@tangle-network/agent-eval 0.145.4 → 0.145.5
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 +9 -0
- package/README.md +5 -0
- package/dist/analyst/index.d.ts +1 -1
- package/dist/analyst/index.js +1 -1
- package/dist/{benchmark-command-BANo4dTm.js → benchmark-command-C44VE5rA.js} +2 -2
- package/dist/{benchmark-command-BANo4dTm.js.map → benchmark-command-C44VE5rA.js.map} +1 -1
- package/dist/benchmarks/index.d.ts +1 -1
- package/dist/benchmarks/index.js +1 -1
- package/dist/campaign/index.d.ts +2 -2
- package/dist/campaign/index.js +2 -2
- package/dist/{campaign-FY23WYzZ.js → campaign-DGWSftLe.js} +2 -2
- package/dist/{campaign-FY23WYzZ.js.map → campaign-DGWSftLe.js.map} +1 -1
- package/dist/cli.js +1 -1
- package/dist/contract/index.d.ts +1 -1
- package/dist/contract/index.js +1 -1
- package/dist/{index-Dytg99g1.d.ts → index-DuKK7tJR.d.ts} +2 -2
- package/dist/{index-Dytg99g1.d.ts.map → index-DuKK7tJR.d.ts.map} +1 -1
- package/dist/openapi.json +1 -1
- package/dist/{skillopt-optimization-method-CkySQ0ca.d.ts → skillopt-optimization-method-Bb1OTHZW.d.ts} +8 -1
- package/dist/skillopt-optimization-method-Bb1OTHZW.d.ts.map +1 -0
- package/dist/{skillopt-optimization-method-CxrOjp1S.js → skillopt-optimization-method-S7sdjzvh.js} +25 -2
- package/dist/skillopt-optimization-method-S7sdjzvh.js.map +1 -0
- package/package.json +1 -1
- package/dist/skillopt-optimization-method-CkySQ0ca.d.ts.map +0 -1
- package/dist/skillopt-optimization-method-CxrOjp1S.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-
|
|
3
|
+
import { n as runAnalystBenchmarkCommand } from "./benchmark-command-C44VE5rA.js";
|
|
4
4
|
import { a as runRpcBatch, o as runRpcOnce, p as handleVersion, r as startServerAsync, s as buildOpenApi } from "./server-ulsOdrTI.js";
|
|
5
5
|
import { writeFileSync } from "node:fs";
|
|
6
6
|
//#region src/cli-config.ts
|
package/dist/contract/index.d.ts
CHANGED
|
@@ -7,7 +7,7 @@ import { C as JudgeDimension, H as SurfaceProposer, M as OptimizerConfig, R as S
|
|
|
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-CRi-Ufrj.js";
|
|
8
8
|
import { C as analyzeRuns, S as SummarizeExecutionOptions, b as AnalyzeRunsOptions, w as summarizeExecution, x as ExecutionReport } from "../engine-otFpE2gF.js";
|
|
9
9
|
import { $ as OptimizationTokenUsage, At as llmJudge, Ct as fsCampaignStorage, D as runEval, E as RunEvalOptions, G as OptimizationMethodComparison, H as CompareOptimizationMethodsOptions, J as OptimizationMethodProvenance, K as OptimizationMethodInput, Ot as LlmJudgeDimension, Q as OptimizationPackageSource, U as ComparisonCost, W as OptimizationMethod, Y as OptimizationMethodResult, at as RunCampaignOptions, b as RunImprovementLoopResult, it as CampaignCellFailureReceipt, kt as LlmJudgeOptions, ot as runCampaign, tt as compareOptimizationMethods, wt as inMemoryCampaignStorage, x as runImprovementLoop, xt as CampaignStorage, y as RunImprovementLoopOptions } from "../provenance-C_ln89b0.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-
|
|
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-Bb1OTHZW.js";
|
|
11
