@tangle-network/agent-eval 0.144.11 → 0.144.13

package/dist/pipelines/index.js

@@ -1,123 +1,9 @@
-import { t as executionTrackByLane } from "../execution-tracks-CpgFPpS5.js";
-import { L as pearsonR, g as interRaterReliability } from "../statistics-ByxzSiOM.js";
+import { a as budgetBreachView, i as failureClusterView, n as computeToolUseMetrics, r as judgeAgreementView, t as toolWasteView } from "../tool-waste-BDdBZG1F.js";
+import { t as compareToBaseline } from "../baseline-BhPRQBVn.js";
 import { s as isToolSpan } from "../schema-CRhEY1SO.js";
+import { i as hasCapturedToolArgs, n as argHash, o as llmSpans, s as runFailureClass, t as aggregateLlm } from "../query-Di7eEQ79.js";
 import { t as buildTrajectory } from "../trajectory-D_7rLrvE.js";
-import { t as compareToBaseline } from "../baseline-C-GocmIW.js";
-import { i as hasCapturedToolArgs, l as toolSpans, n as argHash, o as llmSpans, s as runFailureClass, t as aggregateLlm } from "../query-Di7eEQ79.js";
-import { n as DEFAULT_RULES, r as classifyFailure, t as computeToolUseMetrics } from "../tool-use-metrics-DEGMKycK.js";
-//#region src/pipelines/budget-breach.ts
-async function budgetBreachView(store, options = {}) {
-	const runs = await store.listRuns({
-		scenarioId: options.scenarioId,
-		variantId: options.variantId
-	});
-	const findings = [];
-	const byDimension = {};
-	const byScenario = {};
-	const byVariant = {};
-	for (const run of runs) {
-		const entries = await store.budget(run.runId);
-		for (const e of entries) {
-			if (!e.breached) continue;
-			const excessRatio = e.limit > 0 ? e.consumed / e.limit : Infinity;
-			findings.push({
-				runId: run.runId,
-				scenarioId: run.scenarioId,
-				variantId: run.variantId,
-				dimension: e.dimension,
-				limit: e.limit,
-				consumed: e.consumed,
-				excessRatio,
-				timestamp: e.timestamp
-			});
-			byDimension[e.dimension] = (byDimension[e.dimension] ?? 0) + 1;
-			byScenario[run.scenarioId] = (byScenario[run.scenarioId] ?? 0) + 1;
-			if (run.variantId) byVariant[run.variantId] = (byVariant[run.variantId] ?? 0) + 1;
-		}
-	}
-	const breachedRuns = new Set(findings.map((f) => f.runId));
-	return {
-		findings,
-		byDimension,
-		byScenario,
-		byVariant,
-		totalRuns: runs.length,
-		breachedRunRatio: runs.length > 0 ? breachedRuns.size / runs.length : 0
-	};
-}
-//#endregion
-//#region src/pipelines/failure-cluster.ts
-/**
-* FailureClusterView — groups failed runs by (failureClass, triggerTool,
-* argHash-prefix) so weekly reviews can prioritize the top-N clusters.
-*
-* Each cluster includes: N runs, scenarios affected, representative
-* error message, a proposed mitigation hint (rule → action table).
-*/
-async function failureClusterView(store, options = {}) {
-	const rules = options.rules ?? DEFAULT_RULES;
-	const minSize = options.minClusterSize ?? 1;
-	const runs = await store.listRuns();
-	const clusters = /* @__PURE__ */ new Map();
-	let totalFailures = 0;
-	for (const run of runs) {
-		if (run.status === "completed" && run.outcome?.pass !== false) continue;
-		totalFailures++;
-		const spans = await store.spans({ runId: run.runId });
-		const cls = classifyFailure({
-			run,
-			spans,
-			events: await store.events({ runId: run.runId })
-		}, rules);
-		let toolName;
-		let argPrefix;
-		let dimension;
-		if (cls.triggerSpanId) {
-			const trig = spans.find((s) => s.spanId === cls.triggerSpanId);
-			if (trig?.kind === "tool") {
-				toolName = trig.toolName;
-				if (hasCapturedToolArgs(trig)) argPrefix = argHash(trig.args).slice(0, 16);
-			} else if (trig?.kind === "judge") dimension = trig.dimension;
-		}
-		if (!toolName) {
-			const errored = (await toolSpans(store, run.runId)).filter((t) => t.status === "error").pop();
-			if (errored) {
-				toolName = errored.toolName;
-				if (hasCapturedToolArgs(errored)) argPrefix = argHash(errored.args).slice(0, 16);
-			}
-		}
-		if (!dimension) {
-			const judge = spans.find((s) => s.kind === "judge" && typeof s.dimension === "string");
-			if (judge?.kind === "judge") dimension = judge.dimension;
-		}
-		const key = `${cls.failureClass}|${toolName ?? ""}|${argPrefix ?? ""}|${dimension ?? ""}`;
-		let cluster = clusters.get(key);
-		if (!cluster) {
-			cluster = {
-				failureClass: cls.failureClass,
-				toolName,
-				argPrefix,
-				dimension,
-				runCount: 0,
-				scenarioIds: [],
-				exampleRunId: run.runId,
-				exampleError: firstErrorMessage(spans) ?? cls.reason
-			};
-			clusters.set(key, cluster);
-		}
-		cluster.runCount++;
-		if (!cluster.scenarioIds.includes(run.scenarioId)) cluster.scenarioIds.push(run.scenarioId);
-	}
-	return {
-		clusters: [...clusters.values()].filter((c) => c.runCount >= minSize).sort((a, b) => b.runCount - a.runCount),
-		totalFailures,
-		totalRuns: runs.length
-	};
-}
-function firstErrorMessage(spans) {
-	return spans.find((s) => s.status === "error")?.error;
-}
-//#endregion
+import { t as executionTrackByLane } from "../execution-tracks-CpgFPpS5.js";
 //#region src/pipelines/first-divergence.ts
 async function firstDivergenceView(store, runA, runB, options = {}) {
 	const [a, b] = await Promise.all([buildTrajectory(store, runA), buildTrajectory(store, runB)]);
@@ -168,79 +54,6 @@ function describeDifference(a, b) {
 	return `name "${a.span.name}" vs "${b.span.name}"`;
 }
 //#endregion
-//#region src/pipelines/judge-agreement.ts
-/**
-* JudgeAgreementView — pairwise agreement between judges across the
-* corpus, grouped by dimension.
-*
-* Output drives two workflows:
-*   - Judge robustness audit: "does Claude agree with GPT at κ ≥ 0.6?"
-*   - Calibration tracking: κ vs golden human labels over time (by
-*     providing a `humanGoldenJudgeId`).
-*/
-async function judgeAgreementView(store) {
-	const all = (await store.spans({ kind: "judge" })).filter((s) => s.kind === "judge");
-	if (all.length === 0) return {
-		pairs: [],
-		dimensions: [],
-		judgeIds: []
-	};
-	const byDimension = /* @__PURE__ */ new Map();
-	for (const s of all) {
-		const arr = byDimension.get(s.dimension) ?? [];
-		arr.push(s);
-		byDimension.set(s.dimension, arr);
-	}
-	const judgeIds = [...new Set(all.map((s) => s.judgeId))].sort();
-	const pairs = [];
-	for (const [dim, spans] of byDimension) {
-		const byJudge = /* @__PURE__ */ new Map();
-		for (const s of spans) {
-			const m = byJudge.get(s.judgeId) ?? /* @__PURE__ */ new Map();
-			m.set(s.targetSpanId, s.score);
-			byJudge.set(s.judgeId, m);
-		}
-		const judgesHere = [...byJudge.keys()];
-		for (let i = 0; i < judgesHere.length; i++) for (let j = i + 1; j < judgesHere.length; j++) {
-			const judgeI = judgesHere[i];
-			const judgeJ = judgesHere[j];
-			const a = byJudge.get(judgeI);
-			const b = byJudge.get(judgeJ);
-			const common = [];
-			for (const [target, scoreA] of a) {
-				const scoreB = b.get(target);
-				if (scoreB !== void 0) common.push([scoreA, scoreB]);
-			}
-			if (common.length < 2) continue;
-			const judgeScores = common.map(([scoreA, scoreB]) => [{
-				judgeName: judgeI,
-				dimension: dim,
-				score: scoreA,
-				reasoning: ""
-			}, {
-				judgeName: judgeJ,
-				dimension: dim,
-				score: scoreB,
-				reasoning: ""
-			}]);
-			const k = interRaterReliability(judgeScores[0].map((_, k2) => judgeScores.map((pair) => pair[k2])));
-			pairs.push({
-				judgeA: judgeI,
-				judgeB: judgeJ,
-				dimension: dim,
-				commonItems: common.length,
-				pearson: pearsonR(common.map((c) => c[0]), common.map((c) => c[1])),
-				krippendorff: k
-			});
-		}
-	}
-	return {
-		pairs: pairs.sort((a, b) => b.commonItems - a.commonItems),
-		dimensions: [...byDimension.keys()].sort(),
-		judgeIds
-	};
-}
-//#endregion
 //#region src/pipelines/regression.ts
 /**
 * RegressionView — compares a candidate slice to a baseline slice on a
@@ -439,99 +252,6 @@ function callsAreSerial(previous, next) {
 	return previous.span.endedAt !== void 0 && previous.span.endedAt <= next.span.startedAt;
 }
 //#endregion
-//#region src/pipelines/tool-waste.ts
-/**
-* ToolWasteView — fraction of tool calls whose results weren't used
-* downstream. Without a "used" signal we fall back to structural
-* proxies: error calls, duplicate calls, and tool calls followed by
-* zero subsequent LLM spans are all considered waste.
-*
-* Consumers can pass a `usageOracle` that inspects a tool span and
-* returns true iff the tool's result appears in a later LLM message,
-* artifact, or state mutation — that's the canonical definition; the
-* default heuristic is a reasonable fallback.
-*/
-async function toolWasteView(store, options = {}) {
-	const runs = options.runId ? [options.runId] : (await store.listRuns()).map((r) => r.runId);
-	const byRun = [];
-	let totalCalls = 0;
-	let totalWasted = 0;
-	for (const runId of runs) {
-		const tools = await toolSpans(store, runId);
-		if (tools.length === 0) {
-			byRun.push({
-				runId,
-				wastedCalls: 0,
-				totalCalls: 0,
-				wasteRate: 0
-			});
-			continue;
-		}
-		const sortedLlm = [...await llmSpans(store, runId)].sort((a, b) => a.startedAt - b.startedAt);
-		const startTimes = sortedLlm.map((l) => l.startedAt);
-		const suffixText = buildSuffixText(sortedLlm);
-		let wasted = 0;
-		for (const t of tools) {
-			if (t.status === "error") {
-				wasted++;
-				continue;
-			}
-			const cutoff = upperBound(startTimes, t.startedAt);
-			if (options.usageOracle) {
-				if (!options.usageOracle(t, { llm: sortedLlm.slice(cutoff) })) wasted++;
-			} else {
-				const resultStr = stringify(t.result);
-				if (resultStr === "") continue;
-				if (!(suffixText[cutoff] ?? "").includes(resultStr.slice(0, 120))) wasted++;
-			}
-		}
-		const wasteRate = wasted / tools.length;
-		byRun.push({
-			runId,
-			wastedCalls: wasted,
-			totalCalls: tools.length,
-			wasteRate
-		});
-		totalCalls += tools.length;
-		totalWasted += wasted;
-	}
-	return {
-		byRun,
-		overallWasteRate: totalCalls > 0 ? totalWasted / totalCalls : 0
-	};
-}
-/**
-* Build per-position suffix haystacks: result[i] is the concatenation of every
-* string message content in spans[i..end]. Built back-to-front so each entry
-* reuses the next one — O(total message text) rather than O(spans²).
