vieval 0.0.11 → 0.0.12

package/dist/{cli-CHFCF8UR.mjs → cli-uzS81IPd.mjs}

@@ -1,4 +1,4 @@
-import { a as createOpenTelemetryRuntime, c as detectCliConfigMode, d as loadVievalCliConfig, n as consumeModuleRegistrations, o as createNoopTelemetryRuntime, r as endModuleRegistration, t as beginModuleRegistration, u as loadRawVievalConfig } from "./registry-BSyjwZFx.mjs";
+import { a as createOpenTelemetryRuntime, c as detectCliConfigMode, d as loadVievalCliConfig, n as consumeModuleRegistrations, o as createNoopTelemetryRuntime, r as endModuleRegistration, t as beginModuleRegistration, u as loadRawVievalConfig } from "./registry-BK7k6X81.mjs";
 import { createSchedulerRuntime } from "./core/scheduler/index.mjs";
 import { RunnerExecutionError, collectEvalEntries, createFilesystemTaskCacheRuntime, createRunnerRuntimeContext, createRunnerSchedule, createTaskExecutionContext, runScheduledTasks } from "./core/runner/index.mjs";
 import process from "node:process";
@@ -27,40 +27,44 @@ const supportedWorkspaceConfigFileNames = [
 	"vieval.config.cjs",
 	"vieval.config.json"
 ];
-async function isReadableFile(filePath) {
+/**
+* Loads and validates comparison-mode data from `vieval.config.*`.
+*/
+async function loadVievalComparisonConfig(options = {}) {
+	const cwd = options.cwd ?? process.cwd();
 	try {
-		await access(filePath);
-		return true;
-	} catch {
-		return false;
+		const loaded = await loadRawVievalConfig({
+			configFilePath: options.configFilePath,
+			cwd
+		});
+		if (loaded.configFilePath == null || loaded.config == null) throw new Error("Failed to find vieval config. Expected vieval.config.*");
+		assertComparisonMode(loaded.config);
+		const selectedComparison = selectComparisonConfig(loaded.config.comparisons, options.comparisonId);
+		const configDirectory = dirname(loaded.configFilePath);
+		const explicitMethods = (selectedComparison.methods ?? []).map((method, index) => normalizeMethodShape(method, configDirectory, index));
+		const discoveredMethods = await discoverMethodsFromWorkspaceGlobs({
+			comparison: selectedComparison,
+			configDirectory
+		});
+		const methods = [...explicitMethods, ...discoveredMethods];
+		if (methods.length === 0) throw new Error("Comparison config resolved zero methods. Configure methods or includesWorkspaces.");
+		validateMethodIdsAreUnique(methods);
+		return {
+			config: {
+				benchmark: normalizeBenchmark(selectedComparison),
+				methods
+			},
+			configFilePath: loaded.configFilePath
+		};
+	} catch (error) {
+		const errorMessage = errorMessageFrom(error) ?? "Unknown comparison config loading error.";
+		const resolvedPath = options.configFilePath ?? "vieval.config";
+		throw new Error(`Failed to load comparison config "${resolvedPath}": ${errorMessage}`);
 	}
 }
-function normalizeGlobInput(patterns) {
-	if (patterns == null) return [];
-	return (typeof patterns === "string" ? [patterns] : patterns).map((pattern) => pattern.trim()).filter((pattern) => pattern.length > 0);
-}
-function normalizeMethodShape(method, configDirectory, index) {
-	const id = method.id.trim();
-	const workspace = method.workspace.trim();
-	const project = method.project.trim();
-	const configFilePath = method.configFilePath?.trim();
-	if (id.length === 0) throw new Error(`Comparison method #${index + 1} is missing id.`);
-	if (workspace.length === 0) throw new Error(`Comparison method "${id}" is missing workspace.`);
-	if (project.length === 0) throw new Error(`Comparison method "${id}" is missing project.`);
-	const resolvedWorkspace = isAbsolute(workspace) ? workspace : resolve(configDirectory, workspace);
-	return {
-		configFilePath: configFilePath == null || configFilePath.length === 0 ? void 0 : isAbsolute(configFilePath) ? configFilePath : resolve(configDirectory, configFilePath),
-		id,
-		project,
-		workspace: resolvedWorkspace
-	};
-}
-async function findWorkspaceConfigFile(workspaceDirectory) {
-	for (const fileName of supportedWorkspaceConfigFileNames) {
-		const candidate = join(workspaceDirectory, fileName);
-		if (await isReadableFile(candidate)) return candidate;
-	}
-	return null;
+function assertComparisonMode(config) {
+	const mode = detectCliConfigMode(config);
+	if (mode !== "comparisons") throw new Error(`Expected comparison-mode config, but received ${mode}-mode config.`);
 }
 function createDiscoveredMethodId(configDirectory, workspace, projectName) {
 	const relativeWorkspace = relative(configDirectory, workspace);
@@ -92,24 +96,20 @@ async function discoverMethodsFromWorkspaceGlobs(args) {
 	}
 	return methods;
 }
-function validateMethodIdsAreUnique(methods) {
-	const methodIds = methods.map((method) => method.id);
-	const duplicatedMethodId = methodIds.find((methodId, index) => methodIds.indexOf(methodId) !== index);
-	if (duplicatedMethodId != null) throw new Error(`Duplicate comparison method id "${duplicatedMethodId}".`);
-}
-function assertComparisonMode(config) {
-	const mode = detectCliConfigMode(config);
-	if (mode !== "comparisons") throw new Error(`Expected comparison-mode config, but received ${mode}-mode config.`);
+async function findWorkspaceConfigFile(workspaceDirectory) {
+	for (const fileName of supportedWorkspaceConfigFileNames) {
+		const candidate = join(workspaceDirectory, fileName);
+		if (await isReadableFile(candidate)) return candidate;
+	}
+	return null;
 }
-function selectComparisonConfig(comparisons, comparisonId) {
-	if (comparisons.length === 0) throw new Error("Comparison config requires at least one comparisons entry.");
-	if (comparisonId == null || comparisonId.trim().length === 0) {
-		if (comparisons.length > 1) throw new Error(`Multiple comparisons found. Provide --comparison. Available ids: ${comparisons.map((item) => item.id).join(", ")}`);
-		return comparisons[0];
+async function isReadableFile(filePath) {
+	try {
+		await access(filePath);
+		return true;
+	} catch {
+		return false;
 	}
-	const selected = comparisons.find((item) => item.id === comparisonId);
-	if (selected == null) throw new Error(`Unknown comparison id "${comparisonId}".`);
-	return selected;
 }
 function normalizeBenchmark(comparison) {
 	const benchmarkId = comparison.benchmark.id.trim();
@@ -121,40 +121,40 @@ function normalizeBenchmark(comparison) {
 		sharedCaseNamespace
 	};
 }
-/**
-* Loads and validates comparison-mode data from `vieval.config.*`.
-*/
-async function loadVievalComparisonConfig(options = {}) {
-	const cwd = options.cwd ?? process.cwd();
-	try {
-		const loaded = await loadRawVievalConfig({
-			configFilePath: options.configFilePath,
-			cwd
-		});
-		if (loaded.configFilePath == null || loaded.config == null) throw new Error("Failed to find vieval config. Expected vieval.config.*");
-		assertComparisonMode(loaded.config);
-		const selectedComparison = selectComparisonConfig(loaded.config.comparisons, options.comparisonId);
-		const configDirectory = dirname(loaded.configFilePath);
-		const explicitMethods = (selectedComparison.methods ?? []).map((method, index) => normalizeMethodShape(method, configDirectory, index));
-		const discoveredMethods = await discoverMethodsFromWorkspaceGlobs({
-			comparison: selectedComparison,
-			configDirectory
-		});
-		const methods = [...explicitMethods, ...discoveredMethods];
-		if (methods.length === 0) throw new Error("Comparison config resolved zero methods. Configure methods or includesWorkspaces.");
-		validateMethodIdsAreUnique(methods);
-		return {
-			config: {
-				benchmark: normalizeBenchmark(selectedComparison),
-				methods
-			},
-			configFilePath: loaded.configFilePath
-		};
-	} catch (error) {
-		const errorMessage = errorMessageFrom(error) ?? "Unknown comparison config loading error.";
-		const resolvedPath = options.configFilePath ?? "vieval.config";
-		throw new Error(`Failed to load comparison config "${resolvedPath}": ${errorMessage}`);
+function normalizeGlobInput(patterns) {
+	if (patterns == null) return [];
+	return (typeof patterns === "string" ? [patterns] : patterns).map((pattern) => pattern.trim()).filter((pattern) => pattern.length > 0);
+}
+function normalizeMethodShape(method, configDirectory, index) {
+	const id = method.id.trim();
+	const workspace = method.workspace.trim();
+	const project = method.project.trim();
+	const configFilePath = method.configFilePath?.trim();
+	if (id.length === 0) throw new Error(`Comparison method #${index + 1} is missing id.`);
+	if (workspace.length === 0) throw new Error(`Comparison method "${id}" is missing workspace.`);
+	if (project.length === 0) throw new Error(`Comparison method "${id}" is missing project.`);
+	const resolvedWorkspace = isAbsolute(workspace) ? workspace : resolve(configDirectory, workspace);
+	return {
+		configFilePath: configFilePath == null || configFilePath.length === 0 ? void 0 : isAbsolute(configFilePath) ? configFilePath : resolve(configDirectory, configFilePath),
+		id,
+		project,
+		workspace: resolvedWorkspace
+	};
+}
+function selectComparisonConfig(comparisons, comparisonId) {
+	if (comparisons.length === 0) throw new Error("Comparison config requires at least one comparisons entry.");
+	if (comparisonId == null || comparisonId.trim().length === 0) {
+		if (comparisons.length > 1) throw new Error(`Multiple comparisons found. Provide --comparison. Available ids: ${comparisons.map((item) => item.id).join(", ")}`);
+		return comparisons[0];
 	}
+	const selected = comparisons.find((item) => item.id === comparisonId);
+	if (selected == null) throw new Error(`Unknown comparison id "${comparisonId}".`);
+	return selected;
+}
+function validateMethodIdsAreUnique(methods) {
+	const methodIds = methods.map((method) => method.id);
+	const duplicatedMethodId = methodIds.find((methodId, index) => methodIds.indexOf(methodId) !== index);
+	if (duplicatedMethodId != null) throw new Error(`Duplicate comparison method id "${duplicatedMethodId}".`);
 }
 //#endregion
 //#region src/cli/report-records.ts
@@ -242,51 +242,39 @@ function encodeJsonl(records) {
 	if (records.length === 0) return "";
 	return `${records.map((record) => JSON.stringify(record)).join("\n")}\n`;
 }
-function normalizeCaseEventName(eventName) {
-	if (eventName === "task.case.start" || eventName === "CaseStarted") return "start";
-	if (eventName === "task.case.metric") return "metric";
-	if (eventName === "task.case.score") return "score";
-	if (eventName === "task.case.end" || eventName === "CaseEnded") return "end";
+function addRecordScores(summary, record) {
+	for (const [kind, score] of Object.entries(record.scores)) {
+		if (!Number.isFinite(score)) continue;
+		summary[kind] ??= {
+			average: 0,
+			count: 0,
+			sum: 0
+		};
+		summary[kind].count += 1;
+		summary[kind].sum += score;
+	}
 }
-function extractEventIds(event, args) {
+function applyCaseEnd(draft, event) {
 	const data = asRecord(event.data);
-	return {
-		attemptId: stringFrom(data?.attemptId) ?? event.attemptId ?? args.attemptId,
-		caseId: stringFrom(data?.caseId) ?? event.caseId ?? "",
-		experimentId: stringFrom(data?.experimentId) ?? event.experimentId ?? args.experimentId,
-		projectName: stringFrom(data?.projectName) ?? event.projectName ?? event.projectId ?? args.projectName,
-		runId: stringFrom(data?.runId) ?? event.runId ?? args.runId,
-		taskId: stringFrom(data?.taskId) ?? event.taskId ?? "",
-		workspaceId: stringFrom(data?.workspaceId) ?? event.workspaceId ?? args.workspaceId
-	};
+	draft.caseName = extractCaseName(event) ?? draft.caseName;
+	draft.endedAt = stringFrom(data?.endedAt) ?? event.timestamp ?? draft.endedAt;
+	draft.output = data != null && "output" in data ? data.output : draft.output;
+	draft.state = normalizeState(stringFrom(data?.state)) ?? "failed";
+	draft.scores.exact ??= draft.state === "passed" ? 1 : 0;
 }
-function getOrCreateDraft(drafts, ids, event, args) {
-	const key = createCaseKey(ids.taskId, ids.caseId);
-	const existing = drafts.get(key);
-	if (existing != null) return existing;
-	const draft = {
-		attemptId: ids.attemptId,
-		caseId: ids.caseId,
-		caseName: extractCaseName(event) ?? ids.caseId,
-		experimentId: ids.experimentId,
-		metrics: {},
-		projectName: ids.projectName || args.projectName,
-		retryCount: 0,
-		runId: ids.runId,
-		scores: {},
-		startCount: 0,
-		taskId: ids.taskId,
-		workspaceId: ids.workspaceId
-	};
-	drafts.set(key, draft);
-	return draft;
+function applyCaseMetric(draft, event) {
+	const data = asRecord(event.data);
+	const name = stringFrom(data?.name);
+	if (name == null) return;
+	const value = data?.value;
+	if (isCaseMetricValue(value)) draft.metrics[name] = value;
 }
-function applyIdentity(draft, ids, event, args) {
-	draft.attemptId = ids.attemptId || args.attemptId;
-	draft.experimentId = ids.experimentId || args.experimentId;
-	draft.projectName = extractExplicitProjectName(event) ?? draft.projectName;
-	draft.runId = ids.runId || args.runId;
-	draft.workspaceId = ids.workspaceId || args.workspaceId;
+function applyCaseScore(draft, event) {
+	const data = asRecord(event.data);
+	const kind = stringFrom(data?.kind) ?? stringFrom(data?.name) ?? stringFrom(data?.["vieval.score.kind"]);
+	const score = numberFrom(data?.score) ?? numberFrom(data?.value) ?? numberFrom(data?.["vieval.score.value"]);
+	if (kind == null || score == null) return;
+	draft.scores[kind] = score;
 }
 function applyCaseStart(draft, event) {
 	const data = asRecord(event.data);
@@ -307,66 +295,55 @@ function applyCaseStart(draft, event) {
 	}
 	draft.retryCount = Math.max(draft.retryCount, draft.startCount - 1);
 }
-function applyCaseMetric(draft, event) {
-	const data = asRecord(event.data);
-	const name = stringFrom(data?.name);
-	if (name == null) return;
-	const value = data?.value;
-	if (isCaseMetricValue(value)) draft.metrics[name] = value;
+function applyIdentity(draft, ids, event, args) {
+	draft.attemptId = ids.attemptId || args.attemptId;
+	draft.experimentId = ids.experimentId || args.experimentId;
+	draft.projectName = extractExplicitProjectName(event) ?? draft.projectName;
+	draft.runId = ids.runId || args.runId;
+	draft.workspaceId = ids.workspaceId || args.workspaceId;
 }
-function applyCaseScore(draft, event) {
-	const data = asRecord(event.data);
-	const kind = stringFrom(data?.kind) ?? stringFrom(data?.name) ?? stringFrom(data?.["vieval.score.kind"]);
-	const score = numberFrom(data?.score) ?? numberFrom(data?.value) ?? numberFrom(data?.["vieval.score.value"]);
-	if (kind == null || score == null) return;
-	draft.scores[kind] = score;
+function asRecord(value) {
+	if (value == null || typeof value !== "object" || Array.isArray(value)) return;
+	return value;
 }
-function applyCaseEnd(draft, event) {
-	const data = asRecord(event.data);
-	draft.caseName = extractCaseName(event) ?? draft.caseName;
-	draft.endedAt = stringFrom(data?.endedAt) ?? event.timestamp ?? draft.endedAt;
-	draft.output = data != null && "output" in data ? data.output : draft.output;
-	draft.state = normalizeState(stringFrom(data?.state)) ?? "failed";
-	draft.scores.exact ??= draft.state === "passed" ? 1 : 0;
+/**
+* Normalizes duration timestamps.
+*
+* Before:
+* - `startedAt="2026-05-08T00:00:00.000Z"`, `endedAt="2026-05-08T00:00:01.250Z"`
+* - `startedAt="bad"`, `endedAt="2026-05-08T00:00:01.250Z"`
+*
+* After:
+* - `1250`
+* - `0`
+*/
+function calculateDurationMs(startedAt, endedAt) {
+	const started = Date.parse(startedAt);
+	const ended = Date.parse(endedAt);
+	if (!Number.isFinite(started) || !Number.isFinite(ended)) return 0;
+	return Math.max(0, ended - started);
 }
-function toCaseRecord(draft) {
-	const startedAt = draft.startedAt ?? draft.endedAt ?? "";
-	const endedAt = draft.endedAt ?? startedAt;
+function createCaseKey(taskId, caseId) {
+	return `${taskId}\u0000${caseId}`;
+}
+function extractCaseName(event) {
+	const data = asRecord(event.data);
+	return stringFrom(data?.caseName) ?? stringFrom(data?.name);
+}
+function extractEventIds(event, args) {
+	const data = asRecord(event.data);
 	return {
-		attemptId: draft.attemptId,
-		caseId: draft.caseId,
-		caseName: draft.caseName,
-		durationMs: calculateDurationMs(startedAt, endedAt),
-		endedAt,
-		experimentId: draft.experimentId,
-		...draft.input === void 0 ? {} : { input: draft.input },
-		metrics: draft.metrics,
-		...draft.output === void 0 ? {} : { output: draft.output },
-		projectName: draft.projectName,
-		retryCount: draft.retryCount,
-		runId: draft.runId,
-		schemaVersion: 1,
-		scores: draft.scores,
-		startedAt,
-		state: draft.state ?? "failed",
-		taskId: draft.taskId,
-		workspaceId: draft.workspaceId
+		attemptId: stringFrom(data?.attemptId) ?? event.attemptId ?? args.attemptId,
+		caseId: stringFrom(data?.caseId) ?? event.caseId ?? "",
+		experimentId: stringFrom(data?.experimentId) ?? event.experimentId ?? args.experimentId,
+		projectName: stringFrom(data?.projectName) ?? event.projectName ?? event.projectId ?? args.projectName,
+		runId: stringFrom(data?.runId) ?? event.runId ?? args.runId,
+		taskId: stringFrom(data?.taskId) ?? event.taskId ?? "",
+		workspaceId: stringFrom(data?.workspaceId) ?? event.workspaceId ?? args.workspaceId
 	};
 }
-function addRecordScores(summary, record) {
-	for (const [kind, score] of Object.entries(record.scores)) {
-		if (!Number.isFinite(score)) continue;
-		summary[kind] ??= {
-			average: 0,
-			count: 0,
-			sum: 0
-		};
-		summary[kind].count += 1;
-		summary[kind].sum += score;
-	}
-}
-function finalizeSummaryGroups(groups) {
-	return Object.fromEntries(Object.entries(groups).map(([key, summary]) => [key, finalizeScoreSummary(summary)]));
+function extractExplicitProjectName(event) {
+	return stringFrom(asRecord(event.data)?.projectName) ?? event.projectName ?? event.projectId;
 }
 function finalizeScoreSummary(summary) {
 	return Object.fromEntries(Object.entries(summary).map(([kind, bucket]) => [kind, {
@@ -375,6 +352,9 @@ function finalizeScoreSummary(summary) {
 		sum: bucket.sum
 	}]));
 }
+function finalizeSummaryGroups(groups) {
+	return Object.fromEntries(Object.entries(groups).map(([key, summary]) => [key, finalizeScoreSummary(summary)]));
+}
 function getGroupValue(record, key) {
 	if (Object.hasOwn(record.metrics, key)) return {
 		exists: true,
@@ -386,49 +366,69 @@ function getGroupValue(record, key) {
 		value: directValue
 	} : { exists: false };
 }
-function extractCaseName(event) {
-	const data = asRecord(event.data);
-	return stringFrom(data?.caseName) ?? stringFrom(data?.name);
-}
-function extractExplicitProjectName(event) {
-	return stringFrom(asRecord(event.data)?.projectName) ?? event.projectName ?? event.projectId;
+function getOrCreateDraft(drafts, ids, event, args) {
+	const key = createCaseKey(ids.taskId, ids.caseId);
+	const existing = drafts.get(key);
+	if (existing != null) return existing;
+	const draft = {
+		attemptId: ids.attemptId,
+		caseId: ids.caseId,
+		caseName: extractCaseName(event) ?? ids.caseId,
+		experimentId: ids.experimentId,
+		metrics: {},
+		projectName: ids.projectName || args.projectName,
+		retryCount: 0,
+		runId: ids.runId,
+		scores: {},
+		startCount: 0,
+		taskId: ids.taskId,
+		workspaceId: ids.workspaceId
+	};
+	drafts.set(key, draft);
+	return draft;
 }
-function createCaseKey(taskId, caseId) {
-	return `${taskId}\u0000${caseId}`;
+function isCaseMetricValue(value) {
+	if (value == null || typeof value === "boolean" || typeof value === "number" || typeof value === "string") return true;
+	return Array.isArray(value);
 }
-/**
-* Normalizes duration timestamps.
