chore: sync VERSION file with release v1.1.233 [skip ci]

feat: enhance concurrency queue with health check and admin endpoints

.env.example

@@ -61,6 +61,9 @@ PROXY_USE_IPV4=true
 # ⏱️ 请求超时配置
 REQUEST_TIMEOUT=600000  # 请求超时设置（毫秒），默认10分钟
 
+# 🔧 请求体大小配置
+REQUEST_MAX_SIZE_MB=60
+
 # 📈 使用限制
 DEFAULT_TOKEN_LIMIT=1000000
 
@@ -75,6 +78,8 @@ TOKEN_USAGE_RETENTION=2592000000
 HEALTH_CHECK_INTERVAL=60000
 TIMEZONE_OFFSET=8  # UTC偏移小时数，默认+8（中国时区）
 METRICS_WINDOW=5  # 实时指标统计窗口（分钟），可选1-60，默认5分钟
+# 启动时清理残留的并发排队计数器（默认true，多实例部署时建议设为false）
+CLEAR_CONCURRENCY_QUEUES_ON_STARTUP=true
 
 # 🎨 Web 界面配置
 WEB_TITLE=Claude Relay Service

CLAUDE.md

@@ -22,6 +22,7 @@ Claude Relay Service 是一个多平台 AI API 中转服务，支持 **Claude (
 - **权限控制**: API Key支持权限配置（all/claude/gemini/openai等），控制可访问的服务类型
 - **客户端限制**: 基于User-Agent的客户端识别和限制，支持ClaudeCode、Gemini-CLI等预定义客户端
 - **模型黑名单**: 支持API Key级别的模型访问限制
+- **并发请求排队**: 当API Key并发数超限时，请求进入队列等待而非立即返回429，支持配置最大排队数、超时时间，适用于Claude Code Agent并行工具调用场景
 
 ### 主要服务组件
 
@@ -60,6 +61,7 @@ Claude Relay Service 是一个多平台 AI API 中转服务，支持 **Claude (
 
 - **apiKeyService.js**: API Key管理，验证、限流、使用统计、成本计算
 - **userService.js**: 用户管理系统，支持用户注册、登录、API Key管理
+- **userMessageQueueService.js**: 用户消息串行队列，防止同账户并发用户消息触发限流
 - **pricingService.js**: 定价服务，模型价格管理和成本计算
 - **costInitService.js**: 成本数据初始化服务
 - **webhookService.js**: Webhook通知服务
@@ -185,12 +187,17 @@ npm run service:stop          # 停止服务
 - `CLAUDE_OVERLOAD_HANDLING_MINUTES`: Claude 529错误处理持续时间（分钟，0表示禁用）
 - `STICKY_SESSION_TTL_HOURS`: 粘性会话TTL（小时，默认1）
 - `STICKY_SESSION_RENEWAL_THRESHOLD_MINUTES`: 粘性会话续期阈值（分钟，默认0）
+- `USER_MESSAGE_QUEUE_ENABLED`: 启用用户消息串行队列（默认false）
+- `USER_MESSAGE_QUEUE_DELAY_MS`: 用户消息请求间隔（毫秒，默认200）
+- `USER_MESSAGE_QUEUE_TIMEOUT_MS`: 队列等待超时（毫秒，默认5000，锁持有时间短无需长等待）
+- `USER_MESSAGE_QUEUE_LOCK_TTL_MS`: 锁TTL（毫秒，默认5000，请求发送后立即释放无需长TTL）
 - `METRICS_WINDOW`: 实时指标统计窗口（分钟，1-60，默认5）
 - `MAX_API_KEYS_PER_USER`: 每用户最大API Key数量（默认1）
 - `ALLOW_USER_DELETE_API_KEYS`: 允许用户删除自己的API Keys（默认false）
 - `DEBUG_HTTP_TRAFFIC`: 启用HTTP请求/响应调试日志（默认false，仅开发环境）
 - `PROXY_USE_IPV4`: 代理使用IPv4（默认true）
 - `REQUEST_TIMEOUT`: 请求超时时间（毫秒，默认600000即10分钟）
+- `CLEAR_CONCURRENCY_QUEUES_ON_STARTUP`: 启动时清理残留的并发排队计数器（默认true，多实例部署时建议设为false）
 
 #### AWS Bedrock配置（可选）
 - `CLAUDE_CODE_USE_BEDROCK`: 启用Bedrock（设置为1启用）
@@ -337,6 +344,35 @@ npm run setup  # 自动生成密钥并创建管理员账户
 11. **速率限制未清理**: rateLimitCleanupService每5分钟自动清理过期限流状态
 12. **成本统计不准确**: 运行 `npm run init:costs` 初始化成本数据，检查pricingService是否正确加载模型价格
 13. **缓存命中率低**: 查看缓存监控统计，调整LRU缓存大小配置
+14. **用户消息队列超时**: 优化后锁持有时间已从分钟级降到毫秒级（请求发送后立即释放），默认 `USER_MESSAGE_QUEUE_TIMEOUT_MS=5000` 已足够。如仍有超时，检查网络延迟或禁用此功能（`USER_MESSAGE_QUEUE_ENABLED=false`）
+15. **并发请求排队问题**:
+   - 排队超时：检查 `concurrentRequestQueueTimeoutMs` 配置是否合理（默认10秒）
+   - 排队数过多：调整 `concurrentRequestQueueMaxSize` 和 `concurrentRequestQueueMaxSizeMultiplier`
+   - 查看排队统计：访问 `/admin/concurrency-queue/stats` 接口查看 entered/success/timeout/cancelled/socket_changed/rejected_overload 统计
+   - 排队计数泄漏：系统重启时自动清理，或访问 `/admin/concurrency-queue` DELETE 接口手动清理
+   - Socket 身份验证失败：查看 `socket_changed` 统计，如果频繁发生，检查代理配置或客户端连接稳定性
+   - 健康检查拒绝：查看 `rejected_overload` 统计，表示队列过载时的快速失败次数
+
+### 代理配置要求（并发请求排队）
+
+使用并发请求排队功能时，需要正确配置代理（如 Nginx）的超时参数：
+
+- **推荐配置**: `proxy_read_timeout >= max(2 × concurrentRequestQueueTimeoutMs, 60s)`
+  - 当前默认排队超时 10 秒，Nginx 默认 `proxy_read_timeout = 60s` 已满足要求
+  - 如果调整排队超时到 60 秒，推荐代理超时 ≥ 120 秒
+- **Nginx 配置示例**:
+  ```nginx
+  location /api/ {
+      proxy_read_timeout 120s;  # 排队超时 60s 时推荐 120s
+      proxy_connect_timeout 10s;
+      # ...其他配置
+  }
+  ```
+- **企业防火墙环境**:
+  - 某些企业防火墙可能静默关闭长时间无数据的连接（20-40 秒）
+  - 如遇此问题，联系网络管理员调整空闲连接超时策略
+  - 或降低 `concurrentRequestQueueTimeoutMs` 配置
+- **后续升级说明**: 如有需要，后续版本可能提供可选的轻量级心跳机制
 
 ### 调试工具
 
@@ -449,6 +485,15 @@ npm run setup  # 自动生成密钥并创建管理员账户
 - **缓存优化**: 多层LRU缓存（解密缓存、账户缓存），全局缓存监控和统计
 - **成本追踪**: 实时token使用统计（input/output/cache_create/cache_read）和成本计算（基于pricingService）
 - **并发控制**: Redis Sorted Set实现的并发计数，支持自动过期清理
+- **并发请求排队**: 当API Key并发超限时，请求进入队列等待而非直接返回429
+  - **工作原理**: 采用「先占后检查」模式，每次轮询尝试占位，超限则释放继续等待
+  - **指数退避**: 初始200ms，指数增长至最大2秒，带±20%抖动防惊群效应
+  - **智能清理**: 排队计数有TTL保护（超时+30秒），进程崩溃也能自动清理
+  - **Socket身份验证**: 使用UUID token + socket对象引用双重验证，避免HTTP Keep-Alive连接复用导致的身份混淆
+  - **健康检查**: P90等待时间超过阈值时快速失败（返回429），避免新请求在过载时继续排队
+  - **配置参数**: `concurrentRequestQueueEnabled`（默认false）、`concurrentRequestQueueMaxSize`（默认3）、`concurrentRequestQueueMaxSizeMultiplier`（默认0）、`concurrentRequestQueueTimeoutMs`（默认10秒）、`concurrentRequestQueueMaxRedisFailCount`（默认5）、`concurrentRequestQueueHealthCheckEnabled`（默认true）、`concurrentRequestQueueHealthThreshold`（默认0.8）
+  - **最大排队数**: max(固定值, 并发限制×倍数)，例如并发限制=10、倍数=2时最大排队数=20
+  - **适用场景**: Claude Code Agent并行工具调用、批量请求处理
 - **客户端识别**: 基于User-Agent的客户端限制，支持预定义客户端（ClaudeCode、Gemini-CLI等）
 - **错误处理**: 529错误自动标记账户过载状态，配置时长内自动排除该账户
 
@@ -508,8 +553,16 @@ npm run setup  # 自动生成密钥并创建管理员账户
   - `overload:{accountId}` - 账户过载状态（529错误）
 - **并发控制**:
   - `concurrency:{accountId}` - Redis Sorted Set实现的并发计数
+- **并发请求排队**:
+  - `concurrency:queue:{apiKeyId}` - API Key级别的排队计数器（TTL由 `concurrentRequestQueueTimeoutMs` + 30秒缓冲决定）
+  - `concurrency:queue:stats:{apiKeyId}` - 排队统计（entered/success/timeout/cancelled）
+  - `concurrency:queue:wait_times:{apiKeyId}` - 按API Key的等待时间记录（用于P50/P90/P99计算）
+  - `concurrency:queue:wait_times:global` - 全局等待时间记录
 - **Webhook配置**:
   - `webhook_config:{id}` - Webhook配置
+- **用户消息队列**:
+  - `user_msg_queue_lock:{accountId}` - 用户消息队列锁（当前持有者requestId）
+  - `user_msg_queue_last:{accountId}` - 上次请求完成时间戳（用于延迟计算）
 - **系统信息**:
   - `system_info` - 系统状态缓存
   - `model_pricing` - 模型价格数据（pricingService）

VERSION

@@ -1 +1 @@
-1.1.228
+1.1.233

config/config.example.js

@@ -203,6 +203,15 @@ const config = {
   development: {
     debug: process.env.DEBUG === 'true',
     hotReload: process.env.HOT_RELOAD === 'true'
+  },
+
+  // 📬 用户消息队列配置
+  // 优化说明：锁在请求发送成功后立即释放（而非请求完成后），因为 Claude API 限流基于请求发送时刻计算
+  userMessageQueue: {
+    enabled: process.env.USER_MESSAGE_QUEUE_ENABLED === 'true', // 默认关闭
+    delayMs: parseInt(process.env.USER_MESSAGE_QUEUE_DELAY_MS) || 200, // 请求间隔（毫秒）
+    timeoutMs: parseInt(process.env.USER_MESSAGE_QUEUE_TIMEOUT_MS) || 5000, // 队列等待超时（毫秒），锁持有时间短，无需长等待
+    lockTtlMs: parseInt(process.env.USER_MESSAGE_QUEUE_LOCK_TTL_MS) || 5000 // 锁TTL（毫秒），5秒足以覆盖请求发送
   }
 }
 

docker-compose.yml

@@ -21,6 +21,9 @@ services:
       - PORT=3000
       - HOST=0.0.0.0
       
+      # 🔧 请求体大小配置
+      - REQUEST_MAX_SIZE_MB=60
+      
       # 🔐 安全配置（必填）
       - JWT_SECRET=${JWT_SECRET}  # 必填：至少32字符的随机字符串
       - ENCRYPTION_KEY=${ENCRYPTION_KEY}  # 必填：32字符的加密密钥

src/app.js

@@ -584,6 +584,20 @@ class Application {
 
         // 使用 Lua 脚本批量清理所有过期项
         for (const key of keys) {
+          // 跳过非 Sorted Set 类型的键（这些键有各自的清理逻辑）
+          // - concurrency:queue:stats:* 是 Hash 类型
+          // - concurrency:queue:wait_times:* 是 List 类型
+          // - concurrency:queue:* (不含stats/wait_times) 是 String 类型
+          if (
+            key.startsWith('concurrency:queue:stats:') ||
+            key.startsWith('concurrency:queue:wait_times:') ||
+            (key.startsWith('concurrency:queue:') &&
+              !key.includes(':stats:') &&
+              !key.includes(':wait_times:'))
+          ) {
+            continue
+          }
+
           try {
             const cleaned = await redis.client.eval(
               `
@@ -625,6 +639,28 @@ class Application {
     }, 60000) // 每分钟执行一次
 
     logger.info('🔢 Concurrency cleanup task started (running every 1 minute)')
+
+    // 📬 启动用户消息队列服务
+    const userMessageQueueService = require('./services/userMessageQueueService')
+    // 先清理服务重启后残留的锁，防止旧锁阻塞新请求
+    userMessageQueueService.cleanupStaleLocks().then(() => {
+      // 然后启动定时清理任务
+      userMessageQueueService.startCleanupTask()
+    })
+
+    // 🚦 清理服务重启后残留的并发排队计数器
+    // 多实例部署时建议关闭此开关，避免新实例启动时清空其他实例的队列计数
+    // 可通过 DELETE /admin/concurrency/queue 接口手动清理
+    const clearQueuesOnStartup = process.env.CLEAR_CONCURRENCY_QUEUES_ON_STARTUP !== 'false'
+    if (clearQueuesOnStartup) {
+      redis.clearAllConcurrencyQueues().catch((error) => {
+        logger.error('❌ Error clearing concurrency queues on startup:', error)
+      })
+    } else {
+      logger.info(
+        '🚦 Skipping concurrency queue cleanup on startup (CLEAR_CONCURRENCY_QUEUES_ON_STARTUP=false)'
+      )
+    }
   }
 
   setupGracefulShutdown() {
@@ -661,6 +697,15 @@ class Application {
             logger.error('❌ Error stopping rate limit cleanup service:', error)
           }
 
+          // 停止用户消息队列清理服务
+          try {
+            const userMessageQueueService = require('./services/userMessageQueueService')
+            userMessageQueueService.stopCleanupTask()
+            logger.info('📬 User message queue service stopped')
+          } catch (error) {
+            logger.error('❌ Error stopping user message queue service:', error)
+          }
+
           // 停止费用排序索引服务
           try {
             const costRankService = require('./services/costRankService')

src/middleware/auth.js

@@ -8,6 +8,102 @@ const redis = require('../models/redis')
 const ClientValidator = require('../validators/clientValidator')
 const ClaudeCodeValidator = require('../validators/clients/claudeCodeValidator')
 const claudeRelayConfigService = require('../services/claudeRelayConfigService')
+const { calculateWaitTimeStats } = require('../utils/statsHelper')
+
+// 工具函数
+function sleep(ms) {
+  return new Promise((resolve) => setTimeout(resolve, ms))
+}
+
+/**
+ * 检查排队是否过载，决定是否应该快速失败
+ * 详见 design.md Decision 7: 排队健康检查与快速失败
+ *
+ * @param {string} apiKeyId - API Key ID
+ * @param {number} timeoutMs - 排队超时时间（毫秒）
+ * @param {Object} queueConfig - 队列配置
+ * @param {number} maxQueueSize - 最大排队数
+ * @returns {Promise<Object>} { reject: boolean, reason?: string, estimatedWaitMs?: number, timeoutMs?: number }
+ */
+async function shouldRejectDueToOverload(apiKeyId, timeoutMs, queueConfig, maxQueueSize) {
+  try {
+    // 如果健康检查被禁用，直接返回不拒绝
+    if (!queueConfig.concurrentRequestQueueHealthCheckEnabled) {
+      return { reject: false, reason: 'health_check_disabled' }
+    }
+
+    // 🔑 先检查当前队列长度
+    const currentQueueCount = await redis.getConcurrencyQueueCount(apiKeyId).catch(() => 0)
+
+    // 队列为空，说明系统已恢复，跳过健康检查
+    if (currentQueueCount === 0) {
+      return { reject: false, reason: 'queue_empty', currentQueueCount: 0 }
+    }
+
+    // 🔑 关键改进：只有当队列接近满载时才进行健康检查
+    // 队列长度 <= maxQueueSize * 0.5 时，认为系统有足够余量，跳过健康检查
+    // 这避免了在队列较短时过于保守地拒绝请求
+    // 使用 ceil 确保小队列（如 maxQueueSize=3）时阈值为 2，即队列 <=1 时跳过
+    const queueLoadThreshold = Math.ceil(maxQueueSize * 0.5)
+    if (currentQueueCount <= queueLoadThreshold) {
+      return {
+        reject: false,
+        reason: 'queue_not_loaded',
+        currentQueueCount,
+        queueLoadThreshold,
+        maxQueueSize
+      }
+    }
+
+    // 获取该 API Key 的等待时间样本
+    const waitTimes = await redis.getQueueWaitTimes(apiKeyId)
+    const stats = calculateWaitTimeStats(waitTimes)
+
+    // 样本不足（< 10），跳过健康检查，避免冷启动误判
+    if (!stats || stats.sampleCount < 10) {
+      return { reject: false, reason: 'insufficient_samples', sampleCount: stats?.sampleCount || 0 }
+    }
+
+    // P90 不可靠时也跳过（虽然 sampleCount >= 10 时 p90Unreliable 应该是 false）
+    if (stats.p90Unreliable) {
+      return { reject: false, reason: 'p90_unreliable', sampleCount: stats.sampleCount }
+    }
+
+    // 计算健康阈值：P90 >= 超时时间 × 阈值 时拒绝
+    const threshold = queueConfig.concurrentRequestQueueHealthThreshold || 0.8
+    const maxAllowedP90 = timeoutMs * threshold
+
+    if (stats.p90 >= maxAllowedP90) {
+      return {
+        reject: true,
+        reason: 'queue_overloaded',
+        estimatedWaitMs: stats.p90,
+        timeoutMs,
+        threshold,
+        sampleCount: stats.sampleCount,
+        currentQueueCount,
+        maxQueueSize
+      }
+    }
+
+    return { reject: false, p90: stats.p90, sampleCount: stats.sampleCount, currentQueueCount }
+  } catch (error) {
+    // 健康检查出错时不阻塞请求，记录警告并继续
+    logger.warn(`Health check failed for ${apiKeyId}:`, error.message)
+    return { reject: false, reason: 'health_check_error', error: error.message }
+  }
+}
+
+// 排队轮询配置常量（可通过配置文件覆盖）
+// 性能权衡：初始间隔越短响应越快，但 Redis QPS 越高
+// 当前配置：100 个等待者时约 250-300 QPS（指数退避后）
+const QUEUE_POLLING_CONFIG = {
+  pollIntervalMs: 200, // 初始轮询间隔（毫秒）- 平衡响应速度和 Redis 压力
+  maxPollIntervalMs: 2000, // 最大轮询间隔（毫秒）- 长时间等待时降低 Redis 压力
+  backoffFactor: 1.5, // 指数退避系数
+  jitterRatio: 0.2, // 抖动比例（±20%）- 防止惊群效应
+  maxRedisFailCount: 5 // 连续 Redis 失败阈值（从 3 提高到 5，提高网络抖动容忍度）
+}
 
 const FALLBACK_CONCURRENCY_CONFIG = {
   leaseSeconds: 300,
@@ -128,9 +224,223 @@ function isTokenCountRequest(req) {
   return false
 }
 