-*/
-function buildSuffixText(spans) {
-	const result = new Array(spans.length + 1);
-	result[spans.length] = "";
-	for (let i = spans.length - 1; i >= 0; i--) result[i] = `${spans[i].messages.map((m) => typeof m.content === "string" ? m.content : "").join("\n")}\n${result[i + 1]}`;
-	return result;
-}
-/** Index of the first element strictly greater than `target` in a sorted array. */
-function upperBound(sorted, target) {
-	let lo = 0;
-	let hi = sorted.length;
-	while (lo < hi) {
-		const mid = lo + hi >>> 1;
-		if (sorted[mid] <= target) lo = mid + 1;
-		else hi = mid;
-	}
-	return lo;
-}
-function stringify(v) {
-	if (v === null || v === void 0) return "";
-	if (typeof v === "string") return v;
-	try {
-		return JSON.stringify(v);
-	} catch {
-		return String(v);
-	}
-}
-//#endregion
 export { budgetBreachView, computeToolUseMetrics, failureClusterView, firstDivergenceView, judgeAgreementView, regressionView, stuckLoopView, toolWasteView };
 
 //# sourceMappingURL=index.js.map

package/dist/pipelines/index.js.map

@@ -1 +1 @@
-{"version":3,"file":"index.js","names":[],"sources":["../../src/pipelines/budget-breach.ts","../../src/pipelines/failure-cluster.ts","../../src/pipelines/first-divergence.ts","../../src/pipelines/judge-agreement.ts","../../src/pipelines/regression.ts","../../src/pipelines/stuck-loop.ts","../../src/pipelines/tool-waste.ts"],"sourcesContent":["/**\n * BudgetBreachView — aggregates breach events across the corpus.\n *\n * Answers: which dimensions get hit most often? Which scenarios are\n * underbudgeted? Which variants trigger the most breaches?\n */\n\nimport type { BudgetSpec } from '../trace/schema'\nimport type { TraceStore } from '../trace/store'\n\nexport interface BudgetBreachFinding {\n  runId: string\n  scenarioId: string\n  variantId?: string\n  dimension: keyof BudgetSpec\n  limit: number\n  consumed: number\n  excessRatio: number\n  timestamp: number\n}\n\nexport interface BudgetBreachReport {\n  findings: BudgetBreachFinding[]\n  byDimension: Record<string, number>\n  byScenario: Record<string, number>\n  byVariant: Record<string, number>\n  totalRuns: number\n  breachedRunRatio: number\n}\n\nexport async function budgetBreachView(\n  store: TraceStore,\n  options: { scenarioId?: string; variantId?: string } = {},\n): Promise<BudgetBreachReport> {\n  const runs = await store.listRuns({\n    scenarioId: options.scenarioId,\n    variantId: options.variantId,\n  })\n  const findings: BudgetBreachFinding[] = []\n  const byDimension: Record<string, number> = {}\n  const byScenario: Record<string, number> = {}\n  const byVariant: Record<string, number> = {}\n\n  for (const run of runs) {\n    const entries = await store.budget(run.runId)\n    for (const e of entries) {\n      if (!e.breached) continue\n      const excessRatio = e.limit > 0 ? e.consumed / e.limit : Infinity\n      findings.push({\n        runId: run.runId,\n        scenarioId: run.scenarioId,\n        variantId: run.variantId,\n        dimension: e.dimension,\n        limit: e.limit,\n        consumed: e.consumed,\n        excessRatio,\n        timestamp: e.timestamp,\n      })\n      byDimension[e.dimension] = (byDimension[e.dimension] ?? 0) + 1\n      byScenario[run.scenarioId] = (byScenario[run.scenarioId] ?? 0) + 1\n      if (run.variantId) byVariant[run.variantId] = (byVariant[run.variantId] ?? 0) + 1\n    }\n  }\n\n  const breachedRuns = new Set(findings.map((f) => f.runId))\n  return {\n    findings,\n    byDimension,\n    byScenario,\n    byVariant,\n    totalRuns: runs.length,\n    breachedRunRatio: runs.length > 0 ? breachedRuns.size / runs.length : 0,\n  }\n}\n","/**\n * FailureClusterView — groups failed runs by (failureClass, triggerTool,\n * argHash-prefix) so weekly reviews can prioritize the top-N clusters.\n *\n * Each cluster includes: N runs, scenarios affected, representative\n * error message, a proposed mitigation hint (rule → action table).\n */\n\nimport { classifyFailure, DEFAULT_RULES, type FailureRule } from '../failure-taxonomy'\nimport { argHash, hasCapturedToolArgs, toolSpans } from '../trace/query'\nimport type { FailureClass, Span } from '../trace/schema'\nimport type { TraceStore } from '../trace/store'\n\nexport interface FailureCluster {\n  failureClass: FailureClass\n  /** Tool name when the trigger was a tool span, else undefined. */\n  toolName?: string\n  /** First 16 chars of argHash — clusters similar args. */\n  argPrefix?: string\n  /**\n   * Source dimension when the trigger was a judge span (e.g. `'format'`,\n   * `'safety'`, `'correctness'`). Lets cross-template aggregators\n   * group failures by the dimension that fired without overloading\n   * `argPrefix`. Optional — clusters without this field deserialize cleanly.\n   */\n  dimension?: string\n  runCount: number\n  scenarioIds: string[]\n  exampleError?: string\n  exampleRunId: string\n}\n\nexport interface FailureClusterReport {\n  clusters: FailureCluster[]\n  totalFailures: number\n  totalRuns: number\n}\n\nexport async function failureClusterView(\n  store: TraceStore,\n  options: { rules?: FailureRule[]; minClusterSize?: number } = {},\n): Promise<FailureClusterReport> {\n  const rules = options.rules ?? DEFAULT_RULES\n  const minSize = options.minClusterSize ?? 1\n  const runs = await store.listRuns()\n\n  type Key = string\n  const clusters = new Map<Key, FailureCluster>()\n  let totalFailures = 0\n\n  for (const run of runs) {\n    if (run.status === 'completed' && run.outcome?.pass !== false) continue\n    totalFailures++\n    const spans = await store.spans({ runId: run.runId })\n    const events = await store.events({ runId: run.runId })\n    const cls = classifyFailure({ run, spans, events }, rules)\n\n    let toolName: string | undefined\n    let argPrefix: string | undefined\n    let dimension: string | undefined\n    if (cls.triggerSpanId) {\n      const trig = spans.find((s) => s.spanId === cls.triggerSpanId)\n      if (trig?.kind === 'tool') {\n        toolName = trig.toolName\n        if (hasCapturedToolArgs(trig)) argPrefix = argHash(trig.args).slice(0, 16)\n      } else if (trig?.kind === 'judge') {\n        dimension = trig.dimension\n      }\n    }\n    // Fallback: look at the last errored tool span\n    if (!toolName) {\n      const ts = await toolSpans(store, run.runId)\n      const errored = ts.filter((t) => t.status === 'error').pop()\n      if (errored) {\n        toolName = errored.toolName\n        if (hasCapturedToolArgs(errored)) argPrefix = argHash(errored.args).slice(0, 16)\n      }\n    }\n    // Secondary signal: any judge span on the failed run carries a\n    // dimension. Useful when the rule classified by judge score but\n    // didn't surface the trigger span (or surfaced a non-judge span).\n    if (!dimension) {\n      const judge = spans.find((s) => s.kind === 'judge' && typeof s.dimension === 'string')\n      if (judge?.kind === 'judge') dimension = judge.dimension\n    }\n\n    const key = `${cls.failureClass}|${toolName ?? ''}|${argPrefix ?? ''}|${dimension ?? ''}`\n    let cluster = clusters.get(key)\n    if (!cluster) {\n      cluster = {\n        failureClass: cls.failureClass,\n        toolName,\n        argPrefix,\n        dimension,\n        runCount: 0,\n        scenarioIds: [],\n        exampleRunId: run.runId,\n        exampleError: firstErrorMessage(spans) ?? cls.reason,\n      }\n      clusters.set(key, cluster)\n    }\n    cluster.runCount++\n    if (!cluster.scenarioIds.includes(run.scenarioId)) cluster.scenarioIds.push(run.scenarioId)\n  }\n\n  const arr = [...clusters.values()]\n    .filter((c) => c.runCount >= minSize)\n    .sort((a, b) => b.runCount - a.runCount)\n\n  return { clusters: arr, totalFailures, totalRuns: runs.length }\n}\n\nfunction firstErrorMessage(spans: Span[]): string | undefined {\n  const errored = spans.find((s) => s.status === 'error')\n  return errored?.error\n}\n","/**\n * FirstDivergenceView — aligns two trajectories by step index, reports\n * the first step where they differ.\n *\n * \"Differ\" is configurable — default is (kind, toolName if tool, model\n * if llm). Use this view to attribute \"why is variant B better?\" to a\n * specific step rather than an aggregate mean delta.\n */\n\nimport type { TraceStore } from '../trace/store'\nimport { buildTrajectory, type Trajectory, type TrajectoryStep } from '../trajectory'\n\nexport interface DivergenceReport {\n  runA: string\n  runB: string\n  firstDivergenceIndex: number | null\n  aStep?: TrajectoryStep\n  bStep?: TrajectoryStep\n  reason?: string\n  /** Common prefix length (steps that matched). */\n  commonPrefixLen: number\n}\n\nexport interface DivergenceOptions {\n  /** Returns true if two steps are considered equal. Default: kind + tool/model match. */\n  stepEquals?: (a: TrajectoryStep, b: TrajectoryStep) => boolean\n}\n\nexport async function firstDivergenceView(\n  store: TraceStore,\n  runA: string,\n  runB: string,\n  options: DivergenceOptions = {},\n): Promise<DivergenceReport> {\n  const [a, b] = await Promise.all([buildTrajectory(store, runA), buildTrajectory(store, runB)])\n  const eq = options.stepEquals ?? defaultStepEquals\n  const minLen = Math.min(a.steps.length, b.steps.length)\n  for (let i = 0; i < minLen; i++) {\n    const aStep = a.steps[i]!\n    const bStep = b.steps[i]!\n    if (!eq(aStep, bStep)) {\n      return {\n        runA,\n        runB,\n        firstDivergenceIndex: i,\n        aStep,\n        bStep,\n        reason: describeDifference(aStep, bStep),\n        commonPrefixLen: i,\n      }\n    }\n  }\n  if (a.steps.length === b.steps.length) {\n    return { runA, runB, firstDivergenceIndex: null, commonPrefixLen: minLen }\n  }\n  const longer: Trajectory = a.steps.length > b.steps.length ? a : b\n  const extra = longer.steps.length - minLen\n  // minLen === 0 means one trajectory is empty: divergence is the first step\n  // itself (index 0), and the absent side has no step to surface.\n  const reason =\n    minLen === 0\n      ? `one trajectory is empty; the other has ${extra} step(s) starting at index 0`\n      : `one trajectory has ${extra} more step(s) after index ${minLen - 1}`\n  return {\n    runA,\n    runB,\n    firstDivergenceIndex: minLen,\n    aStep: a.steps[minLen],\n    bStep: b.steps[minLen],\n    reason,\n    commonPrefixLen: minLen,\n  }\n}\n\nfunction defaultStepEquals(a: TrajectoryStep, b: TrajectoryStep): boolean {\n  if (a.span.kind !