-*
-* Before:
-* - `startedAt="2026-05-08T00:00:00.000Z"`, `endedAt="2026-05-08T00:00:01.250Z"`
-* - `startedAt="bad"`, `endedAt="2026-05-08T00:00:01.250Z"`
-*
-* After:
-* - `1250`
-* - `0`
-*/
-function calculateDurationMs(startedAt, endedAt) {
-	const started = Date.parse(startedAt);
-	const ended = Date.parse(endedAt);
-	if (!Number.isFinite(started) || !Number.isFinite(ended)) return 0;
-	return Math.max(0, ended - started);
+function normalizeCaseEventName(eventName) {
+	if (eventName === "task.case.start" || eventName === "CaseStarted") return "start";
+	if (eventName === "task.case.metric") return "metric";
+	if (eventName === "task.case.score") return "score";
+	if (eventName === "task.case.end" || eventName === "CaseEnded") return "end";
 }
 function normalizeState(value) {
 	if (value === "failed" || value === "passed" || value === "skipped" || value === "timeout") return value;
 }
-function isCaseMetricValue(value) {
-	if (value == null || typeof value === "boolean" || typeof value === "number" || typeof value === "string") return true;
-	return Array.isArray(value);
-}
-function asRecord(value) {
-	if (value == null || typeof value !== "object" || Array.isArray(value)) return;
-	return value;
+function numberFrom(value) {
+	return typeof value === "number" && Number.isFinite(value) ? value : void 0;
 }
 function stringFrom(value) {
 	return typeof value === "string" ? value : void 0;
 }
-function numberFrom(value) {
-	return typeof value === "number" && Number.isFinite(value) ? value : void 0;
+function toCaseRecord(draft) {
+	const startedAt = draft.startedAt ?? draft.endedAt ?? "";
+	const endedAt = draft.endedAt ?? startedAt;
+	return {
+		attemptId: draft.attemptId,
+		caseId: draft.caseId,
+		caseName: draft.caseName,
+		durationMs: calculateDurationMs(startedAt, endedAt),
+		endedAt,
+		experimentId: draft.experimentId,
+		...draft.input === void 0 ? {} : { input: draft.input },
+		metrics: draft.metrics,
+		...draft.output === void 0 ? {} : { output: draft.output },
+		projectName: draft.projectName,
+		retryCount: draft.retryCount,
+		runId: draft.runId,
+		schemaVersion: 1,
+		scores: draft.scores,
+		startedAt,
+		state: draft.state ?? "failed",
+		taskId: draft.taskId,
+		workspaceId: draft.workspaceId
+	};
 }
 //#endregion
 //#region src/cli/report-selectors.ts
@@ -493,6 +493,28 @@ const reportCasesHelpText = `
     --group-by     Case field, score name, or metric name used for grouped score summaries
 `;
 /**
+* Builds filtered case inspection output.
+*
+* Use when:
+* - `vieval report cases` needs deterministic JSON/table output
+* - tests need pure filtering and grouping behavior without process I/O
+*
+* Expects:
+* - `where` filters use `key=value`
+* - lookup keys may target direct case fields, score names, or metric names
+*
+* Returns:
+* - filtered records plus grouped score summaries when `groupBy` is present
+*/
+function buildReportCasesOutput(records, options) {
+	const whereFilters = (options.where ?? []).map(parseSelector);
+	const filteredRecords = records.filter((record) => matchesWhereFilters(record, whereFilters));
+	return {
+		groups: options.groupBy == null ? void 0 : buildCaseGroups(filteredRecords, options.groupBy),
+		records: [...filteredRecords]
+	};
+}
+/**
 * Reads normalized case records from one report run directory or report root.
 *
 * Use when:
@@ -524,28 +546,6 @@ async function readCaseRecordsFromReport(reportPath) {
 	return records;
 }
 /**
-* Builds filtered case inspection output.
-*
-* Use when:
-* - `vieval report cases` needs deterministic JSON/table output
-* - tests need pure filtering and grouping behavior without process I/O
-*
-* Expects:
-* - `where` filters use `key=value`
-* - lookup keys may target direct case fields, score names, or metric names
-*
-* Returns:
-* - filtered records plus grouped score summaries when `groupBy` is present
-*/
-function buildReportCasesOutput(records, options) {
-	const whereFilters = (options.where ?? []).map(parseSelector);
-	const filteredRecords = records.filter((record) => matchesWhereFilters(record, whereFilters));
-	return {
-		groups: options.groupBy == null ? void 0 : buildCaseGroups(filteredRecords, options.groupBy),
-		records: [...filteredRecords]
-	};
-}
-/**
 * Runs the `vieval report cases` command.
 *
 * Call stack:
@@ -583,66 +583,16 @@ async function runReportCasesCli(argv) {
 		process.exitCode = 1;
 	}
 }
-function normalizeCliArgv$6(argv) {
-	const normalizedArgv = argv[0] === "--" ? argv.slice(1) : [...argv];
-	if (normalizedArgv[0] === "report" && normalizedArgv[1] === "cases") return normalizedArgv.slice(2);
-	if (normalizedArgv[0] === "cases") return normalizedArgv.slice(1);
-	return normalizedArgv;
-}
-function parseReportCasesCliArguments(argv) {
-	const cli = meow(reportCasesHelpText, {
-		argv: normalizeCliArgv$6(argv),
-		flags: {
-			format: {
-				default: "table",
-				type: "string"
-			},
-			groupBy: { type: "string" },
-			where: {
-				isMultiple: true,
-				type: "string"
-			}
-		},
-		importMeta: import.meta
-	});
-	const reportPath = cli.input[0];
-	if (reportPath == null || reportPath.length === 0) throw new Error("Missing required <reportPath> argument.");
-	return {
-		format: normalizeReportCasesFormat(cli.flags.format),
-		groupBy: cli.flags.groupBy,
-		reportPath,
-		where: cli.flags.where
-	};
-}
-function normalizeReportCasesFormat(value) {
-	const normalized = value.toLowerCase();
-	if (normalized === "json") return "json";
-	if (normalized === "jsonl") return "jsonl";
-	return "table";
-}
-async function resolveCaseRecordPaths(reportPath) {
-	const absoluteReportPath = resolve(reportPath);
-	const directCaseFilePath = resolve(absoluteReportPath, "cases.jsonl");
-	if (existsSync(absoluteReportPath) && absoluteReportPath.endsWith(".jsonl")) return [absoluteReportPath];
-	if (existsSync(directCaseFilePath)) return [directCaseFilePath];
-	return (await glob("**/cases.jsonl", {
-		absolute: true,
-		cwd: absoluteReportPath
-	})).sort((left, right) => left.localeCompare(right));
-}
-function matchesWhereFilters(record, whereFilters) {
-	return whereFilters.every((parsed) => {
-		const resolved = getCaseSelectorValue(record, parsed.key);
-		return resolved.exists && String(resolved.value) === parsed.value;
-	});
-}
-function parseSelector(selector) {
-	const separatorIndex = selector.indexOf("=");
-	if (separatorIndex <= 0 || separatorIndex === selector.length - 1) throw new Error(`Invalid selector "${selector}". Expected "key=value".`);
-	return {
-		key: selector.slice(0, separatorIndex).trim(),
-		value: selector.slice(separatorIndex + 1).trim()
-	};
+function addScores(summary, scores) {
+	for (const [scoreName, value] of Object.entries(scores)) {
+		summary[scoreName] ??= {
+			average: 0,
+			count: 0,
+			sum: 0
+		};
+		summary[scoreName].count += 1;
+		summary[scoreName].sum += value;
+	}
 }
 function buildCaseGroups(records, groupBy) {
 	const groups = {};
@@ -662,17 +612,6 @@ function buildCaseGroups(records, groupBy) {
 		scores: finalizeScores(group.scores)
 	}]));
 }
-function addScores(summary, scores) {
-	for (const [scoreName, value] of Object.entries(scores)) {
-		summary[scoreName] ??= {
-			average: 0,
-			count: 0,
-			sum: 0
-		};
-		summary[scoreName].count += 1;
-		summary[scoreName].sum += value;
-	}
-}
 function finalizeScores(summary) {
 	return Object.fromEntries(Object.entries(summary).sort(([left], [right]) => left.localeCompare(right)).map(([scoreName, bucket]) => [scoreName, {
 		average: bucket.count === 0 ? 0 : bucket.sum / bucket.count,
@@ -691,6 +630,67 @@ function formatCasesTable(output) {
 	}
 	return lines.join("\n");
 }
+function matchesWhereFilters(record, whereFilters) {
+	return whereFilters.every((parsed) => {
+		const resolved = getCaseSelectorValue(record, parsed.key);
+		return resolved.exists && String(resolved.value) === parsed.value;
+	});
+}
+function normalizeCliArgv$6(argv) {
+	const normalizedArgv = argv[0] === "--" ? argv.slice(1) : [...argv];
+	if (normalizedArgv[0] === "report" && normalizedArgv[1] === "cases") return normalizedArgv.slice(2);
+	if (normalizedArgv[0] === "cases") return normalizedArgv.slice(1);
+	return normalizedArgv;
+}
+function normalizeReportCasesFormat(value) {
+	const normalized = value.toLowerCase();
+	if (normalized === "json") return "json";
+	if (normalized === "jsonl") return "jsonl";
+	return "table";
+}
+function parseReportCasesCliArguments(argv) {
+	const cli = meow(reportCasesHelpText, {
+		argv: normalizeCliArgv$6(argv),
+		flags: {
+			format: {
+				default: "table",
+				type: "string"
+			},
+			groupBy: { type: "string" },
+			where: {
+				isMultiple: true,
+				type: "string"
+			}
+		},
+		importMeta: import.meta
+	});
+	const reportPath = cli.input[0];
+	if (reportPath == null || reportPath.length === 0) throw new Error("Missing required <reportPath> argument.");
+	return {
+		format: normalizeReportCasesFormat(cli.flags.format),
+		groupBy: cli.flags.groupBy,
+		reportPath,
+		where: cli.flags.where
+	};
+}
+function parseSelector(selector) {
+	const separatorIndex = selector.indexOf("=");
+	if (separatorIndex <= 0 || separatorIndex === selector.length - 1) throw new Error(`Invalid selector "${selector}". Expected "key=value".`);
+	return {
+		key: selector.slice(0, separatorIndex).trim(),
+		value: selector.slice(separatorIndex + 1).trim()
+	};
+}
+async function resolveCaseRecordPaths(reportPath) {
+	const absoluteReportPath = resolve(reportPath);
+	const directCaseFilePath = resolve(absoluteReportPath, "cases.jsonl");
+	if (existsSync(absoluteReportPath) && absoluteReportPath.endsWith(".jsonl")) return [absoluteReportPath];
+	if (existsSync(directCaseFilePath)) return [directCaseFilePath];
+	return (await glob("**/cases.jsonl", {
+		absolute: true,
+		cwd: absoluteReportPath
+	})).sort((left, right) => left.localeCompare(right));
+}
 //#endregion
 //#region src/cli/report-compare.ts
 /**
@@ -735,17 +735,26 @@ function buildCompareReportArtifact(args) {
 		reportPath: args.reportPath
 	};
 }
+/**
+* Writes compare report artifact as JSON.
+*/
+async function writeCompareReportArtifact(args) {
+	const outputPath = resolve(args.outputPath);
+	await mkdir(dirname(outputPath), { recursive: true });
+	await writeFile(outputPath, `${JSON.stringify(args.artifact, null, 2)}\n`, "utf-8");
+	return outputPath;
+}
 function countCasesForProject(caseRecords, projectName) {
 	return caseRecords.filter((record) => record.projectName === projectName).length;
 }
-function countDistinctCasesForProject(caseRecords, projectName) {
-	return countDistinctCases(caseRecords.filter((record) => record.projectName === projectName));
-}
 function countDistinctCases(caseRecords) {
 	const caseKeys = /* @__PURE__ */ new Set();
 	for (const record of caseRecords) caseKeys.add(`${record.projectName}:${record.taskId}:${record.caseId}`);
 	return caseKeys.size;
 }
+function countDistinctCasesForProject(caseRecords, projectName) {
+	return countDistinctCases(caseRecords.filter((record) => record.projectName === projectName));
+}
 function createWeightedAverage(projects, selectAverage) {
 	let weightedScoreTotal = 0;
 	let weightTotal = 0;
@@ -758,15 +767,6 @@ function createWeightedAverage(projects, selectAverage) {
 	if (weightTotal === 0) return null;
 	return weightedScoreTotal / weightTotal;
 }
-/**
-* Writes compare report artifact as JSON.
-*/
-async function writeCompareReportArtifact(args) {
-	const outputPath = resolve(args.outputPath);
-	await mkdir(dirname(outputPath), { recursive: true });
-	await writeFile(outputPath, `${JSON.stringify(args.artifact, null, 2)}\n`, "utf-8");
-	return outputPath;
-}
 //#endregion
 //#region src/cli/discovery.ts
 /**
@@ -927,21 +927,22 @@ function buildLocalOtlpProjection(args) {
 		}] }] }
 	};
 }
-function toAttributes(attributes) {
-	return Object.entries(attributes).filter(([, value]) => value !== void 0).sort(([leftKey], [rightKey]) => leftKey.localeCompare(rightKey)).map(([key, value]) => ({
-		key,
-		value: toAnyValue(value)
-	}));
+function collectProjectNames(records) {
+	return [...new Set(records.map((record) => record.projectName))].sort((left, right) => left.localeCompare(right));
 }
-function toAnyValue(value) {
-	if (Array.isArray(value)) return { arrayValue: { values: value.map((item) => toAnyValue(item)) } };
-	if (isAttributeScalar(value)) {
-		if (typeof value === "boolean") return { boolValue: value };
-		if (typeof value === "number") return Number.isFinite(value) ? { doubleValue: value } : { stringValue: String(value) };
-		if (value == null) return { stringValue: "null" };
-		return { stringValue: value };
-	}
-	return { stringValue: stableStringify(value) };
+function collectScoreKinds(records) {
+	return [...new Set(records.flatMap((record) => Object.keys(record.scores)))].sort((left, right) => left.localeCompare(right));
+}
+function collectTasks(records) {
+	const tasks = /* @__PURE__ */ new Map();
+	for (const record of records) tasks.set(`${record.projectName}\0${record.taskId}`, {
+		projectName: record.projectName,
+		taskId: record.taskId
+	});
+	return [...tasks.values()].sort((left, right) => {
+		const projectOrder = left.projectName.localeCompare(right.projectName);
+		return projectOrder === 0 ? left.taskId.localeCompare(right.taskId) : projectOrder;
+	});
 }
 function isAttributeScalar(value) {
 	return value == null || typeof value === "boolean" || typeof value === "number" || typeof value === "string";
@@ -958,43 +959,32 @@ function isoToUnixNano(value) {
 	if (!Number.isFinite(unixMilliseconds)) return "0";
 	return String(BigInt(unixMilliseconds) * 1000000n);
 }
-function collectScoreKinds(records) {
-	return [...new Set(records.flatMap((record) => Object.keys(record.scores)))].sort((left, right) => left.localeCompare(right));
-}
-function collectProjectNames(records) {
-	return [...new Set(records.map((record) => record.projectName))].sort((left, right) => left.localeCompare(right));
+function toAnyValue(value) {
+	if (Array.isArray(value)) return { arrayValue: { values: value.map((item) => toAnyValue(item)) } };
+	if (isAttributeScalar(value)) {
+		if (typeof value === "boolean") return { boolValue: value };
+		if (typeof value === "number") return Number.isFinite(value) ? { doubleValue: value } : { stringValue: String(value) };
+		if (value == null) return { stringValue: "null" };
+		return { stringValue: value };
+	}
+	return { stringValue: stableStringify(value) };
 }
-function collectTasks(records) {
-	const tasks = /* @__PURE__ */ new Map();
-	for (const record of records) tasks.set(`${record.projectName}\0${record.taskId}`, {
-		projectName: record.projectName,
-		taskId: record.taskId
-	});
-	return [...tasks.values()].sort((left, right) => {
-		const projectOrder = left.projectName.localeCompare(right.projectName);
-		return projectOrder === 0 ? left.taskId.localeCompare(right.taskId) : projectOrder;
-	});
+function toAttributes(attributes) {
+	return Object.entries(attributes).filter(([, value]) => value !== void 0).sort(([leftKey], [rightKey]) => leftKey.localeCompare(rightKey)).map(([key, value]) => ({
+		key,
+		value: toAnyValue(value)
+	}));
 }
 //#endregion
 //#region src/cli/report-artifacts.ts
 /**
-* Resolves one or more `run-summary.json` paths from a report location.
+* Reads all run artifacts found under `reportPath`.
 *
 * Use when:
-* - callers may pass a run directory, summary file path, or a report root
-*
-* Returns:
-* - sorted absolute summary file paths
+* - callers need multi-run analysis from a directory root
 */
-async function resolveRunSummaryPaths(reportPath) {
-	const absoluteReportPath = resolve(reportPath);
-	const directSummaryPath = resolve(absoluteReportPath, "run-summary.json");
-	if (existsSync(absoluteReportPath) && absoluteReportPath.endsWith(".json")) return [absoluteReportPath];
-	if (existsSync(directSummaryPath)) return [directSummaryPath];
-	return (await glob("**/run-summary.json", {
-		absolute: true,
-		cwd: absoluteReportPath
-	})).sort((left, right) => left.localeCompare(right));
+async function readReportArtifacts(reportPath) {
+	return (await resolveRunSummaryPaths(reportPath)).map((summaryFilePath) => readReportRunArtifact(summaryFilePath));
 }
 /**
 * Reads one run report artifact set from `run-summary.json` and sibling `events.jsonl`.
@@ -1031,13 +1021,23 @@ function readReportRunArtifact(summaryFilePath) {
 	};
 }
 /**
-* Reads all run artifacts found under `reportPath`.
+* Resolves one or more `run-summary.json` paths from a report location.
 *
 * Use when:
-* - callers need multi-run analysis from a directory root
+* - callers may pass a run directory, summary file path, or a report root
+*
+* Returns:
+* - sorted absolute summary file paths
 */
-async function readReportArtifacts(reportPath) {
-	return (await resolveRunSummaryPaths(reportPath)).map((summaryFilePath) => readReportRunArtifact(summaryFilePath));
+async function resolveRunSummaryPaths(reportPath) {
+	const absoluteReportPath = resolve(reportPath);
+	const directSummaryPath = resolve(absoluteReportPath, "run-summary.json");
+	if (existsSync(absoluteReportPath) && absoluteReportPath.endsWith(".json")) return [absoluteReportPath];
+	if (existsSync(directSummaryPath)) return [directSummaryPath];
+	return (await glob("**/run-summary.json", {
+		absolute: true,
+		cwd: absoluteReportPath
+	})).sort((left, right) => left.localeCompare(right));
 }
 /**
 * Creates a compact summary row for one run artifact.
@@ -1138,14 +1138,14 @@ function sanitizeIdentitySegment$1(value) {
 */
 function createNoopReporter() {
 	return {
-		onRunStart(_payload) {},
-		onTaskQueued(_payload) {},
-		onTaskStart(_payload) {},
-		onCaseStart(_payload) {},
+		dispose() {},
 		onCaseEnd(_payload) {},
-		onTaskEnd(_payload) {},
+		onCaseStart(_payload) {},
 		onRunEnd(_payload) {},
-		dispose() {}
+		onRunStart(_payload) {},
+		onTaskEnd(_payload) {},
+		onTaskQueued(_payload) {},
+		onTaskStart(_payload) {}
 	};
 }
 //#endregion
@@ -1154,72 +1154,91 @@ const POINTER = "❯";
 const TREE_NODE_END = "└";
 const TREE_NODE_MIDDLE = "├";
 var SummaryReporterStateMachine = class {
-	options;
-	taskCounters = createCounterState();
 	caseCounters = createCounterState();
-	tasks = /* @__PURE__ */ new Map();
+	options;
 	queueOrderCounter = 0;
 	startedAtMs = 0;
 	startTime = "";
+	taskCounters = createCounterState();
+	tasks = /* @__PURE__ */ new Map();
 	constructor(options) {
 		this.options = options;
 	}
 	/**
-	* Handles run startup.
