@tangle-network/agent-eval 0.144.3 → 0.144.5
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 +14 -0
- package/dist/analyst/index.d.ts +363 -16
- package/dist/analyst/index.d.ts.map +1 -1
- package/dist/analyst/index.js +8 -8
- package/dist/{benchmark-CYtcIF2V.js → benchmark-B181aMF9.js} +2 -2
- package/dist/{benchmark-CYtcIF2V.js.map → benchmark-B181aMF9.js.map} +1 -1
- package/dist/{benchmark-CP6kWfj8.d.ts → benchmark-Fmo42QVE.d.ts} +3 -3
- package/dist/{benchmark-CP6kWfj8.d.ts.map → benchmark-Fmo42QVE.d.ts.map} +1 -1
- package/dist/{benchmark-command-CQPKRUr-.js → benchmark-command-CA_NFOmy.js} +919 -35
- package/dist/benchmark-command-CA_NFOmy.js.map +1 -0
- package/dist/benchmarks/index.d.ts +1 -1
- package/dist/benchmarks/index.js +1 -1
- package/dist/{benchmarks-CkG1bWFa.js → benchmarks-CWbsj0t4.js} +2 -2
- package/dist/{benchmarks-CkG1bWFa.js.map → benchmarks-CWbsj0t4.js.map} +1 -1
- package/dist/campaign/index.d.ts +4 -4
- package/dist/campaign/index.js +2 -2
- package/dist/{campaign-DjGFyPxH.js → campaign-ClpnD7Ug.js} +2 -2
- package/dist/{campaign-DjGFyPxH.js.map → campaign-ClpnD7Ug.js.map} +1 -1
- package/dist/cli.js +1 -1
- package/dist/{client-Bbht4xxl.d.ts → client-DAb7MWtL.d.ts} +2 -2
- package/dist/{client-Bbht4xxl.d.ts.map → client-DAb7MWtL.d.ts.map} +1 -1
- package/dist/{completion-verifier-EJERfFwF.d.ts → completion-verifier-VvpHRu78.d.ts} +3 -3
- package/dist/{completion-verifier-EJERfFwF.d.ts.map → completion-verifier-VvpHRu78.d.ts.map} +1 -1
- package/dist/contract/index.d.ts +6 -6
- package/dist/contract/index.js +4 -4
- package/dist/control.d.ts +2 -2
- package/dist/{default-registry-iXfu2trt.d.ts → default-registry-BwbZ9N9v.d.ts} +5 -5
- package/dist/{default-registry-iXfu2trt.d.ts.map → default-registry-BwbZ9N9v.d.ts.map} +1 -1
- package/dist/{default-registry-SOyHB6qG.js → default-registry-RLNNoeEP.js} +3 -3
- package/dist/{default-registry-SOyHB6qG.js.map → default-registry-RLNNoeEP.js.map} +1 -1
- package/dist/{dspy-rlm-engine-IRCG8kdi.js → dspy-rlm-engine-19FQEMBK.js} +2 -2
- package/dist/{dspy-rlm-engine-IRCG8kdi.js.map → dspy-rlm-engine-19FQEMBK.js.map} +1 -1
- package/dist/{exact-types-BygCBR4L.d.ts → exact-types-BQ7W90C4.d.ts} +2 -2
- package/dist/{exact-types-BygCBR4L.d.ts.map → exact-types-BQ7W90C4.d.ts.map} +1 -1
- package/dist/{extract-usage-7l1Xq5ti.js → extract-usage-BW27f3XW.js} +2 -2
- package/dist/{extract-usage-7l1Xq5ti.js.map → extract-usage-BW27f3XW.js.map} +1 -1
- package/dist/{feedback-trajectory-CSIkRLQX.d.ts → feedback-trajectory-WK7x4mhy.d.ts} +3 -3
- package/dist/{feedback-trajectory-CSIkRLQX.d.ts.map → feedback-trajectory-WK7x4mhy.d.ts.map} +1 -1
- package/dist/hosted/index.d.ts +2 -2
- package/dist/{index-BuHs_OnD.d.ts → index-CpxZSlB7.d.ts} +6 -6
- package/dist/{index-BuHs_OnD.d.ts.map → index-CpxZSlB7.d.ts.map} +1 -1
- package/dist/{index-D_qTihaQ.d.ts → index-CsuAo2-J.d.ts} +4 -4
- package/dist/{index-D_qTihaQ.d.ts.map → index-CsuAo2-J.d.ts.map} +1 -1
- package/dist/index.d.ts +13 -13
- package/dist/index.js +10 -10
- package/dist/{kind-factory-Bvwe3pup.js → kind-factory-BHIgPmzS.js} +2 -2
- package/dist/{kind-factory-Bvwe3pup.js.map → kind-factory-BHIgPmzS.js.map} +1 -1