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-D0nPhkSy.js";
|
|
12
12
|
import { c as EvalRunGenerationSnapshot, g as TraceSpanEvent, n as HostedTenant, o as EvalRunCellScore, s as EvalRunEvent } from "../client-KUbGulm_.js";
|
|
13
13
|
import { i as InMemoryOutcomeStore, n as FileSystemOutcomeStore, o as OutcomeStore, r as FileSystemOutcomeStoreOptions, t as DeploymentOutcome } from "../outcome-store-BYHIuO0e.js";
|
package/dist/contract/index.js
CHANGED
|
@@ -15,7 +15,7 @@ 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-S7sdjzvh.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";
|
|
@@ -9,7 +9,7 @@ import { d as CompletionVerdict, f as CorrectnessChecker, h as ProducedState, n
|
|
|
9
9
|
import { g as PairedArmsComparison } from "./pre-registration-DeRvl9sE.js";
|
|
10
10
|
import { d as PairedBootstrapResult } from "./paired-promotion-decision-CGzg0cI_.js";
|
|
11
11
|
import { lt as PlanCampaignRunOptions, st as CampaignRunPlan, xt as CampaignStorage } from "./provenance-C_ln89b0.js";
|
|
12
|
-
import "./skillopt-optimization-method-
|
|
12
|
+
import "./skillopt-optimization-method-Bb1OTHZW.js";
|
|
13
13
|
import "./promotion-policy-D0nPhkSy.js";
|
|
14
14
|
import "./statistical-heldout-DC7cHtFD.js";
|
|
15
15
|
import { D as LedgerHash, m as LedgerTrustedHeadRemoval, p as LedgerTrustedHead } from "./index-B8Ui1mr1.js";
|
|
@@ -1511,4 +1511,4 @@ declare function verifyCodeSurface(surface: CodeSurface, worktreeDir?: string):
|
|
|
1511
1511
|
declare function resolveWorktreePath(surface: CodeSurface, worktreeDir?: string): string;
|
|
1512
1512
|
//#endregion
|
|
1513
1513
|
export { SearchPlannedEvent as $, CrossSurfacePairEvidence as $n, RunProfileMatrixResult as $t, SearchAccountingAudit as A, CrossSurfaceAdditionRejectionReason as An, ProfileMatrixSegmentResult as At, SearchCostAccounting as B, CrossSurfaceComponentEvidence as Bn, ScoreboardRow as Bt, surfaceHash as C, discoverEvalFixtures as Cn, tangleTracesRoot as Ct, FileSearchLedger as D, analyzeCrossSurfaceInteractions as Dn, ProfileMatrixCoverage as Dt, acquireSingleRunLock as E, planEvalFixtureRun as En, FinalizedProfileMatrixResult as Et, SearchCandidateRegisteredEvent as F, CrossSurfaceCandidateComparison as Fn, runProfileMatrixSegment as Ft, SearchLedgerEvent as G, CrossSurfaceIneligibilityReason as Gn, renderScoreboardMarkdown as Gt, SearchLedger as H, CrossSurfaceDistribution as Hn, UserStory as Ht, SearchCandidateSlot as I, CrossSurfaceCandidateEvidence as In, PlaybackContext as It, SearchLedgerTrustedHeadMode as J, CrossSurfaceInteractionPath as Jn, userStoryScoreboard as Jt, SearchLedgerHash as K, CrossSurfaceInteractionAwareSelection as Kn, scoreUserStory as Kt, SearchCandidateSlotClosedEvent as L, CrossSurfaceCandidateOutcome as Ln, PlaybackDriver as Lt, SearchAttemptAccounting as M, CrossSurfaceBestSingleSelection as Mn, SegmentedProfileMatrixError as Mt, SearchCandidateDecidedEvent as N, CrossSurfaceBootstrapPolicy as Nn, createProfileMatrixPlan as Nt, OpenSearchLedgerOptions as O, AnalyzeCrossSurfaceInteractionsInput as On, ProfileMatrixPlan