+	* Releases reporter resources.
 	*
 	* Use when:
-	* - a new CLI run is starting and the summary state must reset
+	* - CLI cleanup runs from a `finally` block
 	*
 	* Expects:
-	* - `totalTasks` matches the scheduled task count for the run
+	* - repeated calls are safe
 	*
 	* Returns:
 	* - no direct value
 	*/
-	onRunStart(payload) {
-		this.tasks.clear();
-		this.queueOrderCounter = 0;
-		resetCounterState(this.taskCounters, payload.totalTasks);
-		resetCounterState(this.caseCounters, 0);
-		this.startedAtMs = this.options.getNow();
-		this.startTime = formatTimeString(new Date(this.options.getWallClockNow()));
-	}
+	dispose() {}
 	/**
-	* Handles task queue events.
+	* Builds the current live summary window rows.
 	*
 	* Use when:
-	* - a scheduled task becomes visible in the live summary before it starts
+	* - the live reporter or tests need a snapshot of the active window
 	*
 	* Expects:
-	* - `taskId` is stable across later lifecycle events
+	* - `maxRows`, when present, keeps footer rows visible
 	*
 	* Returns:
-	* - no direct value
+	* - terminal rows in display order
 	*/
-	onTaskQueued(payload) {
-		const task = this.getOrCreateTaskState(payload.taskId);
-		if (task.state === "finished") return;
-		task.displayName = payload.displayName ?? task.displayName;
-		task.projectName = payload.projectName ?? task.projectName;
-		this.syncTaskTotalCases(task, payload.totalCases);
-	}
-	/**
-	* Handles task start events.
-	*
-	* Use when:
-	* - a queued task begins executing
+	getWindowRows(options) {
+		const activeRows = this.createActiveRows();
+		const footerRows = this.createFooterRows();
+		const maxRows = options?.maxRows;
+		const footerBlock = [...footerRows, ""];
+		if (maxRows == null || maxRows <= 0) return [...[
+			"",
+			...activeRows,
+			...activeRows.length > 0 ? [""] : []
+		], ...footerBlock];
+		if (maxRows <= footerBlock.length) return footerBlock.slice(-maxRows);
+		return [...createBoundedActiveBlock(activeRows, Math.max(0, maxRows - footerBlock.length)), ...footerBlock];
+	}
+	/**
+	* Handles case completion.
+	*
+	* Use when:
+	* - a running case settles and counters must advance
 	*
 	* Expects:
-	* - the task was previously queued or can be synthesized from its identifier
+	* - duplicate completion for the same `caseId` is ignored
 	*
 	* Returns:
 	* - no direct value
 	*/
-	onTaskStart(payload) {
+	onCaseEnd(payload) {
 		const task = this.getOrCreateTaskState(payload.taskId);
 		if (task.state === "finished") return;
-		task.state = "running";
-		task.startedAt ??= this.options.getNow();
+		if (task.settledCaseIds.has(payload.caseId)) {
+			task.runningCases.delete(payload.caseId);
+			return;
+		}
+		task.settledCaseIds.add(payload.caseId);
+		task.runningCases.delete(payload.caseId);
+		task.completedCases += 1;
+		this.syncTaskTotalCases(task);
+		this.caseCounters.completed += 1;
+		if (payload.state === "passed") {
+			this.caseCounters.passed += 1;
+			return;
+		}
+		if (payload.state === "failed") {
+			this.caseCounters.failed += 1;
+			return;
+		}
+		if (payload.state === "timeout") {
+			this.caseCounters.timeout += 1;
+			return;
+		}
+		this.caseCounters.skipped += 1;
 	}
 	/**
 	* Handles case start events.
@@ -1258,42 +1277,43 @@ var SummaryReporterStateMachine = class {
 		this.syncTaskTotalCases(task);
 	}
 	/**
-	* Handles case completion.
+	* Handles run completion.
 	*
 	* Use when:
-	* - a running case settles and counters must advance
+	* - the caller has final task totals and wants the footer normalized
 	*
 	* Expects:
-	* - duplicate completion for the same `caseId` is ignored
+	* - payload counters are final terminal task totals
 	*
 	* Returns:
 	* - no direct value
 	*/
-	onCaseEnd(payload) {
-		const task = this.getOrCreateTaskState(payload.taskId);
-		if (task.state === "finished") return;
-		if (task.settledCaseIds.has(payload.caseId)) {
-			task.runningCases.delete(payload.caseId);
-			return;
-		}
-		task.settledCaseIds.add(payload.caseId);
-		task.runningCases.delete(payload.caseId);
-		task.completedCases += 1;
-		this.syncTaskTotalCases(task);
-		this.caseCounters.completed += 1;
-		if (payload.state === "passed") {
-			this.caseCounters.passed += 1;
-			return;
-		}
-		if (payload.state === "failed") {
-			this.caseCounters.failed += 1;
-			return;
-		}
-		if (payload.state === "timeout") {
-			this.caseCounters.timeout += 1;
-			return;
-		}
-		this.caseCounters.skipped += 1;
+	onRunEnd(payload) {
+		this.taskCounters.total = payload.totalTasks;
+		this.taskCounters.passed = payload.passedTasks;
+		this.taskCounters.failed = payload.failedTasks;
+		this.taskCounters.skipped = payload.skippedTasks;
+		this.taskCounters.completed = payload.passedTasks + payload.failedTasks + payload.skippedTasks;
+	}
+	/**
+	* Handles run startup.
+	*
+	* Use when:
+	* - a new CLI run is starting and the summary state must reset
+	*
+	* Expects:
+	* - `totalTasks` matches the scheduled task count for the run
+	*
+	* Returns:
+	* - no direct value
+	*/
+	onRunStart(payload) {
+		this.tasks.clear();
+		this.queueOrderCounter = 0;
+		resetCounterState(this.taskCounters, payload.totalTasks);
+		resetCounterState(this.caseCounters, 0);
+		this.startedAtMs = this.options.getNow();
+		this.startTime = formatTimeString(new Date(this.options.getWallClockNow()));
 	}
 	/**
 	* Handles task completion.
@@ -1326,61 +1346,41 @@ var SummaryReporterStateMachine = class {
 		this.taskCounters.skipped += 1;
 	}
 	/**
-	* Handles run completion.
+	* Handles task queue events.
 	*
 	* Use when:
-	* - the caller has final task totals and wants the footer normalized
+	* - a scheduled task becomes visible in the live summary before it starts
 	*
 	* Expects:
-	* - payload counters are final terminal task totals
+	* - `taskId` is stable across later lifecycle events
 	*
 	* Returns:
 	* - no direct value
 	*/
-	onRunEnd(payload) {
-		this.taskCounters.total = payload.totalTasks;
-		this.taskCounters.passed = payload.passedTasks;
-		this.taskCounters.failed = payload.failedTasks;
-		this.taskCounters.skipped = payload.skippedTasks;
-		this.taskCounters.completed = payload.passedTasks + payload.failedTasks + payload.skippedTasks;
+	onTaskQueued(payload) {
+		const task = this.getOrCreateTaskState(payload.taskId);
+		if (task.state === "finished") return;
+		task.displayName = payload.displayName ?? task.displayName;
+		task.projectName = payload.projectName ?? task.projectName;
+		this.syncTaskTotalCases(task, payload.totalCases);
 	}
 	/**
-	* Releases reporter resources.
+	* Handles task start events.
 	*
 	* Use when:
-	* - CLI cleanup runs from a `finally` block
+	* - a queued task begins executing
 	*
 	* Expects:
-	* - repeated calls are safe
+	* - the task was previously queued or can be synthesized from its identifier
 	*
 	* Returns:
 	* - no direct value
 	*/
-	dispose() {}
-	/**
-	* Builds the current live summary window rows.
-	*
-	* Use when:
-	* - the live reporter or tests need a snapshot of the active window
-	*
-	* Expects:
-	* - `maxRows`, when present, keeps footer rows visible
-	*
-	* Returns:
-	* - terminal rows in display order
-	*/
-	getWindowRows(options) {
-		const activeRows = this.createActiveRows();
-		const footerRows = this.createFooterRows();
-		const maxRows = options?.maxRows;
-		const footerBlock = [...footerRows, ""];
-		if (maxRows == null || maxRows <= 0) return [...[
-			"",
-			...activeRows,
-			...activeRows.length > 0 ? [""] : []
-		], ...footerBlock];
-		if (maxRows <= footerBlock.length) return footerBlock.slice(-maxRows);
-		return [...createBoundedActiveBlock(activeRows, Math.max(0, maxRows - footerBlock.length)), ...footerBlock];
+	onTaskStart(payload) {
+		const task = this.getOrCreateTaskState(payload.taskId);
+		if (task.state === "finished") return;
+		task.state = "running";
+		task.startedAt ??= this.options.getNow();
 	}
 	createActiveRows() {
 		const activeTasks = Array.from(this.tasks.values()).filter((task) => task.state !== "finished").sort(compareActiveTasks);
@@ -1449,6 +1449,49 @@ var SummaryReporterStateMachine = class {
 	}
 };
 /**
+* Creates the live summary reporter state machine for `vieval` CLI runs.
+*
+* Use when:
+* - the CLI wants Vitest-style active rows and live counters
+* - tests need a deterministic reporter surface without touching the terminal
+*
+* Expects:
+* - queue/start/end events describe task lifecycle in order
+* - `getNow()` remains monotonic within one run
+* - `getWallClockNow()` returns the wall-clock run start timestamp
+*
+* Returns:
+* - a reporter compatible with the base CLI lifecycle plus `getWindowRows()`
+*
+* Call stack:
+*
+* {@link createSummaryReporter}
+*   -> {@link SummaryReporterStateMachine.onTaskQueued}
+*   -> {@link SummaryReporterStateMachine.onCaseStart}
+*   -> {@link SummaryReporterStateMachine.getWindowRows}
+*/
+function createSummaryReporter(options) {
+	return new SummaryReporterStateMachine(options);
+}
+function compareActiveTasks(left, right) {
+	const leftProject = left.projectName ?? "";
+	const rightProject = right.projectName ?? "";
+	if (leftProject !== rightProject) return leftProject.localeCompare(rightProject);
+	const displayNameOrder = left.displayName.localeCompare(right.displayName);
+	if (displayNameOrder !== 0) return displayNameOrder;
+	return left.queueOrder - right.queueOrder;
+}
+function countRunningCases(tasks) {
+	let runningCount = 0;
+	for (const task of tasks) runningCount += task.runningCases.size;
+	return runningCount;
+}
+function countRunningTasks(tasks) {
+	let runningCount = 0;
+	for (const task of tasks) if (task.state === "running") runningCount += 1;
+	return runningCount;
+}
+/**
 * Creates the active task block while keeping room for summary footer rows.
 *
 * Use when:
@@ -1481,31 +1524,6 @@ function createBoundedActiveBlock(activeRows, maxRows) {
 		c.dim(`   ${TREE_NODE_END} ... ${hiddenRows} more running rows hidden`)
 	];
 }
-/**
-* Creates the live summary reporter state machine for `vieval` CLI runs.
-*
-* Use when:
-* - the CLI wants Vitest-style active rows and live counters
-* - tests need a deterministic reporter surface without touching the terminal
-*
-* Expects:
-* - queue/start/end events describe task lifecycle in order
-* - `getNow()` remains monotonic within one run
-* - `getWallClockNow()` returns the wall-clock run start timestamp
-*
-* Returns:
-* - a reporter compatible with the base CLI lifecycle plus `getWindowRows()`
-*
-* Call stack:
-*
-* {@link createSummaryReporter}
-*   -> {@link SummaryReporterStateMachine.onTaskQueued}
-*   -> {@link SummaryReporterStateMachine.onCaseStart}
-*   -> {@link SummaryReporterStateMachine.getWindowRows}
-*/
-function createSummaryReporter(options) {
-	return new SummaryReporterStateMachine(options);
-}
 function createCounterState() {
 	return {
 		completed: 0,
@@ -1516,29 +1534,17 @@ function createCounterState() {
 		total: 0
 	};
 }
-function resetCounterState(counter, total) {
-	counter.completed = 0;
-	counter.failed = 0;
-	counter.passed = 0;
-	counter.skipped = 0;
-	counter.timeout = 0;
-	counter.total = total;
-}
-function sumTaskCaseTotals(tasks) {
-	let total = 0;
-	for (const task of tasks) total += task.totalCases;
-	return total;
+function estimateTaskDurationMs(task, now) {
+	if (task.startedAt == null) return;
+	return estimateTotalDurationMs(task.completedCases, task.totalCases, Math.max(0, now - task.startedAt));
 }
-function compareActiveTasks(left, right) {
-	const leftProject = left.projectName ?? "";
-	const rightProject = right.projectName ?? "";
-	if (leftProject !== rightProject) return leftProject.localeCompare(rightProject);
-	const displayNameOrder = left.displayName.localeCompare(right.displayName);
-	if (displayNameOrder !== 0) return displayNameOrder;
-	return left.queueOrder - right.queueOrder;
+function estimateTotalDurationMs(completedCount, totalCount, elapsedDurationMs) {
+	if (completedCount === 0 || totalCount === 0) return;
+	const averageDurationMs = elapsedDurationMs / completedCount;
+	return Math.round(averageDurationMs * totalCount);
 }
-function padSummaryTitle(label) {
-	return `${c.dim(label.padEnd(8))} `;
+function formatActiveConcurrencyState(options) {
+	return [options.taskRunningCount > 0 ? c.bold(c.yellow(`${options.taskRunningCount} ${pluralize("task", options.taskRunningCount)} running`)) : c.dim("0 tasks running"), options.caseRunningCount > 0 ? c.bold(c.yellow(`${options.caseRunningCount} ${pluralize("case", options.caseRunningCount)} running`)) : c.dim("0 cases running")].join(c.dim(" | "));
 }
 function formatCounterState(counter, runningCount, timing) {
 	const plannedCount = Math.max(0, counter.total - counter.completed - runningCount);
@@ -1551,19 +1557,6 @@ function formatCounterState(counter, runningCount, timing) {
 		counter.skipped > 0 ? c.yellow(`${counter.skipped} skipped`) : c.dim(`${counter.skipped} skipped`)
 	].join(c.dim(" | ")) + c.gray(` (${counter.total})`) + formatTimingSuffix(timing);
 }
-function formatActiveConcurrencyState(options) {
-	return [options.taskRunningCount > 0 ? c.bold(c.yellow(`${options.taskRunningCount} ${pluralize("task", options.taskRunningCount)} running`)) : c.dim("0 tasks running"), options.caseRunningCount > 0 ? c.bold(c.yellow(`${options.caseRunningCount} ${pluralize("case", options.caseRunningCount)} running`)) : c.dim("0 cases running")].join(c.dim(" | "));
-}
-function pluralize(noun, count) {
-	return count === 1 ? noun : `${noun}s`;
-}
-function formatRetrySuffix(activeCase) {
-	if (activeCase.retryIndex == null || activeCase.retryIndex <= 0 || activeCase.autoRetry == null || activeCase.autoRetry <= 0) return "";
-	return c.dim(` retry ${activeCase.retryIndex}/${activeCase.autoRetry}`);
-}
-function formatTimeString(date) {
-	return date.toTimeString().split(" ")[0] ?? "";
-}
 function formatDuration$2(durationMs) {
 	return formatHumanDuration(durationMs);
 }
@@ -1595,24 +1588,9 @@ function formatProjectBadge(projectName, isTTY) {
 	const background = backgroundPool[projectName.split("").reduce((accumulator, character, index) => accumulator + character.charCodeAt(0) + index, 0) % backgroundPool.length];
 	return `${c.black(background(` ${projectName} `))} `;
 }
-function countRunningCases(tasks) {
-	let runningCount = 0;
-	for (const task of tasks) runningCount += task.runningCases.size;
-	return runningCount;
-}
-function countRunningTasks(tasks) {
-	let runningCount = 0;
-	for (const task of tasks) if (task.state === "running") runningCount += 1;
-	return runningCount;
-}
-function estimateTaskDurationMs(task, now) {
-	if (task.startedAt == null) return;
-	return estimateTotalDurationMs(task.completedCases, task.totalCases, Math.max(0, now - task.startedAt));
-}
-function estimateTotalDurationMs(completedCount, totalCount, elapsedDurationMs) {
-	if (completedCount === 0 || totalCount === 0) return;
-	const averageDurationMs = elapsedDurationMs / completedCount;
-	return Math.round(averageDurationMs * totalCount);
+function formatRetrySuffix(activeCase) {
+	if (activeCase.retryIndex == null || activeCase.retryIndex <= 0 || activeCase.autoRetry == null || activeCase.autoRetry <= 0) return "";
+	return c.dim(` retry ${activeCase.retryIndex}/${activeCase.autoRetry}`);
 }
 function formatTaskProgressSuffix(task, now) {
 	const elapsedDurationMs = task.startedAt == null ? 0 : Math.max(0, now - task.startedAt);
@@ -1621,11 +1599,33 @@ function formatTaskProgressSuffix(task, now) {
 		estimatedDurationMs: estimateTaskDurationMs(task, now)
 	})}`;
 }
+function formatTimeString(date) {
+	return date.toTimeString().split(" ")[0] ?? "";
+}
 function formatTimingSuffix(timing) {
 	const parts = [`elapsed ${formatHumanDuration(timing.elapsedDurationMs)}`];
 	if (timing.estimatedDurationMs != null) parts.push(`estimated ${formatHumanDuration(timing.estimatedDurationMs)}`);
 	return ` (${parts.join(", ")})`;
 }
+function padSummaryTitle(label) {
+	return `${c.dim(label.padEnd(8))} `;
+}
+function pluralize(noun, count) {
+	return count === 1 ? noun : `${noun}s`;
+}
+function resetCounterState(counter, total) {
+	counter.completed = 0;
+	counter.failed = 0;
+	counter.passed = 0;
+	counter.skipped = 0;
+	counter.timeout = 0;
+	counter.total = total;
+}
+function sumTaskCaseTotals(tasks) {
+	let total = 0;
+	for (const task of tasks) total += task.totalCases;
+	return total;
+}
 //#endregion
 //#region src/cli/reporters/index.ts
 /**
@@ -1676,14 +1676,14 @@ const SYNC_END = `${ESC}?2026l`;
 *       -> {@link WindowRenderer.renderWindow}
 */
 var WindowRenderer = class {
+	bufferedOutput = "";
+	finished = false;
 	options;
 	renderInterval;
 	renderScheduled = false;
 	renderScheduleVersion = 0;
-	windowHeight = 0;
 	started = false;
-	finished = false;
-	bufferedOutput = "";
+	windowHeight = 0;
 	constructor(options) {
 		if (options.createInterval && options.clearInterval) {
 			this.options = {
@@ -1714,26 +1714,41 @@ var WindowRenderer = class {
 		};
 	}
 	/**
-	* Starts the periodic refresh loop.
+	* Stops the renderer and clears any visible window state.
 	*
 	* Use when:
-	* - the live reporter is about to emit in-place updates
+	* - cleanup needs to happen from a `finally` block or interrupted run
 	*
 	* Expects:
-	* - repeated calls are harmless and keep the existing timer
+	* - callers may invoke it more than once
 	*
 	* Returns:
 	* - no direct value
 	*/
-	start() {
-		if (this.started && !this.finished) return;
-		this.started = true;
-		this.finished = false;
+	dispose() {
+		this.finish();
+	}
+	/**
+	* Clears the rendered window and stops the refresh loop.
+	*
+	* Use when:
+	* - the live reporter is transitioning to final static output
+	*
+	* Expects:
+	* - repeated calls are safe
+	*
+	* Returns:
+	* - no direct value
+	*/
+	finish() {
+		if (this.finished) return;
+		this.finished = true;
+		this.started = false;
 		this.renderScheduleVersion += 1;
-		if (!this.renderInterval) {
-			this.renderInterval = this.options.createInterval(() => this.schedule(), this.options.intervalMs);
-			this.renderInterval.unref?.();
-		}
+		this.renderScheduled = false;
+		this.stopInterval();
+		this.clearWindow();
+		this.flushBufferedOutput();
 	}
 	/**
 	* Queues a render if one is not already in flight.
