@tangle-network/agent-eval 0.145.20 → 0.145.22
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.
- package/CHANGELOG.md +28 -0
- package/dist/analyst/index.d.ts +13 -13
- package/dist/analyst/index.js +4 -4
- package/dist/{backend-integrity-BffcHGdm.d.ts → backend-integrity-HVmWyOaX.d.ts} +2 -2
- package/dist/{backend-integrity-BffcHGdm.d.ts.map → backend-integrity-HVmWyOaX.d.ts.map} +1 -1
- package/dist/{benchmark-DbiZbPdH.d.ts → benchmark-DTQ1RApl.d.ts} +3 -3
- package/dist/{benchmark-DbiZbPdH.d.ts.map → benchmark-DTQ1RApl.d.ts.map} +1 -1
- package/dist/{benchmark-command-CkgLXHoN.js → benchmark-command-CDB_yKqS.js} +7 -7
- package/dist/{benchmark-command-CkgLXHoN.js.map → benchmark-command-CDB_yKqS.js.map} +1 -1
- package/dist/benchmarks/index.d.ts +4 -4
- package/dist/benchmarks/index.js +3 -3
- package/dist/campaign/index.d.ts +8 -8
- package/dist/campaign/index.js +5 -5
- package/dist/{campaign-Ct6r9oNe.js → campaign-CbQXcARJ.js} +6 -6
- package/dist/{campaign-Ct6r9oNe.js.map → campaign-CbQXcARJ.js.map} +1 -1
- package/dist/{capture-fetch-hykwHfDI.d.ts → capture-fetch-BZiO2cEH.d.ts} +2 -2
- package/dist/{capture-fetch-hykwHfDI.d.ts.map → capture-fetch-BZiO2cEH.d.ts.map} +1 -1
- package/dist/{chat-client-Bp4Ebuuc.js → chat-client-BJmwfnjN.js} +2 -2
- package/dist/{chat-client-Bp4Ebuuc.js.map → chat-client-BJmwfnjN.js.map} +1 -1
- package/dist/cli.js +2 -2
- package/dist/{client-D_TIV9pJ.d.ts → client-YDkU5QST.d.ts} +2 -2
- package/dist/{client-D_TIV9pJ.d.ts.map → client-YDkU5QST.d.ts.map} +1 -1
- package/dist/contract/index.d.ts +10 -10
- package/dist/contract/index.js +5 -5
- package/dist/{default-registry-cpHtP2Og.d.ts → default-registry-DONui3mA.d.ts} +6 -6
- package/dist/{default-registry-cpHtP2Og.d.ts.map → default-registry-DONui3mA.d.ts.map} +1 -1
- package/dist/{define-agent-eval-BXusiVTQ.js → define-agent-eval-BGxmy4W6.js} +3 -3
- package/dist/{define-agent-eval-BXusiVTQ.js.map → define-agent-eval-BGxmy4W6.js.map} +1 -1
- package/dist/{define-agent-eval-BgG5DIOS.d.ts → define-agent-eval-DrpgTVPp.d.ts} +5 -5
- package/dist/{define-agent-eval-BgG5DIOS.d.ts.map → define-agent-eval-DrpgTVPp.d.ts.map} +1 -1
- package/dist/{dspy-rlm-engine-CqwhQQBw.js → dspy-rlm-engine-B2nst4NR.js} +2 -2
- package/dist/{dspy-rlm-engine-CqwhQQBw.js.map → dspy-rlm-engine-B2nst4NR.js.map} +1 -1
- package/dist/{engine-DJqRKbhs.d.ts → engine-Bc3seRPe.d.ts} +4 -4
- package/dist/{engine-DJqRKbhs.d.ts.map → engine-Bc3seRPe.d.ts.map} +1 -1
- package/dist/{eval-campaign-aNqpefCS.js → eval-campaign-DeGLwACc.js} +2 -2
- package/dist/{eval-campaign-aNqpefCS.js.map → eval-campaign-DeGLwACc.js.map} +1 -1
- package/dist/{exact-types-CzbhVDr2.d.ts → exact-types-CYFbvA5U.d.ts} +2 -2
- package/dist/{exact-types-CzbhVDr2.d.ts.map → exact-types-CYFbvA5U.d.ts.map} +1 -1
- package/dist/experiment/index.d.ts +2 -2
- package/dist/{external-optimizer-contracts-lixrOZdX.d.ts → external-optimizer-contracts-CKEY98bK.d.ts} +2 -2
- package/dist/{external-optimizer-contracts-lixrOZdX.d.ts.map → external-optimizer-contracts-CKEY98bK.d.ts.map} +1 -1
- package/dist/{external-optimizer-process-D-9iX64j.js → external-optimizer-process-DYNizz8W.js} +2 -2
