord, O as LabeledScenarioSource, R as Scenario, S as JudgeConfig, T as LabelTrust, a as CampaignResult, b as GenerationRecord, d as DispatchContext, j as MutableSurface, k as LabeledScenarioStore, l as CodeSurface, p as Gate, u as ComponentSurface, w as JudgeScore } from "./types-Dk7PB7vh.js";
7
+ import { A as LabeledScenarioWrite, D as LabeledScenarioSampleArgs, E as LabeledScenarioRecord, O as LabeledScenarioSource, R as Scenario, S as JudgeConfig, T as LabelTrust, a as CampaignResult, b as GenerationRecord, d as DispatchContext, j as MutableSurface, k as LabeledScenarioStore, l as CodeSurface, p as Gate, u as ComponentSurface, w as JudgeScore } from "./types-KEqL1pZc.js";
8
8
  import { E as RiskDifferenceResult, _ as McNemarResult, d as EProcessState, v as PairedBootstrapOptions, y as PairedBootstrapResult } from "./statistics-B4u_CiFd.js";
9
9
  import { A as PairedMcNemarEvidence, D as PairedDecisionMethod, j as PairedPromotionDecision, k as PairedDecisionStatistic } from "./summary-report-o3eJ3gxG.js";
10
- import { rn as PlanCampaignRunOptions, sn as CampaignStorage, tn as CampaignRunPlan } from "./skillopt-optimization-method-D1dqGzzH.js";
10
+ import { rn as PlanCampaignRunOptions, sn as CampaignStorage, tn as CampaignRunPlan } from "./skillopt-optimization-method-CwRvi-rL.js";
11
11
  import { D as LedgerHash, m as LedgerTrustedHeadRemoval, p as LedgerTrustedHead } from "./index-DEb46kc6.js";
12
12
  import { AgentProfile, AgentProfile as AgentProfile$1, HarnessType, HarnessType as HarnessType$1 } from "@tangle-network/agent-interface";
13
13
  //#region src/campaign/analyst-surface.d.ts
@@ -1342,7 +1342,9 @@ interface ProfileSummary {
1342
1342
  records: number;
1343
1343
  /** Mean across scored records, or null when the profile has no task labels. */
1344
1344
  meanComposite: number | null;
1345
- totalCostUsd: number;
1345
+ /** Total cost, or null when any call's cost was not captured. */
1346
+ totalCostUsd: number | null;
1347
+ costProvenance: CostProvenance;
1346
1348
  /** Per-profile integrity verdict — surfaces a single profile that ran stub
1347
1349
  * even when the matrix as a whole looks real. */
1348
1350
  integrity: BackendIntegrityReport;
@@ -2138,4 +2140,4 @@ declare function verifyCodeSurface(surface: CodeSurface, worktreeDir?: string):
2138
2140
  declare function resolveWorktreePath(surface: CodeSurface, worktreeDir?: string): string;
2139
2141
  //#endregion
2140
2142
  export { SearchPlannedEvent as $, EvalFixtureScenario as $n, pairRunRecords as $r, HarnessType$1 as $t, SearchAccountingAudit as A, SequentialDecideFn as An, CrossSurfaceInteractionTask as Ar, UserStory as At, SearchCostAccounting as B, SignedManifest as Bn, CrossSurfaceTaskRow as Br, RunProfileMatrixOptions as Bt, surfaceHash as C, HeldoutSignificance as Cn, CrossSurfaceEligibility as Cr, tangleTracesRoot as Ct, FileSearchLedger as D, dimensionRegressions as Dn, CrossSurfaceInteractionEffect as Dr, ScoreboardRenderOptions as Dt, acquireSingleRunLock as E, detectScale as En, CrossSurfaceInteractionAwareSelection as Er, PlaybackStep as Et, SearchCandidateRegisteredEvent as F, SequentialPairedGateOptions as Fn, CrossSurfacePairwiseEntry as Fr, scoreboardSummary as Ft, SearchLedgerEvent as G, signManifest as Gn, PairArmsResult as Gr, BackendIntegrityReport as Gt, SearchLedger as H, canonicalize as Hn, MatchedPair as Hr, ScenarioRollup as Ht, SearchCandidateSlot as I, sequentialDecide as In, CrossSurfaceRankedSingle as Ir, userStoryScoreboard as It, SearchLedgerTrustedHeadMode as J, neutralizationGate as Jn, PairedArmsComparison as Jr, summarizeAgentReceiptIntegrity as Jt, SearchLedgerHash as K, verifyManifest as Kn, PairRunRecordsResult as Kr, assertRealAgentReceipts as Kt, SearchCandidateSlotClosedEvent as L, sequentialPairedGate as Ln, CrossSurfaceRelativeCost as Lr, ProfileDispatchFn as Lt, SearchAttemptAccounting as M, SequentialDecision as Mn, CrossSurfacePairCompatibility as Mr, makePlaybackDispatch as Mt, SearchCandidateDecidedEvent as N, SequentialObservation as Nn, CrossSurfacePairEvidence as Nr, renderScoreboardMarkdown as Nt, OpenSearchLedgerOptions as O, heldoutSignificance as On, CrossSurfaceInteractionPath as Or, ScoreboardRow as Ot, SearchCandidateLineage as P, SequentialPairedGate as Pn, CrossSurfacePairIncompatibilityReason as Pr, scoreUserStory as Pt, SearchPlan as Q, EvalFixtureRunPlan as Qn, pairArms as Qr, HARNESS_NATIVE_MODEL as Qt, SearchCandidateSurface as R, HypothesisManifest as Rn, CrossSurfaceSelectionPolicy as Rr, ProfileMatrixError as Rt, surfaceContentHash as S, DimensionRegression as Sn, CrossSurfaceDistribution as Sr, resolveRunDir as St, SingleRunLockOptions as T, PairedHoldout as Tn, CrossSurfaceIneligibilityReason as Tr, PlaybackDriver as Tt, SearchLedgerAppendResult as U, evaluateHypothesis as Un, MatchedRunRecordPair as Ur, runProfileMatrix as Ut, SearchFailureReason as V, SignedManifestAlgo as Vn, ComparePairedArmsOptions as Vr, RunProfileMatrixResult as Vt, SearchLedgerEntry as W, hashJson as Wn, PairArmsOptions as Wr, BackendIntegrityError as Wt, SearchOperationKind as X, EvalFixtureFile as Xn, PairedMetricDelta as Xr, AgentProfile$1 as Xt, SearchModelIdentity as Y, EvalFixture as Yn, PairedCorrectness as Yr, summarizeBackendIntegrity as Yt, SearchOperationRecordedEvent as Z, EvalFixtureLoadOptions as Zn, comparePairedArms as Zr, CODING_HARNESSES as Zt, assertCodeSurfaceIdentity as _, RolloutCall as _n, CrossSurfaceCandidateOutcome as _r, compareRankKeys as _t, WorktreeAdapterError as a, harnessAxisOf as an, loadEvalFixtureScenarios as ar, SearchSurfaceKind as at, componentSurfaceIdentityMaterial as b, classifyUngroundedLiterals as bn, CrossSurfaceComponentEvidence as br, scoreDiscrimination as bt, verifyCodeSurface as c, RuntimeEventLike as cn, AnalyzeCrossSurfaceInteractionsInput as cr, SearchTokenAccounting as ct, PhoenixEvaluationResultLike as d, neutralizeText as dn, CrossSurfaceAttemptCompleteness as dr, SearchLedgerConflictError as dt, BuildTraceAnalystSurfaceDispatchOptions as ei, ProfileAxisSpec as en, EvalFixtureValidationMode as er, SearchPlannedOperation as et, PhoenixEvaluatorLike as f, FsLabeledScenarioStore as fn, CrossSurfaceBestSingleSelection as fr, SearchLedgerError as ft, isTransientTransportFailure as g, RolloutArgumentDiffOptions as gn, CrossSurfaceCandidateEvidence as gr, campaignMeanComposite as gt, TransientFailureOptions as h, RolloutArgumentDiff as hn, CrossSurfaceCandidateComparison as hr, campaignBreakdown as ht, WorktreeAdapter as i, traceAnalystQualityJudge as ii, expandProfileAxes as in, loadEvalFixture as ir, SearchSurfaceEvidence as it, SearchArtifactRef as j, SequentialDecideOptions as jn, CrossSurfaceNaiveStackSelection as jr, UserStoryVerdict as jt, SEARCH_LEDGER_SCHEMA as k, pairHoldout as kn, CrossSurfaceInteractionReport as kr, ScoreboardSummary as kt, AutoevalsScoreLike as l, ToolCallEventLike as ln, CrossSurfaceAdditionDecision as lr, openSearchLedger as lt, phoenixEvaluatorJudge as m, LabeledScenarioStoreError as mn, CrossSurfaceCandidate as mr, CampaignBreakdown as mt, GitWorktreeAdapterOptions as n, TraceAnalystScenario as ni, agentProfileId as nn, PlanEvalFixtureRunOptions as nr, SearchSourceRef as nt, gitWorktreeAdapter as o, ArtifactEventLike as on, planEvalFixtureRun as or, SearchTaskAttemptedEvent as ot, autoevalsScorerJudge as p, FsLabeledScenarioStoreOptions as pn, CrossSurfaceBootstrapPolicy as pr, SearchLedgerIntegrityError as pt, SearchLedgerReplay as q, NeutralizationGateOptions as qn, PairedArmRow as qr, assertRealBackend as qt, Worktree as r, buildTraceAnalystSurfaceDispatch as ri, agentProfileModelId as rn, discoverEvalFixtures as rr, SearchSurfaceEffect as rt, resolveWorktreePath as s, ProposalEventLike as sn, analyzeCrossSurfaceInteractions as sr, SearchTaskOutcome as st, CodeSurfaceVerification as t, TraceAnalystArtifact as ti, agentProfileHash as tn, LoadEvalFixtureScenariosOptions as tr, SearchPlannedTask as tt, AutoevalsScorerLike as u, extractProducedState as un, CrossSurfaceAdditionRejectionReason as ur, validateSearchLedgerEvent as ut, assertComponentSurface as v, ScoredRollout as vn, CrossSurfaceCandidateSummary as vr, DiscriminationScore as vt, SingleRunLock as w, HeldoutSignificanceOptions as wn, CrossSurfaceEvidenceBreakdown as wr, PlaybackContext as wt, renderSurfaceDiff as x, rolloutArgumentDiff as xn, CrossSurfaceCompositionStep as xr, selectDiscriminative as xt, codeSurfaceIdentityMaterial as y, UngroundedLiteralReport as yn, CrossSurfaceComponent as yr, ScenarioSignal as yt, SearchCompletedEvent as z, HypothesisResult as zn, CrossSurfaceSelections as zr, ProfileSummary as zt };
2141