as Ot, SearchCandidateLineage as P, CrossSurfaceCandidate as Pn, finalizeProfileMatrix as Pt, SearchPlan as Q, CrossSurfacePairCompatibility as Qn, RunProfileMatrixOptions as Qt, SearchCandidateSurface as R, CrossSurfaceCandidateSummary as Rn, PlaybackStep as Rt, surfaceContentHash as S, PlanEvalFixtureRunOptions as Sn, resolveRunDir as St, SingleRunLockOptions as T, loadEvalFixtureScenarios as Tn, FinalizeProfileMatrixOptions as Tt, SearchLedgerAppendResult as U, CrossSurfaceEligibility as Un, UserStoryVerdict as Ut, SearchFailureReason as V, CrossSurfaceCompositionStep as Vn, ScoreboardSummary as Vt, SearchLedgerEntry as W, CrossSurfaceEvidenceBreakdown as Wn, makePlaybackDispatch as Wt, SearchOperationKind as X, CrossSurfaceInteractionTask as Xn, ProfileMatrixError as Xt, SearchModelIdentity as Y, CrossSurfaceInteractionReport as Yn, ProfileDispatchFn as Yt, SearchOperationRecordedEvent as Z, CrossSurfaceNaiveStackSelection as Zn, ProfileSummary as Zt, assertCodeSurfaceIdentity as _, EvalFixtureLoadOptions as _n, compareRankKeys as _t, WorktreeAdapterError as a, LabeledScenarioStoreError as an, CrossSurfaceSelections as ar, SearchSurfaceKind as at, componentSurfaceIdentityMaterial as b, EvalFixtureValidationMode as bn, scoreDiscrimination as bt, verifyCodeSurface as c, RolloutCall as cn, TraceAnalystArtifact as cr, SearchTokenAccounting as ct, PhoenixEvaluationResultLike as d, classifyUngroundedLiterals as dn, traceAnalystQualityJudge as dr, SearchLedgerConflictError as dt, ScenarioRollup as en, CrossSurfacePairIncompatibilityReason as er, SearchPlannedOperation as et, PhoenixEvaluatorLike as f, rolloutArgumentDiff as fn, SearchLedgerError as ft, isTransientTransportFailure as g, EvalFixtureFile as gn, campaignMeanComposite as gt, TransientFailureOptions as h, EvalFixture as hn, campaignBreakdown as ht, WorktreeAdapter as i, FsLabeledScenarioStoreOptions as in, CrossSurfaceSelectionPolicy as ir, SearchSurfaceEvidence as it, SearchArtifactRef as j, CrossSurfaceAttemptCompleteness as jn, RunProfileMatrixSegmentOptions as jt, SEARCH_LEDGER_SCHEMA as k, CrossSurfaceAdditionDecision as kn, ProfileMatrixRow as kt, AutoevalsScoreLike as l, ScoredRollout as ln, TraceAnalystScenario as lr, openSearchLedger as lt, phoenixEvaluatorJudge as m, neutralizationGate as mn, CampaignBreakdown as mt, GitWorktreeAdapterOptions as n, neutralizeText as nn, CrossSurfaceRankedSingle as nr, SearchSourceRef as nt, gitWorktreeAdapter as o, RolloutArgumentDiff as on, CrossSurfaceTaskRow as or, SearchTaskAttemptedEvent as ot, autoevalsScorerJudge as p, NeutralizationGateOptions as pn, SearchLedgerIntegrityError as pt, SearchLedgerReplay as q, CrossSurfaceInteractionEffect as qn, scoreboardSummary as qt, Worktree as r, FsLabeledScenarioStore as rn, CrossSurfaceRelativeCost as rr, SearchSurfaceEffect as rt, resolveWorktreePath as s, RolloutArgumentDiffOptions as sn, BuildTraceAnalystSurfaceDispatchOptions as sr, SearchTaskOutcome as st, CodeSurfaceVerification as t, runProfileMatrix as tn, CrossSurfacePairwiseEntry as tr, SearchPlannedTask as