@@ -1758,41 +1773,26 @@ var WindowRenderer = class {
 		});
 	}
 	/**
-	* Clears the rendered window and stops the refresh loop.
+	* Starts the periodic refresh loop.
 	*
 	* Use when:
-	* - the live reporter is transitioning to final static output
+	* - the live reporter is about to emit in-place updates
 	*
 	* Expects:
-	* - repeated calls are safe
+	* - repeated calls are harmless and keep the existing timer
 	*
 	* Returns:
 	* - no direct value
 	*/
-	finish() {
-		if (this.finished) return;
-		this.finished = true;
-		this.started = false;
+	start() {
+		if (this.started && !this.finished) return;
+		this.started = true;
+		this.finished = false;
 		this.renderScheduleVersion += 1;
-		this.renderScheduled = false;
-		this.stopInterval();
-		this.clearWindow();
-		this.flushBufferedOutput();
-	}
-	/**
-	* Stops the renderer and clears any visible window state.
-	*
-	* Use when:
-	* - cleanup needs to happen from a `finally` block or interrupted run
-	*
-	* Expects:
-	* - callers may invoke it more than once
-	*
-	* Returns:
-	* - no direct value
-	*/
-	dispose() {
-		this.finish();
+		if (!this.renderInterval) {
+			this.renderInterval = this.options.createInterval(() => this.schedule(), this.options.intervalMs);
+			this.renderInterval.unref?.();
+		}
 	}
 	/**
 	* Alias for disposal to match Vitest's renderer lifecycle naming.
@@ -1830,6 +1830,20 @@ var WindowRenderer = class {
 		}
 		this.bufferedOutput += message;
 	}
+	clearWindow() {
+		if (!this.options.supportsAnsiWindowing || this.windowHeight === 0) return;
+		this.writeOutput(`${CARRIAGE_RETURN}${CLEAR_LINE}`);
+		for (let rowIndex = 1; rowIndex < this.windowHeight; rowIndex += 1) this.writeOutput(`${CARRIAGE_RETURN}${MOVE_CURSOR_ONE_ROW_UP}${CLEAR_LINE}`);
+		this.windowHeight = 0;
+	}
+	flushBufferedOutput() {
+		if (this.bufferedOutput.length === 0) return;
+		this.writeOutput(this.bufferedOutput);
+		this.bufferedOutput = "";
+	}
+	isActiveWindowMode() {
+		return this.started && !this.finished && this.options.supportsAnsiWindowing;
+	}
 	renderWindow() {
 		const windowContent = this.options.getWindow();
 		const rowCount = getRenderedRowCount(windowContent, this.options.getColumns());
@@ -1847,12 +1861,6 @@ var WindowRenderer = class {
 		this.writeOutput("\n");
 		this.windowHeight = 0;
 	}
-	clearWindow() {
-		if (!this.options.supportsAnsiWindowing || this.windowHeight === 0) return;
-		this.writeOutput(`${CARRIAGE_RETURN}${CLEAR_LINE}`);
-		for (let rowIndex = 1; rowIndex < this.windowHeight; rowIndex += 1) this.writeOutput(`${CARRIAGE_RETURN}${MOVE_CURSOR_ONE_ROW_UP}${CLEAR_LINE}`);
-		this.windowHeight = 0;
-	}
 	stopInterval() {
 		if (!this.renderInterval) return;
 		this.renderInterval.clear();
@@ -1861,14 +1869,6 @@ var WindowRenderer = class {
 	writeOutput(message) {
 		this.options.writeOutput(message);
 	}
-	flushBufferedOutput() {
-		if (this.bufferedOutput.length === 0) return;
-		this.writeOutput(this.bufferedOutput);
-		this.bufferedOutput = "";
-	}
-	isActiveWindowMode() {
-		return this.started && !this.finished && this.options.supportsAnsiWindowing;
-	}
 };
 function defaultCreateInterval(callback, intervalMs) {
 	const timer = globalThis.setInterval(callback, intervalMs);
@@ -1895,40 +1895,6 @@ function getTextDisplayWidth(text) {
 }
 //#endregion
 //#region src/cli/reporters/vitest-compat-reporter.ts
-function isReporterReferenceTuple(reference) {
-	return Array.isArray(reference);
-}
-function isAbsoluteLikePath(value) {
-	return value.startsWith("/") || value.startsWith("./") || value.startsWith("../") || /^[A-Z]:[\\/]/i.test(value);
-}
-async function loadReporterModule(path) {
-	if (isAbsoluteLikePath(path)) return import(pathToFileURL(path).href);
-	return import(path);
-}
-function normalizeReporterReference(reference) {
-	if (isReporterReferenceTuple(reference)) return {
-		options: reference[1],
-		value: reference[0]
-	};
-	return {
-		options: void 0,
-		value: reference
-	};
-}
-function createReporterInstance(moduleValue, options) {
-	const value = moduleValue.default ?? moduleValue;
-	if (value == null) return null;
-	if (typeof value === "function") return new value(options);
-	if (typeof value === "object") return value;
-	return null;
-}
-async function emitToReporters(reporters, callback) {
-	await Promise.all(reporters.map(async (reporter) => {
-		try {
-			await callback(reporter);
-		} catch {}
-	}));
-}
 /**
 * Creates a project-level vitest-compatible reporter bridge.
 *
@@ -2020,82 +1986,337 @@ async function createVievalVitestCompatReporterBridge(options) {
 		}
 	};
 }
-//#endregion
-//#region src/cli/run.ts
-/**
-* Returns true when output contains at least one failing project/task/case outcome.
-*/
-function hasRunFailures(output) {
-	return output.projects.some((project) => {
-		if (project.errorMessage != null) return true;
-		if (project.caseSummary != null && (project.caseSummary.failed > 0 || project.caseSummary.timeout > 0)) return true;
-		return (project.caseFailures?.length ?? 0) > 0;
-	});
-}
-function resolveCappedConcurrency(defaultConcurrency, cliConcurrency, fallback) {
-	const effectiveDefault = defaultConcurrency ?? fallback;
-	if (cliConcurrency == null) return effectiveDefault;
-	return Math.min(effectiveDefault, cliConcurrency);
-}
-function resolveOptionalRuntimeTaskConcurrency(defaultConcurrency, cliConcurrency) {
-	return cliConcurrency ?? defaultConcurrency;
+function createReporterInstance(moduleValue, options) {
+	const value = moduleValue.default ?? moduleValue;
+	if (value == null) return null;
+	if (typeof value === "function") return new value(options);
+	if (typeof value === "object") return value;
+	return null;
 }
-function resolveWorkspaceConcurrency(loadedConfig, options) {
-	return resolveCappedConcurrency(loadedConfig.concurrency?.workspace, options.workspaceConcurrency, 1);
+async function emitToReporters(reporters, callback) {
+	await Promise.all(reporters.map(async (reporter) => {
+		try {
+			await callback(reporter);
+		} catch {}
+	}));
 }
-function resolveProjectConcurrency(project, options) {
-	return resolveCappedConcurrency(project.concurrency?.project, options.projectConcurrency, Number.POSITIVE_INFINITY);
+function isAbsoluteLikePath(value) {
+	return value.startsWith("/") || value.startsWith("./") || value.startsWith("../") || /^[A-Z]:[\\/]/i.test(value);
 }
-function resolveTaskConcurrency(project, options) {
-	return resolveCappedConcurrency(project.concurrency?.task, options.taskConcurrency, 1);
+function isReporterReferenceTuple(reference) {
+	return Array.isArray(reference);
 }
-function resolveScheduledTaskConcurrency(project, options) {
-	return Math.min(resolveProjectConcurrency(project, options), resolveTaskConcurrency(project, options));
+async function loadReporterModule(path) {
+	if (isAbsoluteLikePath(path)) return import(pathToFileURL(path).href);
+	return import(path);
 }
-function resolveRuntimeTaskConcurrency(taskConcurrency, project, options) {
-	const attempt = resolveOptionalRuntimeTaskConcurrency(taskConcurrency?.attempt ?? project.concurrency?.attempt, options.attemptConcurrency);
-	const caseConcurrency = resolveOptionalRuntimeTaskConcurrency(taskConcurrency?.case ?? project.concurrency?.case, options.caseConcurrency);
-	if (attempt == null && caseConcurrency == null) return;
-	return {
-		attempt,
-		case: caseConcurrency
+function normalizeReporterReference(reference) {
+	if (isReporterReferenceTuple(reference)) return {
+		options: reference[1],
+		value: reference[0]
 	};
-}
-function createScheduledTaskWithRuntimeConcurrency(task, project, options) {
-	const taskDefinition = task.entry.task;
-	if (taskDefinition == null) return task;
-	const concurrency = resolveRuntimeTaskConcurrency(taskDefinition.concurrency, project, options);
 	return {
-		...task,
-		entry: {
-			...task.entry,
-			task: {
-				...taskDefinition,
-				concurrency
+		options: void 0,
+		value: reference
+	};
+}
+//#endregion
+//#region src/cli/run.ts
+/**
+* Formats CLI run output as human-readable lines.
+*/
+function formatVievalCliRunOutput(output) {
+	const colorEnabled = shouldUseColor();
+	const colors = createColorPalette(colorEnabled);
+	const lines = [];
+	lines.push(` ${colors.dim("RUN")}  ${colors.yellow("vieval")}`);
+	lines.push(` ${colors.dim("Config")}  ${output.configFilePath ?? "(not found, using defaults)"}`);
+	lines.push("");
+	let passedProjects = 0;
+	let skippedProjects = 0;
+	let failedProjects = 0;
+	let totalTasks = 0;
+	let executedTasks = 0;
+	function formatMatrixSummary(summary) {
+		if (summary == null) return null;
+		const runAxesLabel = summary.runAxes.length === 0 ? "-" : summary.runAxes.join("|");
+		const evalAxesLabel = summary.evalAxes.length === 0 ? "-" : summary.evalAxes.join("|");
+		return `matrix run ${summary.runRows} [${runAxesLabel}] / eval ${summary.evalRows} [${evalAxesLabel}]`;
+	}
+	function formatScheduleBreakdown(project) {
+		const summary = project.matrixSummary;
+		if (summary == null) return null;
+		if (project.taskCount <= 0 || project.entryCount <= 0 || summary.runRows <= 0 || summary.evalRows <= 0) return null;
+		const denominator = project.entryCount * summary.runRows * summary.evalRows;
+		if (denominator <= 0 || project.taskCount % denominator !== 0) return null;
+		const providerCount = project.taskCount / denominator;
+		return [
+			colors.dim("schedule "),
+			colors.yellow(String(project.entryCount)),
+			colors.dim(" entries × "),
+			colors.yellow(String(providerCount)),
+			colors.dim(" inferenceExecutors × "),
+			colors.yellow(String(summary.runRows)),
+			colors.dim(" run rows × "),
+			colors.yellow(String(summary.evalRows)),
+			colors.dim(" eval rows = "),
+			colors.green(String(project.taskCount)),
+			colors.dim(" tasks")
+		].join("");
+	}
+	for (const project of output.projects) {
+		totalTasks += project.taskCount;
+		executedTasks += project.result?.overall.runCount ?? 0;
+		const badge = createProjectBadge(project.name, colors, colorEnabled);
+		const isFailed = project.errorMessage != null;
+		const hasFailedCases = (project.caseSummary?.failed ?? 0) > 0 || (project.caseSummary?.timeout ?? 0) > 0 || (project.caseFailures?.length ?? 0) > 0;
+		if (isFailed) {
+			failedProjects += 1;
+			lines.push(` ${colors.red("❯")} ${badge}${formatDuration$1(project.durationMs, colors)}`);
+			lines.push(`   ${project.errorMessage}`);
+			continue;
+		}
+		if (!project.executed) {
+			skippedProjects += 1;
+			const countLabel = colors.dim(`(${project.taskCount} tasks)`);
+			const detailsLabel = colors.dim(` ${project.discoveredEvalFileCount} files, ${project.entryCount} entries, 0 runs, hybrid n/a`);
+			const matrixSummary = formatMatrixSummary(project.matrixSummary);
+			lines.push(` ${colors.dim("○")} ${badge}${countLabel}${detailsLabel}${formatDuration$1(project.durationMs, colors)}`);
+			if (matrixSummary != null) lines.push(`   ${colors.dim(matrixSummary)}`);
+			const scheduleBreakdown = formatScheduleBreakdown(project);
+			if (scheduleBreakdown != null) lines.push(`   ${scheduleBreakdown}`);
+			continue;
+		}
+		if (hasFailedCases) failedProjects += 1;
+		else passedProjects += 1;
+		const hybridAverageLabel = formatHybridAverage(project.result?.overall.hybridAverage);
+		const runCount = project.result?.overall.runCount ?? 0;
+		const countLabel = colors.dim(`(${project.taskCount} tasks)`);
+		const caseSummaryLabel = project.caseSummary == null ? "" : `, cases ${project.caseSummary.passed} passed | ${project.caseSummary.failed} failed | ${project.caseSummary.timeout} timeout`;
+		const detailsLabel = colors.dim(` ${project.discoveredEvalFileCount} files, ${project.entryCount} entries, ${runCount} runs${caseSummaryLabel}, hybrid ${hybridAverageLabel}`);
+		const matrixSummary = formatMatrixSummary(project.matrixSummary);
+		lines.push(` ${hasFailedCases ? colors.red("❯") : colors.green("✓")} ${badge}${countLabel}${detailsLabel}${formatDuration$1(project.durationMs, colors)}`);
+		if (matrixSummary != null) lines.push(`   ${colors.dim(matrixSummary)}`);
+		const scheduleBreakdown = formatScheduleBreakdown(project);
+		if (scheduleBreakdown != null) lines.push(`   ${scheduleBreakdown}`);
+		if ((project.caseFailures?.length ?? 0) > 0) {
+			lines.push(`   ${colors.red("Failed cases:")}`);
+			for (const failure of project.caseFailures.slice(0, 5)) {
+				lines.push(`   ${colors.red(`- ${failure.caseName} (${failure.taskId})`)}`);
+				for (const line of failure.errorMessage.split("\n")) lines.push(`     ${colors.red(line)}`);
 			}
+			if (project.caseFailures.length > 5) lines.push(`   ${colors.dim(`... ${project.caseFailures.length - 5} more failed cases`)}`);
+		}
+	}
+	lines.push("");
+	if (failedProjects > 0 || skippedProjects > 0) {
+		const summarySegments = [`${colors.green(String(passedProjects))} passed`];
+		if (skippedProjects > 0) summarySegments.push(`${colors.dim(String(skippedProjects))} skipped`);
+		if (failedProjects > 0) summarySegments.push(`${colors.red(String(failedProjects))} failed`);
+		lines.push(` ${colors.dim("Projects")}  ${summarySegments.join(" | ")} (${output.projects.length})`);
+	} else lines.push(` ${colors.dim("Projects")}  ${colors.green(String(passedProjects))} passed (${output.projects.length})`);
+	lines.push(` ${colors.dim("Tasks")}     ${executedTasks} executed / ${totalTasks} scheduled`);
+	return lines.join("\n");
+}
+/**
+* Returns true when output contains at least one failing project/task/case outcome.
+*/
+function hasRunFailures(output) {
+	return output.projects.some((project) => {
+		if (project.errorMessage != null) return true;
+		if (project.caseSummary != null && (project.caseSummary.failed > 0 || project.caseSummary.timeout > 0)) return true;
+		return (project.caseFailures?.length ?? 0) > 0;
+	});
+}
+/**
+* Runs vieval orchestration from config and returns project-level summaries.
