@kenz1117/dsh-ui-usage-billing 1.2.6 → 1.2.7

package/lib/index.js

@@ -331,7 +331,7 @@ const MODEL_CATALOG = [
 		key: "flash",
 		name: "DeepSeek V4.1 Flash",
 		provider: "DeepSeek",
-		colorVar: "dsw-static-blue-500",
+		colorVar: "ds-blue",
 		price: {
 			currency: "CNY",
 			input: 2,
@@ -349,7 +349,7 @@ const MODEL_CATALOG = [
 		key: "flash-vision-exp",
 		name: "DeepSeek V4 Flash Vision (Exp)",
 		provider: "DeepSeek",
-		colorVar: "dsw-static-blue-500",
+		colorVar: "ds-blue",
 		retired: true,
 		price: {
 			currency: "CNY",
@@ -398,7 +398,7 @@ const MODEL_CATALOG = [
 		key: "glm-5.3",
 		name: "GLM-5.3",
 		provider: "智谱 AI",
-		colorVar: "dsw-static-blue-500",
+		colorVar: "ds-blue",
 		price: {
 			currency: "CNY",
 			input: 8,
@@ -464,7 +464,7 @@ const MODEL_CATALOG = [
 		key: "glm-5-turbo",
 		name: "GLM-5-Turbo",
 		provider: "智谱 AI",
-		colorVar: "dsw-static-blue-500",
+		colorVar: "ds-blue",
 		price: {
 			currency: "CNY",
 			input: 5,
@@ -606,7 +606,7 @@ const MODEL_CATALOG = [
 		key: "qwen-plus",
 		name: "Qwen3.5-Plus",
 		provider: "阿里通义",
-		colorVar: "dsw-static-blue-500",
+		colorVar: "ds-blue",
 		price: {
 			currency: "CNY",
 			input: .8,
@@ -659,7 +659,7 @@ const MODEL_CATALOG = [
 		key: "doubao",
 		name: "Doubao Seed-2.0 Pro",
 		provider: "字节豆包",
-		colorVar: "dsw-static-red-500",
+		colorVar: "ds-red",
 		price: {
 			currency: "CNY",
 			input: 3.2,
@@ -695,7 +695,7 @@ const MODEL_CATALOG = [
 		key: "doubao-seed-evolving",
 		name: "Doubao-Seed-Evolving",
 		provider: "字节豆包",
-		colorVar: "dsw-static-red-500",
+		colorVar: "ds-red",
 		price: {
 			currency: "CNY",
 			input: 6,
@@ -803,7 +803,7 @@ const MODEL_CATALOG = [
 		key: "minimax",
 		name: "MiniMax-M3",
 		provider: "MiniMax",
-		colorVar: "dsw-static-amber-500",
+		colorVar: "ds-amber",
 		price: {
 			currency: "CNY",
 			input: 2.1,
@@ -827,7 +827,7 @@ const MODEL_CATALOG = [
 		key: "minimax-m2.7-highspeed",
 		name: "MiniMax-M2.7-highspeed",
 		provider: "MiniMax",
-		colorVar: "dsw-static-amber-500",
+		colorVar: "ds-amber",
 		price: {
 			currency: "CNY",
 			input: 4.2,
@@ -875,7 +875,7 @@ const MODEL_CATALOG = [
 		key: "yi",
 		name: "Yi-Lightning",
 		provider: "零一万物",
-		colorVar: "dsw-static-green-500",
+		colorVar: "ds-green",
 		price: {
 			currency: "CNY",
 			input: .99,
@@ -937,7 +937,7 @@ const MODEL_CATALOG = [
 		key: "gpt-6-astra",
 		name: "GPT-6 Astra",
 		provider: "OpenAI",
-		colorVar: "dsw-static-green-500",
+		colorVar: "ds-green",
 		price: {
 			currency: "USD",
 			input: 10,
@@ -949,7 +949,7 @@ const MODEL_CATALOG = [
 		key: "gpt-5.6-sol",
 		name: "GPT-5.6 Sol",
 		provider: "OpenAI",
-		colorVar: "dsw-static-green-500",
+		colorVar: "ds-green",
 		price: {
 			currency: "USD",
 			input: 5,
@@ -979,7 +979,7 @@ const MODEL_CATALOG = [
 		key: "gpt-5.6-luna",
 		name: "GPT-5.6 Luna",
 		provider: "OpenAI",
-		colorVar: "dsw-static-green-500",
+		colorVar: "ds-green",
 		price: {
 			currency: "USD",
 			input: .2,
@@ -1053,7 +1053,7 @@ const MODEL_CATALOG = [
 		key: "llama",
 		name: "Llama 4 Maverick",
 		provider: "Meta",