tt, AutoevalsScorerLike as u, UngroundedLiteralReport as un, buildTraceAnalystSurfaceDispatch as ur, validateSearchLedgerEvent as ut, assertComponentSurface as v, EvalFixtureRunPlan as vn, DiscriminationScore as vt, SingleRunLock as w, loadEvalFixture as wn, CreateProfileMatrixPlanOptions as wt, renderSurfaceDiff as x, LoadEvalFixtureScenariosOptions as xn, selectDiscriminative as xt, codeSurfaceIdentityMaterial as y, EvalFixtureScenario as yn, ScenarioSignal as yt, SearchCompletedEvent as z, CrossSurfaceComponent as zn, ScoreboardRenderOptions as zt };
|
|
1514
|
-
//# sourceMappingURL=index-
|
|
1514
|
+
//# sourceMappingURL=index-DuKK7tJR.d.ts.map
|
|
@@ -1 +1 @@
|
|
|
1
|
-
{"version":3,"file":"index-
|
|
1
|
+
{"version":3,"file":"index-DuKK7tJR.d.ts","names":[],"sources":["../src/campaign/analyst-surface.ts","../src/campaign/cross-surface-types.ts","../src/campaign/cross-surface-interaction.ts","../src/campaign/fixtures.ts","../src/campaign/gates/neutralization-gate.ts","../src/campaign/grounded-reflection.ts","../src/campaign/labeled-store/fs-adapter.ts","../src/campaign/neutralize.ts","../src/campaign/presets/run-profile-matrix.ts","../src/campaign/presets/playback.ts","../src/campaign/presets/segmented-profile-matrix.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;;;;KCtDU;;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;;;;;;;;;;;;;;;;;;;;;;;;;;;;;UC9BF;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;;;;;;cC6ClB,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;;;EAGA;;;;EAIA;;EAEA,WAAW;;EAEX;;EAEA;;;;;;;;EAQA;;;EAGA;;EAEA;;;EAGA;;EAEA;;EAEA,eAAe;EACf,gBAAgB;;;EAGhB,UAAU;;EAEV,YAAY;;;EAGZ,aAAa,UAAU;;;;EAIvB;;;;;;;EAOA,cACE,UAAU,WACV,UAAU;IACL;IAAgB;;;;;;;EAMvB,aAAa;IAAS,SAAS;IAAc,UAAU;IAAW;;;EAElE;;EAEA;;UAGe;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;;;;;iBA6LhC,iBAAiB,kBAAkB,UAAU,WACjE,MAAM,wBAAwB,WAAW,aACxC,QAAQ,uBAAuB,WAAW;;;;;UCtX5B;;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;;;cC1KK,oCAAoC;EAC/C,YAAY;;UAKG;EACf;EACA;EACA;EACA;EACA;;UAGe,kBAAkB,kBAAkB,UAAU;WACpD;WACA;WACA;WACA;WACA,mBAAmB;WACnB,oBAAoB;WACpB,iBAAiB,YAAY,WAAW;WACxC;WACA;WACA,UAAU,YAAY,wBAAwB,WAAW;WACzD;;WAEA;WACA,WAAW,YAAY,wBAAwB,WAAW;WAC1D;WACA;WACA,aAAa,UAAU;WACvB,cACP,UAAU,WACV,UAAU;IACL;IAAgB;;WACd,eAAe;;UAGT,+BAA+B,kBAAkB,UAAU,mBAClE,KACN,wBAAwB,WAAW;;EAgBrC;;UAGe;EACf;EACA;EACA;EACA;EACA;;UAGe,+BAA+B,kBAAkB,UAAU;EAC1E,MAAM,kBAAkB,WAAW;;EAEnC;;EAEA,yBAAyB;EACzB,UAAU,kBAAkB,WAAW;EACvC;EACA,UAAU;EACV;EACA;EACA;EACA,eAAe,wBAAwB,WAAW;EAClD,gBAAgB,wBAAwB,WAAW;EACnD,YAAY;;UAGG,2BAA2B,WAAW,kBAAkB;EACvE;EACA;EACA,QAAQ,uBAAuB,WAAW;EAC1C,UAAU;;UAGK,6BAA6B,kBAAkB,UAAU;EACxE,MAAM,kBAAkB,WAAW;EACnC;EACA,UAAU;EACV;EACA;EACA,YAAY;;UAGG,6BAA6B,WAAW,kBAAkB,kBACjE,uBAAuB,WAAW;EAC1C,UAAU;;iBAGI,wBAAwB,kBAAkB,UAAU,WAClE,MAAM,+BAA+B,WAAW,aAC/C,kBAAkB,WAAW;iBA0HV,wBAAwB,kBAAkB,UAAU,WACxE,MAAM,+BAA+B,WAAW,aAC/C,QAAQ,2BAA2B,WAAW;iBAkD3B,sBAAsB,kBAAkB,UAAU,WACtE,MAAM,6BAA6B,WAAW,aAC7C,QAAQ,6BAA6B,WAAW;;;;;;;;iBCrTnC;;;;;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/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": "0.145.