+*
+* Call stack:
+*
+* {@link runVievalCli}
+*   -> {@link loadVievalCliConfig}
+*   -> {@link discoverEvalFiles}
+*   -> {@link collectEvalEntries}
+*   -> {@link createRunnerSchedule}
+*   -> {@link runScheduledTasks} (optional)
+*
+* Use when:
+* - running eval collection and scheduling from a single command
+* - keeping business-agent eval files near their implementation packages
+*/
+async function runVievalCli(options = {}) {
+	const loadedConfig = await loadVievalCliConfig({
+		configFilePath: options.configFilePath,
+		cwd: options.cwd
+	});
+	const telemetry = loadedConfig.reporting?.openTelemetry?.enabled === true ? createOpenTelemetryRuntime() : createNoopTelemetryRuntime();
+	const onOpenTelemetryRunEnd = loadedConfig.reporting?.openTelemetry?.enabled === true ? loadedConfig.reporting.openTelemetry.onRunEnd : void 0;
+	const restoreEnvironment = applyRunEnvironment(loadedConfig.env);
+	let runError;
+	let runEndError;
+	let output;
+	let reporter;
+	try {
+		const selectedProjects = filterProjectsByName(loadedConfig.projects, options.project ?? []);
+		const preparedProjects = await Promise.all(selectedProjects.map(async (project) => prepareProject(project)));
+		const identity = createRunIdentity(options, preparedProjects);
+		const eventRecorder = createEventRecorder(identity);
+		const runReporter = createReporterWithEventCapture(createRunReporter(options.reporter), eventRecorder.record);
+		reporter = runReporter;
+		output = await telemetry.withSpan("vieval.run", {
+			"vieval.attempt.id": identity.attemptId,
+			"vieval.experiment.id": identity.experimentId,
+			"vieval.run.id": identity.runId,
+			"vieval.workspace.id": identity.workspaceId
+		}, async () => {
+			const workspaceScheduler = createSchedulerRuntime({ concurrency: { workspace: resolveWorkspaceConcurrency(loadedConfig, options) } });
+			const executableProjects = preparedProjects.filter((project) => project.kind === "prepared").map((project) => project.prepared);
+			const totalTasks = preparedProjects.reduce((sum, project) => {
+				if (project.kind === "prepared") return sum + project.prepared.tasks.length;
+				return sum + project.summary.taskCount;
+			}, 0);
+			const skippedSummaryTasks = preparedProjects.reduce((sum, project) => {
+				if (project.kind === "summary") return sum + project.summary.taskCount;
+				return sum;
+			}, 0);
+			const reporterCounters = {
+				failedTasks: 0,
+				passedTasks: 0,
+				skippedTasks: 0
+			};
+			runReporter.onRunStart({ totalTasks });
+			for (const project of executableProjects) for (const task of project.tasks) runReporter.onTaskQueued(createTaskQueuePayload(task, project.name));
+			const projectSummaries = (await Promise.all(preparedProjects.map(async (preparedProject, index) => {
+				if (preparedProject.kind === "summary") return {
+					index,
+					summary: preparedProject.summary
+				};
+				return {
+					index,
+					summary: await telemetry.withSpan("vieval.project", {
+						"vieval.project.name": preparedProject.prepared.name,
+						"vieval.run.id": identity.runId
+					}, async () => await workspaceScheduler.runCase({
+						experimentId: identity.experimentId,
+						projectName: preparedProject.prepared.name,
+						scope: "workspace",
+						workspaceId: identity.workspaceId
+					}, async () => executePreparedProject(preparedProject.prepared, identity, options.cacheProjectName, telemetry, runReporter, reporterCounters, eventRecorder.record, options)))
+				};
+			}))).sort((left, right) => left.index - right.index).map((item) => item.summary);
+			runReporter.onRunEnd({
+				failedTasks: reporterCounters.failedTasks,
+				passedTasks: reporterCounters.passedTasks,
+				skippedTasks: reporterCounters.skippedTasks + skippedSummaryTasks,
+				totalTasks
+			});
+			const output = {
+				attemptId: identity.attemptId,
+				configFilePath: loadedConfig.configFilePath,
+				experimentId: identity.experimentId,
+				projects: projectSummaries,
+				reportDirectory: null,
+				runId: identity.runId,
+				workspaceId: identity.workspaceId
+			};
+			if (options.reportOut != null) output.reportDirectory = await writeRunReportArtifacts(output, eventRecorder.events, identity, options.reportOut);
+			return output;
+		});
+	} catch (error) {
+		runError = error;
+	} finally {
+		if (onOpenTelemetryRunEnd != null) try {
+			await onOpenTelemetryRunEnd();
+		} catch (error) {
+			if (runError == null) runEndError = error;
+		}
+		reporter?.dispose();
+		restoreEnvironment();
+	}
+	if (runError != null) throw runError;
+	if (runEndError != null) throw runEndError;
+	if (output == null) throw new Error("Vieval run finished without output.");
+	return output;
+}
+function applyRunEnvironment(env) {
+	const envEntries = Object.entries(env);
+	if (envEntries.length === 0) return () => {};
+	const snapshot = /* @__PURE__ */ new Map();
+	for (const [key, value] of envEntries) {
+		snapshot.set(key, {
+			existed: Object.hasOwn(process.env, key),
+			value: process.env[key]
+		});
+		if (value == null) {
+			delete process.env[key];
+			continue;
+		}
+		process.env[key] = value;
+	}
+	return () => {
+		for (const [key, previous] of snapshot.entries()) {
+			if (previous.existed) {
+				if (previous.value == null) {
+					delete process.env[key];
+					continue;
+				}
+				process.env[key] = previous.value;
+				continue;
+			}
+			delete process.env[key];
 		}
 	};
 }
-function resolveCliRuntimeConcurrency(options) {
-	if (options.attemptConcurrency == null && options.caseConcurrency == null) return;
+function cloneScheduledTaskMatrix(task) {
 	return {
-		attempt: options.attemptConcurrency,
-		case: options.caseConcurrency
+		eval: { ...task.matrix.eval },
+		meta: { ...task.matrix.meta },
+		run: { ...task.matrix.run }
 	};
 }
-function shouldUseColor() {
-	if (process.env.NO_COLOR != null) return false;
-	const forceColor = process.env.FORCE_COLOR;
-	if (forceColor != null) return forceColor !== "0";
-	return process.stdout.isTTY === true;
+function createAutoTaskExecutor(reporter, projectName, recordEvent, projectCaseCounters, projectCaseFailures, vitestCompatReporter) {
+	return async (task, context) => {
+		const taskDefinition = task.entry.task;
+		if (taskDefinition == null) throw new Error(`Missing eval task definition for entry "${task.entry.id}".`);
+		const output = await taskDefinition.run({
+			cache: context.cache,
+			models: context.models,
+			reporterHooks: resolveTaskReporterHooks(task, context, reporter, projectName, recordEvent, projectCaseCounters, projectCaseFailures, vitestCompatReporter),
+			task,
+			telemetry: context.telemetry
+		});
+		return {
+			entryId: task.entry.id,
+			id: task.id,
+			inferenceExecutorId: task.inferenceExecutor.id,
+			matrix: task.matrix,
+			scores: [...output.scores]
+		};
+	};
+}
+function createCliTaskExecutionContext(task, models, cacheRootDirectory, cacheProjectName, workspaceId, telemetry, reporter, projectName, recordEvent, projectCaseCounters, projectCaseFailures, runtimeConcurrency, vitestCompatReporter) {
+	return {
+		...createTaskExecutionContext({
+			cache: createFilesystemTaskCacheRuntime({
+				cacheRootDirectory,
+				projectName: cacheProjectName,
+				workspaceId
+			}),
+			models,
+			task
+		}),
+		reporterHooks: createTaskReporterHooks(task, reporter, projectName, recordEvent, projectCaseCounters, projectCaseFailures, vitestCompatReporter),
+		runtimeConcurrency,
+		telemetry
+	};
 }
 function createColorPalette(enabled) {
 	if (!enabled) return {
-		black: (value) => value,
 		bgCyan: (value) => value,
 		bgGreen: (value) => value,
 		bgMagenta: (value) => value,
 		bgYellow: (value) => value,
+		black: (value) => value,
 		dim: (value) => value,
 		gray: (value) => value,
 		green: (value) => value,
@@ -2103,11 +2324,11 @@ function createColorPalette(enabled) {
 		yellow: (value) => value
 	};
 	return {
-		black: (value) => c.black(value),
 		bgCyan: (value) => c.bgCyan(value),
 		bgGreen: (value) => c.bgGreen(value),
 		bgMagenta: (value) => c.bgMagenta(value),
 		bgYellow: (value) => c.bgYellow(value),
+		black: (value) => c.black(value),
 		dim: (value) => c.dim(value),
 		gray: (value) => c.gray(value),
 		green: (value) => c.green(value),
@@ -2115,67 +2336,6 @@ function createColorPalette(enabled) {
 		yellow: (value) => c.yellow(value)
 	};
 }
-function createProjectBadge(name, colors, colorEnabled) {
-	if (!colorEnabled || !c.isColorSupported) return `|${name}| `;
-	const labelColorPool = [
-		colors.bgYellow,
-		colors.bgCyan,
-		colors.bgGreen,
-		colors.bgMagenta
-	];
-	const background = labelColorPool[name.split("").reduce((accumulator, char, index) => accumulator + char.charCodeAt(0) + index, 0) % labelColorPool.length];
-	return `${colors.black(background(` ${name} `))} `;
-}
-function formatDuration$1(durationMs, colors) {
-	if (durationMs == null) return "";
-	const rounded = Math.round(durationMs);
-	return (rounded > 1e3 ? colors.yellow : colors.green)(` ${rounded}${colors.dim("ms")}`);
-}
-function formatHybridAverage(hybridAverage) {
-	if (hybridAverage == null) return "n/a";
-	return hybridAverage.toFixed(3).replace(/\.?0+$/, "");
-}
-function filterProjectsByName(projects, names) {
-	if (names.length === 0) return [...projects];
-	const nameSet = new Set(names);
-	return projects.filter((project) => nameSet.has(project.name));
-}
-function sanitizeIdentitySegment(value) {
-	const normalized = value.trim();
-	if (normalized.length === 0) return "default";
-	return normalized.replace(/[^\w.-]+/g, "-");
-}
-function createExperimentMatrixRows(tasks) {
-	const rows = /* @__PURE__ */ new Set();
-	for (const task of tasks) {
-		const runRowId = task.matrix.meta.runRowId;
-		const evalRowId = task.matrix.meta.evalRowId;
-		if (runRowId !== "default" && evalRowId !== "default") {
-			rows.add(`run:${runRowId}+eval:${evalRowId}`);
-			continue;
-		}
-		if (runRowId !== "default") rows.add(`run:${runRowId}`);
-		if (evalRowId !== "default") rows.add(`eval:${evalRowId}`);
-	}
-	return [...rows].sort((left, right) => left.localeCompare(right));
-}
-function resolveExperimentId(options, preparedProjects) {
-	if (options.experiment != null) return sanitizeIdentitySegment(options.experiment);
-	const matrixRows = /* @__PURE__ */ new Set();
-	for (const project of preparedProjects) project.experimentMatrixRows.forEach((row) => matrixRows.add(row));
-	if (matrixRows.size === 0) return "default-experiment";
-	return sanitizeIdentitySegment(`matrix-${[...matrixRows].sort().join("--")}`);
-}
-function createRunIdentity(options, preparedProjects) {
-	const workspaceId = sanitizeIdentitySegment(options.workspace ?? "default-workspace");
-	const experimentId = resolveExperimentId(options, preparedProjects);
-	return {
-		attemptId: sanitizeIdentitySegment(options.attempt ?? `attempt-${(/* @__PURE__ */ new Date()).toISOString().replace(/[:.]/g, "-")}`),
-		experimentId,
-		runId: `run-${Date.now()}-${randomUUID().slice(0, 8)}`,
-		workspaceId
-	};
-}
 function createEventRecorder(identity) {
 	const events = [];
 	const taskProjectMap = /* @__PURE__ */ new Map();
@@ -2204,6 +2364,50 @@ function createEventRecorder(identity) {
 		}
 	};
 }
+function createExperimentMatrixRows(tasks) {
+	const rows = /* @__PURE__ */ new Set();
+	for (const task of tasks) {
+		const runRowId = task.matrix.meta.runRowId;
+		const evalRowId = task.matrix.meta.evalRowId;
+		if (runRowId !== "default" && evalRowId !== "default") {
+			rows.add(`run:${runRowId}+eval:${evalRowId}`);
+			continue;
+		}
+		if (runRowId !== "default") rows.add(`run:${runRowId}`);
+		if (evalRowId !== "default") rows.add(`eval:${evalRowId}`);
+	}
+	return [...rows].sort((left, right) => left.localeCompare(right));
+}
+function createProjectBadge(name, colors, colorEnabled) {
+	if (!colorEnabled || !c.isColorSupported) return `|${name}| `;
+	const labelColorPool = [
+		colors.bgYellow,
+		colors.bgCyan,
+		colors.bgGreen,
+		colors.bgMagenta
+	];
+	const background = labelColorPool[name.split("").reduce((accumulator, char, index) => accumulator + char.charCodeAt(0) + index, 0) % labelColorPool.length];
+	return `${colors.black(background(` ${name} `))} `;
+}
+function createProjectMatrixSummary(tasks) {
+	if (tasks.length === 0) return null;
+	const runAxes = /* @__PURE__ */ new Set();
+	const evalAxes = /* @__PURE__ */ new Set();
+	const runRows = /* @__PURE__ */ new Set();
+	const evalRows = /* @__PURE__ */ new Set();
+	for (const task of tasks) {
+		Object.keys(task.matrix.run).forEach((axis) => runAxes.add(axis));
+		Object.keys(task.matrix.eval).forEach((axis) => evalAxes.add(axis));
+		runRows.add(task.matrix.meta.runRowId);
+		evalRows.add(task.matrix.meta.evalRowId);
+	}
+	return {
+		evalAxes: [...evalAxes].sort(),
+		evalRows: evalRows.size,
+		runAxes: [...runAxes].sort(),
+		runRows: runRows.size
+	};
+}
 function createReporterWithEventCapture(reporter, recordEvent) {
 	return {
 		dispose() {
@@ -2239,38 +2443,16 @@ function createReporterWithEventCapture(reporter, recordEvent) {
 		}
 	};
 }
-function applyRunEnvironment(env) {
-	const envEntries = Object.entries(env);
-	if (envEntries.length === 0) return () => {};
-	const snapshot = /* @__PURE__ */ new Map();
-	for (const [key, value] of envEntries) {
-		snapshot.set(key, {
-			existed: Object.hasOwn(process.env, key),
-			value: process.env[key]
-		});
-		if (value == null) {
-			delete process.env[key];
-			continue;
-		}
-		process.env[key] = value;
-	}
-	return () => {
-		for (const [key, previous] of snapshot.entries()) {
-			if (previous.existed) {
-				if (previous.value == null) {
-					delete process.env[key];
-					continue;
-				}
-				process.env[key] = previous.value;
-				continue;
-			}
-			delete process.env[key];
-		}
+function createRunIdentity(options, preparedProjects) {
+	const workspaceId = sanitizeIdentitySegment(options.workspace ?? "default-workspace");
+	const experimentId = resolveExperimentId(options, preparedProjects);
+	return {
+		attemptId: sanitizeIdentitySegment(options.attempt ?? `attempt-${(/* @__PURE__ */ new Date()).toISOString().replace(/[:.]/g, "-")}`),
+		experimentId,
+		runId: `run-${Date.now()}-${randomUUID().slice(0, 8)}`,
+		workspaceId
 	};
 }
-function isSummaryReporter(reporter) {
-	return "getWindowRows" in reporter;
-}
 function createRunReporter(options) {
 	const getRows = options?.getRows ?? (() => process.stdout.rows);
 	const reporter = createCliReporter({
@@ -2342,21 +2524,23 @@ function createRunReporter(options) {
 		}
 	};
 }
-/**
-* Normalizes terminal row count into the live reporter window height.
-*
-* Before:
-* - undefined
-* - 4
-* - 40
-*
-* After:
-* - 23
-* - 6
-* - 39
-*/
-function normalizeLiveReporterMaxRows(rows) {
-	return Math.max(6, (rows == null || !Number.isFinite(rows) || rows <= 0 ? 24 : Math.floor(rows)) - 1);
+function createScheduledTaskWithRuntimeConcurrency(task, project, options) {
+	const taskDefinition = task.entry.task;
+	if (taskDefinition == null) return task;
+	const concurrency = resolveRuntimeTaskConcurrency(taskDefinition.concurrency, project, options);
+	return {
+		...task,
+		entry: {
+			...task.entry,
+			task: {
+				...taskDefinition,
+				concurrency
+			}
+		}
+	};
+}
+function createTaskCaseReporterId(payload) {
+	return `${payload.index}:${encodeURIComponent(payload.name)}`;
 }
 function createTaskQueuePayload(task, projectName) {
 	return {
@@ -2365,9 +2549,6 @@ function createTaskQueuePayload(task, projectName) {
 		taskId: task.id
 	};
 }
-function createTaskCaseReporterId(payload) {
-	return `${payload.index}:${encodeURIComponent(payload.name)}`;
-}
 function createTaskReporterHooks(task, reporter, projectName, recordEvent, projectCaseCounters, projectCaseFailures, vitestCompatReporter) {
 	function syncCaseTotal(total) {
 		reporter.onTaskQueued({
@@ -2415,8 +2596,8 @@ function createTaskReporterHooks(task, reporter, projectName, recordEvent, proje
 			reporter.onCaseStart({
 				autoRetry: payload.autoRetry,
 				caseId,
-				input: payload.input,
 				caseName: payload.name,
+				input: payload.input,
 				retryIndex: payload.retryIndex,
 				taskId: task.id
 			});
@@ -2434,144 +2615,6 @@ function createTaskReporterHooks(task, reporter, projectName, recordEvent, proje
 		}
 	};
 }
-function createCliTaskExecutionContext(task, models, cacheRootDirectory, cacheProjectName, workspaceId, telemetry, reporter, projectName, recordEvent, projectCaseCounters, projectCaseFailures, runtimeConcurrency, vitestCompatReporter) {
-	return {
-		...createTaskExecutionContext({
-			cache: createFilesystemTaskCacheRuntime({
-				cacheRootDirectory,
-				projectName: cacheProjectName,
-				workspaceId
-			}),
-			models,
-			task
-		}),
-		reporterHooks: createTaskReporterHooks(task, reporter, projectName, recordEvent, projectCaseCounters, projectCaseFailures, vitestCompatReporter),
-		runtimeConcurrency,
-		telemetry
-	};
-}
-function resolveTaskReporterHooks(task, context, reporter, projectName, recordEvent, projectCaseCounters, projectCaseFailures, vitestCompatReporter) {
-	return context.reporterHooks ?? createTaskReporterHooks(task, reporter, projectName, recordEvent, projectCaseCounters, projectCaseFailures, vitestCompatReporter);
-}
-function getFailedTaskId(error) {
-	if (error instanceof RunnerExecutionError) return error.taskId;
-	return null;
-}
-function createAutoTaskExecutor(reporter, projectName, recordEvent, projectCaseCounters, projectCaseFailures, vitestCompatReporter) {
-	return async (task, context) => {
-		const taskDefinition = task.entry.task;
-		if (taskDefinition == null) throw new Error(`Missing eval task definition for entry "${task.entry.id}".`);
-		const output = await taskDefinition.run({
-			cache: context.cache,
-			models: context.models,
-			reporterHooks: resolveTaskReporterHooks(task, context, reporter, projectName, recordEvent, projectCaseCounters, projectCaseFailures, vitestCompatReporter),
-			task,
-			telemetry: context.telemetry
-		});
-		return {
-			entryId: task.entry.id,
-			id: task.id,
-			matrix: task.matrix,
-			inferenceExecutorId: task.inferenceExecutor.id,
-			scores: [...output.scores]
-		};
-	};
-}
-function cloneScheduledTaskMatrix(task) {
-	return {
-		eval: { ...task.matrix.eval },
-		meta: { ...task.matrix.meta },
-		run: { ...task.matrix.run }
-	};
-}
-function createProjectMatrixSummary(tasks) {
-	if (tasks.length === 0) return null;
-	const runAxes = /* @__PURE__ */ new Set();
-	const evalAxes = /* @__PURE__ */ new Set();
-	const runRows = /* @__PURE__ */ new Set();
-	const evalRows = /* @__PURE__ */ new Set();
-	for (const task of tasks) {
-		Object.keys(task.matrix.run).forEach((axis) => runAxes.add(axis));
-		Object.keys(task.matrix.eval).forEach((axis) => evalAxes.add(axis));
-		runRows.add(task.matrix.meta.runRowId);
-		evalRows.add(task.matrix.meta.evalRowId);
-	}
-	return {
-		evalAxes: [...evalAxes].sort(),
-		evalRows: evalRows.size,
-		runAxes: [...runAxes].sort(),
-		runRows: runRows.size
-	};
-}
-async function prepareProject(project) {
-	const startedAt = Date.now();
-	try {
-		const runtimeContext = await createRunnerRuntimeContext({
-			cwd: project.root,
-			fallbackProjectRootDirectory: project.root
-		});
-		const evalFilePaths = await discoverEvalFiles({
-			exclude: project.exclude,
-			include: project.include,
-			root: project.root
-		});
-		const entries = collectEvalEntries(await loadEvalModulesWithVitestRuntime(evalFilePaths, project.root), runtimeContext);
-		const tasks = createRunnerSchedule({
-			evalMatrix: project.evalMatrix,
-			entries,
-			inferenceExecutors: project.inferenceExecutors,
-			runMatrix: project.runMatrix
-		});
-		const canAutoExecuteEntryTasks = entries.some((entry) => entry.task != null) && project.models.length > 0;
-		if (project.executor == null && !canAutoExecuteEntryTasks) return {
-			experimentMatrixRows: createExperimentMatrixRows(tasks),
-			kind: "summary",
-			summary: {
-				caseSummary: null,
-				caseFailures: [],
-				discoveredEvalFileCount: evalFilePaths.length,
-				durationMs: Date.now() - startedAt,
-				entryCount: entries.length,
-				errorMessage: null,
-				executed: false,
-				matrixSummary: createProjectMatrixSummary(tasks),
-				name: project.name,
-				result: null,
-				taskCount: tasks.length
-			}
-		};
-		return {
-			experimentMatrixRows: createExperimentMatrixRows(tasks),
-			kind: "prepared",
-			prepared: {
-				discoveredEvalFileCount: evalFilePaths.length,
-				entryCount: entries.length,
-				name: project.name,
-				project,
-				startedAt,
-				tasks
-			}
-		};
-	} catch (error) {
-		return {
-			experimentMatrixRows: [],
-			kind: "summary",
-			summary: {
-				caseSummary: null,
-				caseFailures: [],
-				discoveredEvalFileCount: 0,
-				durationMs: Date.now() - startedAt,
-				entryCount: 0,
-				errorMessage: errorMessageFrom(error) ?? "Unknown project execution error.",
-				executed: false,
-				matrixSummary: null,
-				name: project.name,
-				result: null,
-				taskCount: 0
-			}
-		};
-	}
-}
 async function executePreparedProject(prepared, identity, cacheProjectName, telemetry, reporter, counters, recordEvent, options) {
 	const settledTaskIds = /* @__PURE__ */ new Set();
 	const projectCaseCounters = {
@@ -2607,6 +2650,7 @@ async function executePreparedProject(prepared, identity, cacheProjectName, tele
 			createExecutionContext(task) {
 				return createCliTaskExecutionContext(task, prepared.project.models, resolve(prepared.project.root, ".vieval", "cache"), cacheProjectName ?? prepared.name, identity.workspaceId, telemetry, reporter, prepared.name, recordEvent, projectCaseCounters, projectCaseFailures, resolveCliRuntimeConcurrency(options), vitestCompatReporter);
 			},
+			maxConcurrency: resolveScheduledTaskConcurrency(prepared.project, options),
 			onTaskEnd(task, state) {
 				settledTaskIds.add(task.id);
 				reporter.onTaskEnd({
@@ -2626,11 +2670,11 @@ async function executePreparedProject(prepared, identity, cacheProjectName, tele
 			onTaskStart(task) {
 				reporter.onTaskStart({ taskId: task.id });
 				vitestCompatReporter?.onTaskStart({ taskId: task.id });
-			},
-			maxConcurrency: resolveScheduledTaskConcurrency(prepared.project, options)
+			}
 		});
 		await vitestCompatReporter?.onRunEnd({ failed: false });
 		return {
+			caseFailures: projectCaseFailures,
 			caseSummary: {
 				failed: projectCaseCounters.failed,
 				passed: projectCaseCounters.passed,
@@ -2638,7 +2682,6 @@ async function executePreparedProject(prepared, identity, cacheProjectName, tele
 				timeout: projectCaseCounters.timeout,
 				total: projectCaseCounters.seenCaseIds.size
 			},
-			caseFailures: projectCaseFailures,
 			discoveredEvalFileCount: prepared.discoveredEvalFileCount,
 			durationMs: Date.now() - prepared.startedAt,
 			entryCount: prepared.entryCount,
@@ -2671,238 +2714,195 @@ async function executePreparedProject(prepared, identity, cacheProjectName, tele
 				state: "skipped",
 				taskId: task.id
 			});
-			await vitestCompatReporter?.onTaskEnd({
-				state: "skipped",
-				taskId: task.id
-			});
-		}
-		await vitestCompatReporter?.onRunEnd({ failed: true });
-		return {
-			caseSummary: {
-				failed: projectCaseCounters.failed,
-				passed: projectCaseCounters.passed,
-				skipped: projectCaseCounters.skipped,
-				timeout: projectCaseCounters.timeout,
-				total: projectCaseCounters.seenCaseIds.size
-			},
-			caseFailures: projectCaseFailures,
-			discoveredEvalFileCount: prepared.discoveredEvalFileCount,
-			durationMs: Date.now() - prepared.startedAt,
-			entryCount: prepared.entryCount,
-			errorMessage: errorMessageFrom(error) ?? "Unknown project execution error.",
-			executed: false,
-			matrixSummary: createProjectMatrixSummary(prepared.tasks),
-			name: prepared.name,
-			result: null,
-			taskCount: prepared.tasks.length
-		};
-	}
-}
-/**
-* Runs vieval orchestration from config and returns project-level summaries.
