@tangle-network/agent-eval 0.161.1 → 0.170.0

package/dist/{llm-judge-BhasIPFT.js → llm-judge-DbJdo8Nj.js}

@@ -1,18 +1,18 @@
 import { i as JudgeError, s as ValidationError, t as AgentEvalError } from "./errors-Dngq5h35.js";
-import { i as hashCanonical, r as canonicalString } from "./canonical-IL-Bu-14.js";
-import { s as weightedComposite, t as confidenceInterval } from "./descriptive-jDOuI6mz.js";
-import { r as pairedBootstrap } from "./paired-tests-BHIhYVdu.js";
-import { i as modelHasSnapshot } from "./run-record-BC0ebuRP.js";
-import { n as contentHash } from "./verdict-cache-CdVVTVmn.js";
-import { a as campaignCellCostProvenance, o as campaignCellExecutionEvidence, t as detectRewardHacking, u as projectCampaignCellQuality } from "./reward-hacking-t4lB1yt8.js";
+import { a as hashCanonical, i as compareCodeUnits, r as canonicalString } from "./canonical-DPyQ_rpt.js";
+import { s as weightedComposite, t as confidenceInterval } from "./descriptive-1V17A-qa.js";
+import { r as pairedBootstrap } from "./paired-tests-C8iCsioC.js";
+import { i as modelHasSnapshot } from "./run-record-DLORoL7t.js";
+import { n as contentHash } from "./verdict-cache-B3eCVQtY.js";
+import { a as campaignCellCostProvenance, o as campaignCellExecutionEvidence, t as detectRewardHacking, u as projectCampaignCellQuality } from "./reward-hacking-SkxYgT0x.js";
 import { i as CostLedger, t as CostAccountingIncompleteError } from "./cost-ledger-B1qx30B4.js";
-import { u as mapConcurrent } from "./ledger-core-BOzlRygb.js";
-import { b as createRunCostLedger, h as isRecord, m as isExternalTextCandidate, x as fsCampaignStorage } from "./external-optimizer-subprocess-BIWbHpgD.js";
+import { u as mapConcurrent } from "./ledger-core-PIfjCbKn.js";
+import { S as fsCampaignStorage, g as isRecord, h as isExternalTextCandidate, x as createRunCostLedger } from "./external-optimizer-subprocess-Cex8Da2i.js";
 import { t as DEFAULT_REDACTION_RULES } from "./redact-7Aq1ukl-.js";
-import { a as heldoutSignificance, i as dimensionRegressions, o as pairHoldout } from "./power-preflight-DEw-uC7q.js";
-import { a as combineAbortSignals, n as clamp01, r as assertProposalFindings } from "./run-score-lDzV0X8j.js";
-import { c as deepFreezeCanonicalJson } from "./types-BI4fT3HN.js";
-import { d as stripFencedJson, o as costReceiptFromLlm, s as costReceiptFromLlmError, u as maximumChargeForLlmRequest } from "./llm-client-hgDieDNN.js";
+import { a as heldoutSignificance, i as dimensionRegressions, o as pairHoldout } from "./power-preflight-CFXm0Vjo.js";
+import { r as combineAbortSignals, t as assertProposalFindings } from "./proposal-findings-bko3GGy-.js";
+import { l as deepFreezeCanonicalJson } from "./types-CiWITkGo.js";
+import { d as stripFencedJson, o as costReceiptFromLlm, s as costReceiptFromLlmError, u as maximumChargeForLlmRequest } from "./llm-client-BFMRpmqb.js";
 import { createHash } from "node:crypto";
 import { writeFileSync } from "node:fs";
 import { z } from "zod";
@@ -1166,7 +1166,7 @@ async function runCampaign(opts) {
 	if (firstCellFailure) throw firstCellFailure.cause;
 	if (failedLane) throw firstLaneError ?? failedLane.reason;
 	const endedAt = now();
-	cellsRef.sort((a, b) => a.cellId.localeCompare(b.cellId));
+	cellsRef.sort((a, b) => compareCodeUnits(a.cellId, b.cellId));
 	const aggregates = computeAggregates(cellsRef, judges, seed, costLedger.summary({ tags: { runDir: opts.runDir } }));
 	return {
 		manifestHash,
@@ -1665,7 +1665,7 @@ var Dataset = class Dataset {
 	* contamination-verifiable archives.
 	*/
 	toJsonl() {
-		return `${this.scenarios.slice().sort((a, b) => a.id.localeCompare(b.id)).map((s) => canonicalString(s)).join("\n")}\n`;
+		return `${this.scenarios.slice().sort((a, b) => compareCodeUnits(a.id, b.id)).map((s) => canonicalString(s)).join("\n")}\n`;
 	}
 	static fromJsonl(jsonl, manifest) {
 		const scenarios = [];
@@ -1689,7 +1689,7 @@ var Dataset = class Dataset {
 * coerced field would let two different datasets share one.