+/**
+ * 等待并发槽位（排队机制核心）
+ *
+ * 采用「先占后检查」模式避免竞态条件：
+ * - 每次轮询时尝试 incrConcurrency 占位
+ * - 如果超限则 decrConcurrency 释放并继续等待
+ * - 成功获取槽位后返回，调用方无需再次 incrConcurrency
+ *
+ * ⚠️ 重要清理责任说明：
+ * - 排队计数：此函数的 finally 块负责调用 decrConcurrencyQueue 清理
+ * - 并发槽位：当返回 acquired=true 时，槽位已被占用（通过 incrConcurrency）
+ *   调用方必须在请求结束时调用 decrConcurrency 释放槽位
+ *   （已在 authenticateApiKey 的 finally 块中处理）
+ *
+ * @param {Object} req - Express 请求对象
+ * @param {Object} res - Express 响应对象
+ * @param {string} apiKeyId - API Key ID
+ * @param {Object} queueOptions - 配置参数
+ * @returns {Promise<Object>} { acquired: boolean, reason?: string, waitTimeMs: number }
+ */
+async function waitForConcurrencySlot(req, res, apiKeyId, queueOptions) {
+  const {
+    concurrencyLimit,
+    requestId,
+    leaseSeconds,
+    timeoutMs,
+    pollIntervalMs,
+    maxPollIntervalMs,
+    backoffFactor,
+    jitterRatio,
+    maxRedisFailCount: configMaxRedisFailCount
+  } = queueOptions
+
+  let clientDisconnected = false
+  // 追踪轮询过程中是否临时占用了槽位（用于异常时清理）
+  // 工作流程：
+  // 1. incrConcurrency 成功且 count <= limit 时，设置 internalSlotAcquired = true
+  // 2. 统计记录完成后，设置 internalSlotAcquired = false 并返回（所有权转移给调用方）
+  // 3. 如果在步骤 1-2 之间发生异常，finally 块会检测到 internalSlotAcquired = true 并释放槽位
+  let internalSlotAcquired = false
+
+  // 监听客户端断开事件
+  // ⚠️ 重要：必须监听 socket 的事件，而不是 req 的事件！
+  // 原因：对于 POST 请求，当 body-parser 读取完请求体后，req（IncomingMessage 可读流）
+  // 的 'close' 事件会立即触发，但这不代表客户端断开连接！客户端仍在等待响应。
+  // socket 的 'close' 事件才是真正的连接关闭信号。
+  const { socket } = req
+  const onSocketClose = () => {
+    clientDisconnected = true
+    logger.debug(
+      `🔌 [Queue] Socket closed during queue wait for API key ${apiKeyId}, requestId: ${requestId}`
+    )
+  }
+
+  if (socket) {
+    socket.once('close', onSocketClose)
+  }
+
+  // 检查 socket 是否在监听器注册前已被销毁（边界情况）
+  if (socket?.destroyed) {
+    clientDisconnected = true
+  }
+
+  const startTime = Date.now()
+  let pollInterval = pollIntervalMs
+  let redisFailCount = 0
+  // 优先使用配置中的值，否则使用默认值
+  const maxRedisFailCount = configMaxRedisFailCount || QUEUE_POLLING_CONFIG.maxRedisFailCount
+
+  try {
+    while (Date.now() - startTime < timeoutMs) {
+      // 检测客户端是否断开（双重检查：事件标记 + socket 状态）
+      // socket.destroyed 是同步检查，确保即使事件处理有延迟也能及时检测
+      if (clientDisconnected || socket?.destroyed) {
+        redis
+          .incrConcurrencyQueueStats(apiKeyId, 'cancelled')
+          .catch((e) => logger.warn('Failed to record cancelled stat:', e))
+        return {
+          acquired: false,
+          reason: 'client_disconnected',
+          waitTimeMs: Date.now() - startTime
+        }
+      }
+
+      // 尝试获取槽位（先占后检查）
+      try {
+        const count = await redis.incrConcurrency(apiKeyId, requestId, leaseSeconds)
+        redisFailCount = 0 // 重置失败计数
+
+        if (count <= concurrencyLimit) {
+          // 成功获取槽位！
+          const waitTimeMs = Date.now() - startTime
+
+          // 槽位所有权转移说明：
+          // 1. 此时槽位已通过 incrConcurrency 获取
+          // 2. 先标记 internalSlotAcquired = true，确保异常时 finally 块能清理
+          // 3. 统计操作完成后，清除标记并返回，所有权转移给调用方
+          // 4. 调用方（authenticateApiKey）负责在请求结束时释放槽位
+
+          // 标记槽位已获取（用于异常时 finally 块清理）
+          internalSlotAcquired = true
+
+          // 记录统计（非阻塞，fire-and-forget 模式）
+          // ⚠️ 设计说明：
+          // - 故意不 await 这些 Promise，因为统计记录不应阻塞请求处理
+          // - 每个 Promise 都有独立的 .catch()，确保单个失败不影响其他
+          // - 外层 .catch() 是防御性措施，处理 Promise.all 本身的异常
+          // - 即使统计记录在函数返回后才完成/失败，也是安全的（仅日志记录）
+          // - 统计数据丢失可接受，不影响核心业务逻辑
+          Promise.all([
+            redis
+              .recordQueueWaitTime(apiKeyId, waitTimeMs)
+              .catch((e) => logger.warn('Failed to record queue wait time:', e)),
+            redis
+              .recordGlobalQueueWaitTime(waitTimeMs)
+              .catch((e) => logger.warn('Failed to record global wait time:', e)),
+            redis
+              .incrConcurrencyQueueStats(apiKeyId, 'success')
+              .catch((e) => logger.warn('Failed to increment success stats:', e))
+          ]).catch((e) => logger.warn('Failed to record queue stats batch:', e))
+
+          // 成功返回前清除标记（所有权转移给调用方，由其负责释放）
+          internalSlotAcquired = false
+          return { acquired: true, waitTimeMs }
+        }
+
+        // 超限，释放槽位继续等待
+        try {
+          await redis.decrConcurrency(apiKeyId, requestId)
+        } catch (decrError) {
+          // 释放失败时记录警告但继续轮询
+          // 下次 incrConcurrency 会自然覆盖同一 requestId 的条目
+          logger.warn(
+            `Failed to release slot during polling for ${apiKeyId}, will retry:`,
+            decrError
+          )
+        }
+      } catch (redisError) {
+        redisFailCount++
+        logger.error(
+          `Redis error in queue polling (${redisFailCount}/${maxRedisFailCount}):`,
+          redisError
+        )
+
+        if (redisFailCount >= maxRedisFailCount) {
+          // 连续 Redis 失败，放弃排队
+          return {
+            acquired: false,
+            reason: 'redis_error',
+            waitTimeMs: Date.now() - startTime
+          }
+        }
+      }
+
+      // 指数退避等待
+      await sleep(pollInterval)
+
+      // 计算下一次轮询间隔（指数退避 + 抖动）
+      // 1. 先应用指数退避
+      let nextInterval = pollInterval * backoffFactor
+      // 2. 添加抖动防止惊群效应（±jitterRatio 范围内的随机偏移）
+      //    抖动范围：[-jitterRatio, +jitterRatio]，例如 jitterRatio=0.2 时为 ±20%
+      //    这是预期行为：负抖动可使间隔略微缩短，正抖动可使间隔略微延长
+      //    目的是分散多个等待者的轮询时间点，避免同时请求 Redis
+      const jitter = nextInterval * jitterRatio * (Math.random() * 2 - 1)
+      nextInterval = nextInterval + jitter
+      // 3. 确保在合理范围内：最小 1ms，最大 maxPollIntervalMs
+      //    Math.max(1, ...) 保证即使负抖动也不会产生 ≤0 的间隔
+      pollInterval = Math.max(1, Math.min(nextInterval, maxPollIntervalMs))
+    }
+
+    // 超时
+    redis
+      .incrConcurrencyQueueStats(apiKeyId, 'timeout')
+      .catch((e) => logger.warn('Failed to record timeout stat:', e))
+    return { acquired: false, reason: 'timeout', waitTimeMs: Date.now() - startTime }
+  } finally {
+    // 确保清理：
+    // 1. 减少排队计数（排队计数在调用方已增加，这里负责减少）
+    try {
+      await redis.decrConcurrencyQueue(apiKeyId)
+    } catch (cleanupError) {
+      // 清理失败记录错误（可能导致计数泄漏，但有 TTL 保护）
+      logger.error(
+        `Failed to decrement queue count in finally block for ${apiKeyId}:`,
+        cleanupError
+      )
+    }
+
+    // 2. 如果内部获取了槽位但未正常返回（异常路径），释放槽位
+    if (internalSlotAcquired) {
+      try {
+        await redis.decrConcurrency(apiKeyId, requestId)
+        logger.warn(
+          `⚠️ Released orphaned concurrency slot in finally block for ${apiKeyId}, requestId: ${requestId}`
+        )
+      } catch (slotCleanupError) {
+        logger.error(
+          `Failed to release orphaned concurrency slot for ${apiKeyId}:`,
+          slotCleanupError
+        )
+      }
+    }
+
+    // 清理 socket 事件监听器
+    if (socket) {
+      socket.removeListener('close', onSocketClose)
+    }
+  }
+}
+
 // 🔑 API Key验证中间件（优化版）
 const authenticateApiKey = async (req, res, next) => {
   const startTime = Date.now()
+  let authErrored = false
+  let concurrencyCleanup = null
+  let hasConcurrencySlot = false
 
   try {
     // 安全提取API Key，支持多种格式（包括Gemini CLI支持）
@@ -265,33 +575,86 @@ const authenticateApiKey = async (req, res, next) => {
       }
       const requestId = uuidv4()
 
+      // ⚠️ 优化后的 Connection: close 设置策略
+      // 问题背景：HTTP Keep-Alive 使多个请求共用同一个 TCP 连接
+      // 当第一个请求正在处理，第二个请求进入排队时，它们共用同一个 socket
+      // 如果客户端超时关闭连接，两个请求都会受影响
+      // 优化方案：只有在请求实际进入排队时才设置 Connection: close
+      // 未排队的请求保持 Keep-Alive，避免不必要的 TCP 握手开销
+      // 详见 design.md Decision 2: Connection: close 设置时机
+      // 注意：Connection: close 将在下方代码实际进入排队时设置（第 637 行左右）
+
+      // ============================================================
+      // 🔒 并发槽位状态管理说明
+      // ============================================================
+      // 此函数中有两个关键状态变量：
+      // - hasConcurrencySlot: 当前是否持有并发槽位
+      // - concurrencyCleanup: 错误时调用的清理函数
+      //
+      // 状态转换流程：
+      // 1. incrConcurrency 成功 → hasConcurrencySlot=true, 设置临时清理函数
+      // 2. 若超限 → 释放槽位，hasConcurrencySlot=false, concurrencyCleanup=null
+      // 3. 若排队成功 → hasConcurrencySlot=true, 升级为完整清理函数（含 interval 清理）
+      // 4. 请求结束（res.close/req.close）→ 调用 decrementConcurrency 释放
+      // 5. 认证错误 → finally 块调用 concurrencyCleanup 释放
+      //
+      // 为什么需要两种清理函数？
+      // - 临时清理：在排队/认证过程中出错时使用，只释放槽位
+      // - 完整清理：请求正常开始后使用，还需清理 leaseRenewInterval
+      // ============================================================
+      const setTemporaryConcurrencyCleanup = () => {
+        concurrencyCleanup = async () => {
+          if (!hasConcurrencySlot) {
+            return
+          }
+          hasConcurrencySlot = false
+          try {
+            await redis.decrConcurrency(validation.keyData.id, requestId)
+          } catch (cleanupError) {
+            logger.error(
+              `Failed to decrement concurrency after auth error for key ${validation.keyData.id}:`,
+              cleanupError
+            )
+          }
+        }
+      }
+
       const currentConcurrency = await redis.incrConcurrency(
         validation.keyData.id,
         requestId,
         leaseSeconds
       )
+      hasConcurrencySlot = true
+      setTemporaryConcurrencyCleanup()
       logger.api(
         `📈 Incremented concurrency for key: ${validation.keyData.id} (${validation.keyData.name}), current: ${currentConcurrency}, limit: ${concurrencyLimit}`
       )
 
       if (currentConcurrency > concurrencyLimit) {
-        // 如果超过限制，立即减少计数（添加 try-catch 防止异常导致并发泄漏）
+        // 1. 先释放刚占用的槽位
         try {
-          const newCount = await redis.decrConcurrency(validation.keyData.id, requestId)
+          await redis.decrConcurrency(validation.keyData.id, requestId)
-          logger.api(
-            `📉 Decremented concurrency (429 rejected) for key: ${validation.keyData.id} (${validation.keyData.name}), new count: ${newCount}`
-          )
         } catch (error) {
           logger.error(
             `Failed to decrement concurrency after limit exceeded for key ${validation.keyData.id}:`,
             error
           )
         }
+        hasConcurrencySlot = false
+        concurrencyCleanup = null
+
+        // 2. 获取排队配置
+        const queueConfig = await claudeRelayConfigService.getConfig()
+
+        // 3. 排队功能未启用，直接返回 429（保持现有行为）
+        if (!queueConfig.concurrentRequestQueueEnabled) {
           logger.security(
             `🚦 Concurrency limit exceeded for key: ${validation.keyData.id} (${
               validation.keyData.name
             }), current: ${currentConcurrency - 1}, limit: ${concurrencyLimit}`
           )
+          // 建议客户端在短暂延迟后重试（并发场景下通常很快会有槽位释放）
+          res.set('Retry-After', '1')
           return res.status(429).json({
             error: 'Concurrency limit exceeded',
             message: `Too many concurrent requests. Limit: ${concurrencyLimit} concurrent requests`,
@@ -300,6 +663,260 @@ const authenticateApiKey = async (req, res, next) => {
           })
         }
 
+        // 4. 计算最大排队数
+        const maxQueueSize = Math.max(
+          concurrencyLimit * queueConfig.concurrentRequestQueueMaxSizeMultiplier,
+          queueConfig.concurrentRequestQueueMaxSize
+        )
+
+        // 4.5 排队健康检查：过载时快速失败
+        // 详见 design.md Decision 7: 排队健康检查与快速失败
+        const overloadCheck = await shouldRejectDueToOverload(
+          validation.keyData.id,
+          queueConfig.concurrentRequestQueueTimeoutMs,
+          queueConfig,
+          maxQueueSize
+        )
+        if (overloadCheck.reject) {
+          // 使用健康检查返回的当前排队数，避免重复调用 Redis
+          const currentQueueCount = overloadCheck.currentQueueCount || 0
+          logger.api(
+            `🚨 Queue overloaded for key: ${validation.keyData.id} (${validation.keyData.name}), ` +
+              `P90=${overloadCheck.estimatedWaitMs}ms, timeout=${overloadCheck.timeoutMs}ms, ` +
+              `threshold=${overloadCheck.threshold}, samples=${overloadCheck.sampleCount}, ` +
+              `concurrency=${concurrencyLimit}, queue=${currentQueueCount}/${maxQueueSize}`
+          )
+          // 记录被拒绝的过载统计
+          redis
+            .incrConcurrencyQueueStats(validation.keyData.id, 'rejected_overload')
+            .catch((e) => logger.warn('Failed to record rejected_overload stat:', e))
+          // 返回 429 + Retry-After，让客户端稍后重试
+          const retryAfterSeconds = 30
+          res.set('Retry-After', String(retryAfterSeconds))
+          return res.status(429).json({
+            error: 'Queue overloaded',
+            message: `Queue is overloaded. Estimated wait time (${overloadCheck.estimatedWaitMs}ms) exceeds threshold. Limit: ${concurrencyLimit} concurrent requests, queue: ${currentQueueCount}/${maxQueueSize}. Please retry later.`,
+            currentConcurrency: concurrencyLimit,
+            concurrencyLimit,
+            queueCount: currentQueueCount,
+            maxQueueSize,
+            estimatedWaitMs: overloadCheck.estimatedWaitMs,
+            timeoutMs: overloadCheck.timeoutMs,
+            queueTimeoutMs: queueConfig.concurrentRequestQueueTimeoutMs,
+            retryAfterSeconds
+          })
+        }
+
+        // 5. 尝试进入排队（原子操作：先增加再检查，避免竞态条件）
+        let queueIncremented = false
+        try {
+          const newQueueCount = await redis.incrConcurrencyQueue(
+            validation.keyData.id,
+            queueConfig.concurrentRequestQueueTimeoutMs
+          )
+          queueIncremented = true
+
+          if (newQueueCount > maxQueueSize) {
+            // 超过最大排队数，立即释放并返回 429
+            await redis.decrConcurrencyQueue(validation.keyData.id)
+            queueIncremented = false
+            logger.api(
+              `🚦 Concurrency queue full for key: ${validation.keyData.id} (${validation.keyData.name}), ` +
+                `queue: ${newQueueCount - 1}, maxQueue: ${maxQueueSize}`
+            )
+            // 队列已满，建议客户端在排队超时时间后重试
+            const retryAfterSeconds = Math.ceil(queueConfig.concurrentRequestQueueTimeoutMs / 1000)
+            res.set('Retry-After', String(retryAfterSeconds))
+            return res.status(429).json({
+              error: 'Concurrency queue full',
+              message: `Too many requests waiting in queue. Limit: ${concurrencyLimit} concurrent requests, queue: ${newQueueCount - 1}/${maxQueueSize}, timeout: ${retryAfterSeconds}s`,
+              currentConcurrency: concurrencyLimit,
+              concurrencyLimit,
+              queueCount: newQueueCount - 1,
+              maxQueueSize,
+              queueTimeoutMs: queueConfig.concurrentRequestQueueTimeoutMs,
+              retryAfterSeconds
+            })
+          }
+
+          // 6. 已成功进入排队，记录统计并开始等待槽位
+          logger.api(
+            `⏳ Request entering queue for key: ${validation.keyData.id} (${validation.keyData.name}), ` +
+              `queue position: ${newQueueCount}`
+          )
+          redis
+            .incrConcurrencyQueueStats(validation.keyData.id, 'entered')
+            .catch((e) => logger.warn('Failed to record entered stat:', e))
+
+          // ⚠️ 仅在请求实际进入排队时设置 Connection: close
+          // 详见 design.md Decision 2: Connection: close 设置时机
+          // 未排队的请求保持 Keep-Alive，避免不必要的 TCP 握手开销
+          if (!res.headersSent) {
+            res.setHeader('Connection', 'close')
+            logger.api(
+              `🔌 [Queue] Set Connection: close for queued request, key: ${validation.keyData.id}`
+            )
+          }
+
+          // ⚠️ 记录排队开始时的 socket 标识，用于排队完成后验证
+          // 问题背景：HTTP Keep-Alive 连接复用时，长时间排队可能导致 socket 被其他请求使用
+          // 验证方法：使用 UUID token + socket 对象引用双重验证
+          // 详见 design.md Decision 1: Socket 身份验证机制
+          req._crService = req._crService || {}
+          req._crService.queueToken = uuidv4()
+          req._crService.originalSocket = req.socket
+          req._crService.startTime = Date.now()
+          const savedToken = req._crService.queueToken
+          const savedSocket = req._crService.originalSocket
+
+          // ⚠️ 重要：在调用前将 queueIncremented 设为 false
+          // 因为 waitForConcurrencySlot 的 finally 块会负责清理排队计数
+          // 如果在调用后设置，当 waitForConcurrencySlot 抛出异常时
+          // 外层 catch 块会重复减少计数（finally 已经减过一次）
+          queueIncremented = false
+
+          const slot = await waitForConcurrencySlot(req, res, validation.keyData.id, {
+            concurrencyLimit,
+            requestId,
+            leaseSeconds,
+            timeoutMs: queueConfig.concurrentRequestQueueTimeoutMs,
+            pollIntervalMs: QUEUE_POLLING_CONFIG.pollIntervalMs,
+            maxPollIntervalMs: QUEUE_POLLING_CONFIG.maxPollIntervalMs,
+            backoffFactor: QUEUE_POLLING_CONFIG.backoffFactor,
+            jitterRatio: QUEUE_POLLING_CONFIG.jitterRatio,
+            maxRedisFailCount: queueConfig.concurrentRequestQueueMaxRedisFailCount
+          })
+
+          // 7. 处理排队结果
+          if (!slot.acquired) {
+            if (slot.reason === 'client_disconnected') {
+              // 客户端已断开，不返回响应（连接已关闭）
+              logger.api(
+                `🔌 Client disconnected while queuing for key: ${validation.keyData.id} (${validation.keyData.name})`
+              )
+              return
+            }
+
+            if (slot.reason === 'redis_error') {
+              // Redis 连续失败，返回 503
+              logger.error(
+                `❌ Redis error during queue wait for key: ${validation.keyData.id} (${validation.keyData.name})`
+              )
+              return res.status(503).json({
+                error: 'Service temporarily unavailable',
+                message: 'Failed to acquire concurrency slot due to internal error'
+              })
+            }
+            // 排队超时（使用 api 级别，与其他排队日志保持一致）
+            logger.api(
+              `⏰ Queue timeout for key: ${validation.keyData.id} (${validation.keyData.name}), waited: ${slot.waitTimeMs}ms`
+            )
+            // 已等待超时，建议客户端稍后重试
+            // ⚠️ Retry-After 策略优化：
+            // - 请求已经等了完整的 timeout 时间，说明系统负载较高
+            // - 过早重试（如固定 5 秒）会加剧拥塞，导致更多超时
+            // - 合理策略：使用 timeout 时间的一半作为重试间隔
+            // - 最小值 5 秒，最大值 30 秒，避免极端情况
+            const timeoutSeconds = Math.ceil(queueConfig.concurrentRequestQueueTimeoutMs / 1000)
+            const retryAfterSeconds = Math.max(5, Math.min(30, Math.ceil(timeoutSeconds / 2)))
+            res.set('Retry-After', String(retryAfterSeconds))
+            return res.status(429).json({
+              error: 'Queue timeout',
+              message: `Request timed out waiting for concurrency slot. Limit: ${concurrencyLimit} concurrent requests, maxQueue: ${maxQueueSize}, Queue timeout: ${timeoutSeconds}s, waited: ${slot.waitTimeMs}ms`,
+              currentConcurrency: concurrencyLimit,
+              concurrencyLimit,
+              maxQueueSize,
+              queueTimeoutMs: queueConfig.concurrentRequestQueueTimeoutMs,
+              waitTimeMs: slot.waitTimeMs,
+              retryAfterSeconds
+            })
+          }
+
+          // 8. 排队成功，slot.acquired 表示已在 waitForConcurrencySlot 中获取到槽位
+          logger.api(
+            `✅ Queue wait completed for key: ${validation.keyData.id} (${validation.keyData.name}), ` +
+              `waited: ${slot.waitTimeMs}ms`
+          )
+          hasConcurrencySlot = true
+          setTemporaryConcurrencyCleanup()
+
+          // 9. ⚠️ 关键检查：排队等待结束后，验证客户端是否还在等待响应
+          // 长时间排队后，客户端可能在应用层已放弃（如 Claude Code 的超时机制），
+          // 但 TCP 连接仍然存活。此时继续处理请求是浪费资源。
+          // 注意：如果发送了心跳，headersSent 会是 true，但这是正常的
+          const postQueueSocket = req.socket
+          // 只检查连接是否真正断开（destroyed/writableEnded/socketDestroyed）
+          // headersSent 在心跳场景下是正常的，不应该作为放弃的依据
+          if (res.destroyed || res.writableEnded || postQueueSocket?.destroyed) {
+            logger.warn(
+              `⚠️ Client no longer waiting after queue for key: ${validation.keyData.id} (${validation.keyData.name}), ` +
+                `waited: ${slot.waitTimeMs}ms | destroyed: ${res.destroyed}, ` +
+                `writableEnded: ${res.writableEnded}, socketDestroyed: ${postQueueSocket?.destroyed}`
+            )
+            // 释放刚获取的槽位
+            hasConcurrencySlot = false
+            await redis
+              .decrConcurrency(validation.keyData.id, requestId)
+              .catch((e) => logger.error('Failed to release slot after client abandoned:', e))
+            // 不返回响应（客户端已不在等待）
+            return
+          }
+
+          // 10. ⚠️ 关键检查：验证 socket 身份是否改变
+          // HTTP Keep-Alive 连接复用可能导致排队期间 socket 被其他请求使用
+          // 验证方法：UUID token + socket 对象引用双重验证
+          // 详见 design.md Decision 1: Socket 身份验证机制
+          const queueData = req._crService
+          const socketIdentityChanged =
+            !queueData ||
+            queueData.queueToken !== savedToken ||
+            queueData.originalSocket !== savedSocket
+
+          if (socketIdentityChanged) {
+            logger.error(
+              `❌ [Queue] Socket identity changed during queue wait! ` +
+                `key: ${validation.keyData.id} (${validation.keyData.name}), ` +
+                `waited: ${slot.waitTimeMs}ms | ` +
+                `tokenMatch: ${queueData?.queueToken === savedToken}, ` +
+                `socketMatch: ${queueData?.originalSocket === savedSocket}`
+            )
+            // 释放刚获取的槽位
+            hasConcurrencySlot = false
+            await redis
+              .decrConcurrency(validation.keyData.id, requestId)
+              .catch((e) => logger.error('Failed to release slot after socket identity change:', e))
+            // 记录 socket_changed 统计
+            redis
+              .incrConcurrencyQueueStats(validation.keyData.id, 'socket_changed')
+              .catch((e) => logger.warn('Failed to record socket_changed stat:', e))
+            // 不返回响应（socket 已被其他请求使用）
+            return
+          }
+        } catch (queueError) {
+          // 异常时清理资源，防止泄漏
+          // 1. 清理排队计数（如果还没被 waitForConcurrencySlot 的 finally 清理）
+          if (queueIncremented) {
+            await redis
+              .decrConcurrencyQueue(validation.keyData.id)
+              .catch((e) => logger.error('Failed to cleanup queue count after error:', e))
+          }
+
+          // 2. 防御性清理：如果 waitForConcurrencySlot 内部获取了槽位但在返回前异常
+          //    虽然这种情况极少发生（统计记录的异常会被内部捕获），但为了安全起见
+          //    尝试释放可能已获取的槽位。decrConcurrency 使用 ZREM，即使成员不存在也安全
+          if (hasConcurrencySlot) {
+            hasConcurrencySlot = false
+            await redis
+              .decrConcurrency(validation.keyData.id, requestId)
+              .catch((e) =>
+                logger.error('Failed to cleanup concurrency slot after queue error:', e)
+              )
+          }
+
+          throw queueError
+        }
+      }
+
       const renewIntervalMs =
         renewIntervalSeconds > 0 ? Math.max(renewIntervalSeconds * 1000, 15000) : 0
 
@@ -358,6 +975,7 @@ const authenticateApiKey = async (req, res, next) => {
       const decrementConcurrency = async () => {
         if (!concurrencyDecremented) {
           concurrencyDecremented = true
+          hasConcurrencySlot = false
           if (leaseRenewInterval) {
             clearInterval(leaseRenewInterval)
             leaseRenewInterval = null
@@ -372,6 +990,11 @@ const authenticateApiKey = async (req, res, next) => {
           }
         }
       }
+      // 升级为完整清理函数（包含 leaseRenewInterval 清理逻辑）
+      // 此时请求已通过认证，后续由 res.close/req.close 事件触发清理
+      if (hasConcurrencySlot) {
+        concurrencyCleanup = decrementConcurrency
+      }
 
       // 监听最可靠的事件（避免重复监听）
       // res.on('close') 是最可靠的，会在连接关闭时触发
@@ -697,6 +1320,7 @@ const authenticateApiKey = async (req, res, next) => {
 
     return next()
   } catch (error) {
+    authErrored = true
     const authDuration = Date.now() - startTime
     logger.error(`❌ Authentication middleware error (${authDuration}ms):`, {
       error: error.message,
@@ -710,6 +1334,14 @@ const authenticateApiKey = async (req, res, next) => {
       error: 'Authentication error',
       message: 'Internal server error during authentication'
     })
+  } finally {
+    if (authErrored && typeof concurrencyCleanup === 'function') {
+      try {
+        await concurrencyCleanup()
+      } catch (cleanupError) {
+        logger.error('Failed to cleanup concurrency after auth error:', cleanupError)
+      }
+    }
   }
 }
 
@@ -1388,7 +2020,8 @@ const globalRateLimit = async (req, res, next) =>
 
 // 📊 请求大小限制中间件
 const requestSizeLimit = (req, res, next) => {
-  const maxSize = 60 * 1024 * 1024 // 60MB
+  const MAX_SIZE_MB = parseInt(process.env.REQUEST_MAX_SIZE_MB || '60', 10)
+  const maxSize = MAX_SIZE_MB * 1024 * 1024
   const contentLength = parseInt(req.headers['content-length'] || '0')
 
   if (contentLength > maxSize) {

src/models/redis.js

@@ -50,6 +50,18 @@ function getWeekStringInTimezone(date = new Date()) {
   return `${year}-W${String(weekNumber).padStart(2, '0')}`
 }
 
+// 并发队列相关常量
+const QUEUE_STATS_TTL_SECONDS = 86400 * 7 // 统计计数保留 7 天
+const WAIT_TIME_TTL_SECONDS = 86400 // 等待时间样本保留 1 天（滚动窗口，无需长期保留）
+// 等待时间样本数配置（提高统计置信度）
+// - 每 API Key 从 100 提高到 500：提供更稳定的 P99 估计
+// - 全局从 500 提高到 2000：支持更高精度的 P99.9 分析
+// - 内存开销约 12-20KB（Redis quicklist 每元素 1-10 字节），可接受
+// 详见 design.md Decision 5: 等待时间统计样本数
+const WAIT_TIME_SAMPLES_PER_KEY = 500 // 每个 API Key 保留的等待时间样本数
+const WAIT_TIME_SAMPLES_GLOBAL = 2000 // 全局保留的等待时间样本数
+const QUEUE_TTL_BUFFER_SECONDS = 30 // 排队计数器TTL缓冲时间
+
 class RedisClient {
   constructor() {
     this.client = null
@@ -2556,4 +2568,593 @@ redisClient.getDateStringInTimezone = getDateStringInTimezone
 redisClient.getHourInTimezone = getHourInTimezone
 redisClient.getWeekStringInTimezone = getWeekStringInTimezone
 
+// ============== 用户消息队列相关方法 ==============
+
+/**
+ * 尝试获取用户消息队列锁
+ * 使用 Lua 脚本保证原子性
+ * @param {string} accountId - 账户ID
+ * @param {string} requestId - 请求ID
+ * @param {number} lockTtlMs - 锁 TTL（毫秒）
+ * @param {number} delayMs - 请求间隔（毫秒）
+ * @returns {Promise<{acquired: boolean, waitMs: number}>}
+ *   - acquired: 是否成功获取锁
+ *   - waitMs: 需要等待的毫秒数（-1表示被占用需等待，>=0表示需要延迟的毫秒数）
+ */
+redisClient.acquireUserMessageLock = async function (accountId, requestId, lockTtlMs, delayMs) {
+  const lockKey = `user_msg_queue_lock:${accountId}`
+  const lastTimeKey = `user_msg_queue_last:${accountId}`
+
+  const script = `
+    local lockKey = KEYS[1]
+    local lastTimeKey = KEYS[2]
+    local requestId = ARGV[1]
+    local lockTtl = tonumber(ARGV[2])
+    local delayMs = tonumber(ARGV[3])
+
+    -- 检查锁是否空闲
+    local currentLock = redis.call('GET', lockKey)
+    if currentLock == false then
+      -- 检查是否需要延迟
+      local lastTime = redis.call('GET', lastTimeKey)
+      local now = redis.call('TIME')
+      local nowMs = tonumber(now[1]) * 1000 + math.floor(tonumber(now[2]) / 1000)
+
+      if lastTime then
+        local elapsed = nowMs - tonumber(lastTime)
+        if elapsed < delayMs then
+          -- 需要等待的毫秒数
+          return {0, delayMs - elapsed}
+        end
+      end
+
+      -- 获取锁
+      redis.call('SET', lockKey, requestId, 'PX', lockTtl)
+      return {1, 0}
+    end
+
+    -- 锁被占用，返回等待
+    return {0, -1}
+  `
+
+  try {
+    const result = await this.client.eval(
+      script,
+      2,
+      lockKey,
+      lastTimeKey,
+      requestId,
+      lockTtlMs,
+      delayMs
+    )
+    return {
+      acquired: result[0] === 1,
+      waitMs: result[1]
+    }
+  } catch (error) {
+    logger.error(`Failed to acquire user message lock for account ${accountId}:`, error)
+    // 返回 redisError 标记，让上层能区分 Redis 故障和正常锁占用
+    return { acquired: false, waitMs: -1, redisError: true, errorMessage: error.message }
+  }
+}
+
+/**
+ * 释放用户消息队列锁并记录完成时间
+ * @param {string} accountId - 账户ID
+ * @param {string} requestId - 请求ID
+ * @returns {Promise<boolean>} 是否成功释放
+ */
+redisClient.releaseUserMessageLock = async function (accountId, requestId) {
+  const lockKey = `user_msg_queue_lock:${accountId}`
+  const lastTimeKey = `user_msg_queue_last:${accountId}`
+
+  const script = `
+    local lockKey = KEYS[1]
+    local lastTimeKey = KEYS[2]
+    local requestId = ARGV[1]
+
+    -- 验证锁持有者
+    local currentLock = redis.call('GET', lockKey)
+    if currentLock == requestId then
+      -- 记录完成时间
+      local now = redis.call('TIME')
+      local nowMs = tonumber(now[1]) * 1000 + math.floor(tonumber(now[2]) / 1000)
+      redis.call('SET', lastTimeKey, nowMs, 'EX', 60)  -- 60秒后过期
+
+      -- 删除锁
+      redis.call('DEL', lockKey)
+      return 1
+    end
+    return 0
+  `
+
+  try {
+    const result = await this.client.eval(script, 2, lockKey, lastTimeKey, requestId)
+    return result === 1
+  } catch (error) {
+    logger.error(`Failed to release user message lock for account ${accountId}:`, error)
+    return false
+  }
+}
+
+/**
+ * 强制释放用户消息队列锁（用于清理孤儿锁）
+ * @param {string} accountId - 账户ID
+ * @returns {Promise<boolean>} 是否成功释放
+ */
+redisClient.forceReleaseUserMessageLock = async function (accountId) {
+  const lockKey = `user_msg_queue_lock:${accountId}`
+
+  try {
+    await this.client.del(lockKey)
+    return true
+  } catch (error) {
+    logger.error(`Failed to force release user message lock for account ${accountId}:`, error)
+    return false
+  }
+}
+
+/**
+ * 获取用户消息队列统计信息（用于调试）
+ * @param {string} accountId - 账户ID
+ * @returns {Promise<Object>} 队列统计
+ */
+redisClient.getUserMessageQueueStats = async function (accountId) {
+  const lockKey = `user_msg_queue_lock:${accountId}`
+  const lastTimeKey = `user_msg_queue_last:${accountId}`
+
+  try {
+    const [lockHolder, lastTime, lockTtl] = await Promise.all([
+      this.client.get(lockKey),
+      this.client.get(lastTimeKey),
+      this.client.pttl(lockKey)
+    ])
+
+    return {
+      accountId,
+      isLocked: !!lockHolder,
+      lockHolder,
+      lockTtlMs: lockTtl > 0 ? lockTtl : 0,
+      lockTtlRaw: lockTtl, // 原始 PTTL 值：>0 有TTL，-1 无过期时间，-2 键不存在
+      lastCompletedAt: lastTime ? new Date(parseInt(lastTime)).toISOString() : null
+    }
+  } catch (error) {
+    logger.error(`Failed to get user message queue stats for account ${accountId}:`, error)
+    return {
+      accountId,
+      isLocked: false,
+      lockHolder: null,
+      lockTtlMs: 0,
+      lockTtlRaw: -2,
+      lastCompletedAt: null
+    }
+  }
+}
+
+/**
+ * 扫描所有用户消息队列锁（用于清理任务）
+ * @returns {Promise<string[]>} 账户ID列表
+ */
+redisClient.scanUserMessageQueueLocks = async function () {
+  const accountIds = []
+  let cursor = '0'
+  let iterations = 0
+  const MAX_ITERATIONS = 1000 // 防止无限循环
+
+  try {
+    do {
+      const [newCursor, keys] = await this.client.scan(
+        cursor,
+        'MATCH',
+        'user_msg_queue_lock:*',
+        'COUNT',
+        100
+      )
+      cursor = newCursor
+      iterations++
+
+      for (const key of keys) {
+        const accountId = key.replace('user_msg_queue_lock:', '')
+        accountIds.push(accountId)
+      }
+
+      // 防止无限循环
+      if (iterations >= MAX_ITERATIONS) {
+        logger.warn(
+          `📬 User message queue: SCAN reached max iterations (${MAX_ITERATIONS}), stopping early`,
+          { foundLocks: accountIds.length }
+        )
+        break
+      }
+    } while (cursor !== '0')
+
+    if (accountIds.length > 0) {
+      logger.debug(
+        `📬 User message queue: scanned ${accountIds.length} lock(s) in ${iterations} iteration(s)`
+      )
+    }
+
+    return accountIds
+  } catch (error) {
+    logger.error('Failed to scan user message queue locks:', error)
+    return []
+  }
+}
+
+// ============================================
+// 🚦 API Key 并发请求排队方法
+// ============================================
+
+/**
+ * 增加排队计数（使用 Lua 脚本确保原子性）
+ * @param {string} apiKeyId - API Key ID
+ * @param {number} [timeoutMs=60000] - 排队超时时间（毫秒），用于计算 TTL
+ * @returns {Promise<number>} 增加后的排队数量
+ */
+redisClient.incrConcurrencyQueue = async function (apiKeyId, timeoutMs = 60000) {
+  const key = `concurrency:queue:${apiKeyId}`
+  try {
+    // 使用 Lua 脚本确保 INCR 和 EXPIRE 原子执行，防止进程崩溃导致计数器泄漏
+    // TTL = 超时时间 + 缓冲时间（确保键不会在请求还在等待时过期）
+    const ttlSeconds = Math.ceil(timeoutMs / 1000) + QUEUE_TTL_BUFFER_SECONDS
+    const script = `
+      local count = redis.call('INCR', KEYS[1])
+      redis.call('EXPIRE', KEYS[1], ARGV[1])
+      return count
+    `
+    const count = await this.client.eval(script, 1, key, String(ttlSeconds))
+    logger.database(
+      `🚦 Incremented queue count for key ${apiKeyId}: ${count} (TTL: ${ttlSeconds}s)`
+    )
+    return parseInt(count)
+  } catch (error) {
+    logger.error(`Failed to increment concurrency queue for ${apiKeyId}:`, error)
+    throw error
+  }
+}
+
+/**
+ * 减少排队计数（使用 Lua 脚本确保原子性）
+ * @param {string} apiKeyId - API Key ID
+ * @returns {Promise<number>} 减少后的排队数量
+ */
+redisClient.decrConcurrencyQueue = async function (apiKeyId) {
+  const key = `concurrency:queue:${apiKeyId}`
+  try {
+    // 使用 Lua 脚本确保 DECR 和 DEL 原子执行，防止进程崩溃导致计数器残留
+    const script = `
+      local count = redis.call('DECR', KEYS[1])
+      if count <= 0 then
+        redis.call('DEL', KEYS[1])
+        return 0
+      end
+      return count
+    `
+    const count = await this.client.eval(script, 1, key)
+    const result = parseInt(count)
+    if (result === 0) {
+      logger.database(`🚦 Queue count for key ${apiKeyId} is 0, removed key`)
+    } else {
+      logger.database(`🚦 Decremented queue count for key ${apiKeyId}: ${result}`)
+    }
+    return result
+  } catch (error) {
+    logger.error(`Failed to decrement concurrency queue for ${apiKeyId}:`, error)
+    throw error
+  }
+}
+
+/**
+ * 获取排队计数
+ * @param {string} apiKeyId - API Key ID
+ * @returns {Promise<number>} 当前排队数量
+ */
+redisClient.getConcurrencyQueueCount = async function (apiKeyId) {
+  const key = `concurrency:queue:${apiKeyId}`
+  try {
+    const count = await this.client.get(key)
+    return parseInt(count || 0)
+  } catch (error) {
+    logger.error(`Failed to get concurrency queue count for ${apiKeyId}:`, error)
+    return 0
+  }
+}
+
+/**
+ * 清空排队计数
+ * @param {string} apiKeyId - API Key ID
+ * @returns {Promise<boolean>} 是否成功清空
+ */
+redisClient.clearConcurrencyQueue = async function (apiKeyId) {
+  const key = `concurrency:queue:${apiKeyId}`
+  try {
+    await this.client.del(key)
+    logger.database(`🚦 Cleared queue count for key ${apiKeyId}`)
+    return true
+  } catch (error) {
+    logger.error(`Failed to clear concurrency queue for ${apiKeyId}:`, error)
+    return false
+  }
+}
+
+/**
+ * 扫描所有排队计数器
+ * @returns {Promise<string[]>} API Key ID 列表
+ */
+redisClient.scanConcurrencyQueueKeys = async function () {
+  const apiKeyIds = []
+  let cursor = '0'
+  let iterations = 0
+  const MAX_ITERATIONS = 1000
+
+  try {
+    do {
+      const [newCursor, keys] = await this.client.scan(
+        cursor,
+        'MATCH',
+        'concurrency:queue:*',
+        'COUNT',
+        100
+      )
+      cursor = newCursor
+      iterations++
+
+      for (const key of keys) {
+        // 排除统计和等待时间相关的键
+        if (
+          key.startsWith('concurrency:queue:stats:') ||
+          key.startsWith('concurrency:queue:wait_times:')
+        ) {
+          continue
+        }
+        const apiKeyId = key.replace('concurrency:queue:', '')
+        apiKeyIds.push(apiKeyId)
+      }
+
+      if (iterations >= MAX_ITERATIONS) {
+        logger.warn(
+          `🚦 Concurrency queue: SCAN reached max iterations (${MAX_ITERATIONS}), stopping early`,
+          { foundQueues: apiKeyIds.length }
+        )
+        break
+      }
+    } while (cursor !== '0')
+
+    return apiKeyIds
+  } catch (error) {
+    logger.error('Failed to scan concurrency queue keys:', error)
+    return []
+  }
+}
+
+/**
+ * 清理所有排队计数器（用于服务重启）
+ * @returns {Promise<number>} 清理的计数器数量
+ */
+redisClient.clearAllConcurrencyQueues = async function () {
+  let cleared = 0
+  let cursor = '0'
+  let iterations = 0
+  const MAX_ITERATIONS = 1000
+
+  try {
+    do {
+      const [newCursor, keys] = await this.client.scan(
+        cursor,
+        'MATCH',
+        'concurrency:queue:*',
+        'COUNT',
+        100
+      )
+      cursor = newCursor
+      iterations++
+
+      // 只删除排队计数器，保留统计数据
+      const queueKeys = keys.filter(
+        (key) =>
+          !key.startsWith('concurrency:queue:stats:') &&
+          !key.startsWith('concurrency:queue:wait_times:')
+      )
+
+      if (queueKeys.length > 0) {
+        await this.client.del(...queueKeys)
+        cleared += queueKeys.length
+      }
+
+      if (iterations >= MAX_ITERATIONS) {
+        break
+      }
+    } while (cursor !== '0')
+
+    if (cleared > 0) {
+      logger.info(`🚦 Cleared ${cleared} concurrency queue counter(s) on startup`)
+    }
+    return cleared
+  } catch (error) {
+    logger.error('Failed to clear all concurrency queues:', error)
+    return 0
+  }
+}
+
+/**
+ * 增加排队统计计数（使用 Lua 脚本确保原子性）
+ * @param {string} apiKeyId - API Key ID
+ * @param {string} field - 统计字段 (entered/success/timeout/cancelled)
+ * @returns {Promise<number>} 增加后的计数
+ */
+redisClient.incrConcurrencyQueueStats = async function (apiKeyId, field) {
+  const key = `concurrency:queue:stats:${apiKeyId}`
+  try {
+    // 使用 Lua 脚本确保 HINCRBY 和 EXPIRE 原子执行
+    // 防止在两者之间崩溃导致统计键没有 TTL（内存泄漏）
+    const script = `
+      local count = redis.call('HINCRBY', KEYS[1], ARGV[1], 1)
+      redis.call('EXPIRE', KEYS[1], ARGV[2])
+      return count
+    `
+    const count = await this.client.eval(script, 1, key, field, String(QUEUE_STATS_TTL_SECONDS))
+    return parseInt(count)
+  } catch (error) {
+    logger.error(`Failed to increment queue stats ${field} for ${apiKeyId}:`, error)
+    return 0
+  }
+}
+
+/**
+ * 获取排队统计
+ * @param {string} apiKeyId - API Key ID
+ * @returns {Promise<Object>} 统计数据
+ */
+redisClient.getConcurrencyQueueStats = async function (apiKeyId) {
+  const key = `concurrency:queue:stats:${apiKeyId}`
+  try {
+    const stats = await this.client.hgetall(key)
+    return {
+      entered: parseInt(stats?.entered || 0),
+      success: parseInt(stats?.success || 0),
+      timeout: parseInt(stats?.timeout || 0),
+      cancelled: parseInt(stats?.cancelled || 0),
+      socket_changed: parseInt(stats?.socket_changed || 0),
+      rejected_overload: parseInt(stats?.rejected_overload || 0)
+    }
+  } catch (error) {
+    logger.error(`Failed to get queue stats for ${apiKeyId}:`, error)
+    return {
+      entered: 0,
+      success: 0,
+      timeout: 0,
+      cancelled: 0,
+      socket_changed: 0,
+      rejected_overload: 0
+    }
+  }
+}
+
+/**
+ * 记录排队等待时间（按 API Key 分开存储）
+ * @param {string} apiKeyId - API Key ID
+ * @param {number} waitTimeMs - 等待时间（毫秒）
+ * @returns {Promise<void>}
+ */
+redisClient.recordQueueWaitTime = async function (apiKeyId, waitTimeMs) {
+  const key = `concurrency:queue:wait_times:${apiKeyId}`
+  try {
+    // 使用 Lua 脚本确保原子性，同时设置 TTL 防止内存泄漏
+    const script = `
+      redis.call('LPUSH', KEYS[1], ARGV[1])
+      redis.call('LTRIM', KEYS[1], 0, ARGV[2])
+      redis.call('EXPIRE', KEYS[1], ARGV[3])
+      return 1
+    `
+    await this.client.eval(
+      script,
+      1,
+      key,
+      waitTimeMs,
+      WAIT_TIME_SAMPLES_PER_KEY - 1,
+      WAIT_TIME_TTL_SECONDS
+    )
+  } catch (error) {
+    logger.error(`Failed to record queue wait time for ${apiKeyId}:`, error)
+  }
+}
+
+/**
+ * 记录全局排队等待时间
+ * @param {number} waitTimeMs - 等待时间（毫秒）
+ * @returns {Promise<void>}
+ */
+redisClient.recordGlobalQueueWaitTime = async function (waitTimeMs) {
+  const key = 'concurrency:queue:wait_times:global'
+  try {
+    // 使用 Lua 脚本确保原子性，同时设置 TTL 防止内存泄漏
+    const script = `
+      redis.call('LPUSH', KEYS[1], ARGV[1])
+      redis.call('LTRIM', KEYS[1], 0, ARGV[2])
+      redis.call('EXPIRE', KEYS[1], ARGV[3])
+      return 1
+    `
+    await this.client.eval(
+      script,
+      1,
+      key,
+      waitTimeMs,
+      WAIT_TIME_SAMPLES_GLOBAL - 1,
+      WAIT_TIME_TTL_SECONDS
+    )
+  } catch (error) {
+    logger.error('Failed to record global queue wait time:', error)
+  }
+}
+
+/**
+ * 获取全局等待时间列表
+ * @returns {Promise<number[]>} 等待时间列表
+ */
+redisClient.getGlobalQueueWaitTimes = async function () {
+  const key = 'concurrency:queue:wait_times:global'
+  try {
+    const samples = await this.client.lrange(key, 0, -1)
+    return samples.map(Number)
+  } catch (error) {
+    logger.error('Failed to get global queue wait times:', error)
+    return []
+  }
+}
+
+/**
+ * 获取指定 API Key 的等待时间列表
+ * @param {string} apiKeyId - API Key ID
+ * @returns {Promise<number[]>} 等待时间列表
+ */
+redisClient.getQueueWaitTimes = async function (apiKeyId) {
+  const key = `concurrency:queue:wait_times:${apiKeyId}`
+  try {
+    const samples = await this.client.lrange(key, 0, -1)
+    return samples.map(Number)
+  } catch (error) {
+    logger.error(`Failed to get queue wait times for ${apiKeyId}:`, error)
+    return []
+  }
+}
+
+/**
+ * 扫描所有排队统计键
+ * @returns {Promise<string[]>} API Key ID 列表
+ */
+redisClient.scanConcurrencyQueueStatsKeys = async function () {
+  const apiKeyIds = []
+  let cursor = '0'
+  let iterations = 0
+  const MAX_ITERATIONS = 1000
+
+  try {
+    do {
+      const [newCursor, keys] = await this.client.scan(
+        cursor,
+        'MATCH',
+        'concurrency:queue:stats:*',
+        'COUNT',
+        100
+      )
+      cursor = newCursor
+      iterations++
+
+      for (const key of keys) {
+        const apiKeyId = key.replace('concurrency:queue:stats:', '')
+        apiKeyIds.push(apiKeyId)
+      }
+
+      if (iterations >= MAX_ITERATIONS) {
+        break
+      }
+    } while (cursor !== '0')
+
+    return apiKeyIds
+  } catch (error) {
+    logger.error('Failed to scan concurrency queue stats keys:', error)
+    return []
+  }
+}
+
 module.exports = redisClient