- //# sourceMappingURL=index-DCP4I2Qx.d.ts.map
2143
+ //# sourceMappingURL=index-DY83M1WL.d.ts.map
@@ -1 +1 @@
1
- {"version":3,"file":"index-DCP4I2Qx.d.ts","names":[],"sources":["../src/campaign/analyst-surface.ts","../src/paired-arms.ts","../src/campaign/cross-surface-types.ts","../src/campaign/cross-surface-interaction.ts","../src/campaign/fixtures.ts","../src/campaign/gates/neutralization-gate.ts","../src/pre-registration.ts","../src/campaign/gates/sequential.ts","../src/campaign/gates/statistical-heldout.ts","../src/campaign/grounded-reflection.ts","../src/campaign/labeled-store/fs-adapter.ts","../src/campaign/neutralize.ts","../src/produced-state.ts","../src/agent-profile.ts","../src/integrity/backend-integrity.ts","../src/campaign/presets/run-profile-matrix.ts","../src/campaign/presets/playback.ts","../src/campaign/run-dir.ts","../src/campaign/scenario-selection.ts","../src/campaign/score-utils.ts","../src/campaign/search-ledger-errors.ts","../src/campaign/search-ledger.ts","../src/campaign/single-run-lock.ts","../src/campaign/surface-identity.ts","../src/campaign/transient-failure.ts","../src/campaign/upstream-evaluators.ts","../src/campaign/worktree/index.ts"],"mappings":";;;;;;;;;;;;;UAWiB,6BAA6B;EAC5C;EACA,YAAY;EACZ,YAAY;EACZ,yBAAyB;EACzB,kBAAkB;;UAGH;EACf,mBAAmB;EACnB,QAAQ;EACR,MAAM;;UAGS;EACf,QAAQ;IACN;IACA,YAAY;IACZ;IACA,QAAQ;MACN,QAAQ;;iBAGE,iCACd,SAAS,2CAET,SAAS,gBACT,UAAU,sBACV,SAAS,oBACN,QAAQ;iBAgBG,4BAA4B,YAC1C,sBACA;;;;;UClBe;;;EAGf;;;;;EAKA;;EAEA;;EAEA;;EAEA,UAAU;;UAGK;;EAEf;;EAEA;;;UAIe;EACf;;;;EAIA;EACA,UAAU;EACV,WAAW;;UAGI;;EAEf,OAAO;;;EAGP,kBAAkB;;EAElB,mBAAmB;;;;;;;;;;;;;;;;;;;;iBAqBL,SAAS,eAAe,gBAAgB,MAAM,kBAAkB;;UA8G/D;;EAEf;;EAEA;;EAEA,SAAS;;EAET,gBAAgB;;;UAID;EACf;;EAEA;;EAEA;;EAEA;;EAEA;;;;EAIA,aAAa;;EAEb;IAAY;IAAW;;;UAGR,iCAAiC;;;;EAIhD;;EAEA,YAAY;;UAGG;EACf;EACA;EACA;;;EAGA,aAAa;EACb,cAAc;;;;;;;;;;;;;;;iBAgBA,kBACd,eAAe,gBACf,MAAM,2BACL;UAmEc;EACf;EACA;EACA,UAAU;EACV,WAAW;;UAGI;EACf,OAAO;EACP,kBAAkB;EAClB,mBAAmB;;;;;;;;;iBAeL,eACd,uBAAuB,aACvB,wBAAwB,cACvB;;;;KCjXS;;UAGK;EACf;EACA;;EAEA;;;UAIe;EACf;EACA;EACA;EACA;;;UAIe;EACf;;EAEA;;EAEA;;;;;;;UAQe;EACf;EACA;;EAEA;EACA,cAAc;EACd;EACA;;;;;;EAMA,MAAM;EACN,mBAAmB;;EAEnB;;UAGe;EACf;EACA;EACA;;;UAIe;EACf;EACA;EACA;EACA;;EAEA,kCAAkC;;EAElC;;UAGe,qCACf,aAAa,sBAAsB;EAEnC,qBAAqB;EACrB,qBAAqB;EACrB,eAAe;EACf;;EAEA;;EAEA;;EAEA;;EAEA;EACA,WAAW;EACX,WAAW;;UAGI;EACf;EACA;EACA;EACA;EACA;EACA;;UAGe;EACf;EACA;EACA;EACA;;UAGe;EACf,aAAa;EACb;EACA;EACA;EACA;;KAGU;UAWK;EACf;EACA,SAAS;;UAGM;EACf;EACA;EACA;EACA;EACA;EACA;EACA;EACA;;UAGe;EACf,WAAW;EACX,SAAS;EACT,OAAO;EACP,OAAO,eAAe;EACtB,QAAQ;EACR,QAAQ;;EAER,sBAAsB;;EAEtB,aAAa;;UAGE;EACf;EACA;EACA;;EAEA;;UAGe;EACf;EACA;EACA;EACA;EACA;EACA,QAAQ;EACR,cAAc,eAAe;;UAGd;EACf;EACA;EACA;EACA;EACA;;UAGe;EACf;;EAEA;EACA;;UAGe;EACf,SAAS;EACT;EACA;EACA;EACA;EACA,eAAe;EACf,gBAAgB;;KAGN;UAYK;EACf;EACA,SAAS;EACT;;UAGe;EACf;EACA;EACA;EACA;EACA;EACA;EACA;EACA,4BAA4B,iCAAiC;EAC7D,wBAAwB,eAAe;EACvC,QAAQ;EACR,QAAQ;EACR,aAAa;EACb,eAAe;;UAGA;EACf;EACA;EACA;;UAGe;EACf;EACA;EACA,SAAS;;UAGM;;EAEf;EACA;;KAGU;UAaK;EACf;EACA;EACA;EACA;EACA;EACA,uBAAuB;EACvB;EACA;EACA,SAAS;;UAGM;EACf;EACA;EACA,YAAY;EACZ;;;UAIe;EACf;EACA;EACA;EACA;EACA,OAAO;;UAGQ;;EAEf;;EAEA;EACA;;EAEA;EACA;;EAEA,gBAAgB;;EAEhB,OAAO;;UAGQ;EACf,YAAY;EACZ,YAAY;EACZ,kBAAkB;;UAGH,8BACf,aAAa,sBAAsB;EAEnC;EACA;EACA;EACA;;EAEA,MAAM;EACN,iBAAiB;EACjB,iBAAiB;EACjB,YAAY;EACZ,UAAU;EACV,YAAY;;;;;;;;;iBCnRE,gCAAgC,aAAa,qBAC3D,OAAO,qCAAqC,QAC3C,8BAA8B;;;KC7CrB;UAEK;EACf;EACA;EACA;;UAGe;EACf;EACA;EACA;EACA;EACA;EACA;EACA,OAAO;EACP;;UAGe,4BAA4B;EAC3C;EACA;EACA;EACA;EACA;EACA;EACA;EACA;;UAGe;;EAEf,aAAa;;EAEb;;UAGe,wCAAwC;EACvD;;UAGe,0BAA0B,6BACjC,KACN,uBAAuB,qBAAqB;EAG9C;EACA,aAAa;EACb;EACA;EACA,UAAU;;KAGA,qBAAqB;EAC/B,UAAU,MAAM,KAAK;;;iBAIP,qBAAqB;;;;;iBA4BrB,gBACd,kBACA,cACA,UAAS,yBACR;;iBA4Ca,yBACd,kBACA,UAAS,kCACR;;;;;iBAuBa,mBAAmB,qBACjC,SAAS,0BAA0B,aAClC;;;UChJc,0BAA0B,kBAAkB,WAAW;EACtE,WAAW;;;;;EAKX;;;;;;iBAuBc,mBAAmB,WAAW,kBAAkB,UAC9D,SAAS,0BAA0B,aAClC,KAAK,WAAW;;;;;;;;;;;;;UC9CF;EACf;;EAEA;;EAEA;;EAEA;;EAEA;;EAEA;;EAEA;;EAEA;;EAEA;;EAEA;EACA;;;;;;;;;;KAWU;UAEK,uBAAuB;;EAEtC;;;;;;;;;EASA,OAAO;;UAGQ;EACf,UAAU;EACV;EACA;EACA;;;EAGA;;EAEA,kBAAkB;EAGlB;;;;;;;;;;;iBAYc,aAAa;;;;;;;;;;;;;;;;;;;;;;iBA8BP,SAAS,GAAG,KAAK,IAAI;;;;;;;;;;iBAkBrB,aAAa,GAAG,qBAAqB,QAAQ;;;;;;;iBAW7C,eAAe,GAAG,iBAAiB;;;;;iBAWnC,mBACpB,UAAU,gBACV;EAAY;EAAW;EAAgB;IACtC,QAAQ;;;KCpHC;UAEK;EACf,UAAU;;EAEV;;EAEA;;;EAGA;;UAGe;;;EAGf;;;;EAIA;;;EAGA;;EAEA;;;;EAIA;;;EAGA;;;EAGA,kBAAkB;;EAElB;;UAGe,qBAAqB,qBAAqB,kBAAkB,WAAW,kBAC9E,KAAK,WAAW;;;;;;;;EAQxB,QAAQ,gBAAgB;;EAExB,SAAS;IAAkB,UAAU;;;;;;;;;;;;;;;;;iBA0MvB,qBAAqB,qBAAqB,kBAAkB,WAAW,UACrF,SAAS,8BACR,qBAAqB,WAAW;UAsFlB;;EAEf;;EAEA;;EAEA;;EAEA;;UAGe;GACd;IAAQ,SAAS;;IAAyB;IAAe;;;EAE1D,SAAS;;;;;;;;;;;;;;;;;;;;;;;;;iBA0BK,iBAAiB,UAAS,0BAA+B;;;UCtXxD;;EAEf;;EAEA;;EAEA;;;;;;;;;;;;;iBAcc,YACd,WAAW,YAAY,eAAe,cACtC,UAAU,YAAY,eAAe,cACrC,aAAa,aACb,SAAS,GAAG,oCACX;UAsDc;EACf,QAAQ;;;;;;;;;;;;EAYR,WAAW;;;;EAIX,iBAAiB;;;;;;;EAOjB,UAAU;;EAEV,mBAAmB;;EAEnB,SAAS;;;;EAIT;;EAEA;;EAEA;;EAEA,gBAAgB;;EAEhB;;;;EAIA;;EAEA;;UAGe;EACf;EACA;EACA;EACA;;EAEA;EACA;;;;;;;;;;;;;;;;;;;;;;;;;;;;iBA6Bc,oBACd,QAAQ,eACR,OAAM,6BACL;UAkEc;EACf;;;EAGA,WAAW;;;;EAIX;;EAEA;IAAM;IAAa;;;EAEnB,mBAAmB;;EAEnB,SAAS;;EAET;;;;;EAKA;EACA;EACA;;;;;iBAMc,YAAY;;;;;;;;;;;;;;;;;;;iBAsBZ,qBACd,WAAW,YAAY,eAAe,cACtC,UAAU,YAAY,eAAe,cACrC,aAAa,aACb,8BACA;EACE;EACA;EACA;EACA;;;EAGA;IAED;;;;;;;;;;;;;;;;;;;;;;;;;;;;;UC7Tc;WACN;WACA,MAAM,SAAS;;;UAIT;;WAEN;;WAEA;WACA,gBAAgB;;UAGV;;WAEN;;WAEA;;UAGM;;WAEN;;WAEA,eAAe;;WAEf,eAAe;;;;;;;;iBASV,oBACd,mBAAmB,iBACnB,OAAM,6BACL;UAsCc;;WAEN;;WAEA;;;;;;;;;iBAUK,2BACd,cACA,MAAM,KAAK,0DACV;;;UCjFc;;EAEf;;;EAGA;;EAEA;;;cAIW,kCAAkC;WAE3B;EADlB,YACkB,cAChB;;;;;;cAiBS,kCAAkC;mBAIhB;mBAHZ;mBACA;EAEjB,YAA6B,SAAS;EAKhC,QAAQ,OAAO,uBAAuB;EAYtC,OAAO,MAAM,4BAA4B,QAAQ;EA8CjD,QAAQ;IACZ;IACA;IACA,UAAU;IACV,SAAS,OAAO;;UAkCV;UAiCA;UAmBA;UAkBA;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;iBCtNM,eAAe;;;;UCZd;EACf;EACA;;;;;;;UAQe;EACf;EACA;EACA;EACA;EACA;EACA;;;UAIe;EACf;EACA;EACA;EACA;EAIA;;;;;;;;KASU,mBACR,oBACA,oBACA;EACE;;;;;;;;;;;;iBAmBU,qBAAqB,iBAAiB,qBAAqB;;;;;;;;;;cCjE9D,2BAA2B;UAOvB;;;EAGf,MAAM;;EAEN,qBAAqB;;;EAGrB;;;;;;EAMA;;;;;;;cAQW;;;;;;;;;;;;;;;;;;iBAmBG,kBAAkB,MAAM,kBAAkB;;;;;;;;iBAuD1C,cACd,SAAS,KAAK;EACX,SAAS;EAAa;;;;;;;;;iBAiBX,eAAe,SAAS;;;;;iBASxB,oBAAoB,SAAS;;;;;;;;;;;iBA0C7B,iBAAiB,SAAS;;;UChKzB;;EAEf;;EAEA;;EAEA;;EAEA;;EAEA;;EAEA;;EAEA;;EAEA;;EAEA;;;;;;;cAQW,8BAA8B;WAGvB,QAAQ;EAF1B,YACE,iBACgB,QAAQ;;;;;;;iBAWZ,0BACd,SAAS,cAAc,aACtB;;iBAWa,+BACd,UAAU,cAAc,eACvB;;;;;;;;iBA0Ga,kBACd,SAAS,cAAc,YACvB;EAAQ;IACP;;iBAMa,wBACd,UAAU,cAAc,cACxB;EAAQ;IACP;;;;;;cC/HU,2BAA2B;EACtC,YAAY;;;;;KAQF,kBAAkB,kBAAkB,UAAU,cACxD,SAAS,gBACT,UAAU,WACV,KAAK,oBACF,QAAQ;UAEI,wBAAwB,kBAAkB,UAAU;;EAEnE,UAAU;;EAEV,WAAW;;EAEX,UAAU,kBAAkB,WAAW;;EAEvC,SAAS,YAAY,WAAW;;;EAGhC;;;EAGA;;;;EAIA;;EAEA,WAAW;;EAEX;;EAEA;;;;;;;;EAQA;;;EAGA;;;EAGA;;EAEA;;EAEA,eAAe;EACf,gBAAgB;;;EAGhB,UAAU;;EAEV,YAAY;;;EAGZ,aAAa,UAAU;;;;EAIvB;;;;;;;EAOA,cACE,UAAU,WACV,UAAU;IACL;IAAgB;;;UAGR;EACf;EACA;EACA;;EAEA;;EAEA;EACA;;;EAGA,WAAW;;UAGI;EACf;EACA;;UAGe,uBAAuB,WAAW,kBAAkB;EACnE;EACA;;;;EAIA,SAAS;EACT,WAAW,eAAe;EAC1B,YAAY,eAAe;;EAE3B,YAAY,eAAe;;EAE3B,WAAW;;EAEX,WAAW,eAAe,eAAe,WAAW;;;;;iBA0GhC,iBAAiB,kBAAkB,UAAU,WACjE,MAAM,wBAAwB,WAAW,aACxC,QAAQ,uBAAuB,WAAW;;;;;UC/Q5B;;EAEf;;EAEA,UAAU;;;;;;;UAQK,kBAAkB;;EAEjC;;EAEA,OAAO;;EAEP,cAAc;;;UAIC,wBAAwB;EACvC,SAAS;;;;;;;;;;;UAYM,eAAe,eAAe,YAAY;EACzD,IAAI,OAAO,QAAQ,KAAK,kBAAkB,iBAAiB;;;;;;;iBAQ7C,qBAAqB,eAAe,WAClD,QAAQ,eAAe,UACtB,kBAAkB,QAAQ;;UAQZ,yBAAyB;EACxC;;;;;;;;iBASoB,eACpB,OAAO,WACP,OAAO,eACP,kBAAkB,qBACjB,QAAQ;;UAUM;EACf;EACA;EACA;EACA;EACA;EACA;;;;;;;iBAQc,oBAAoB,mBAAmB,qBAAqB;;UAkB3D;;EAEf;;EAEA;;EAEA;;EAEA;;EAEA;;EAEA;;;iBAIc,kBAAkB,eAAe,kBAAkB;UAwBlD;;EAEf;;EAEA,OAAO;;EAEP;;;;;;;;;iBAkBc,yBACd,eAAe,iBACf,OAAM;;;;;;;;iBCnMQ;;;;;iBAQA,cAAc,gBAAgB;;;;;;;;;;;;;;;;;UCD7B;EACf;;EAEA;;UAGe;EACf;;EAEA;;EAEA;EACA;;EAEA;;;;;;;;;;iBA4Cc,oBACd,SAAS,kBACT;EAAS;IACR;;;;;;;;;;;iBA0Ba,qBACd,SAAS,kBACT,WACA;EAAS;;;;;;;iBC3FK,sBAAsB,WAAW,kBAAkB,UACjE,UAAU,eAAe,WAAW;;;;iBA0CtB,gBAAgB,sBAAsB;UAWrC;;EAEf,YAAY;;;;;EAKZ,WAAW;IAAQ;IAAoB;IAAmB;IAAgB;;;;;iBAK5D,kBAAkB,WAAW,kBAAkB,UAC7D,UAAU,eAAe,WAAW,aACnC;;;;cC/EU,0BAA0B;;cAG1B,mCAAmC;;cAGnC,kCAAkC;;;cCgClC;KAED,mBAAmB;KAEnB;;;UAYK;EACf;EACA;EACA,QAAQ;EACR;;;;UAKe;EACf;EACA;;UAGe;EACf;EACA;;UAGe;EACf;EACA,MAAM;EACN,UAAU;;UAGK;;EAEf;EACA;EACA;EACA;EACA,gBAAgB;;KAGN;UAOK;EACf;EACA,QAAQ;EACR,WAAW;;;EAGX;;UAGe;EACf;EACA,MAAM;;UAGS;EACf;;;EAGA;;UAGe;;EAEf,gBAAgB;;EAEhB,OAAO;;EAEP,YAAY;;KAGF;EAEN;EACA;EACA;EACA;;EAGA;EACA;;KAGM;EAEN;EACA;EACA;;EAGA;;EAEA;EACA;;UAGW;EACf,QAAQ;EACR,MAAM;;UAGS;EACf;EACA;;KAGU;EAEN;EACA;EACA,SAAS;;EAGT;EACA;EACA,SAAS;EACT,SAAS;;EAGT;EACA,SAAS;EACT,OAAO;IAAwB;;;KAGzB;EAEN;EACA;EACA;EACA;EACA;;EAGA;EACA;;;;UAKW;EACf;EACA;EACA;EACA,QAAQ;EACR,UAAU;;UAGF;EACR;EACA;EACA,WAAW;;UAGI,2BAA2B;EAC1C;EACA,MAAM;;UAGS,uCAAuC;EACtD;EACA;EACA;EACA;EACA,SAAS;EACT,UAAU;;UAGK,uCAAuC;EACtD;EACA;EACA;EACA,QAAQ;;UAGO,iCAAiC;EAChD;EACA;EACA;EACA;EACA;IACE;IACA,QAAQ;;EAEV;IACE,OAAO;IACP,OAAO;IACP,WAAW;;EAEb,SAAS;EACT,YAAY;EACZ,iBAAiB;;UAGF,qCAAqC;EACpD;EACA;EACA,eAAe;EACf;IAEM;IACA,OAAO;IACP,QAAQ;;IAGR;IACA,QAAQ;;EAEd;IACM;;IACA;IAAmB,SAAS;;IAC5B;IAAkB,SAAS;;EACjC,YAAY;;UAGG,oCAAoC;EACnD;EACA;EACA;IACM;;IAEA;IACA,QAAQ;;;UAIC,6BAA6B;EAC5C;EACA;IAEM;IACA;;IAGA;IACA,QAAQ;;;KAIJ,oBACR,qBACA,iCACA,iCACA,2BACA,+BACA,8BACA;UAEa;EACf,eAAe;EACf;EACA;EACA,cAAc;EACd,OAAO;EACP,WAAW;;KAGD;EAEN;EACA;EACA;EACA;EACA;;EAGA;EACA;EACA;EACA;EACA;EACA;EACA;;UAGW;EACf;EACA;EACA;EACA;EACA;EACA;EACA;IAAY;IAAgB;IAAgB;;EAC5C;IAAqB;IAAmB;IAAiB;;EACzD;IAAa;IAAkB;IAAkB;;EACjD;IACE;IACA;IACA;IACA;IACA;IACA;;EAEF;EACA;EACA,YAAY;EACZ,UAAU;;UAGK;EACf,SAAS;EACT,MAAM;EACN,YAAY;EACZ,sBAAsB;EACtB,UAAU;EACV,YAAY;EACZ,WAAW;EACX,YAAY;EACZ,OAAO;;UAGQ;EACf,OAAO;;EAEP;EACA,QAAQ;;;;iBA+ZM,0BAA0B,iBAAiB;;;;;;;;;;;;;;;KAsB/C;UAEK;EACf;EACA;EACA,cAAc;;UAGC;WACN;WACA;;WAEA;EACT,OAAO,OAAO,oBAAoB,QAAQ;EAC1C,UAAU,QAAQ;;EAElB,eAAe,QAAQ;;;EAGvB,kBAAkB,QAAQ;;;;;;;EAO1B,oBAAoB,QAAQ;;;;iBAKd,iBAAiB,SAAS,0BAA0B;;cA8BvD,4BAA4B;WAC9B;WACA;WACA;mBACQ;mBACA;EAMjB,YAAY,cAAc,oBAAoB,cAAa;EAerD,UAAU,QAAQ;EAIlB,OAAO,OAAO,oBAAoB,QAAQ;EAU1C,eAAe,QAAQ;EAIvB,kBAAkB,QAAQ;EAI1B,oBAAoB,QAAQ;;;;;;;;;;;;;;;;;;;UC/3BnB;;WAEN;;WAEA;;WAEA;;WAEA;;UAGM;;EAEf;;;;;;;iBAwBc,qBAAqB,MAAM,uBAAuB;;;;iBCpDlD,0BAA0B,2BAA2B,WAAW;;iBAmChE,uBAAuB,2BAA2B,WAAW;;iBA8B7D,iCAAiC,SAAS;;;;iBAa1C,4BAA4B,SAAS;;iBAgBrC,mBAAmB,SAAS;;iBAW5B,YAAY,SAAS;;iBAKrB,kBACd,eAAe,gBACf,iBAAiB;;;;;;;;;;;;;;;;;;;;;UCpGF;;;;;;;WAON;;WAEA,yBAAyB;;;;;;iBAYpB,4BACd,oCACA,OAAM;;;UCnCS;EACf;EACA;EACA;;UAGe,qBAAqB,gBAAgB;EACpD;EACA;EACA;EACA,SACE,QAAQ,SACR,SAAS,4BACR,QAAQ;;UAGI;EACf;EACA;EACA,WAAW;;KAGD,oBAAoB,eAAe,4BAC7C,OAAO,QACP,SAAS,8BACN,qBAAqB,QAAQ;UAEjB;WACN,QAAQ;WACR;;KAGN,sBAAsB,WAAW,KACpC,iBAAiB;EAGjB;;UAGQ,qBAAqB,kBAAkB;EAC/C;EACA;EACA;EACA,aAAa,UAAU;;EAEvB,eAAe;;iBAGD,sBACd,gBAAgB,yBAChB,WACA,kBAAkB,WAAW,UAE7B,WAAW,qBAAqB,UAChC,SAAS,qBAAqB;EAC5B,SAAS;IAAS,UAAU;IAAW,UAAU;MAAc;EAC/D,WAAW,sBAAsB;IAElC,YAAY,WAAW;iBA8CV,qBACd,eAAe,yBACf,WACA,kBAAkB,WAAW,UAE7B,QAAQ,oBAAoB,SAC5B,SAAS,qBAAqB;EAC5B;EACA,SAAS;IAAS,UAAU;IAAW,UAAU;MAAc;;EAEzD;EAAc;;EACd;EAAa,UAAU,sBAAsB;KAEpD,YAAY,WAAW;;;KCxFrB,qBAAqB;KACrB,iBAAiB,SAAS;KAC1B,aAAa,gBAAgB,aAAa,MAAM,mBAAmB;UAcvD;;WAEN;;WAEA;;WAEA;;WAEA;;WAEA;;UAGM;;EAEf,OAAO;IAAQ;IAAiB;MAAkB,QAAQ;;;EAG1D,SAAS,UAAU,UAAU,kBAAkB,QAAQ;;EAEvD,QAAQ,UAAU,WAAW;;;cAIlB,6BAA6B;WAG7B;EAFX,YACE,iBACS;;UAOI;;EAEf;;EAEA;;EAEA;;;;EAIA,MAAM;;UAobS;;EAEf;;EAEA;;EAEA;;;EAGA,YAAY;;;;;iBAoJE,mBAAmB,MAAM,4BAA4B;;;;iBA0FrD,kBACd,SAAS,aACT,uBACC;;;iBAOa,oBAAoB,SAAS,aAAa"}
1
+ {"version":3,"file":"index-DY83M1WL.d.ts","names":[],"sources":["../src/campaign/analyst-surface.ts","../src/paired-arms.ts","../src/campaign/cross-surface-types.ts","../src/campaign/cross-surface-interaction.ts","../src/campaign/fixtures.ts","../src/campaign/gates/neutralization-gate.ts","../src/pre-registration.ts","../src/campaign/gates/sequential.ts","../src/campaign/gates/statistical-heldout.ts","../src/campaign/grounded-reflection.ts","../src/campaign/labeled-store/fs-adapter.ts","../src/campaign/neutralize.ts","../src/produced-state.ts","../src/agent-profile.ts","../src/integrity/backend-integrity.ts","../src/campaign/presets/run-profile-matrix.ts","../src/campaign/presets/playback.ts","../src/campaign/run-dir.ts","../src/campaign/scenario-selection.ts","../src/campaign/score-utils.ts","../src/campaign/search-ledger-errors.ts","../src/campaign/search-ledger.ts","../src/campaign/single-run-lock.ts","../src/campaign/surface-identity.ts","../src/campaign/transient-failure.ts","../src/campaign/upstream-evaluators.ts","../src/campaign/worktree/index.ts"],"mappings":";;;;;;;;;;;;;UAWiB,6BAA6B;EAC5C;EACA,YAAY;EACZ,YAAY;EACZ,yBAAyB;EACzB,kBAAkB;;UAGH;EACf,mBAAmB;EACnB,QAAQ;EACR,MAAM;;UAGS;EACf,QAAQ;IACN;IACA,YAAY;IACZ;IACA,QAAQ;MACN,QAAQ;;iBAGE,iCACd,SAAS,2CAET,SAAS,gBACT,UAAU,sBACV,SAAS,oBACN,QAAQ;iBAgBG,4BAA4B,YAC1C,sBACA;;;;;UClBe;;;EAGf;;;;;EAKA;;EAEA;;EAEA;;EAEA,UAAU;;UAGK;;EAEf;;EAEA;;;UAIe;EACf;;;;EAIA;EACA,UAAU;EACV,WAAW;;UAGI;;EAEf,OAAO;;;EAGP,kBAAkB;;EAElB,mBAAmB;;;;;;;;;;;;;;;;;;;;iBAqBL,SAAS,eAAe,gBAAgB,MAAM,kBAAkB;;UA8G/D;;EAEf;;EAEA;;EAEA,SAAS;;EAET,gBAAgB;;;UAID;EACf;;EAEA;;EAEA;;EAEA;;EAEA;;;;EAIA,aAAa;;EAEb;IAAY;IAAW;;;UAGR,iCAAiC;;;;EAIhD;;EAEA,YAAY;;UAGG;EACf;EACA;EACA;;;EAGA,aAAa;EACb,cAAc;;;;;;;;;;;;;;;iBAgBA,kBACd,eAAe,gBACf,MAAM,2BACL;UAmEc;EACf;EACA;EACA,UAAU;EACV,WAAW;;UAGI;EACf,OAAO;EACP,kBAAkB;EAClB,mBAAmB;;;;;;;;;iBAeL,eACd,uBAAuB,aACvB,wBAAwB,cACvB;;;;KCjXS;;UAGK;EACf;EACA;;EAEA;;;UAIe;EACf;EACA;EACA;EACA;;;UAIe;EACf;;EAEA;;EAEA;;;;;;;UAQe;EACf;EACA;;EAEA;EACA,cAAc;EACd;EACA;;;;;;EAMA,MAAM;EACN,mBAAmB;;EAEnB;;UAGe;EACf;EACA;EACA;;;UAIe;EACf;EACA;EACA;EACA;;EAEA,kCAAkC;;EAElC;;UAGe,qCACf,aAAa,sBAAsB;EAEnC,qBAAqB;EACrB,qBAAqB;EACrB,eAAe;EACf;;EAEA;;EAEA;;EAEA;;EAEA;EACA,WAAW;EACX,WAAW;;UAGI;EACf;EACA;EACA;EACA;EACA;EACA;;UAGe;EACf;EACA;EACA;EACA;;UAGe;EACf,aAAa;EACb;EACA;EACA;EACA;;KAGU;UAWK;EACf;EACA,SAAS;;UAGM;EACf;EACA;EACA;EACA;EACA;EACA;EACA;EACA;;UAGe;EACf,WAAW;EACX,SAAS;EACT,OAAO;EACP,OAAO,eAAe;EACtB,QAAQ;EACR,QAAQ;;EAER,sBAAsB;;EAEtB,aAAa;;UAGE;EACf;EACA;EACA;;EAEA;;UAGe;EACf;EACA;EACA;EACA;EACA;EACA,QAAQ;EACR,cAAc,eAAe;;UAGd;EACf;EACA;EACA;EACA;EACA;;UAGe;EACf;;EAEA;EACA;;UAGe;EACf,SAAS;EACT;EACA;EACA;EACA;EACA,eAAe;EACf,gBAAgB;;KAGN;UAYK;EACf;EACA,SAAS;EACT;;UAGe;EACf;EACA;EACA;EACA;EACA;EACA;EACA;EACA,4BAA4B,iCAAiC;EAC7D,wBAAwB,eAAe;EACvC,QAAQ;EACR,QAAQ;EACR,aAAa;EACb,eAAe;;UAGA;EACf;EACA;EACA;;UAGe;EACf;EACA;EACA,SAAS;;UAGM;;EAEf;EACA;;KAGU;UAaK;EACf;EACA;EACA;EACA;EACA;EACA,uBAAuB;EACvB;EACA;EACA,SAAS;;UAGM;EACf;EACA;EACA,YAAY;EACZ;;;UAIe;EACf;EACA;EACA;EACA;EACA,OAAO;;UAGQ;;EAEf;;EAEA;EACA;;EAEA;EACA;;EAEA,gBAAgB;;EAEhB,OAAO;;UAGQ;EACf,YAAY;EACZ,YAAY;EACZ,kBAAkB;;UAGH,8BACf,aAAa,sBAAsB;EAEnC;EACA;EACA;EACA;;EAEA,MAAM;EACN,iBAAiB;EACjB,iBAAiB;EACjB,YAAY;EACZ,UAAU;EACV,YAAY;;;;;;;;;iBCnRE,gCAAgC,aAAa,qBAC3D,OAAO,qCAAqC,QAC3C,8BAA8B;;;KC7CrB;UAEK;EACf;EACA;EACA;;UAGe;EACf;EACA;EACA;EACA;EACA;EACA;EACA,OAAO;EACP;;UAGe,4BAA4B;EAC3C;EACA;EACA;EACA;EACA;EACA;EACA;EACA;;UAGe;;EAEf,aAAa;;EAEb;;UAGe,wCAAwC;EACvD;;UAGe,0BAA0B,6BACjC,KACN,uBAAuB,qBAAqB;EAG9C;EACA,aAAa;EACb;EACA;EACA,UAAU;;KAGA,qBAAqB;EAC/B,UAAU,MAAM,KAAK;;;iBAIP,qBAAqB;;;;;iBA4BrB,gBACd,kBACA,cACA,UAAS,yBACR;;iBA4Ca,yBACd,kBACA,UAAS,kCACR;;;;;iBAuBa,mBAAmB,qBACjC,SAAS,0BAA0B,aAClC;;;UChJc,0BAA0B,kBAAkB,WAAW;EACtE,WAAW;;;;;EAKX;;;;;;iBAuBc,mBAAmB,WAAW,kBAAkB,UAC9D,SAAS,0BAA0B,aAClC,KAAK,WAAW;;;;;;;;;;;;;UC9CF;EACf;;EAEA;;EAEA;;EAEA;;EAEA;;EAEA;;EAEA;;EAEA;;EAEA;;EAEA;EACA;;;;;;;;;;KAWU;UAEK,uBAAuB;;EAEtC;;;;;;;;;EASA,OAAO;;UAGQ;EACf,UAAU;EACV;EACA;EACA;;;EAGA;;EAEA,kBAAkB;EAGlB;;;;;;;;;;;iBAYc,aAAa;;;;;;;;;;;;;;;;;;;;;;iBA8BP,SAAS,GAAG,KAAK,IAAI;;;;;;;;;;iBAkBrB,aAAa,GAAG,qBAAqB,QAAQ;;;;;;;iBAW7C,eAAe,GAAG,iBAAiB;;;;;iBAWnC,mBACpB,UAAU,gBACV;EAAY;EAAW;EAAgB;IACtC,QAAQ;;;KCpHC;UAEK;EACf,UAAU;;EAEV;;EAEA;;;EAGA;;UAGe;;;EAGf;;;;EAIA;;;EAGA;;EAEA;;;;EAIA;;;EAGA;;;EAGA,kBAAkB;;EAElB;;UAGe,qBAAqB,qBAAqB,kBAAkB,WAAW,kBAC9E,KAAK,WAAW;;;;;;;;EAQxB,QAAQ,gBAAgB;;EAExB,SAAS;IAAkB,UAAU;;;;;;;;;;;;;;;;;iBA0MvB,qBAAqB,qBAAqB,kBAAkB,WAAW,UACrF,SAAS,8BACR,qBAAqB,WAAW;UAsFlB;;EAEf;;EAEA;;EAEA;;EAEA;;UAGe;GACd;IAAQ,SAAS;;IAAyB;IAAe;;;EAE1D,SAAS;;;;;;;;;;;;;;;;;;;;;;;;;iBA0BK,iBAAiB,UAAS,0BAA+B;;;UCtXxD;;EAEf;;EAEA;;EAEA;;;;;;;;;;;;;iBAcc,YACd,WAAW,YAAY,eAAe,cACtC,UAAU,YAAY,eAAe,cACrC,aAAa,aACb,SAAS,GAAG,oCACX;UAsDc;EACf,QAAQ;;;;;;;;;;;;EAYR,WAAW;;;;EAIX,iBAAiB;;;;;;;EAOjB,UAAU;;EAEV,mBAAmB;;EAEnB,SAAS;;;;EAIT;;EAEA;;EAEA;;EAEA,gBAAgB;;EAEhB;;;;EAIA;;EAEA;;UAGe;EACf;EACA;EACA;EACA;;EAEA;EACA;;;;;;;;;;;;;;;;;;;;;;;;;;;;iBA6Bc,oBACd,QAAQ,eACR,OAAM,6BACL;UAkEc;EACf;;;EAGA,WAAW;;;;EAIX;;EAEA;IAAM;IAAa;;;EAEnB,mBAAmB;;EAEnB,SAAS;;EAET;;;;;EAKA;EACA;EACA;;;;;iBAMc,YAAY;;;;;;;;;;;;;;;;;;;iBAsBZ,qBACd,WAAW,YAAY,eAAe,cACtC,UAAU,YAAY,eAAe,cACrC,aAAa,aACb,8BACA;EACE;EACA;EACA;EACA;;;EAGA;IAED;;;;;;;;;;;;;;;;;;;;;;;;;;;;;UC7Tc;WACN;WACA,MAAM,SAAS;;;UAIT;;WAEN;;WAEA;WACA,gBAAgB;;UAGV;;WAEN;;WAEA;;UAGM;;WAEN;;WAEA,eAAe;;WAEf,eAAe;;;;;;;;iBASV,oBACd,mBAAmB,iBACnB,OAAM,6BACL;UAsCc;;WAEN;;WAEA;;;;;;;;;iBAUK,2BACd,cACA,MAAM,KAAK,0DACV;;;UCjFc;;EAEf;;;EAGA;;EAEA;;;cAIW,kCAAkC;WAE3B;EADlB,YACkB,cAChB;;;;;;cAiBS,kCAAkC;mBAIhB;mBAHZ;mBACA;EAEjB,YAA6B,SAAS;EAKhC,QAAQ,OAAO,uBAAuB;EAYtC,OAAO,MAAM,4BAA4B,QAAQ;EA8CjD,QAAQ;IACZ;IACA;IACA,UAAU;IACV,SAAS,OAAO;;UAkCV;UAiCA;UAmBA;UAkBA;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;iBCtNM,eAAe;;;;UCZd;EACf;EACA;;;;;;;UAQe;EACf;EACA;EACA;EACA;EACA;EACA;;;UAIe;EACf;EACA;EACA;EACA;EAIA;;;;;;;;KASU,mBACR,oBACA,oBACA;EACE;;;;;;;;;;;;iBAmBU,qBAAqB,iBAAiB,qBAAqB;;;;;;;;;;cCjE9D,2BAA2B;UAOvB;;;EAGf,MAAM;;EAEN,qBAAqB;;;EAGrB;;;;;;EAMA;;;;;;;cAQW;;;;;;;;;;;;;;;;;;iBAmBG,kBAAkB,MAAM,kBAAkB;;;;;;;;iBAuD1C,cACd,SAAS,KAAK;EACX,SAAS;EAAa;;;;;;;;;iBAiBX,eAAe,SAAS;;;;;iBASxB,oBAAoB,SAAS;;;;;;;;;;;iBA0C7B,iBAAiB,SAAS;;;UChKzB;;EAEf;;EAEA;;EAEA;;EAEA;;EAEA;;EAEA;;EAEA;;EAEA;;EAEA;;;;;;;cAQW,8BAA8B;WAGvB,QAAQ;EAF1B,YACE,iBACgB,QAAQ;;;;;;;iBAWZ,0BACd,SAAS,cAAc,aACtB;;iBAWa,+BACd,UAAU,cAAc,eACvB;;;;;;;;iBA0Ga,kBACd,SAAS,cAAc,YACvB;EAAQ;IACP;;iBAMa,wBACd,UAAU,cAAc,cACxB;EAAQ;IACP;;;;;;cC9HU,2BAA2B;EACtC,YAAY;;;;;KAQF,kBAAkB,kBAAkB,UAAU,cACxD,SAAS,gBACT,UAAU,WACV,KAAK,oBACF,QAAQ;UAEI,wBAAwB,kBAAkB,UAAU;;EAEnE,UAAU;;EAEV,WAAW;;EAEX,UAAU,kBAAkB,WAAW;;EAEvC,SAAS,YAAY,WAAW;;;EAGhC;;;EAGA;;;;EAIA;;EAEA,WAAW;;EAEX;;EAEA;;;;;;;;EAQA;;;EAGA;;;EAGA;;EAEA;;EAEA,eAAe;EACf,gBAAgB;;;EAGhB,UAAU;;EAEV,YAAY;;;EAGZ,aAAa,UAAU;;;;EAIvB;;;;;;;EAOA,cACE,UAAU,WACV,UAAU;IACL;IAAgB;;;UAGR;EACf;EACA;EACA;;EAEA;;EAEA;;EAEA;EACA,gBAAgB;;;EAGhB,WAAW;;UAGI;EACf;EACA;;UAGe,uBAAuB,WAAW,kBAAkB;EACnE;EACA;;;;EAIA,SAAS;EACT,WAAW,eAAe;EAC1B,YAAY,eAAe;;EAE3B,YAAY,eAAe;;EAE3B,WAAW;;EAEX,WAAW,eAAe,eAAe,WAAW;;;;;iBA0GhC,iBAAiB,kBAAkB,UAAU,WACjE,MAAM,wBAAwB,WAAW,aACxC,QAAQ,uBAAuB,WAAW;;;;;UClR5B;;EAEf;;EAEA,UAAU;;;;;;;UAQK,kBAAkB;;EAEjC;;EAEA,OAAO;;EAEP,cAAc;;;UAIC,wBAAwB;EACvC,SAAS;;;;;;;;;;;UAYM,eAAe,eAAe,YAAY;EACzD,IAAI,OAAO,QAAQ,KAAK,kBAAkB,iBAAiB;;;;;;;iBAQ7C,qBAAqB,eAAe,WAClD,QAAQ,eAAe,UACtB,kBAAkB,QAAQ;;UAQZ,yBAAyB;EACxC;;;;;;;;iBASoB,eACpB,OAAO,WACP,OAAO,eACP,kBAAkB,qBACjB,QAAQ;;UAUM;EACf;EACA;EACA;EACA;EACA;EACA;;;;;;;iBAQc,oBAAoB,mBAAmB,qBAAqB;;UAkB3D;;EAEf;;EAEA;;EAEA;;EAEA;;EAEA;;EAEA;;;iBAIc,kBAAkB,eAAe,kBAAkB;UAwBlD;;EAEf;;EAEA,OAAO;;EAEP;;;;;;;;;iBAkBc,yBACd,eAAe,iBACf,OAAM;;;;;;;;iBCnMQ;;;;;iBAQA,cAAc,gBAAgB;;;;;;;;;;;;;;;;;UCD7B;EACf;;EAEA;;UAGe;EACf;;EAEA;;EAEA;EACA;;EAEA;;;;;;;;;;iBA4Cc,oBACd,SAAS,kBACT;EAAS;IACR;;;;;;;;;;;iBA0Ba,qBACd,SAAS,kBACT,WACA;EAAS;;;;;;;iBC3FK,sBAAsB,WAAW,kBAAkB,UACjE,UAAU,eAAe,WAAW;;;;iBA0CtB,gBAAgB,sBAAsB;UAWrC;;EAEf,YAAY;;;;;EAKZ,WAAW;IAAQ;IAAoB;IAAmB;IAAgB;;;;;iBAK5D,kBAAkB,WAAW,kBAAkB,UAC7D,UAAU,eAAe,WAAW,aACnC;;;;cC/EU,0BAA0B;;cAG1B,mCAAmC;;cAGnC,kCAAkC;;;cCgClC;KAED,mBAAmB;KAEnB;;;UAYK;EACf;EACA;EACA,QAAQ;EACR;;;;UAKe;EACf;EACA;;UAGe;EACf;EACA;;UAGe;EACf;EACA,MAAM;EACN,UAAU;;UAGK;;EAEf;EACA;EACA;EACA;EACA,gBAAgB;;KAGN;UAOK;EACf;EACA,QAAQ;EACR,WAAW;;;EAGX;;UAGe;EACf;EACA,MAAM;;UAGS;EACf;;;EAGA;;UAGe;;EAEf,gBAAgB;;EAEhB,OAAO;;EAEP,YAAY;;KAGF;EAEN;EACA;EACA;EACA;;EAGA;EACA;;KAGM;EAEN;EACA;EACA;;EAGA;;EAEA;EACA;;UAGW;EACf,QAAQ;EACR,MAAM;;UAGS;EACf;EACA;;KAGU;EAEN;EACA;EACA,SAAS;;EAGT;EACA;EACA,SAAS;EACT,SAAS;;EAGT;EACA,SAAS;EACT,OAAO;IAAwB;;;KAGzB;EAEN;EACA;EACA;EACA;EACA;;EAGA;EACA;;;;UAKW;EACf;EACA;EACA;EACA,QAAQ;EACR,UAAU;;UAGF;EACR;EACA;EACA,WAAW;;UAGI,2BAA2B;EAC1C;EACA,MAAM;;UAGS,uCAAuC;EACtD;EACA;EACA;EACA;EACA,SAAS;EACT,UAAU;;UAGK,uCAAuC;EACtD;EACA;EACA;EACA,QAAQ;;UAGO,iCAAiC;EAChD;EACA;EACA;EACA;EACA;IACE;IACA,QAAQ;;EAEV;IACE,OAAO;IACP,OAAO;IACP,WAAW;;EAEb,SAAS;EACT,YAAY;EACZ,iBAAiB;;UAGF,qCAAqC;EACpD;EACA;EACA,eAAe;EACf;IAEM;IACA,OAAO;IACP,QAAQ;;IAGR;IACA,QAAQ;;EAEd;IACM;;IACA;IAAmB,SAAS;;IAC5B;IAAkB,SAAS;;EACjC,YAAY;;UAGG,oCAAoC;EACnD;EACA;EACA;IACM;;IAEA;IACA,QAAQ;;;UAIC,6BAA6B;EAC5C;EACA;IAEM;IACA;;IAGA;IACA,QAAQ;;;KAIJ,oBACR,qBACA,iCACA,iCACA,2BACA,+BACA,8BACA;UAEa;EACf,eAAe;EACf;EACA;EACA,cAAc;EACd,OAAO;EACP,WAAW;;KAGD;EAEN;EACA;EACA;EACA;EACA;;EAGA;EACA;EACA;EACA;EACA;EACA;EACA;;UAGW;EACf;EACA;EACA;EACA;EACA;EACA;EACA;IAAY;IAAgB;IAAgB;;EAC5C;IAAqB;IAAmB;IAAiB;;EACzD;IAAa;IAAkB;IAAkB;;EACjD;IACE;IACA;IACA;IACA;IACA;IACA;;EAEF;EACA;EACA,YAAY;EACZ,UAAU;;UAGK;EACf,SAAS;EACT,MAAM;EACN,YAAY;EACZ,sBAAsB;EACtB,UAAU;EACV,YAAY;EACZ,WAAW;EACX,YAAY;EACZ,OAAO;;UAGQ;EACf,OAAO;;EAEP;EACA,QAAQ;;;;iBA+ZM,0BAA0B,iBAAiB;;;;;;;;;;;;;;;KAsB/C;UAEK;EACf;EACA;EACA,cAAc;;UAGC;WACN;WACA;;WAEA;EACT,OAAO,OAAO,oBAAoB,QAAQ;EAC1C,UAAU,QAAQ;;EAElB,eAAe,QAAQ;;;EAGvB,kBAAkB,QAAQ;;;;;;;EAO1B,oBAAoB,QAAQ;;;;iBAKd,iBAAiB,SAAS,0BAA0B;;cA8BvD,4BAA4B;WAC9B;WACA;WACA;mBACQ;mBACA;EAMjB,YAAY,cAAc,oBAAoB,cAAa;EAerD,UAAU,QAAQ;EAIlB,OAAO,OAAO,oBAAoB,QAAQ;EAU1C,eAAe,QAAQ;EAIvB,kBAAkB,QAAQ;EAI1B,oBAAoB,QAAQ;;;;;;;;;;;;;;;;;;;UC/3BnB;;WAEN;;WAEA;;WAEA;;WAEA;;UAGM;;EAEf;;;;;;;iBAwBc,qBAAqB,MAAM,uBAAuB;;;;iBCpDlD,0BAA0B,2BAA2B,WAAW;;iBAmChE,uBAAuB,2BAA2B,WAAW;;iBA8B7D,iCAAiC,SAAS;;;;iBAa1C,4BAA4B,SAAS;;iBAgBrC,mBAAmB,SAAS;;iBAW5B,YAAY,SAAS;;iBAKrB,kBACd,eAAe,gBACf,iBAAiB;;;;;;;;;;;;;;;;;;;;;UCpGF;;;;;;;WAON;;WAEA,yBAAyB;;;;;;iBAYpB,4BACd,oCACA,OAAM;;;UCnCS;EACf;EACA;EACA;;UAGe,qBAAqB,gBAAgB;EACpD;EACA;EACA;EACA,SACE,QAAQ,SACR,SAAS,4BACR,QAAQ;;UAGI;EACf;EACA;EACA,WAAW;;KAGD,oBAAoB,eAAe,4BAC7C,OAAO,QACP,SAAS,8BACN,qBAAqB,QAAQ;UAEjB;WACN,QAAQ;WACR;;KAGN,sBAAsB,WAAW,KACpC,iBAAiB;EAGjB;;UAGQ,qBAAqB,kBAAkB;EAC/C;EACA;EACA;EACA,aAAa,UAAU;;EAEvB,eAAe;;iBAGD,sBACd,gBAAgB,yBAChB,WACA,kBAAkB,WAAW,UAE7B,WAAW,qBAAqB,UAChC,SAAS,qBAAqB;EAC5B,SAAS;IAAS,UAAU;IAAW,UAAU;MAAc;EAC/D,WAAW,sBAAsB;IAElC,YAAY,WAAW;iBA8CV,qBACd,eAAe,yBACf,WACA,kBAAkB,WAAW,UAE7B,QAAQ,oBAAoB,SAC5B,SAAS,qBAAqB;EAC5B;EACA,SAAS;IAAS,UAAU;IAAW,UAAU;MAAc;;EAEzD;EAAc;;EACd;EAAa,UAAU,sBAAsB;KAEpD,YAAY,WAAW;;;KCxFrB,qBAAqB;KACrB,iBAAiB,SAAS;KAC1B,aAAa,gBAAgB,aAAa,MAAM,mBAAmB;UAcvD;;WAEN;;WAEA;;WAEA;;WAEA;;WAEA;;UAGM;;EAEf,OAAO;IAAQ;IAAiB;MAAkB,QAAQ;;;EAG1D,SAAS,UAAU,UAAU,kBAAkB,QAAQ;;EAEvD,QAAQ,UAAU,WAAW;;;cAIlB,6BAA6B;WAG7B;EAFX,YACE,iBACS;;UAOI;;EAEf;;EAEA;;EAEA;;;;EAIA,MAAM;;UAobS;;EAEf;;EAEA;;EAEA;;;EAGA,YAAY;;;;;iBAoJE,mBAAmB,MAAM,4BAA4B;;;;iBA0FrD,kBACd,SAAS,aACT,uBACC;;;iBAOa,oBAAoB,SAAS,aAAa"}
package/dist/index.d.ts CHANGED
@@ -2,7 +2,7 @@ import { a as JudgeError, c as ReplayError, i as ConfigError, l as ValidationErr
2
2
  import { C as verifyAgentProfileCell, S as validateAgentProfileCell, _ as buildAgentInterfaceProfileCell, a as AgentProfileCellSchemaVersion, b as requireAgentProfileCell, c as AgentProfileHarness, d as AgentProfileKind, f as AgentProfileSource, g as assertRunAgentProfileCell, h as agentProfileCellKey, i as AgentProfileCellInput, l as AgentProfileJson, m as agentProfileCellHashMaterial, n as AgentInterfaceProfileLike, o as AgentProfileCellValidationError, p as AgentProfileSourceInput, r as AgentProfileCell, s as AgentProfileDimensionValue, t as AGENT_PROFILE_KINDS, u as AgentProfileJsonObject, v as buildAgentProfileCell, x as toAgentProfileJson, y as groupRunsByAgentProfileCell } from "./agent-profile-cell-Cw0PVwDr.js";
3
3
  import { t as DefaultVerdict } from "./verdict-Dps8_okt.js";
4
4
  import { a as MultiLayerVerifier, c as VerifyContext, i as LayerStatus, l as VerifyOptions, n as Layer, o as Severity, r as LayerResult, s as VerificationReport, t as Finding, u as gradeSemanticStatus } from "./multi-layer-verifier-BHY1gWAc.js";
5
- import { $ as RunScore, B as ConceptSpec, E as FindingSubjectKind, G as SemanticConceptJudgeOptions, H as DEFAULT_COMPLEXITY_WEIGHTS, J as runSemanticConceptJudge, K as SemanticConceptJudgeResult, L as defineTraceAnalyst, M as DspyRlmTraceEngineOptions, N as createDspyRlmTraceEngine, Q as DEFAULT_RUN_SCORE_WEIGHTS, R as ConceptComplexity, T as FindingSubject, U as SEMANTIC_CONCEPT_JUDGE_VERSION, V as ConceptWeightStrategy, W as SemanticConceptJudgeInput, X as RunCriticOptions, Y as RunCritic, Z as RunTrace, _ as FindingsDiff, b as defaultIsMaterial, c as DEFAULT_TRACE_ANALYST_KINDS, d as IMPROVEMENT_KIND_SPEC, et as RunScoreWeights, f as FAILURE_MODE_KIND_SPEC, g as DiffPolicy, h as emitControlIntegrityFindings, l as KNOWLEDGE_POISONING_KIND_SPEC, m as ControlIntegrityAnalyst, n as SkillUsageAnalyst, nt as clamp01, p as CONTROL_INTEGRITY_ANALYST, q as createSemanticConceptJudge, t as SKILL_USAGE_ANALYST, tt as aggregateRunScore, u as KNOWLEDGE_GAP_KIND_SPEC, v as FindingsStore, x as diffFindings, y as PersistedFinding, z as ConceptFinding } from "./skill-usage-CNZTk_LT.js";
5
+ import { $ as RunScore, B as ConceptSpec, E as FindingSubjectKind, G as SemanticConceptJudgeOptions, H as DEFAULT_COMPLEXITY_WEIGHTS, J as runSemanticConceptJudge, K as SemanticConceptJudgeResult, L as defineTraceAnalyst, M as DspyRlmTraceEngineOptions, N as createDspyRlmTraceEngine, Q as DEFAULT_RUN_SCORE_WEIGHTS, R as ConceptComplexity, T as FindingSubject, U as SEMANTIC_CONCEPT_JUDGE_VERSION, V as ConceptWeightStrategy, W as SemanticConceptJudgeInput, X as RunCriticOptions, Y as RunCritic, Z as RunTrace, _ as FindingsDiff, b as defaultIsMaterial, c as DEFAULT_TRACE_ANALYST_KINDS, d as IMPROVEMENT_KIND_SPEC, et as RunScoreWeights, f as FAILURE_MODE_KIND_SPEC, g as DiffPolicy, h as emitControlIntegrityFindings, l as KNOWLEDGE_POISONING_KIND_SPEC, m as ControlIntegrityAnalyst, n as SkillUsageAnalyst, nt as clamp01, p as CONTROL_INTEGRITY_ANALYST, q as createSemanticConceptJudge, t as SKILL_USAGE_ANALYST, tt as aggregateRunScore, u as KNOWLEDGE_GAP_KIND_SPEC, v as FindingsStore, x as diffFindings, y as PersistedFinding, z as ConceptFinding } from "./skill-usage-CRqaHi0R.js";