 */
 async function hashScenarios(scenarios) {
-	return hashCanonical(scenarios.slice().sort((a, b) => a.id.localeCompare(b.id))).slice(7);
+	return hashCanonical(scenarios.slice().sort((a, b) => compareCodeUnits(a.id, b.id))).slice(7);
 }
 /** Splitmix-ish deterministic shuffle — small, self-contained, no deps. */
 function seededShuffle(items, seed) {
@@ -3141,8 +3141,38 @@ function assertComponentSurface(surface) {
 		if (typeof content !== "string") throw new TypeError(`ComponentSurface component '${name}' must be a string`);
 	}
 }
-/** Return deterministic identity material independent of component key order. */
+/**
+* Deterministic identity material for a component surface.
+*
+* `canonicalString` orders keys by UTF-16 code unit (RFC 8785), which is a
+* property of the value alone. The previous material ordered them with
+* `localeCompare`, which reads the host's collation — so the same surface
+* could produce two different identities on two machines, and the stored
+* identity would stop matching a recomputation of the identical surface.
+*/
 function componentSurfaceIdentityMaterial(surface) {
+	assertComponentSurface(surface);
+	return canonicalString({
+		schema: "tangle.component-surface",
+		components: surface.components
+	});
+}
+/**
+* The retired material builder, kept PRIVATE and reachable only from
+* {@link surfaceHashMatches}.
+*
+* A surface identity recorded before this release was minted from these bytes.
+* The verify path tries the current material first and falls back to this one,
+* so a stored identity still matches its own surface; nothing mints from it.
+*
+* Every component surface's identity moves, not only one whose names sort
+* differently under the host's collation: RFC 8785 also orders the two
+* top-level keys, so `components` precedes `schema` where this builder emitted
+* them in literal order. The retention window therefore covers every stored
+* component-surface identity, which is why this builder is kept rather than
+* scoped to the mixed-case case.
+*/
+function retiredComponentSurfaceIdentityMaterial(surface) {
 	assertComponentSurface(surface);
 	return JSON.stringify({
 		schema: "tangle.component-surface",
@@ -3175,6 +3205,25 @@ function surfaceContentHash(surface) {
 function surfaceHash(surface) {
 	return surfaceContentHash(surface).slice(7, 23);
 }
+/**
+* Whether `storedHash` is the loop key of `surface`, under the current identity
+* material or the retired one.
+*
+* A stored key is 16 hex characters with no room for a scheme tag, so the
+* scheme cannot be read off the value the way an `agent-profile-cell` id names
+* its own. The verify path therefore tries both, which gives the same property:
+* a key minted by an earlier release still matches its own surface, and a
+* surface that was actually edited matches neither.
+*
+* Only a component surface can differ between the two; a prompt or code surface
+* produces identical material under both, so the second comparison is a no-op
+* for them.
+*/
+function surfaceHashMatches(surface, storedHash) {
+	if (surfaceHash(surface) === storedHash) return true;
+	if (typeof surface === "string" || surface.kind !== "components") return false;
+	return createHash("sha256").update(retiredComponentSurfaceIdentityMaterial(surface)).digest("hex").slice(0, 16) === storedHash;
+}
 /** Canonical customer-visible description of the exact before/after surfaces. */
 function renderSurfaceDiff(winnerSurface, baselineSurface) {
 	if (typeof winnerSurface === "string" && typeof baselineSurface === "string") return [
@@ -4995,7 +5044,7 @@ function resolveSelectedParent(returned, measuredByHash, generation) {
 	const parent = returned;
 	const measured = measuredByHash.get(parent.surfaceHash);
 	if (!measured) throw new Error(`runOptimization: selectParent returned surface "${parent.surfaceHash}" in generation ${generation}, which this run has not measured to completion; a parent must be a scored surface from the frontier`);
-	if (surfaceHash(parent.surface) !== parent.surfaceHash) throw new Error(`runOptimization: selectParent returned a parent whose surface does not match its surfaceHash "${parent.surfaceHash}" (generation ${generation})`);
