aicodeswitch 5.2.10 → 5.2.12

package/dist/server/performance-tracker.js

@@ -0,0 +1,377 @@
+"use strict";
+var __awaiter = (this && this.__awaiter) || function (thisArg, _arguments, P, generator) {
+    function adopt(value) { return value instanceof P ? value : new P(function (resolve) { resolve(value); }); }
+    return new (P || (P = Promise))(function (resolve, reject) {
+        function fulfilled(value) { try { step(generator.next(value)); } catch (e) { reject(e); } }
+        function rejected(value) { try { step(generator["throw"](value)); } catch (e) { reject(e); } }
+        function step(result) { result.done ? resolve(result.value) : adopt(result.value).then(fulfilled, rejected); }
+        step((generator = generator.apply(thisArg, _arguments || [])).next());
+    });
+};
+var __importDefault = (this && this.__importDefault) || function (mod) {
+    return (mod && mod.__esModule) ? mod : { "default": mod };
+};
+Object.defineProperty(exports, "__esModule", { value: true });
+exports.ServicePerformanceTracker = void 0;
+/**
+ * ServicePerformanceTracker - 服务性能统计全局聚合模块
+ *
+ * 设计要点（详见 docs/PRD/service-performance-tpm.md）：
+ * - 全局统计，与 AUTH 模式无关；普通路由 + AccessKey 路由流量统一采集。
+ * - 两个数据点：TTFT（首 Token 返回时间）、TPM（每分钟吐 token 数）。
+ * - 三级聚合：供应商 → 服务 → 模型；上卷基于 sum+count 加权，avg 由 sum/count 派生。
+ * - 走势按小时桶（键 "YYYY-MM-DD HH"，保留 72 桶）。
+ * - 内存增量 + debounce(5s) flush + 原子写（tmp+rename）。
+ *
+ * recordPerformance 为纯内存同步操作，可在请求完成路径无开销调用。
+ */
+const path_1 = __importDefault(require("path"));
+const promises_1 = __importDefault(require("fs/promises"));
+const HOURLY_BUCKET_LIMIT = 72; // 保留最近 72 个小时桶（约 3 天）
+class ServicePerformanceTracker {
+    constructor(dataPath) {
+        Object.defineProperty(this, "dataPath", {
+            enumerable: true,
+            configurable: true,
+            writable: true,
+            value: void 0
+        });
+        Object.defineProperty(this, "file", {
+            enumerable: true,
+            configurable: true,
+            writable: true,
+            value: { vendors: {} }
+        });
+        Object.defineProperty(this, "dirty", {
+            enumerable: true,
+            configurable: true,
+            writable: true,
+            value: false
+        });
+        Object.defineProperty(this, "flushTimer", {
+            enumerable: true,
+            configurable: true,
+            writable: true,
+            value: null
+        });
+        Object.defineProperty(this, "FLUSH_INTERVAL", {
+            enumerable: true,
+            configurable: true,
+            writable: true,
+            value: 5000
+        }); // 5s
+        this.dataPath = dataPath;
+    }
+    /** 初始化：加载已有数据文件 */
+    initialize() {
+        return __awaiter(this, void 0, void 0, function* () {
+            try {
+                const raw = yield promises_1.default.readFile(this.filePath, 'utf-8');
+                const parsed = JSON.parse(raw);
+                if (parsed && parsed.vendors)
+                    this.file = parsed;
+            }
+            catch (_a) {
+                // 首次启动或文件损坏，使用空桶
+                this.file = { vendors: {} };
+            }
+        });
+    }
+    /**
+     * 记录一次请求的性能数据点（三级同步聚合）。
+     * 纯内存操作，不阻塞调用方。
+     */
+    recordPerformance(vendorId, vendorName, serviceId, serviceName, model, metrics, timestamp = Date.now()) {
+        if (!vendorId || !serviceId || !model)
+            return;
+        const hour = this.formatHourKey(timestamp);
+        // 模型级（最细，维护极值）
+        const modelAgg = this.ensureModel(vendorId, vendorName, serviceId, serviceName, model);
+        this.accumulate(modelAgg, metrics, hour, /* withExtremes */ true);
+        // 服务级上卷
+        const serviceAgg = this.ensureService(vendorId, vendorName, serviceId, serviceName);
+        this.accumulate(serviceAgg, metrics, hour, /* withExtremes */ false);
+        // 供应商级上卷
+        const vendorAgg = this.ensureVendor(vendorId, vendorName);
+        this.accumulate(vendorAgg, metrics, hour, /* withExtremes */ false);
+        this.dirty = true;
+    }
+    // ---------------- 读取（派生视图） ----------------