- package/dist/{external-optimizer-process-D-9iX64j.js.map → external-optimizer-process-DYNizz8W.js.map} +1 -1
- package/dist/{external-optimizer-subprocess-x5AbhAYX.js → external-optimizer-subprocess-r6dP9DAj.js} +2 -2
- package/dist/{external-optimizer-subprocess-x5AbhAYX.js.map → external-optimizer-subprocess-r6dP9DAj.js.map} +1 -1
- package/dist/{feedback-trajectory-D9vYSob_.d.ts → feedback-trajectory-Cg9EDxKB.d.ts} +3 -3
- package/dist/{feedback-trajectory-D9vYSob_.d.ts.map → feedback-trajectory-Cg9EDxKB.d.ts.map} +1 -1
- package/dist/hosted/index.d.ts +2 -2
- package/dist/index-BgHYx1QG.d.ts +1 -0
- package/dist/{index-YrUFx3FU.d.ts → index-CevPFhDT.d.ts} +7 -7
- package/dist/{index-YrUFx3FU.d.ts.map → index-CevPFhDT.d.ts.map} +1 -1
- package/dist/{index-CM-e2LiV.d.ts → index-DGrcBMbC.d.ts} +9 -9
- package/dist/{index-CM-e2LiV.d.ts.map → index-DGrcBMbC.d.ts.map} +1 -1
- package/dist/index.d.ts +24 -25
- package/dist/index.d.ts.map +1 -1
- package/dist/index.js +18 -19
- package/dist/index.js.map +1 -1
- package/dist/{integrity-OrcI9Nau.d.ts → integrity-DDoeWibF.d.ts} +2 -2
- package/dist/{integrity-OrcI9Nau.d.ts.map → integrity-DDoeWibF.d.ts.map} +1 -1
- package/dist/{llm-client-d0-2TT1g.js → llm-client-_UE4fU7Z.js} +45 -2
- package/dist/{llm-client-d0-2TT1g.js.map → llm-client-_UE4fU7Z.js.map} +1 -1
- package/dist/{llm-judge-DqAkIQA1.js → llm-judge-B7ffRF8w.js} +3 -3
- package/dist/{llm-judge-DqAkIQA1.js.map → llm-judge-B7ffRF8w.js.map} +1 -1
- package/dist/meta-eval/index.d.ts +1 -1
- package/dist/metrics-Cl0L1KUy.js.map +1 -1
- package/dist/multishot/index.d.ts +1 -1
- package/dist/openapi.json +1 -1
- package/dist/{produced-state-B9skdU8u.js → produced-state-D3k45S9a.js} +3 -3
- package/dist/{produced-state-B9skdU8u.js.map → produced-state-D3k45S9a.js.map} +1 -1
- package/dist/{promotion-policy-u3wj6w3U.d.ts → promotion-policy-BKi5OK5l.d.ts} +2 -2
- package/dist/{promotion-policy-u3wj6w3U.d.ts.map → promotion-policy-BKi5OK5l.d.ts.map} +1 -1
- package/dist/{provenance-AACfvbUR.d.ts → provenance-Bb7Zmyj2.d.ts} +6 -6
- package/dist/{provenance-AACfvbUR.d.ts.map → provenance-Bb7Zmyj2.d.ts.map} +1 -1
- package/dist/{registry-B_1Frl8a.d.ts → registry-CZgEmVXb.d.ts} +4 -4
- package/dist/{registry-B_1Frl8a.d.ts.map → registry-CZgEmVXb.d.ts.map} +1 -1
- package/dist/{researcher-Du-oniHp.d.ts → researcher-CC305Ed-.d.ts} +3 -3
- package/dist/{researcher-Du-oniHp.d.ts.map → researcher-CC305Ed-.d.ts.map} +1 -1
- package/dist/rl.d.ts +2 -2
- package/dist/rl.js +1 -1
- package/dist/{semantic-concept-judge-C6M-qOeb.js → semantic-concept-judge-BsY2Q0Oq.js} +2 -2
- package/dist/{semantic-concept-judge-C6M-qOeb.js.map → semantic-concept-judge-BsY2Q0Oq.js.map} +1 -1
- package/dist/{series-convergence-D1cL1f-4.d.ts → series-convergence-C9G-GNYK.d.ts} +2 -2
- package/dist/{series-convergence-D1cL1f-4.d.ts.map → series-convergence-C9G-GNYK.d.ts.map} +1 -1
- package/dist/{server-ulsOdrTI.js → server-cQXve8i4.js} +2 -2
- package/dist/{server-ulsOdrTI.js.map → server-cQXve8i4.js.map} +1 -1