src/routes/admin/claudeRelayConfig.js

@@ -40,7 +40,14 @@ router.put('/claude-relay-config', authenticateAdmin, async (req, res) => {
       claudeCodeOnlyEnabled,
       globalSessionBindingEnabled,
       sessionBindingErrorMessage,
-      sessionBindingTtlDays
+      sessionBindingTtlDays,
+      userMessageQueueEnabled,
+      userMessageQueueDelayMs,
+      userMessageQueueTimeoutMs,
+      concurrentRequestQueueEnabled,
+      concurrentRequestQueueMaxSize,
+      concurrentRequestQueueMaxSizeMultiplier,
+      concurrentRequestQueueTimeoutMs
     } = req.body
 
     // 验证输入
@@ -78,15 +85,117 @@ router.put('/claude-relay-config', authenticateAdmin, async (req, res) => {
       }
     }
 
+    // 验证用户消息队列配置
+    if (userMessageQueueEnabled !== undefined && typeof userMessageQueueEnabled !== 'boolean') {
+      return res.status(400).json({ error: 'userMessageQueueEnabled must be a boolean' })
+    }
+
+    if (userMessageQueueDelayMs !== undefined) {
+      if (
+        typeof userMessageQueueDelayMs !== 'number' ||
+        userMessageQueueDelayMs < 0 ||
+        userMessageQueueDelayMs > 10000
+      ) {
+        return res
+          .status(400)
+          .json({ error: 'userMessageQueueDelayMs must be a number between 0 and 10000' })
+      }
+    }
+
+    if (userMessageQueueTimeoutMs !== undefined) {
+      if (
+        typeof userMessageQueueTimeoutMs !== 'number' ||
+        userMessageQueueTimeoutMs < 1000 ||
+        userMessageQueueTimeoutMs > 300000
+      ) {
+        return res
+          .status(400)
+          .json({ error: 'userMessageQueueTimeoutMs must be a number between 1000 and 300000' })
+      }
+    }
+
+    // 验证并发请求排队配置
+    if (
+      concurrentRequestQueueEnabled !== undefined &&
+      typeof concurrentRequestQueueEnabled !== 'boolean'
+    ) {
+      return res.status(400).json({ error: 'concurrentRequestQueueEnabled must be a boolean' })
+    }
+
+    if (concurrentRequestQueueMaxSize !== undefined) {
+      if (
+        typeof concurrentRequestQueueMaxSize !== 'number' ||
+        !Number.isInteger(concurrentRequestQueueMaxSize) ||
+        concurrentRequestQueueMaxSize < 1 ||
+        concurrentRequestQueueMaxSize > 100
+      ) {
+        return res
+          .status(400)
+          .json({ error: 'concurrentRequestQueueMaxSize must be an integer between 1 and 100' })
+      }
+    }
+
+    if (concurrentRequestQueueMaxSizeMultiplier !== undefined) {
+      // 使用 Number.isFinite() 同时排除 NaN、Infinity、-Infinity 和非数字类型
+      if (
+        !Number.isFinite(concurrentRequestQueueMaxSizeMultiplier) ||
+        concurrentRequestQueueMaxSizeMultiplier < 0 ||
+        concurrentRequestQueueMaxSizeMultiplier > 10
+      ) {
+        return res.status(400).json({
+          error: 'concurrentRequestQueueMaxSizeMultiplier must be a finite number between 0 and 10'
+        })
+      }
+    }
+
+    if (concurrentRequestQueueTimeoutMs !== undefined) {
+      if (
+        typeof concurrentRequestQueueTimeoutMs !== 'number' ||
+        !Number.isInteger(concurrentRequestQueueTimeoutMs) ||
+        concurrentRequestQueueTimeoutMs < 5000 ||
+        concurrentRequestQueueTimeoutMs > 300000
+      ) {
+        return res.status(400).json({
+          error:
+            'concurrentRequestQueueTimeoutMs must be an integer between 5000 and 300000 (5 seconds to 5 minutes)'
+        })
+      }
+    }
+
     const updateData = {}
-    if (claudeCodeOnlyEnabled !== undefined)
+    if (claudeCodeOnlyEnabled !== undefined) {
       updateData.claudeCodeOnlyEnabled = claudeCodeOnlyEnabled
-    if (globalSessionBindingEnabled !== undefined)
+    }
+    if (globalSessionBindingEnabled !== undefined) {
       updateData.globalSessionBindingEnabled = globalSessionBindingEnabled
-    if (sessionBindingErrorMessage !== undefined)
+    }
+    if (sessionBindingErrorMessage !== undefined) {
       updateData.sessionBindingErrorMessage = sessionBindingErrorMessage
-    if (sessionBindingTtlDays !== undefined)
+    }
+    if (sessionBindingTtlDays !== undefined) {
       updateData.sessionBindingTtlDays = sessionBindingTtlDays
+    }
+    if (userMessageQueueEnabled !== undefined) {
+      updateData.userMessageQueueEnabled = userMessageQueueEnabled
+    }
+    if (userMessageQueueDelayMs !== undefined) {
+      updateData.userMessageQueueDelayMs = userMessageQueueDelayMs
+    }
+    if (userMessageQueueTimeoutMs !== undefined) {
+      updateData.userMessageQueueTimeoutMs = userMessageQueueTimeoutMs
+    }
+    if (concurrentRequestQueueEnabled !== undefined) {
+      updateData.concurrentRequestQueueEnabled = concurrentRequestQueueEnabled
+    }
+    if (concurrentRequestQueueMaxSize !== undefined) {
+      updateData.concurrentRequestQueueMaxSize = concurrentRequestQueueMaxSize
+    }
+    if (concurrentRequestQueueMaxSizeMultiplier !== undefined) {
+      updateData.concurrentRequestQueueMaxSizeMultiplier = concurrentRequestQueueMaxSizeMultiplier
+    }
+    if (concurrentRequestQueueTimeoutMs !== undefined) {
+      updateData.concurrentRequestQueueTimeoutMs = concurrentRequestQueueTimeoutMs
+    }
 
     const updatedConfig = await claudeRelayConfigService.updateConfig(
       updateData,

src/routes/admin/concurrency.js

@@ -7,26 +7,40 @@ const express = require('express')
 const router = express.Router()
 const redis = require('../../models/redis')
 const logger = require('../../utils/logger')
+const { authenticateAdmin } = require('../../middleware/auth')
+const { calculateWaitTimeStats } = require('../../utils/statsHelper')
 
 /**
  * GET /admin/concurrency
  * 获取所有并发状态
  */
-router.get('/concurrency', async (req, res) => {
+router.get('/concurrency', authenticateAdmin, async (req, res) => {
   try {
     const status = await redis.getAllConcurrencyStatus()
 
+    // 为每个 API Key 获取排队计数
+    const statusWithQueue = await Promise.all(
+      status.map(async (s) => {
+        const queueCount = await redis.getConcurrencyQueueCount(s.apiKeyId)
+        return {
+          ...s,
+          queueCount
+        }
+      })
+    )
+
     // 计算汇总统计
     const summary = {
-      totalKeys: status.length,
+      totalKeys: statusWithQueue.length,
-      totalActiveRequests: status.reduce((sum, s) => sum + s.activeCount, 0),
+      totalActiveRequests: statusWithQueue.reduce((sum, s) => sum + s.activeCount, 0),
-      totalExpiredRequests: status.reduce((sum, s) => sum + s.expiredCount, 0)
+      totalExpiredRequests: statusWithQueue.reduce((sum, s) => sum + s.expiredCount, 0),
+      totalQueuedRequests: statusWithQueue.reduce((sum, s) => sum + s.queueCount, 0)
     }
 
     res.json({
       success: true,
       summary,
-      concurrencyStatus: status
+      concurrencyStatus: statusWithQueue
     })
   } catch (error) {
     logger.error('❌ Failed to get concurrency status:', error)
@@ -39,17 +53,171 @@ router.get('/concurrency', async (req, res) => {
 })
 
 /**
- * GET /admin/concurrency/:apiKeyId
+ * GET /admin/concurrency-queue/stats
- * 获取特定 API Key 的并发状态详情
+ * 获取排队统计信息
  */
-router.get('/concurrency/:apiKeyId', async (req, res) => {
+router.get('/concurrency-queue/stats', authenticateAdmin, async (req, res) => {
   try {
-    const { apiKeyId } = req.params
+    // 获取所有有统计数据的 API Key
-    const status = await redis.getConcurrencyStatus(apiKeyId)
+    const statsKeys = await redis.scanConcurrencyQueueStatsKeys()
+    const queueKeys = await redis.scanConcurrencyQueueKeys()
+
+    // 合并所有相关的 API Key
+    const allApiKeyIds = [...new Set([...statsKeys, ...queueKeys])]
+
+    // 获取各 API Key 的详细统计
+    const perKeyStats = await Promise.all(
+      allApiKeyIds.map(async (apiKeyId) => {
+        const [queueCount, stats, waitTimes] = await Promise.all([
+          redis.getConcurrencyQueueCount(apiKeyId),
+          redis.getConcurrencyQueueStats(apiKeyId),
+          redis.getQueueWaitTimes(apiKeyId)
+        ])
+
+        return {
+          apiKeyId,
+          currentQueueCount: queueCount,
+          stats,
+          waitTimeStats: calculateWaitTimeStats(waitTimes)
+        }
+      })
+    )
+
+    // 获取全局等待时间统计
+    const globalWaitTimes = await redis.getGlobalQueueWaitTimes()
+    const globalWaitTimeStats = calculateWaitTimeStats(globalWaitTimes)
+
+    // 计算全局汇总
+    const globalStats = {
+      totalEntered: perKeyStats.reduce((sum, s) => sum + s.stats.entered, 0),
+      totalSuccess: perKeyStats.reduce((sum, s) => sum + s.stats.success, 0),
+      totalTimeout: perKeyStats.reduce((sum, s) => sum + s.stats.timeout, 0),
+      totalCancelled: perKeyStats.reduce((sum, s) => sum + s.stats.cancelled, 0),
+      totalSocketChanged: perKeyStats.reduce((sum, s) => sum + (s.stats.socket_changed || 0), 0),
+      totalRejectedOverload: perKeyStats.reduce(
+        (sum, s) => sum + (s.stats.rejected_overload || 0),
+        0
+      ),
+      currentTotalQueued: perKeyStats.reduce((sum, s) => sum + s.currentQueueCount, 0),
+      // 队列资源利用率指标
+      peakQueueSize:
+        perKeyStats.length > 0 ? Math.max(...perKeyStats.map((s) => s.currentQueueCount)) : 0,
+      avgQueueSize:
+        perKeyStats.length > 0
+          ? Math.round(
+              perKeyStats.reduce((sum, s) => sum + s.currentQueueCount, 0) / perKeyStats.length
+            )
+          : 0,
+      activeApiKeys: perKeyStats.filter((s) => s.currentQueueCount > 0).length
+    }
+
+    // 计算成功率
+    if (globalStats.totalEntered > 0) {
+      globalStats.successRate = Math.round(
+        (globalStats.totalSuccess / globalStats.totalEntered) * 100
+      )
+      globalStats.timeoutRate = Math.round(
+        (globalStats.totalTimeout / globalStats.totalEntered) * 100
+      )
+      globalStats.cancelledRate = Math.round(
+        (globalStats.totalCancelled / globalStats.totalEntered) * 100
+      )
+    }
+
+    // 从全局等待时间统计中提取关键指标
+    if (globalWaitTimeStats) {
+      globalStats.avgWaitTimeMs = globalWaitTimeStats.avg
+      globalStats.p50WaitTimeMs = globalWaitTimeStats.p50
+      globalStats.p90WaitTimeMs = globalWaitTimeStats.p90
+      globalStats.p99WaitTimeMs = globalWaitTimeStats.p99
+      // 多实例采样策略标记（详见 design.md Decision 9）
+      // 全局 P90 仅用于可视化和监控，不用于系统决策
+      // 健康检查使用 API Key 级别的 P90（每 Key 独立采样）
+      globalWaitTimeStats.globalP90ForVisualizationOnly = true
+    }
 
     res.json({
       success: true,
-      concurrencyStatus: status
+      globalStats,
+      globalWaitTimeStats,
+      perKeyStats
+    })
+  } catch (error) {
+    logger.error('❌ Failed to get queue stats:', error)
+    res.status(500).json({
+      success: false,
+      error: 'Failed to get queue stats',
+      message: error.message
+    })
+  }
+})
+
+/**
+ * DELETE /admin/concurrency-queue/:apiKeyId
+ * 清理特定 API Key 的排队计数
+ */
+router.delete('/concurrency-queue/:apiKeyId', authenticateAdmin, async (req, res) => {
+  try {
+    const { apiKeyId } = req.params
+    await redis.clearConcurrencyQueue(apiKeyId)
+
+    logger.warn(`🧹 Admin ${req.admin?.username || 'unknown'} cleared queue for key ${apiKeyId}`)
+
+    res.json({
+      success: true,
+      message: `Successfully cleared queue for API key ${apiKeyId}`
+    })
+  } catch (error) {
+    logger.error(`❌ Failed to clear queue for ${req.params.apiKeyId}:`, error)
+    res.status(500).json({
+      success: false,
+      error: 'Failed to clear queue',
+      message: error.message
+    })
+  }
+})
+
+/**
+ * DELETE /admin/concurrency-queue
+ * 清理所有排队计数
+ */
+router.delete('/concurrency-queue', authenticateAdmin, async (req, res) => {
+  try {
+    const cleared = await redis.clearAllConcurrencyQueues()
+
+    logger.warn(`🧹 Admin ${req.admin?.username || 'unknown'} cleared ALL queues`)
+
+    res.json({
+      success: true,
+      message: 'Successfully cleared all queues',
+      cleared
+    })
+  } catch (error) {
+    logger.error('❌ Failed to clear all queues:', error)
+    res.status(500).json({
+      success: false,
+      error: 'Failed to clear all queues',
+      message: error.message
+    })
+  }
+})
+
+/**
+ * GET /admin/concurrency/:apiKeyId
+ * 获取特定 API Key 的并发状态详情
+ */
+router.get('/concurrency/:apiKeyId', authenticateAdmin, async (req, res) => {
+  try {
+    const { apiKeyId } = req.params
+    const status = await redis.getConcurrencyStatus(apiKeyId)
+    const queueCount = await redis.getConcurrencyQueueCount(apiKeyId)
+
+    res.json({
+      success: true,
+      concurrencyStatus: {
+        ...status,
+        queueCount
+      }
     })
   } catch (error) {
     logger.error(`❌ Failed to get concurrency status for ${req.params.apiKeyId}:`, error)
@@ -65,7 +233,7 @@ router.get('/concurrency/:apiKeyId', async (req, res) => {
  * DELETE /admin/concurrency/:apiKeyId
  * 强制清理特定 API Key 的并发计数
  */
-router.delete('/concurrency/:apiKeyId', async (req, res) => {
+router.delete('/concurrency/:apiKeyId', authenticateAdmin, async (req, res) => {
   try {
     const { apiKeyId } = req.params
     const result = await redis.forceClearConcurrency(apiKeyId)
@@ -93,7 +261,7 @@ router.delete('/concurrency/:apiKeyId', async (req, res) => {
  * DELETE /admin/concurrency
  * 强制清理所有并发计数
  */
-router.delete('/concurrency', async (req, res) => {
+router.delete('/concurrency', authenticateAdmin, async (req, res) => {
   try {
     const result = await redis.forceClearAllConcurrency()
 