6
6
  import { C as PendingCostCall, D as costForUsage, E as costForTokenPricing, O as modelPriceKey, S as PaidCallResult, T as RunPaidCallInput, _ as CostReservationExceededError, a as CostChannel, b as CustomTokenPricing, c as CostLedgerHandle, d as CostLedgerPersistenceError, f as CostLedgerSummary, g as CostReceiptInput, h as CostReceiptCaptureError, i as CostCeilingReachedError, l as CostLedgerOptions, m as CostReceipt, n as CostAccountingIncompleteError, o as CostLedger, p as CostProvenance, r as CostCallConflictError, s as CostLedgerFilter, t as ChannelRollup, u as CostLedgerPersistence, v as CostResult, w as PendingCostCallView, x as MaximumCharge, y as CostUsage } from "./cost-ledger-FuQvHxPm.js";
7
7
  import { C as TraceEvent, D as isSandboxSpan, E as isRetrievalSpan, O as isToolSpan, S as ToolSpan, T as isLlmSpan, _ as Span, a as FAILURE_CLASSES, b as SpanStatus, c as JudgeSpan, d as RetrievalSpan, f as Run, g as SandboxSpan, h as RunStatus, i as EventKind, l as LlmSpan, n as BudgetLedgerEntry, o as FailureClass, p as RunLayer, r as BudgetSpec, s as GenericSpan, t as Artifact, u as Message, v as SpanBase, w as isJudgeSpan, x as TRACE_SCHEMA_VERSION, y as SpanKind } from "./schema-BtVldJ3T.js";
8
8
  import { a as RunRecord, c as RunTaskFailure, d as isRunRecord, f as modelHasSnapshot, g as validateRunRecord, h as runTaskScore, i as RunOutcome, l as RunTerminalOutcome, m as roundTripRunRecord, n as RunCostProvenance, o as RunRecordValidationError, p as parseRunRecordSafe, r as RunJudgeMetadata, s as RunSplitTag, t as JudgeScoresRecord, u as RunTokenUsage } from "./run-record-Tb3TTtUn.js";
@@ -11,24 +11,24 @@ import { $ as assertLlmRoute, A as ChatClient, B as SandboxSdkTransportOpts, C a
11
11
  import { a as RunFilter, i as InMemoryTraceStore, n as FileSystemTraceStore, o as SpanFilter, r as FileSystemTraceStoreOptions, s as TraceStore, t as EventFilter } from "./store-CT9YIIve.js";
12
12
  import { a as TraceEmitterOptions, i as TraceEmitter, n as RunCompleteHookContext, o as llmSpanFromProvider, r as SpanHandle, t as RunCompleteHook } from "./emitter-DGQGoLyj.js";
13
13
  import { a as RunIntegrityReport, i as RunIntegrityIssueCode, n as RunIntegrityExpectations, o as assertRunCaptured, r as RunIntegrityIssue, s as throwIfRunIncomplete, t as RunIntegrityError } from "./integrity-DRXobPEs.js";
14
- import { $ as traceAnalystOnRunComplete, A as stringField, At as OtlpSpanRoleInput, B as TraceInsightReadiness, Bt as exportRunAsOtlp, C as ProjectedOtlpSpan, Ct as createOtelTracingStore, D as inferOtlpKind, Dt as OtelExporter, E as firstStringAttr, Et as OtelExportConfig, F as TraceInsightFinding, Ft as traceSpanKindToOpenInferenceKind, G as defaultTraceInsightPanel, H as TraceInsightTask, I as TraceInsightPanelRole, It as OTEL_AGENT_EVAL_SCOPE, J as inferDomainKeywords, K as describeTraceInsightScope, L as TraceInsightPromptInput, Lt as OtlpExport, M as flattenOtlpExportToNdjson, Mt as applyToolSpanOtlpAttributes, N as OtlpFlatLine, Nt as classifyOtlpSpanRole, O as projectOtlpFlatLine, Ot as createOtelExporter, P as TraceInsightContext, Pt as isOtlpModelCall, Q as TraceAnalystHookOptions, R as TraceInsightQualityGate, Rt as OtlpResourceSpans, S as otlpToTraceRunRecords, St as redactValue, T as extractOtlpAttributes, Tt as ExportableSpan, U as buildTraceInsightContext, V as TraceInsightSuite, W as buildTraceInsightPrompt, X as scoreTraceInsightReadiness, Y as planTraceInsightQuestions, Z as tokenizeDomainWords, _ as OtlpTraceRunRecord, _t as DEFAULT_REDACTION_RULES, a as ReplayFetchOptions, at as TraceFileMissingError, b as otlpRowsToTraceRunRecords, bt as RedactionRule, c as ToolTraceMissingError, ct as AnalyzeTracesInput, d as OtlpFileTraceStoreOptions, dt as analyzeTraces, et as SpanNotFoundError, f as ToolSpansToTraceAnalysisStoreOptions, ft as createBoundedTraceAnalysisStore, g as OtlpToRunRecordsOptions, gt as convertTraceStoresToOtlp, h as TRACE_ANALYST_ACTOR_DESCRIPTION_VERSION, ht as TracesToOtlpResult, i as ReplayCacheStats, it as TraceFileMalformedError, j as FlattenOtlpOptions, jt as ToolSpanOtlpInput, k as readOtlpStatus, kt as OtlpSpanRole, l as toolSpansToTraceAnalysisStore, lt as AnalyzeTracesOptions, m as TRACE_ANALYST_ACTOR_DESCRIPTION, mt as TraceStoreToOtlpOptions, n as ReplayCacheEntry, nt as TraceAnalysisStoreContractError, o as createReplayFetch, ot as TraceFileTooLargeError, p as otlpTextToTraceAnalysisStore, pt as TraceStoreSource, q as domainEvidencePattern, r as ReplayCacheMissError, rt as TraceAnalysisValidationError, s as iterateRawCalls, st as TraceNotFoundError, t as ReplayCache, tt as TraceAnalysisLimitError, u as OtlpFileTraceStore, ut as AnalyzeTracesResult, v as TraceAggregate, vt as REDACTION_VERSION, w as asString, wt as otelRunCompleteHook, x as otlpToRunRecords, xt as redactString, y as otlpRowsToRunRecords, yt as RedactionReport, z as TraceInsightQuestion, zt as OtlpSpan } from "./replay-BqTgoioO.js";
14
+ import { $ as traceAnalystOnRunComplete, A as stringField, At as OtlpSpanRoleInput, B as TraceInsightReadiness, Bt as exportRunAsOtlp, C as ProjectedOtlpSpan, Ct as createOtelTracingStore, D as inferOtlpKind, Dt as OtelExporter, E as firstStringAttr, Et as OtelExportConfig, F as TraceInsightFinding, Ft as traceSpanKindToOpenInferenceKind, G as defaultTraceInsightPanel, H as TraceInsightTask, I as TraceInsightPanelRole, It as OTEL_AGENT_EVAL_SCOPE, J as inferDomainKeywords, K as describeTraceInsightScope, L as TraceInsightPromptInput, Lt as OtlpExport, M as flattenOtlpExportToNdjson, Mt as applyToolSpanOtlpAttributes, N as OtlpFlatLine, Nt as classifyOtlpSpanRole, O as projectOtlpFlatLine, Ot as createOtelExporter, P as TraceInsightContext, Pt as isOtlpModelCall, Q as TraceAnalystHookOptions, R as TraceInsightQualityGate, Rt as OtlpResourceSpans, S as otlpToTraceRunRecords, St as redactValue, T as extractOtlpAttributes, Tt as ExportableSpan, U as buildTraceInsightContext, V as TraceInsightSuite, W as buildTraceInsightPrompt, X as scoreTraceInsightReadiness, Y as planTraceInsightQuestions, Z as tokenizeDomainWords, _ as OtlpTraceRunRecord, _t as DEFAULT_REDACTION_RULES, a as ReplayFetchOptions, at as TraceFileMissingError, b as otlpRowsToTraceRunRecords, bt as RedactionRule, c as ToolTraceMissingError, ct as AnalyzeTracesInput, d as OtlpFileTraceStoreOptions, dt as analyzeTraces, et as SpanNotFoundError, f as ToolSpansToTraceAnalysisStoreOptions, ft as createBoundedTraceAnalysisStore, g as OtlpToRunRecordsOptions, gt as convertTraceStoresToOtlp, h as TRACE_ANALYST_ACTOR_DESCRIPTION_VERSION, ht as TracesToOtlpResult, i as ReplayCacheStats, it as TraceFileMalformedError, j as FlattenOtlpOptions, jt as ToolSpanOtlpInput, k as readOtlpStatus, kt as OtlpSpanRole, l as toolSpansToTraceAnalysisStore, lt as AnalyzeTracesOptions, m as TRACE_ANALYST_ACTOR_DESCRIPTION, mt as TraceStoreToOtlpOptions, n as ReplayCacheEntry, nt as TraceAnalysisStoreContractError, o as createReplayFetch, ot as TraceFileTooLargeError, p as otlpTextToTraceAnalysisStore, pt as TraceStoreSource, q as domainEvidencePattern, r as ReplayCacheMissError, rt as TraceAnalysisValidationError, s as iterateRawCalls, st as TraceNotFoundError, t as ReplayCache, tt as TraceAnalysisLimitError, u as OtlpFileTraceStore, ut as AnalyzeTracesResult, v as TraceAggregate, vt as REDACTION_VERSION, w as asString, wt as otelRunCompleteHook, x as otlpToRunRecords, xt as redactString, y as otlpRowsToRunRecords, yt as RedactionReport, z as TraceInsightQuestion, zt as OtlpSpan } from "./replay-BuJM6kLh.js";
15
15
  import { C as TOOL_LATENCY_MS, D as asNumber, E as applyLlmSpanOtlpAttributes, O as contextInputTokens, S as TOOL_ARGS_CAPTURED, T as TOOL_NAME_ATTR_KEYS, _ as LlmSpanOtlpInput, a as LLM_CACHE_WRITE_TOKEN_ATTR_KEYS, b as RUN_COST_ATTR_KEYS, c as LLM_COST_USD, d as LLM_MODEL_ATTR_KEYS, f as LLM_MODEL_NAME, g as LLM_REASONING_TOKEN_ATTR_KEYS, h as LLM_REASONING_TOKENS, i as LLM_CACHE_WRITE_TOKENS, k as firstNumberAttr, l as LLM_INPUT_TOKENS, m as LLM_OUTPUT_TOKEN_ATTR_KEYS, n as LLM_CACHED_TOKENS, o as LLM_CONTEXT_TOKENS, p as LLM_OUTPUT_TOKENS, r as LLM_CACHED_TOKEN_ATTR_KEYS, s as LLM_COST_ATTR_KEYS, t as INPUT_VALUE, u as LLM_INPUT_TOKEN_ATTR_KEYS, v as OPENINFERENCE_SPAN_KIND, w as TOOL_NAME, x as SPAN_KIND_ATTR_KEYS, y as OUTPUT_VALUE } from "./attribute-vocabulary-DLJ6303h.js";
16
16