- package/dist/multishot/index.d.ts +1 -1
- package/dist/multishot/index.js +1 -1
- package/dist/multishot/index.js.map +1 -1
- package/dist/openapi.json +1 -1
- package/dist/{replay-DQ-55DC_.d.ts → replay-B7S7Pdbw.d.ts} +3 -3
- package/dist/{replay-DQ-55DC_.d.ts.map → replay-B7S7Pdbw.d.ts.map} +1 -1
- package/dist/{replay-CqOsGjzU.js → replay-GW61ezMW.js} +4 -4
- package/dist/{replay-CqOsGjzU.js.map → replay-GW61ezMW.js.map} +1 -1
- package/dist/rl.d.ts +1 -1
- package/dist/{run-evidence-H1vRpIdT.d.ts → run-evidence-j5Ynww6L.d.ts} +2 -2
- package/dist/{run-evidence-H1vRpIdT.d.ts.map → run-evidence-j5Ynww6L.d.ts.map} +1 -1
- package/dist/{semantic-concept-judge-Do5aM9wP.js → semantic-concept-judge-DKRtp2sY.js} +2 -2
- package/dist/{semantic-concept-judge-Do5aM9wP.js.map → semantic-concept-judge-DKRtp2sY.js.map} +1 -1
- package/dist/{skill-usage-DtpLou9L.d.ts → skill-usage-DUvvudWR.d.ts} +5 -5
- package/dist/{skill-usage-DtpLou9L.d.ts.map → skill-usage-DUvvudWR.d.ts.map} +1 -1
- package/dist/{skillopt-optimization-method-CvSJGdm3.d.ts → skillopt-optimization-method-CGz9ywhM.d.ts} +4 -4
- package/dist/{skillopt-optimization-method-CvSJGdm3.d.ts.map → skillopt-optimization-method-CGz9ywhM.d.ts.map} +1 -1
- package/dist/{store-otlp-D4I90_vR.js → store-otlp-CKtTpRhv.js} +2 -2
- package/dist/{store-otlp-D4I90_vR.js.map → store-otlp-CKtTpRhv.js.map} +1 -1
- package/dist/{tool-groups-CMmsgTzj.d.ts → tool-groups-ByZiqpVk.d.ts} +4 -4
- package/dist/{tool-groups-CMmsgTzj.d.ts.map → tool-groups-ByZiqpVk.d.ts.map} +1 -1
- package/dist/traces.d.ts +3 -3
- package/dist/traces.js +4 -4
- package/dist/{types-y8jrxXWd.d.ts → types-CZt1PBIk.d.ts} +19 -1
- package/dist/{types-y8jrxXWd.d.ts.map → types-CZt1PBIk.d.ts.map} +1 -1
- package/dist/{types-DcJxgsLy.d.ts → types-Dcoaqcsc.d.ts} +2 -2
- package/dist/{types-DcJxgsLy.d.ts.map → types-Dcoaqcsc.d.ts.map} +1 -1
- package/dist/{usage-receipt-CgxMEBZq.js → usage-receipt-EVI8B8Xu.js} +8 -1
- package/dist/usage-receipt-EVI8B8Xu.js.map +1 -0
- package/dist/wire/index.d.ts +1 -1
- package/docs/prime-analyst.md +120 -0
- package/docs/trace-analysis.md +1 -1
- package/package.json +3 -3
- package/dist/benchmark-command-CQPKRUr-.js.map +0 -1
- package/dist/usage-receipt-CgxMEBZq.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-CA_NFOmy.js";
|
|
4
4
|
import { a as runRpcBatch, o as runRpcOnce, p as handleVersion, r as startServerAsync, s as buildOpenApi } from "./server-Df00sdwz.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-ooo9FWns.js";
|
|
2
|
-
import { _ as GateDecision, j as MutableSurface } from "./types-
|
|
2
|
+
import { _ as GateDecision, j as MutableSurface } from "./types-Dcoaqcsc.js";
|
|
3
3
|
import { o as InsightReport } from "./insight-report-D5m1z0_n.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-DAb7MWtL.d.ts.map
|
|
@@ -1 +1 @@
|
|
|
1
|
-
{"version":3,"file":"client-
|
|
1
|
+
{"version":3,"file":"client-DAb7MWtL.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"}
|
|
@@ -1,8 +1,8 @@
|
|
|
1
1
|
import { t as DefaultVerdict } from "./verdict-Dps8_okt.js";
|
|
2
2
|
import { c as CostLedgerHandle } from "./cost-ledger-FuQvHxPm.js";