-		colorVar: "dsw-static-red-500",
+		colorVar: "ds-red",
 		price: {
 			currency: "USD",
 			input: .2,
@@ -1077,7 +1077,7 @@ const MODEL_CATALOG = [
 		key: "claude-opus-4-6",
 		name: "Claude Opus 4.6",
 		provider: "Anthropic",
-		colorVar: "dsw-static-red-500",
+		colorVar: "ds-red",
 		price: {
 			currency: "USD",
 			input: 5,
@@ -1113,7 +1113,7 @@ const MODEL_CATALOG = [
 		key: "claude-opus-5",
 		name: "Claude Opus 5",
 		provider: "Anthropic",
-		colorVar: "dsw-static-red-500",
+		colorVar: "ds-red",
 		price: {
 			currency: "USD",
 			input: 5,
@@ -1137,7 +1137,7 @@ const MODEL_CATALOG = [
 		key: "mistral-large-2512",
 		name: "Mistral Large 3",
 		provider: "Mistral AI",
-		colorVar: "dsw-static-violet-500",
+		colorVar: "ds-violet",
 		price: {
 			currency: "USD",
 			input: .5,
@@ -1173,7 +1173,7 @@ const MODEL_CATALOG = [
 		key: "command-a-03-2025",
 		name: "Command A",
 		provider: "Cohere",
-		colorVar: "dsw-static-cyan-500",
+		colorVar: "ds-cyan",
 		price: {
 			currency: "USD",
 			input: 2.5,
@@ -1197,7 +1197,7 @@ const MODEL_CATALOG = [
 		key: "longcat-2.0",
 		name: "LongCat 2.0",
 		provider: "美团",
-		colorVar: "dsw-static-amber-500",
+		colorVar: "ds-amber",
 		price: {
 			currency: "CNY",
 			input: 4,
@@ -1210,7 +1210,7 @@ const MODEL_CATALOG = [
 		key: "minicpm-v-4.5",
 		name: "MiniCPM-V 4.5",
 		provider: "面壁智能",
-		colorVar: "dsw-static-green-500",
+		colorVar: "ds-green",
 		price: {
 			currency: "CNY",
 			input: 1,
@@ -1236,7 +1236,7 @@ const MODEL_CATALOG = [
 		key: "dots-3-note-preview",
 		name: "Dots3-Note Preview",
 		provider: "小红书",
-		colorVar: "dsw-static-red-500",
+		colorVar: "ds-red",
 		price: {
 			currency: "CNY",
 			input: 2,
@@ -1403,7 +1403,7 @@ const MODEL_CATALOG = [
 		key: "doubao-seed-2.0-code",
 		name: "Doubao Seed-2.0 Code",
 		provider: "字节豆包",
-		colorVar: "dsw-static-red-500",
+		colorVar: "ds-red",
 		price: {
 			currency: "CNY",
 			input: 3.2,
@@ -2047,9 +2047,151 @@ function workspaceNameOf(cwd) {
 }
 /** TTFT 尖峰阈值（毫秒）：超过计为一次尖峰样本，用于定位服务端抖动。 */
 const PERF_SPIKE_MS = 1e4;
+/** 全零摘要（合并的恒等元）。 */
+const EMPTY_MODEL_DIGEST = {
+	samples: 0,
+	ttftSum: 0,
+	ttftMax: 0,
+	ttftSpikes: 0,
+	ttftP50: 0,
+	ttftP90: 0,
+	tpsCount: 0,
+	tpsSum: 0,
+	latencyCount: 0,
+	latencySum: 0,
+	estimated: 0
+};
+/** 从样本集构建模型级摘要（样本侧精确：P50/P90 为精确插值）。 */
+function modelDigestOf(samples) {
+	const ttfts = [];
+	let ttftSum = 0;
+	let ttftMax = 0;
+	let ttftSpikes = 0;
+	let tpsCount = 0;
+	let tpsSum = 0;
+	let latencyCount = 0;
+	let latencySum = 0;
+	let estimated = 0;
+	for (const sample of samples) {
+		ttfts.push(sample.ttftMs);
+		ttftSum += sample.ttftMs;
+		if (sample.ttftMs > ttftMax) ttftMax = sample.ttftMs;
+		if (sample.ttftMs > 1e4) ttftSpikes += 1;
+		if (sample.tps !== void 0) {
+			tpsCount += 1;
+			tpsSum += sample.tps;
+		}
+		if (sample.latencyMs !== void 0) {
+			latencyCount += 1;
+			latencySum += sample.latencyMs;
+		}
+		if (sample.estimated) estimated += 1;
+	}
+	return {
+		samples: samples.length,
+		ttftSum,
+		ttftMax,
+		ttftSpikes,
+		ttftP50: percentile(ttfts, .5),