- package/dist/{skillopt-optimization-method-DO6OFgcz.d.ts → skillopt-optimization-method-C9MrMgW9.d.ts} +5 -5
- package/dist/{skillopt-optimization-method-DO6OFgcz.d.ts.map → skillopt-optimization-method-C9MrMgW9.d.ts.map} +1 -1
- package/dist/{skillopt-optimization-method-CcWZBaSz.js → skillopt-optimization-method-Cv3K4DaS.js} +4 -4
- package/dist/{skillopt-optimization-method-CcWZBaSz.js.map → skillopt-optimization-method-Cv3K4DaS.js.map} +1 -1
- package/dist/{statistical-heldout-CBGPSZ5X.d.ts → statistical-heldout-fvxtHnKQ.d.ts} +2 -2
- package/dist/{statistical-heldout-CBGPSZ5X.d.ts.map → statistical-heldout-fvxtHnKQ.d.ts.map} +1 -1
- package/dist/{store-tool-spans-C9c4R7ca.d.ts → store-tool-spans-CV0hRsUp.d.ts} +4 -4
- package/dist/{store-tool-spans-C9c4R7ca.d.ts.map → store-tool-spans-CV0hRsUp.d.ts.map} +1 -1
- package/dist/{tool-groups-CZotz-e_.d.ts → tool-groups-CDDXNKhd.d.ts} +3 -3
- package/dist/tool-groups-CDDXNKhd.d.ts.map +1 -0
- package/dist/trace-repair/index.d.ts +1 -1
- package/dist/traces.d.ts +7 -7
- package/dist/{types-C4bSVIr7.d.ts → types-CWcPguwd.d.ts} +2 -2
- package/dist/{types-C4bSVIr7.d.ts.map → types-CWcPguwd.d.ts.map} +1 -1
- package/dist/{types-BjsNDR49.d.ts → types-DABZgDGV.d.ts} +3 -3
- package/dist/{types-BjsNDR49.d.ts.map → types-DABZgDGV.d.ts.map} +1 -1
- package/dist/{types-Cx3YUh2r.d.ts → types-XBRdmbj8.d.ts} +61 -2
- package/dist/types-XBRdmbj8.d.ts.map +1 -0
- package/dist/{types-vXyshMwx.d.ts → types-oAsa-fs4.d.ts} +2 -2
- package/dist/{types-vXyshMwx.d.ts.map → types-oAsa-fs4.d.ts.map} +1 -1
- package/dist/wire/index.d.ts +2 -2
- package/dist/wire/index.js +1 -1
- package/package.json +4 -3
- package/dist/index-BKShPTwZ.d.ts +0 -1
- package/dist/tool-groups-CZotz-e_.d.ts.map +0 -1
- package/dist/types-Cx3YUh2r.d.ts.map +0 -1
package/dist/cli.js
CHANGED
|
@@ -1,7 +1,7 @@
|
|
|
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-
|
|
4
|
-
import { a as runRpcBatch, o as runRpcOnce, p as handleVersion, r as startServerAsync, s as buildOpenApi } from "./server-
|
|
3
|
+
import { n as runAnalystBenchmarkCommand } from "./benchmark-command-CDB_yKqS.js";
|
|
4
|
+
import { a as runRpcBatch, o as runRpcOnce, p as handleVersion, r as startServerAsync, s as buildOpenApi } from "./server-cQXve8i4.js";
|
|
5
5
|
import { writeFileSync } from "node:fs";
|
|
6
6
|
//#region src/cli-config.ts
|
|
7
7
|
function resolveCliLlmConfig(env = process.env) {
|
|
@@ -1,5 +1,5 @@
|
|
|
1
1
|
import { l as RunTerminalOutcome } from "./run-record-DVV82Gwh.js";
|
|
2
|
-
import { _ as GateDecision, j as MutableSurface } from "./types-
|
|
2
|
+
import { _ as GateDecision, j as MutableSurface } from "./types-DABZgDGV.js";
|
|
3
3
|
import { o as InsightReport } from "./insight-report-BeT8KCgI.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-
|
|
202
|
+
//# sourceMappingURL=client-YDkU5QST.d.ts.map
|
|
@@ -1 +1 @@
|
|
|
1
|
-
{"version":3,"file":"client-
|
|
1
|
+
{"version":3,"file":"client-YDkU5QST.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"}
|
package/dist/contract/index.d.ts
CHANGED
|
@@ -1,17 +1,17 @@
|
|
|
1
1
|
import { c as CostLedgerHandle } from "../cost-ledger-DbQdN3nO.js";