+    /** 全部供应商一览（vendorRollup 派生） */
+    getVendorsOverview() {
+        return Object.entries(this.file.vendors).map(([vendorId, v]) => ({
+            vendorId,
+            vendorName: v.vendorName,
+            derived: this.derive(v.vendorRollup),
+        }));
+    }
+    /** 全部 API 服务平铺一览（含所属供应商），用于「API 服务」维度对比 */
+    getServicesOverview() {
+        const out = [];
+        for (const [vendorId, v] of Object.entries(this.file.vendors)) {
+            for (const [serviceId, s] of Object.entries(v.services)) {
+                out.push({
+                    serviceId,
+                    serviceName: s.serviceName,
+                    vendorId,
+                    vendorName: v.vendorName,
+                    derived: this.derive(s.serviceRollup),
+                });
+            }
+        }
+        return out;
+    }
+    /** 某供应商：自身 rollup + 其下所有服务 rollup */
+    getVendorDetail(vendorId) {
+        const v = this.file.vendors[vendorId];
+        if (!v)
+            return null;
+        return {
+            vendorName: v.vendorName,
+            derived: this.derive(v.vendorRollup),
+            hourly: this.trendFrom(vendorId, undefined, undefined),
+            services: Object.entries(v.services).map(([serviceId, s]) => ({
+                serviceId,
+                serviceName: s.serviceName,
+                derived: this.derive(s.serviceRollup),
+            })),
+        };
+    }
+    /** 某服务：自身 rollup + 其下所有模型 */
+    getServiceDetail(serviceId) {
+        const found = this.locateService(serviceId);
+        if (!found)
+            return null;
+        const { vendorId, vendorName, serviceEntry } = found;
+        return {
+            vendorId,
+            vendorName,
+            serviceName: serviceEntry.serviceName,
+            derived: this.derive(serviceEntry.serviceRollup),
+            hourly: this.trendFrom(vendorId, serviceId, undefined),
+            models: Object.entries(serviceEntry.models).map(([model, agg]) => ({
+                model,
+                derived: this.derive(agg),
+            })),
+        };
+    }
+    /** 单模型：派生 + 小时走势 + 极值 */
+    getModelDetail(serviceId, model) {
+        const found = this.locateService(serviceId);
+        if (!found)
+            return null;
+        const agg = found.serviceEntry.models[model];
+        if (!agg)
+            return null;
+        return {
+            derived: this.derive(agg),
+            hourly: this.trendFrom(found.vendorId, serviceId, model),
+        };
+    }
+    // ---------------- 持久化 ----------------
+    flush() {
+        return __awaiter(this, void 0, void 0, function* () {
+            if (!this.dirty)
+                return;
+            this.dirty = false;
+            yield this.save();
+        });
+    }
+    startAutoFlush() {
+        if (this.flushTimer)
+            return;
+        this.flushTimer = setInterval(() => {
+            this.flush().catch(err => console.error('[PerformanceTracker] Auto flush error:', err));
+        }, this.FLUSH_INTERVAL);
+    }
+    stopAutoFlush() {
+        if (this.flushTimer) {
+            clearInterval(this.flushTimer);
+            this.flushTimer = null;
+        }
+    }
+    get filePath() {
+        return path_1.default.join(this.dataPath, 'service-performance.json');
+    }
+    save() {
+        return __awaiter(this, void 0, void 0, function* () {
+            const tmp = this.filePath + '.tmp';
+            yield promises_1.default.mkdir(path_1.default.dirname(this.filePath), { recursive: true });
+            yield promises_1.default.writeFile(tmp, JSON.stringify(this.file, null, 2), 'utf-8');
+            yield promises_1.default.rename(tmp, this.filePath);
+        });
+    }
+    // ---------------- 内部：结构创建 ----------------
+    ensureVendor(vendorId, vendorName) {
+        let v = this.file.vendors[vendorId];
+        if (!v) {
+            v = {
+                vendorName,
+                vendorRollup: this.emptyAggregate(),
+                services: {},
+            };
+            this.file.vendors[vendorId] = v;
+        }
+        if (vendorName && !v.vendorName)