-*
-* Call stack:
-*
-* {@link runVievalCli}
-*   -> {@link loadVievalCliConfig}
-*   -> {@link discoverEvalFiles}
-*   -> {@link collectEvalEntries}
-*   -> {@link createRunnerSchedule}
-*   -> {@link runScheduledTasks} (optional)
-*
-* Use when:
-* - running eval collection and scheduling from a single command
-* - keeping business-agent eval files near their implementation packages
-*/
-async function runVievalCli(options = {}) {
-	const loadedConfig = await loadVievalCliConfig({
-		configFilePath: options.configFilePath,
-		cwd: options.cwd
-	});
-	const telemetry = loadedConfig.reporting?.openTelemetry?.enabled === true ? createOpenTelemetryRuntime() : createNoopTelemetryRuntime();
-	const onOpenTelemetryRunEnd = loadedConfig.reporting?.openTelemetry?.enabled === true ? loadedConfig.reporting.openTelemetry.onRunEnd : void 0;
-	const restoreEnvironment = applyRunEnvironment(loadedConfig.env);
-	let runError;
-	let runEndError;
-	let output;
-	let reporter;
-	try {
-		const selectedProjects = filterProjectsByName(loadedConfig.projects, options.project ?? []);
-		const preparedProjects = await Promise.all(selectedProjects.map(async (project) => prepareProject(project)));
-		const identity = createRunIdentity(options, preparedProjects);
-		const eventRecorder = createEventRecorder(identity);
-		const runReporter = createReporterWithEventCapture(createRunReporter(options.reporter), eventRecorder.record);
-		reporter = runReporter;
-		output = await telemetry.withSpan("vieval.run", {
-			"vieval.attempt.id": identity.attemptId,
-			"vieval.experiment.id": identity.experimentId,
-			"vieval.run.id": identity.runId,
-			"vieval.workspace.id": identity.workspaceId
-		}, async () => {
-			const workspaceScheduler = createSchedulerRuntime({ concurrency: { workspace: resolveWorkspaceConcurrency(loadedConfig, options) } });
-			const executableProjects = preparedProjects.filter((project) => project.kind === "prepared").map((project) => project.prepared);
-			const totalTasks = preparedProjects.reduce((sum, project) => {
-				if (project.kind === "prepared") return sum + project.prepared.tasks.length;
-				return sum + project.summary.taskCount;
-			}, 0);
-			const skippedSummaryTasks = preparedProjects.reduce((sum, project) => {
-				if (project.kind === "summary") return sum + project.summary.taskCount;
-				return sum;
-			}, 0);
-			const reporterCounters = {
-				failedTasks: 0,
-				passedTasks: 0,
-				skippedTasks: 0
-			};
-			runReporter.onRunStart({ totalTasks });
-			for (const project of executableProjects) for (const task of project.tasks) runReporter.onTaskQueued(createTaskQueuePayload(task, project.name));
-			const projectSummaries = (await Promise.all(preparedProjects.map(async (preparedProject, index) => {
-				if (preparedProject.kind === "summary") return {
-					index,
-					summary: preparedProject.summary
-				};
-				return {
-					index,
-					summary: await telemetry.withSpan("vieval.project", {
-						"vieval.project.name": preparedProject.prepared.name,
-						"vieval.run.id": identity.runId
-					}, async () => await workspaceScheduler.runCase({
-						experimentId: identity.experimentId,
-						projectName: preparedProject.prepared.name,
-						scope: "workspace",
-						workspaceId: identity.workspaceId
-					}, async () => executePreparedProject(preparedProject.prepared, identity, options.cacheProjectName, telemetry, runReporter, reporterCounters, eventRecorder.record, options)))
-				};
-			}))).sort((left, right) => left.index - right.index).map((item) => item.summary);
-			runReporter.onRunEnd({
-				failedTasks: reporterCounters.failedTasks,
-				passedTasks: reporterCounters.passedTasks,
-				skippedTasks: reporterCounters.skippedTasks + skippedSummaryTasks,
-				totalTasks
+			await vitestCompatReporter?.onTaskEnd({
+				state: "skipped",
+				taskId: task.id
 			});
-			const output = {
-				attemptId: identity.attemptId,
-				configFilePath: loadedConfig.configFilePath,
-				experimentId: identity.experimentId,
-				projects: projectSummaries,
-				reportDirectory: null,
-				runId: identity.runId,
-				workspaceId: identity.workspaceId
-			};
-			if (options.reportOut != null) output.reportDirectory = await writeRunReportArtifacts(output, eventRecorder.events, identity, options.reportOut);
-			return output;
-		});
-	} catch (error) {
-		runError = error;
-	} finally {
-		if (onOpenTelemetryRunEnd != null) try {
-			await onOpenTelemetryRunEnd();
-		} catch (error) {
-			if (runError == null) runEndError = error;
 		}
-		reporter?.dispose();
-		restoreEnvironment();
+		await vitestCompatReporter?.onRunEnd({ failed: true });
+		return {
+			caseFailures: projectCaseFailures,
+			caseSummary: {
+				failed: projectCaseCounters.failed,
+				passed: projectCaseCounters.passed,
+				skipped: projectCaseCounters.skipped,
+				timeout: projectCaseCounters.timeout,
+				total: projectCaseCounters.seenCaseIds.size
+			},
+			discoveredEvalFileCount: prepared.discoveredEvalFileCount,
+			durationMs: Date.now() - prepared.startedAt,
+			entryCount: prepared.entryCount,
+			errorMessage: errorMessageFrom(error) ?? "Unknown project execution error.",
+			executed: false,
+			matrixSummary: createProjectMatrixSummary(prepared.tasks),
+			name: prepared.name,
+			result: null,
+			taskCount: prepared.tasks.length
+		};
 	}
-	if (runError != null) throw runError;
-	if (runEndError != null) throw runEndError;
-	if (output == null) throw new Error("Vieval run finished without output.");
-	return output;
+}
+function filterProjectsByName(projects, names) {
+	if (names.length === 0) return [...projects];
+	const nameSet = new Set(names);
+	return projects.filter((project) => nameSet.has(project.name));
+}
+function formatDuration$1(durationMs, colors) {
+	if (durationMs == null) return "";
+	const rounded = Math.round(durationMs);
+	return (rounded > 1e3 ? colors.yellow : colors.green)(` ${rounded}${colors.dim("ms")}`);
+}
+function formatHybridAverage(hybridAverage) {
+	if (hybridAverage == null) return "n/a";
+	return hybridAverage.toFixed(3).replace(/\.?0+$/, "");
+}
+function getFailedTaskId(error) {
+	if (error instanceof RunnerExecutionError) return error.taskId;
+	return null;
+}
+function isSummaryReporter(reporter) {
+	return "getWindowRows" in reporter;
 }
 /**
-* Formats CLI run output as human-readable lines.
+* Normalizes terminal row count into the live reporter window height.
+*
+* Before:
+* - undefined
+* - 4
+* - 40
+*
+* After:
+* - 23
+* - 6
+* - 39
 */
-function formatVievalCliRunOutput(output) {
-	const colorEnabled = shouldUseColor();
-	const colors = createColorPalette(colorEnabled);
-	const lines = [];
-	lines.push(` ${colors.dim("RUN")}  ${colors.yellow("vieval")}`);
-	lines.push(` ${colors.dim("Config")}  ${output.configFilePath ?? "(not found, using defaults)"}`);
-	lines.push("");
-	let passedProjects = 0;
-	let skippedProjects = 0;
-	let failedProjects = 0;
-	let totalTasks = 0;
-	let executedTasks = 0;
-	function formatMatrixSummary(summary) {
-		if (summary == null) return null;
-		const runAxesLabel = summary.runAxes.length === 0 ? "-" : summary.runAxes.join("|");
-		const evalAxesLabel = summary.evalAxes.length === 0 ? "-" : summary.evalAxes.join("|");
-		return `matrix run ${summary.runRows} [${runAxesLabel}] / eval ${summary.evalRows} [${evalAxesLabel}]`;
-	}
-	function formatScheduleBreakdown(project) {
-		const summary = project.matrixSummary;
-		if (summary == null) return null;
-		if (project.taskCount <= 0 || project.entryCount <= 0 || summary.runRows <= 0 || summary.evalRows <= 0) return null;
-		const denominator = project.entryCount * summary.runRows * summary.evalRows;
-		if (denominator <= 0 || project.taskCount % denominator !== 0) return null;
-		const providerCount = project.taskCount / denominator;
-		return [
-			colors.dim("schedule "),
-			colors.yellow(String(project.entryCount)),
-			colors.dim(" entries × "),
-			colors.yellow(String(providerCount)),
-			colors.dim(" inferenceExecutors × "),
-			colors.yellow(String(summary.runRows)),
-			colors.dim(" run rows × "),
-			colors.yellow(String(summary.evalRows)),
-			colors.dim(" eval rows = "),
-			colors.green(String(project.taskCount)),
-			colors.dim(" tasks")
-		].join("");
-	}
-	for (const project of output.projects) {
-		totalTasks += project.taskCount;
-		executedTasks += project.result?.overall.runCount ?? 0;
-		const badge = createProjectBadge(project.name, colors, colorEnabled);
-		const isFailed = project.errorMessage != null;
-		const hasFailedCases = (project.caseSummary?.failed ?? 0) > 0 || (project.caseSummary?.timeout ?? 0) > 0 || (project.caseFailures?.length ?? 0) > 0;
-		if (isFailed) {
-			failedProjects += 1;
-			lines.push(` ${colors.red("❯")} ${badge}${formatDuration$1(project.durationMs, colors)}`);
-			lines.push(`   ${project.errorMessage}`);
-			continue;
-		}
-		if (!project.executed) {
-			skippedProjects += 1;
-			const countLabel = colors.dim(`(${project.taskCount} tasks)`);
-			const detailsLabel = colors.dim(` ${project.discoveredEvalFileCount} files, ${project.entryCount} entries, 0 runs, hybrid n/a`);
-			const matrixSummary = formatMatrixSummary(project.matrixSummary);
-			lines.push(` ${colors.dim("○")} ${badge}${countLabel}${detailsLabel}${formatDuration$1(project.durationMs, colors)}`);
-			if (matrixSummary != null) lines.push(`   ${colors.dim(matrixSummary)}`);
-			const scheduleBreakdown = formatScheduleBreakdown(project);
-			if (scheduleBreakdown != null) lines.push(`   ${scheduleBreakdown}`);
-			continue;
-		}
-		if (hasFailedCases) failedProjects += 1;
-		else passedProjects += 1;
-		const hybridAverageLabel = formatHybridAverage(project.result?.overall.hybridAverage);
-		const runCount = project.result?.overall.runCount ?? 0;
-		const countLabel = colors.dim(`(${project.taskCount} tasks)`);
-		const caseSummaryLabel = project.caseSummary == null ? "" : `, cases ${project.caseSummary.passed} passed | ${project.caseSummary.failed} failed | ${project.caseSummary.timeout} timeout`;
-		const detailsLabel = colors.dim(` ${project.discoveredEvalFileCount} files, ${project.entryCount} entries, ${runCount} runs${caseSummaryLabel}, hybrid ${hybridAverageLabel}`);
-		const matrixSummary = formatMatrixSummary(project.matrixSummary);
-		lines.push(` ${hasFailedCases ? colors.red("❯") : colors.green("✓")} ${badge}${countLabel}${detailsLabel}${formatDuration$1(project.durationMs, colors)}`);
-		if (matrixSummary != null) lines.push(`   ${colors.dim(matrixSummary)}`);
-		const scheduleBreakdown = formatScheduleBreakdown(project);
-		if (scheduleBreakdown != null) lines.push(`   ${scheduleBreakdown}`);
-		if ((project.caseFailures?.length ?? 0) > 0) {
-			lines.push(`   ${colors.red("Failed cases:")}`);
-			for (const failure of project.caseFailures.slice(0, 5)) {
-				lines.push(`   ${colors.red(`- ${failure.caseName} (${failure.taskId})`)}`);
-				for (const line of failure.errorMessage.split("\n")) lines.push(`     ${colors.red(line)}`);
+function normalizeLiveReporterMaxRows(rows) {
+	return Math.max(6, (rows == null || !Number.isFinite(rows) || rows <= 0 ? 24 : Math.floor(rows)) - 1);
+}
+async function prepareProject(project) {
+	const startedAt = Date.now();
+	try {
+		const runtimeContext = await createRunnerRuntimeContext({
+			cwd: project.root,
+			fallbackProjectRootDirectory: project.root
+		});
+		const evalFilePaths = await discoverEvalFiles({
+			exclude: project.exclude,
+			include: project.include,
+			root: project.root
+		});
+		const entries = collectEvalEntries(await loadEvalModulesWithVitestRuntime(evalFilePaths, project.root), runtimeContext);
+		const tasks = createRunnerSchedule({
+			entries,
+			evalMatrix: project.evalMatrix,
+			inferenceExecutors: project.inferenceExecutors,
+			runMatrix: project.runMatrix
+		});
+		const canAutoExecuteEntryTasks = entries.some((entry) => entry.task != null) && project.models.length > 0;
+		if (project.executor == null && !canAutoExecuteEntryTasks) return {
+			experimentMatrixRows: createExperimentMatrixRows(tasks),
+			kind: "summary",
+			summary: {
+				caseFailures: [],
+				caseSummary: null,
+				discoveredEvalFileCount: evalFilePaths.length,
+				durationMs: Date.now() - startedAt,
+				entryCount: entries.length,
+				errorMessage: null,
+				executed: false,
+				matrixSummary: createProjectMatrixSummary(tasks),
+				name: project.name,
+				result: null,
+				taskCount: tasks.length
 			}
-			if (project.caseFailures.length > 5) lines.push(`   ${colors.dim(`... ${project.caseFailures.length - 5} more failed cases`)}`);
-		}
+		};
+		return {
+			experimentMatrixRows: createExperimentMatrixRows(tasks),
+			kind: "prepared",
+			prepared: {
+				discoveredEvalFileCount: evalFilePaths.length,
+				entryCount: entries.length,
+				name: project.name,
+				project,
+				startedAt,
+				tasks
+			}
+		};
+	} catch (error) {
+		return {
+			experimentMatrixRows: [],
+			kind: "summary",
+			summary: {
+				caseFailures: [],
+				caseSummary: null,
+				discoveredEvalFileCount: 0,
+				durationMs: Date.now() - startedAt,
+				entryCount: 0,
+				errorMessage: errorMessageFrom(error) ?? "Unknown project execution error.",
+				executed: false,
+				matrixSummary: null,
+				name: project.name,
+				result: null,
+				taskCount: 0
+			}
+		};
 	}
-	lines.push("");
-	if (failedProjects > 0 || skippedProjects > 0) {
-		const summarySegments = [`${colors.green(String(passedProjects))} passed`];
-		if (skippedProjects > 0) summarySegments.push(`${colors.dim(String(skippedProjects))} skipped`);
-		if (failedProjects > 0) summarySegments.push(`${colors.red(String(failedProjects))} failed`);
-		lines.push(` ${colors.dim("Projects")}  ${summarySegments.join(" | ")} (${output.projects.length})`);
-	} else lines.push(` ${colors.dim("Projects")}  ${colors.green(String(passedProjects))} passed (${output.projects.length})`);
-	lines.push(` ${colors.dim("Tasks")}     ${executedTasks} executed / ${totalTasks} scheduled`);
-	return lines.join("\n");
+}
+function resolveCappedConcurrency(defaultConcurrency, cliConcurrency, fallback) {
+	const effectiveDefault = defaultConcurrency ?? fallback;
+	if (cliConcurrency == null) return effectiveDefault;
+	return Math.min(effectiveDefault, cliConcurrency);
+}
+function resolveCliRuntimeConcurrency(options) {
+	if (options.attemptConcurrency == null && options.caseConcurrency == null) return;
+	return {
+		attempt: options.attemptConcurrency,
+		case: options.caseConcurrency
+	};
+}
+function resolveExperimentId(options, preparedProjects) {
+	if (options.experiment != null) return sanitizeIdentitySegment(options.experiment);
+	const matrixRows = /* @__PURE__ */ new Set();
+	for (const project of preparedProjects) project.experimentMatrixRows.forEach((row) => matrixRows.add(row));
+	if (matrixRows.size === 0) return "default-experiment";
+	return sanitizeIdentitySegment(`matrix-${[...matrixRows].sort().join("--")}`);
+}
+function resolveOptionalRuntimeTaskConcurrency(defaultConcurrency, cliConcurrency) {
+	return cliConcurrency ?? defaultConcurrency;
+}
+function resolveProjectConcurrency(project, options) {
+	return resolveCappedConcurrency(project.concurrency?.project, options.projectConcurrency, Number.POSITIVE_INFINITY);
+}
+function resolveRuntimeTaskConcurrency(taskConcurrency, project, options) {
+	const attempt = resolveOptionalRuntimeTaskConcurrency(taskConcurrency?.attempt ?? project.concurrency?.attempt, options.attemptConcurrency);
+	const caseConcurrency = resolveOptionalRuntimeTaskConcurrency(taskConcurrency?.case ?? project.concurrency?.case, options.caseConcurrency);
+	if (attempt == null && caseConcurrency == null) return;
+	return {
+		attempt,
+		case: caseConcurrency
+	};
+}
+function resolveScheduledTaskConcurrency(project, options) {
+	return Math.min(resolveProjectConcurrency(project, options), resolveTaskConcurrency(project, options));
+}
+function resolveTaskConcurrency(project, options) {
+	return resolveCappedConcurrency(project.concurrency?.task, options.taskConcurrency, 1);
+}
+function resolveTaskReporterHooks(task, context, reporter, projectName, recordEvent, projectCaseCounters, projectCaseFailures, vitestCompatReporter) {
+	return context.reporterHooks ?? createTaskReporterHooks(task, reporter, projectName, recordEvent, projectCaseCounters, projectCaseFailures, vitestCompatReporter);
+}
+function resolveWorkspaceConcurrency(loadedConfig, options) {
+	return resolveCappedConcurrency(loadedConfig.concurrency?.workspace, options.workspaceConcurrency, 1);
+}
+function sanitizeIdentitySegment(value) {
+	const normalized = value.trim();
+	if (normalized.length === 0) return "default";
+	return normalized.replace(/[^\w.-]+/g, "-");
+}
+function shouldUseColor() {
+	if (process.env.NO_COLOR != null) return false;
+	const forceColor = process.env.FORCE_COLOR;
+	if (forceColor != null) return forceColor !== "0";
+	return process.stdout.isTTY === true;
 }
 //#endregion
 //#region src/cli/compare.ts
@@ -2918,17 +2918,12 @@ const compareHelpText = `
     --output      Optional output artifact path
     --format      Console output format: table | json (default: table)
 `;
-function normalizeCliArgv$5(argv) {
-	const normalizedArgv = argv[0] === "--" ? argv.slice(1) : [...argv];
-	if (normalizedArgv[0] === "compare") return normalizedArgv.slice(1);
-	return normalizedArgv;
-}
 function parseCompareCliArguments(argv) {
 	const cli = meow(compareHelpText, {
 		argv: normalizeCliArgv$5(argv),
 		flags: {
-			config: { type: "string" },
 			comparison: { type: "string" },
+			config: { type: "string" },
 			format: {
 				default: "table",
 				type: "string"
@@ -3009,6 +3004,11 @@ async function runCompareCliOrExit(argv) {
 		process.exitCode = 1;
 	}
 }
+function normalizeCliArgv$5(argv) {
+	const normalizedArgv = argv[0] === "--" ? argv.slice(1) : [...argv];
+	if (normalizedArgv[0] === "compare") return normalizedArgv.slice(1);
+	return normalizedArgv;
+}
 //#endregion
 //#region package.json
 var name = "vieval";
@@ -3034,14 +3034,6 @@ const evalRunHelpText = `
     --report-out Report output root directory
     --json       Print machine-readable JSON output
 `;
-function normalizeCliArgv$4(argv) {
-	const normalizedArgv = argv[0] === "--" ? argv.slice(1) : [...argv];
-	return normalizedArgv[0] === "run" ? normalizedArgv.slice(1) : normalizedArgv;
-}
-function normalizeProjectNames(projectNames) {
-	if (typeof projectNames === "string") return [projectNames];
-	return projectNames ?? [];
-}
 /**
 * Parses `vieval run` CLI arguments into one normalized execution payload.