+	if (!surfaceHashMatches(parent.surface, parent.surfaceHash)) throw new Error(`runOptimization: selectParent returned a parent whose surface does not match its surfaceHash "${parent.surfaceHash}" (generation ${generation})`);
 	return measured;
 }
 function immutableProposalSnapshot(value, label) {
@@ -5013,7 +5062,7 @@ function deepFreeze(value, seen = /* @__PURE__ */ new WeakSet()) {
 }
 function validatedPremeasuredBaseline(args) {
 	const { input } = args;
-	if (input.surfaceHash !== surfaceHash(args.baselineSurface)) throw new Error("runOptimization: premeasured baseline surface hash does not match baselineSurface");
+	if (!surfaceHashMatches(args.baselineSurface, input.surfaceHash)) throw new Error("runOptimization: premeasured baseline surface hash does not match baselineSurface");
 	const campaign = input.campaign;
 	if (campaign.reps !== args.reps) throw new Error(`runOptimization: premeasured baseline reps ${campaign.reps} do not match requested reps ${args.reps}`);
 	if (campaign.seed !== args.seed) throw new Error(`runOptimization: premeasured baseline seed ${campaign.seed} does not match requested seed ${args.seed}`);
@@ -5355,7 +5404,7 @@ function buildLoopProvenanceRecord(args) {
 			const measured = surfaceByHash.get(c.surfaceHash);
 			if (measured === void 0) throw new Error("buildLoopProvenanceRecord: measured candidate is missing its surface");
 			const { surface, campaign } = measured;
-			if (surfaceHash(surface) !== c.surfaceHash) throw new Error("buildLoopProvenanceRecord: candidate surface hash does not match its surface bytes");
+			if (!surfaceHashMatches(surface, c.surfaceHash)) throw new Error("buildLoopProvenanceRecord: candidate surface hash does not match its surface bytes");
 			if (campaign.splitDigest !== args.baselineSearchCampaign.splitDigest) throw new Error("buildLoopProvenanceRecord: candidate campaign does not match the search split");
 			const entry = {
 				generation: gen.generationIndex,
@@ -5424,7 +5473,7 @@ function buildLoopProvenanceRecord(args) {
 					lift: holdoutMeasurement.neutralized - holdoutMeasurement.baseline
 				} } : {}
 			},
-			costReceiptsDigest: canonicalDigest([...args.costReceipts].sort((left, right) => left.callId.localeCompare(right.callId)))
+			costReceiptsDigest: canonicalDigest([...args.costReceipts].sort((left, right) => compareCodeUnits(left.callId, right.callId)))
 		},
 		baselineSearchComposite,
 		gate: {
@@ -5482,7 +5531,7 @@ function campaignMeasurementDigest(campaign) {
 		reps: campaign.reps,
 		runDir: campaign.runDir,
 		scenarios: campaign.scenarios,
-		cells: [...campaign.cells].sort((left, right) => left.cellId.localeCompare(right.cellId)).map((cell) => ({
+		cells: [...campaign.cells].sort((left, right) => compareCodeUnits(left.cellId, right.cellId)).map((cell) => ({
 			manifestHash: cell.manifestHash ?? null,
 			cellId: cell.cellId,
 			scenarioId: cell.scenarioId,
@@ -5805,6 +5854,35 @@ async function emitLoopProvenance(args) {
 	};
 }
 //#endregion
+//#region src/run-score.ts
+const DEFAULT_RUN_SCORE_WEIGHTS = {
+	success: 4,
+	goalProgress: 2,
+	repoGroundedness: 1.5,
+	driftPenalty: -1.5,
+	toolUseQuality: 1,
+	patchQuality: 1.25,
+	testReality: 1.5,
+	finalGate: 3,
+	reviewerBlockers: -2,
+	costUsd: -.2,
+	wallSeconds: -.1
+};
+function aggregateRunScore(score, weights = {}) {
+	const w = {
+		...DEFAULT_RUN_SCORE_WEIGHTS,
+		...weights
+	};
+	return w.success * clamp01(score.success) + w.goalProgress * clamp01(score.goalProgress) + w.repoGroundedness * clamp01(score.repoGroundedness) + w.driftPenalty * clamp01(score.driftPenalty) + w.toolUseQuality * clamp01(score.toolUseQuality) + w.patchQuality * clamp01(score.patchQuality) + w.testReality * clamp01(score.testReality) + w.finalGate * clamp01(score.finalGate) + w.reviewerBlockers * clamp01(score.reviewerBlockers) + w.costUsd * Math.max(0, finiteOrZero(score.costUsd)) + w.wallSeconds * Math.max(0, finiteOrZero(score.wallSeconds) / 60);
+}
+function clamp01(value) {
+	if (!Number.isFinite(value)) return 0;
+	return Math.max(0, Math.min(1, value));
+}
+function finiteOrZero(value) {
+	return Number.isFinite(value) ? value : 0;
+}
+//#endregion
 //#region src/campaign/transient-failure.ts
 const BASE_TRANSIENT = /\b50[234]\b|no stream output|produced no stream|admission timed out|admission_rejected|queue_timeout|fetch failed|ECONNRESET|This operation was aborted/i;