@@ -118,7 +286,7 @@ router.delete('/concurrency', async (req, res) => {
  * POST /admin/concurrency/cleanup
  * 清理过期的并发条目（不影响活跃请求）
  */
-router.post('/concurrency/cleanup', async (req, res) => {
+router.post('/concurrency/cleanup', authenticateAdmin, async (req, res) => {
   try {
     const { apiKeyId } = req.body
     const result = await redis.cleanupExpiredConcurrency(apiKeyId || null)

src/routes/api.js

@@ -190,12 +190,42 @@ async function handleMessagesRequest(req, res) {
     )
 
     if (isStream) {
+      // 🔍 检查客户端连接是否仍然有效（可能在并发排队等待期间断开）
+      if (res.destroyed || res.socket?.destroyed || res.writableEnded) {
+        logger.warn(
+          `⚠️ Client disconnected before stream response could start for key: ${req.apiKey?.name || 'unknown'}`
+        )
+        return undefined
+      }
+
       // 流式响应 - 只使用官方真实usage数据
       res.setHeader('Content-Type', 'text/event-stream')
       res.setHeader('Cache-Control', 'no-cache')
       res.setHeader('Connection', 'keep-alive')
       res.setHeader('Access-Control-Allow-Origin', '*')
       res.setHeader('X-Accel-Buffering', 'no') // 禁用 Nginx 缓冲
+      // ⚠️ 检查 headers 是否已发送（可能在排队心跳时已设置）
+      if (!res.headersSent) {
+        res.setHeader('Content-Type', 'text/event-stream')
+        res.setHeader('Cache-Control', 'no-cache')
+        // ⚠️ 关键修复：尊重 auth.js 提前设置的 Connection: close
+        // 当并发队列功能启用时，auth.js 会设置 Connection: close 来禁用 Keep-Alive
+        // 这里只在没有设置过 Connection 头时才设置 keep-alive
+        const existingConnection = res.getHeader('Connection')
+        if (!existingConnection) {
+          res.setHeader('Connection', 'keep-alive')
+        } else {
+          logger.api(
+            `🔌 [STREAM] Preserving existing Connection header: ${existingConnection} for key: ${req.apiKey?.name || 'unknown'}`
+          )
+        }
+        res.setHeader('Access-Control-Allow-Origin', '*')
+        res.setHeader('X-Accel-Buffering', 'no') // 禁用 Nginx 缓冲
+      } else {
+        logger.debug(
+          `📤 [STREAM] Headers already sent, skipping setHeader for key: ${req.apiKey?.name || 'unknown'}`
+        )
+      }
 
       // 禁用 Nagle 算法，确保数据立即发送
       if (res.socket && typeof res.socket.setNoDelay === 'function') {
@@ -657,12 +687,61 @@ async function handleMessagesRequest(req, res) {
         }
       }, 1000) // 1秒后检查
     } else {
+      // 🔍 检查客户端连接是否仍然有效（可能在并发排队等待期间断开）
+      if (res.destroyed || res.socket?.destroyed || res.writableEnded) {
+        logger.warn(
+          `⚠️ Client disconnected before non-stream request could start for key: ${req.apiKey?.name || 'unknown'}`
+        )
+        return undefined
+      }
+
       // 非流式响应 - 只使用官方真实usage数据
       logger.info('📄 Starting non-streaming request', {
         apiKeyId: req.apiKey.id,
         apiKeyName: req.apiKey.name
       })
 
+      // 📊 监听 socket 事件以追踪连接状态变化
+      const nonStreamSocket = res.socket
+      let _clientClosedConnection = false
+      let _socketCloseTime = null
+
+      if (nonStreamSocket) {
+        const onSocketEnd = () => {
+          _clientClosedConnection = true
+          _socketCloseTime = Date.now()
+          logger.warn(
+            `⚠️ [NON-STREAM] Socket 'end' event - client sent FIN | key: ${req.apiKey?.name}, ` +
+              `requestId: ${req.requestId}, elapsed: ${Date.now() - startTime}ms`
+          )
+        }
+        const onSocketClose = () => {
+          _clientClosedConnection = true
+          logger.warn(
+            `⚠️ [NON-STREAM] Socket 'close' event | key: ${req.apiKey?.name}, ` +
+              `requestId: ${req.requestId}, elapsed: ${Date.now() - startTime}ms, ` +
+              `hadError: ${nonStreamSocket.destroyed}`
+          )
+        }
+        const onSocketError = (err) => {
+          logger.error(
+            `❌ [NON-STREAM] Socket error | key: ${req.apiKey?.name}, ` +
+              `requestId: ${req.requestId}, error: ${err.message}`
+          )
+        }
+
+        nonStreamSocket.once('end', onSocketEnd)
+        nonStreamSocket.once('close', onSocketClose)
+        nonStreamSocket.once('error', onSocketError)
+
+        // 清理监听器（在响应结束后）
+        res.once('finish', () => {
+          nonStreamSocket.removeListener('end', onSocketEnd)
+          nonStreamSocket.removeListener('close', onSocketClose)
+          nonStreamSocket.removeListener('error', onSocketError)
+        })
+      }
+
       // 生成会话哈希用于sticky会话
       const sessionHash = sessionHelper.generateSessionHash(req.body)
 
@@ -867,6 +946,15 @@ async function handleMessagesRequest(req, res) {
         bodyLength: response.body ? response.body.length : 0
       })
 
+      // 🔍 检查客户端连接是否仍然有效
+      // 在长时间请求过程中，客户端可能已经断开连接（超时、用户取消等）
+      if (res.destroyed || res.socket?.destroyed || res.writableEnded) {
+        logger.warn(
+          `⚠️ Client disconnected before non-stream response could be sent for key: ${req.apiKey?.name || 'unknown'}`
+        )
+        return undefined
+      }
+
       res.status(response.statusCode)
 
       // 设置响应头，避免 Content-Length 和 Transfer-Encoding 冲突
@@ -932,10 +1020,12 @@ async function handleMessagesRequest(req, res) {
           logger.warn('⚠️ No usage data found in Claude API JSON response')
         }
 
+        // 使用 Express 内建的 res.json() 发送响应（简单可靠）
         res.json(jsonData)
       } catch (parseError) {
         logger.warn('⚠️ Failed to parse Claude API response as JSON:', parseError.message)
         logger.info('📄 Raw response body:', response.body)
+        // 使用 Express 内建的 res.send() 发送响应（简单可靠）
         res.send(response.body)
       }
 

src/services/bedrockRelayService.js

@@ -6,6 +6,7 @@ const {
 const { fromEnv } = require('@aws-sdk/credential-providers')
 const logger = require('../utils/logger')
 const config = require('../../config/config')
+const userMessageQueueService = require('./userMessageQueueService')
 
 class BedrockRelayService {
   constructor() {
@@ -69,7 +70,68 @@ class BedrockRelayService {
 
   // 处理非流式请求
   async handleNonStreamRequest(requestBody, bedrockAccount = null) {
+    const accountId = bedrockAccount?.id
+    let queueLockAcquired = false
+    let queueRequestId = null
+
     try {
+      // 📬 用户消息队列处理
+      if (userMessageQueueService.isUserMessageRequest(requestBody)) {
+        // 校验 accountId 非空，避免空值污染队列锁键
+        if (!accountId || accountId === '') {
+          logger.error('❌ accountId missing for queue lock in Bedrock handleNonStreamRequest')
+          throw new Error('accountId missing for queue lock')
+        }
+        const queueResult = await userMessageQueueService.acquireQueueLock(accountId)
+        if (!queueResult.acquired && !queueResult.skipped) {
+          // 区分 Redis 后端错误和队列超时
+          const isBackendError = queueResult.error === 'queue_backend_error'
+          const errorCode = isBackendError ? 'QUEUE_BACKEND_ERROR' : 'QUEUE_TIMEOUT'
+          const errorType = isBackendError ? 'queue_backend_error' : 'queue_timeout'
+          const errorMessage = isBackendError
+            ? 'Queue service temporarily unavailable, please retry later'
+            : 'User message queue wait timeout, please retry later'
+          const statusCode = isBackendError ? 500 : 503
+
+          // 结构化性能日志，用于后续统计
+          logger.performance('user_message_queue_error', {
+            errorType,
+            errorCode,
+            accountId,
+            statusCode,
+            backendError: isBackendError ? queueResult.errorMessage : undefined
+          })
+
+          logger.warn(
+            `📬 User message queue ${errorType} for Bedrock account ${accountId}`,
+            isBackendError ? { backendError: queueResult.errorMessage } : {}
+          )
+          return {
+            statusCode,
+            headers: {
+              'Content-Type': 'application/json',
+              'x-user-message-queue-error': errorType
+            },
+            body: JSON.stringify({
+              type: 'error',
+              error: {
+                type: errorType,
+                code: errorCode,
+                message: errorMessage
+              }
+            }),
+            success: false
+          }
+        }
+        if (queueResult.acquired && !queueResult.skipped) {
+          queueLockAcquired = true
+          queueRequestId = queueResult.requestId
+          logger.debug(
+            `📬 User message queue lock acquired for Bedrock account ${accountId}, requestId: ${queueRequestId}`
+          )
+        }
+      }
+
       const modelId = this._selectModel(requestBody, bedrockAccount)
       const region = this._selectRegion(modelId, bedrockAccount)
       const client = this._getBedrockClient(region, bedrockAccount)
@@ -90,6 +152,23 @@ class BedrockRelayService {
       const response = await client.send(command)
       const duration = Date.now() - startTime
 
+      // 📬 请求已发送成功，立即释放队列锁（无需等待响应处理完成）
+      // 因为限流基于请求发送时刻计算（RPM），不是请求完成时刻
+      if (queueLockAcquired && queueRequestId && accountId) {
+        try {
+          await userMessageQueueService.releaseQueueLock(accountId, queueRequestId)
+          queueLockAcquired = false // 标记已释放，防止 finally 重复释放
+          logger.debug(
+            `📬 User message queue lock released early for Bedrock account ${accountId}, requestId: ${queueRequestId}`
+          )
+        } catch (releaseError) {
+          logger.error(
+            `❌ Failed to release user message queue lock early for Bedrock account ${accountId}:`,
+            releaseError.message
+          )
+        }
+      }
+
       // 解析响应
       const responseBody = JSON.parse(new TextDecoder().decode(response.body))
       const claudeResponse = this._convertFromBedrockFormat(responseBody)
@@ -106,12 +185,94 @@ class BedrockRelayService {
     } catch (error) {
       logger.error('❌ Bedrock非流式请求失败:', error)
       throw this._handleBedrockError(error)
+    } finally {
+      // 📬 释放用户消息队列锁（兜底，正常情况下已在请求发送后提前释放）
+      if (queueLockAcquired && queueRequestId && accountId) {
+        try {
+          await userMessageQueueService.releaseQueueLock(accountId, queueRequestId)
+          logger.debug(
+            `📬 User message queue lock released in finally for Bedrock account ${accountId}, requestId: ${queueRequestId}`
+          )
+        } catch (releaseError) {
+          logger.error(
+            `❌ Failed to release user message queue lock for Bedrock account ${accountId}:`,
+            releaseError.message
+          )
+        }
+      }
     }
   }
 
   // 处理流式请求
   async handleStreamRequest(requestBody, bedrockAccount = null, res) {
+    const accountId = bedrockAccount?.id
+    let queueLockAcquired = false
+    let queueRequestId = null
+
     try {
+      // 📬 用户消息队列处理
+      if (userMessageQueueService.isUserMessageRequest(requestBody)) {
+        // 校验 accountId 非空，避免空值污染队列锁键
+        if (!accountId || accountId === '') {
+          logger.error('❌ accountId missing for queue lock in Bedrock handleStreamRequest')
+          throw new Error('accountId missing for queue lock')
+        }
+        const queueResult = await userMessageQueueService.acquireQueueLock(accountId)
+        if (!queueResult.acquired && !queueResult.skipped) {
+          // 区分 Redis 后端错误和队列超时
+          const isBackendError = queueResult.error === 'queue_backend_error'
+          const errorCode = isBackendError ? 'QUEUE_BACKEND_ERROR' : 'QUEUE_TIMEOUT'
+          const errorType = isBackendError ? 'queue_backend_error' : 'queue_timeout'
+          const errorMessage = isBackendError
+            ? 'Queue service temporarily unavailable, please retry later'
+            : 'User message queue wait timeout, please retry later'
+          const statusCode = isBackendError ? 500 : 503
+
+          // 结构化性能日志，用于后续统计
+          logger.performance('user_message_queue_error', {
+            errorType,
+            errorCode,
+            accountId,
+            statusCode,
+            stream: true,
+            backendError: isBackendError ? queueResult.errorMessage : undefined
+          })
+
+          logger.warn(
+            `📬 User message queue ${errorType} for Bedrock account ${accountId} (stream)`,
+            isBackendError ? { backendError: queueResult.errorMessage } : {}
+          )
+          if (!res.headersSent) {
+            const existingConnection = res.getHeader ? res.getHeader('Connection') : null
+            res.writeHead(statusCode, {
+              'Content-Type': 'text/event-stream',
+              'Cache-Control': 'no-cache',
+              Connection: existingConnection || 'keep-alive',
+              'x-user-message-queue-error': errorType
+            })
+          }
+          const errorEvent = `event: error\ndata: ${JSON.stringify({
+            type: 'error',
+            error: {
+              type: errorType,
+              code: errorCode,
+              message: errorMessage
+            }
+          })}\n\n`
+          res.write(errorEvent)
+          res.write('data: [DONE]\n\n')
+          res.end()
+          return { success: false, error: errorType }
+        }
+        if (queueResult.acquired && !queueResult.skipped) {
+          queueLockAcquired = true
+          queueRequestId = queueResult.requestId
+          logger.debug(
+            `📬 User message queue lock acquired for Bedrock account ${accountId} (stream), requestId: ${queueRequestId}`
+          )
+        }
+      }
+
       const modelId = this._selectModel(requestBody, bedrockAccount)
       const region = this._selectRegion(modelId, bedrockAccount)
       const client = this._getBedrockClient(region, bedrockAccount)
@@ -131,11 +292,35 @@ class BedrockRelayService {
       const startTime = Date.now()
       const response = await client.send(command)
 
+      // 📬 请求已发送成功，立即释放队列锁（无需等待响应处理完成）
+      // 因为限流基于请求发送时刻计算（RPM），不是请求完成时刻
+      if (queueLockAcquired && queueRequestId && accountId) {
+        try {
+          await userMessageQueueService.releaseQueueLock(accountId, queueRequestId)
+          queueLockAcquired = false // 标记已释放，防止 finally 重复释放
+          logger.debug(
+            `📬 User message queue lock released early for Bedrock stream account ${accountId}, requestId: ${queueRequestId}`
+          )
+        } catch (releaseError) {
+          logger.error(
+            `❌ Failed to release user message queue lock early for Bedrock stream account ${accountId}:`,
+            releaseError.message
+          )
+        }
+      }
+
       // 设置SSE响应头
+      // ⚠️ 关键修复：尊重 auth.js 提前设置的 Connection: close
+      const existingConnection = res.getHeader ? res.getHeader('Connection') : null
+      if (existingConnection) {
+        logger.debug(
+          `🔌 [Bedrock Stream] Preserving existing Connection header: ${existingConnection}`
+        )
+      }
       res.writeHead(200, {
         'Content-Type': 'text/event-stream',
         'Cache-Control': 'no-cache',
-        Connection: 'keep-alive',
+        Connection: existingConnection || 'keep-alive',
         'Access-Control-Allow-Origin': '*',
         'Access-Control-Allow-Headers': 'Content-Type, Authorization'
       })
@@ -191,6 +376,21 @@ class BedrockRelayService {
       res.end()
 
       throw this._handleBedrockError(error)
+    } finally {
+      // 📬 释放用户消息队列锁（兜底，正常情况下已在请求发送后提前释放）
+      if (queueLockAcquired && queueRequestId && accountId) {
+        try {
+          await userMessageQueueService.releaseQueueLock(accountId, queueRequestId)
+          logger.debug(
+            `📬 User message queue lock released in finally for Bedrock stream account ${accountId}, requestId: ${queueRequestId}`
+          )
+        } catch (releaseError) {
+          logger.error(
+            `❌ Failed to release user message queue lock for Bedrock stream account ${accountId}:`,
+            releaseError.message
+          )
+        }
+      }
     }
   }
 

src/services/ccrRelayService.js

@@ -3,6 +3,8 @@ const ccrAccountService = require('./ccrAccountService')
 const logger = require('../utils/logger')
 const config = require('../../config/config')
 const { parseVendorPrefixedModel } = require('../utils/modelHelper')
+const userMessageQueueService = require('./userMessageQueueService')
+const { isStreamWritable } = require('../utils/streamHelper')
 
 class CcrRelayService {
   constructor() {
@@ -21,8 +23,67 @@ class CcrRelayService {
   ) {
     let abortController = null
     let account = null
+    let queueLockAcquired = false
+    let queueRequestId = null
 
     try {
+      // 📬 用户消息队列处理
+      if (userMessageQueueService.isUserMessageRequest(requestBody)) {
+        // 校验 accountId 非空，避免空值污染队列锁键
+        if (!accountId || accountId === '') {
+          logger.error('❌ accountId missing for queue lock in CCR relayRequest')
+          throw new Error('accountId missing for queue lock')
+        }
+        const queueResult = await userMessageQueueService.acquireQueueLock(accountId)
+        if (!queueResult.acquired && !queueResult.skipped) {
+          // 区分 Redis 后端错误和队列超时
+          const isBackendError = queueResult.error === 'queue_backend_error'
+          const errorCode = isBackendError ? 'QUEUE_BACKEND_ERROR' : 'QUEUE_TIMEOUT'
+          const errorType = isBackendError ? 'queue_backend_error' : 'queue_timeout'
+          const errorMessage = isBackendError
+            ? 'Queue service temporarily unavailable, please retry later'
+            : 'User message queue wait timeout, please retry later'
+          const statusCode = isBackendError ? 500 : 503
+
+          // 结构化性能日志，用于后续统计
+          logger.performance('user_message_queue_error', {
+            errorType,
+            errorCode,
+            accountId,
+            statusCode,
+            backendError: isBackendError ? queueResult.errorMessage : undefined
+          })
+
+          logger.warn(
+            `📬 User message queue ${errorType} for CCR account ${accountId}`,
+            isBackendError ? { backendError: queueResult.errorMessage } : {}
+          )
+          return {
+            statusCode,
+            headers: {
+              'Content-Type': 'application/json',
+              'x-user-message-queue-error': errorType
+            },
+            body: JSON.stringify({
+              type: 'error',
+              error: {
+                type: errorType,
+                code: errorCode,
+                message: errorMessage
+              }
+            }),
+            accountId
+          }
+        }
+        if (queueResult.acquired && !queueResult.skipped) {
+          queueLockAcquired = true
+          queueRequestId = queueResult.requestId
+          logger.debug(
+            `📬 User message queue lock acquired for CCR account ${accountId}, requestId: ${queueRequestId}`
+          )
+        }
+      }
+
       // 获取账户信息
       account = await ccrAccountService.getAccount(accountId)
       if (!account) {
@@ -162,6 +223,23 @@ class CcrRelayService {
       )
       const response = await axios(requestConfig)
 
+      // 📬 请求已发送成功，立即释放队列锁（无需等待响应处理完成）
+      // 因为 Claude API 限流基于请求发送时刻计算（RPM），不是请求完成时刻
+      if (queueLockAcquired && queueRequestId && accountId) {
+        try {
+          await userMessageQueueService.releaseQueueLock(accountId, queueRequestId)
+          queueLockAcquired = false // 标记已释放，防止 finally 重复释放
+          logger.debug(
+            `📬 User message queue lock released early for CCR account ${accountId}, requestId: ${queueRequestId}`
+          )
+        } catch (releaseError) {
+          logger.error(
+            `❌ Failed to release user message queue lock early for CCR account ${accountId}:`,
+            releaseError.message
+          )
+        }
+      }
+
       // 移除监听器（请求成功完成）
       if (clientRequest) {
         clientRequest.removeListener('close', handleClientDisconnect)
@@ -233,6 +311,21 @@ class CcrRelayService {
       )
 
       throw error
+    } finally {
+      // 📬 释放用户消息队列锁（兜底，正常情况下已在请求发送后提前释放）
+      if (queueLockAcquired && queueRequestId && accountId) {
+        try {
+          await userMessageQueueService.releaseQueueLock(accountId, queueRequestId)
+          logger.debug(
+            `📬 User message queue lock released in finally for CCR account ${accountId}, requestId: ${queueRequestId}`
+          )
+        } catch (releaseError) {
+          logger.error(
+            `❌ Failed to release user message queue lock for CCR account ${accountId}:`,
+            releaseError.message
+          )
+        }
+      }
     }
   }
 