== b.span.kind) return false\n  if (a.span.kind === 'tool' && b.span.kind === 'tool') return a.span.toolName === b.span.toolName\n  if (a.span.kind === 'llm' && b.span.kind === 'llm') return a.span.model === b.span.model\n  if (a.span.kind === 'judge' && b.span.kind === 'judge')\n    return a.span.dimension === b.span.dimension\n  return a.span.name === b.span.name\n}\n\nfunction describeDifference(a: TrajectoryStep, b: TrajectoryStep): string {\n  if (a.span.kind !== b.span.kind) return `kind ${a.span.kind} vs ${b.span.kind}`\n  if (a.span.kind === 'tool' && b.span.kind === 'tool' && a.span.toolName !== b.span.toolName) {\n    return `tool ${a.span.toolName} vs ${b.span.toolName}`\n  }\n  if (a.span.kind === 'llm' && b.span.kind === 'llm' && a.span.model !== b.span.model) {\n    return `model ${a.span.model} vs ${b.span.model}`\n  }\n  return `name \"${a.span.name}\" vs \"${b.span.name}\"`\n}\n","/**\n * JudgeAgreementView — pairwise agreement between judges across the\n * corpus, grouped by dimension.\n *\n * Output drives two workflows:\n *   - Judge robustness audit: \"does Claude agree with GPT at κ ≥ 0.6?\"\n *   - Calibration tracking: κ vs golden human labels over time (by\n *     providing a `humanGoldenJudgeId`).\n */\n\nimport { interRaterReliability, pearsonR } from '../statistics'\nimport type { JudgeSpan } from '../trace/schema'\nimport type { TraceStore } from '../trace/store'\n\nexport interface JudgePair {\n  judgeA: string\n  judgeB: string\n  dimension: string\n  /** Number of (targetSpanId, dimension) tuples both judges scored. */\n  commonItems: number\n  pearson: number\n  krippendorff: number\n}\n\nexport interface JudgeAgreementReport {\n  pairs: JudgePair[]\n  dimensions: string[]\n  judgeIds: string[]\n}\n\nexport async function judgeAgreementView(store: TraceStore): Promise<JudgeAgreementReport> {\n  const all = (await store.spans({ kind: 'judge' })).filter(\n    (s): s is JudgeSpan => s.kind === 'judge',\n  )\n  if (all.length === 0) return { pairs: [], dimensions: [], judgeIds: [] }\n\n  const byDimension = new Map<string, JudgeSpan[]>()\n  for (const s of all) {\n    const arr = byDimension.get(s.dimension) ?? []\n    arr.push(s)\n    byDimension.set(s.dimension, arr)\n  }\n\n  const judgeIds = [...new Set(all.map((s) => s.judgeId))].sort()\n  const pairs: JudgePair[] = []\n  for (const [dim, spans] of byDimension) {\n    const byJudge = new Map<string, Map<string, number>>()\n    for (const s of spans) {\n      const m = byJudge.get(s.judgeId) ?? new Map<string, number>()\n      m.set(s.targetSpanId, s.score)\n      byJudge.set(s.judgeId, m)\n    }\n    const judgesHere = [...byJudge.keys()]\n    for (let i = 0; i < judgesHere.length; i++) {\n      for (let j = i + 1; j < judgesHere.length; j++) {\n        const judgeI = judgesHere[i]!\n        const judgeJ = judgesHere[j]!\n        const a = byJudge.get(judgeI)!\n        const b = byJudge.get(judgeJ)!\n        const common: Array<[number, number]> = []\n        for (const [target, scoreA] of a) {\n          const scoreB = b.get(target)\n          if (scoreB !== undefined) common.push([scoreA, scoreB])\n        }\n        if (common.length < 2) continue\n        const judgeScores = common.map(\n          ([scoreA, scoreB]) =>\n            [\n              { judgeName: judgeI, dimension: dim, score: scoreA, reasoning: '' },\n              { judgeName: judgeJ, dimension: dim, score: scoreB, reasoning: '' },\n            ] as const,\n        )\n        const k = interRaterReliability(\n          judgeScores[0]!.map((_, k2) => judgeScores.map((pair) => pair[k2]!)),\n        )\n        pairs.push({\n          judgeA: judgeI,\n          judgeB: judgeJ,\n          dimension: dim,\n          commonItems: common.length,\n          pearson: pearsonR(\n            common.map((c) => c[0]),\n            common.map((c) => c[1]),\n          ),\n          krippendorff: k,\n        })\n      }\n    }\n  }\n\n  return {\n    pairs: pairs.sort((a, b) => b.commonItems - a.commonItems),\n    dimensions: [...byDimension.keys()].sort(),\n    judgeIds,\n  }\n}\n","/**\n * RegressionView — compares a candidate slice to a baseline slice on a\n * named metric. Delegates the statistics (Welch's t-test, Cohen's d,\n * IQR stability) to `baseline.ts`.\n *\n * This is the entry point for CI regression gates: \"given runs tagged\n * release=A and release=B, did any metric regress?\"\n */\n\nimport { type BaselineOptions, type BaselineReport, compareToBaseline } from '../baseline'\nimport { aggregateLlm, llmSpans, runFailureClass } from '../trace/query'\nimport type { Run } from '../trace/schema'\nimport type { RunFilter, TraceStore } from '../trace/store'\n\nexport interface RegressionSpec {\n  metric: string\n  higherIsBetter: boolean\n  /** Extract a scalar from a run. Default extractors handle common metrics. */\n  extract?: (run: Run, store: TraceStore) => Promise<number | null>\n}\n\nexport interface RegressionOptions extends BaselineOptions {\n  baseline: RunFilter\n  candidate: RunFilter\n}\n\nexport async function regressionView(\n  store: TraceStore,\n  metrics: RegressionSpec[],\n  options: RegressionOptions,\n): Promise<BaselineReport> {\n  const baselineRuns = await store.listRuns(options.baseline)\n  const candidateRuns = await store.listRuns(options.candidate)\n  const samples = await Promise.all(\n    metrics.map(async (m) => {\n      const extract = m.extract ?? defaultExtract(m.metric)\n      const baseline = await extractAll(baselineRuns, extract, store)\n      const candidate = await extractAll(candidateRuns, extract, store)\n      return { metric: m.metric, higherIsBetter: m.higherIsBetter, baseline, candidate }\n    }),\n  )\n  return compareToBaseline(samples, options)\n}\n\nasync function extractAll(\n  runs: Run[],\n  extract: (r: Run, s: TraceStore) => Promise<number | null>,\n  store: TraceStore,\n): Promise<number[]> {\n  const out: number[] = []\n  for (const r of runs) {\n    const v = await extract(r, store)\n    if (v !== null && Number.isFinite(v)) out.push(v)\n  }\n  return out\n}\n\nfunction defaultExtract(metric: string): (run: Run, store: TraceStore) => Promise<number | null> {\n  return async (run, store) => {\n    switch (metric) {\n      case 'score':\n      case 'overallScore':\n        return run.outcome?.score ?? null\n      case 'pass':\n        return run.outcome?.pass === true ? 1 : 0\n      case 'durationMs':\n        return run.endedAt && run.startedAt ? run.endedAt - run.startedAt : null\n      case 'costUsd': {\n        const llm = await llmSpans(store, run.runId)\n        return aggregateLlm(llm).costUsd\n      }\n      case 'inputTokens': {\n        const llm = await llmSpans(store, run.runId)\n        return aggregateLlm(llm).inputTokens\n      }\n      case 'outputTokens': {\n        const llm = await llmSpans(store, run.runId)\n        return aggregateLlm(llm).outputTokens\n      }\n      case 'failureClass': {\n        return runFailureClass(run) === 'success' ? 1 : 0\n      }\n      default:\n        return null\n    }\n  }\n}\n","/**\n * StuckLoopView — detects when an agent calls the same tool with the\n * same (or structurally similar) arguments ≥ N times in a short window.\n *\n * Rationale: agents that loop are the number-one production failure\n * mode on long-horizon flows. The view returns (runId, toolName,\n * argHash, occurrences, windowMs) for each detected loop plus a\n * fraction of runs affected.\n */\n\nimport { executionTrackByLane } from '../trace/execution-tracks'\nimport { argHash, hasCapturedToolArgs } from '../trace/query'\nimport { isToolSpan, type Span, type ToolSpan } from '../trace/schema'\nimport type { TraceStore } from '../trace/store'\n\nconst DEFAULT_MAX_WINDOW_MS = 60_000\nconst DEFAULT_MAX_INTERVENING_TOOL_CALLS = 0\n\ninterface ScopedToolCall {\n  span: ToolSpan\n  scopeSpanId: string | null\n  laneSpanId: string | null\n}\n\ninterface IndexedToolCall extends ScopedToolCall {\n  toolCallIndex: number\n  trackId: string\n}\n\nexport interface StuckLoopFinding {\n  runId: string\n  toolName: string\n  argHash: string\n  /** Calls in this episode's densest qualifying interval, not the whole-run total. */\n  occurrences: number\n  spanIds: string[]\n  /** Nearest agent ancestor, or the direct parent when ancestry is incomplete. */\n  scopeSpanId?: string\n  /** Milliseconds between first and last call in the loop. */\n  windowMs: number\n}\n\nexport interface StuckLoopReport {\n  findings: StuckLoopFinding[]\n  affectedRunRatio: number\n  totalRuns: number\n}\n\nexport interface StuckLoopOptions {\n  /** Minimum call count to flag a loop (default 3). */\n  minOccurrences?: number\n  /** Maximum time between the first and last repeated call (default 60 seconds). */\n  maxWindowMs?: number\n  /**\n   * Maximum other tool calls allowed between adjacent repeats (default 0).\n   * Set to 1 to detect alternating patterns such as A,B,A,B,A.\n   */\n  maxInterveningToolCalls?: number\n  /** Filter to a specific runId; omit to scan the entire corpus. */\n  runId?: string\n}\n\nexport async function stuckLoopView(\n  store: TraceStore,\n  options: StuckLoopOptions = {},\n): Promise<StuckLoopReport> {\n  const minOccurrences = options.minOccurrences ?? 