  import { a as judgeSpans, c as runsForScenario, i as hasCapturedToolArgs, l as toolSpans, n as argHash, o as llmSpans, r as groupBy, s as runFailureClass, t as aggregateLlm } from "./query-CJ_DX8vl.js";
17
17
  import { A as TRACE_ANALYST_TRUNCATION_MARKER_PREFIX, C as DEFAULT_TRACE_ANALYST_BUDGETS, D as SearchSpanResult, E as QueryTracesPage, F as TraceAnalystSpanStatus, I as TraceAnalystTraceSummary, L as ViewSpansResult, M as TraceAnalystFilters, N as TraceAnalystSpan, O as SearchTraceResult, P as TraceAnalystSpanKind, R as ViewTraceOversized, S as TraceAnalysisStoreContext, T as ErrorCluster, _ as makeFinding, a as AnalystInputKind, b as TRACE_ANALYSIS_LIMITS, c as AnalystRunInputs, d as AnalystSeverity, f as AnalystUsageReceipt, g as computeFindingId, h as ProposalFindingOrigin, i as AnalystFinding, j as TraceAnalystByteBudgets, k as SpanMatchRecord, l as AnalystRunResult, m as ProposalFinding, n as AnalystContext, o as AnalystRequirements, p as EvidenceRef, r as AnalystCost, s as AnalystRunEvent, t as Analyst, u as AnalystRunSummary, v as makeProposalFinding, w as DatasetOverview, x as TraceAnalysisStore, y as BoundedTraceAnalysisStoreOptions, z as ViewTraceResult } from "./types-D3jh6F98.js";
18
- import { a as createTraceAnalyst, c as runTraceAnalyst, f as BehavioralMetrics, g as computeTraceMetrics, h as SuboptimalSignal, i as TraceAnalystDefinition, m as SuboptimalCode, n as buildDefaultAnalystRegistry, o as renderPriorFindings, p as BehavioralTokenSequence, r as CreateTraceAnalystOptions, s as renderUpstreamFindings, t as DefaultAnalystRegistryOptions } from "./default-registry-Ci7wAAR8.js";
18
+ import { a as createTraceAnalyst, c as runTraceAnalyst, f as BehavioralMetrics, g as computeTraceMetrics, h as SuboptimalSignal, i as TraceAnalystDefinition, m as SuboptimalCode, n as buildDefaultAnalystRegistry, o as renderPriorFindings, p as BehavioralTokenSequence, r as CreateTraceAnalystOptions, s as renderUpstreamFindings, t as DefaultAnalystRegistryOptions } from "./default-registry-D3uqKbo6.js";
19
19
  import { a as ExactAnalystRunPolicySnapshot, c as ExactAnalystSnapshot, d as ExactExecutionComponentSnapshot, i as ExactAnalystRunEvent, l as ExactCapableAnalyst, n as ExactAnalystExecutionPlanSnapshot, o as ExactAnalystRunResult, r as ExactAnalystRunCompletion, s as ExactAnalystRunSummary, t as ExactAnalystBudgetSnapshot, u as ExactExecutionComponentIdentity } from "./exact-types-B0lJV3tu.js";
20
20
  import { A as ExactAnalystRunExecutionError, C as jsonHasKeys, D as AnalystRegistryOptions, E as AnalystRegistry, M as RegistryRunOpts, N as assertExactRegistryRunOpts, O as BudgetPolicy, S as containsAll, T as AnalystHooks, _ as ValidationContext, a as ProducedProposal, b as byteLengthRange, c as SatisfiedBy, d as createLlmCorrectnessChecker, f as createTokenRecallChecker, g as ArtifactValidator, h as Artifact$1, i as LlmCorrectnessCheckerOpts, j as ExactRegistryRunOpts, k as ExactAnalystBudgetPolicy, l as TaskGold, m as verifyCompletion, n as CompletionVerdict, o as ProducedState, p as parseCorrectnessResponse, r as CorrectnessChecker, s as RequirementCheck, t as CompletionRequirement, u as completionVerdict, v as ValidationIssue, w as regexMatch, x as composeValidators, y as ValidationResult } from "./completion-verifier-D15NHYSk.js";
21
- import { R as Scenario$1, S as JudgeConfig$1, w as JudgeScore$1 } from "./types-Dk7PB7vh.js";
21
+ import { R as Scenario$1, S as JudgeConfig$1, w as JudgeScore$1 } from "./types-KEqL1pZc.js";
22
22
  import { a as DatasetScenario, c as SliceOptions, i as DatasetProvenance, l as hashScenarios, n as DatasetDifficulty, o as DatasetSplit, r as DatasetManifest, s as HoldoutLockedError, t as Dataset } from "./dataset-v_Y5902-.js";
23
23
  import { $ as pairedCohensDz, A as WeightedCompositeInput, At as positionalBias, B as eProcess, C as RankTestMethod, Ct as GoldenItem, D as ScoreRiskDifferenceResult, Dt as calibrateJudge, E as RiskDifferenceResult, Et as VerbosityBiasResult, F as cliffsDelta, G as mannWhitneyU, H as interRaterReliability, I as cohensD, J as mcnemarRequiredN, K as mcnemar, L as confidenceInterval, M as WilcoxonSignedRankResult, Mt as verbosityBias, N as benjaminiHochberg, O as SignTestAlternative, Ot as calibrateJudgeContinuous, P as bonferroni, Q as pairedBootstrap, R as corpusInterRaterAgreement, S as ProportionInterval, St as ContinuousCalibrationResult, T as RankTestOptions, Tt as SelfPreferenceResult, U as interpretCliffs, V as holm, W as isBinaryOutcomeVector, X as normalizeScores, Y as mulberry32, Z as pairedBinaryScale, _ as McNemarResult, _t as wilson, a as CorpusAgreementReport, at as pairedSignTest, b as PairedSignTestResult, bt as ContinuousAgreement, c as DEFAULT_PERMUTATIONS, ct as passAtK, d as EProcessState, dt as requiredPairedSampleSize, et as pairedDeltaTieFraction, f as EProcessStep, ft as requiredSampleSize, g as MannWhitneyResult, gt as wilcoxonSignedRank, h as MANN_WHITNEY_EXACT_MAX_WORK, ht as weightedMean, i as CorpusAgreementPerDimension, it as pairedRiskDifferenceScore, j as WeightedCompositeResult, jt as selfPreference, k as WILCOXON_EXACT_MAX_N, kt as continuousAgreement, l as EProcess, lt as pearsonR, m as MANN_WHITNEY_EXACT_MAX_STATES, mt as weightedComposite, n as CliffsMagnitude, nt as pairedRiskDifference, o as CorpusScoreRecord, ot as pairedTTest, p as ExactRiskDifferenceResult, pt as spearmanR, q as mcnemarPower, r as CorpusAgreementOptions, rt as pairedRiskDifferenceExact, s as DECISION_PAIRED_DELTA_STATISTIC, st as partialCredit, t as BOOTSTRAP_GATE_MIN_N, tt as pairedMde, u as EProcessOptions, ut as ranks, v as PairedBootstrapOptions, vt as CalibrationResult, w as RankTestMethodRequest, wt as PositionalBiasResult, x as PairedTTestResult, xt as ContinuousAgreementOptions, y as PairedBootstrapResult, yt as CandidateScore, z as corpusInterRaterAgreementFromJudgeScores } from "./statistics-B4u_CiFd.js";
24
24
  import { A as PairedMcNemarEvidence, C as HeldOutGate, D as PairedDecisionMethod, E as SplitCoverage, M as PairedPromotionDecisionOptions, N as decidePairedPromotion, O as PairedDecisionShape, P as pairedDecisionShape, S as GateEvidence, T as HeldOutGateRejectionCode, _ as paretoChart, a as ParetoPoint, b as DeltaStatistic, c as ResearchReportCandidate, d as ResearchReportOptions, f as ResearchReportRecommendation, g as gainHistogram, h as SummaryTableRow, i as ParetoFigureSpec, j as PairedPromotionDecision, k as PairedDecisionStatistic, l as ResearchReportDecision, m as SummaryTableOptions, n as GainDistributionFigureSpec, o as RESEARCH_REPORT_HARD_PAIR_FLOOR, p as SummaryTable, r as GainDistributionOptions, s as ResearchReport, t as GainDistributionBin, u as ResearchReportMethodology, v as researchReport, w as HeldOutGateConfig, x as GateDecision, y as summaryTable } from "./summary-report-o3eJ3gxG.js";
25
25
  import { a as FailureClassification, c as classifyFailure, i as DEFAULT_RULES, o as FailureContext, s as FailureRule } from "./failure-cluster-CqcvCcdR.js";
26
- import { C as MakeEvalToolsConfig, S as EvalToolDef, T as toOpenAiTool, _ as BuildTraceAnalysisToolsOptions, a as TraceAnalysisEngineRequest, b as buildTraceAnalysisToolDescriptors, c as resolveTraceAnalystLimits, d as RawAnalystEvidence, i as TraceAnalysisEngine, o as TraceAnalysisEngineResult, p as RawAnalystFinding, r as DEFAULT_TRACE_ANALYST_LIMITS, s as TraceAnalystLimits, v as TRACE_ANALYST_TOOL_NAMESPACE, w as makeEvalTools, x as traceAnalystFunctionGroup, y as TraceAnalysisToolDescriptor } from "./tool-groups-DVQTy9lq.js";
26
+ import { C as MakeEvalToolsConfig, S as EvalToolDef, T as toOpenAiTool, _ as BuildTraceAnalysisToolsOptions, a as TraceAnalysisEngineRequest, b as buildTraceAnalysisToolDescriptors, c as resolveTraceAnalystLimits, d as RawAnalystEvidence, i as TraceAnalysisEngine, o as TraceAnalysisEngineResult, p as RawAnalystFinding, r as DEFAULT_TRACE_ANALYST_LIMITS, s as TraceAnalystLimits, v as TRACE_ANALYST_TOOL_NAMESPACE, w as makeEvalTools, x as traceAnalystFunctionGroup, y as TraceAnalysisToolDescriptor } from "./tool-groups-Bqy4A3QB.js";
27
27
  import { A as isTrainableSplit, D as assertRolloutLine, E as assertMintedLines, T as assertMinted, b as RolloutSplit, h as RolloutLine, i as ChatToolCall, j as validateRolloutLine, k as isRolloutLine, n as ChatMessage, o as MintedRolloutLine, p as RolloutCapture, s as MintedRolloutOutcome, u as ROLLOUT_SCHEMA, w as ToolDef, x as RolloutStep, y as RolloutRole } from "./schema-Cef2cFmb.js";
28
28
  import { C as Unavailable, D as showMeasured, E as isUnavailable, S as SupervisorRunTreeGapCode, _ as SupervisorRunReport, b as SupervisorRunTree, f as SUPERVISOR_RUN_SCHEMA, g as SupervisorRunReader, h as SupervisorRunNodeRole, p as SourceLimits, r as Measured, v as SupervisorRunRollup, x as SupervisorRunTreeGap, y as SupervisorRunSources } from "./types-Dea6tiVI.js";