@@ -248,7 +341,77 @@ class CcrRelayService {
     options = {}
   ) {
     let account = null
+    let queueLockAcquired = false
+    let queueRequestId = null
+
     try {
+      // 📬 用户消息队列处理
+      if (userMessageQueueService.isUserMessageRequest(requestBody)) {
+        // 校验 accountId 非空，避免空值污染队列锁键
+        if (!accountId || accountId === '') {
+          logger.error(
+            '❌ accountId missing for queue lock in CCR relayStreamRequestWithUsageCapture'
+          )
+          throw new Error('accountId missing for queue lock')
+        }
+        const queueResult = await userMessageQueueService.acquireQueueLock(accountId)
+        if (!queueResult.acquired && !queueResult.skipped) {
+          // 区分 Redis 后端错误和队列超时
+          const isBackendError = queueResult.error === 'queue_backend_error'
+          const errorCode = isBackendError ? 'QUEUE_BACKEND_ERROR' : 'QUEUE_TIMEOUT'
+          const errorType = isBackendError ? 'queue_backend_error' : 'queue_timeout'
+          const errorMessage = isBackendError
+            ? 'Queue service temporarily unavailable, please retry later'
+            : 'User message queue wait timeout, please retry later'
+          const statusCode = isBackendError ? 500 : 503
+
+          // 结构化性能日志，用于后续<E5908E><E7BBAD>计
+          logger.performance('user_message_queue_error', {
+            errorType,
+            errorCode,
+            accountId,
+            statusCode,
+            stream: true,
+            backendError: isBackendError ? queueResult.errorMessage : undefined
+          })
+
+          logger.warn(
+            `📬 User message queue ${errorType} for CCR account ${accountId} (stream)`,
+            isBackendError ? { backendError: queueResult.errorMessage } : {}
+          )
+          if (!responseStream.headersSent) {
+            const existingConnection = responseStream.getHeader
+              ? responseStream.getHeader('Connection')
+              : null
+            responseStream.writeHead(statusCode, {
+              'Content-Type': 'text/event-stream',
+              'Cache-Control': 'no-cache',
+              Connection: existingConnection || 'keep-alive',
+              'x-user-message-queue-error': errorType
+            })
+          }
+          const errorEvent = `event: error\ndata: ${JSON.stringify({
+            type: 'error',
+            error: {
+              type: errorType,
+              code: errorCode,
+              message: errorMessage
+            }
+          })}\n\n`
+          responseStream.write(errorEvent)
+          responseStream.write('data: [DONE]\n\n')
+          responseStream.end()
+          return
+        }
+        if (queueResult.acquired && !queueResult.skipped) {
+          queueLockAcquired = true
+          queueRequestId = queueResult.requestId
+          logger.debug(
+            `📬 User message queue lock acquired for CCR account ${accountId} (stream), requestId: ${queueRequestId}`
+          )
+        }
+      }
+
       // 获取账户信息
       account = await ccrAccountService.getAccount(accountId)
       if (!account) {
@@ -296,14 +459,53 @@ class CcrRelayService {
         accountId,
         usageCallback,
         streamTransformer,
-        options
+        options,
+        // 📬 回调：在收到响应头时释放队列锁
+        async () => {
+          if (queueLockAcquired && queueRequestId && accountId) {
+            try {
+              await userMessageQueueService.releaseQueueLock(accountId, queueRequestId)
+              queueLockAcquired = false // 标记已释放，防止 finally 重复释放
+              logger.debug(
+                `📬 User message queue lock released early for CCR stream account ${accountId}, requestId: ${queueRequestId}`
+              )
+            } catch (releaseError) {
+              logger.error(
+                `❌ Failed to release user message queue lock early for CCR stream account ${accountId}:`,
+                releaseError.message
+              )
+            }
+          }
+        }
       )
 
       // 更新最后使用时间
       await this._updateLastUsedTime(accountId)
     } catch (error) {
+      // 客户端主动断开连接是正常情况，使用 INFO 级别
+      if (error.message === 'Client disconnected') {
+        logger.info(
+          `🔌 CCR stream relay ended: Client disconnected (Account: ${account?.name || accountId})`
+        )
+      } else {
         logger.error(`❌ CCR stream relay failed (Account: ${account?.name || accountId}):`, error)
+      }
       throw error
+    } finally {
+      // 📬 释放用户消息队列锁（兜底，正常情况下已在收到响应头后提前释放）
+      if (queueLockAcquired && queueRequestId && accountId) {
+        try {
+          await userMessageQueueService.releaseQueueLock(accountId, queueRequestId)
+          logger.debug(
+            `📬 User message queue lock released in finally for CCR stream account ${accountId}, requestId: ${queueRequestId}`
+          )
+        } catch (releaseError) {
+          logger.error(
+            `❌ Failed to release user message queue lock for CCR stream account ${accountId}:`,
+            releaseError.message
+          )
+        }
+      }
     }
   }
 
@@ -317,7 +519,8 @@ class CcrRelayService {
     accountId,
     usageCallback,
     streamTransformer = null,
-    requestOptions = {}
+    requestOptions = {},
+    onResponseHeaderReceived = null
   ) {
     return new Promise((resolve, reject) => {
       let aborted = false
@@ -380,8 +583,11 @@ class CcrRelayService {
       // 发送请求
       const request = axios(requestConfig)
 
+      // 注意：使用 .then(async ...) 模式处理响应
+      // - 内部的 releaseQueueLock 有独立的 try-catch，不会导致未捕获异常
+      // - queueLockAcquired = false 的赋值会在 finally 执行前完成（JS 单线程保证）
       request
-        .then((response) => {
+        .then(async (response) => {
           logger.debug(`🌊 CCR stream response status: ${response.status}`)
 
           // 错误响应处理
@@ -404,10 +610,13 @@ class CcrRelayService {
 
             // 设置错误响应的状态码和响应头
             if (!responseStream.headersSent) {
+              const existingConnection = responseStream.getHeader
+                ? responseStream.getHeader('Connection')
+                : null
               const errorHeaders = {
                 'Content-Type': response.headers['content-type'] || 'application/json',
                 'Cache-Control': 'no-cache',
-                Connection: 'keep-alive'
+                Connection: existingConnection || 'keep-alive'
               }
               // 避免 Transfer-Encoding 冲突，让 Express 自动处理
               delete errorHeaders['Transfer-Encoding']
@@ -417,13 +626,13 @@ class CcrRelayService {
 
             // 直接透传错误数据，不进行包装
             response.data.on('data', (chunk) => {
-              if (!responseStream.destroyed) {
+              if (isStreamWritable(responseStream)) {
                 responseStream.write(chunk)
               }
             })
 
             response.data.on('end', () => {
-              if (!responseStream.destroyed) {
+              if (isStreamWritable(responseStream)) {
                 responseStream.end()
               }
               resolve() // 不抛出异常，正常完成流处理
@@ -431,6 +640,19 @@ class CcrRelayService {
             return
           }
 
+          // 📬 收到成功响应头（HTTP 200），调用回调释放队列锁
+          // 此时请求已被 Claude API 接受并计入 RPM 配额，无需等待响应完成
+          if (onResponseHeaderReceived && typeof onResponseHeaderReceived === 'function') {
+            try {
+              await onResponseHeaderReceived()
+            } catch (callbackError) {
+              logger.error(
+                `❌ Failed to execute onResponseHeaderReceived callback for CCR stream account ${accountId}:`,
+                callbackError.message
+              )
+            }
+          }
+
           // 成功响应，检查并移除错误状态
           ccrAccountService.isAccountRateLimited(accountId).then((isRateLimited) => {
             if (isRateLimited) {
@@ -444,11 +666,20 @@ class CcrRelayService {
           })
 
           // 设置响应头
+          // ⚠️ 关键修复：尊重 auth.js 提前设置的 Connection: close
           if (!responseStream.headersSent) {
+            const existingConnection = responseStream.getHeader
+              ? responseStream.getHeader('Connection')
+              : null
+            if (existingConnection) {
+              logger.debug(
+                `🔌 [CCR Stream] Preserving existing Connection header: ${existingConnection}`
+              )
+            }
             const headers = {
               'Content-Type': 'text/event-stream',
               'Cache-Control': 'no-cache',
-              Connection: 'keep-alive',
+              Connection: existingConnection || 'keep-alive',
               'Access-Control-Allow-Origin': '*',
               'Access-Control-Allow-Headers': 'Cache-Control'
             }
@@ -487,12 +718,17 @@ class CcrRelayService {
                   }
 
                   // 写入到响应流
-                  if (outputLine && !responseStream.destroyed) {
+                  if (outputLine && isStreamWritable(responseStream)) {
                     responseStream.write(`${outputLine}\n`)
+                  } else if (outputLine) {
+                    // 客户端连接已断开，记录警告
+                    logger.warn(
+                      `⚠️ [CCR] Client disconnected during stream, skipping data for account: ${accountId}`
+                    )
                   }
                 } else {
                   // 空行也需要传递
-                  if (!responseStream.destroyed) {
+                  if (isStreamWritable(responseStream)) {
                     responseStream.write('\n')
                   }
                 }
@@ -503,10 +739,6 @@ class CcrRelayService {
           })
 
           response.data.on('end', () => {
-            if (!responseStream.destroyed) {
-              responseStream.end()
-            }
-
             // 如果收集到使用统计数据，调用回调
             if (usageCallback && Object.keys(collectedUsage).length > 0) {
               try {
@@ -518,12 +750,26 @@ class CcrRelayService {
               }
             }
 
+            if (isStreamWritable(responseStream)) {
+              // 等待数据完全 flush 到客户端后再 resolve
+              responseStream.end(() => {
+                logger.debug(
+                  `🌊 CCR stream response completed and flushed | bytesWritten: ${responseStream.bytesWritten || 'unknown'}`
+                )
                 resolve()
               })
+            } else {
+              // 连接已断开，记录警告
+              logger.warn(
+                `⚠️ [CCR] Client disconnected before stream end, data may not have been received | account: ${accountId}`
+              )
+              resolve()
+            }
+          })
 
           response.data.on('error', (err) => {
             logger.error('❌ Stream data error:', err)
-            if (!responseStream.destroyed) {
+            if (isStreamWritable(responseStream)) {
               responseStream.end()
             }
             reject(err)
@@ -555,7 +801,7 @@ class CcrRelayService {
             }
           }
 
-          if (!responseStream.destroyed) {
+          if (isStreamWritable(responseStream)) {
             responseStream.write(`data: ${JSON.stringify(errorResponse)}\n\n`)
             responseStream.end()
           }

src/services/claudeConsoleRelayService.js

@@ -9,6 +9,8 @@ const {
   sanitizeErrorMessage,
   isAccountDisabledError
 } = require('../utils/errorSanitizer')
+const userMessageQueueService = require('./userMessageQueueService')
+const { isStreamWritable } = require('../utils/streamHelper')
 
 class ClaudeConsoleRelayService {
   constructor() {
@@ -29,8 +31,68 @@ class ClaudeConsoleRelayService {
     let account = null
     const requestId = uuidv4() // 用于并发追踪
     let concurrencyAcquired = false
+    let queueLockAcquired = false
+    let queueRequestId = null
 
     try {
+      // 📬 用户消息队列处理：如果是用户消息请求，需要获取队列锁
+      if (userMessageQueueService.isUserMessageRequest(requestBody)) {
+        // 校验 accountId 非空，避免空值污染队列锁键
+        if (!accountId || accountId === '') {
+          logger.error('❌ accountId missing for queue lock in console relayRequest')
+          throw new Error('accountId missing for queue lock')
+        }
+        const queueResult = await userMessageQueueService.acquireQueueLock(accountId)
+        if (!queueResult.acquired && !queueResult.skipped) {
+          // 区分 Redis 后端错误和队列超时
+          const isBackendError = queueResult.error === 'queue_backend_error'
+          const errorCode = isBackendError ? 'QUEUE_BACKEND_ERROR' : 'QUEUE_TIMEOUT'
+          const errorType = isBackendError ? 'queue_backend_error' : 'queue_timeout'
+          const errorMessage = isBackendError
+            ? 'Queue service temporarily unavailable, please retry later'
+            : 'User message queue wait timeout, please retry later'
+          const statusCode = isBackendError ? 500 : 503
+
+          // 结构化性能日志，用于后续统计
+          logger.performance('user_message_queue_error', {
+            errorType,
+            errorCode,
+            accountId,
+            statusCode,
+            apiKeyName: apiKeyData.name,
+            backendError: isBackendError ? queueResult.errorMessage : undefined
+          })
+
+          logger.warn(
+            `📬 User message queue ${errorType} for console account ${accountId}, key: ${apiKeyData.name}`,
+            isBackendError ? { backendError: queueResult.errorMessage } : {}
+          )
+          return {
+            statusCode,
+            headers: {
+              'Content-Type': 'application/json',
+              'x-user-message-queue-error': errorType
+            },
+            body: JSON.stringify({
+              type: 'error',
+              error: {
+                type: errorType,
+                code: errorCode,
+                message: errorMessage
+              }
+            }),
+            accountId
+          }
+        }
+        if (queueResult.acquired && !queueResult.skipped) {
+          queueLockAcquired = true
+          queueRequestId = queueResult.requestId
+          logger.debug(
+            `📬 User message queue lock acquired for console account ${accountId}, requestId: ${queueRequestId}`
+          )
+        }
+      }
+
       // 获取账户信息
       account = await claudeConsoleAccountService.getAccount(accountId)
       if (!account) {
@@ -203,6 +265,23 @@ class ClaudeConsoleRelayService {
       )
       const response = await axios(requestConfig)
 
+      // 📬 请求已发送成功，立即释放队列锁（无需等待响应处理完成）
+      // 因为 Claude API 限流基于请求发送时刻计算（RPM），不是请求完成时刻
+      if (queueLockAcquired && queueRequestId && accountId) {
+        try {
+          await userMessageQueueService.releaseQueueLock(accountId, queueRequestId)
+          queueLockAcquired = false // 标记已释放，防止 finally 重复释放
+          logger.debug(
+            `📬 User message queue lock released early for console account ${accountId}, requestId: ${queueRequestId}`
+          )
+        } catch (releaseError) {
+          logger.error(
+            `❌ Failed to release user message queue lock early for console account ${accountId}:`,
+            releaseError.message
+          )
+        }
+      }
+
       // 移除监听器（请求成功完成）
       if (clientRequest) {
         clientRequest.removeListener('close', handleClientDisconnect)
@@ -366,6 +445,21 @@ class ClaudeConsoleRelayService {
           )
         }
       }
+
+      // 📬 释放用户消息队列锁（兜底，正常情况下已在请求发送后提前释放）
+      if (queueLockAcquired && queueRequestId && accountId) {
+        try {
+          await userMessageQueueService.releaseQueueLock(accountId, queueRequestId)
+          logger.debug(
+            `📬 User message queue lock released in finally for console account ${accountId}, requestId: ${queueRequestId}`
+          )
+        } catch (releaseError) {
+          logger.error(
+            `❌ Failed to release user message queue lock for account ${accountId}:`,
+            releaseError.message
+          )
+        }
+      }
     }
   }
 
@@ -384,8 +478,71 @@ class ClaudeConsoleRelayService {
     const requestId = uuidv4() // 用于并发追踪
     let concurrencyAcquired = false
     let leaseRefreshInterval = null // 租约刷新定时器
+    let queueLockAcquired = false
+    let queueRequestId = null
 
     try {
+      // 📬 用户消息队列处理：如果是用户消息请求，需要获取队列锁
+      if (userMessageQueueService.isUserMessageRequest(requestBody)) {
+        // 校验 accountId 非空，避免空值污染队列锁键
+        if (!accountId || accountId === '') {
+          logger.error(
+            '❌ accountId missing for queue lock in console relayStreamRequestWithUsageCapture'
+          )
+          throw new Error('accountId missing for queue lock')
+        }
+        const queueResult = await userMessageQueueService.acquireQueueLock(accountId)
+        if (!queueResult.acquired && !queueResult.skipped) {
+          // 区分 Redis 后端错误和队列超时
+          const isBackendError = queueResult.error === 'queue_backend_error'
+          const errorCode = isBackendError ? 'QUEUE_BACKEND_ERROR' : 'QUEUE_TIMEOUT'
+          const errorType = isBackendError ? 'queue_backend_error' : 'queue_timeout'
+          const errorMessage = isBackendError
+            ? 'Queue service temporarily unavailable, please retry later'
+            : 'User message queue wait timeout, please retry later'
+          const statusCode = isBackendError ? 500 : 503
+
+          // 结构化性能日志，用于后续统计
+          logger.performance('user_message_queue_error', {
+            errorType,
+            errorCode,
+            accountId,
+            statusCode,
+            stream: true,
+            apiKeyName: apiKeyData.name,
+            backendError: isBackendError ? queueResult.errorMessage : undefined
+          })
+
+          logger.warn(
+            `📬 User message queue ${errorType} for console account ${accountId} (stream), key: ${apiKeyData.name}`,
+            isBackendError ? { backendError: queueResult.errorMessage } : {}
+          )
+          if (!responseStream.headersSent) {
+            const existingConnection = responseStream.getHeader
+              ? responseStream.getHeader('Connection')
+              : null
+            responseStream.writeHead(statusCode, {
+              'Content-Type': 'text/event-stream',
+              'Cache-Control': 'no-cache',
+              Connection: existingConnection || 'keep-alive',
+              'x-user-message-queue-error': errorType
+            })
+          }
+          const errorEvent = `event: error\ndata: ${JSON.stringify({ type: 'error', error: { type: errorType, code: errorCode, message: errorMessage } })}\n\n`
+          responseStream.write(errorEvent)
+          responseStream.write('data: [DONE]\n\n')
+          responseStream.end()
+          return
+        }
+        if (queueResult.acquired && !queueResult.skipped) {
+          queueLockAcquired = true
+          queueRequestId = queueResult.requestId
+          logger.debug(
+            `📬 User message queue lock acquired for console account ${accountId} (stream), requestId: ${queueRequestId}`
+          )
+        }
+      }
+
       // 获取账户信息
       account = await claudeConsoleAccountService.getAccount(accountId)
       if (!account) {
@@ -483,16 +640,40 @@ class ClaudeConsoleRelayService {
         accountId,
         usageCallback,
         streamTransformer,
-        options
+        options,
+        // 📬 回调：在收到响应头时释放队列锁
+        async () => {
+          if (queueLockAcquired && queueRequestId && accountId) {
+            try {
+              await userMessageQueueService.releaseQueueLock(accountId, queueRequestId)
+              queueLockAcquired = false // 标记已释放，防止 finally 重复释放
+              logger.debug(
+                `📬 User message queue lock released early for console stream account ${accountId}, requestId: ${queueRequestId}`
+              )
+            } catch (releaseError) {
+              logger.error(
+                `❌ Failed to release user message queue lock early for console stream account ${accountId}:`,
+                releaseError.message
+              )
+            }
+          }
+        }
       )
 
       // 更新最后使用时间
       await this._updateLastUsedTime(accountId)
     } catch (error) {
+      // 客户端主动断开连接是正常情况，使用 INFO 级别
+      if (error.message === 'Client disconnected') {
+        logger.info(
+          `🔌 Claude Console stream relay ended: Client disconnected (Account: ${account?.name || accountId})`
+        )
+      } else {
         logger.error(
           `❌ Claude Console stream relay failed (Account: ${account?.name || accountId}):`,
           error
         )
+      }
       throw error
     } finally {
       // 🛑 清理租约刷新定时器
@@ -517,6 +698,21 @@ class ClaudeConsoleRelayService {
           )
         }
       }
+
+      // 📬 释放用户消息队列锁（兜底，正常情况下已在收到响应头后提前释放）
+      if (queueLockAcquired && queueRequestId && accountId) {
+        try {
+          await userMessageQueueService.releaseQueueLock(accountId, queueRequestId)
+          logger.debug(
+            `📬 User message queue lock released in finally for console stream account ${accountId}, requestId: ${queueRequestId}`
+          )
+        } catch (releaseError) {
+          logger.error(
+            `❌ Failed to release user message queue lock for stream account ${accountId}:`,
+            releaseError.message
+          )
+        }
+      }
     }
   }
 
@@ -530,7 +726,8 @@ class ClaudeConsoleRelayService {
     accountId,
     usageCallback,
     streamTransformer = null,
-    requestOptions = {}
+    requestOptions = {},
+    onResponseHeaderReceived = null
   ) {
     return new Promise((resolve, reject) => {
       let aborted = false
@@ -593,8 +790,11 @@ class ClaudeConsoleRelayService {
       // 发送请求
       const request = axios(requestConfig)
 
+      // 注意：使用 .then(async ...) 模式处理响应
+      // - 内部的 releaseQueueLock 有独立的 try-catch，不会导致未捕获异常
+      // - queueLockAcquired = false 的赋值会在 finally 执行前完成（JS 单线程保证）
       request
-        .then((response) => {
+        .then(async (response) => {
           logger.debug(`🌊 Claude Console Claude stream response status: ${response.status}`)
 
           // 错误响应处理
@@ -682,7 +882,7 @@ class ClaudeConsoleRelayService {
                   `🧹 [Stream] [SANITIZED] Error response to client: ${JSON.stringify(sanitizedError)}`
                 )
 
-                if (!responseStream.destroyed) {
+                if (isStreamWritable(responseStream)) {
                   responseStream.write(JSON.stringify(sanitizedError))
                   responseStream.end()
                 }
@@ -690,7 +890,7 @@ class ClaudeConsoleRelayService {
                 const sanitizedText = sanitizeErrorMessage(errorDataForCheck)
                 logger.error(`🧹 [Stream] [SANITIZED] Error response to client: ${sanitizedText}`)
 