|
|
2
2
|
import { a as RunRecord, n as RunCostProvenance, s as RunSplitTag } from "../run-record-DVV82Gwh.js";
|
|
3
|
-
import { p as ChatClient, v as CreateChatClientOpts, w as createChatClient } from "../types-
|
|
4
|
-
import { _ as ProposalFinding, i as AnalystFinding, v as ProposalFindingOrigin, x as makeProposalFinding } from "../types-
|
|
5
|
-
import { n as buildDefaultAnalystRegistry, t as DefaultAnalystRegistryOptions } from "../default-registry-
|
|
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, 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-
|
|
3
|
+
import { p as ChatClient, v as CreateChatClientOpts, w as createChatClient } from "../types-XBRdmbj8.js";
|
|
4
|
+
import { _ as ProposalFinding, i as AnalystFinding, v as ProposalFindingOrigin, x as makeProposalFinding } from "../types-oAsa-fs4.js";
|
|
5
|
+
import { n as buildDefaultAnalystRegistry, t as DefaultAnalystRegistryOptions } from "../default-registry-DONui3mA.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, 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-DABZgDGV.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-BeT8KCgI.js";
|
|
8
|
-
import { C as analyzeRuns, S as SummarizeExecutionOptions, b as AnalyzeRunsOptions, w as summarizeExecution, x as ExecutionReport } from "../engine-
|
|
9
|
-
import { $ as OptimizationTokenUsage, D as runEval, Dt as fsCampaignStorage, E as RunEvalOptions, G as OptimizationMethodComparison, H as CompareOptimizationMethodsOptions, J as OptimizationMethodProvenance, K as OptimizationMethodInput, Mt as LlmJudgeDimension, Nt as LlmJudgeOptions, Ot as inMemoryCampaignStorage, Pt as llmJudge, Q as OptimizationPackageSource, Tt as CampaignStorage, U as ComparisonCost, W as OptimizationMethod, Y as OptimizationMethodResult, b as RunImprovementLoopResult, bt as runCampaign, tt as compareOptimizationMethods, vt as CampaignCellFailureReceipt, x as runImprovementLoop, y as RunImprovementLoopOptions, yt as RunCampaignOptions } from "../provenance-
|
|
10
|
-
import { B as runReferenceEquivalenceJudge, C as ExternalTextOptimizerContext, E as ExternalTextEvaluationResponse, F as ReferenceEquivalenceJudgeInput, I as ReferenceEquivalenceJudgeOptions, L as ReferenceEquivalenceJudgeResult, N as REFERENCE_EQUIVALENCE_INPUT_LIMITS, P as REFERENCE_EQUIVALENCE_JUDGE_VERSION, R as ReferenceEquivalenceScenario, S as ExternalTextOptimizationMethodConfig, T as ExternalOptimizationExample, _ as DefaultProductionGateOptions, a as GepaAdaptiveEngineRun, b as composeGate, c as GepaOptimizationMethodConfig, d as gepaOptimizationMethod, f as OpenAICompatibleOptimizerModel, g as DefaultProductionGateCheck, h as heldOutGate, i as skillOptOptimizationMethod, j as campaignSplitDigest, l as GepaOptimizationRecipe, m as HeldOutGateOptions, n as SkillOptRunnerCommand, o as GepaEngineOptions, p as OptimizerModelBudget, r as SkillOptTrainerConfig, s as GepaEngineRun, t as SkillOptOptimizationMethodConfig, u as GepaRunnerCommand, v as DefaultProductionRewardHackingOptions, w as ExternalTextOptimizerResult, x as externalTextOptimizationMethod, y as defaultProductionGate, z as createReferenceEquivalenceJudge } from "../skillopt-optimization-method-