+		ttftP90: percentile(ttfts, .9),
+		tpsCount,
+		tpsSum,
+		latencyCount,
+		latencySum,
+		estimated
+	};
+}
+/** 合并两个模型级摘要：计数/求和/极值精确；分位数按样本数加权。 */
+function mergeModelDigest(a, b) {
+	const n = a.samples + b.samples;
+	const weighted = (x, y) => n === 0 ? 0 : (x * a.samples + y * b.samples) / n;
+	return {
+		samples: n,
+		ttftSum: a.ttftSum + b.ttftSum,
+		ttftMax: Math.max(a.ttftMax, b.ttftMax),
+		ttftSpikes: a.ttftSpikes + b.ttftSpikes,
+		ttftP50: weighted(a.ttftP50, b.ttftP50),
+		ttftP90: weighted(a.ttftP90, b.ttftP90),
+		tpsCount: a.tpsCount + b.tpsCount,
+		tpsSum: a.tpsSum + b.tpsSum,
+		latencyCount: a.latencyCount + b.latencyCount,
+		latencySum: a.latencySum + b.latencySum,
+		estimated: a.estimated + b.estimated
+	};
+}
+/** 从样本集构建小时级摘要（均值可无损重建）。 */
+function hourDigestOf(samples) {
+	let ttftSum = 0;
+	let tpsCount = 0;
+	let tpsSum = 0;
+	for (const sample of samples) {
+		ttftSum += sample.ttftMs;
+		if (sample.tps !== void 0) {
+			tpsCount += 1;
+			tpsSum += sample.tps;
+		}
+	}
+	return {
+		samples: samples.length,
+		ttftSum,
+		tpsCount,
+		tpsSum
+	};
+}
+/** 全零小时摘要（合并的恒等元）。 */
+const EMPTY_HOUR_DIGEST = {
+	samples: 0,
+	ttftSum: 0,
+	tpsCount: 0,
+	tpsSum: 0
+};
+/** 合并两个小时级摘要（纯字段加法，无损）。 */
+function mergeHourDigest(a, b) {
+	return {
+		samples: a.samples + b.samples,
+		ttftSum: a.ttftSum + b.ttftSum,
+		tpsCount: a.tpsCount + b.tpsCount,
+		tpsSum: a.tpsSum + b.tpsSum
+	};
+}
+/**
+* 样本集 → 账本摘要形态（模型 → 摘要；小时 → 模型 → 摘要）。serializeFold
+* 与账本迁移共用同一转换，两处口径不会漂移。
+* @param samples - 一个会话的全部性能样本。
+* @returns `{ byModel, byHourModel }` 的 JSON-safe 记录。
+*/
+function perfDigestsOf(samples) {
+	const byModelSamples = /* @__PURE__ */ new Map();
+	const byHourModelSamples = /* @__PURE__ */ new Map();
+	for (const sample of samples) {
+		let modelSamples = byModelSamples.get(sample.model);
+		if (modelSamples === void 0) {
+			modelSamples = [];
+			byModelSamples.set(sample.model, modelSamples);
+		}
+		modelSamples.push(sample);
+		let hourModels = byHourModelSamples.get(sample.hour);
+		if (hourModels === void 0) {
+			hourModels = /* @__PURE__ */ new Map();
+			byHourModelSamples.set(sample.hour, hourModels);
+		}
+		let hourModelSamples = hourModels.get(sample.model);
+		if (hourModelSamples === void 0) {
+			hourModelSamples = [];
+			hourModels.set(sample.model, hourModelSamples);
+		}
+		hourModelSamples.push(sample);
+	}
+	return {
+		byModel: Object.fromEntries([...byModelSamples].map(([model, list]) => [model, modelDigestOf(list)])),
+		byHourModel: Object.fromEntries([...byHourModelSamples].map(([hour, models]) => [hour, Object.fromEntries([...models].map(([model, list]) => [model, hourDigestOf(list)]))]))
+	};
+}
 /**
 * 账本迁移注册表。首条迁移给 1.0.6 及更早的行回填 foldVersion = 1（它们全部出自
-* header 归因算法）；此后新写入的行总带当前 {@link FOLD_VERSION}。
+* header 归因算法）；perf 摘要化迁移把 v15 及更早行的逐样本 `perf` 数组转换为
+* `perfDigest`/`perfHourDigest` 摘要（日志已删的行不再被重折，迁移是它们唯一的
+* 收缩机会）；此后新写入的行总带当前 {@link FOLD_VERSION} 与摘要形态。
 */
 const LEDGER_MIGRATIONS = [{
 	id: "fold-version-backfill",
@@ -2061,6 +2203,26 @@ const LEDGER_MIGRATIONS = [{
 		}
 		return changed;
 	}
+}, {
+	id: "perf-sample-to-digest",