+            v.vendorName = vendorName;
+        return v.vendorRollup;
+    }
+    ensureService(vendorId, vendorName, serviceId, serviceName) {
+        var _a;
+        const v = (_a = this.file.vendors[vendorId]) !== null && _a !== void 0 ? _a : (this.file.vendors[vendorId] = {
+            vendorName, vendorRollup: this.emptyAggregate(), services: {},
+        });
+        let s = v.services[serviceId];
+        if (!s) {
+            s = {
+                serviceName,
+                serviceRollup: this.emptyAggregate(),
+                models: {},
+                updatedAt: Date.now(),
+            };
+            v.services[serviceId] = s;
+        }
+        if (serviceName && !s.serviceName)
+            s.serviceName = serviceName;
+        s.updatedAt = Date.now();
+        return s.serviceRollup;
+    }
+    ensureModel(vendorId, vendorName, serviceId, serviceName, model) {
+        // 确保供应商 + 服务节点存在（service rollup 由 recordPerformance 单独累加）
+        this.ensureService(vendorId, vendorName, serviceId, serviceName);
+        const s = this.file.vendors[vendorId].services[serviceId];
+        let m = s.models[model];
+        if (!m) {
+            m = this.emptyAggregate();
+            s.models[model] = m;
+        }
+        return m;
+    }
+    // ---------------- 内部：累加 ----------------
+    emptyAggregate() {
+        return {
+            precise: this.emptyBucket(),
+            estimated: this.emptyBucket(),
+            errorCount: 0,
+            hourly: {},
+        };
+    }
+    emptyBucket() {
+        return { count: 0, sumTtftMs: 0, sumTps: 0, totalOutputTokens: 0 };
+    }
+    accumulate(agg, m, hour, withExtremes) {
+        var _a;
+        if (m.isError) {
+            agg.errorCount += 1;
+            return;
+        }
+        const bucket = m.timingAccuracy === 'precise' ? agg.precise : agg.estimated;
+        const hasTtft = m.timingAccuracy === 'precise' && typeof m.ttftMs === 'number';
+        const hasTps = typeof m.tokensPerSecond === 'number';
+        bucket.count += 1;
+        if (hasTtft)
+            bucket.sumTtftMs += m.ttftMs;
+        if (hasTps)
+            bucket.sumTps += m.tokensPerSecond;
+        if (m.outputTokens)
+            bucket.totalOutputTokens += m.outputTokens;
+        // 小时桶（仅精确样本计入走势，避免估算样本污染）
+        if (m.timingAccuracy === 'precise') {
+            const hb = (_a = agg.hourly[hour]) !== null && _a !== void 0 ? _a : (agg.hourly[hour] = this.emptyBucket());
+            hb.count += 1;
+            if (hasTtft)
+                hb.sumTtftMs += m.ttftMs;
+            if (hasTps)
+                hb.sumTps += m.tokensPerSecond;
+            if (m.outputTokens)
+                hb.totalOutputTokens += m.outputTokens;
+            this.trimHourly(agg.hourly);
+        }
+        // 极值（仅模型级、仅精确样本）
+        if (withExtremes && m.timingAccuracy === 'precise') {
+            if (hasTtft) {
+                if (agg.minTtftMs === undefined || m.ttftMs < agg.minTtftMs)
+                    agg.minTtftMs = m.ttftMs;
+                if (agg.maxTtftMs === undefined || m.ttftMs > agg.maxTtftMs)
+                    agg.maxTtftMs = m.ttftMs;
+            }
+            if (hasTps) {
+                if (agg.minTps === undefined || m.tokensPerSecond < agg.minTps)
+                    agg.minTps = m.tokensPerSecond;
+                if (agg.maxTps === undefined || m.tokensPerSecond > agg.maxTps)
+                    agg.maxTps = m.tokensPerSecond;
+            }
+        }
+    }
+    trimHourly(hourly) {
+        const keys = Object.keys(hourly);
+        if (keys.length <= HOURLY_BUCKET_LIMIT)
+            return;
+        keys.sort(); // "YYYY-MM-DD HH" 字典序即时间序
+        const drop = keys.length - HOURLY_BUCKET_LIMIT;
+        for (let i = 0; i < drop; i++)
+            delete hourly[keys[i]];
+    }
+    // ---------------- 内部：派生 ----------------
+    derive(agg) {
+        const p = agg.precise;
+        const count = p.count;
+        const avgTtftMs = count > 0 ? p.sumTtftMs / count : 0;
+        const avgTps = count > 0 ? p.sumTps / count : 0;
+        return {
+            count,