|
|
11
|
-
import { a as ObjectiveSource, c as PromotionPolicy, d as paretoSignificanceGate, i as EvidenceVector, l as buildEvidenceVector, n as AxisVerdict, o as ParetoSignificanceGateOptions, r as BuildEvidenceVectorOptions, s as PromotionObjective, t as AxisEvidence, u as paretoPolicy } from "../promotion-policy-
|
|
12
|
-
import { c as EvalRunGenerationSnapshot, g as TraceSpanEvent, n as HostedTenant, o as EvalRunCellScore, s as EvalRunEvent } from "../client-
|
|
8
|
+
import { C as analyzeRuns, S as SummarizeExecutionOptions, b as AnalyzeRunsOptions, w as summarizeExecution, x as ExecutionReport } from "../engine-Bc3seRPe.js";
|
|
9
|
+
import { $ as OptimizationTokenUsage, D as runEval, Dt as fsCampaignStorage, E as RunEvalOptions, G as OptimizationMethodComparison, H as CompareOptimizationMethodsOptions, J as OptimizationMethodProvenance, K as OptimizationMethodInput, Mt as LlmJudgeDimension, Nt as LlmJudgeOptions, Ot as inMemoryCampaignStorage, Pt as llmJudge, Q as OptimizationPackageSource, Tt as CampaignStorage, U as ComparisonCost, W as OptimizationMethod, Y as OptimizationMethodResult, b as RunImprovementLoopResult, bt as runCampaign, tt as compareOptimizationMethods, vt as CampaignCellFailureReceipt, x as runImprovementLoop, y as RunImprovementLoopOptions, yt as RunCampaignOptions } from "../provenance-Bb7Zmyj2.js";
|
|
10
|
+
import { B as runReferenceEquivalenceJudge, C as ExternalTextOptimizerContext, E as ExternalTextEvaluationResponse, F as ReferenceEquivalenceJudgeInput, I as ReferenceEquivalenceJudgeOptions, L as ReferenceEquivalenceJudgeResult, N as REFERENCE_EQUIVALENCE_INPUT_LIMITS, P as REFERENCE_EQUIVALENCE_JUDGE_VERSION, R as ReferenceEquivalenceScenario, S as ExternalTextOptimizationMethodConfig, T as ExternalOptimizationExample, _ as DefaultProductionGateOptions, a as GepaAdaptiveEngineRun, b as composeGate, c as GepaOptimizationMethodConfig, d as gepaOptimizationMethod, f as OpenAICompatibleOptimizerModel, g as DefaultProductionGateCheck, h as heldOutGate, i as skillOptOptimizationMethod, j as campaignSplitDigest, l as GepaOptimizationRecipe, m as HeldOutGateOptions, n as SkillOptRunnerCommand, o as GepaEngineOptions, p as OptimizerModelBudget, r as SkillOptTrainerConfig, s as GepaEngineRun, t as SkillOptOptimizationMethodConfig, u as GepaRunnerCommand, v as DefaultProductionRewardHackingOptions, w as ExternalTextOptimizerResult, x as externalTextOptimizationMethod, y as defaultProductionGate, z as createReferenceEquivalenceJudge } from "../skillopt-optimization-method-C9MrMgW9.js";
|
|
11
|
+
import { a as ObjectiveSource, c as PromotionPolicy, d as paretoSignificanceGate, i as EvidenceVector, l as buildEvidenceVector, n as AxisVerdict, o as ParetoSignificanceGateOptions, r as BuildEvidenceVectorOptions, s as PromotionObjective, t as AxisEvidence, u as paretoPolicy } from "../promotion-policy-BKi5OK5l.js";
|
|
12
|
+
import { c as EvalRunGenerationSnapshot, g as TraceSpanEvent, n as HostedTenant, o as EvalRunCellScore, s as EvalRunEvent } from "../client-YDkU5QST.js";