+	apply(document) {
+		let changed = false;
+		for (const session of document.sessions) {
+			const fold = session.fold;
+			if (Array.isArray(fold.perf)) {
+				const { byModel, byHourModel } = perfDigestsOf(fold.perf);
+				fold.perfDigest = byModel;
+				fold.perfHourDigest = byHourModel;
+				delete fold.perf;
+				changed = true;
+			} else if (fold.perfDigest === void 0 || fold.perfHourDigest === void 0) {
+				fold.perfDigest ??= {};
+				fold.perfHourDigest ??= {};
+				changed = true;
+			}
+		}
+		return changed;
+	}
 }];
 /**
 * 在加载边界对账本文档应用未执行的迁移，并记录已应用 id 供写回。
@@ -2084,6 +2246,7 @@ function runLedgerMigrations(document, migrations = LEDGER_MIGRATIONS) {
 }
 /** Serialize Map/Set-heavy fold state into a JSON-safe ledger entry. */
 function serializeFold(fold) {
+	const { byModel: perfDigest, byHourModel: perfHourDigest } = perfDigestsOf(fold.perf);
 	return {
 		total: fold.total,
 		byModel: Object.fromEntries(fold.byModel),
@@ -2096,7 +2259,8 @@ function serializeFold(fold) {
 		unpricedModels: [...fold.unpricedModels],
 		planCalls: Object.fromEntries(fold.planCalls),
 		turns: fold.turns,
-		perf: fold.perf,
+		perfDigest,
+		perfHourDigest,
 		roles: fold.roles,
 		lastActive: fold.lastActive,
 		...fold.title !== void 0 ? { title: fold.title } : {}
@@ -2116,7 +2280,9 @@ function deserializeFold(fold) {
 		unpricedModels: new Set(fold.unpricedModels),
 		planCalls: new Map(Object.entries(fold.planCalls)),
 		turns: fold.turns,
-		perf: fold.perf,
+		perf: [],
+		perfDigest: new Map(Object.entries(fold.perfDigest)),
+		perfHourDigest: new Map(Object.entries(fold.perfHourDigest).map(([hour, models]) => [hour, new Map(Object.entries(models))])),
 		roles: fold.roles,
 		lastActive: fold.lastActive,
 		...fold.title !== void 0 ? { title: fold.title } : {}
@@ -2132,7 +2298,7 @@ function ledgerSessionsOf(value) {
 		const row = entry;
 		if (typeof row.id !== "string" || row.id === "" || row.fold === null || typeof row.fold !== "object") return false;
 		const fold = row.fold;
-		return fold.total !== void 0 && fold.byModel !== void 0 && fold.byDay !== void 0 && fold.byDayModels !== void 0 && fold.bySite !== void 0 && Array.isArray(fold.unpricedModels) && fold.planCalls !== void 0 && Array.isArray(fold.turns) && Array.isArray(fold.perf) && fold.roles !== void 0 && typeof fold.lastActive === "number";
+		return fold.total !== void 0 && fold.byModel !== void 0 && fold.byDay !== void 0 && fold.byDayModels !== void 0 && fold.bySite !== void 0 && Array.isArray(fold.unpricedModels) && fold.planCalls !== void 0 && Array.isArray(fold.turns) && fold.perfDigest !== void 0 && fold.perfHourDigest !== void 0 && fold.roles !== void 0 && typeof fold.lastActive === "number";
 	});
 }
 /**
@@ -2227,6 +2393,8 @@ function freshFold() {
 		planCalls: /* @__PURE__ */ new Map(),
 		turns: [],
 		perf: [],
+		perfDigest: /* @__PURE__ */ new Map(),
+		perfHourDigest: /* @__PURE__ */ new Map(),
 		roles: {
 			userChars: 0,
 			toolChars: 0,
@@ -2468,11 +2636,6 @@ function mergeUsageInto(acc, cell) {
 	acc.officialCost += cell.officialCost;