|
|
3
3
|
import { A as ChatClient, _t as RawProviderSink } from "./types-BjMFz88h.js";
|
|
4
|
-
import { c as AnalystRunInputs, i as AnalystFinding, l as AnalystRunResult, n as AnalystContext, s as AnalystRunEvent, t as Analyst, u as AnalystRunSummary } from "./types-
|
|
5
|
-
import { i as ExactAnalystRunEvent, o as ExactAnalystRunResult, u as ExactExecutionComponentIdentity } from "./exact-types-
|
|
4
|
+
import { c as AnalystRunInputs, i as AnalystFinding, l as AnalystRunResult, n as AnalystContext, s as AnalystRunEvent, t as Analyst, u as AnalystRunSummary } from "./types-CZt1PBIk.js";
|
|
5
|
+
import { i as ExactAnalystRunEvent, o as ExactAnalystRunResult, u as ExactExecutionComponentIdentity } from "./exact-types-BQ7W90C4.js";
|
|
6
6
|
//#region src/analyst/registry.d.ts
|
|
7
7
|
interface AnalystHooks {
|
|
8
8
|
/** Legacy runs may mutate ctx; exact runs provide a frozen observational context. */
|
|
@@ -411,4 +411,4 @@ declare function createTokenRecallChecker(opts?: {
|
|
|
411
411
|
}): CorrectnessChecker;
|
|
412
412
|
//#endregion
|
|
413
413
|
export { ExactAnalystRunExecutionError as A, jsonHasKeys as C, AnalystRegistryOptions as D, AnalystRegistry as E, RegistryRunOpts as M, assertExactRegistryRunOpts as N, BudgetPolicy as O, containsAll as S, AnalystHooks as T, ValidationContext as _, ProducedProposal as a, byteLengthRange as b, SatisfiedBy as c, createLlmCorrectnessChecker as d, createTokenRecallChecker as f, ArtifactValidator as g, Artifact as h, LlmCorrectnessCheckerOpts as i, ExactRegistryRunOpts as j, ExactAnalystBudgetPolicy as k, TaskGold as l, verifyCompletion as m, CompletionVerdict as n, ProducedState as o, parseCorrectnessResponse as p, CorrectnessChecker as r, RequirementCheck as s, CompletionRequirement as t, completionVerdict as u, ValidationIssue as v, regexMatch as w, composeValidators as x, ValidationResult as y };
|
|
414
|
-
//# sourceMappingURL=completion-verifier-
|
|
414
|
+
//# sourceMappingURL=completion-verifier-VvpHRu78.d.ts.map
|
package/dist/{completion-verifier-EJERfFwF.d.ts.map → completion-verifier-VvpHRu78.d.ts.map}
RENAMED
|
@@ -1 +1 @@
|
|
|
1
|
-
{"version":3,"file":"completion-verifier-
|
|
1
|
+
{"version":3,"file":"completion-verifier-VvpHRu78.d.ts","names":[],"sources":["../src/analyst/registry.ts","../src/artifact-validator.ts","../src/completion-verifier.ts"],"mappings":";;;;;;UAqDiB;;EAEf,iBAAiB;IACf,SAAS;IACT,KAAK;IACL;aACS;;EAEX,gBAAgB;IACd,SAAS;IACT,SAAS;IACT,UAAU;IACV;aACS;;;;;;EAMX,SAAS;IACP,SAAS;IACT,OAAO;IACP;MACE,+BAA+B,QAAQ;;EAE3C,YAAY;IAAQ,QAAQ;aAA4B;;UAGzC;;EAEf;;EAEA,UAAU;;;;;;;EAOV,YAAY;IACV,SAAS;IACT;IACA;IACA;;;UAIa;;EAEf,OAAO;;EAEP,OAAO,aAAa,SAAS;;EAE7B,QAAQ;;EAER,gBAAgB;;EAEhB,gBAAgB;;EAEhB,eAAe;;UAGA;;EAEf;;EAEA;;EAEA,SAAS;;EAET;;EAEA,SAAS;;EAET,aAAa;;EAEb;;EAEA,OAAO;;;;;;;;;EASP,gBAAgB,cAAc,kBAAkB,eAAe,cAAc;;;;;;EAM7E;;;KAIU;WAEG;WACA;;WAGA;WACA;;WAEA,SAAS,SAAS;;;;;;;;;;UAWhB;WACN;WACA,QAAQ;WACR;WACA,QAAQ;WACR,YAAY;WACZ,oBAAoB;WACpB;WACA,MAAM,SAAS;WACf,eACL,cAAc,kBACd,SAAS,eAAe,cAAc;WAEjC;WACA;WACA;WACA;;;cAqCE,sCAAsC;WACxC;WACA,QAAQ;EAEjB,YAAY,iBAAiB,QAAQ,uBAAuB,UAAU;;cAU3D;mBACM;mBACA;EAEjB,YAAY,UAAS;EAIrB,SAAS,SAAS;EAsBlB,QAAQ;IACN;IACA;IACA;IACA,MAAM;;EAUF,IACJ,eACA,QAAQ,kBACR,UAAS,kBACR,QAAQ;;EAUL,SACJ,eACA,QAAQ,kBACR,SAAS,uBACR,QAAQ;;EAQJ,eACL,eACA,QAAQ,kBACR,SAAS,uBACR,eAAe;;;;;;;;;;;;EAmBX,UACL,eACA,QAAQ,kBACR,UAAS,kBACR,eAAe;UAIV;UA8BA;UAwFO;UA2aP;UAaA;UAUA;;;iBAiGM,2BAA2B,yBAAyB,SAAS;;;;;;;;;;;;;;;;UC79B5D;;EAEf;;EAEA;;EAEA;;EAEA,QAAQ;;EAER,WAAW;;UAGI;EACf;EACA;;EAEA,iBAAiB;;EAEjB,QAAQ;;UAGO;EACf;EACA;;EAEA;;UAGe;EACf;;EAEA;EACA,QAAQ;;EAER,WAAW;;UAGI;;EAEf;;EAEA;;EAEA,SAAS,UAAU,UAAU,SAAS,oBAAoB,QAAQ;;;;;;iBAWpD,kBACd,YAAY,qBACZ;EAAY;EAAe;IAC1B;;iBAgCa,WAAW,cAAc,SAAS,SAAS;;iBAkB3C,YAAY,cAAc,0BAA0B;;iBAuCpD,gBAAgB,cAAc,aAAa,cAAc;;iBAoBzD,YACd,cACA,oBACA;EAAY;IACX;;;;KCnJS;UAEK;;EAEf;;EAEA;;EAEA;;EAEA,cAAc;;UAGC;EACf;EACA,cAAc;;UAGC;EACf;EACA;EACA;;EAEA;;;UAIe;;EAEf,WAAW;;EAEX,WAAW;;EAEX;;UAGe;EACf;EACA;;EAEA;;;;;;EAMA;;EAEA;;;;;;;;;EASA;;EAEA;;EAEA;;;;;UAMe,0BAA0B;EACzC;EACA,cAAc;;EAEd;;EAEA;;EAEA;;;;;;;;;iBAUc,kBAAkB;EAChC;EACA,cAAc;IACZ;;;;;;KAoCQ,sBACV,aAAa,uBACb,oBACG;EAAU;EAAkB;;;;;;;;;iBAqLX,iBACpB,MAAM,UACN,OAAO,eACP,kBAAkB,qBACjB,QAAQ;UAoFM;EACf;;EAEA,aAAa;EACb;EACA,WAAW;EACX,SAAS;;EAET;;;;;;EAMA;;;;;;EAMA,UAAU;;;;;;;;;iBAUI,yBAAyB;EAAgB;EAAkB;;;;;;;;iBAiC3D,4BACd,MAAM,YACN,OAAM,4BACL;;;;;;;;;;;;;;;iBA6Ia,yBACd;EAAQ;EAAoB;IAC3B"}
|
package/dist/contract/index.d.ts
CHANGED
|
@@ -1,13 +1,13 @@
|
|
|
1
1
|
import { c as CostLedgerHandle, f as CostLedgerSummary, m as CostReceipt } from "../cost-ledger-FuQvHxPm.js";
|
|
2
2
|
import { a as RunRecord, n as RunCostProvenance, s as RunSplitTag } from "../run-record-ooo9FWns.js";
|
|
3
3
|
import { A as ChatClient, F as CreateChatClientOpts, V as createChatClient } from "../types-BjMFz88h.js";
|
|
4
|
-
import { h as ProposalFindingOrigin, i as AnalystFinding, m as ProposalFinding, v 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, 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-
|
|
4
|
+
import { h as ProposalFindingOrigin, i as AnalystFinding, m as ProposalFinding, v as makeProposalFinding } from "../types-CZt1PBIk.js";
|
|
5
|
+
import { n as buildDefaultAnalystRegistry, t as DefaultAnalystRegistryOptions } from "../default-registry-BwbZ9N9v.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-Dcoaqcsc.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-D5m1z0_n.js";
|
|
8
|
-
import { A as summarizeExecution, D as ExecutionReport, E as AnalyzeRunsOptions, O as SummarizeExecutionOptions, k as analyzeRuns } from "../tool-groups-
|
|
9
|
-
import { A as RunEvalOptions, An as LlmJudgeDimension, B as OptimizerModelBudget, Bt as ComparisonCost, C as RunImprovementLoopOptions, Cn as REFERENCE_EQUIVALENCE_JUDGE_VERSION, D as RunOptimizationOptions, Dn as ReferenceEquivalenceScenario, E as PremeasuredOptimizationBaseline, En as ReferenceEquivalenceJudgeResult, 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, Mn as llmJudge, Mt as composeGate, N as GepaEngineOptions, Nt as externalTextOptimizationMethod, On as createReferenceEquivalenceJudge, P as GepaEngineRun, Pt as ExternalTextOptimizationMethodConfig, Qt as compareOptimizationMethods, R as gepaOptimizationMethod, Rt as ExternalTextEvaluationResponse, Sn as REFERENCE_EQUIVALENCE_INPUT_LIMITS, T as runImprovementLoop, Tn