|
|
5
|
+
"version": "0.145.5",
|
|
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",
|
|
@@ -66,6 +66,13 @@ interface ExternalTextEvaluationResponse {
|
|
|
66
66
|
dimensions: Record<string, number>;
|
|
67
67
|
notes?: string;
|
|
68
68
|
artifact?: unknown;
|
|
69
|
+
/** The optimizer score is a finite protocol value; Eval remains the
|
|
70
|
+
* authority for whether the underlying cell actually produced a score. */
|
|
71
|
+
status?: 'scored' | 'failed';
|
|
72
|
+
error?: {
|
|
73
|
+
stage: 'dispatch' | 'judge' | 'unknown';
|
|
74
|
+
message: string;
|
|
75
|
+
};
|
|
69
76
|
};
|
|
70
77
|
}
|
|
71
78
|
declare function decodeExternalTextCandidate(candidate: ExternalTextCandidate): MutableSurface;
|
|
@@ -424,4 +431,4 @@ interface SkillOptOptimizationMethodConfig<TScenario extends Scenario, TArtifact
|
|
|
424
431
|
declare function skillOptOptimizationMethod<TScenario extends Scenario, TArtifact>(config: SkillOptOptimizationMethodConfig<TScenario, TArtifact>): OptimizationMethod<TScenario, TArtifact>;
|
|
425
432
|
//#endregion
|
|
426
433
|
export { campaignScenarioIdentity as A, runReferenceEquivalenceJudge as B, ExternalTextOptimizerContext as C, decodeExternalTextCandidate as D, ExternalTextEvaluationResponse as E, ReferenceEquivalenceJudgeInput as F, ReferenceEquivalenceJudgeOptions as I, ReferenceEquivalenceJudgeResult as L, campaignSplitDigestFromIdentities as M, REFERENCE_EQUIVALENCE_INPUT_LIMITS as N, assertCampaignDesign as O, REFERENCE_EQUIVALENCE_JUDGE_VERSION as P, ReferenceEquivalenceScenario as R, ExternalTextOptimizationMethodConfig as S, ExternalOptimizationExample as T, DefaultProductionGateOptions as _, GepaAdaptiveEngineRun as a, composeGate as b, GepaOptimizationMethodConfig as c, gepaOptimizationMethod as d, OpenAICompatibleOptimizerModel as f, DefaultProductionGateCheck as g, heldOutGate as h, skillOptOptimizationMethod as i, campaignSplitDigest as j, assertCampaignSplitIdentity as k, GepaOptimizationRecipe as l, HeldOutGateOptions as m, SkillOptRunnerCommand as n, GepaEngineOptions as o, OptimizerModelBudget as p, SkillOptTrainerConfig as r, GepaEngineRun as s, SkillOptOptimizationMethodConfig as t, GepaRunnerCommand as u, DefaultProductionRewardHackingOptions as v, ExternalTextOptimizerResult as w, externalTextOptimizationMethod as x, defaultProductionGate as y, createReferenceEquivalenceJudge as z };
|
|
427
|
-
//# sourceMappingURL=skillopt-optimization-method-
|
|
434
|
+
//# sourceMappingURL=skillopt-optimization-method-Bb1OTHZW.d.ts.map
|
|
@@ -0,0 +1 @@
|
|
|
1
|
+