3\n  const maxWindowMs = options.maxWindowMs ?? DEFAULT_MAX_WINDOW_MS\n  const maxInterveningToolCalls =\n    options.maxInterveningToolCalls ?? DEFAULT_MAX_INTERVENING_TOOL_CALLS\n  if (!Number.isInteger(minOccurrences) || minOccurrences < 1) {\n    throw new RangeError('minOccurrences must be a positive integer')\n  }\n  if (!Number.isFinite(maxWindowMs) || maxWindowMs < 0) {\n    throw new RangeError('maxWindowMs must be a finite non-negative number')\n  }\n  if (!Number.isInteger(maxInterveningToolCalls) || maxInterveningToolCalls < 0) {\n    throw new RangeError('maxInterveningToolCalls must be a non-negative integer')\n  }\n  const runs = options.runId\n    ? [{ runId: options.runId }]\n    : (await store.listRuns()).map((r) => ({ runId: r.runId }))\n\n  const findings: StuckLoopFinding[] = []\n  for (const { runId } of runs) {\n    const spans = await store.spans({ runId })\n    const spansById = new Map(spans.map((span) => [span.spanId, span]))\n    const scopedTools: ScopedToolCall[] = spans\n      .filter(isToolSpan)\n      .map((span, sourceIndex) => ({ span, sourceIndex }))\n      .sort((a, b) => a.span.startedAt - b.span.startedAt || a.sourceIndex - b.sourceIndex)\n      .map(({ span }) => ({ span, ...executionScope(span, spansById) }))\n    const trackByLane = executionTrackByLane(\n      scopedTools.map((call) => {\n        const direct = call.laneSpanId === null || call.laneSpanId === call.scopeSpanId\n        const timed = direct\n          ? call.span\n          : call.laneSpanId\n            ? spansById.get(call.laneSpanId)\n            : undefined\n        return {\n          key: executionKey(call),\n          scopeKey: JSON.stringify(call.scopeSpanId),\n          start: timed?.startedAt ?? null,\n          end: timed?.endedAt ?? null,\n        }\n      }),\n    )\n    const nextToolIndexByTrack = new Map<string, number>()\n    const orderedTools: IndexedToolCall[] = scopedTools.map((call) => {\n      const trackId = trackByLane.get(executionKey(call))!\n      const toolCallIndex = nextToolIndexByTrack.get(trackId) ?? 0\n      nextToolIndexByTrack.set(trackId, toolCallIndex + 1)\n      return { ...call, toolCallIndex, trackId }\n    })\n    const byKey = new Map<\n      string,\n      {\n        calls: IndexedToolCall[]\n        argHash: string\n        toolName: string\n        scopeSpanId: string | null\n      }\n    >()\n    for (const call of orderedTools) {\n      if (!hasCapturedToolArgs(call.span)) continue\n      const h = argHash(call.span.args)\n      const key = JSON.stringify([call.trackId, call.span.toolName, h])\n      const bucket = byKey.get(key) ?? {\n        calls: [],\n        argHash: h,\n        toolName: call.span.toolName,\n        scopeSpanId: call.scopeSpanId,\n      }\n      bucket.calls.push(call)\n      byKey.set(key, bucket)\n    }\n    for (const { calls, argHash: h, toolName, scopeSpanId } of byKey.values()) {\n      if (calls.length < minOccurrences) continue\n      let episodeStart = 0\n      for (let episodeEnd = 1; episodeEnd <= calls.length; episodeEnd += 1) {\n        const previous = calls[episodeEnd - 1]!\n        const next = calls[episodeEnd]\n        const episodeEnded =\n          next === undefined ||\n          next.span.startedAt - previous.span.startedAt > maxWindowMs ||\n          next.toolCallIndex - previous.toolCallIndex - 1 > maxInterveningToolCalls ||\n          !callsAreSerial(previous, next)\n        if (!episodeEnded) continue\n\n        const episode = calls.slice(episodeStart, episodeEnd)\n        let left = 0\n        let bestStart = 0\n        let bestEnd = -1\n        for (let right = 0; right < episode.length; right += 1) {\n          while (episode[right]!.span.startedAt - episode[left]!.span.startedAt > maxWindowMs) {\n            left += 1\n          }\n          if (right - left > bestEnd - bestStart) {\n            bestStart = left\n            bestEnd = right\n          }\n        }\n        if (bestEnd - bestStart + 1 >= minOccurrences) {\n          const loop = episode.slice(bestStart, bestEnd + 1)\n          const first = loop[0]!.span.startedAt\n          const last = loop[loop.length - 1]!.span.startedAt\n          findings.push({\n            runId,\n            toolName,\n            argHash: h,\n            occurrences: loop.length,\n            spanIds: loop.map((call) => call.span.spanId),\n            ...(scopeSpanId ? { scopeSpanId } : {}),\n            windowMs: last - first,\n          })\n        }\n        episodeStart = episodeEnd\n      }\n    }\n  }\n\n  const affectedRuns = new Set(findings.map((f) => f.runId))\n  return {\n    findings,\n    affectedRunRatio: runs.length > 0 ? affectedRuns.size / runs.length : 0,\n    totalRuns: runs.length,\n  }\n}\n\nfunction laneKey(call: Pick<ScopedToolCall, 'scopeSpanId' | 'laneSpanId'>): string {\n  return JSON.stringify([call.scopeSpanId, call.laneSpanId])\n}\n\nfunction executionKey(call: ScopedToolCall): string {\n  return call.laneSpanId === null || call.laneSpanId === call.scopeSpanId\n    ? JSON.stringify([call.scopeSpanId, call.span.spanId])\n    : laneKey(call)\n}\n\nfunction executionScope(\n  span: ToolSpan,\n  spansById: ReadonlyMap<string, Span>,\n): { scopeSpanId: string | null; laneSpanId: string | null } {\n  const directParent = span.parentSpanId\n  if (!directParent) return { scopeSpanId: null, laneSpanId: null }\n\n  let currentId: string | undefined = directParent\n  let laneSpanId: string | null = null\n  const seen = new Set<string>()\n  while (currentId && !seen.has(currentId)) {\n    seen.add(currentId)\n    const current = spansById.get(currentId)\n    if (!current) return { scopeSpanId: directParent, laneSpanId: directParent }\n    if (current.kind === 'agent') {\n      return { scopeSpanId: current.spanId, laneSpanId: laneSpanId ?? current.spanId }\n    }\n    laneSpanId = current.spanId\n    currentId = current.parentSpanId\n  }\n  return { scopeSpanId: directParent, laneSpanId: directParent }\n}\n\nfunction callsAreSerial(previous: IndexedToolCall, next: IndexedToolCall): boolean {\n  return previous.span.endedAt !== undefined && previous.span.endedAt <= next.span.startedAt\n}\n","/**\n * ToolWasteView — fraction of tool calls whose results weren't used\n * downstream. Without a \"used\" signal we fall back to structural\n * proxies: error calls, duplicate calls, and tool calls followed by\n * zero subsequent LLM spans are all considered waste.\n *\n * Consumers can pass a `usageOracle` that inspects a tool span and\n * returns true iff the tool's result appears in a later LLM message,\n * artifact, or state mutation — that's the canonical definition; the\n * default heuristic is a reasonable fallback.\n */\n\nimport { computeToolUseMetrics } from '../tool-use-metrics'\nimport { llmSpans, toolSpans } from '../trace/query'\nimport type { LlmSpan, ToolSpan } from '../trace/schema'\nimport type { TraceStore } from '../trace/store'\n\nexport interface ToolWasteFinding {\n  runId: string\n  wastedCalls: number\n  totalCalls: number\n  wasteRate: number\n}\n\nexport interface ToolWasteReport {\n  byRun: ToolWasteFinding[]\n  overallWasteRate: number\n}\n\nexport interface ToolWasteOptions {\n  runId?: string\n  usageOracle?: (tool: ToolSpan, later: { llm: Awaited<ReturnType<typeof llmSpans>> }) => boolean\n}\n\nexport async function toolWasteView(\n  store: TraceStore,\n  options: ToolWasteOptions = {},\n): Promise<ToolWasteReport> {\n  const runs = options.runId ? [options.runId] : (await store.listRuns()).map((r) => r.runId)\n\n  const byRun: ToolWasteFinding[] = []\n  let totalCalls = 0\n  let totalWasted = 0\n  for (const runId of runs) {\n    const tools = await toolSpans(store, runId)\n    if (tools.length === 0) {\n      byRun.push({ runId, wastedCalls: 0, totalCalls: 0, wasteRate: 0 })\n      continue\n    }\n    const llms = await llmSpans(store, runId)\n    // Sort LLM spans once by start time, then build a suffix index of the\n    // concatenated message text. `suffixText[i]` is the haystack of every\n    // string message content in spans[i..]. Per tool we binary-search the\n    // first span started strictly after the tool, then test that one suffix\n    // — turning the per-tool O(llms × messages × content) scan into a single\n    // O(log llms) lookup over precomputed text.\n    const sortedLlm = [...llms].sort((a, b) => a.startedAt - b.startedAt)\n    const startTimes = sortedLlm.map((l) => l.startedAt)\n    const suffixText = buildSuffixText(sortedLlm)\n    let wasted = 0\n    for (const t of tools) {\n      if (t.status === 'error') {\n        wasted++\n        continue\n      }\n      // First LLM span started strictly after this tool (upper-bound search).\n      const cutoff = upperBound(startTimes, t.startedAt)\n      if (options.usageOracle) {\n        if (!options.usageOracle(t, { llm: sortedLlm.slice(cutoff) })) wasted++\n      } else {\n        // Default heuristic: a tool whose result is NOT mentioned in any\n        // later LLM input message is likely wasted. An empty/null result has\n        // no payload to propagate downstream — there is nothing to find in a\n        // later message, so it is not evidence of waste; skip it.\n        const resultStr = stringify(t.result)\n        if (resultStr === '') continue\n        const haystack = suffixText[cutoff] ?? ''\n        const used = haystack.includes(resultStr.slice(0, 120))\n        if (!used) wasted++\n      }\n    }\n    const wasteRate = wasted / tools.length\n    byRun.push({ runId, wastedCalls: wasted, totalCalls: tools.length, wasteRate })\n    totalCalls += tools.length\n    totalWasted += wasted\n  }\n  return { byRun, overallWasteRate: totalCalls > 0 ? totalWasted / totalCalls : 0 }\n}\n\n/**\n * Build per-position suffix haystacks: result[i] is the concatenation of every\n * string message content in spans[i..end]. Built back-to-front so each entry\n * reuses the next one — O(total message text) rather than O(spans²).\n */\nfunction buildSuffixText(spans: LlmSpan[]): string[] {\n  const result = new Array<string>(spans.length + 1)\n  result[spans.length] = ''\n  for (let i = spans.length - 1; i >= 0; i--) {\n    const own = spans[i]!.messages.map((m) =>\n      typeof m.content === 'string' ? m.content : '',\n    ).join('\\n')\n    result[i] = `${own}\\n${result[i + 1]}`\n  }\n  return result\n}\n\n/** Index of the first element strictly greater than `target` in a sorted array. */\nfunction upperBound(sorted: number[], target: number): number {\n  let lo = 0\n  let hi = sorted.length\n  while (lo < hi) {\n    const mid = (lo + hi) >>> 1\n    if (sorted[mid]! <= target) lo = mid + 1\n    else hi = mid\n  }\n  return lo\n}\n\nfunction stringify(v: unknown): string {\n  if (v === null || v === undefined) return ''\n  if (typeof v === 'string') return v\n  try {\n    return JSON.stringify(v)\n  } catch {\n    return String(v)\n  }\n}\n\n// Re-export for convenience in consumers that want both descriptive and usage metrics.