as ReferenceEquivalenceJudgeOptions, 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, _n as campaignSplitDigest, an as RunCampaignOptions, dt as DefaultProductionGateCheck, fn as fsCampaignStorage, ft as DefaultProductionGateOptions, i as skillOptOptimizationMethod, j as runEval, jn as LlmJudgeOptions, kn as runReferenceEquivalenceJudge, lt as HeldOutGateOptions, mt as defaultProductionGate, n as SkillOptRunnerCommand, ot as PowerPreflight, p as LoopProvenanceRecord, pn as inMemoryCampaignStorage, pt as DefaultProductionRewardHackingOptions, q as PromotionObjective, r as SkillOptTrainerConfig, sn as runCampaign, t as SkillOptOptimizationMethodConfig, tn as CampaignCellFailureReceipt, un as CampaignStorage, ut as heldOutGate, w as RunImprovementLoopResult, wn as ReferenceEquivalenceJudgeInput, z as OpenAICompatibleOptimizerModel, zt as CompareOptimizationMethodsOptions } from "../skillopt-optimization-method-
|
|
10
|
-
import { c as EvalRunGenerationSnapshot, g as TraceSpanEvent, n as HostedTenant, o as EvalRunCellScore, s as EvalRunEvent } from "../client-
|
|
8
|
+
import { A as summarizeExecution, D as ExecutionReport, E as AnalyzeRunsOptions, O as SummarizeExecutionOptions, k as analyzeRuns } from "../tool-groups-ByZiqpVk.js";
|
|
9
|
+
import { A as RunEvalOptions, An as LlmJudgeDimension, B as OptimizerModelBudget, Bt as ComparisonCost, C as RunImprovementLoopOptions, Cn as REFERENCE_EQUIVALENCE_JUDGE_VERSION, D as RunOptimizationOptions, Dn as ReferenceEquivalenceScenario, E as PremeasuredOptimizationBaseline, En as ReferenceEquivalenceJudgeResult, 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, Mn as llmJudge, Mt as composeGate, N as GepaEngineOptions, Nt as externalTextOptimizationMethod, On as createReferenceEquivalenceJudge, P as GepaEngineRun, Pt as ExternalTextOptimizationMethodConfig, Qt as compareOptimizationMethods, R as gepaOptimizationMethod, Rt as ExternalTextEvaluationResponse, Sn as REFERENCE_EQUIVALENCE_INPUT_LIMITS, T as runImprovementLoop, Tn as ReferenceEquivalenceJudgeOptions, 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, _n as campaignSplitDigest, an as RunCampaignOptions, dt as DefaultProductionGateCheck, fn as fsCampaignStorage, ft as DefaultProductionGateOptions, i as skillOptOptimizationMethod, j as runEval, jn as LlmJudgeOptions, kn as runReferenceEquivalenceJudge, lt as HeldOutGateOptions, mt as defaultProductionGate, n as SkillOptRunnerCommand, ot as PowerPreflight, p as LoopProvenanceRecord, pn as inMemoryCampaignStorage, pt as DefaultProductionRewardHackingOptions, q as PromotionObjective, r as SkillOptTrainerConfig, sn as runCampaign, t as SkillOptOptimizationMethodConfig, tn as CampaignCellFailureReceipt, un as CampaignStorage, ut as heldOutGate, w as RunImprovementLoopResult, wn as ReferenceEquivalenceJudgeInput, z as OpenAICompatibleOptimizerModel, zt as CompareOptimizationMethodsOptions } from "../skillopt-optimization-method-CGz9ywhM.js";
|
|
10
|
+
import { c as EvalRunGenerationSnapshot, g as TraceSpanEvent, n as HostedTenant, o as EvalRunCellScore, s as EvalRunEvent } from "../client-DAb7MWtL.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