{"version":3,"file":"skillopt-optimization-method-Bb1OTHZW.d.ts","names":[],"sources":["../src/reference-equivalence-judge.ts","../src/campaign/coverage.ts","../src/campaign/external-text-evaluation.ts","../src/campaign/external-text-optimization-contract.ts","../src/campaign/external-text-optimization.ts","../src/campaign/gates/compose.ts","../src/campaign/gates/default-production-gate.ts","../src/campaign/gates/heldout-gate.ts","../src/campaign/optimizer-model.ts","../src/campaign/gepa-optimization-method.ts","../src/campaign/skillopt-optimization-method.ts"],"mappings":";;;;;;;;cAOa;cAEA;WACX;WACA;WACA;;UAGe,qCAAqC;EACpD;EACA;;UAGe;EACf;EACA;EACA;;UAGe;;EAEf,MAAM;;EAEN;;EAEA,SAAS;;EAET,aAAa;;UAGE,wCAAwC;EACvD;EACA;EACA;EACA;;;iBA0Bc,gCACd,SAAS,mCACR,oBAAoB;;iBA2CD,6BACpB,OAAO,gCACP,SAAS,mCACR,QAAQ;;;;iBCxGK,qBAAqB,kBAAkB,UACrD,oBAAoB,aACpB;;iBA2Bc,yBAAyB,kBAAkB,UACzD,UAAU,YACT,2BAA2B,KAAK;;iBAUnB,kCACd,oBAAoB,4BACpB;;iBAgBc,oBAAoB,kBAAkB,UACpD,oBAAoB,aACpB;;iBAOc,4BACd,oBAAoB,4BACpB,cACA;;;UCpEe;EACf;EACA;;UAGe;EACf;EACA;IACE;IACA,YAAY;IACZ;IACA;;;IAGA;IACA;MACE;MACA;;;;iBAwHU,4BAA4B,WAAW,wBAAwB;;;UC3I9D;WACN;WACA;WACA;WACA;WACA;WACA,eAAe;WACf,mBAAmB;WACnB,uBAAuB;WACvB;WACA;;WAEA;WACA;WACA;WACA,QAAQ;;WAER,MAAM;WACN,WACP,SAFa,kCAGV,QAAQ;;UAGE;EACf,eAAe;EACf;EACA;IACM;;IACA;;IACA;IAAkB;;;;;;;;;UAqBT,qCACf,kBAAkB,UAClB;EAEA;EACA,QAAQ,KAAK;EACb;EACA;EACA;EACA;;EAEA;;EAEA;;EAEA,SAAS;EACT;EACA;EACA,oBAAoB,UAAU;EAC9B,oBAAoB,UAAU,WAAW,UAAU;EACnD,MAAM,SAAS,iCAAiC,QAAQ;;;;;;;;;;iBCN1C,+BAA+B,kBAAkB,UAAU,WACzE,QAAQ,qCAAqC,WAAW,aACvD,mBAAmB,WAAW;;;;;;iBClEjB,YAAY,qBAAqB,kBAAkB,WAAW,aACzE,OAAO,MAAM,KAAK,WAAW,cAC/B,KAAK,WAAW;;;KCaP;KAOA,wCAAwC,KAClD;EAGA,SAAS,YAAY;;UAGN;;;EAGf,kBAAkB;;;;;EAKlB;;EAEA;;EAEA;;EAEA;;;;EAIA;;;;EAIA;;;;;EAKA;;;;EAIA;;;EAGA;;;;EAIA,iBAAiB;;;EAGjB,aAAa;;EAEb,gBAAgB;;EAEhB,SAAS;;;;;EAKT,iBAAiB;;;;;iBAMH,sBAAsB,WAAW,kBAAkB,UACjE,SAAS,+BACR,KAAK,WAAW;;;UC/EF,mBAAmB,kBAAkB,WAAW;EAC/D,WAAW;;;EAGX;;EAEA;;;;EAIA;;EAEA;;EAEA;;;;;;;;iBASc,YAAY,WAAW,kBAAkB,UACvD,SAAS,mBAAmB,aAC3B,KAAK,WAAW;;;KCvCP,uBAAuB;;UAGlB;EACf;EACA,QAAQ;;EAER,MAAM;;EAEN;;;;;UCmDe;;EAEf;;EAEA;;EAEA;;EAEA;;EAEA;;;;;EAKA,eAAe;;;UAIA,sBAAsB;;EAErC;;;KAIU,wBAAwB;;;;;;;KAQxB;EAEN;EACA,KAAK;;EAGL;EACA,eAAe;;EAGf;EACA,eAAe;;EAEf;;EAEA;EACA;EACA;EACA;EACA;EACA;EACA;;EAGA;EACA,eAAe;EACf;;EAGA;EACA,eAAe;EACf;;EAGA;EACA,kBAAkB;EAClB,cAAc;EACd;;;KAIM,oBAAoB;UAEf,6BAA6B,kBAAkB,UAAU;;EAExE;;EAEA,QAAQ;;EAER;;EAEA;;EAEA;;;;;EAKA;;EAEA;;EAEA;;EAEA,2BAA2B,QAAQ;;EAEnC;;;;;;EAMA,YAAY;;;;;;EAMZ,oBAAoB,UAAU;;EAE9B,oBAAoB,UAAU,WAAW,UAAU;;EAEnD,SAAS;;;;;EAKT;EACA,SAAS;;;;;;;;;;iBAWK,uBAAuB,kBAAkB,UAAU,WACjE,QAAQ,6BAA6B,WAAW,aAC/C,mBAAmB,WAAW;;;UC5IhB;EACf;EACA;EACA;EACA;EACA;EACA;EACA;EACA;EACA;EACA;EACA;EACA;EACA;EACA;EACA;EACA;;;;;EAKA,YAAY;;KAGF,wBAAwB;UAEnB,iCAAiC,kBAAkB,UAAU;EAC5E;;EAEA;EACA;;EAEA;EACA,SAAS;;;;;EAKT,WAAW;;EAEX;;EAEA;EACA;;EAEA;;EAEA,2BAA2B,QAAQ;EACnC;EACA,oBAAoB,UAAU;EAC9B,oBAAoB,UAAU,WAAW,UAAU;EACnD,SAAS;EACT,SAAS;;;iBAIK,2BAA2B,kBAAkB,UAAU,WACrE,QAAQ,iCAAiC,WAAW,aACnD,mBAAmB,WAAW"}
|
package/dist/{skillopt-optimization-method-CxrOjp1S.js → skillopt-optimization-method-S7sdjzvh.js}
RENAMED
|
@@ -355,7 +355,9 @@ function createExternalTextEvaluator(args) {
|
|
|
355
355
|
signal
|
|
356
356
|
});
|
|
357
357
|
cached.set(cacheKey, result);
|
|
358
|
-
result.