29
- import { B as BenchmarkSource, F as BenchmarkDatasetItem, H as deterministicSplit, I as BenchmarkEvaluation, L as BenchmarkFamily, N as BENCHMARK_SPLIT_SEED, P as BenchmarkAdapter, R as BenchmarkResponder, V as BenchmarkTaskKind, t as index_d_exports, z as BenchmarkScenario } from "./index-CKI1CXTL.js";
30
- import { $ as Objective, At as CanarySeverity, Cn as ReferenceEquivalenceJudgeResult, Ct as scoreRedTeamOutput, Dn as LlmJudgeDimension, Dt as CanaryKind, En as runReferenceEquivalenceJudge, Et as CanaryEvaluation, On as LlmJudgeOptions, Ot as CanaryOptions, Q as Direction, Sn as ReferenceEquivalenceJudgeOptions, St as redTeamReport, Tn as createReferenceEquivalenceJudge, Tt as CanaryAlert, _t as RedTeamCategory, at as scalarScore, bn as REFERENCE_EQUIVALENCE_JUDGE_VERSION, bt as RedTeamReport, et as ParetoResult, gt as RedTeamCase, ht as DEFAULT_RED_TEAM_CORPUS, it as paretoFrontierWithCrowding, jt as runCanaries, kn as llmJudge, kt as CanaryReport, nt as dominates, rt as paretoFrontier, tt as crowdingDistance, vt as RedTeamFinding, wn as ReferenceEquivalenceScenario, wt as toolNamesForRun, xn as ReferenceEquivalenceJudgeInput, xt as redTeamDataset, yn as REFERENCE_EQUIVALENCE_INPUT_LIMITS, yt as RedTeamPayload } from "./skillopt-optimization-method-D1dqGzzH.js";
31
- import { $r as pairRunRecords, $t as HarnessType, Bn as SignedManifest, Gn as signManifest, Gr as PairArmsResult, Gt as BackendIntegrityReport, Hn as canonicalize, Hr as MatchedPair, Jr as PairedArmsComparison, Jt as summarizeAgentReceiptIntegrity, Kn as verifyManifest, Kr as PairRunRecordsResult, Kt as assertRealAgentReceipts, Qr as pairArms, Qt as HARNESS_NATIVE_MODEL, Rn as HypothesisManifest, Un as evaluateHypothesis, Ur as MatchedRunRecordPair, Vn as SignedManifestAlgo, Vr as ComparePairedArmsOptions, Wn as hashJson, Wr as PairArmsOptions, Wt as BackendIntegrityError, Xr as PairedMetricDelta, Xt as AgentProfile, Yr as PairedCorrectness, Yt as summarizeBackendIntegrity, Zr as comparePairedArms, Zt as CODING_HARNESSES, an as harnessAxisOf, cn as RuntimeEventLike, en as ProfileAxisSpec, in as expandProfileAxes, ln as ToolCallEventLike, nn as agentProfileId, on as ArtifactEventLike, qr as PairedArmRow, qt as assertRealBackend, rn as agentProfileModelId, sn as ProposalEventLike, tn as agentProfileHash, un as extractProducedState, zn as HypothesisResult } from "./index-DCP4I2Qx.js";
29
+ import { B as BenchmarkSource, F as BenchmarkDatasetItem, H as deterministicSplit, I as BenchmarkEvaluation, L as BenchmarkFamily, N as BENCHMARK_SPLIT_SEED, P as BenchmarkAdapter, R as BenchmarkResponder, V as BenchmarkTaskKind, t as index_d_exports, z as BenchmarkScenario } from "./index-CaOjqCli.js";
30
+ import { $ as Objective, At as CanarySeverity, Cn as ReferenceEquivalenceJudgeResult, Ct as scoreRedTeamOutput, Dn as LlmJudgeDimension, Dt as CanaryKind, En as runReferenceEquivalenceJudge, Et as CanaryEvaluation, On as LlmJudgeOptions, Ot as CanaryOptions, Q as Direction, Sn as ReferenceEquivalenceJudgeOptions, St as redTeamReport, Tn as createReferenceEquivalenceJudge, Tt as CanaryAlert, _t as RedTeamCategory, at as scalarScore, bn as REFERENCE_EQUIVALENCE_JUDGE_VERSION, bt as RedTeamReport, et as ParetoResult, gt as RedTeamCase, ht as DEFAULT_RED_TEAM_CORPUS, it as paretoFrontierWithCrowding, jt as runCanaries, kn as llmJudge, kt as CanaryReport, nt as dominates, rt as paretoFrontier, tt as crowdingDistance, vt as RedTeamFinding, wn as ReferenceEquivalenceScenario, wt as toolNamesForRun, xn as ReferenceEquivalenceJudgeInput, xt as redTeamDataset, yn as REFERENCE_EQUIVALENCE_INPUT_LIMITS, yt as RedTeamPayload } from "./skillopt-optimization-method-CwRvi-rL.js";
31
+ import { $r as pairRunRecords, $t as HarnessType, Bn as SignedManifest, Gn as signManifest, Gr as PairArmsResult, Gt as BackendIntegrityReport, Hn as canonicalize, Hr as MatchedPair, Jr as PairedArmsComparison, Jt as summarizeAgentReceiptIntegrity, Kn as verifyManifest, Kr as PairRunRecordsResult, Kt as assertRealAgentReceipts, Qr as pairArms, Qt as HARNESS_NATIVE_MODEL, Rn as HypothesisManifest, Un as evaluateHypothesis, Ur as MatchedRunRecordPair, Vn as SignedManifestAlgo, Vr as ComparePairedArmsOptions, Wn as hashJson, Wr as PairArmsOptions, Wt as BackendIntegrityError, Xr as PairedMetricDelta, Xt as AgentProfile, Yr as PairedCorrectness, Yt as summarizeBackendIntegrity, Zr as comparePairedArms, Zt as CODING_HARNESSES, an as harnessAxisOf, cn as RuntimeEventLike, en as ProfileAxisSpec, in as expandProfileAxes, ln as ToolCallEventLike, nn as agentProfileId, on as ArtifactEventLike, qr as PairedArmRow, qt as assertRealBackend, rn as agentProfileModelId, sn as ProposalEventLike, tn as agentProfileHash, un as extractProducedState, zn as HypothesisResult } from "./index-DY83M1WL.js";
32
32
  import { _ as vitestTestParser, a as DockerSandboxDriver, c as SandboxHarness, d as SubprocessSandboxDriver, f as SubprocessSandboxDriverOptions, g as pytestTestParser, h as jestTestParser, i as runTestGradedScenario, l as SandboxHarnessResult, m as composeParsers, n as TestGradedRunResult, o as HarnessConfig, p as TestOutputParser, r as TestGradedScenario, s as SandboxDriver, t as TestGradedRunOptions, u as SandboxResult } from "./test-graded-scenario-D1TaI2va.js";
33
33
  import { $ as ControlRunResult, A as feedbackTrajectoryToOptimizerRow, B as AnalystReviewCounts, C as analystRunToReviewRequests, D as feedbackTrajectoriesToDatasetScenarios, E as createFeedbackTrajectory, F as serializeFeedbackTrajectoriesJsonl, G as analystFindingDigest, H as AnalystReviewQuality, I as summarizePreferenceMemory, J as ControlActionOutcome, K as analystRunDigest, L as withAssignedFeedbackSplit, M as renderPreferenceMemoryMarkdown, N as replayFeedbackTrajectories, O as feedbackTrajectoriesToOptimizerRows, P as replayFeedbackTrajectory, Q as ControlEvalResult, R as AnalystFindingDigest, S as analystRunToFeedbackTrajectory, T as controlRunToFeedbackTrajectory, U as AnalystReviewSource, V as AnalystReviewDecision, W as AnalystRunDigest, X as ControlContext, Y as ControlBudget, Z as ControlDecision, _ as FeedbackTrajectoryStore, a as FeedbackLabel, at as StopDecision, b as PreferenceMemoryEntry, c as FeedbackOptimizerRow, ct as runAgentControlLoop, d as FeedbackReplayResult, dt as subjectiveEval, et as ControlRuntimeConfig, f as FeedbackSeverity, g as FeedbackTrajectoryFilter, h as FeedbackTrajectory, i as FeedbackAttempt, it as ControlStopPolicies, j as parseFeedbackTrajectoriesJsonl, k as feedbackTrajectoryToDatasetScenario, l as FeedbackOutcome, lt as stopOnNoProgress, m as FeedbackTask, n as AnalystReviewRequest, nt as ControlSeverity, o as FeedbackLabelKind, ot as allCriticalPassed, p as FeedbackSplitPolicy, q as ControlActionFailureMode, r as FeedbackArtifactType, rt as ControlStep, s as FeedbackLabelSource, st as objectiveEval, t as AnalystFeedbackTrajectoryOptions, tt as ControlRuntimeError, u as FeedbackReplayAdapter, ut as stopOnRepeatedAction, v as FileSystemFeedbackTrajectoryStore, w as assignFeedbackSplit, x as ProposedSideEffect, y as InMemoryFeedbackTrajectoryStore, z as AnalystMissedIssue } from "./feedback-trajectory-BCHqzLh3.js";
34
34
  import { A as ActionExecutionPolicy, C as ReviewMemoryStore, D as inMemoryReviewStore, E as createLlmReviewer, M as evaluateActionPolicy, O as jsonlReviewStore, S as ReviewMemoryEntry, T as VerifyFn, _ as ProposeReviewReport, a as ProposeReviewControlAction, b as ReviewFn, c as ProposeReviewControlState, d as LlmJsonCall, f as LlmReviewerConfig, g as ProposeReviewConfig, h as ProposeOutput, i as scoreFromEvals, j as ActionPolicyDecision, k as runProposeReview, l as controlFailureClassFromVerification, m as ProposeInput, n as RunEvidenceMetadata, o as ProposeReviewControlConfig, p as ProposeFn, r as controlRunToRunRecord, s as ProposeReviewControlResult, t as ControlRunToRunRecordOptions, u as runProposeReviewAsControlLoop, v as ProposeReviewShot, w as Verification, x as ReviewInput, y as Review } from "./run-evidence-8Ou28QSa.js";
package/dist/index.js CHANGED
@@ -15,9 +15,9 @@ import { c as assertMintedLines, f as isRolloutLine, i as ROLLOUT_SCHEMA, m as v
15
15
  import { a as scoreOrigin, n as observedScore, o as trainingReward, r as observedSplitScore, s as trainingScore, t as isRealnessGated } from "./reward-nw2xZGZG.js";
16
16
  import { a as clamp01, i as aggregateRunScore, r as DEFAULT_RUN_SCORE_WEIGHTS } from "./proposal-findings-2GIUo1et.js";
17
17