|
|
13
13
|
import { i as InMemoryOutcomeStore, n as FileSystemOutcomeStore, o as OutcomeStore, r as FileSystemOutcomeStoreOptions, t as DeploymentOutcome } from "../outcome-store-BYHIuO0e.js";
|
|
14
|
-
import { a as DefinedAgentEval, c as SelfImproveOptions, d as SelfImproveRunError, f as selfImprove, i as DefineAgentEvalOptions, l as SelfImproveProgressEvent, n as AgentEvalEvaluateOptions, o as defineAgentEval, r as AgentEvalImproveOptions, s as SelfImproveBudget, t as AgentEvalAgent, u as SelfImproveResult } from "../define-agent-eval-
|
|
14
|
+
import { a as DefinedAgentEval, c as SelfImproveOptions, d as SelfImproveRunError, f as selfImprove, i as DefineAgentEvalOptions, l as SelfImproveProgressEvent, n as AgentEvalEvaluateOptions, o as defineAgentEval, r as AgentEvalImproveOptions, s as SelfImproveBudget, t as AgentEvalAgent, u as SelfImproveResult } from "../define-agent-eval-DrpgTVPp.js";
|
|
15
15
|
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";
|
|
16
16
|
//#region src/contract/measured-comparison.d.ts
|
|
17
17
|
interface SealCandidateBenchmarkSuiteOptions {
|
package/dist/contract/index.js
CHANGED
|
@@ -1,21 +1,21 @@
|
|
|
1
1
|
import { s as ValidationError } from "../errors-Dngq5h35.js";
|
|
2
2
|
import { r as pairedBootstrap } from "../paired-tests-BHIhYVdu.js";
|
|
3
|
-
import { a as summarizeExecution, i as analyzeRuns, n as SelfImproveRunError, r as selfImprove, t as defineAgentEval } from "../define-agent-eval-
|
|
3
|
+
import { a as summarizeExecution, i as analyzeRuns, n as SelfImproveRunError, r as selfImprove, t as defineAgentEval } from "../define-agent-eval-BGxmy4W6.js";
|
|
4
4
|
import { a as parseRunRecordSafe } from "../run-record-D2lDdSAz.js";
|
|
5
|
-
import { G as runEval, K as runCampaign, f as runImprovementLoop, nt as campaignSplitDigest, t as llmJudge, x as compareOptimizationMethods, z as defaultProductionGate } from "../llm-judge-
|
|
5
|
+
import { G as runEval, K as runCampaign, f as runImprovementLoop, nt as campaignSplitDigest, t as llmJudge, x as compareOptimizationMethods, z as defaultProductionGate } from "../llm-judge-B7ffRF8w.js";
|
|
6
6
|
import { i as isModelPriced, n as estimateCost } from "../metrics-Cl0L1KUy.js";
|
|
7
7
|
import { i as CostLedger } from "../cost-ledger-BSe92yAV.js";
|
|
8
8
|
import { d as mapConcurrentRange } from "../ledger-core-DTae9rv_.js";
|
|
9
|
-
import { S as inMemoryCampaignStorage, x as fsCampaignStorage } from "../external-optimizer-subprocess-
|
|
9
|
+
import { S as inMemoryCampaignStorage, x as fsCampaignStorage } from "../external-optimizer-subprocess-r6dP9DAj.js";
|
|
10
10
|
import { a as heldoutSignificance, s as decidePairedPromotion } from "../power-preflight-DEw-uC7q.js";
|
|
11
11
|
import { r as makeProposalFinding } from "../types-BI4fT3HN.js";
|
|
12
12
|
import { $ as classifyOtlpSpanRole, et as isOtlpModelCall } from "../kind-factory-CPmSd58s.js";
|
|
13
|
-
import { n as buildDefaultAnalystRegistry, t as createChatClient } from "../chat-client-
|
|
13
|
+
import { n as buildDefaultAnalystRegistry, t as createChatClient } from "../chat-client-BJmwfnjN.js";