|
|
358
|
+
result.then((response) => {
|
|
359
|
+
if (response.info.status === "failed" && cached.get(cacheKey) === result) cached.delete(cacheKey);
|
|
360
|
+
}, () => {
|
|
359
361
|
if (cached.get(cacheKey) === result) cached.delete(cacheKey);
|
|
360
362
|
});
|
|
361
363
|
return result;
|
|
@@ -405,6 +407,26 @@ async function scoreOneScenario(args) {
|
|
|
405
407
|
maxConcurrency: 1,
|
|
406
408
|
...args.signal ? { signal: args.signal } : {}
|
|
407
409
|
});
|
|
410
|
+
const failedCell = campaign.cells.find((cell) => cell.error !== void 0);
|
|
411
|
+
if (failedCell) {
|
|
412
|
+
const stage = failedCell.errorStage ?? "unknown";
|
|
413
|
+
const message = failedCell.error?.slice(0, 2e3) ?? "cell failed without an error message";
|
|
414
|
+
const response = {
|
|
415
|
+
score: 0,
|
|
416
|
+
info: {
|
|
417
|
+
scenarioId: failedCell.scenarioId,
|
|
418
|
+
dimensions: {},
|
|
419
|
+
status: "failed",
|
|
420
|
+
notes: `Evaluation failed at ${stage}; optimizer penalty score 0 is not a measured zero. ${message}`,
|
|
421
|
+
error: {
|
|
422
|
+
stage,
|
|
423
|
+
message
|
|
424
|
+
}
|
|
425
|
+
}
|
|
426
|
+
};
|
|
427
|
+
if (JSON.stringify(response).length > args.maxEvidenceChars) throw new Error(`${args.label} failed evaluation evidence for '${args.scenario.id}' exceeds maxEvidenceChars`);
|
|
428
|
+
return response;
|
|
429
|
+
}
|
|
408
430
|
const breakdown = campaignBreakdown(campaign);
|
|
409
431
|
const row = breakdown.scenarios[0];
|
|
410
432
|
if (!row) throw new Error(`${args.label} evaluation produced no score for '${args.scenario.id}'`);
|
|
@@ -415,6 +437,7 @@ async function scoreOneScenario(args) {
|
|
|
415
437
|
info: {
|
|
416
438
|
scenarioId: row.scenarioId,
|
|
417
439
|
dimensions: breakdown.dimensions,
|
|
440
|
+
status: "scored",
|
|
418
441
|
...row.notes ? { notes: row.notes } : {},
|
|
419
442
|
...artifact !== void 0 ? { artifact } : {}
|
|
420
443
|
}
|
|
@@ -1996,4 +2019,4 @@ function skillOptOptimizationMethod(config) {
|
|
|
1996
2019
|
//#endregion
|
|
1997
2020
|
export { externalTextOptimizationMethod as a, REFERENCE_EQUIVALENCE_INPUT_LIMITS as c, runReferenceEquivalenceJudge as d, composeGate as i, REFERENCE_EQUIVALENCE_JUDGE_VERSION as l, gepaOptimizationMethod as n, decodeExternalTextCandidate as o, heldOutGate as r, readExternalOptimizerObservationArtifact as s, skillOptOptimizationMethod as t, createReferenceEquivalenceJudge as u };
|
|
1998
2021
|
|
|
1999
|
-
//# sourceMappingURL=skillopt-optimization-method-
|
|
2022
|
+
//# sourceMappingURL=skillopt-optimization-method-S7sdjzvh.js.map
|