|
package/dist/contract/index.js
CHANGED
|
@@ -1,16 +1,16 @@
|
|
|
1
1
|
import { c as ValidationError } from "../errors-D-LKuDhb.js";
|
|
2
2
|
import { i as estimateCost, o as isModelPriced } from "../metrics-C9YY1OcL.js";
|
|
3
3
|
import { i as CostLedger } from "../cost-ledger-DMFxsLKr.js";
|
|
4
|
-
import { C as createChatClient, t as buildDefaultAnalystRegistry } from "../default-registry-
|
|
4
|
+
import { C as createChatClient, t as buildDefaultAnalystRegistry } from "../default-registry-RLNNoeEP.js";
|
|
5
5
|
import { LLM_MODEL_ATTR_KEYS, SPAN_KIND_ATTR_KEYS } from "../trace-attributes.js";
|
|
6
|
-
import { nt as classifyOtlpSpanRole, rt as isOtlpModelCall } from "../kind-factory-
|
|
7
|
-
import { s as makeProposalFinding } from "../usage-receipt-
|
|
6
|
+
import { nt as classifyOtlpSpanRole, rt as isOtlpModelCall } from "../kind-factory-BHIgPmzS.js";
|
|
7
|
+
import { s as makeProposalFinding } from "../usage-receipt-EVI8B8Xu.js";
|
|
8
8
|
import { i as inMemoryCampaignStorage, n as createRunCostLedger, r as fsCampaignStorage } from "../single-run-lock-D5iN0Xzb.js";
|
|
9
9
|
import { m as mapConcurrentRange } from "../ledger-core-DXZIqu17.js";
|
|
10
10
|
import { A as assertOptimizationResult, B as surfaceContentHash, Ct as decidePairedPromotion, M as compareOptimizationMethods, O as composeGate, S as defaultProductionGate, T as heldoutSignificance, V as surfaceHash, X as resolveRunDir, Y as runCampaign, _ as buildEvidenceVector, a as emitLoopProvenance, at as REFERENCE_EQUIVALENCE_JUDGE_VERSION, b as powerPreflight, d as runImprovementLoop, g as gepaOptimizationMethod, gt as llmJudge, it as REFERENCE_EQUIVALENCE_INPUT_LIMITS, k as externalTextOptimizationMethod, o as loopProvenanceArgsFromResult, ot as createReferenceEquivalenceJudge, p as runEval, st as runReferenceEquivalenceJudge, t as skillOptOptimizationMethod, tt as campaignSplitDigest, v as paretoPolicy, x as heldOutGate, y as paretoSignificanceGate } from "../skillopt-optimization-method-C4FX42dy.js";
|
|
11
11
|
import { E as pairedBootstrap } from "../statistics-ByxzSiOM.js";
|
|
12
12
|
import { i as parseRunRecordSafe, r as modelHasSnapshot } from "../run-record-CWN8-VsV.js";
|
|
13
|
-
import { a as recordAggregateMeasurements, i as readTaskFailureLabels, o as summarizeExecutionMeasurements, s as summarizeTraceErrors, t as extractUsage } from "../extract-usage-
|
|
13
|
+
import { a as recordAggregateMeasurements, i as readTaskFailureLabels, o as summarizeExecutionMeasurements, s as summarizeTraceErrors, t as extractUsage } from "../extract-usage-BW27f3XW.js";
|
|
14
14
|
import { n as summarizeExecution, t as analyzeRuns } from "../analyze-runs-DWIvOAGk.js";
|
|
15
15
|
import { c as campaignCellTaskScore, l as campaignCellToRunRecord, o as campaignCellExecutionEvidence, s as campaignCellJudgeDimensions } from "../reward-hacking-CyuzxKly.js";
|
|
16
16
|
import { n as InMemoryOutcomeStore, t as FileSystemOutcomeStore } from "../outcome-store-ChBKlTd_.js";
|
package/dist/control.d.ts
CHANGED
|
@@ -1,3 +1,3 @@
|
|
|
1
|
-