 *
@@ -3058,9 +3050,12 @@ function normalizeProjectNames(projectNames) {
 function parseCliArguments(argv) {
 	const cli = meow(evalRunHelpText, {
 		argv: normalizeCliArgv$4(argv),
-		importMeta: import.meta,
 		flags: {
+			attempt: { type: "string" },
+			attemptConcurrency: { type: "number" },
+			caseConcurrency: { type: "number" },
 			config: { type: "string" },
+			experiment: { type: "string" },
 			json: {
 				default: false,
 				type: "boolean"
@@ -3069,16 +3064,13 @@ function parseCliArguments(argv) {
 				isMultiple: true,
 				type: "string"
 			},
-			workspace: { type: "string" },
-			experiment: { type: "string" },
-			attempt: { type: "string" },
-			workspaceConcurrency: { type: "number" },
 			projectConcurrency: { type: "number" },
+			reportOut: { type: "string" },
 			taskConcurrency: { type: "number" },
-			attemptConcurrency: { type: "number" },
-			caseConcurrency: { type: "number" },
-			reportOut: { type: "string" }
-		}
+			workspace: { type: "string" },
+			workspaceConcurrency: { type: "number" }
+		},
+		importMeta: import.meta
 	});
 	return {
 		attempt: cli.flags.attempt,
@@ -3131,198 +3123,57 @@ async function runEvalRunCli(argv) {
 			attemptConcurrency: parsed.attemptConcurrency,
 			caseConcurrency: parsed.caseConcurrency,
 			configFilePath: parsed.configFilePath,
-			experiment: parsed.experiment,
-			project: parsed.project,
-			projectConcurrency: parsed.projectConcurrency,
-			reportOut: parsed.reportOut,
-			taskConcurrency: parsed.taskConcurrency,
-			workspace: parsed.workspace,
-			workspaceConcurrency: parsed.workspaceConcurrency
-		});
-		if (parsed.json) {
-			process.stdout.write(`${JSON.stringify(output, null, 2)}\n`);
-			if (hasRunFailures(output)) process.exitCode = 1;
-			return;
-		}
-		process.stdout.write(`${formatVievalCliRunOutput(output)}\n`);
-		if (hasRunFailures(output)) process.exitCode = 1;
-	} catch (error) {
-		const errorMessage = errorMessageFrom(error) ?? "Unknown CLI failure.";
-		process.stderr.write(`[${name}] ${errorMessage}\n`);
-		process.exitCode = 1;
-	}
-}
-//#endregion
-//#region src/cli/report-analyze.ts
-const reportAnalyzeHelpText = `
-  Analyze generated vieval report artifacts.
-
-  Usage
-    $ vieval report analyze <reportPath> [options]
-
-  Options
-    --format       Output format: table | json | jsonl | csv (default: table)
-    --workspace    Workspace id filter
-    --project      Project name filter (exact)
-    --experiment   Experiment id filter
-    --attempt      Attempt id filter
-    --run          Run id filter
-    --task-state   Keep runs containing at least one task in this state
-    --case-state   Keep runs containing at least one case in this state
-    --contains     Keep runs containing this text in event name or payload
-    --error-contains Keep runs containing this text in project errors or event payload
-    --run-matrix   Keep runs matching run-matrix selector "key=value[,key=value]"
-    --eval-matrix  Keep runs matching eval-matrix selector "key=value[,key=value]"
-`;
-function normalizeCliArgv$3(argv) {
-	const normalizedArgv = argv[0] === "--" ? argv.slice(1) : [...argv];
-	if (normalizedArgv[0] === "report" && normalizedArgv[1] === "analyze") return normalizedArgv.slice(2);
-	if (normalizedArgv[0] === "analyze") return normalizedArgv.slice(1);
-	return normalizedArgv;
-}
-function parseReportAnalyzeCliArguments(argv) {
-	const cli = meow(reportAnalyzeHelpText, {
-		argv: normalizeCliArgv$3(argv),
-		flags: {
-			attempt: { type: "string" },
-			caseState: { type: "string" },
-			contains: { type: "string" },
-			evalMatrix: { type: "string" },
-			errorContains: { type: "string" },
-			experiment: { type: "string" },
-			format: {
-				default: "table",
-				type: "string"
-			},
-			project: { type: "string" },
-			runMatrix: { type: "string" },
-			run: { type: "string" },
-			taskState: { type: "string" },
-			workspace: { type: "string" }
-		},
-		importMeta: import.meta
-	});
-	const reportPath = cli.input[0];
-	if (reportPath == null || reportPath.length === 0) throw new Error("Missing required <reportPath> argument.");
-	const normalizedFormat = cli.flags.format.toLowerCase();
-	const format = normalizedFormat === "json" ? "json" : normalizedFormat === "jsonl" ? "jsonl" : normalizedFormat === "csv" ? "csv" : "table";
-	return {
-		attempt: cli.flags.attempt,
-		caseState: normalizeStateFilter(cli.flags.caseState),
-		contains: cli.flags.contains,
-		evalMatrix: parseMatrixSelector(cli.flags.evalMatrix),
-		errorContains: cli.flags.errorContains,
-		experiment: cli.flags.experiment,
-		format,
-		project: cli.flags.project,
-		reportPath,
-		runMatrix: parseMatrixSelector(cli.flags.runMatrix),
-		run: cli.flags.run,
-		taskState: normalizeStateFilter(cli.flags.taskState),
-		workspace: cli.flags.workspace
-	};
-}
-function normalizeStateFilter(value) {
-	if (value == null) return;
-	const normalized = value.trim().toLowerCase();
-	if (normalized === "passed" || normalized === "failed" || normalized === "skipped") return normalized;
-	throw new Error(`Unsupported state filter "${value}". Expected "passed", "failed", or "skipped".`);
-}
-function parseMatrixSelector(value) {
-	if (value == null) return;
-	const selector = {};
-	const segments = value.split(",").map((segment) => segment.trim()).filter((segment) => segment.length > 0);
-	for (const segment of segments) {
-		const separatorIndex = segment.indexOf("=");
-		if (separatorIndex <= 0 || separatorIndex === segment.length - 1) throw new Error(`Invalid matrix selector segment "${segment}". Expected "key=value".`);
-		const key = segment.slice(0, separatorIndex).trim();
-		const parsedValue = segment.slice(separatorIndex + 1).trim();
-		if (key.length === 0 || parsedValue.length === 0) throw new Error(`Invalid matrix selector segment "${segment}". Expected "key=value".`);
-		selector[key] = parsedValue;
-	}
-	return selector;
-}
-function filterAnalyzeRows(rows, parsed) {
-	return rows.filter((row) => {
-		if (parsed.workspace != null && row.workspaceId !== parsed.workspace) return false;
-		if (parsed.experiment != null && row.experimentId !== parsed.experiment) return false;
-		if (parsed.attempt != null && row.attemptId !== parsed.attempt) return false;
-		if (parsed.run != null && row.runId !== parsed.run) return false;
-		if (parsed.project != null && !row.projectNames.includes(parsed.project)) return false;
-		return true;
-	});
-}
-function includesNeedle(value, needle) {
-	const normalizedNeedle = needle.trim().toLowerCase();
-	if (normalizedNeedle.length === 0) return true;
-	return JSON.stringify(value).toLowerCase().includes(normalizedNeedle);
-}
-function hasTaskState(artifact, targetState) {
-	return artifact.events.some((event) => {
-		if (event.event !== "TaskEnded") return false;
-		return event.data?.state === targetState;
-	});
-}
-function hasCaseState(artifact, targetState) {
-	return artifact.events.some((event) => {
-		if (event.event !== "CaseEnded") return false;
-		return event.data?.state === targetState;
-	});
-}
-function matchesMatrixSelector(matrix, selector) {
-	return Object.entries(selector).every(([key, expectedValue]) => String(matrix[key]) === expectedValue);
-}
-function hasRunMatrixMatch(artifact, selector) {
-	return artifact.summary.projects.some((project) => project.result?.runs.some((run) => matchesMatrixSelector(run.matrix.run, selector)) === true);
-}
-function hasEvalMatrixMatch(artifact, selector) {
-	return artifact.summary.projects.some((project) => project.result?.runs.some((run) => matchesMatrixSelector(run.matrix.eval, selector)) === true);
-}
-function matchesOutcomeFilters(artifact, parsed) {
-	if (parsed.runMatrix != null && !hasRunMatrixMatch(artifact, parsed.runMatrix)) return false;
-	if (parsed.evalMatrix != null && !hasEvalMatrixMatch(artifact, parsed.evalMatrix)) return false;
-	if (parsed.taskState != null && !hasTaskState(artifact, parsed.taskState)) return false;
-	if (parsed.caseState != null && !hasCaseState(artifact, parsed.caseState)) return false;
-	if (parsed.contains != null) {
-		if (!artifact.events.some((event) => includesNeedle({
-			data: event.data,
-			event: event.event
-		}, parsed.contains))) return false;
-	}
-	if (parsed.errorContains != null) {
-		if (!(artifact.summary.projects.map((project) => project.errorMessage).filter((errorMessage) => errorMessage != null).some((errorMessage) => includesNeedle(errorMessage, parsed.errorContains)) || artifact.events.some((event) => includesNeedle(event.data, parsed.errorContains)))) return false;
-	}
-	return true;
-}
-async function readReportAnalyzeOutput(parsed) {
-	const artifacts = await readReportArtifacts(parsed.reportPath);
-	const rows = artifacts.map((artifact) => summarizeReportRunArtifact(artifact));
-	const identityFilteredRows = filterAnalyzeRows(rows, parsed);
-	const rowByDirectory = new Map(identityFilteredRows.map((row) => [row.reportDirectory, row]));
-	const filteredRows = artifacts.filter((artifact) => rowByDirectory.has(artifact.reportDirectory)).filter((artifact) => matchesOutcomeFilters(artifact, parsed)).map((artifact) => rowByDirectory.get(artifact.reportDirectory)).filter((row) => row != null);
-	return {
-		experimentSummaries: buildExperimentSummaries(filteredRows),
-		filteredRunCount: filteredRows.length,
-		runs: filteredRows,
-		totalRunCount: rows.length
-	};
-}
-function roundMetric(value) {
-	return Number(value.toFixed(6));
-}
-function computeAverage(values) {
-	if (values.length === 0) return 0;
-	return values.reduce((sum, value) => sum + value, 0) / values.length;
+			experiment: parsed.experiment,
+			project: parsed.project,
+			projectConcurrency: parsed.projectConcurrency,
+			reportOut: parsed.reportOut,
+			taskConcurrency: parsed.taskConcurrency,
+			workspace: parsed.workspace,
+			workspaceConcurrency: parsed.workspaceConcurrency
+		});
+		if (parsed.json) {
+			process.stdout.write(`${JSON.stringify(output, null, 2)}\n`);
+			if (hasRunFailures(output)) process.exitCode = 1;
+			return;
+		}
+		process.stdout.write(`${formatVievalCliRunOutput(output)}\n`);
+		if (hasRunFailures(output)) process.exitCode = 1;
+	} catch (error) {
+		const errorMessage = errorMessageFrom(error) ?? "Unknown CLI failure.";
+		process.stderr.write(`[${name}] ${errorMessage}\n`);
+		process.exitCode = 1;
+	}
 }
-function computeStandardDeviation(values) {
-	if (values.length === 0) return 0;
-	const average = computeAverage(values);
-	const variance = computeAverage(values.map((value) => (value - average) ** 2));
-	return Math.sqrt(variance);
+function normalizeCliArgv$4(argv) {
+	const normalizedArgv = argv[0] === "--" ? argv.slice(1) : [...argv];
+	return normalizedArgv[0] === "run" ? normalizedArgv.slice(1) : normalizedArgv;
 }
-function createExperimentGroupKey(row) {
-	return `${row.workspaceId ?? "unknown-workspace"}::${row.experimentId ?? "unknown-experiment"}`;
+function normalizeProjectNames(projectNames) {
+	if (typeof projectNames === "string") return [projectNames];
+	return projectNames ?? [];
 }
+//#endregion
+//#region src/cli/report-analyze.ts
+const reportAnalyzeHelpText = `
+  Analyze generated vieval report artifacts.
+
+  Usage
+    $ vieval report analyze <reportPath> [options]
+
+  Options
+    --format       Output format: table | json | jsonl | csv (default: table)
+    --workspace    Workspace id filter
+    --project      Project name filter (exact)
+    --experiment   Experiment id filter
+    --attempt      Attempt id filter
+    --run          Run id filter
+    --task-state   Keep runs containing at least one task in this state
+    --case-state   Keep runs containing at least one case in this state
+    --contains     Keep runs containing this text in event name or payload
+    --error-contains Keep runs containing this text in project errors or event payload
+    --run-matrix   Keep runs matching run-matrix selector "key=value[,key=value]"
+    --eval-matrix  Keep runs matching eval-matrix selector "key=value[,key=value]"
+`;
 /**
 * Builds experiment-level rollups from filtered run rows.
 *
@@ -3383,13 +3234,13 @@ function buildExperimentSummaries(rows) {
 		const stdevAttemptSuccessRate = computeStandardDeviation(attemptSuccessRates);
 		return {
 			attemptCount: attemptToRuns.size,
-			attemptSummaries,
 			attemptSuccessRateStats: {
 				avg: roundMetric(avgAttemptSuccessRate),
 				max: roundMetric(maxAttemptSuccessRate),
 				min: roundMetric(minAttemptSuccessRate),
 				stdev: roundMetric(stdevAttemptSuccessRate)
 			},
+			attemptSummaries,
 			experimentId,
 			failedProjects,
 			runCount: groupRows.length,
@@ -3404,16 +3255,94 @@ function buildExperimentSummaries(rows) {
 		return left.experimentId.localeCompare(right.experimentId);
 	});
 }
-function formatTableOutput$1(output) {
-	const header = "Run ID | Workspace | Experiment | Attempt | Projects(executed/total) | FailedProjects | Tasks | Events";
-	const lines = output.runs.map((row) => {
-		return `${row.runId ?? "n/a"} | ${row.workspaceId ?? "n/a"} | ${row.experimentId ?? "n/a"} | ${row.attemptId ?? "n/a"} | ${`${row.executedProjects}/${row.totalProjects}`} | ${row.failedProjects} | ${row.totalTasks} | ${row.eventsCount}`;
+function parseReportAnalyzeCliArguments(argv) {
+	const cli = meow(reportAnalyzeHelpText, {
+		argv: normalizeCliArgv$3(argv),
+		flags: {
+			attempt: { type: "string" },
+			caseState: { type: "string" },
+			contains: { type: "string" },
+			errorContains: { type: "string" },
+			evalMatrix: { type: "string" },
+			experiment: { type: "string" },
+			format: {
+				default: "table",
+				type: "string"
+			},
+			project: { type: "string" },
+			run: { type: "string" },
+			runMatrix: { type: "string" },
+			taskState: { type: "string" },
+			workspace: { type: "string" }
+		},
+		importMeta: import.meta
+	});
+	const reportPath = cli.input[0];
+	if (reportPath == null || reportPath.length === 0) throw new Error("Missing required <reportPath> argument.");
+	const normalizedFormat = cli.flags.format.toLowerCase();
+	const format = normalizedFormat === "json" ? "json" : normalizedFormat === "jsonl" ? "jsonl" : normalizedFormat === "csv" ? "csv" : "table";
+	return {
+		attempt: cli.flags.attempt,
+		caseState: normalizeStateFilter(cli.flags.caseState),
+		contains: cli.flags.contains,
+		errorContains: cli.flags.errorContains,
+		evalMatrix: parseMatrixSelector(cli.flags.evalMatrix),
+		experiment: cli.flags.experiment,
+		format,
+		project: cli.flags.project,
+		reportPath,
+		run: cli.flags.run,
+		runMatrix: parseMatrixSelector(cli.flags.runMatrix),
+		taskState: normalizeStateFilter(cli.flags.taskState),
+		workspace: cli.flags.workspace
+	};
+}
+async function runReportAnalyzeCli(argv) {
+	try {
+		const parsed = parseReportAnalyzeCliArguments(argv);
+		const output = await readReportAnalyzeOutput(parsed);
+		if (parsed.format === "json") {
+			process.stdout.write(`${JSON.stringify(output, null, 2)}\n`);
+			return;
+		}
+		if (parsed.format === "jsonl") {
+			const jsonl = output.runs.map((run) => JSON.stringify(run)).join("\n");
+			process.stdout.write(`${jsonl}${jsonl.length > 0 ? "\n" : ""}`);
+			return;
+		}
+		if (parsed.format === "csv") {
+			process.stdout.write(`${formatCsvOutput(output)}\n`);
+			return;
+		}
+		process.stdout.write(`${formatTableOutput$1(output)}\n`);
+	} catch (error) {
+		const errorMessage = errorMessageFrom(error) ?? "Unknown report analyze failure.";
+		process.stderr.write(`[vieval report analyze] ${errorMessage}\n`);
+		process.exitCode = 1;
+	}
+}
+function computeAverage(values) {
+	if (values.length === 0) return 0;
+	return values.reduce((sum, value) => sum + value, 0) / values.length;
+}
+function computeStandardDeviation(values) {
+	if (values.length === 0) return 0;
+	const average = computeAverage(values);
+	const variance = computeAverage(values.map((value) => (value - average) ** 2));
+	return Math.sqrt(variance);
+}
+function createExperimentGroupKey(row) {
+	return `${row.workspaceId ?? "unknown-workspace"}::${row.experimentId ?? "unknown-experiment"}`;
+}
+function filterAnalyzeRows(rows, parsed) {
+	return rows.filter((row) => {
+		if (parsed.workspace != null && row.workspaceId !== parsed.workspace) return false;
+		if (parsed.experiment != null && row.experimentId !== parsed.experiment) return false;
+		if (parsed.attempt != null && row.attemptId !== parsed.attempt) return false;
+		if (parsed.run != null && row.runId !== parsed.run) return false;
+		if (parsed.project != null && !row.projectNames.includes(parsed.project)) return false;
+		return true;
 	});
-	return [
-		`ANALYZE vieval report: ${output.filteredRunCount}/${output.totalRunCount} runs (${output.experimentSummaries.length} experiment groups)`,
-		header,
-		...lines
-	].join("\n");
 }
 function formatCsvOutput(output) {
 	return [[
@@ -3446,29 +3375,100 @@ function formatCsvOutput(output) {
 		].join(",");
 	})].join("\n");
 }
-async function runReportAnalyzeCli(argv) {
-	try {
-		const parsed = parseReportAnalyzeCliArguments(argv);
-		const output = await readReportAnalyzeOutput(parsed);
-		if (parsed.format === "json") {
-			process.stdout.write(`${JSON.stringify(output, null, 2)}\n`);
-			return;
-		}
-		if (parsed.format === "jsonl") {
-			const jsonl = output.runs.map((run) => JSON.stringify(run)).join("\n");
-			process.stdout.write(`${jsonl}${jsonl.length > 0 ? "\n" : ""}`);
-			return;
-		}
-		if (parsed.format === "csv") {
-			process.stdout.write(`${formatCsvOutput(output)}\n`);
-			return;
-		}
-		process.stdout.write(`${formatTableOutput$1(output)}\n`);
-	} catch (error) {
-		const errorMessage = errorMessageFrom(error) ?? "Unknown report analyze failure.";
-		process.stderr.write(`[vieval report analyze] ${errorMessage}\n`);
-		process.exitCode = 1;
+function formatTableOutput$1(output) {
+	const header = "Run ID | Workspace | Experiment | Attempt | Projects(executed/total) | FailedProjects | Tasks | Events";
+	const lines = output.runs.map((row) => {
+		return `${row.runId ?? "n/a"} | ${row.workspaceId ?? "n/a"} | ${row.experimentId ?? "n/a"} | ${row.attemptId ?? "n/a"} | ${`${row.executedProjects}/${row.totalProjects}`} | ${row.failedProjects} | ${row.totalTasks} | ${row.eventsCount}`;
+	});
+	return [
+		`ANALYZE vieval report: ${output.filteredRunCount}/${output.totalRunCount} runs (${output.experimentSummaries.length} experiment groups)`,
+		header,
+		...lines
+	].join("\n");
+}
+function hasCaseState(artifact, targetState) {
+	return artifact.events.some((event) => {
+		if (event.event !== "CaseEnded") return false;
+		return event.data?.state === targetState;
+	});
+}
+function hasEvalMatrixMatch(artifact, selector) {
+	return artifact.summary.projects.some((project) => project.result?.runs.some((run) => matchesMatrixSelector(run.matrix.eval, selector)) === true);
+}
+function hasRunMatrixMatch(artifact, selector) {
+	return artifact.summary.projects.some((project) => project.result?.runs.some((run) => matchesMatrixSelector(run.matrix.run, selector)) === true);
+}
+function hasTaskState(artifact, targetState) {
+	return artifact.events.some((event) => {
+		if (event.event !== "TaskEnded") return false;
+		return event.data?.state === targetState;
+	});
+}
+function includesNeedle(value, needle) {
+	const normalizedNeedle = needle.trim().toLowerCase();
+	if (normalizedNeedle.length === 0) return true;
+	return JSON.stringify(value).toLowerCase().includes(normalizedNeedle);
+}
+function matchesMatrixSelector(matrix, selector) {
+	return Object.entries(selector).every(([key, expectedValue]) => String(matrix[key]) === expectedValue);
+}
+function matchesOutcomeFilters(artifact, parsed) {
+	if (parsed.runMatrix != null && !hasRunMatrixMatch(artifact, parsed.runMatrix)) return false;
+	if (parsed.evalMatrix != null && !hasEvalMatrixMatch(artifact, parsed.evalMatrix)) return false;
+	if (parsed.taskState != null && !hasTaskState(artifact, parsed.taskState)) return false;
+	if (parsed.caseState != null && !hasCaseState(artifact, parsed.caseState)) return false;
+	if (parsed.contains != null) {
+		if (!artifact.events.some((event) => includesNeedle({
+			data: event.data,
+			event: event.event
+		}, parsed.contains))) return false;
+	}
+	if (parsed.errorContains != null) {
+		if (!(artifact.summary.projects.map((project) => project.errorMessage).filter((errorMessage) => errorMessage != null).some((errorMessage) => includesNeedle(errorMessage, parsed.errorContains)) || artifact.events.some((event) => includesNeedle(event.data, parsed.errorContains)))) return false;
+	}
+	return true;
+}
+function normalizeCliArgv$3(argv) {
+	const normalizedArgv = argv[0] === "--" ? argv.slice(1) : [...argv];
+	if (normalizedArgv[0] === "report" && normalizedArgv[1] === "analyze") return normalizedArgv.slice(2);
+	if (normalizedArgv[0] === "analyze") return normalizedArgv.slice(1);
+	return normalizedArgv;
+}
+function normalizeStateFilter(value) {
+	if (value == null) return;
+	const normalized = value.trim().toLowerCase();
+	if (normalized === "passed" || normalized === "failed" || normalized === "skipped") return normalized;
+	throw new Error(`Unsupported state filter "${value}". Expected "passed", "failed", or "skipped".`);
+}
+function parseMatrixSelector(value) {
+	if (value == null) return;
+	const selector = {};
+	const segments = value.split(",").map((segment) => segment.trim()).filter((segment) => segment.length > 0);
+	for (const segment of segments) {
+		const separatorIndex = segment.indexOf("=");
+		if (separatorIndex <= 0 || separatorIndex === segment.length - 1) throw new Error(`Invalid matrix selector segment "${segment}". Expected "key=value".`);
+		const key = segment.slice(0, separatorIndex).trim();
+		const parsedValue = segment.slice(separatorIndex + 1).trim();
+		if (key.length === 0 || parsedValue.length === 0) throw new Error(`Invalid matrix selector segment "${segment}". Expected "key=value".`);
+		selector[key] = parsedValue;
 	}
+	return selector;
+}
+async function readReportAnalyzeOutput(parsed) {
+	const artifacts = await readReportArtifacts(parsed.reportPath);
+	const rows = artifacts.map((artifact) => summarizeReportRunArtifact(artifact));
+	const identityFilteredRows = filterAnalyzeRows(rows, parsed);
+	const rowByDirectory = new Map(identityFilteredRows.map((row) => [row.reportDirectory, row]));
+	const filteredRows = artifacts.filter((artifact) => rowByDirectory.has(artifact.reportDirectory)).filter((artifact) => matchesOutcomeFilters(artifact, parsed)).map((artifact) => rowByDirectory.get(artifact.reportDirectory)).filter((row) => row != null);
+	return {
+		experimentSummaries: buildExperimentSummaries(filteredRows),
+		filteredRunCount: filteredRows.length,
+		runs: filteredRows,
+		totalRunCount: rows.length
+	};
+}
+function roundMetric(value) {
+	return Number(value.toFixed(6));
 }
 //#endregion
 //#region src/cli/report-case-compare.ts
@@ -3545,6 +3545,50 @@ function buildCaseComparison(args) {
 	};
 }
 /**
+* Formats a case comparison as a compact human-readable table.