\nexport { computeToolUseMetrics }\n"],"mappings":";;;;;;;;AA8BA,eAAsB,iBACpB,OACA,UAAuD,CAAC,GAC3B;CAC7B,MAAM,OAAO,MAAM,MAAM,SAAS;EAChC,YAAY,QAAQ;EACpB,WAAW,QAAQ;CACrB,CAAC;CACD,MAAM,WAAkC,CAAC;CACzC,MAAM,cAAsC,CAAC;CAC7C,MAAM,aAAqC,CAAC;CAC5C,MAAM,YAAoC,CAAC;CAE3C,KAAK,MAAM,OAAO,MAAM;EACtB,MAAM,UAAU,MAAM,MAAM,OAAO,IAAI,KAAK;EAC5C,KAAK,MAAM,KAAK,SAAS;GACvB,IAAI,CAAC,EAAE,UAAU;GACjB,MAAM,cAAc,EAAE,QAAQ,IAAI,EAAE,WAAW,EAAE,QAAQ;GACzD,SAAS,KAAK;IACZ,OAAO,IAAI;IACX,YAAY,IAAI;IAChB,WAAW,IAAI;IACf,WAAW,EAAE;IACb,OAAO,EAAE;IACT,UAAU,EAAE;IACZ;IACA,WAAW,EAAE;GACf,CAAC;GACD,YAAY,EAAE,cAAc,YAAY,EAAE,cAAc,KAAK;GAC7D,WAAW,IAAI,eAAe,WAAW,IAAI,eAAe,KAAK;GACjE,IAAI,IAAI,WAAW,UAAU,IAAI,cAAc,UAAU,IAAI,cAAc,KAAK;EAClF;CACF;CAEA,MAAM,eAAe,IAAI,IAAI,SAAS,KAAK,MAAM,EAAE,KAAK,CAAC;CACzD,OAAO;EACL;EACA;EACA;EACA;EACA,WAAW,KAAK;EAChB,kBAAkB,KAAK,SAAS,IAAI,aAAa,OAAO,KAAK,SAAS;CACxE;AACF;;;;;;;;;;ACnCA,eAAsB,mBACpB,OACA,UAA8D,CAAC,GAChC;CAC/B,MAAM,QAAQ,QAAQ,SAAS;CAC/B,MAAM,UAAU,QAAQ,kBAAkB;CAC1C,MAAM,OAAO,MAAM,MAAM,SAAS;CAGlC,MAAM,2BAAW,IAAI,IAAyB;CAC9C,IAAI,gBAAgB;CAEpB,KAAK,MAAM,OAAO,MAAM;EACtB,IAAI,IAAI,WAAW,eAAe,IAAI,SAAS,SAAS,OAAO;EAC/D;EACA,MAAM,QAAQ,MAAM,MAAM,MAAM,EAAE,OAAO,IAAI,MAAM,CAAC;EAEpD,MAAM,MAAM,gBAAgB;GAAE;GAAK;GAAO,QAAA,MADrB,MAAM,OAAO,EAAE,OAAO,IAAI,MAAM,CAAC;EACL,GAAG,KAAK;EAEzD,IAAI;EACJ,IAAI;EACJ,IAAI;EACJ,IAAI,IAAI,eAAe;GACrB,MAAM,OAAO,MAAM,MAAM,MAAM,EAAE,WAAW,IAAI,aAAa;GAC7D,IAAI,MAAM,SAAS,QAAQ;IACzB,WAAW,KAAK;IAChB,IAAI,oBAAoB,IAAI,GAAG,YAAY,QAAQ,KAAK,IAAI,CAAC,CAAC,MAAM,GAAG,EAAE;GAC3E,OAAO,IAAI,MAAM,SAAS,SACxB,YAAY,KAAK;EAErB;EAEA,IAAI,CAAC,UAAU;GAEb,MAAM,WAAU,MADC,UAAU,OAAO,IAAI,KAAK,EAAA,CACxB,QAAQ,MAAM,EAAE,WAAW,OAAO,CAAC,CAAC,IAAI;GAC3D,IAAI,SAAS;IACX,WAAW,QAAQ;IACnB,IAAI,oBAAoB,OAAO,GAAG,YAAY,QAAQ,QAAQ,IAAI,CAAC,CAAC,MAAM,GAAG,EAAE;GACjF;EACF;EAIA,IAAI,CAAC,WAAW;GACd,MAAM,QAAQ,MAAM,MAAM,MAAM,EAAE,SAAS,WAAW,OAAO,EAAE,cAAc,QAAQ;GACrF,IAAI,OAAO,SAAS,SAAS,YAAY,MAAM;EACjD;EAEA,MAAM,MAAM,GAAG,IAAI,aAAa,GAAG,YAAY,GAAG,GAAG,aAAa,GAAG,GAAG,aAAa;EACrF,IAAI,UAAU,SAAS,IAAI,GAAG;EAC9B,IAAI,CAAC,SAAS;GACZ,UAAU;IACR,cAAc,IAAI;IAClB;IACA;IACA;IACA,UAAU;IACV,aAAa,CAAC;IACd,cAAc,IAAI;IAClB,cAAc,kBAAkB,KAAK,KAAK,IAAI;GAChD;GACA,SAAS,IAAI,KAAK,OAAO;EAC3B;EACA,QAAQ;EACR,IAAI,CAAC,QAAQ,YAAY,SAAS,IAAI,UAAU,GAAG,QAAQ,YAAY,KAAK,IAAI,UAAU;CAC5F;CAMA,OAAO;EAAE,UAJG,CAAC,GAAG,SAAS,OAAO,CAAC,CAAC,CAC/B,QAAQ,MAAM,EAAE,YAAY,OAAO,CAAC,CACpC,MAAM,GAAG,MAAM,EAAE,WAAW,EAAE,QAEZ;EAAG;EAAe,WAAW,KAAK;CAAO;AAChE;AAEA,SAAS,kBAAkB,OAAmC;CAE5D,OADgB,MAAM,MAAM,MAAM,EAAE,WAAW,OAClC,CAAC,EAAE;AAClB;;;ACvFA,eAAsB,oBACpB,OACA,MACA,MACA,UAA6B,CAAC,GACH;CAC3B,MAAM,CAAC,GAAG,KAAK,MAAM,QAAQ,IAAI,CAAC,gBAAgB,OAAO,IAAI,GAAG,gBAAgB,OAAO,IAAI,CAAC,CAAC;CAC7F,MAAM,KAAK,QAAQ,cAAc;CACjC,MAAM,SAAS,KAAK,IAAI,EAAE,MAAM,QAAQ,EAAE,MAAM,MAAM;CACtD,KAAK,IAAI,IAAI,GAAG,IAAI,QAAQ,KAAK;EAC/B,MAAM,QAAQ,EAAE,MAAM;EACtB,MAAM,QAAQ,EAAE,MAAM;EACtB,IAAI,CAAC,GAAG,OAAO,KAAK,GAClB,OAAO;GACL;GACA;GACA,sBAAsB;GACtB;GACA;GACA,QAAQ,mBAAmB,OAAO,KAAK;GACvC,iBAAiB;EACnB;CAEJ;CACA,IAAI,EAAE,MAAM,WAAW,EAAE,MAAM,QAC7B,OAAO;EAAE;EAAM;EAAM,sBAAsB;EAAM,iBAAiB;CAAO;CAG3E,MAAM,SADqB,EAAE,MAAM,SAAS,EAAE,MAAM,SAAS,IAAI,EAAA,CAC5C,MAAM,SAAS;CAGpC,MAAM,SACJ,WAAW,IACP,0CAA0C,MAAM,gCAChD,sBAAsB,MAAM,4BAA4B,SAAS;CACvE,OAAO;EACL;EACA;EACA,sBAAsB;EACtB,OAAO,EAAE,MAAM;EACf,OAAO,EAAE,MAAM;EACf;EACA,iBAAiB;CACnB;AACF;AAEA,SAAS,kBAAkB,GAAmB,GAA4B;CACxE,IAAI,EAAE,KAAK,SAAS,EAAE,KAAK,MAAM,OAAO;CACxC,IAAI,EAAE,KAAK,SAAS,UAAU,EAAE,KAAK,SAAS,QAAQ,OAAO,EAAE,KAAK,aAAa,EAAE,KAAK;CACxF,IAAI,EAAE,KAAK,SAAS,SAAS,EAAE,KAAK,SAAS,OAAO,OAAO,EAAE,KAAK,UAAU,EAAE,KAAK;CACnF,IAAI,EAAE,KAAK,SAAS,WAAW,EAAE,KAAK,SAAS,SAC7C,OAAO,EAAE,KAAK,cAAc,EAAE,KAAK;CACrC,OAAO,EAAE,KAAK,SAAS,EAAE,KAAK;AAChC;AAEA,SAAS,mBAAmB,GAAmB,GAA2B;CACxE,IAAI,EAAE,KAAK,SAAS,EAAE,KAAK,MAAM,OAAO,QAAQ,EAAE,KAAK,KAAK,MAAM,EAAE,KAAK;CACzE,IAAI,EAAE,KAAK,SAAS,UAAU,EAAE,KAAK,SAAS,UAAU,EAAE,KAAK,aAAa,EAAE,KAAK,UACjF,OAAO,QAAQ,EAAE,KAAK,SAAS,MAAM,EAAE,KAAK;CAE9C,IAAI,EAAE,KAAK,SAAS,SAAS,EAAE,KAAK,SAAS,SAAS,EAAE,KAAK,UAAU,EAAE,KAAK,OAC5E,OAAO,SAAS,EAAE,KAAK,MAAM,MAAM,EAAE,KAAK;CAE5C,OAAO,SAAS,EAAE,KAAK,KAAK,QAAQ,EAAE,KAAK,KAAK;AAClD;;;;;;;;;;;;AC9DA,eAAsB,mBAAmB,OAAkD;CACzF,MAAM,OAAO,MAAM,MAAM,MAAM,EAAE,MAAM,QAAQ,CAAC,EAAA,CAAG,QAChD,MAAsB,EAAE,SAAS,OACpC;CACA,IAAI,IAAI,WAAW,GAAG,OAAO;EAAE,OAAO,CAAC;EAAG,YAAY,CAAC;EAAG,UAAU,CAAC;CAAE;CAEvE,MAAM,8BAAc,IAAI,IAAyB;CACjD,KAAK,MAAM,KAAK,KAAK;EACnB,MAAM,MAAM,YAAY,IAAI,EAAE,SAAS,KAAK,CAAC;EAC7C,IAAI,KAAK,CAAC;EACV,YAAY,IAAI,EAAE,WAAW,GAAG;CAClC;CAEA,MAAM,WAAW,CAAC,GAAG,IAAI,IAAI,IAAI,KAAK,MAAM,EAAE,OAAO,CAAC,CAAC,CAAC,CAAC,KAAK;CAC9D,MAAM,QAAqB,CAAC;CAC5B,KAAK,MAAM,CAAC,KAAK,UAAU,aAAa;EACtC,MAAM,0BAAU,IAAI,IAAiC;EACrD,KAAK,MAAM,KAAK,OAAO;GACrB,MAAM,IAAI,QAAQ,IAAI,EAAE,OAAO,qBAAK,IAAI,IAAoB;GAC5D,EAAE,IAAI,EAAE,cAAc,EAAE,KAAK;GAC7B,QAAQ,IAAI,EAAE,SAAS,CAAC;EAC1B;EACA,MAAM,aAAa,CAAC,GAAG,QAAQ,KAAK,CAAC;EACrC,KAAK,IAAI,IAAI,GAAG,IAAI,WAAW,QAAQ,KACrC,KAAK,IAAI,IAAI,IAAI,GAAG,IAAI,WAAW,QAAQ,KAAK;GAC9C,MAAM,SAAS,WAAW;GAC1B,MAAM,SAAS,WAAW;GAC1B,MAAM,IAAI,QAAQ,IAAI,MAAM;GAC5B,MAAM,IAAI,QAAQ,IAAI,MAAM;GAC5B,MAAM,SAAkC,CAAC;GACzC,KAAK,MAAM,CAAC,QAAQ,WAAW,GAAG;IAChC,MAAM,SAAS,EAAE,IAAI,MAAM;IAC3B,IAAI,WAAW,KAAA,GAAW,OAAO,KAAK,CAAC,QAAQ,MAAM,CAAC;GACxD;GACA,IAAI,OAAO,SAAS,GAAG;GACvB,MAAM,cAAc,OAAO,KACxB,CAAC,QAAQ,YACR,CACE;IAAE,WAAW;IAAQ,WAAW;IAAK,OAAO;IAAQ,WAAW;GAAG,GAClE;IAAE,WAAW;IAAQ,WAAW;IAAK,OAAO;IAAQ,WAAW;GAAG,CACpE,CACJ;GACA,MAAM,IAAI,sBACR,YAAY,EAAE,CAAE,KAAK,GAAG,OAAO,YAAY,KAAK,SAAS,KAAK,GAAI,CAAC,CACrE;GACA,MAAM,KAAK;IACT,QAAQ;IACR,QAAQ;IACR,WAAW;IACX,aAAa,OAAO;IACpB,SAAS,SACP,OAAO,KAAK,MAAM,EAAE,EAAE,GACtB,OAAO,KAAK,MAAM,EAAE,EAAE,CACxB;IACA,cAAc;GAChB,CAAC;EACH;CAEJ;CAEA,OAAO;EACL,OAAO,MAAM,MAAM,GAAG,MAAM,EAAE,cAAc,EAAE,WAAW;EACzD,YAAY,CAAC,GAAG,YAAY,KAAK,CAAC,CAAC,CAAC,KAAK;EACzC;CACF;AACF;;;;;;;;;;;ACrEA,eAAsB,eACpB,OACA,SACA,SACyB;CACzB,MAAM,eAAe,MAAM,MAAM,SAAS,QAAQ,QAAQ;CAC1D,MAAM,gBAAgB,MAAM,MAAM,SAAS,QAAQ,SAAS;CAS5D,OAAO,kBAAkB,MARH,QAAQ,IAC5B,QAAQ,IAAI,OAAO,MAAM;EACvB,MAAM,UAAU,EAAE,WAAW,eAAe,EAAE,MAAM;EACpD,MAAM,WAAW,MAAM,WAAW,cAAc,SAAS,KAAK;EAC9D,MAAM,YAAY,MAAM,WAAW,eAAe,SAAS,KAAK;EAChE,OAAO;GAAE,QAAQ,EAAE;GAAQ,gBAAgB,EAAE;GAAgB;GAAU;EAAU;CACnF,CAAC,CACH,GACkC,OAAO;AAC3C;AAEA,eAAe,WACb,MACA,SACA,OACmB;CACnB,MAAM,MAAgB,CAAC;CACvB,KAAK,MAAM,KAAK,MAAM;EACpB,MAAM,IAAI,MAAM,QAAQ,GAAG,KAAK;EAChC,IAAI,MAAM,QAAQ,OAAO,SAAS,CAAC,GAAG,IAAI,KAAK,CAAC;CAClD;CACA,OAAO;AACT;AAEA,SAAS,eAAe,QAAyE;CAC/F,OAAO,OAAO,KAAK,UAAU;EAC3B,QAAQ,QAAR;GACE,KAAK;GACL,KAAK,gBACH,OAAO,IAAI,SAAS,SAAS;GAC/B,KAAK,QACH,OAAO,IAAI,SAAS,SAAS,OAAO,IAAI;GAC1C,KAAK,cACH,OAAO,IAAI,WAAW,IAAI,YAAY,IAAI,UAAU,IAAI,YAAY;GACtE,KAAK,WAEH,OAAO,aAAa,MADF,SAAS,OAAO,IAAI,KAAK,CACpB,CAAC,CAAC;GAE3B,KAAK,eAEH,OAAO,aAAa,MADF,SAAS,OAAO,IAAI,KAAK,CACpB,CAAC,CAAC;GAE3B,KAAK,gBAEH,OAAO,aAAa,MADF,SAAS,OAAO,IAAI,KAAK,CACpB,CAAC,CAAC;GAE3B,KAAK,gBACH,OAAO,gBAAgB,GAAG,MAAM,YAAY,IAAI;GAElD,SACE,OAAO;EACX;CACF;AACF;;;;;;;;;;;;ACvEA,MAAM,wBAAwB;AAC9B,MAAM,qCAAqC;AA8C3C,eAAsB,cACpB,OACA,UAA4B,CAAC,GACH;CAC1B,MAAM,iBAAiB,QAAQ,kBAAkB;CACjD,MAAM,cAAc,QAAQ,eAAe;CAC3C,MAAM,0BACJ,QAAQ,2BAA2B;CACrC,IAAI,CAAC,OAAO,UAAU,cAAc,KAAK,iBAAiB,GACxD,MAAM,IAAI,WAAW,2CAA2C;CAElE,IAAI,CAAC,OAAO,SAAS,WAAW,KAAK,cAAc,GACjD,MAAM,IAAI,WAAW,kDAAkD;CAEzE,IAAI,CAAC,OAAO,UAAU,uBAAuB,KAAK,0BAA0B,GAC1E,MAAM,IAAI,WAAW,wDAAwD;CAE/E,MAAM,OAAO,QAAQ,QACjB,CAAC,EAAE,OAAO,QAAQ,MAAM,CAAC,KACxB,MAAM,MAAM,SAAS,EAAA,CAAG,KAAK,OAAO,EAAE,OAAO,EAAE,MAAM,EAAE;CAE5D,MAAM,WAA+B,CAAC;CACtC,KAAK,MAAM,EAAE,WAAW,MAAM;EAC5B,MAAM,QAAQ,MAAM,MAAM,MAAM,EAAE,MAAM,CAAC;EACzC,MAAM,YAAY,IAAI,IAAI,MAAM,KAAK,SAAS,CAAC,KAAK,QAAQ,IAAI,CAAC,CAAC;EAClE,MAAM,cAAgC,MACnC,OAAO,UAAU,CAAC,CAClB,KAAK,MAAM,iBAAiB;GAAE;GAAM;EAAY,EAAE,CAAC,CACnD,MAAM,GAAG,MAAM,EAAE,KAAK,YAAY,EAAE,KAAK,aAAa,EAAE,cAAc,EAAE,WAAW,CAAC,CACpF,KAAK,EAAE,YAAY;GAAE;GAAM,GAAG,eAAe,MAAM,SAAS;EAAE,EAAE;EACnE,MAAM,cAAc,qBAClB,YAAY,KAAK,SAAS;GAExB,MAAM,QADS,KAAK,eAAe,QAAQ,KAAK,eAAe,KAAK,cAEhE,KAAK,OACL,KAAK,aACH,UAAU,IAAI,KAAK,UAAU,IAC7B,KAAA;GACN,OAAO;IACL,KAAK,aAAa,IAAI;IACtB,UAAU,KAAK,UAAU,KAAK,WAAW;IACzC,OAAO,OAAO,aAAa;IAC3B,KAAK,OAAO,WAAW;GACzB;EACF,CAAC,CACH;EACA,MAAM,uCAAuB,IAAI,IAAoB;EACrD,MAAM,eAAkC,YAAY,KAAK,SAAS;GAChE,MAAM,UAAU,YAAY,IAAI,aAAa,IAAI,CAAC;GAClD,MAAM,gBAAgB,qBAAqB,IAAI,OAAO,KAAK;GAC3D,qBAAqB,IAAI,SAAS,gBAAgB,CAAC;GACnD,OAAO;IAAE,GAAG;IAAM;IAAe;GAAQ;EAC3C,CAAC;EACD,MAAM,wBAAQ,IAAI,IAQhB;EACF,KAAK,MAAM,QAAQ,cAAc;GAC/B,IAAI,CAAC,oBAAoB,KAAK,IAAI,GAAG;GACrC,MAAM,IAAI,QAAQ,KAAK,KAAK,IAAI;GAChC,MAAM,MAAM,KAAK,UAAU;IAAC,KAAK;IAAS,KAAK,KAAK;IAAU;GAAC,CAAC;GAChE,MAAM,SAAS,MAAM,IAAI,GAAG,KAAK;IAC/B,OAAO,CAAC;IACR,SAAS;IACT,UAAU,KAAK,KAAK;IACpB,aAAa,KAAK;GACpB;GACA,OAAO,MAAM,KAAK,IAAI;GACtB,MAAM,IAAI,KAAK,MAAM;EACvB;EACA,KAAK,MAAM,EAAE,OAAO,SAAS,GAAG,UAAU,iBAAiB,MAAM,OAAO,GAAG;GACzE,IAAI,MAAM,SAAS,gBAAgB;GACnC,IAAI,eAAe;GACnB,KAAK,IAAI,aAAa,GAAG,cAAc,MAAM,QAAQ,cAAc,GAAG;IACpE,MAAM,WAAW,MAAM,aAAa;IACpC,MAAM,OAAO,MAAM;IAMnB,IAAI,EAJF,SAAS,KAAA,KACT,KAAK,KAAK,YAAY,SAAS,KAAK,YAAY,eAChD,KAAK,gBAAgB,SAAS,gBAAgB,IAAI,2BAClD,CAAC,eAAe,UAAU,IAAI,IACb;IAEnB,MAAM,UAAU,MAAM,MAAM,cAAc,UAAU;IACpD,IAAI,OAAO;IACX,IAAI,YAAY;IAChB,IAAI,UAAU;IACd,KAAK,IAAI,QAAQ,GAAG,QAAQ,QAAQ,QAAQ,SAAS,GAAG;KACtD,OAAO,QAAQ,MAAM,CAAE,KAAK,YAAY,QAAQ,KAAK,CAAE,KAAK,YAAY,aACtE,QAAQ;KAEV,IAAI,QAAQ,OAAO,UAAU,WAAW;MACtC,YAAY;MACZ,UAAU;KACZ;IACF;IACA,IAAI,UAAU,YAAY,KAAK,gBAAgB;KAC7C,MAAM,OAAO,QAAQ,MAAM,WAAW,UAAU,CAAC;KACjD,MAAM,QAAQ,KAAK,EAAE,CAAE,KAAK;KAC5B,MAAM,OAAO,KAAK,KAAK,SAAS,EAAE,CAAE,KAAK;KACzC,SAAS,KAAK;MACZ;MACA;MACA,SAAS;MACT,aAAa,KAAK;MAClB,SAAS,KAAK,KAAK,SAAS,KAAK,KAAK,MAAM;MAC5C,GAAI,cAAc,EAAE,YAAY,IAAI,CAAC;MACrC,UAAU,OAAO;KACnB,CAAC;IACH;IACA,eAAe;GACjB;EACF;CACF;CAEA,MAAM,eAAe,IAAI,IAAI,SAAS,KAAK,MAAM,EAAE,KAAK,CAAC;CACzD,OAAO;EACL;EACA,kBAAkB,KAAK,SAAS,IAAI,aAAa,OAAO,KAAK,SAAS;EACtE,WAAW,KAAK;CAClB;AACF;AAEA,SAAS,QAAQ,MAAkE;CACjF,OAAO,KAAK,UAAU,CAAC,KAAK,aAAa,KAAK,UAAU,CAAC;AAC3D;AAEA,SAAS,aAAa,MAA8B;CAClD,OAAO,KAAK,eAAe,QAAQ,KAAK,eAAe,KAAK,cACxD,KAAK,UAAU,CAAC,KAAK,aAAa,KAAK,KAAK,MAAM,CAAC,IACnD,QAAQ,IAAI;AAClB;AAEA,SAAS,eACP,MACA,WAC2D;CAC3D,MAAM,eAAe,KAAK;CAC1B,IAAI,CAAC,cAAc,OAAO;EAAE,aAAa;EAAM,YAAY;CAAK;CAEhE,IAAI,YAAgC;CACpC,IAAI,aAA4B;CAChC,MAAM,uBAAO,IAAI,IAAY;CAC7B,OAAO,aAAa,CAAC,KAAK,IAAI,SAAS,GAAG;EACxC,KAAK,IAAI,SAAS;EAClB,MAAM,UAAU,UAAU,IAAI,SAAS;EACvC,IAAI,CAAC,SAAS,OAAO;GAAE,aAAa;GAAc,YAAY;EAAa;EAC3E,IAAI,QAAQ,SAAS,SACnB,OAAO;GAAE,aAAa,QAAQ;GAAQ,YAAY,cAAc,QAAQ;EAAO;EAEjF,aAAa,QAAQ;EACrB,YAAY,QAAQ;CACtB;CACA,OAAO;EAAE,aAAa;EAAc,YAAY;CAAa;AAC/D;AAEA,SAAS,eAAe,UAA2B,MAAgC;CACjF,OAAO,SAAS,KAAK,YAAY,KAAA,KAAa,SAAS,KAAK,WAAW,KAAK,KAAK;AACnF;;;;;;;;;;;;;;AC/LA,eAAsB,cACpB,OACA,UAA4B,CAAC,GACH;CAC1B,MAAM,OAAO,QAAQ,QAAQ,CAAC,QAAQ,KAAK,KAAK,MAAM,MAAM,SAAS,EAAA,CAAG,KAAK,MAAM,EAAE,KAAK;CAE1F,MAAM,QAA4B,CAAC;CACnC,IAAI,aAAa;CACjB,IAAI,cAAc;CAClB,KAAK,MAAM,SAAS,MAAM;EACxB,MAAM,QAAQ,MAAM,UAAU,OAAO,KAAK;EAC1C,IAAI,MAAM,WAAW,GAAG;GACtB,MAAM,KAAK;IAAE;IAAO,aAAa;IAAG,YAAY;IAAG,WAAW;GAAE,CAAC;GACjE;EACF;EAQA,MAAM,YAAY,CAAC,GAAG,MAPH,SAAS,OAAO,KAAK,CAOd,CAAC,CAAC,MAAM,GAAG,MAAM,EAAE,YAAY,EAAE,SAAS;EACpE,MAAM,aAAa,UAAU,KAAK,MAAM,EAAE,SAAS;EACnD,MAAM,aAAa,gBAAgB,SAAS;EAC5C,IAAI,SAAS;EACb,KAAK,MAAM,KAAK,OAAO;GACrB,IAAI,EAAE,WAAW,SAAS;IACxB;IACA;GACF;GAEA,MAAM,SAAS,WAAW,YAAY,EAAE,SAAS;GACjD,IAAI,QAAQ,aACN;QAAA,CAAC,QAAQ,YAAY,GAAG,EAAE,KAAK,UAAU,MAAM,MAAM,EAAE,CAAC,GAAG;GAAA,OAC1D;IAKL,MAAM,YAAY,UAAU,EAAE,MAAM;IACpC,IAAI,cAAc,IAAI;IAGtB,IAAI,EAFa,WAAW,WAAW,GAAA,CACjB,SAAS,UAAU,MAAM,GAAG,GAAG,CAC7C,GAAG;GACb;EACF;EACA,MAAM,YAAY,SAAS,MAAM;EACjC,MAAM,KAAK;GAAE;GAAO,aAAa;GAAQ,YAAY,MAAM;GAAQ;EAAU,CAAC;EAC9E,cAAc,MAAM;EACpB,eAAe;CACjB;CACA,OAAO;EAAE;EAAO,kBAAkB,aAAa,IAAI,cAAc,aAAa;CAAE;AAClF;;;;;;AAOA,SAAS,gBAAgB,OAA4B;CACnD,MAAM,SAAS,IAAI,MAAc,MAAM,SAAS,CAAC;CACjD,OAAO,MAAM,UAAU;CACvB,KAAK,IAAI,IAAI,MAAM,SAAS,GAAG,KAAK,GAAG,KAIrC,OAAO,KAAK,GAHA,MAAM,EAAE,CAAE,SAAS,KAAK,MAClC,OAAO,EAAE,YAAY,WAAW,EAAE,UAAU,EAC9C,CAAC,CAAC,KAAK,IACU,EAAE,IAAI,OAAO,IAAI;CAEpC,OAAO;AACT;;AAGA,SAAS,WAAW,QAAkB,QAAwB;CAC5D,IAAI,KAAK;CACT,IAAI,KAAK,OAAO;CAChB,OAAO,KAAK,IAAI;EACd,MAAM,MAAO,KAAK,OAAQ;EAC1B,IAAI,OAAO,QAAS,QAAQ,KAAK,MAAM;OAClC,KAAK;CACZ;CACA,OAAO;AACT;AAEA,SAAS,UAAU,GAAoB;CACrC,IAAI,MAAM,QAAQ,MAAM,KAAA,GAAW,OAAO;CAC1C,IAAI,OAAO,MAAM,UAAU,OAAO;CAClC,IAAI;EACF,OAAO,KAAK,UAAU,CAAC;CACzB,QAAQ;EACN,OAAO,OAAO,CAAC;CACjB;AACF"}