-                if (!responseStream.destroyed) {
+                if (isStreamWritable(responseStream)) {
                   responseStream.write(sanitizedText)
                   responseStream.end()
                 }
@@ -701,6 +901,19 @@ class ClaudeConsoleRelayService {
             return
           }
 
+          // 📬 收到成功响应头（HTTP 200），调用回调释放队列锁
+          // 此时请求已被 Claude API 接受并计入 RPM 配额，无需等待响应完成
+          if (onResponseHeaderReceived && typeof onResponseHeaderReceived === 'function') {
+            try {
+              await onResponseHeaderReceived()
+            } catch (callbackError) {
+              logger.error(
+                `❌ Failed to execute onResponseHeaderReceived callback for console stream account ${accountId}:`,
+                callbackError.message
+              )
+            }
+          }
+
           // 成功响应，检查并移除错误状态
           claudeConsoleAccountService.isAccountRateLimited(accountId).then((isRateLimited) => {
             if (isRateLimited) {
@@ -714,11 +927,22 @@ class ClaudeConsoleRelayService {
           })
 
           // 设置响应头
+          // ⚠️ 关键修复：尊重 auth.js 提前设置的 Connection: close
+          // 当并发队列功能启用时，auth.js 会设置 Connection: close 来禁用 Keep-Alive
           if (!responseStream.headersSent) {
+            const existingConnection = responseStream.getHeader
+              ? responseStream.getHeader('Connection')
+              : null
+            const connectionHeader = existingConnection || 'keep-alive'
+            if (existingConnection) {
+              logger.debug(
+                `🔌 [Console Stream] Preserving existing Connection header: ${existingConnection}`
+              )
+            }
             responseStream.writeHead(200, {
               'Content-Type': 'text/event-stream',
               'Cache-Control': 'no-cache',
-              Connection: 'keep-alive',
+              Connection: connectionHeader,
               'X-Accel-Buffering': 'no'
             })
           }
@@ -744,20 +968,33 @@ class ClaudeConsoleRelayService {
               buffer = lines.pop() || ''
 
               // 转发数据并解析usage
-              if (lines.length > 0 && !responseStream.destroyed) {
+              if (lines.length > 0) {
+                // 检查流是否可写（客户端连接是否有效）
+                if (isStreamWritable(responseStream)) {
                   const linesToForward = lines.join('\n') + (lines.length > 0 ? '\n' : '')
 
                   // 应用流转换器如果有
+                  let dataToWrite = linesToForward
                   if (streamTransformer) {
                     const transformed = streamTransformer(linesToForward)
                     if (transformed) {
-                    responseStream.write(transformed)
+                      dataToWrite = transformed
-                  }
                     } else {
-                  responseStream.write(linesToForward)
+                      dataToWrite = null
+                    }
                   }
 
-                // 解析SSE数据寻找usage信息
+                  if (dataToWrite) {
+                    responseStream.write(dataToWrite)
+                  }
+                } else {
+                  // 客户端连接已断开，记录警告（但仍继续解析usage）
+                  logger.warn(
+                    `⚠️ [Console] Client disconnected during stream, skipping ${lines.length} lines for account: ${account?.name || accountId}`
+                  )
+                }
+
+                // 解析SSE数据寻找usage信息（无论连接状态如何）
                 for (const line of lines) {
                   if (line.startsWith('data:')) {
                     const jsonStr = line.slice(5).trimStart()
@@ -865,7 +1102,7 @@ class ClaudeConsoleRelayService {
                 `❌ Error processing Claude Console stream data (Account: ${account?.name || accountId}):`,
                 error
               )
-              if (!responseStream.destroyed) {
+              if (isStreamWritable(responseStream)) {
                 // 如果有 streamTransformer（如测试请求），使用前端期望的格式
                 if (streamTransformer) {
                   responseStream.write(
@@ -888,7 +1125,7 @@ class ClaudeConsoleRelayService {
           response.data.on('end', () => {
             try {
               // 处理缓冲区中剩余的数据
-              if (buffer.trim() && !responseStream.destroyed) {
+              if (buffer.trim() && isStreamWritable(responseStream)) {
                 if (streamTransformer) {
                   const transformed = streamTransformer(buffer)
                   if (transformed) {
@@ -937,12 +1174,33 @@ class ClaudeConsoleRelayService {
               }
 
               // 确保流正确结束
-              if (!responseStream.destroyed) {
+              if (isStreamWritable(responseStream)) {
-                responseStream.end()
+                // 📊 诊断日志：流结束前状态
+                logger.info(
+                  `📤 [STREAM] Ending response | destroyed: ${responseStream.destroyed}, ` +
+                    `socketDestroyed: ${responseStream.socket?.destroyed}, ` +
+                    `socketBytesWritten: ${responseStream.socket?.bytesWritten || 0}`
+                )
+
+                // 禁用 Nagle 算法确保数据立即发送
+                if (responseStream.socket && !responseStream.socket.destroyed) {
+                  responseStream.socket.setNoDelay(true)
                 }
 
-              logger.debug('🌊 Claude Console Claude stream response completed')
+                // 等待数据完全 flush 到客户端后再 resolve
+                responseStream.end(() => {
+                  logger.info(
+                    `✅ [STREAM] Response ended and flushed | socketBytesWritten: ${responseStream.socket?.bytesWritten || 'unknown'}`
+                  )
                   resolve()
+                })
+              } else {
+                // 连接已断开，记录警告
+                logger.warn(
+                  `⚠️ [Console] Client disconnected before stream end, data may not have been received | account: ${account?.name || accountId}`
+                )
+                resolve()
+              }
             } catch (error) {
               logger.error('❌ Error processing stream end:', error)
               reject(error)
@@ -954,7 +1212,7 @@ class ClaudeConsoleRelayService {
               `❌ Claude Console stream error (Account: ${account?.name || accountId}):`,
               error
             )
-            if (!responseStream.destroyed) {
+            if (isStreamWritable(responseStream)) {
               // 如果有 streamTransformer（如测试请求），使用前端期望的格式
               if (streamTransformer) {
                 responseStream.write(
@@ -1002,14 +1260,17 @@ class ClaudeConsoleRelayService {
 
           // 发送错误响应
           if (!responseStream.headersSent) {
+            const existingConnection = responseStream.getHeader
+              ? responseStream.getHeader('Connection')
+              : null
             responseStream.writeHead(error.response?.status || 500, {
               'Content-Type': 'text/event-stream',
               'Cache-Control': 'no-cache',
-              Connection: 'keep-alive'
+              Connection: existingConnection || 'keep-alive'
             })
           }
 
-          if (!responseStream.destroyed) {
+          if (isStreamWritable(responseStream)) {
             // 如果有 streamTransformer（如测试请求），使用前端期望的格式
             if (streamTransformer) {
               responseStream.write(
@@ -1179,7 +1440,7 @@ class ClaudeConsoleRelayService {
           'Cache-Control': 'no-cache'
         })
       }
-      if (!responseStream.destroyed && !responseStream.writableEnded) {
+      if (isStreamWritable(responseStream)) {
         responseStream.write(
           `data: ${JSON.stringify({ type: 'test_complete', success: false, error: error.message })}\n\n`
         )

src/services/claudeRelayConfigService.js

@@ -15,6 +15,20 @@ const DEFAULT_CONFIG = {
   globalSessionBindingEnabled: false,
   sessionBindingErrorMessage: '你的本地session已污染，请清理后使用。',
   sessionBindingTtlDays: 30, // 会话绑定 TTL（天），默认30天
+  // 用户消息队列配置
+  userMessageQueueEnabled: false, // 是否启用用户消息队列（默认关闭）
+  userMessageQueueDelayMs: 200, // 请求间隔（毫秒）
+  userMessageQueueTimeoutMs: 5000, // 队列等待超时（毫秒），优化后锁持有时间短无需长等待
+  userMessageQueueLockTtlMs: 5000, // 锁TTL（毫秒），请求发送后立即释放无需长TTL
+  // 并发请求排队配置
+  concurrentRequestQueueEnabled: false, // 是否启用并发请求排队（默认关闭）
+  concurrentRequestQueueMaxSize: 3, // 固定最小排队数（默认3）
+  concurrentRequestQueueMaxSizeMultiplier: 0, // 并发数的倍数（默认0，仅使用固定值）
+  concurrentRequestQueueTimeoutMs: 10000, // 排队超时（毫秒，默认10秒）
+  concurrentRequestQueueMaxRedisFailCount: 5, // 连续 Redis 失败阈值（默认5次）
+  // 排队健康检查配置
+  concurrentRequestQueueHealthCheckEnabled: true, // 是否启用排队健康检查（默认开启）
+  concurrentRequestQueueHealthThreshold: 0.8, // 健康检查阈值（P90 >= 超时 × 阈值时拒绝新请求）
   updatedAt: null,
   updatedBy: null
 }
@@ -100,7 +114,8 @@ class ClaudeRelayConfigService {
 
       logger.info(`✅ Claude relay config updated by ${updatedBy}:`, {
         claudeCodeOnlyEnabled: updatedConfig.claudeCodeOnlyEnabled,
-        globalSessionBindingEnabled: updatedConfig.globalSessionBindingEnabled
+        globalSessionBindingEnabled: updatedConfig.globalSessionBindingEnabled,
+        concurrentRequestQueueEnabled: updatedConfig.concurrentRequestQueueEnabled
       })
 
       return updatedConfig
@@ -316,11 +331,11 @@ class ClaudeRelayConfigService {
 
   /**
    * 验证新会话请求
-   * @param {Object} requestBody - 请求体
+   * @param {Object} _requestBody - 请求体（预留参数，当前未使用）
    * @param {string} originalSessionId - 原始会话ID
    * @returns {Promise<Object>} { valid: boolean, error?: string, binding?: object, isNewSession?: boolean }
    */
-  async validateNewSession(requestBody, originalSessionId) {
+  async validateNewSession(_requestBody, originalSessionId) {
     const cfg = await this.getConfig()
 
     if (!cfg.globalSessionBindingEnabled) {

src/services/claudeRelayService.js

@@ -15,6 +15,8 @@ const ClaudeCodeValidator = require('../validators/clients/claudeCodeValidator')
 const { formatDateWithTimezone } = require('../utils/dateHelper')
 const requestIdentityService = require('./requestIdentityService')
 const { createClaudeTestPayload } = require('../utils/testPayloadHelper')
+const userMessageQueueService = require('./userMessageQueueService')
+const { isStreamWritable } = require('../utils/streamHelper')
 
 class ClaudeRelayService {
   constructor() {
@@ -148,6 +150,9 @@ class ClaudeRelayService {
     options = {}
   ) {
     let upstreamRequest = null
+    let queueLockAcquired = false
+    let queueRequestId = null
+    let selectedAccountId = null
 
     try {
       // 调试日志：查看API Key数据
@@ -192,11 +197,70 @@ class ClaudeRelayService {
       }
       const { accountId } = accountSelection
       const { accountType } = accountSelection
+      selectedAccountId = accountId
 
       logger.info(
         `📤 Processing API request for key: ${apiKeyData.name || apiKeyData.id}, account: ${accountId} (${accountType})${sessionHash ? `, session: ${sessionHash}` : ''}`
       )
 
+      // 📬 用户消息队列处理：如果是用户消息请求，需要获取队列锁
+      if (userMessageQueueService.isUserMessageRequest(requestBody)) {
+        // 校验 accountId 非空，避免空值污染队列锁键
+        if (!accountId || accountId === '') {
+          logger.error('❌ accountId missing for queue lock in relayRequest')
+          throw new Error('accountId missing for queue lock')
+        }
+        const queueResult = await userMessageQueueService.acquireQueueLock(accountId)
+        if (!queueResult.acquired && !queueResult.skipped) {
+          // 区分 Redis 后端错误和队列超时
+          const isBackendError = queueResult.error === 'queue_backend_error'
+          const errorCode = isBackendError ? 'QUEUE_BACKEND_ERROR' : 'QUEUE_TIMEOUT'
+          const errorType = isBackendError ? 'queue_backend_error' : 'queue_timeout'
+          const errorMessage = isBackendError
+            ? 'Queue service temporarily unavailable, please retry later'
+            : 'User message queue wait timeout, please retry later'
+          const statusCode = isBackendError ? 500 : 503
+
+          // 结构化性能日志，用于后续统计
+          logger.performance('user_message_queue_error', {
+            errorType,
+            errorCode,
+            accountId,
+            statusCode,
+            apiKeyName: apiKeyData.name,
+            backendError: isBackendError ? queueResult.errorMessage : undefined
+          })
+
+          logger.warn(
+            `📬 User message queue ${errorType} for account ${accountId}, key: ${apiKeyData.name}`,
+            isBackendError ? { backendError: queueResult.errorMessage } : {}
+          )
+          return {
+            statusCode,
+            headers: {
+              'Content-Type': 'application/json',
+              'x-user-message-queue-error': errorType
+            },
+            body: JSON.stringify({
+              type: 'error',
+              error: {
+                type: errorType,
+                code: errorCode,
+                message: errorMessage
+              }
+            }),
+            accountId
+          }
+        }
+        if (queueResult.acquired && !queueResult.skipped) {
+          queueLockAcquired = true
+          queueRequestId = queueResult.requestId
+          logger.debug(
+            `📬 User message queue lock acquired for account ${accountId}, requestId: ${queueRequestId}`
+          )
+        }
+      }
+
       // 获取账户信息
       let account = await claudeAccountService.getAccount(accountId)
 
@@ -271,6 +335,23 @@ class ClaudeRelayService {
         options
       )
 
+      // 📬 请求已发送成功，立即释放队列锁（无需等待响应处理完成）
+      // 因为 Claude API 限流基于请求发送时刻计算（RPM），不是请求完成时刻
+      if (queueLockAcquired && queueRequestId && selectedAccountId) {
+        try {
+          await userMessageQueueService.releaseQueueLock(selectedAccountId, queueRequestId)
+          queueLockAcquired = false // 标记已释放，防止 finally 重复释放
+          logger.debug(
+            `📬 User message queue lock released early for account ${selectedAccountId}, requestId: ${queueRequestId}`
+          )
+        } catch (releaseError) {
+          logger.error(
+            `❌ Failed to release user message queue lock early for account ${selectedAccountId}:`,
+            releaseError.message
+          )
+        }
+      }
+
       response.accountId = accountId
       response.accountType = accountType
 
@@ -539,6 +620,21 @@ class ClaudeRelayService {
         error.message
       )
       throw error
+    } finally {
+      // 📬 释放用户消息队列锁（兜底，正常情况下已在请求发送后提前释放）
+      if (queueLockAcquired && queueRequestId && selectedAccountId) {
+        try {
+          await userMessageQueueService.releaseQueueLock(selectedAccountId, queueRequestId)
+          logger.debug(
+            `📬 User message queue lock released in finally for account ${selectedAccountId}, requestId: ${queueRequestId}`
+          )
+        } catch (releaseError) {
+          logger.error(
+            `❌ Failed to release user message queue lock for account ${selectedAccountId}:`,
+            releaseError.message
+          )
+        }
+      }
     }
   }
 
@@ -962,6 +1058,8 @@ class ClaudeRelayService {
 
     logger.info(`🔗 指纹是这个: ${headers['User-Agent']}`)
 
+    logger.info(`🔗 指纹是这个: ${headers['User-Agent']}`)
+
     // 根据模型和客户端传递的 anthropic-beta 动态设置 header
     const modelId = requestPayload?.model || body?.model
     const clientBetaHeader = clientHeaders?.['anthropic-beta']
@@ -1057,8 +1155,6 @@ class ClaudeRelayService {
         timeout: config.requestTimeout || 600000
       }
 
-      console.log(options.path)
-
       const req = https.request(options, (res) => {
         let responseData = Buffer.alloc(0)
 
@@ -1112,7 +1208,6 @@ class ClaudeRelayService {
       }
 
       req.on('error', async (error) => {
-        console.error(': ❌ ', error)
         logger.error(`❌ Claude API request error (Account: ${accountId}):`, error.message, {
           code: error.code,
           errno: error.errno,
@@ -1163,6 +1258,10 @@ class ClaudeRelayService {
     streamTransformer = null,
     options = {}
   ) {
+    let queueLockAcquired = false
+    let queueRequestId = null
+    let selectedAccountId = null
+
     try {
       // 调试日志：查看API Key数据（流式请求）
       logger.info('🔍 [Stream] API Key data received:', {
@@ -1206,6 +1305,73 @@ class ClaudeRelayService {
       }
       const { accountId } = accountSelection
       const { accountType } = accountSelection
+      selectedAccountId = accountId
+
+      // 📬 用户消息队列处理：如果是用户消息请求，需要获取队列锁
+      if (userMessageQueueService.isUserMessageRequest(requestBody)) {
+        // 校验 accountId 非空，避免空值污染队列锁键
+        if (!accountId || accountId === '') {
+          logger.error('❌ accountId missing for queue lock in relayStreamRequestWithUsageCapture')
+          throw new Error('accountId missing for queue lock')
+        }
+        const queueResult = await userMessageQueueService.acquireQueueLock(accountId)
+        if (!queueResult.acquired && !queueResult.skipped) {
+          // 区分 Redis 后端错误和队列超时
+          const isBackendError = queueResult.error === 'queue_backend_error'
+          const errorCode = isBackendError ? 'QUEUE_BACKEND_ERROR' : 'QUEUE_TIMEOUT'
+          const errorType = isBackendError ? 'queue_backend_error' : 'queue_timeout'
+          const errorMessage = isBackendError
+            ? 'Queue service temporarily unavailable, please retry later'
+            : 'User message queue wait timeout, please retry later'
+          const statusCode = isBackendError ? 500 : 503
+
+          // 结构化性能日志，用于后续统计
+          logger.performance('user_message_queue_error', {
+            errorType,
+            errorCode,
+            accountId,
+            statusCode,
+            stream: true,
+            apiKeyName: apiKeyData.name,
+            backendError: isBackendError ? queueResult.errorMessage : undefined
+          })
+
+          logger.warn(
+            `📬 User message queue ${errorType} for account ${accountId} (stream), key: ${apiKeyData.name}`,
+            isBackendError ? { backendError: queueResult.errorMessage } : {}
+          )
+          if (!responseStream.headersSent) {
+            const existingConnection = responseStream.getHeader
+              ? responseStream.getHeader('Connection')
+              : null
+            responseStream.writeHead(statusCode, {
+              'Content-Type': 'text/event-stream',
+              'Cache-Control': 'no-cache',
+              Connection: existingConnection || 'keep-alive',
+              'x-user-message-queue-error': errorType
+            })
+          }
+          const errorEvent = `event: error\ndata: ${JSON.stringify({
+            type: 'error',
+            error: {
+              type: errorType,
+              code: errorCode,
+              message: errorMessage
+            }
+          })}\n\n`
+          responseStream.write(errorEvent)
+          responseStream.write('data: [DONE]\n\n')
+          responseStream.end()
+          return
+        }
+        if (queueResult.acquired && !queueResult.skipped) {
+          queueLockAcquired = true
+          queueRequestId = queueResult.requestId
+          logger.debug(
+            `📬 User message queue lock acquired for account ${accountId} (stream), requestId: ${queueRequestId}`
+          )
+        }
+      }
 
       logger.info(
         `📡 Processing streaming API request with usage capture for key: ${apiKeyData.name || apiKeyData.id}, account: ${accountId} (${accountType})${sessionHash ? `, session: ${sessionHash}` : ''}`
@@ -1272,11 +1438,48 @@ class ClaudeRelayService {
         sessionHash,
         streamTransformer,
         options,
-        isDedicatedOfficialAccount
+        isDedicatedOfficialAccount,
+        // 📬 新增回调：在收到响应头时释放队列锁
+        async () => {
+          if (queueLockAcquired && queueRequestId && selectedAccountId) {
+            try {
+              await userMessageQueueService.releaseQueueLock(selectedAccountId, queueRequestId)
+              queueLockAcquired = false // 标记已释放，防止 finally 重复释放
+              logger.debug(
+                `📬 User message queue lock released early for stream account ${selectedAccountId}, requestId: ${queueRequestId}`
+              )
+            } catch (releaseError) {
+              logger.error(
+                `❌ Failed to release user message queue lock early for stream account ${selectedAccountId}:`,
+                releaseError.message
+              )
+            }
+          }
+        }
       )
     } catch (error) {
+      // 客户端主动断开连接是正常情况，使用 INFO 级别
+      if (error.message === 'Client disconnected') {
+        logger.info(`🔌 Claude stream relay ended: Client disconnected`)
+      } else {
         logger.error(`❌ Claude stream relay with usage capture failed:`, error)
+      }
       throw error
+    } finally {
+      // 📬 释放用户消息队列锁（兜底，正常情况下已在收到响应头后提前释放）
+      if (queueLockAcquired && queueRequestId && selectedAccountId) {
+        try {
+          await userMessageQueueService.releaseQueueLock(selectedAccountId, queueRequestId)
+          logger.debug(
+            `📬 User message queue lock released in finally for stream account ${selectedAccountId}, requestId: ${queueRequestId}`
+          )
+        } catch (releaseError) {
+          logger.error(
+            `❌ Failed to release user message queue lock for stream account ${selectedAccountId}:`,
+            releaseError.message
+          )
+        }
+      }
     }
   }
 