+*
+* Use when:
+* - `vieval report compare` should expose the same information as JSON output
+* - users need a terminal-first overview of group and per-case deltas
+*
+* Expects:
+* - comparison output was produced by {@link buildCaseComparison}
+*
+* Returns:
+* - multi-line text containing aggregate, group, top-change, case, and unmatched summaries
+*/
+function formatCaseComparisonTable(output) {
+	const lines = [
+		"COMPARE  vieval report cases",
+		`Matched   ${output.cases.length}`,
+		`Added     ${output.added.length}`,
+		`Removed   ${output.removed.length}`,
+		`Scores    left=${output.overall.leftAverage.toFixed(3)} right=${output.overall.rightAverage.toFixed(3)} delta=${output.overall.delta.toFixed(3)}`
+	];
+	if (output.groups != null && Object.keys(output.groups).length > 0) {
+		lines.push("Groups");
+		for (const [groupKey, group] of Object.entries(output.groups)) lines.push(`${groupKey}  count=${group.count} left=${group.leftAverage.toFixed(3)} right=${group.rightAverage.toFixed(3)} delta=${group.delta.toFixed(3)}`);
+	}
+	if (output.topImprovements.length > 0) {
+		lines.push("Top improvements");
+		for (const row of output.topImprovements) lines.push(`${row.caseKey}  delta=${row.delta.score.toFixed(3)} left=${row.delta.left.toFixed(3)} right=${row.delta.right.toFixed(3)}`);
+	}
+	if (output.topRegressions.length > 0) {
+		lines.push("Top regressions");
+		for (const row of output.topRegressions) lines.push(`${row.caseKey}  delta=${row.delta.score.toFixed(3)} left=${row.delta.left.toFixed(3)} right=${row.delta.right.toFixed(3)}`);
+	}
+	if (output.cases.length > 0) {
+		lines.push("Cases");
+		for (const row of output.cases) {
+			const changedMetricNames = Object.keys(row.metricsChanged);
+			lines.push(`${row.caseKey}  delta=${row.delta.score.toFixed(3)} changedMetrics=${changedMetricNames.length === 0 ? "none" : changedMetricNames.join(",")}`);
+		}
+	}
+	if (output.added.length > 0) lines.push(`Added cases ${output.added.map((record) => record.caseId).join(",")}`);
+	if (output.removed.length > 0) lines.push(`Removed cases ${output.removed.map((record) => record.caseId).join(",")}`);
+	return lines.join("\n");
+}
+/**
 * Runs the `vieval report compare` command.
 *
 * Call stack:
@@ -3586,6 +3630,55 @@ async function runReportCompareCli(argv) {
 		process.exitCode = 1;
 	}
 }
+function averageScore(records, scoreKind) {
+	const values = records.map((record) => record.scores[scoreKind]).filter((value) => typeof value === "number");
+	if (values.length === 0) return 0;
+	return values.reduce((sum, value) => sum + value, 0) / values.length;
+}
+function buildComparisonGroups(cases, groupBy) {
+	const groupedRows = {};
+	for (const row of cases) {
+		const resolved = getCaseSelectorValue(row.right, groupBy);
+		if (!resolved.exists) continue;
+		const groupKey = `${groupBy}=${String(resolved.value)}`;
+		groupedRows[groupKey] ??= [];
+		groupedRows[groupKey].push(row);
+	}
+	return Object.fromEntries(Object.entries(groupedRows).sort(([left], [right]) => left.localeCompare(right)).map(([groupKey, rows]) => {
+		const leftAverage = rows.reduce((sum, row) => sum + row.delta.left, 0) / rows.length;
+		const rightAverage = rows.reduce((sum, row) => sum + row.delta.right, 0) / rows.length;
+		return [groupKey, {
+			count: rows.length,
+			delta: rightAverage - leftAverage,
+			leftAverage,
+			rightAverage
+		}];
+	}));
+}
+function compareCaseRecords(left, right) {
+	return left.caseId.localeCompare(right.caseId);
+}
+function diffMetrics(left, right) {
+	const changed = {};
+	const metricKeys = [.../* @__PURE__ */ new Set([...Object.keys(left), ...Object.keys(right)])].sort((leftKey, rightKey) => leftKey.localeCompare(rightKey));
+	for (const metricKey of metricKeys) if (stableStringify(left[metricKey]) !== stableStringify(right[metricKey])) changed[metricKey] = {
+		left: left[metricKey],
+		right: right[metricKey]
+	};
+	return changed;
+}
+function getScore(record, scoreKind) {
+	return record.scores[scoreKind] ?? 0;
+}
+function indexRecordsByCaseKey(records, caseKey, side) {
+	const indexed = /* @__PURE__ */ new Map();
+	for (const record of records) {
+		const resolved = resolveCaseKey(record, caseKey);
+		if (indexed.has(resolved)) throw new Error(`Duplicate case key "${resolved}" in ${side} report.`);
+		indexed.set(resolved, record);
+	}
+	return indexed;
+}
 function normalizeCliArgv$2(argv) {
 	const normalizedArgv = argv[0] === "--" ? argv.slice(1) : [...argv];
 	if (normalizedArgv[0] === "report" && normalizedArgv[1] === "compare") return normalizedArgv.slice(2);
@@ -3621,15 +3714,6 @@ function parseReportCompareCliArguments(argv) {
 		scoreKind: cli.flags.scoreKind
 	};
 }
-function indexRecordsByCaseKey(records, caseKey, side) {
-	const indexed = /* @__PURE__ */ new Map();
-	for (const record of records) {
-		const resolved = resolveCaseKey(record, caseKey);
-		if (indexed.has(resolved)) throw new Error(`Duplicate case key "${resolved}" in ${side} report.`);
-		indexed.set(resolved, record);
-	}
-	return indexed;
-}
 function resolveCaseKey(record, caseKey) {
 	if (caseKey != null) {
 		const resolved = getCaseSelectorValue(record, caseKey);
@@ -3641,90 +3725,6 @@ function resolveCaseKey(record, caseKey) {
 	const vievalCaseId = getCaseSelectorValue(record, "vieval.case.id");
 	return vievalCaseId.exists ? String(vievalCaseId.value) : record.caseId;
 }
-function getScore(record, scoreKind) {
-	return record.scores[scoreKind] ?? 0;
-}
-function averageScore(records, scoreKind) {
-	const values = records.map((record) => record.scores[scoreKind]).filter((value) => typeof value === "number");
-	if (values.length === 0) return 0;
-	return values.reduce((sum, value) => sum + value, 0) / values.length;
-}
-function diffMetrics(left, right) {
-	const changed = {};
-	const metricKeys = [.../* @__PURE__ */ new Set([...Object.keys(left), ...Object.keys(right)])].sort((leftKey, rightKey) => leftKey.localeCompare(rightKey));
-	for (const metricKey of metricKeys) if (stableStringify(left[metricKey]) !== stableStringify(right[metricKey])) changed[metricKey] = {
-		left: left[metricKey],
-		right: right[metricKey]
-	};
-	return changed;
-}
-function buildComparisonGroups(cases, groupBy) {
-	const groupedRows = {};
-	for (const row of cases) {
-		const resolved = getCaseSelectorValue(row.right, groupBy);
-		if (!resolved.exists) continue;
-		const groupKey = `${groupBy}=${String(resolved.value)}`;
-		groupedRows[groupKey] ??= [];
-		groupedRows[groupKey].push(row);
-	}
-	return Object.fromEntries(Object.entries(groupedRows).sort(([left], [right]) => left.localeCompare(right)).map(([groupKey, rows]) => {
-		const leftAverage = rows.reduce((sum, row) => sum + row.delta.left, 0) / rows.length;
-		const rightAverage = rows.reduce((sum, row) => sum + row.delta.right, 0) / rows.length;
-		return [groupKey, {
-			count: rows.length,
-			delta: rightAverage - leftAverage,
-			leftAverage,
-			rightAverage
-		}];
-	}));
-}
-function compareCaseRecords(left, right) {
-	return left.caseId.localeCompare(right.caseId);
-}
-/**
-* Formats a case comparison as a compact human-readable table.
-*
-* Use when:
-* - `vieval report compare` should expose the same information as JSON output
-* - users need a terminal-first overview of group and per-case deltas
-*
-* Expects:
-* - comparison output was produced by {@link buildCaseComparison}
-*
-* Returns:
-* - multi-line text containing aggregate, group, top-change, case, and unmatched summaries
-*/
-function formatCaseComparisonTable(output) {
-	const lines = [
-		"COMPARE  vieval report cases",
-		`Matched   ${output.cases.length}`,
-		`Added     ${output.added.length}`,
-		`Removed   ${output.removed.length}`,
-		`Scores    left=${output.overall.leftAverage.toFixed(3)} right=${output.overall.rightAverage.toFixed(3)} delta=${output.overall.delta.toFixed(3)}`
-	];
-	if (output.groups != null && Object.keys(output.groups).length > 0) {
-		lines.push("Groups");
-		for (const [groupKey, group] of Object.entries(output.groups)) lines.push(`${groupKey}  count=${group.count} left=${group.leftAverage.toFixed(3)} right=${group.rightAverage.toFixed(3)} delta=${group.delta.toFixed(3)}`);
-	}
-	if (output.topImprovements.length > 0) {
-		lines.push("Top improvements");
-		for (const row of output.topImprovements) lines.push(`${row.caseKey}  delta=${row.delta.score.toFixed(3)} left=${row.delta.left.toFixed(3)} right=${row.delta.right.toFixed(3)}`);
-	}
-	if (output.topRegressions.length > 0) {
-		lines.push("Top regressions");
-		for (const row of output.topRegressions) lines.push(`${row.caseKey}  delta=${row.delta.score.toFixed(3)} left=${row.delta.left.toFixed(3)} right=${row.delta.right.toFixed(3)}`);
-	}
-	if (output.cases.length > 0) {
-		lines.push("Cases");
-		for (const row of output.cases) {
-			const changedMetricNames = Object.keys(row.metricsChanged);
-			lines.push(`${row.caseKey}  delta=${row.delta.score.toFixed(3)} changedMetrics=${changedMetricNames.length === 0 ? "none" : changedMetricNames.join(",")}`);
-		}
-	}
-	if (output.added.length > 0) lines.push(`Added cases ${output.added.map((record) => record.caseId).join(",")}`);
-	if (output.removed.length > 0) lines.push(`Removed cases ${output.removed.map((record) => record.caseId).join(",")}`);
-	return lines.join("\n");
-}
 //#endregion
 //#region src/cli/report-index.ts
 const reportIndexHelpText = `
@@ -3737,12 +3737,6 @@ const reportIndexHelpText = `
     --output      Output file path (default: <reportPath>/index/runs.jsonl)
     --format      Console output format: table | json | jsonl (default: table)
 `;
-function normalizeCliArgv$1(argv) {
-	const normalizedArgv = argv[0] === "--" ? argv.slice(1) : [...argv];
-	if (normalizedArgv[0] === "report" && normalizedArgv[1] === "index") return normalizedArgv.slice(2);
-	if (normalizedArgv[0] === "index") return normalizedArgv.slice(1);
-	return normalizedArgv;
-}
 function parseReportIndexCliArguments(argv) {
 	const cli = meow(reportIndexHelpText, {
 		argv: normalizeCliArgv$1(argv),
@@ -3764,25 +3758,6 @@ function parseReportIndexCliArguments(argv) {
 		reportPath
 	};
 }
-async function writeIndexFile(parsed) {
-	const rows = (await readReportArtifacts(parsed.reportPath)).map((artifact) => summarizeReportRunArtifact(artifact));
-	const indexFilePath = resolve(parsed.output ?? resolve(parsed.reportPath, "index", "runs.jsonl"));
-	await mkdir(dirname(indexFilePath), { recursive: true });
-	const indexContents = rows.map((row) => JSON.stringify(row)).join("\n");
-	await writeFile(indexFilePath, `${indexContents}${indexContents.length > 0 ? "\n" : ""}`, "utf-8");
-	return {
-		indexFilePath,
-		indexedRunCount: rows.length,
-		rows
-	};
-}
-function formatTableOutput(output) {
-	return [
-		"INDEX  vieval report",
-		`Path      ${output.indexFilePath}`,
-		`Run count ${output.indexedRunCount}`
-	].join("\n");
-}
 async function runReportIndexCli(argv) {
 	try {
 		const parsed = parseReportIndexCliArguments(argv);
@@ -3803,6 +3778,31 @@ async function runReportIndexCli(argv) {
 		process.exitCode = 1;
 	}
 }
+function formatTableOutput(output) {
+	return [
+		"INDEX  vieval report",
+		`Path      ${output.indexFilePath}`,
+		`Run count ${output.indexedRunCount}`
+	].join("\n");
+}
+function normalizeCliArgv$1(argv) {
+	const normalizedArgv = argv[0] === "--" ? argv.slice(1) : [...argv];
+	if (normalizedArgv[0] === "report" && normalizedArgv[1] === "index") return normalizedArgv.slice(2);
+	if (normalizedArgv[0] === "index") return normalizedArgv.slice(1);
+	return normalizedArgv;
+}
+async function writeIndexFile(parsed) {
+	const rows = (await readReportArtifacts(parsed.reportPath)).map((artifact) => summarizeReportRunArtifact(artifact));
+	const indexFilePath = resolve(parsed.output ?? resolve(parsed.reportPath, "index", "runs.jsonl"));
+	await mkdir(dirname(indexFilePath), { recursive: true });
+	const indexContents = rows.map((row) => JSON.stringify(row)).join("\n");
+	await writeFile(indexFilePath, `${indexContents}${indexContents.length > 0 ? "\n" : ""}`, "utf-8");
+	return {
+		indexedRunCount: rows.length,
+		indexFilePath,
+		rows
+	};
+}
 //#endregion
 //#region src/cli/index.ts
 const topLevelHelpText = `
@@ -3823,9 +3823,6 @@ const topLevelHelpText = `
     $ vieval report analyze .vieval/reports/my-run
     $ vieval report index .vieval/reports --output .vieval/reports/index/runs.jsonl
 `;
-function normalizeCliArgv(argv) {
-	return argv[0] === "--" ? argv.slice(1) : [...argv];
-}
 /**
 * Parses top-level `vieval` CLI arguments into one command dispatch payload.
 *
@@ -3843,9 +3840,9 @@ function parseTopLevelCliArguments(argv) {
 	const normalizedArgv = normalizeCliArgv(argv);
 	const command = normalizedArgv[0];
 	meow(topLevelHelpText, {
+		argv: normalizedArgv,
 		autoHelp: false,
 		autoVersion: false,
-		argv: normalizedArgv,
 		importMeta: import.meta
 	});
 	if (command == null || command === "help" || command === "--help" || command === "-h") return {
@@ -3909,7 +3906,10 @@ async function runTopLevelCli(argv) {
 	}
 	await runEvalRunCli(parsed.commandArgv);
 }
+function normalizeCliArgv(argv) {
+	return argv[0] === "--" ? argv.slice(1) : [...argv];
+}
 //#endregion
 export { runTopLevelCli as n, parseTopLevelCliArguments as t };
 
-//# sourceMappingURL=cli-CHFCF8UR.mjs.map
+//# sourceMappingURL=cli-uzS81IPd.mjs.map