+            avgTtftMs,
+            avgTpm: avgTps * 60,
+            minTtftMs: agg.minTtftMs,
+            maxTtftMs: agg.maxTtftMs,
+            minTps: agg.minTps,
+            maxTps: agg.maxTps,
+            errorCount: agg.errorCount,
+            totalOutputTokens: p.totalOutputTokens,
+            successRate: count + agg.errorCount > 0 ? count / (count + agg.errorCount) : 0,
+        };
+    }
+    trendFrom(vendorId, serviceId, model) {
+        var _a, _b;
+        const v = this.file.vendors[vendorId];
+        if (!v)
+            return [];
+        let agg;
+        if (model && serviceId) {
+            agg = (_a = v.services[serviceId]) === null || _a === void 0 ? void 0 : _a.models[model];
+        }
+        else if (serviceId) {
+            agg = (_b = v.services[serviceId]) === null || _b === void 0 ? void 0 : _b.serviceRollup;
+        }
+        else {
+            agg = v.vendorRollup;
+        }
+        if (!agg)
+            return [];
+        return Object.entries(agg.hourly)
+            .sort(([a], [b]) => (a < b ? -1 : a > b ? 1 : 0))
+            .map(([hour, b]) => ({
+            hour,
+            count: b.count,
+            avgTtftMs: b.count > 0 ? b.sumTtftMs / b.count : 0,
+            avgTpm: b.count > 0 ? (b.sumTps / b.count) * 60 : 0,
+        }));
+    }
+    locateService(serviceId) {
+        for (const [vendorId, v] of Object.entries(this.file.vendors)) {
+            const s = v.services[serviceId];
+            if (s)
+                return { vendorId, vendorName: v.vendorName, serviceEntry: s };
+        }
+        return null;
+    }
+    formatHourKey(ts) {
+        const d = new Date(ts);
+        const yyyy = d.getFullYear();
+        const mm = String(d.getMonth() + 1).padStart(2, '0');
+        const dd = String(d.getDate()).padStart(2, '0');
+        const hh = String(d.getHours()).padStart(2, '0');
+        return `${yyyy}-${mm}-${dd} ${hh}`;
+    }
+}
+exports.ServicePerformanceTracker = ServicePerformanceTracker;

package/dist/server/proxy-server.js

@@ -52,6 +52,7 @@ const crypto_1 = __importDefault(require("crypto"));
 const streaming_1 = require("./transformers/streaming");
 const model_rewrite_transform_1 = require("./transformers/model-rewrite-transform");
 const chunk_collector_1 = require("./transformers/chunk-collector");
+const stream_timing_transform_1 = require("./transformers/stream-timing-transform");
 const rules_status_service_1 = require("./rules-status-service");
 const index_1 = require("./conversions/index");
 const stream_converter_adapter_1 = require("./conversions/stream-converter-adapter");
@@ -219,6 +220,12 @@ class ProxyServer {
             writable: true,
             value: null
         });
+        Object.defineProperty(this, "performanceTracker", {
+            enumerable: true,
+            configurable: true,
+            writable: true,
+            value: null
+        });
         // 请求去重缓存：用于防止同一个请求被重复计数（如网络重试）
         // key: requestHash, value: timestamp
         Object.defineProperty(this, "requestDedupeCache", {
@@ -253,6 +260,43 @@ class ProxyServer {
     getAccessKeyModule() {
         return this.accessKeyModule;
     }
+    /** 设置服务性能统计 tracker（全局，与 AUTH 无关） */
+    setPerformanceTracker(tracker) {
+        this.performanceTracker = tracker;
+    }
+    /** 获取服务性能统计 tracker */
+    getPerformanceTracker() {
+        return this.performanceTracker;
+    }
+    /**
+     * 采集一次请求的服务性能数据点（全局，与 AUTH 无关）。
+     * 在两条转发路径的 finalizeLog 公共点调用，覆盖 AccessKey + 普通路由。
+     * 流式：依据 streamTiming 精确计算 TTFT 与生成阶段吞吐；非流式：端到端估算（estimated）。
+     */
+    emitPerformance(params) {
+        const tracker = this.performanceTracker;
+        if (!tracker)
+            return;
+        const { statusCode, startTime, usage, streamTiming, service, vendorId, vendorName, model } = params;
+        const isError = statusCode >= 400;
+        const outputTokens = usage === null || usage === void 0 ? void 0 : usage.outputTokens;
+        const responseMs = Date.now() - startTime;
+        let ttftMs;
+        let tokensPerSecond;
+        let timingAccuracy = 'estimated';