+{"version":3,"file":"index.js","names":[],"sources":["../../src/pipelines/first-divergence.ts","../../src/pipelines/regression.ts","../../src/pipelines/stuck-loop.ts"],"sourcesContent":["/**\n * FirstDivergenceView — aligns two trajectories by step index, reports\n * the first step where they differ.\n *\n * \"Differ\" is configurable — default is (kind, toolName if tool, model\n * if llm). Use this view to attribute \"why is variant B better?\" to a\n * specific step rather than an aggregate mean delta.\n */\n\nimport type { TraceStore } from '../trace/store'\nimport { buildTrajectory, type Trajectory, type TrajectoryStep } from '../trajectory'\n\nexport interface DivergenceReport {\n  runA: string\n  runB: string\n  firstDivergenceIndex: number | null\n  aStep?: TrajectoryStep\n  bStep?: TrajectoryStep\n  reason?: string\n  /** Common prefix length (steps that matched). */\n  commonPrefixLen: number\n}\n\nexport interface DivergenceOptions {\n  /** Returns true if two steps are considered equal. Default: kind + tool/model match. */\n  stepEquals?: (a: TrajectoryStep, b: TrajectoryStep) => boolean\n}\n\nexport async function firstDivergenceView(\n  store: TraceStore,\n  runA: string,\n  runB: string,\n  options: DivergenceOptions = {},\n): Promise<DivergenceReport> {\n  const [a, b] = await Promise.all([buildTrajectory(store, runA), buildTrajectory(store, runB)])\n  const eq = options.stepEquals ?? defaultStepEquals\n  const minLen = Math.min(a.steps.length, b.steps.length)\n  for (let i = 0; i < minLen; i++) {\n    const aStep = a.steps[i]!\n    const bStep = b.steps[i]!\n    if (!eq(aStep, bStep)) {\n      return {\n        runA,\n        runB,\n        firstDivergenceIndex: i,\n        aStep,\n        bStep,\n        reason: describeDifference(aStep, bStep),\n        commonPrefixLen: i,\n      }\n    }\n  }\n  if (a.steps.length === b.steps.length) {\n    return { runA, runB, firstDivergenceIndex: null, commonPrefixLen: minLen }\n  }\n  const longer: Trajectory = a.steps.length > b.steps.length ? a : b\n  const extra = longer.steps.length - minLen\n  // minLen === 0 means one trajectory is empty: divergence is the first step\n  // itself (index 0), and the absent side has no step to surface.\n  const reason =\n    minLen === 0\n      ? `one trajectory is empty; the other has ${extra} step(s) starting at index 0`\n      : `one trajectory has ${extra} more step(s) after index ${minLen - 1}`\n  return {\n    runA,\n    runB,\n    firstDivergenceIndex: minLen,\n    aStep: a.steps[minLen],\n    bStep: b.steps[minLen],\n    reason,\n    commonPrefixLen: minLen,\n  }\n}\n\nfunction defaultStepEquals(a: TrajectoryStep, b: TrajectoryStep): boolean {\n  if (a.span.kind !== b.span.kind) return false\n  if (a.span.kind === 'tool' && b.span.kind === 'tool') return a.span.toolName === b.span.toolName\n  if (a.span.kind === 'llm' && b.span.kind === 'llm') return a.span.model === b.span.model\n  if (a.span.kind === 'judge' && b.span.kind === 'judge')\n    return a.span.dimension === b.span.dimension\n  return a.span.name === b.span.name\n}\n\nfunction describeDifference(a: TrajectoryStep, b: TrajectoryStep): string {\n  if (a.span.kind !== b.span.kind) return `kind ${a.span.kind} vs ${b.span.kind}`\n  if (a.span.kind === 'tool' && b.span.kind === 'tool' && a.span.toolName !== b.span.toolName) {\n    return `tool ${a.span.toolName} vs ${b.span.toolName}`\n  }\n  if (a.span.kind === 'llm' && b.span.kind === 'llm' && a.span.model !== b.span.model) {\n    return `model ${a.span.model} vs ${b.span.model}`\n  }\n  return `name \"${a.span.name}\" vs \"${b.span.name}\"`\n}\n","/**\n * RegressionView — compares a candidate slice to a baseline slice on a\n * named metric. Delegates the statistics (Welch's t-test, Cohen's d,\n * IQR stability) to `baseline.ts`.\n *\n * This is the entry point for CI regression gates: \"given runs tagged\n * release=A and release=B, did any metric regress?\"\n */\n\nimport { type BaselineOptions, type BaselineReport, compareToBaseline } from '../baseline'\nimport { aggregateLlm, llmSpans, runFailureClass } from '../trace/query'\nimport type { Run } from '../trace/schema'\nimport type { RunFilter, TraceStore } from '../trace/store'\n\nexport interface RegressionSpec {\n  metric: string\n  higherIsBetter: boolean\n  /** Extract a scalar from a run. Default extractors handle common metrics. */\n  extract?: (run: Run, store: TraceStore) => Promise<number | null>\n}\n\nexport interface RegressionOptions extends BaselineOptions {\n  baseline: RunFilter\n  candidate: RunFilter\n}\n\nexport async function regressionView(\n  store: TraceStore,\n  metrics: RegressionSpec[],\n  options: RegressionOptions,\n): Promise<BaselineReport> {\n  const baselineRuns = await store.listRuns(options.baseline)\n  const candidateRuns = await store.listRuns(options.candidate)\n  const samples = await Promise.all(\n    metrics.map(async (m) => {\n      const extract = m.extract ?? defaultExtract(m.metric)\n      const baseline = await extractAll(baselineRuns, extract, store)\n      const candidate = await extractAll(candidateRuns, extract, store)\n      return { metric: m.metric, higherIsBetter: m.higherIsBetter, baseline, candidate }\n    }),\n  )\n  return compareToBaseline(samples, options)\n}\n\nasync function extractAll(\n  runs: Run[],\n  extract: (r: Run, s: TraceStore) => Promise<number | null>,\n  store: TraceStore,\n): Promise<number[]> {\n  const out: number[] = []\n  for (const r of runs) {\n    const v = await extract(r, store)\n    if (v !== null && Number.isFinite(v)) out.push(v)\n  }\n  return out\n}\n\nfunction defaultExtract(metric: string): (run: Run, store: TraceStore) => Promise<number | null> {\n  return async (run, store) => {\n    switch (metric) {\n      case 'score':\n      case 'overallScore':\n        return run.outcome?.score ?? null\n      case 'pass':\n        return run.outcome?.pass === true ? 1 : 0\n      case 'durationMs':\n        return run.endedAt && run.startedAt ? run.endedAt - run.startedAt : null\n      case 'costUsd': {\n        const llm = await llmSpans(store, run.runId)\n        return aggregateLlm(llm).costUsd\n      }\n      case 'inputTokens': {\n        const llm = await llmSpans(store, run.runId)\n        return aggregateLlm(llm).inputTokens\n      }\n      case 'outputTokens': {\n        const llm = await llmSpans(store, run.runId)\n        return aggregateLlm(llm).outputTokens\n      }\n      case 'failureClass': {\n        return runFailureClass(run) === 'success' ? 1 : 0\n      }\n      default:\n        return null\n    }\n  }\n}\n","/**\n * StuckLoopView — detects when an agent calls the same tool with the\n * same (or structurally similar) arguments ≥ N times in a short window.\n *\n * Rationale: agents that loop are the number-one production failure\n * mode on long-horizon flows. The view returns (runId, toolName,\n * argHash, occurrences, windowMs) for each detected loop plus a\n * fraction of runs affected.\n */\n\nimport { executionTrackByLane } from '../trace/execution-tracks'\nimport { argHash, hasCapturedToolArgs } from '../trace/query'\nimport { isToolSpan, type Span, type ToolSpan } from '../trace/schema'\nimport type { TraceStore } from '../trace/store'\n\nconst DEFAULT_MAX_WINDOW_MS = 60_000\nconst DEFAULT_MAX_INTERVENING_TOOL_CALLS = 0\n\ninterface ScopedToolCall {\n  span: ToolSpan\n  scopeSpanId: string | null\n  laneSpanId: string | null\n}\n\ninterface IndexedToolCall extends ScopedToolCall {\n  toolCallIndex: number\n  trackId: string\n}\n\nexport interface StuckLoopFinding {\n  runId: string\n  toolName: string\n  argHash: string\n  /** Calls in this episode's densest qualifying interval, not the whole-run total. */\n  occurrences: number\n  spanIds: string[]\n  /** Nearest agent ancestor, or the direct parent when ancestry is incomplete. */\n  scopeSpanId?: string\n  /** Milliseconds between first and last call in the loop. */\n  windowMs: number\n}\n\nexport interface StuckLoopReport {\n  findings: StuckLoopFinding[]\n  affectedRunRatio: number\n  totalRuns: number\n}\n\nexport interface StuckLoopOptions {\n  /** Minimum call count to flag a loop (default 3). */\n  minOccurrences?: number\n  /** Maximum time between the first and last repeated call (default 60 seconds). */\n  maxWindowMs?: number\n  /**\n   * Maximum other tool calls allowed between adjacent repeats (default 0).\n   * Set to 1 to detect alternating patterns such as A,B,A,B,A.\n   */\n  maxInterveningToolCalls?: number\n  /** Filter to a specific runId; omit to scan the entire corpus. */\n  runId?: string\n}\n\nexport async function stuckLoopView(\n  store: TraceStore,\n  options: StuckLoopOptions = {},\n): Promise<StuckLoopReport> {\n  const minOccurrences = options.minOccurrences ?? 