  import { a as RunCritic, d as defaultIsMaterial, f as diffFindings, h as defineTraceAnalyst, i as runSemanticConceptJudge, n as SEMANTIC_CONCEPT_JUDGE_VERSION, o as SKILL_USAGE_ANALYST, p as LockedJsonlAppender, r as createSemanticConceptJudge, s as SkillUsageAnalyst, t as DEFAULT_COMPLEXITY_WEIGHTS, u as FindingsStore } from "./semantic-concept-judge-DJQtFr95.js";
18
- import { t as createDspyRlmTraceEngine } from "./dspy-rlm-engine-Bw8Zz9CX.js";
18
+ import { t as createDspyRlmTraceEngine } from "./dspy-rlm-engine-CJ4sN8Wx.js";
19
19
  import { f as Mutex, p as mapConcurrent } from "./ledger-core-Dxz0Rkwa.js";
20
- import { At as pairedDeltaTest, Ct as paretoFrontier, D as runCanaries, Dt as decidePairedPromotion, Ft as summarizeBackendIntegrity, G as buildReflectionPrompt, It as JudgeParseError, K as parseReflectionResponse, Mt as assertRealAgentReceipts, Nt as assertRealBackend, Ot as pairedDecisionShape, Pt as summarizeAgentReceiptIntegrity, St as dominates, Tt as scalarScore, W as DEFAULT_MUTATION_PRIMITIVES, _t as canonicalJson, at as createReferenceEquivalenceJudge, bt as inMemoryVerdictCache, ct as redTeamDataset, dt as toolNamesForRun, ft as Dataset, gt as cachedJudge, ht as llmJudge, it as REFERENCE_EQUIVALENCE_JUDGE_VERSION, jt as BackendIntegrityError, kt as minimumPairsForPairedDeltaTest, lt as redTeamReport, mt as hashScenarios, ot as runReferenceEquivalenceJudge, pt as HoldoutLockedError, rt as REFERENCE_EQUIVALENCE_INPUT_LIMITS, st as DEFAULT_RED_TEAM_CORPUS, ut as scoreRedTeamOutput, vt as contentHash, wt as paretoFrontierWithCrowding, xt as crowdingDistance, yt as fileVerdictCache, z as surfaceContentHash } from "./skillopt-optimization-method-DV6c4ooy.js";
20
+ import { At as pairedDeltaTest, Ct as paretoFrontier, D as runCanaries, Dt as decidePairedPromotion, Ft as summarizeBackendIntegrity, G as buildReflectionPrompt, It as JudgeParseError, K as parseReflectionResponse, Mt as assertRealAgentReceipts, Nt as assertRealBackend, Ot as pairedDecisionShape, Pt as summarizeAgentReceiptIntegrity, St as dominates, Tt as scalarScore, W as DEFAULT_MUTATION_PRIMITIVES, _t as canonicalJson, at as createReferenceEquivalenceJudge, bt as inMemoryVerdictCache, ct as redTeamDataset, dt as toolNamesForRun, ft as Dataset, gt as cachedJudge, ht as llmJudge, it as REFERENCE_EQUIVALENCE_JUDGE_VERSION, jt as BackendIntegrityError, kt as minimumPairsForPairedDeltaTest, lt as redTeamReport, mt as hashScenarios, ot as runReferenceEquivalenceJudge, pt as HoldoutLockedError, rt as REFERENCE_EQUIVALENCE_INPUT_LIMITS, st as DEFAULT_RED_TEAM_CORPUS, ut as scoreRedTeamOutput, vt as contentHash, wt as paretoFrontierWithCrowding, xt as crowdingDistance, yt as fileVerdictCache, z as surfaceContentHash } from "./skillopt-optimization-method-Di8Xkalt.js";
21
21
  import { $ as selfPreference, A as pairedRiskDifference, B as requiredSampleSize, C as mulberry32, D as pairedCohensDz, E as pairedBootstrap, F as partialCredit, G as wilson, H as weightedComposite, I as passAtK, J as normalCdf, K as studentTCdf, L as pearsonR, M as pairedRiskDifferenceScore, N as pairedSignTest, O as pairedDeltaTieFraction, P as pairedTTest, Q as positionalBias, R as ranks, S as mcnemarRequiredN, T as pairedBinaryScale, U as weightedMean, V as spearmanR, W as wilcoxonSignedRank, X as calibrateJudgeContinuous, Y as calibrateJudge, Z as continuousAgreement, _ as interpretCliffs, a as MANN_WHITNEY_EXACT_MAX_WORK, b as mcnemar, c as bonferroni, d as confidenceInterval, et as verbosityBias, f as corpusInterRaterAgreement, g as interRaterReliability, h as holm, i as MANN_WHITNEY_EXACT_MAX_STATES, j as pairedRiskDifferenceExact, k as pairedMde, l as cliffsDelta, m as eProcess, n as DECISION_PAIRED_DELTA_STATISTIC, o as WILCOXON_EXACT_MAX_N, p as corpusInterRaterAgreementFromJudgeScores, r as DEFAULT_PERMUTATIONS, s as benjaminiHochberg, t as BOOTSTRAP_GATE_MIN_N, u as cohensD, v as isBinaryOutcomeVector, w as normalizeScores, x as mcnemarPower, y as mannWhitneyU, z as requiredPairedSampleSize } from "./statistics-ByxzSiOM.js";
22
22
  import { n as pairArms, r as pairRunRecords, t as comparePairedArms } from "./paired-arms-iZ08VFMN.js";
23
23
  import { n as llmSpanFromProvider, t as TraceEmitter } from "./emitter-CPBAhxum.js";
@@ -32,7 +32,7 @@ import { C as SUPERVISOR_RUN_SCHEMA, T as showMeasured, _ as readClaudeCodeSuper
32
32
  import { A as TRACE_ANALYST_ACTOR_DESCRIPTION, C as domainEvidencePattern, D as tokenizeDomainWords, E as scoreTraceInsightReadiness, O as traceAnalystOnRunComplete, S as describeTraceInsightScope, T as planTraceInsightQuestions, _ as otlpToTraceRunRecords, a as convertTraceStoresToOtlp, b as buildTraceInsightPrompt, c as otelRunCompleteHook, d as captureFetchToRawSink, f as ToolTraceMissingError, g as otlpToRunRecords, h as otlpRowsToTraceRunRecords, i as iterateRawCalls, j as TRACE_ANALYST_ACTOR_DESCRIPTION_VERSION, k as analyzeTraces, l as OTEL_AGENT_EVAL_SCOPE, m as otlpRowsToRunRecords, n as ReplayCacheMissError, o as createOtelExporter, p as toolSpansToTraceAnalysisStore, r as createReplayFetch, s as createOtelTracingStore, t as ReplayCache, u as exportRunAsOtlp, v as flattenOtlpExportToNdjson, w as inferDomainKeywords, x as defaultTraceInsightPanel, y as buildTraceInsightContext } from "./replay-DjUfTrHD.js";
33
33
  import { i as otlpTextToTraceAnalysisStore, n as OtlpFileTraceStore } from "./store-otlp-D4I90_vR.js";
34
34
  import { n as extractUsageFromResponse, r as extractUsageFromSse, t as extractUsage } from "./extract-usage-7l1Xq5ti.js";
35
- import { B as agentProfileModelId, G as createLlmCorrectnessChecker, H as harnessAxisOf, I as CODING_HARNESSES, J as verifyCompletion, K as createTokenRecallChecker, L as HARNESS_NATIVE_MODEL, R as agentProfileHash, U as extractProducedState, V as expandProfileAxes, W as completionVerdict, q as parseCorrectnessResponse, z as agentProfileId } from "./campaign-COFTVy3Q.js";
35
+ import { B as agentProfileModelId, G as createLlmCorrectnessChecker, H as harnessAxisOf, I as CODING_HARNESSES, J as verifyCompletion, K as createTokenRecallChecker, L as HARNESS_NATIVE_MODEL, R as agentProfileHash, U as extractProducedState, V as expandProfileAxes, W as completionVerdict, q as parseCorrectnessResponse, z as agentProfileId } from "./campaign-C7ZGiZYG.js";
36
36
  import { n as iqr, r as welchsTTest, t as compareToBaseline } from "./baseline-C-GocmIW.js";
37
37
  import { a as judgeSpans, c as runsForScenario, i as hasCapturedToolArgs, l as toolSpans, n as argHash, o as llmSpans, r as groupBy, s as runFailureClass, t as aggregateLlm } from "./query-Di7eEQ79.js";
38
38
  import { a as checkBehavioralCanary, i as canaryLeakView, o as checkCanaries, r as HoldoutAuditor, s as runBehavioralCanaries, t as analyzeRuns } from "./analyze-runs-BScZvqMV.js";
@@ -42,7 +42,7 @@ import { a as InMemoryTraceStore, i as FileSystemTraceStore, n as assertRunCaptu
42
42
  import { i as redactValue, n as REDACTION_VERSION, r as redactString, t as DEFAULT_REDACTION_RULES } from "./redact-7Aq1ukl-.js";
43
43
  import { n as DEFAULT_RULES, r as classifyFailure, t as computeToolUseMetrics } from "./tool-use-metrics-DEGMKycK.js";
44
44
  import { t as analyzeSeries } from "./series-convergence-CjO2QdRW.js";
45
- import { _ as deterministicSplit, g as BENCHMARK_SPLIT_SEED, t as benchmarks_exports } from "./benchmarks-Ce8GdJbc.js";
45
+ import { _ as deterministicSplit, g as BENCHMARK_SPLIT_SEED, t as benchmarks_exports } from "./benchmarks-Dg0B56tK.js";
46
46
  import { t as runEvalCampaign } from "./eval-campaign-YdkpWWoT.js";
47
47
  import { n as pairedEvalueSequence, t as evaluateInterimReleaseConfidence } from "./sequential-Br0mAPHA.js";
48
48
  import { accessSync, appendFileSync, constants, cpSync, existsSync, mkdirSync, promises, readFileSync, readdirSync, statSync, writeFileSync } from "node:fs";
@@ -1,5 +1,5 @@
1
1
  import { p as CostProvenance } from "../cost-ledger-FuQvHxPm.js";
2
- import { w as JudgeScore } from "../types-Dk7PB7vh.js";
2
+ import { w as JudgeScore } from "../types-KEqL1pZc.js";
3
3
  import { o as MatrixResult } from "../index-DSC51roc.js";
4
4
  import { AgentProfile } from "@tangle-network/agent-interface";
5
5
  //#region src/multishot/types.d.ts
package/dist/openapi.json CHANGED
@@ -2,7 +2,7 @@
2
2
  "openapi": "3.1.0",
3
3
  "info": {
4
4
  "title": "@tangle-network/agent-eval — wire protocol",
5
- "version": "0.142.1",
5
+ "version": "0.143.0",
6
6
  "description": "HTTP and stdio RPC interface to agent-eval. The TypeScript runtime is the source of truth; this spec is the contract that cross-language clients (Python, Rust, Go) generate from.\n\nWire-protocol version: 1.0.0. Bumps on breaking changes to request/response schemas.",
7
7
  "contact": {
8
8
  "name": "Tangle Network",