+        if (streamTiming && streamTiming.hasTiming()) {
+            timingAccuracy = 'precise';
+            ttftMs = streamTiming.firstEventAt - startTime;
+            const generationMs = streamTiming.lastEventAt - streamTiming.firstEventAt;
+            if (outputTokens && generationMs > 0) {
+                tokensPerSecond = outputTokens / (generationMs / 1000);
+            }
+        }
+        else if (outputTokens && responseMs > 0) {
+            tokensPerSecond = outputTokens / (responseMs / 1000);
+        }
+        tracker.recordPerformance(vendorId !== null && vendorId !== void 0 ? vendorId : service.vendorId, vendorName, service.id, service.name, model, { ttftMs, tokensPerSecond, outputTokens, timingAccuracy, isError });
+    }
     /**
      * 从请求中提取 API Key（支持三种 Header，按优先级依次尝试）
      */
@@ -3544,12 +3588,20 @@ class ProxyServer {
             let downstreamResponseBodyForLog;
             let upstreamRequestForLog;
             let actuallyUsedProxy = false; // 标记是否实际使用了代理
+            // 服务性能打点：流式分支会创建实例并注入 pipeline；finalizeLog 据此判定 precise/estimated
+            let streamTiming = null;
             // 标记规则正在使用
             rules_status_service_1.rulesStatusBroadcaster.markRuleInUse(route.id, rule.id);
             const finalizeLog = (statusCode, error) => __awaiter(this, void 0, void 0, function* () {
-                var _a, _b, _c, _d, _e, _f, _g;
+                var _a, _b, _c, _d, _e, _f, _g, _h;
                 if (logged)
                     return;
+                // 服务性能数据点采集（全局，与 AUTH 无关；独立于 enableLogging 开关）
+                this.emitPerformance({
+                    statusCode, startTime, usage: usageForLog, streamTiming,
+                    service, vendorId: vendor === null || vendor === void 0 ? void 0 : vendor.id, vendorName: vendor === null || vendor === void 0 ? void 0 : vendor.name,
+                    model: rule.targetModel || ((_a = req.body) === null || _a === void 0 ? void 0 : _a.model),
+                });
                 const isError = statusCode >= 400;
                 if (isError) {
                     console.log(`\x1b[31m[Request Error]\x1b[0m client=${targetType}, session=${sessionId}, rule=${rule.id}(${rule.contentType}), vendor=${(vendor === null || vendor === void 0 ? void 0 : vendor.name) || '-'}, service=${service.name}, status=${statusCode}, time=${Date.now() - startTime}ms${error ? `, error=${error}` : ''}`);
@@ -3558,7 +3610,7 @@ class ProxyServer {
                     console.log(`\x1b[33m[Request End]\x1b[0m client=${targetType}, session=${sessionId}, rule=${rule.id}(${rule.contentType}), vendor=${(vendor === null || vendor === void 0 ? void 0 : vendor.name) || '-'}, service=${service.name}, status=${statusCode}, time=${Date.now() - startTime}ms`);
                 }
                 // 检查是否启用日志记录（默认启用）
-                const enableLogging = ((_a = this.config) === null || _a === void 0 ? void 0 : _a.enableLogging) !== false; // 默认为 true
+                const enableLogging = ((_b = this.config) === null || _b === void 0 ? void 0 : _b.enableLogging) !== false; // 默认为 true
                 if (!enableLogging) {
                     return;
                 }
@@ -3585,10 +3637,10 @@ class ProxyServer {
                             targetType,
                             targetServiceId: service.id,
                             targetServiceName: service.name,
-                            targetModel: rule.targetModel || ((_b = req.body) === null || _b === void 0 ? void 0 : _b.model),
+                            targetModel: rule.targetModel || ((_c = req.body) === null || _c === void 0 ? void 0 : _c.model),
                             vendorId: service.vendorId,
                             vendorName: vendor === null || vendor === void 0 ? void 0 : vendor.name,
-                            requestModel: (_c = req.body) === null || _c === void 0 ? void 0 : _c.model,
+                            requestModel: (_d = req.body) === null || _d === void 0 ? void 0 : _d.model,
                             tags: this.buildRelayTags(relayedForLog, useOriginalConfig),