 	if (cell.searchCalls !== void 0) acc.searchCalls = (acc.searchCalls ?? 0) + cell.searchCalls;
 }
-/** 均值（数组非空时调用；空数组按 0 兜底）。 */
-function mean(values) {
-	if (values.length === 0) return 0;
-	return values.reduce((sum, value) => sum + value, 0) / values.length;
-}
 /** 分位数（0..1）：先拷贝排序，再线性插值；空数组返回 0。 */
 function percentile(values, p) {
 	if (values.length === 0) return 0;
@@ -2489,6 +2652,53 @@ function percentile(values, p) {
 	if (a === void 0 || b === void 0) return 0;
 	return a + (b - a) * (idx - lo);
 }
+/** 活样本累加器 → 模型级摘要（跨会话合并前先归一到摘要形态）。 */
+function accumToModelDigest(acc) {
+	return {
+		samples: acc.ttfts.length,
+		ttftSum: acc.ttfts.reduce((sum, value) => sum + value, 0),
+		ttftMax: acc.ttfts.length === 0 ? 0 : Math.max(...acc.ttfts),
+		ttftSpikes: acc.ttfts.filter((ttft) => ttft > PERF_SPIKE_MS).length,
+		ttftP50: percentile(acc.ttfts, .5),
+		ttftP90: percentile(acc.ttfts, .9),
+		tpsCount: acc.tps.length,
+		tpsSum: acc.tps.reduce((sum, value) => sum + value, 0),
+		latencyCount: acc.latencies.length,
+		latencySum: acc.latencies.reduce((sum, value) => sum + value, 0),
+		estimated: acc.estimated
+	};
+}
+/** 活样本小时累加器 → 小时级摘要。 */
+function accumToHourDigest(acc) {
+	return {
+		samples: acc.ttfts.length,
+		ttftSum: acc.ttfts.reduce((sum, value) => sum + value, 0),
+		tpsCount: acc.tps.length,
+		tpsSum: acc.tps.reduce((sum, value) => sum + value, 0)
+	};
+}
+/** 模型级摘要 → 面板行（均值由求和/计数派生）。 */
+function modelPerfOf(digest) {
+	return {
+		samples: digest.samples,
+		ttftAvg: digest.samples === 0 ? 0 : digest.ttftSum / digest.samples,
+		ttftP50: digest.ttftP50,
+		ttftP90: digest.ttftP90,
+		ttftMax: digest.ttftMax,
+		ttftSpikes: digest.ttftSpikes,
+		...digest.tpsCount === 0 ? {} : { tpsAvg: digest.tpsSum / digest.tpsCount },
+		latencyAvg: digest.latencyCount === 0 ? 0 : digest.latencySum / digest.latencyCount,
+		estimatedSamples: digest.estimated
+	};
+}
+/** 小时级摘要 → 面板行（均值由求和/计数派生）。 */
+function hourPerfOf(digest) {
+	return {
+		samples: digest.samples,
+		ttftAvg: digest.samples === 0 ? 0 : digest.ttftSum / digest.samples,
+		...digest.tpsCount === 0 ? {} : { tpsAvg: digest.tpsSum / digest.tpsCount }
+	};
+}
 /**
 * 聚合配置指纹：影响折叠语义的全部配置（订阅豁免、官方名单、路由别名、搜索估值）
 * 的稳定序列化。账本行的复用判定携带该指纹——用户改配置后（如为 grok build 加
@@ -2536,7 +2746,7 @@ function createUsageAggregator(persistence, options = {}) {
 			id,
 			...cwd === void 0 ? {} : { cwd },
 			...stamp === null ? {} : { stamp },
-			foldVersion: 15,
+			foldVersion: 16,
 			fingerprint: foldFingerprint,
 			fold: serializeFold(fold)
 		};
@@ -2633,7 +2843,7 @@ function createUsageAggregator(persistence, options = {}) {
 			try {
 				const cwd = meta.cwd;
 				const ledgerRow = ledger.get(id);
-				if (ledgerRow !== void 0 && ledgerRow.stamp !== void 0 && stamp !== null && ledgerRow.stamp === stamp && (ledgerRow.foldVersion ?? 1) === 15 && ledgerRow.fingerprint === foldFingerprint) {
+				if (ledgerRow !== void 0 && ledgerRow.stamp !== void 0 && stamp !== null && ledgerRow.stamp === stamp && (ledgerRow.foldVersion ?? 1) === 16 && ledgerRow.fingerprint === foldFingerprint) {
 					const fold = deserializeFold(ledgerRow.fold);