|
|
14
14
|
import { LLM_MODEL_ATTR_KEYS, SPAN_KIND_ATTR_KEYS } from "../trace-attributes.js";
|
|
15
15
|
import { t as extractUsage } from "../extract-usage-BrQ8mCLX.js";
|
|
16
16
|
import { n as InMemoryOutcomeStore, t as FileSystemOutcomeStore } from "../outcome-store-ChBKlTd_.js";
|
|
17
17
|
import { i as summarizeTraceErrors, n as recordAggregateMeasurements, r as summarizeExecutionMeasurements, t as readTaskFailureLabels } from "../task-failure-attributes-CpQ4y5RD.js";
|
|
18
|
-
import { a as externalTextOptimizationMethod, c as REFERENCE_EQUIVALENCE_INPUT_LIMITS, d as runReferenceEquivalenceJudge, i as composeGate, l as REFERENCE_EQUIVALENCE_JUDGE_VERSION, n as gepaOptimizationMethod, r as heldOutGate, t as skillOptOptimizationMethod, u as createReferenceEquivalenceJudge } from "../skillopt-optimization-method-
|
|
18
|
+
import { a as externalTextOptimizationMethod, c as REFERENCE_EQUIVALENCE_INPUT_LIMITS, d as runReferenceEquivalenceJudge, i as composeGate, l as REFERENCE_EQUIVALENCE_JUDGE_VERSION, n as gepaOptimizationMethod, r as heldOutGate, t as skillOptOptimizationMethod, u as createReferenceEquivalenceJudge } from "../skillopt-optimization-method-Cv3K4DaS.js";
|
|
19
19
|
import { n as paretoPolicy, r as paretoSignificanceGate, t as buildEvidenceVector } from "../promotion-policy-xzA40Evo.js";
|
|
20
20
|
import { createHash } from "node:crypto";
|
|
21
21
|
import { agentCandidateBenchmarkSuiteSchema, agentCandidateBenchmarkTaskSchema, agentCandidateBundleSchema, agentCandidateEvaluationPolicySchema, agentCandidateExperimentSchema, agentImprovementMeasuredComparisonSchema, agentProfileImprovementExperimentSchema, agentProfileImprovementMeasuredComparisonSchema, agentProfileImprovementRunCellSchema, agentProfileImprovementRunReceiptSchema, agentProfileImprovementSuiteInputsSchema, agentProfileImprovementSuiteSchema, agentProfileImprovementTaskSchema, candidateExecutionEvidenceSchema, canonicalCandidateDigest, canonicalCandidateJson, numbersApproximatelyEqual, omitTopLevelDigest } from "@tangle-network/agent-interface";
|
|
@@ -1,8 +1,8 @@
|
|
|
1
|
-
import { i as AnalystFinding, n as AnalystContext, w as TraceAnalysisStore } from "./types-
|
|
2
|
-
import { l as ExactCapableAnalyst } from "./exact-types-
|
|
3
|
-
import { n as AnalystRegistry, r as AnalystRegistryOptions } from "./registry-
|
|
4
|
-
import { a as TraceAnalystLimits, d as RawAnalystFinding, i as TraceAnalysisEngineResult, n as TraceAnalysisEngine } from "./engine-
|
|
5
|
-
import { t as TraceToolGroupName } from "./tool-groups-
|
|
1
|
+
import { i as AnalystFinding, n as AnalystContext, w as TraceAnalysisStore } from "./types-oAsa-fs4.js";
|
|
2
|
+
import { l as ExactCapableAnalyst } from "./exact-types-CYFbvA5U.js";
|
|
3
|
+
import { n as AnalystRegistry, r as AnalystRegistryOptions } from "./registry-CZgEmVXb.js";
|
|
4
|
+
import { a as TraceAnalystLimits, d as RawAnalystFinding, i as TraceAnalysisEngineResult, n as TraceAnalysisEngine } from "./engine-Bc3seRPe.js";
|
|
5
|
+
import { t as TraceToolGroupName } from "./tool-groups-CDDXNKhd.js";