 const TIMEOUT_PATTERN = /timeout after \d+ ?ms|cli-bridge timeout/i;
@@ -6173,6 +6251,6 @@ function firstString(value) {
 	return typeof value === "string" && value.trim() ? value : void 0;
 }
 //#endregion
-export { scoreRedTeamOutput as $, codeSurfaceIdentityMaterial as A, verifySearchHistoryReceipt as B, openAutoPr as C, costFromLedgerSummary as D, compareOptimizationMethods as E, SearchHistoryRequiredError as F, parseReflectionResponse as G, campaignMeanComposite as H, assertCompleteSearchHistory as I, readGepaCandidatePopulationArtifact as J, recoverTruncatedJson as K, assertSearchHistoryMatchesReplay as L, renderSurfaceDiff as M, surfaceContentHash as N, optimizationTokenUsageFromSummary as O, surfaceHash as P, redTeamReport as Q, createSearchHistoryReceipt as R, paretoFrontierWithCrowding as S, combineComparisonCosts as T, compareRankKeys as U, campaignBreakdown as V, buildReflectionPrompt as W, DEFAULT_RED_TEAM_CORPUS as X, defaultProductionGate as Y, redTeamDataset as Z, labelTrustRank as _, buildLoopProvenanceRecord as a, tangleTracesRoot as at, dominates as b, emitLoopProvenance as c, summarizeBackendIntegrity as ct, provenanceRecordPath as d, campaignScenarioIdentity as dt, runCanaries as et, provenanceSpansPath as f, campaignSplitDigest as ft, isProposedCandidate as g, computeManifestHash as gt, runOptimization as h, cellCachePath as ht, transientDispatchFailure as i, resolveRunDir as it, componentSurfaceIdentityMaterial as j, assertCodeSurfaceIdentity as k, loopProvenanceArgsFromResult as l, assertCampaignDesign as lt, runImprovementLoop as m, readCachedCell as mt, JudgeParseError as n, runCampaign as nt, campaignMeasurementDigest as o, BackendIntegrityError as ot, verifyLoopProvenanceRecord as p, campaignSplitDigestFromIdentities as pt, assertGepaCandidatePopulationSummary as q, isTransientTransportFailure as r, planCampaignRun as rt, canonicalDigest as s, assertRealBackend as st, llmJudge as t, runEval as tt, loopProvenanceSpans as u, assertCampaignSplitIdentity as ut, SearchRecorder as v, assertOptimizationResult as w, paretoFrontier as x, recordCandidatePopulationSearch as y, searchHistoryCoverageRow as z };
+export { redTeamDataset as $, optimizationTokenUsageFromSummary as A, createSearchHistoryReceipt as B, paretoFrontier as C, combineComparisonCosts as D, assertOptimizationResult as E, surfaceContentHash as F, compareRankKeys as G, verifySearchHistoryReceipt as H, surfaceHash as I, recoverTruncatedJson as J, buildReflectionPrompt as K, SearchHistoryRequiredError as L, codeSurfaceIdentityMaterial as M, componentSurfaceIdentityMaterial as N, compareOptimizationMethods as O, renderSurfaceDiff as P, DEFAULT_RED_TEAM_CORPUS as Q, assertCompleteSearchHistory as R, dominates as S, openAutoPr as T, campaignBreakdown as U, searchHistoryCoverageRow as V, campaignMeanComposite as W, readGepaCandidatePopulationArtifact as X, assertGepaCandidatePopulationSummary as Y, defaultProductionGate as Z, runOptimization as _, cellCachePath as _t, aggregateRunScore as a, planCampaignRun as at, SearchRecorder as b, campaignMeasurementDigest as c, BackendIntegrityError as ct, loopProvenanceArgsFromResult as d, assertCampaignDesign as dt, redTeamReport as et, loopProvenanceSpans as f, assertCampaignSplitIdentity as ft, runImprovementLoop as g, readCachedCell as gt, verifyLoopProvenanceRecord as h, campaignSplitDigestFromIdentities as ht, transientDispatchFailure as i, runCampaign as it, assertCodeSurfaceIdentity as j, costFromLedgerSummary as k, canonicalDigest as l, assertRealBackend as lt, provenanceSpansPath as m, campaignSplitDigest as mt, JudgeParseError as n, runCanaries as nt, clamp01 as o, resolveRunDir as ot, provenanceRecordPath as p, campaignScenarioIdentity as pt, parseReflectionResponse as q, isTransientTransportFailure as r, runEval as rt, buildLoopProvenanceRecord as s, tangleTracesRoot as st, llmJudge as t, scoreRedTeamOutput as tt, emitLoopProvenance as u, summarizeBackendIntegrity as ut, isProposedCandidate as v, computeManifestHash as vt, paretoFrontierWithCrowding as w, recordCandidatePopulationSearch as x, labelTrustRank as y, assertSearchHistoryMatchesReplay as z };
 
-//# sourceMappingURL=llm-judge-BhasIPFT.js.map
+//# sourceMappingURL=llm-judge-DbJdo8Nj.js.map