 					folds.push({
 						id,
@@ -2715,7 +2925,7 @@ function createUsageAggregator(persistence, options = {}) {
 			for (const entry of ledger.values()) {
 				if (included.has(entry.id)) continue;
 				try {
-					const stale = (entry.foldVersion ?? 1) < 15 || entry.fingerprint !== foldFingerprint;
+					const stale = (entry.foldVersion ?? 1) < 16 || entry.fingerprint !== foldFingerprint;
 					folds.push({
 						id: entry.id,
 						...entry.cwd === void 0 ? {} : { cwd: entry.cwd },
@@ -2752,6 +2962,8 @@ function createUsageAggregator(persistence, options = {}) {
 		};
 		const perfModel = /* @__PURE__ */ new Map();
 		const perfHourModel = /* @__PURE__ */ new Map();
+		const perfModelDigest = /* @__PURE__ */ new Map();
+		const perfHourDigest = /* @__PURE__ */ new Map();
 		for (const { id: sessionId, cwd, fold, staleLedger } of folds) {
 			mergeUsageInto(total, fold.total);
 			roles.userChars += fold.roles.userChars;
@@ -2798,6 +3010,15 @@ function createUsageAggregator(persistence, options = {}) {
 				hourAccum.ttfts.push(sample.ttftMs);
 				if (sample.tps !== void 0) hourAccum.tps.push(sample.tps);
 			}
+			for (const [model, digest] of fold.perfDigest) perfModelDigest.set(model, mergeModelDigest(perfModelDigest.get(model) ?? EMPTY_MODEL_DIGEST, digest));
+			for (const [hour, models] of fold.perfHourDigest) {
+				let hourModels = perfHourDigest.get(hour);
+				if (hourModels === void 0) {
+					hourModels = /* @__PURE__ */ new Map();
+					perfHourDigest.set(hour, hourModels);
+				}
+				for (const [model, digest] of models) hourModels.set(model, mergeHourDigest(hourModels.get(model) ?? EMPTY_HOUR_DIGEST, digest));
+			}
 			for (const row of fold.turns) turnRows.push({
 				sessionId,
 				...row
@@ -2841,23 +3062,26 @@ function createUsageAggregator(persistence, options = {}) {
 		};
 		const toModelDayRecord = (map) => Object.fromEntries([...map].map(([day, models]) => [day, Object.fromEntries(models)]));
 		const toModelDaySiteRecord = (map) => Object.fromEntries([...map].map(([day, models]) => [day, Object.fromEntries([...models].map(([model, sites]) => [model, Object.fromEntries(sites)]))]));
-		const perf = perfModel.size === 0 ? void 0 : {
-			byModel: Object.fromEntries([...perfModel].map(([model, acc]) => [model, {
-				samples: acc.ttfts.length,
-				ttftAvg: mean(acc.ttfts),
-				ttftP50: percentile(acc.ttfts, .5),
-				ttftP90: percentile(acc.ttfts, .9),
-				ttftMax: Math.max(...acc.ttfts),
-				ttftSpikes: acc.ttfts.filter((ttft) => ttft > PERF_SPIKE_MS).length,
-				...acc.tps.length === 0 ? {} : { tpsAvg: mean(acc.tps) },
-				latencyAvg: acc.latencies.length === 0 ? 0 : mean(acc.latencies),
-				estimatedSamples: acc.estimated
-			}])),
-			byHourModel: Object.fromEntries([...perfHourModel].map(([hour, models]) => [hour, Object.fromEntries([...models].map(([model, acc]) => [model, {
-				samples: acc.ttfts.length,
-				ttftAvg: mean(acc.ttfts),
-				...acc.tps.length === 0 ? {} : { tpsAvg: mean(acc.tps) }
-			}]))]))
+		const modelKeys = /* @__PURE__ */ new Set([...perfModel.keys(), ...perfModelDigest.keys()]);