|
|
6
6
|
//#region src/trace-analyst/behavioral-metrics.d.ts
|
|
7
7
|
type SuboptimalCode = 'monotonic-input-growth' | 'output-length-decay' | 'single-tool-dependency' | 'no-self-verification';
|
|
8
8
|
interface SuboptimalSignal {
|
|
@@ -111,4 +111,4 @@ interface DefaultAnalystRegistryOptions {
|
|
|
111
111
|
declare function buildDefaultAnalystRegistry(options?: DefaultAnalystRegistryOptions): AnalystRegistry;
|
|
112
112
|
//#endregion
|
|
113
113
|
export { createTraceAnalyst as a, runTraceAnalyst as c, deriveEfficiencyFindings as d, TraceAnalystDefinition as i, BehavioralAnalystOptions as l, buildDefaultAnalystRegistry as n, renderPriorFindings as o, CreateTraceAnalystOptions as r, renderUpstreamFindings as s, DefaultAnalystRegistryOptions as t, behavioralAnalyst as u };
|
|
114
|
-
//# sourceMappingURL=default-registry-
|
|
114
|
+
//# sourceMappingURL=default-registry-DONui3mA.d.ts.map
|
|
@@ -1 +1 @@
|
|
|
1
|
-
{"version":3,"file":"default-registry-
|
|
1
|
+
{"version":3,"file":"default-registry-DONui3mA.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;;;;UC1BR;;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"}
|
|
@@ -9,9 +9,9 @@ import { r as welchsTTest } from "./baseline-BhPRQBVn.js";
|
|
|
9
9
|
import { r as observedSplitScore } from "./reward-nw2xZGZG.js";
|
|
10
10
|
import { c as validateRunRecord, i as modelHasSnapshot } from "./run-record-D2lDdSAz.js";
|
|
11
11
|
import { r as paretoChart } from "./summary-report-Blysd6Z2.js";
|
|
12
|
-
import { A as surfaceHash, G as runEval, J as resolveRunDir, f as runImprovementLoop, k as surfaceContentHash, o as emitLoopProvenance, s as loopProvenanceArgsFromResult, y as assertOptimizationResult, z as defaultProductionGate } from "./llm-judge-
|
|
12
|
+
import { A as surfaceHash, G as runEval, J as resolveRunDir, f as runImprovementLoop, k as surfaceContentHash, o as emitLoopProvenance, s as loopProvenanceArgsFromResult, y as assertOptimizationResult, z as defaultProductionGate } from "./llm-judge-B7ffRF8w.js";
|
|
13
13
|
import { c as campaignCellTaskScore, l as campaignCellToRunRecord, o as campaignCellExecutionEvidence, s as campaignCellJudgeDimensions } from "./reward-hacking-DFo2FU5J.js";
|
|
14
|
-
import { S as inMemoryCampaignStorage, b as createRunCostLedger, x as fsCampaignStorage } from "./external-optimizer-subprocess-
|
|
14
|
+
import { S as inMemoryCampaignStorage, b as createRunCostLedger, x as fsCampaignStorage } from "./external-optimizer-subprocess-r6dP9DAj.js";
|
|
15
15
|
import { t as powerPreflight } from "./power-preflight-DEw-uC7q.js";
|
|
16
16
|
import { t as createHostedClient } from "./client-LIuo-KPv.js";
|
|
17
17
|
//#region src/contamination-guard.ts
|
|
@@ -1491,4 +1491,4 @@ function requirePositiveInteger(value, field) {
|
|
|
1491
1491
|
//#endregion
|
|
1492
1492
|
export { summarizeExecution as a, analyzeRuns as i, SelfImproveRunError as n, checkCanaries as o, selfImprove as r, defineAgentEval as t };
|
|
1493
1493
|
|
|
1494
|
-
//# sourceMappingURL=define-agent-eval-
|
|
1494
|
+
//# sourceMappingURL=define-agent-eval-BGxmy4W6.js.map
|