import { $ as ControlRunResult, J as ControlActionOutcome, Q as ControlEvalResult, X as ControlContext, Y as ControlBudget, Z as ControlDecision, at as StopDecision, ct as runAgentControlLoop, dt as subjectiveEval, et as ControlRuntimeConfig, it as ControlStopPolicies, lt as stopOnNoProgress, nt as ControlSeverity, ot as allCriticalPassed, q as ControlActionFailureMode, rt as ControlStep, st as objectiveEval, tt as ControlRuntimeError, ut as stopOnRepeatedAction } from "./feedback-trajectory-
|
|
2
|
-
import { A as ActionExecutionPolicy, M as evaluateActionPolicy, _ as ProposeReviewReport, a as ProposeReviewControlAction, c as ProposeReviewControlState, g as ProposeReviewConfig, i as scoreFromEvals, j as ActionPolicyDecision, k as runProposeReview, n as RunEvidenceMetadata, o as ProposeReviewControlConfig, r as controlRunToRunRecord, s as ProposeReviewControlResult, t as ControlRunToRunRecordOptions, u as runProposeReviewAsControlLoop } from "./run-evidence-
|
|
1
|
+
import { $ as ControlRunResult, J as ControlActionOutcome, Q as ControlEvalResult, X as ControlContext, Y as ControlBudget, Z as ControlDecision, at as StopDecision, ct as runAgentControlLoop, dt as subjectiveEval, et as ControlRuntimeConfig, it as ControlStopPolicies, lt as stopOnNoProgress, nt as ControlSeverity, ot as allCriticalPassed, q as ControlActionFailureMode, rt as ControlStep, st as objectiveEval, tt as ControlRuntimeError, ut as stopOnRepeatedAction } from "./feedback-trajectory-WK7x4mhy.js";
|
|
2
|
+
import { A as ActionExecutionPolicy, M as evaluateActionPolicy, _ as ProposeReviewReport, a as ProposeReviewControlAction, c as ProposeReviewControlState, g as ProposeReviewConfig, i as scoreFromEvals, j as ActionPolicyDecision, k as runProposeReview, n as RunEvidenceMetadata, o as ProposeReviewControlConfig, r as controlRunToRunRecord, s as ProposeReviewControlResult, t as ControlRunToRunRecordOptions, u as runProposeReviewAsControlLoop } from "./run-evidence-j5Ynww6L.js";
|
|
3
3
|
export { type ActionExecutionPolicy, type ActionPolicyDecision, type ControlActionFailureMode, type ControlActionOutcome, type ControlBudget, type ControlContext, type ControlDecision, type ControlEvalResult, type ControlRunResult, type ControlRunToRunRecordOptions, type ControlRuntimeConfig, type ControlRuntimeError, type ControlSeverity, type ControlStep, type ControlStopPolicies, type ProposeReviewConfig, type ProposeReviewControlAction, type ProposeReviewControlConfig, type ProposeReviewControlResult, type ProposeReviewControlState, type ProposeReviewReport, type RunEvidenceMetadata, type StopDecision, allCriticalPassed, controlRunToRunRecord, evaluateActionPolicy, objectiveEval, runAgentControlLoop, runProposeReview, runProposeReviewAsControlLoop, scoreFromEvals, stopOnNoProgress, stopOnRepeatedAction, subjectiveEval };
|
|
@@ -1,7 +1,7 @@
|
|
|
1
|
-
import { N as TraceAnalystSpan, i as AnalystFinding, n as AnalystContext, x as TraceAnalysisStore } from "./types-
|
|
2
|
-
import { l as ExactCapableAnalyst } from "./exact-types-
|
|
3
|
-
import { D as AnalystRegistryOptions, E as AnalystRegistry } from "./completion-verifier-
|
|
4
|
-
import { i as TraceAnalysisEngine, o as TraceAnalysisEngineResult, p as RawAnalystFinding, s as TraceAnalystLimits, t as TraceToolGroupName } from "./tool-groups-