+		const hourKeys = /* @__PURE__ */ new Set([...perfHourModel.keys(), ...perfHourDigest.keys()]);
+		const perf = modelKeys.size === 0 && hourKeys.size === 0 ? void 0 : {
+			byModel: Object.fromEntries([...modelKeys].map((model) => {
+				const live = perfModel.get(model);
+				const restored = perfModelDigest.get(model);
+				const digest = live === void 0 ? restored : mergeModelDigest(accumToModelDigest(live), restored ?? EMPTY_MODEL_DIGEST);
+				return digest === void 0 ? null : [model, modelPerfOf(digest)];
+			}).filter((entry) => entry !== null)),
+			byHourModel: Object.fromEntries([...hourKeys].sort().map((hour) => {
+				const liveModels = perfHourModel.get(hour);
+				const digestModels = perfHourDigest.get(hour);
+				const modelKeysInHour = /* @__PURE__ */ new Set([...liveModels?.keys() ?? [], ...digestModels?.keys() ?? []]);
+				return [hour, Object.fromEntries([...modelKeysInHour].map((model) => {
+					const live = liveModels?.get(model);
+					const restored = digestModels?.get(model);
+					const digest = live === void 0 ? restored : mergeHourDigest(accumToHourDigest(live), restored ?? EMPTY_HOUR_DIGEST);
+					return digest === void 0 ? null : [model, hourPerfOf(digest)];
+				}).filter((entry) => entry !== null))];
+			}))
 		};
 		lastDoc = {
 			version: 4,
@@ -2903,6 +3127,29 @@ function createUsageAggregator(persistence, options = {}) {
 			});
 			return inflight;
 		},
+		async sessionUsageOf(sessionId) {
+			await this.aggregate();
+			const hit = cache.get(sessionId);
+			if (hit !== void 0) return {
+				calls: hit.fold.total.calls,
+				cost: hit.fold.total.cost,
+				input: hit.fold.total.input,
+				output: hit.fold.total.output
+			};
+			const row = ledger.get(sessionId);
+			if (row === void 0) return void 0;
+			try {
+				const fold = deserializeFold(row.fold);
+				return {
+					calls: fold.total.calls,
+					cost: fold.total.cost,
+					input: fold.total.input,
+					output: fold.total.output
+				};
+			} catch {
+				return;
+			}
+		},
 		async flush() {
 			if (inflight !== void 0) await inflight.catch(() => {});
 			await saveLedger(true);
@@ -5468,8 +5715,10 @@ const DEFAULT_BALANCE_API_KEY_ENV = "DEEPSEEK_API_KEY";
 * 下发，供「设置 → 插件信息卡」展示。发布版 lib/index.js 相对包根解析。
 */
 const PACKAGE_VERSION = createRequire(import.meta.url)("../package.json").version ?? "0.0.0";
-/** 统计快照的落盘节流（毫秒）：前端 30 秒轮询，快照最多每 30 秒写一次。 */
-const SNAPSHOT_INTERVAL_MS = 3e4;
+/** 统计快照的落盘节流（毫秒）：快照只服务聚合失败回退与重启首屏（展示级
+*  新鲜度即可），降到 5 分钟一次——原子写含上一版备份是两次全量 JSON 落盘，
+*  重度用户快照上百 KB，跟随前端轮询 30 秒写一次纯属写放大。 */
+const SNAPSHOT_INTERVAL_MS = 3e5;
 /** usage-stats 响应的等待预算（毫秒）：聚合在此预算内完成就回实时数据，
 *  超预算立即回最近快照（stale-while-revalidate），后台折叠继续跑。
 *  轻量用户暖缓存下的聚合是毫秒级，预算几乎不会触发。 */
@@ -6002,6 +6251,11 @@ function apply(ctx, config = {}) {
 					}
 					const sessionId = exec.agent?.id;
 					if (sessionId === void 0) throw new Error("usage_stats 的 session 范围需要 agent 会话上下文");
+					const summary = await aggregator.sessionUsageOf(String(sessionId));
+					if (summary !== void 0) return {
+						range: args.range,
+						...summary
+					};
 					let cost = 0;
 					let calls = 0;
 					let input = 0;