                             responseHeaders: responseHeadersForLog,
                             responseBody: responseBodyForLog,
@@ -3622,7 +3674,7 @@ class ProxyServer {
                                 vendorName: vendor === null || vendor === void 0 ? void 0 : vendor.name,
                                 serviceId: service.id,
                                 serviceName: service.name,
-                                model: rule.targetModel || ((_d = req.body) === null || _d === void 0 ? void 0 : _d.model),
+                                model: rule.targetModel || ((_e = req.body) === null || _e === void 0 ? void 0 : _e.model),
                                 totalTokens: sessionTokens,
                             }).catch(err => console.error('[KeySession] upsert error:', err));
                         }
@@ -3649,10 +3701,10 @@ class ProxyServer {
                             targetType,
                             targetServiceId: service.id,
                             targetServiceName: service.name,
-                            targetModel: rule.targetModel || ((_e = req.body) === null || _e === void 0 ? void 0 : _e.model),
+                            targetModel: rule.targetModel || ((_f = req.body) === null || _f === void 0 ? void 0 : _f.model),
                             vendorId: service.vendorId,
                             vendorName: vendor === null || vendor === void 0 ? void 0 : vendor.name,
-                            requestModel: (_f = req.body) === null || _f === void 0 ? void 0 : _f.model,
+                            requestModel: (_g = req.body) === null || _g === void 0 ? void 0 : _g.model,
                             tags: this.buildRelayTags(relayedForLog, useOriginalConfig),
                         });
                     }
@@ -3665,7 +3717,7 @@ class ProxyServer {
                 const vendors = this.dbManager.getVendors();
                 const vendorForLog = vendors.find(v => v.id === service.vendorId);
                 // 从请求体中提取模型信息
-                const requestModel = (_g = req.body) === null || _g === void 0 ? void 0 : _g.model;
+                const requestModel = (_h = req.body) === null || _h === void 0 ? void 0 : _h.model;
                 const tagsForLog = this.buildRelayTags(relayedForLog, useOriginalConfig);
                 if (extraTagsForLog.length > 0) {
                     tagsForLog.push(...extraTagsForLog);
@@ -4061,6 +4113,8 @@ class ProxyServer {
                     const downstreamChunkCollector = new chunk_collector_1.ChunkCollectorTransform(() => {
                         rules_status_service_1.rulesStatusBroadcaster.refreshRuleInUse(route.id, rule.id);
                     });
+                    // 服务性能打点：记录首/末 SSE 事件时间，用于 TTFT 与生成阶段吞吐
+                    streamTiming = new stream_timing_transform_1.StreamTimingTransform(startTime);
                     const compactResponseSanitizer = rule.contentType === 'compact' && targetType === 'claude-code'
                         ? new ClaudeCompactResponseSanitizer()
                         : null;
@@ -4102,7 +4156,7 @@ class ProxyServer {
                         ensureResponseWritable();
                         return yield new Promise((resolve, reject) => {
                             if (converter) {
-                                const streamStages = [streamSource, parser, eventCollector, converter];
+                                const streamStages = [streamSource, parser, eventCollector, streamTiming, converter];
                                 if (compactResponseSanitizer) {
                                     streamStages.push(compactResponseSanitizer);
                                 }
@@ -4120,7 +4174,7 @@ class ProxyServer {
                                 });
                                 return;
                             }
-                            const streamStages = [streamSource, parser, eventCollector];
+                            const streamStages = [streamSource, parser, eventCollector, streamTiming];
                             if (compactResponseSanitizer) {