3\n  const maxWindowMs = options.maxWindowMs ?? DEFAULT_MAX_WINDOW_MS\n  const maxInterveningToolCalls =\n    options.maxInterveningToolCalls ?? DEFAULT_MAX_INTERVENING_TOOL_CALLS\n  if (!Number.isInteger(minOccurrences) || minOccurrences < 1) {\n    throw new RangeError('minOccurrences must be a positive integer')\n  }\n  if (!Number.isFinite(maxWindowMs) || maxWindowMs < 0) {\n    throw new RangeError('maxWindowMs must be a finite non-negative number')\n  }\n  if (!Number.isInteger(maxInterveningToolCalls) || maxInterveningToolCalls < 0) {\n    throw new RangeError('maxInterveningToolCalls must be a non-negative integer')\n  }\n  const runs = options.runId\n    ? [{ runId: options.runId }]\n    : (await store.listRuns()).map((r) => ({ runId: r.runId }))\n\n  const findings: StuckLoopFinding[] = []\n  for (const { runId } of runs) {\n    const spans = await store.spans({ runId })\n    const spansById = new Map(spans.map((span) => [span.spanId, span]))\n    const scopedTools: ScopedToolCall[] = spans\n      .filter(isToolSpan)\n      .map((span, sourceIndex) => ({ span, sourceIndex }))\n      .sort((a, b) => a.span.startedAt - b.span.startedAt || a.sourceIndex - b.sourceIndex)\n      .map(({ span }) => ({ span, ...executionScope(span, spansById) }))\n    const trackByLane = executionTrackByLane(\n      scopedTools.map((call) => {\n        const direct = call.laneSpanId === null || call.laneSpanId === call.scopeSpanId\n        const timed = direct\n          ? call.span\n          : call.laneSpanId\n            ? spansById.get(call.laneSpanId)\n            : undefined\n        return {\n          key: executionKey(call),\n          scopeKey: JSON.stringify(call.scopeSpanId),\n          start: timed?.startedAt ?? null,\n          end: timed?.endedAt ?? null,\n        }\n      }),\n    )\n    const nextToolIndexByTrack = new Map<string, number>()\n    const orderedTools: IndexedToolCall[] = scopedTools.map((call) => {\n      const trackId = trackByLane.get(executionKey(call))!\n      const toolCallIndex = nextToolIndexByTrack.get(trackId) ?? 0\n      nextToolIndexByTrack.set(trackId, toolCallIndex + 1)\n      return { ...call, toolCallIndex, trackId }\n    })\n    const byKey = new Map<\n      string,\n      {\n        calls: IndexedToolCall[]\n        argHash: string\n        toolName: string\n        scopeSpanId: string | null\n      }\n    >()\n    for (const call of orderedTools) {\n      if (!hasCapturedToolArgs(call.span)) continue\n      const h = argHash(call.span.args)\n      const key = JSON.stringify([call.trackId, call.span.toolName, h])\n      const bucket = byKey.get(key) ?? {\n        calls: [],\n        argHash: h,\n        toolName: call.span.toolName,\n        scopeSpanId: call.scopeSpanId,\n      }\n      bucket.calls.push(call)\n      byKey.set(key, bucket)\n    }\n    for (const { calls, argHash: h, toolName, scopeSpanId } of byKey.values()) {\n      if (calls.length < minOccurrences) continue\n      let episodeStart = 0\n      for (let episodeEnd = 1; episodeEnd <= calls.length; episodeEnd += 1) {\n        const previous = calls[episodeEnd - 1]!\n        const next = calls[episodeEnd]\n        const episodeEnded =\n          next === undefined ||\n          next.span.startedAt - previous.span.startedAt > maxWindowMs ||\n          next.toolCallIndex - previous.toolCallIndex - 1 > maxInterveningToolCalls ||\n          !callsAreSerial(previous, next)\n        if (!episodeEnded) continue\n\n        const episode = calls.slice(episodeStart, episodeEnd)\n        let left = 0\n        let bestStart = 0\n        let bestEnd = -1\n        for (let right = 0; right < episode.length; right += 1) {\n          while (episode[right]!.span.startedAt - episode[left]!.span.startedAt > maxWindowMs) {\n            left += 1\n          }\n          if (right - left > bestEnd - bestStart) {\n            bestStart = left\n            bestEnd = right\n          }\n        }\n        if (bestEnd - bestStart + 1 >= minOccurrences) {\n          const loop = episode.slice(bestStart, bestEnd + 1)\n          const first = loop[0]!.span.startedAt\n          const last = loop[loop.length - 1]!.span.startedAt\n          findings.push({\n            runId,\n            toolName,\n            argHash: h,\n            occurrences: loop.length,\n            spanIds: loop.map((call) => call.span.spanId),\n            ...(scopeSpanId ? { scopeSpanId } : {}),\n            windowMs: last - first,\n          })\n        }\n        episodeStart = episodeEnd\n      }\n    }\n  }\n\n  const affectedRuns = new Set(findings.map((f) => f.runId))\n  return {\n    findings,\n    affectedRunRatio: runs.length > 0 ? affectedRuns.size / runs.length : 0,\n    totalRuns: runs.length,\n  }\n}\n\nfunction laneKey(call: Pick<ScopedToolCall, 'scopeSpanId' | 'laneSpanId'>): string {\n  return JSON.stringify([call.scopeSpanId, call.laneSpanId])\n}\n\nfunction executionKey(call: ScopedToolCall): string {\n  return call.laneSpanId === null || call.laneSpanId === call.scopeSpanId\n    ? JSON.stringify([call.scopeSpanId, call.span.spanId])\n    : laneKey(call)\n}\n\nfunction executionScope(\n  span: ToolSpan,\n  spansById: ReadonlyMap<string, Span>,\n): { scopeSpanId: string | null; laneSpanId: string | null } {\n  const directParent = span.parentSpanId\n  if (!directParent) return { scopeSpanId: null, laneSpanId: null }\n\n  let currentId: string | undefined = directParent\n  let laneSpanId: string | null = null\n  const seen = new Set<string>()\n  while (currentId && !seen.has(currentId)) {\n    seen.add(currentId)\n    const current = spansById.get(currentId)\n    if (!current) return { scopeSpanId: directParent, laneSpanId: directParent }\n    if (current.kind === 'agent') {\n      return { scopeSpanId: current.spanId, laneSpanId: laneSpanId ?? current.spanId }\n    }\n    laneSpanId = current.spanId\n    currentId = current.parentSpanId\n  }\n  return { scopeSpanId: directParent, laneSpanId: directParent }\n}\n\nfunction callsAreSerial(previous: IndexedToolCall, next: IndexedToolCall): boolean {\n  return previous.span.endedAt !== undefined && previous.span.endedAt <= next.span.startedAt\n}\n"],"mappings":";;;;;;;AA4BA,eAAsB,oBACpB,OACA,MACA,MACA,UAA6B,CAAC,GACH;CAC3B,MAAM,CAAC,GAAG,KAAK,MAAM,QAAQ,IAAI,CAAC,gBAAgB,OAAO,IAAI,GAAG,gBAAgB,OAAO,IAAI,CAAC,CAAC;CAC7F,MAAM,KAAK,QAAQ,cAAc;CACjC,MAAM,SAAS,KAAK,IAAI,EAAE,MAAM,QAAQ,EAAE,MAAM,MAAM;CACtD,KAAK,IAAI,IAAI,GAAG,IAAI,QAAQ,KAAK;EAC/B,MAAM,QAAQ,EAAE,MAAM;EACtB,MAAM,QAAQ,EAAE,MAAM;EACtB,IAAI,CAAC,GAAG,OAAO,KAAK,GAClB,OAAO;GACL;GACA;GACA,sBAAsB;GACtB;GACA;GACA,QAAQ,mBAAmB,OAAO,KAAK;GACvC,iBAAiB;EACnB;CAEJ;CACA,IAAI,EAAE,MAAM,WAAW,EAAE,MAAM,QAC7B,OAAO;EAAE;EAAM;EAAM,sBAAsB;EAAM,iBAAiB;CAAO;CAG3E,MAAM,SADqB,EAAE,MAAM,SAAS,EAAE,MAAM,SAAS,IAAI,EAAA,CAC5C,MAAM,SAAS;CAGpC,MAAM,SACJ,WAAW,IACP,0CAA0C,MAAM,gCAChD,sBAAsB,MAAM,4BAA4B,SAAS;CACvE,OAAO;EACL;EACA;EACA,sBAAsB;EACtB,OAAO,EAAE,MAAM;EACf,OAAO,EAAE,MAAM;EACf;EACA,iBAAiB;CACnB;AACF;AAEA,SAAS,kBAAkB,GAAmB,GAA4B;CACxE,IAAI,EAAE,KAAK,SAAS,EAAE,KAAK,MAAM,OAAO;CACxC,IAAI,EAAE,KAAK,SAAS,UAAU,EAAE,KAAK,SAAS,QAAQ,OAAO,EAAE,KAAK,aAAa,EAAE,KAAK;CACxF,IAAI,EAAE,KAAK,SAAS,SAAS,EAAE,KAAK,SAAS,OAAO,OAAO,EAAE,KAAK,UAAU,EAAE,KAAK;CACnF,IAAI,EAAE,KAAK,SAAS,WAAW,EAAE,KAAK,SAAS,SAC7C,OAAO,EAAE,KAAK,cAAc,EAAE,KAAK;CACrC,OAAO,EAAE,KAAK,SAAS,EAAE,KAAK;AAChC;AAEA,SAAS,mBAAmB,GAAmB,GAA2B;CACxE,IAAI,EAAE,KAAK,SAAS,EAAE,KAAK,MAAM,OAAO,QAAQ,EAAE,KAAK,KAAK,MAAM,EAAE,KAAK;CACzE,IAAI,EAAE,KAAK,SAAS,UAAU,EAAE,KAAK,SAAS,UAAU,EAAE,KAAK,aAAa,EAAE,KAAK,UACjF,OAAO,QAAQ,EAAE,KAAK,SAAS,MAAM,EAAE,KAAK;CAE9C,IAAI,EAAE,KAAK,SAAS,SAAS,EAAE,KAAK,SAAS,SAAS,EAAE,KAAK,UAAU,EAAE,KAAK,OAC5E,OAAO,SAAS,EAAE,KAAK,MAAM,MAAM,EAAE,KAAK;CAE5C,OAAO,SAAS,EAAE,KAAK,KAAK,QAAQ,EAAE,KAAK,KAAK;AAClD;;;;;;;;;;;AClEA,eAAsB,eACpB,OACA,SACA,SACyB;CACzB,MAAM,eAAe,MAAM,MAAM,SAAS,QAAQ,QAAQ;CAC1D,MAAM,gBAAgB,MAAM,MAAM,SAAS,QAAQ,SAAS;CAS5D,OAAO,kBAAkB,MARH,QAAQ,IAC5B,QAAQ,IAAI,OAAO,MAAM;EACvB,MAAM,UAAU,EAAE,WAAW,eAAe,EAAE,MAAM;EACpD,MAAM,WAAW,MAAM,WAAW,cAAc,SAAS,KAAK;EAC9D,MAAM,YAAY,MAAM,WAAW,eAAe,SAAS,KAAK;EAChE,OAAO;GAAE,QAAQ,EAAE;GAAQ,gBAAgB,EAAE;GAAgB;GAAU;EAAU;CACnF,CAAC,CACH,GACkC,OAAO;AAC3C;AAEA,eAAe,WACb,MACA,SACA,OACmB;CACnB,MAAM,MAAgB,CAAC;CACvB,KAAK,MAAM,KAAK,MAAM;EACpB,MAAM,IAAI,MAAM,QAAQ,GAAG,KAAK;EAChC,IAAI,MAAM,QAAQ,OAAO,SAAS,CAAC,GAAG,IAAI,KAAK,CAAC;CAClD;CACA,OAAO;AACT;AAEA,SAAS,eAAe,QAAyE;CAC/F,OAAO,OAAO,KAAK,UAAU;EAC3B,QAAQ,QAAR;GACE,KAAK;GACL,KAAK,gBACH,OAAO,IAAI,SAAS,SAAS;GAC/B,KAAK,QACH,OAAO,IAAI,SAAS,SAAS,OAAO,IAAI;GAC1C,KAAK,cACH,OAAO,IAAI,WAAW,IAAI,YAAY,IAAI,UAAU,IAAI,YAAY;GACtE,KAAK,WAEH,OAAO,aAAa,MADF,SAAS,OAAO,IAAI,KAAK,CACpB,CAAC,CAAC;GAE3B,KAAK,eAEH,OAAO,aAAa,MADF,SAAS,OAAO,IAAI,KAAK,CACpB,CAAC,CAAC;GAE3B,KAAK,gBAEH,OAAO,aAAa,MADF,SAAS,OAAO,IAAI,KAAK,CACpB,CAAC,CAAC;GAE3B,KAAK,gBACH,OAAO,gBAAgB,GAAG,MAAM,YAAY,IAAI;GAElD,SACE,OAAO;EACX;CACF;AACF;;;;;;;;;;;;ACvEA,MAAM,wBAAwB;AAC9B,MAAM,qCAAqC;AA8C3C,eAAsB,cACpB,OACA,UAA4B,CAAC,GACH;CAC1B,MAAM,iBAAiB,QAAQ,kBAAkB;CACjD,MAAM,cAAc,QAAQ,eAAe;CAC3C,MAAM,0BACJ,QAAQ,2BAA2B;CACrC,IAAI,CAAC,OAAO,UAAU,cAAc,KAAK,iBAAiB,GACxD,MAAM,IAAI,WAAW,2CAA2C;CAElE,IAAI,CAAC,OAAO,SAAS,WAAW,KAAK,cAAc,GACjD,MAAM,IAAI,WAAW,kDAAkD;CAEzE,IAAI,CAAC,OAAO,UAAU,uBAAuB,KAAK,0BAA0B,GAC1E,MAAM,IAAI,WAAW,wDAAwD;CAE/E,MAAM,OAAO,QAAQ,QACjB,CAAC,EAAE,OAAO,QAAQ,MAAM,CAAC,KACxB,MAAM,MAAM,SAAS,EAAA,CAAG,KAAK,OAAO,EAAE,OAAO,EAAE,MAAM,EAAE;CAE5D,MAAM,WAA+B,CAAC;CACtC,KAAK,MAAM,EAAE,WAAW,MAAM;EAC5B,MAAM,QAAQ,MAAM,MAAM,MAAM,EAAE,MAAM,CAAC;EACzC,MAAM,YAAY,IAAI,IAAI,MAAM,KAAK,SAAS,CAAC,KAAK,QAAQ,IAAI,CAAC,CAAC;EAClE,MAAM,cAAgC,MACnC,OAAO,UAAU,CAAC,CAClB,KAAK,MAAM,iBAAiB;GAAE;GAAM;EAAY,EAAE,CAAC,CACnD,MAAM,GAAG,MAAM,EAAE,KAAK,YAAY,EAAE,KAAK,aAAa,EAAE,cAAc,EAAE,WAAW,CAAC,CACpF,KAAK,EAAE,YAAY;GAAE;GAAM,GAAG,eAAe,MAAM,SAAS;EAAE,EAAE;EACnE,MAAM,cAAc,qBAClB,YAAY,KAAK,SAAS;GAExB,MAAM,QADS,KAAK,eAAe,QAAQ,KAAK,eAAe,KAAK,cAEhE,KAAK,OACL,KAAK,aACH,UAAU,IAAI,KAAK,UAAU,IAC7B,KAAA;GACN,OAAO;IACL,KAAK,aAAa,IAAI;IACtB,UAAU,KAAK,UAAU,KAAK,WAAW;IACzC,OAAO,OAAO,aAAa;IAC3B,KAAK,OAAO,WAAW;GACzB;EACF,CAAC,CACH;EACA,MAAM,uCAAuB,IAAI,IAAoB;EACrD,MAAM,eAAkC,YAAY,KAAK,SAAS;GAChE,MAAM,UAAU,YAAY,IAAI,aAAa,IAAI,CAAC;GAClD,MAAM,gBAAgB,qBAAqB,IAAI,OAAO,KAAK;GAC3D,qBAAqB,IAAI,SAAS,gBAAgB,CAAC;GACnD,OAAO;IAAE,GAAG;IAAM;IAAe;GAAQ;EAC3C,CAAC;EACD,MAAM,wBAAQ,IAAI,IAQhB;EACF,KAAK,MAAM,QAAQ,cAAc;GAC/B,IAAI,CAAC,oBAAoB,KAAK,IAAI,GAAG;GACrC,MAAM,IAAI,QAAQ,KAAK,KAAK,IAAI;GAChC,MAAM,MAAM,KAAK,UAAU;IAAC,KAAK;IAAS,KAAK,KAAK;IAAU;GAAC,CAAC;GAChE,MAAM,SAAS,MAAM,IAAI,GAAG,KAAK;IAC/B,OAAO,CAAC;IACR,SAAS;IACT,UAAU,KAAK,KAAK;IACpB,aAAa,KAAK;GACpB;GACA,OAAO,MAAM,KAAK,IAAI;GACtB,MAAM,IAAI,KAAK,MAAM;EACvB;EACA,KAAK,MAAM,EAAE,OAAO,SAAS,GAAG,UAAU,iBAAiB,MAAM,OAAO,GAAG;GACzE,IAAI,MAAM,SAAS,gBAAgB;GACnC,IAAI,eAAe;GACnB,KAAK,IAAI,aAAa,GAAG,cAAc,MAAM,QAAQ,cAAc,GAAG;IACpE,MAAM,WAAW,MAAM,aAAa;IACpC,MAAM,OAAO,MAAM;IAMnB,IAAI,EAJF,SAAS,KAAA,KACT,KAAK,KAAK,YAAY,SAAS,KAAK,YAAY,eAChD,KAAK,gBAAgB,SAAS,gBAAgB,IAAI,2BAClD,CAAC,eAAe,UAAU,IAAI,IACb;IAEnB,MAAM,UAAU,MAAM,MAAM,cAAc,UAAU;IACpD,IAAI,OAAO;IACX,IAAI,YAAY;IAChB,IAAI,UAAU;IACd,KAAK,IAAI,QAAQ,GAAG,QAAQ,QAAQ,QAAQ,SAAS,GAAG;KACtD,OAAO,QAAQ,MAAM,CAAE,KAAK,YAAY,QAAQ,KAAK,CAAE,KAAK,YAAY,aACtE,QAAQ;KAEV,IAAI,QAAQ,OAAO,UAAU,WAAW;MACtC,YAAY;MACZ,UAAU;KACZ;IACF;IACA,IAAI,UAAU,YAAY,KAAK,gBAAgB;KAC7C,MAAM,OAAO,QAAQ,MAAM,WAAW,UAAU,CAAC;KACjD,MAAM,QAAQ,KAAK,EAAE,CAAE,KAAK;KAC5B,MAAM,OAAO,KAAK,KAAK,SAAS,EAAE,CAAE,KAAK;KACzC,SAAS,KAAK;MACZ;MACA;MACA,SAAS;MACT,aAAa,KAAK;MAClB,SAAS,KAAK,KAAK,SAAS,KAAK,KAAK,MAAM;MAC5C,GAAI,cAAc,EAAE,YAAY,IAAI,CAAC;MACrC,UAAU,OAAO;KACnB,CAAC;IACH;IACA,eAAe;GACjB;EACF;CACF;CAEA,MAAM,eAAe,IAAI,IAAI,SAAS,KAAK,MAAM,EAAE,KAAK,CAAC;CACzD,OAAO;EACL;EACA,kBAAkB,KAAK,SAAS,IAAI,aAAa,OAAO,KAAK,SAAS;EACtE,WAAW,KAAK;CAClB;AACF;AAEA,SAAS,QAAQ,MAAkE;CACjF,OAAO,KAAK,UAAU,CAAC,KAAK,aAAa,KAAK,UAAU,CAAC;AAC3D;AAEA,SAAS,aAAa,MAA8B;CAClD,OAAO,KAAK,eAAe,QAAQ,KAAK,eAAe,KAAK,cACxD,KAAK,UAAU,CAAC,KAAK,aAAa,KAAK,KAAK,MAAM,CAAC,IACnD,QAAQ,IAAI;AAClB;AAEA,SAAS,eACP,MACA,WAC2D;CAC3D,MAAM,eAAe,KAAK;CAC1B,IAAI,CAAC,cAAc,OAAO;EAAE,aAAa;EAAM,YAAY;CAAK;CAEhE,IAAI,YAAgC;CACpC,IAAI,aAA4B;CAChC,MAAM,uBAAO,IAAI,IAAY;CAC7B,OAAO,aAAa,CAAC,KAAK,IAAI,SAAS,GAAG;EACxC,KAAK,IAAI,SAAS;EAClB,MAAM,UAAU,UAAU,IAAI,SAAS;EACvC,IAAI,CAAC,SAAS,OAAO;GAAE,aAAa;GAAc,YAAY;EAAa;EAC3E,IAAI,QAAQ,SAAS,SACnB,OAAO;GAAE,aAAa,QAAQ;GAAQ,YAAY,cAAc,QAAQ;EAAO;EAEjF,aAAa,QAAQ;EACrB,YAAY,QAAQ;CACtB;CACA,OAAO;EAAE,aAAa;EAAc,YAAY;CAAa;AAC/D;AAEA,SAAS,eAAe,UAA2B,MAAgC;CACjF,OAAO,SAAS,KAAK,YAAY,KAAA,KAAa,SAAS,KAAK,WAAW,KAAK,KAAK;AACnF"}

package/dist/power-and-mde-CHIrXJll.js

@@ -0,0 +1,195 @@
+import { s as ValidationError } from "./errors-Dngq5h35.js";
+import { s as zQuantile } from "./internal-BDHPCnjk.js";
+import { u as normalCdf } from "./paired-arms-D-XRF_fy.js";
+//#region src/statistics/multiplicity.ts
+/**
+* Multiple-comparison corrections: family-wise error (Bonferroni, Holm) and
+* false discovery rate (Benjamini-Hochberg), with inclusive rejection
+* boundaries throughout.
+*/
+/**
+* Bonferroni adjustment: multiply every p-value by the test count, clamp at 1.
+*
+* Rejects at `p_adjusted ≤ alpha` — the boundary is inclusive, matching
+* {@link holm}, which uniformly dominates this correction and must therefore
+* never reject less. Validates its inputs on the same terms.
+*/
+function bonferroni(pValues, alpha = .05) {
+	assertAlpha("bonferroni", "alpha", alpha);
+	assertPValues("bonferroni", pValues);
+	const k = pValues.length;
+	const adjusted = pValues.map((p) => Math.min(1, p * k));
+	return {
+		adjusted,
+		significant: adjusted.map((p) => p <= alpha)
+	};
+}
+/**
+* Holm step-down family-wise error adjustment.
+*
+* P-values are sorted from smallest to largest, multiplied by their remaining
+* hypothesis count, and made monotonically non-decreasing before being mapped
+* back to input order. This uniformly dominates plain Bonferroni while keeping
+* strong family-wise error control under arbitrary dependence.
+*/
+function holm(pValues, alpha = .05) {
+	assertAlpha("holm", "alpha", alpha);
+	assertPValues("holm", pValues);
+	const count = pValues.length;
+	if (count === 0) return {
+		adjusted: [],
+		significant: []
+	};
+	const ordered = pValues.map((pValue, index) => ({
+		pValue,
+		index
+	})).sort((a, b) => a.pValue - b.pValue || a.index - b.index);
+	const adjusted = new Array(count);
+	let previous = 0;
+	for (let rank = 0; rank < count; rank++) {
+		const entry = ordered[rank];
+		const stepAdjusted = Math.min(1, entry.pValue * (count - rank));
+		previous = Math.max(previous, stepAdjusted);
+		adjusted[entry.index] = previous;
+	}
+	return {
+		adjusted,
+		significant: adjusted.map((pValue) => pValue <= alpha)
+	};
+}
+/**
+* Benjamini–Hochberg false discovery rate. Returns adjusted q-values and
+* significance at the target FDR; handles ties and preserves q monotonicity.
+*
+* Rejects at `q ≤ fdr` — the BH rule is inclusive at the boundary, so an
+* exactly-`fdr` q-value is a discovery.
+*/
+function benjaminiHochberg(pValues, fdr = .05) {
+	assertAlpha("benjaminiHochberg", "fdr", fdr);
+	assertPValues("benjaminiHochberg", pValues);
+	const n = pValues.length;
+	if (n === 0) return {
+		qValues: [],
+		significant: []
+	};
+	const indexed = pValues.map((p, i) => ({
+		p,
+		i
+	})).sort((a, b) => a.p - b.p);
+	const q = new Array(n);
+	let minRight = 1;
+	for (let k = n - 1; k >= 0; k--) {
+		const rank = k + 1;
+		const entry = indexed[k];
+		const raw = n / rank * entry.p;
+		const bounded = Math.min(minRight, raw);
+		minRight = bounded;
+		q[entry.i] = Math.min(1, bounded);
+	}
+	return {
+		qValues: q,
+		significant: q.map((v) => v <= fdr)
+	};
+}
+function assertAlpha(fn, label, value) {
+	if (!Number.isFinite(value) || value <= 0 || value >= 1) throw new ValidationError(`${fn}: ${label} must be in (0,1), got ${value}`);
+}
+function assertPValues(fn, pValues) {
+	for (const [index, pValue] of pValues.entries()) if (!Number.isFinite(pValue) || pValue < 0 || pValue > 1) throw new ValidationError(`${fn}: pValues[${index}] must be in [0,1], got ${pValue}`);
+}
+//#endregion
+//#region src/statistics/power-and-mde.ts
+/**
+* Required N per arm for a two-sample comparison at target effect size,
+* alpha, and power. Normal-approximation formula:
+*   n = 2 * ( (z_{1-α/2} + z_{1-β}) / d )^2
+* where d is Cohen's d. Returns Infinity for effect ≤ 0.
+*/
+function requiredSampleSize(opts) {
+	const effect = opts.effect;
+	if (!Number.isFinite(effect) || effect <= 0) return Infinity;
+	const alpha = opts.alpha ?? .05;
+	const power = opts.power ?? .8;
+	const n = 2 * ((zQuantile(opts.twoSided ?? true ? 1 - alpha / 2 : 1 - alpha) + zQuantile(power)) / effect) ** 2;
+	return Math.ceil(n);
+}
+/**
+* Required number of paired observations for a target Cohen's dz.
+* Unlike the independent-groups formula, this has no two-arm factor of two.
+*
+* Normal quantiles with no t correction, so treat the result as a LOWER bound:
+* it returns 32 where the exact t-based answer is 34 at dz = 0.5, and 13 where
+* it is 15 at dz = 0.8 — a 6–13 % shortfall precisely in the range a caller
+* consults to decide whether 3–10 repetitions suffice.
+*/
+function requiredPairedSampleSize(opts) {
+	const effect = opts.effect;
+	if (!Number.isFinite(effect) || effect <= 0) return Infinity;
+	const alpha = opts.alpha ?? .05;
+	const power = opts.power ?? .8;
+	const zAlpha = zQuantile(opts.twoSided ?? true ? 1 - alpha / 2 : 1 - alpha);
+	const zBeta = zQuantile(power);
+	return Math.ceil(((zAlpha + zBeta) / effect) ** 2);
+}
+/**
+* Minimum detectable paired effect (standardised units) for a target paired
+* sample size: d_min = (z_{1-α/2} + z_β) / sqrt(n_paired). Multiply by
+* sd(deltas) for score units; treat as a lower bound — Wilcoxon and bootstrap
+* have asymptotic relative efficiency below 1 vs the t-test on heavy tails.
+*/
+function pairedMde(opts) {
+	if (!Number.isFinite(opts.nPaired) || opts.nPaired <= 0) return Infinity;
+	const alpha = opts.alpha ?? .05;
+	const power = opts.power ?? .8;
+	return (zQuantile(opts.twoSided ?? true ? 1 - alpha / 2 : 1 - alpha) + zQuantile(power)) / Math.sqrt(opts.nPaired);
+}
+/**
+* Number of paired observations needed for a McNemar test to reach a target
+* power — the pre-registration companion to {@link mcnemar}. Parametrised by the
+* expected discordant-cell probabilities `p10` (P[treatment wins on a pair]) and
+* `p01` (P[control wins]); concordant pairs carry no information, so the count
+* is driven entirely by the discordant rate. Lachin's (1992) asymptotic normal
+* approximation: with discordant rate `pDisc = p10 + p01` and marginal effect
+* `δ = p10 − p01`,
+*   n = ( z_{1-α/2}·√pDisc + z_{1-β}·√(pDisc − δ²) )² / δ².
+* Returns Infinity when there is no effect (p10 === p01). Asymptotic — at the
+* tiny discordant counts where the exact {@link mcnemar} differs from the normal
+* approximation, treat the result as a lower bound and prefer the discordant-pair
+* floor.
+*/
+function mcnemarRequiredN(opts) {
+	const { p10, p01 } = opts;
+	if (p10 < 0 || p01 < 0 || p10 + p01 > 1) throw new Error(`mcnemarRequiredN: require p10,p01 ≥ 0 and p10+p01 ≤ 1 (got ${p10}, ${p01})`);
+	const delta = p10 - p01;
+	if (delta === 0) return Infinity;
+	const alpha = opts.alpha ?? .05;
+	const power = opts.power ?? .8;
+	const twoSided = opts.twoSided ?? true;
+	const pDisc = p10 + p01;
+	const zAlpha = zQuantile(twoSided ? 1 - alpha / 2 : 1 - alpha);
+	const zBeta = zQuantile(power);
+	const n = (zAlpha * Math.sqrt(pDisc) + zBeta * Math.sqrt(Math.max(0, pDisc - delta * delta))) ** 2 / (delta * delta);
+	return Math.ceil(n);
+}
+/**
+* Power of a McNemar test at a given number of paired observations, the inverse
+* of {@link mcnemarRequiredN} (same Lachin asymptotic model, same parameters).
+* Returns a value in [0, 1]; equals `alpha` when there is no effect.
+*/
+function mcnemarPower(opts) {
+	const { p10, p01, nPairs } = opts;
+	if (p10 < 0 || p01 < 0 || p10 + p01 > 1) throw new Error(`mcnemarPower: require p10,p01 ≥ 0 and p10+p01 ≤ 1 (got ${p10}, ${p01})`);
+	const alpha = opts.alpha ?? .05;
+	const twoSided = opts.twoSided ?? true;
+	const delta = p10 - p01;
+	if (delta === 0 || nPairs <= 0) return alpha;
+	const pDisc = p10 + p01;
+	const zAlpha = zQuantile(twoSided ? 1 - alpha / 2 : 1 - alpha);
+	const denom = Math.sqrt(Math.max(1e-12, pDisc - delta * delta));
+	const zBeta = (Math.sqrt(nPairs) * Math.abs(delta) - zAlpha * Math.sqrt(pDisc)) / denom;
+	return Math.min(1, Math.max(0, normalCdf(zBeta)));
+}
+//#endregion
+export { requiredSampleSize as a, holm as c, requiredPairedSampleSize as i, mcnemarRequiredN as n, benjaminiHochberg as o, pairedMde as r, bonferroni as s, mcnemarPower as t };
+
+//# sourceMappingURL=power-and-mde-CHIrXJll.js.map