|
|
1
|
+
import { N as TraceAnalystSpan, i as AnalystFinding, n as AnalystContext, x as TraceAnalysisStore } from "./types-CZt1PBIk.js";
|
|
2
|
+
import { l as ExactCapableAnalyst } from "./exact-types-BQ7W90C4.js";
|
|
3
|
+
import { D as AnalystRegistryOptions, E as AnalystRegistry } from "./completion-verifier-VvpHRu78.js";
|
|
4
|
+
import { i as TraceAnalysisEngine, o as TraceAnalysisEngineResult, p as RawAnalystFinding, s as TraceAnalystLimits, t as TraceToolGroupName } from "./tool-groups-ByZiqpVk.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-
|
|
118
|
+
//# sourceMappingURL=default-registry-BwbZ9N9v.d.ts.map
|
|
@@ -1 +1 @@
|
|
|
1
|
-
{"version":3,"file":"default-registry-
|
|
1
|
+
{"version":3,"file":"default-registry-BwbZ9N9v.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"}
|
|
@@ -1,8 +1,8 @@
|
|
|
1
1
|
import { n as LlmClient } from "./llm-client-D3EoChAU.js";
|
|
2
2
|
import { LLM_CONTEXT_TOKENS, LLM_INPUT_TOKEN_ATTR_KEYS, LLM_OUTPUT_TOKEN_ATTR_KEYS, TOOL_NAME_ATTR_KEYS } from "./trace-attributes.js";
|
|
3
3
|
import { t as executionTrackByLane } from "./execution-tracks-CpgFPpS5.js";
|
|
4
|
-
import { $ as spanEpochMillis, H as snapshotExactExecutionPlan, R as findingSubjectGrammarPromptFor, U as deepFreezeCanonicalJson, V as snapshotExactExecutionComponentIdentity, t as createTraceAnalyst } from "./kind-factory-
|
|
5
|
-
import { i as validateUsageSettlementTimeout, o as makeFinding, t as assertValidAnalystUsageReceipt } from "./usage-receipt-
|
|
4
|
+
import { $ as spanEpochMillis, H as snapshotExactExecutionPlan, R as findingSubjectGrammarPromptFor, U as deepFreezeCanonicalJson, V as snapshotExactExecutionComponentIdentity, t as createTraceAnalyst } from "./kind-factory-BHIgPmzS.js";
|
|
5
|
+
import { i as validateUsageSettlementTimeout, o as makeFinding, t as assertValidAnalystUsageReceipt } from "./usage-receipt-EVI8B8Xu.js";
|
|
6
6
|
import { i as hashCanonical, r as canonicalString } from "./canonical-D011XM8r.js";
|
|
7
7
|
import { t as analyzeSupervisorRunIntegrity } from "./integrity-DY6tIbl0.js";
|
|
8
8
|
import { o as combineAbortSignals } from "./proposal-findings-2GIUo1et.js";
|
|
@@ -2449,4 +2449,4 @@ function buildDefaultAnalystRegistry(options = {}) {
|
|
|
2449
2449
|
//#endregion
|
|
2450
2450
|
export { createChatClient as C, computeTraceMetrics as S, CONTROL_INTEGRITY_ANALYST as _, analystFindingDigest as a, behavioralAnalyst as b, completedAnalystReviewQuality as c, validateAnalystReviewDecisions as d, DEFAULT_TRACE_ANALYST_KINDS as f, FAILURE_MODE_KIND_SPEC as g, IMPROVEMENT_KIND_SPEC as h, assertExactRegistryRunOpts as i, readAnalystReview as l, KNOWLEDGE_GAP_KIND_SPEC as m, AnalystRegistry as n, analystRunDigest as o, KNOWLEDGE_POISONING_KIND_SPEC as p, ExactAnalystRunExecutionError as r, assertUniqueFindingIds as s, buildDefaultAnalystRegistry as t, snapshotAnalystRun as u, ControlIntegrityAnalyst as v, deriveEfficiencyFindings as x, emitControlIntegrityFindings as y };
|
|
2451
2451
|
|
|
2452
|
-
//# sourceMappingURL=default-registry-
|
|
2452
|
+
//# sourceMappingURL=default-registry-RLNNoeEP.js.map
|