                                 streamStages.push(compactResponseSanitizer);
                             }
@@ -4608,6 +4662,8 @@ class ProxyServer {
             let responseBodyForLog;
             let downstreamResponseBodyForLog;
             let streamChunksForLog;
+            // 服务性能打点：流式分支会创建实例并注入 pipeline
+            let streamTiming = null;
             let responseHeadersForLog;
             let upstreamRequestForLog;
             let relayedForLog = true;
@@ -4629,10 +4685,16 @@ class ProxyServer {
                 requestBody = (0, compact_1.normalizeClaudeCompactRequestBody)(requestBody);
             }
             const finalizeLog = (statusCode, error) => __awaiter(this, void 0, void 0, function* () {
-                var _a, _b, _c, _d, _e;
+                var _a, _b, _c, _d, _e, _f;
                 if (logged)
                     return;
                 logged = true;
+                // 服务性能数据点采集（全局，与 AUTH 无关；独立于 enableLogging 开关）
+                this.emitPerformance({
+                    statusCode, startTime, usage: usageForLog, streamTiming,
+                    service, vendorId: vendor === null || vendor === void 0 ? void 0 : vendor.id, vendorName: vendor === null || vendor === void 0 ? void 0 : vendor.name,
+                    model: rule.targetModel || ((_a = req.body) === null || _a === void 0 ? void 0 : _a.model),
+                });
                 // AccessKey 独立日志处理
                 const accessKeyCtx = req._accessKeyCtx;
                 if (accessKeyCtx && this.accessKeyModule) {
@@ -4651,10 +4713,10 @@ class ProxyServer {
                             ruleId: rule.id,
                             targetServiceId: service.id,
                             targetServiceName: service.name,
-                            targetModel: rule.targetModel || ((_a = req.body) === null || _a === void 0 ? void 0 : _a.model),
+                            targetModel: rule.targetModel || ((_b = req.body) === null || _b === void 0 ? void 0 : _b.model),
                             vendorId: service.vendorId,
                             vendorName: vendor === null || vendor === void 0 ? void 0 : vendor.name,
-                            requestModel: (_b = req.body) === null || _b === void 0 ? void 0 : _b.model,
+                            requestModel: (_c = req.body) === null || _c === void 0 ? void 0 : _c.model,
                             tags: this.buildRelayTags(relayedForLog),
                         });
                         if (usageForLog && statusCode < 400) {
@@ -4680,7 +4742,7 @@ class ProxyServer {
                                 vendorName: vendor === null || vendor === void 0 ? void 0 : vendor.name,
                                 serviceId: service.id,
                                 serviceName: service.name,
-                                model: rule.targetModel || ((_c = req.body) === null || _c === void 0 ? void 0 : _c.model),
+                                model: rule.targetModel || ((_d = req.body) === null || _d === void 0 ? void 0 : _d.model),
                                 totalTokens: sessionTokens,
                             }).catch(err => console.error('[KeySession] upsert error:', err));
                         }
@@ -4704,10 +4766,10 @@ class ProxyServer {
                             ruleId: rule.id,
                             targetServiceId: service.id,
                             targetServiceName: service.name,
-                            targetModel: rule.targetModel || ((_d = req.body) === null || _d === void 0 ? void 0 : _d.model),
+                            targetModel: rule.targetModel || ((_e = req.body) === null || _e === void 0 ? void 0 : _e.model),
                             vendorId: service.vendorId,
                             vendorName: vendor === null || vendor === void 0 ? void 0 : vendor.name,
-                            requestModel: (_e = req.body) === null || _e === void 0 ? void 0 : _e.model,
+                            requestModel: (_f = req.body) === null || _f === void 0 ? void 0 : _f.model,
                             tags: this.buildRelayTags(relayedForLog),
                         });
                     }
@@ -4877,6 +4939,8 @@ class ProxyServer {
                     const downstreamChunkCollector = new chunk_collector_1.ChunkCollectorTransform(() => {
                         rules_status_service_1.rulesStatusBroadcaster.refreshRuleInUse(route.id, rule.id);
                     });
+                    // 服务性能打点：记录首/末 SSE 事件时间
+                    streamTiming = new stream_timing_transform_1.StreamTimingTransform(startTime);
                     responseHeadersForLog = this.normalizeResponseHeaders(responseHeaders);
                     // 流式 model 回写：将上游返回的 model 改写为客户端请求时的原始模型名
                     const originalModel = (_d = req.body) === null || _d === void 0 ? void 0 : _d.model;
@@ -4911,7 +4975,7 @@ class ProxyServer {
                                 return stages;
                             };
                             if (converter) {
-                                const stages = buildStages(streamSource, parser, eventCollector, converter);
+                                const stages = buildStages(streamSource, parser, eventCollector, streamTiming, converter);
                                 stream_1.pipeline(...stages, (error) => {
                                     if (error) {
                                         reject(error);
@@ -4921,7 +4985,7 @@ class ProxyServer {
                                 });
                             }
                             else {
-                                const stages = buildStages(streamSource, parser, eventCollector);
+                                const stages = buildStages(streamSource, parser, eventCollector, streamTiming);
                                 stream_1.pipeline(...stages, (error) => {
                                     if (error) {
                                         reject(error);

package/dist/server/rules-status-service.js

@@ -52,8 +52,8 @@ class RulesStatusBroadcaster extends events_1.EventEmitter {
             enumerable: true,
             configurable: true,
             writable: true,
-            value: 10000
-        }); // 10秒无活动后标记为空闲
+            value: 120000
+        }); // 120秒无活动后标记为空闲（兜底安全网，覆盖 thinking hold 等长静默场景）
         Object.defineProperty(this, "IDLE_DEBOUNCE_DELAY", {
             enumerable: true,
             configurable: true,
@@ -258,16 +258,31 @@ class RulesStatusBroadcaster extends events_1.EventEmitter {
         });
     }
     /**
-     * 刷新规则使用中的不活动定时器（轻量级，仅重置定时器，不修改状态）
-     * 用于 streaming 过程中持续保持 in_use 状态
+     * 刷新规则使用中的不活动定时器（轻量级，仅重置定时器，通常不修改状态）
+     * 用于 streaming 过程中持续保持 in_use 状态。
+     *
+     * 行为：
+     * - status === 'in_use'：重置不活动定时器，并清除可能 pending 的 idle debounce，
+     *   避免已触发的 idle 经 SSE 推送出去（thinking hold 场景的关键修复）。
+     * - status === 'idle'：说明此前已被错误判空闲，但请求仍在出流——重新标记为 in_use
+     *   以便经 SSE 把状态推回"使用中"，实现前端自愈。
+     * - status === 'error' / 'suspended'：早退，这两种终态有独立恢复机制，不应被流式刷新覆盖。
      */
     refreshRuleInUse(routeId, ruleId) {
         const currentStatus = this.ruleStates.get(ruleId);
-        // 仅当状态已经是 in_use 时才刷新定时器
-        if ((currentStatus === null || currentStatus === void 0 ? void 0 : currentStatus.status) !== 'in_use')
+        // 终态有独立恢复机制，刷新不应覆盖
+        if ((currentStatus === null || currentStatus === void 0 ? void 0 : currentStatus.status) === 'error' || (currentStatus === null || currentStatus === void 0 ? void 0 : currentStatus.status) === 'suspended') {
             return;
+        }
         const timeoutKey = `${routeId}:${ruleId}`;
+        // 已被错误判空闲：重新标记为 in_use（内部会清旧定时器/debounce 并 emit statusChanged → SSE 推回使用中）
+        if ((currentStatus === null || currentStatus === void 0 ? void 0 : currentStatus.status) === 'idle') {
+            this.markRuleInUse(routeId, ruleId);
+            return;
+        }
+        // in_use：重置不活动定时器，并清除 pending 的 idle debounce（阻止已触发的 idle 经 SSE 推送）
         this.clearRuleTimeout(timeoutKey);
+        this.clearIdleDebounce(timeoutKey);
         const timeout = setTimeout(() => {
             this.markRuleIdle(routeId, ruleId);
         }, this.INACTIVITY_TIMEOUT);