@@ -1293,7 +1496,8 @@ class ClaudeRelayService {
     sessionHash,
     streamTransformer = null,
     requestOptions = {},
-    isDedicatedOfficialAccount = false
+    isDedicatedOfficialAccount = false,
+    onResponseStart = null // 📬 新增：收到响应头时的回调，用于提前释放队列锁
   ) {
     // 获取账户信息用于统一 User-Agent
     const account = await claudeAccountService.getAccount(accountId)
@@ -1478,7 +1682,6 @@ class ClaudeRelayService {
           })
 
           res.on('end', () => {
-            console.error(': ❌ ', errorData)
             logger.error(
               `❌ Claude API error response (Account: ${account?.name || accountId}):`,
               errorData
@@ -1502,7 +1705,7 @@ class ClaudeRelayService {
                 }
               })()
             }
-            if (!responseStream.destroyed) {
+            if (isStreamWritable(responseStream)) {
               // 解析 Claude API 返回的错误详情
               let errorMessage = `Claude API error: ${res.statusCode}`
               try {
@@ -1540,6 +1743,16 @@ class ClaudeRelayService {
           return
         }
 
+        // 📬 收到成功响应头（HTTP 200），立即调用回调释放队列锁
+        // 此时请求已被 Claude API 接受并计入 RPM 配额，无需等待响应完成
+        if (onResponseStart && typeof onResponseStart === 'function') {
+          try {
+            await onResponseStart()
+          } catch (callbackError) {
+            logger.error('❌ Error in onResponseStart callback:', callbackError.message)
+          }
+        }
+
         let buffer = ''
         const allUsageData = [] // 收集所有的usage事件
         let currentUsageData = {} // 当前正在收集的usage数据
@@ -1557,7 +1770,8 @@ class ClaudeRelayService {
             buffer = lines.pop() || '' // 保留最后的不完整行
 
             // 转发已处理的完整行到客户端
-            if (lines.length > 0 && !responseStream.destroyed) {
+            if (lines.length > 0) {
+              if (isStreamWritable(responseStream)) {
                 const linesToForward = lines.join('\n') + (lines.length > 0 ? '\n' : '')
                 // 如果有流转换器，应用转换
                 if (streamTransformer) {
@@ -1568,6 +1782,12 @@ class ClaudeRelayService {
                 } else {
                   responseStream.write(linesToForward)
                 }
+              } else {
+                // 客户端连接已断开，记录警告（但仍继续解析usage）
+                logger.warn(
+                  `⚠️ [Official] Client disconnected during stream, skipping ${lines.length} lines for account: ${accountId}`
+                )
+              }
             }
 
             for (const line of lines) {
@@ -1671,7 +1891,7 @@ class ClaudeRelayService {
           } catch (error) {
             logger.error('❌ Error processing stream data:', error)
             // 发送错误但不破坏流，让它自然结束
-            if (!responseStream.destroyed) {
+            if (isStreamWritable(responseStream)) {
               responseStream.write('event: error\n')
               responseStream.write(
                 `data: ${JSON.stringify({
@@ -1687,7 +1907,7 @@ class ClaudeRelayService {
         res.on('end', async () => {
           try {
             // 处理缓冲区中剩余的数据
-            if (buffer.trim() && !responseStream.destroyed) {
+            if (buffer.trim() && isStreamWritable(responseStream)) {
               if (streamTransformer) {
                 const transformed = streamTransformer(buffer)
                 if (transformed) {
@@ -1699,8 +1919,16 @@ class ClaudeRelayService {
             }
 
             // 确保流正确结束
-            if (!responseStream.destroyed) {
+            if (isStreamWritable(responseStream)) {
               responseStream.end()
+              logger.debug(
+                `🌊 Stream end called | bytesWritten: ${responseStream.bytesWritten || 'unknown'}`
+              )
+            } else {
+              // 连接已断开，记录警告
+              logger.warn(
+                `⚠️ [Official] Client disconnected before stream end, data may not have been received | account: ${account?.name || accountId}`
+              )
             }
           } catch (error) {
             logger.error('❌ Error processing stream end:', error)
@@ -1898,14 +2126,17 @@ class ClaudeRelayService {
         }
 
         if (!responseStream.headersSent) {
+          const existingConnection = responseStream.getHeader
+            ? responseStream.getHeader('Connection')
+            : null
           responseStream.writeHead(statusCode, {
             'Content-Type': 'text/event-stream',
             'Cache-Control': 'no-cache',
-            Connection: 'keep-alive'
+            Connection: existingConnection || 'keep-alive'
           })
         }
 
-        if (!responseStream.destroyed) {
+        if (isStreamWritable(responseStream)) {
           // 发送 SSE 错误事件
           responseStream.write('event: error\n')
           responseStream.write(
@@ -1925,13 +2156,16 @@ class ClaudeRelayService {
         logger.error(`❌ Claude stream request timeout | Account: ${account?.name || accountId}`)
 
         if (!responseStream.headersSent) {
+          const existingConnection = responseStream.getHeader
+            ? responseStream.getHeader('Connection')
+            : null
           responseStream.writeHead(504, {
             'Content-Type': 'text/event-stream',
             'Cache-Control': 'no-cache',
-            Connection: 'keep-alive'
+            Connection: existingConnection || 'keep-alive'
           })
         }
-        if (!responseStream.destroyed) {
+        if (isStreamWritable(responseStream)) {
           // 发送 SSE 错误事件
           responseStream.write('event: error\n')
           responseStream.write(
@@ -1950,7 +2184,7 @@ class ClaudeRelayService {
       responseStream.on('close', () => {
         logger.debug('🔌 Client disconnected, cleaning up stream')
         if (!req.destroyed) {
-          req.destroy()
+          req.destroy(new Error('Client disconnected'))
         }
       })
 
@@ -2246,10 +2480,13 @@ class ClaudeRelayService {
 
       // 设置响应头
       if (!responseStream.headersSent) {
+        const existingConnection = responseStream.getHeader
+          ? responseStream.getHeader('Connection')
+          : null
         responseStream.writeHead(200, {
           'Content-Type': 'text/event-stream',
           'Cache-Control': 'no-cache',
-          Connection: 'keep-alive',
+          Connection: existingConnection || 'keep-alive',
           'X-Accel-Buffering': 'no'
         })
       }
@@ -2277,7 +2514,7 @@ class ClaudeRelayService {
     } catch (error) {
       logger.error(`❌ Test account connection failed:`, error)
       // 发送错误事件给前端
-      if (!responseStream.destroyed && !responseStream.writableEnded) {
+      if (isStreamWritable(responseStream)) {
         try {
           const errorMsg = error.message || '测试失败'
           responseStream.write(`data: ${JSON.stringify({ type: 'error', error: errorMsg })}\n\n`)

src/services/droidRelayService.js

@@ -336,7 +336,12 @@ class DroidRelayService {
         )
       }
     } catch (error) {
+      // 客户端主动断开连接是正常情况，使用 INFO 级别
+      if (error.message === 'Client disconnected') {
+        logger.info(`🔌 Droid relay ended: Client disconnected`)
+      } else {
         logger.error(`❌ Droid relay error: ${error.message}`, error)
+      }
 
       const status = error?.response?.status
       if (status >= 400 && status < 500) {
@@ -634,7 +639,7 @@ class DroidRelayService {
       // 客户端断开连接时清理
       clientResponse.on('close', () => {
         if (req && !req.destroyed) {
-          req.destroy()
+          req.destroy(new Error('Client disconnected'))
         }
       })
 

src/services/openaiResponsesRelayService.js

@@ -426,9 +426,9 @@ class OpenAIResponsesRelayService {
       const lines = data.split('\n')
 
       for (const line of lines) {
-        if (line.startsWith('data: ')) {
+        if (line.startsWith('data:')) {
           try {
-            const jsonStr = line.slice(6)
+            const jsonStr = line.slice(5).trim()
             if (jsonStr === '[DONE]') {
               continue
             }

src/services/userMessageQueueService.js

@@ -0,0 +1,348 @@
+/**
+ * 用户消息队列服务
+ * 为 Claude 账户实现基于消息类型的串行排队机制
+ *
+ * 当请求的最后一条消息是用户输入（role: user）时，
+ * 同一账户的此类请求需要串行等待，并在请求之间添加延迟
+ */
+
+const { v4: uuidv4 } = require('uuid')
+const redis = require('../models/redis')
+const config = require('../../config/config')
+const logger = require('../utils/logger')
+
+// 清理任务间隔
+const CLEANUP_INTERVAL_MS = 60000 // 1分钟
+
+// 轮询等待配置
+const POLL_INTERVAL_BASE_MS = 50 // 基础轮询间隔
+const POLL_INTERVAL_MAX_MS = 500 // 最大轮询间隔
+const POLL_BACKOFF_FACTOR = 1.5 // 退避因子
+
+class UserMessageQueueService {
+  constructor() {
+    this.cleanupTimer = null
+  }
+
+  /**
+   * 检测请求是否为真正的用户消息请求
+   * 区分真正的用户输入和 tool_result 消息
+   *
+   * Claude API 消息格式：
+   * - 用户文本消息: { role: 'user', content: 'text' } 或 { role: 'user', content: [{ type: 'text', text: '...' }] }
+   * - 工具结果消息: { role: 'user', content: [{ type: 'tool_result', tool_use_id: '...', content: '...' }] }
+   *
+   * @param {Object} requestBody - 请求体
+   * @returns {boolean} - 是否为真正的用户消息（排除 tool_result）
+   */
+  isUserMessageRequest(requestBody) {
+    const messages = requestBody?.messages
+    if (!Array.isArray(messages) || messages.length === 0) {
+      return false
+    }
+    const lastMessage = messages[messages.length - 1]
+
+    // 检查 role 是否为 user
+    if (lastMessage?.role !== 'user') {
+      return false
+    }
+
+    // 检查 content 是否包含 tool_result 类型
+    const { content } = lastMessage
+    if (Array.isArray(content)) {
+      // 如果 content 数组中任何元素是 tool_result，则不是真正的用户消息
+      const hasToolResult = content.some(
+        (block) => block?.type === 'tool_result' || block?.type === 'tool_use_result'
+      )
+      if (hasToolResult) {
+        return false
+      }
+    }
+
+    // role 是 user 且不包含 tool_result，是真正的用户消息
+    return true
+  }
+
+  /**
+   * 获取当前配置（支持 Web 界面配置优先）
+   * @returns {Promise<Object>} 配置对象
+   */
+  async getConfig() {
+    // 默认配置（防止 config.userMessageQueue 未定义）
+    // 注意：优化后的默认值 - 锁持有时间从分钟级降到毫秒级，无需长等待
+    const queueConfig = config.userMessageQueue || {}
+    const defaults = {
+      enabled: queueConfig.enabled ?? false,
+      delayMs: queueConfig.delayMs ?? 200,
+      timeoutMs: queueConfig.timeoutMs ?? 5000, // 从 60000 降到 5000，因为锁持有时间短
+      lockTtlMs: queueConfig.lockTtlMs ?? 5000 // 从 120000 降到 5000，5秒足以覆盖请求发送
+    }
+
+    // 尝试从 claudeRelayConfigService 获取 Web 界面配置
+    try {
+      const claudeRelayConfigService = require('./claudeRelayConfigService')
+      const webConfig = await claudeRelayConfigService.getConfig()
+
+      return {
+        enabled:
+          webConfig.userMessageQueueEnabled !== undefined
+            ? webConfig.userMessageQueueEnabled
+            : defaults.enabled,
+        delayMs:
+          webConfig.userMessageQueueDelayMs !== undefined
+            ? webConfig.userMessageQueueDelayMs
+            : defaults.delayMs,
+        timeoutMs:
+          webConfig.userMessageQueueTimeoutMs !== undefined
+            ? webConfig.userMessageQueueTimeoutMs
+            : defaults.timeoutMs,
+        lockTtlMs:
+          webConfig.userMessageQueueLockTtlMs !== undefined
+            ? webConfig.userMessageQueueLockTtlMs
+            : defaults.lockTtlMs
+      }
+    } catch {
+      // 回退到环境变量配置
+      return defaults
+    }
+  }
+
+  /**
+   * 检查功能是否启用
+   * @returns {Promise<boolean>}
+   */
+  async isEnabled() {
+    const cfg = await this.getConfig()
+    return cfg.enabled === true
+  }
+
+  /**
+   * 获取账户队列锁（阻塞等待）
+   * @param {string} accountId - 账户ID
+   * @param {string} requestId - 请求ID（可选，会自动生成）
+   * @param {number} timeoutMs - 超时时间（可选，使用配置默认值）
+   * @returns {Promise<{acquired: boolean, requestId: string, error?: string}>}
+   */
+  async acquireQueueLock(accountId, requestId = null, timeoutMs = null) {
+    const cfg = await this.getConfig()
+
+    if (!cfg.enabled) {
+      return { acquired: true, requestId: requestId || uuidv4(), skipped: true }
+    }
+
+    const reqId = requestId || uuidv4()
+    const timeout = timeoutMs || cfg.timeoutMs
+    const startTime = Date.now()
+    let retryCount = 0
+
+    logger.debug(`📬 User message queue: attempting to acquire lock for account ${accountId}`, {
+      requestId: reqId,
+      timeoutMs: timeout
+    })
+
+    while (Date.now() - startTime < timeout) {
+      const result = await redis.acquireUserMessageLock(
+        accountId,
+        reqId,
+        cfg.lockTtlMs,
+        cfg.delayMs
+      )
+
+      // 检测 Redis 错误，立即返回系统错误而非继续轮询
+      if (result.redisError) {
+        logger.error(`📬 User message queue: Redis error while acquiring lock`, {
+          accountId,
+          requestId: reqId,
+          errorMessage: result.errorMessage
+        })
+        return {
+          acquired: false,
+          requestId: reqId,
+          error: 'queue_backend_error',
+          errorMessage: result.errorMessage
+        }
+      }
+
+      if (result.acquired) {
+        logger.debug(`📬 User message queue: lock acquired for account ${accountId}`, {
+          requestId: reqId,
+          waitedMs: Date.now() - startTime,
+          retries: retryCount
+        })
+        return { acquired: true, requestId: reqId }
+      }
+
+      // 需要等待
+      if (result.waitMs > 0) {
+        // 需要延迟（上一个请求刚完成）
+        await this._sleep(Math.min(result.waitMs, timeout - (Date.now() - startTime)))
+      } else {
+        // 锁被占用，使用指数退避轮询等待
+        const basePollInterval = Math.min(
+          POLL_INTERVAL_BASE_MS * Math.pow(POLL_BACKOFF_FACTOR, retryCount),
+          POLL_INTERVAL_MAX_MS
+        )
+        // 添加 ±15% 随机抖动，避免高并发下的周期性碰撞
+        const jitter = basePollInterval * (0.85 + Math.random() * 0.3)
+        const pollInterval = Math.min(jitter, POLL_INTERVAL_MAX_MS)
+        await this._sleep(pollInterval)
+        retryCount++
+      }
+    }
+
+    // 超时
+    logger.warn(`📬 User message queue: timeout waiting for lock`, {
+      accountId,
+      requestId: reqId,
+      timeoutMs: timeout
+    })
+
+    return {
+      acquired: false,
+      requestId: reqId,
+      error: 'queue_timeout'
+    }
+  }
+
+  /**
+   * 释放账户队列锁
+   * @param {string} accountId - 账户ID
+   * @param {string} requestId - 请求ID
+   * @returns {Promise<boolean>}
+   */
+  async releaseQueueLock(accountId, requestId) {
+    if (!accountId || !requestId) {
+      return false
+    }
+
+    const released = await redis.releaseUserMessageLock(accountId, requestId)
+
+    if (released) {
+      logger.debug(`📬 User message queue: lock released for account ${accountId}`, {
+        requestId
+      })
+    } else {
+      logger.warn(`📬 User message queue: failed to release lock (not owner?)`, {
+        accountId,
+        requestId
+      })
+    }
+
+    return released
+  }
+
+  /**
+   * 获取队列统计信息
+   * @param {string} accountId - 账户ID
+   * @returns {Promise<Object>}
+   */
+  async getQueueStats(accountId) {
+    return await redis.getUserMessageQueueStats(accountId)
+  }
+
+  /**
+   * 服务启动时清理所有残留的队列锁
+   * 防止服务重启后旧锁阻塞新请求
+   * @returns {Promise<number>} 清理的锁数量
+   */
+  async cleanupStaleLocks() {
+    try {
+      const accountIds = await redis.scanUserMessageQueueLocks()
+      let cleanedCount = 0
+
+      for (const accountId of accountIds) {
+        try {
+          await redis.forceReleaseUserMessageLock(accountId)
+          cleanedCount++
+          logger.debug(`📬 User message queue: cleaned stale lock for account ${accountId}`)
+        } catch (error) {
+          logger.error(
+            `📬 User message queue: failed to clean lock for account ${accountId}:`,
+            error
+          )
+        }
+      }
+
+      if (cleanedCount > 0) {
+        logger.info(`📬 User message queue: cleaned ${cleanedCount} stale lock(s) on startup`)
+      }
+
+      return cleanedCount
+    } catch (error) {
+      logger.error('📬 User message queue: failed to cleanup stale locks on startup:', error)
+      return 0
+    }
+  }
+
+  /**
+   * 启动定时清理任务
+   * 始终启动，每次执行时检查配置以支持运行时动态启用/禁用
+   */
+  startCleanupTask() {
+    if (this.cleanupTimer) {
+      return
+    }
+
+    this.cleanupTimer = setInterval(async () => {
+      // 每次运行时检查配置，以便在运行时动态启用/禁用
+      const currentConfig = await this.getConfig()
+      if (!currentConfig.enabled) {
+        logger.debug('📬 User message queue: cleanup skipped (feature disabled)')
+        return
+      }
+      await this._cleanupOrphanLocks()
+    }, CLEANUP_INTERVAL_MS)
+
+    logger.info('📬 User message queue: cleanup task started')
+  }
+
+  /**
+   * 停止定时清理任务
+   */
+  stopCleanupTask() {
+    if (this.cleanupTimer) {
+      clearInterval(this.cleanupTimer)
+      this.cleanupTimer = null
+      logger.info('📬 User message queue: cleanup task stopped')
+    }
+  }
+
+  /**
+   * 清理孤儿锁
+   * 检测异常情况：锁存在但没有设置过期时间（lockTtlRaw === -1）
+   * 正常情况下所有锁都应该有 TTL，Redis 会自动过期
+   * @private
+   */
+  async _cleanupOrphanLocks() {
+    try {
+      const accountIds = await redis.scanUserMessageQueueLocks()
+
+      for (const accountId of accountIds) {
+        const stats = await redis.getUserMessageQueueStats(accountId)
+
+        // 检测异常情况：锁存在（isLocked=true）但没有过期时间（lockTtlRaw=-1）
+        // 正常创建的锁都带有 PX 过期时间，如果没有说明是异常状态
+        if (stats.isLocked && stats.lockTtlRaw === -1) {
+          logger.warn(
+            `📬 User message queue: cleaning up orphan lock without TTL for account ${accountId}`,
+            { lockHolder: stats.lockHolder }
+          )
+          await redis.forceReleaseUserMessageLock(accountId)
+        }
+      }
+    } catch (error) {
+      logger.error('📬 User message queue: cleanup task error:', error)
+    }
+  }
+
+  /**
+   * 睡眠辅助函数
+   * @param {number} ms - 毫秒
+   * @private
+   */
+  _sleep(ms) {
+    return new Promise((resolve) => setTimeout(resolve, ms))
+  }
+}
+
+module.exports = new UserMessageQueueService()

src/utils/logger.js

@@ -137,6 +137,7 @@ const createLogFormat = (colorize = false) => {
 
 const logFormat = createLogFormat(false)
 const consoleFormat = createLogFormat(true)
+const isTestEnv = process.env.NODE_ENV === 'test' || process.env.JEST_WORKER_ID
 
 // 📁 确保日志目录存在并设置权限
 if (!fs.existsSync(config.logging.dirname)) {
@@ -159,7 +160,8 @@ const createRotateTransport = (filename, level = null) => {
     transport.level = level
   }
 
-  // 监听轮转事件
+  // 监听轮转事件（测试环境关闭以避免 Jest 退出后输出）
+  if (!isTestEnv) {
     transport.on('rotate', (oldFilename, newFilename) => {
       console.log(`📦 Log rotated: ${oldFilename} -> ${newFilename}`)
     })
@@ -171,6 +173,7 @@ const createRotateTransport = (filename, level = null) => {
     transport.on('archive', (zipFilename) => {
       console.log(`🗜️ Log archived: ${zipFilename}`)
     })
+  }
 
   return transport
 }

src/utils/statsHelper.js

@@ -0,0 +1,105 @@
+/**
+ * 统计计算工具函数
+ * 提供百分位数计算、等待时间统计等通用统计功能
+ */
+
+/**
+ * 计算百分位数（使用 nearest-rank 方法）
+ * @param {number[]} sortedArray - 已排序的数组（升序）
+ * @param {number} percentile - 百分位数 (0-100)
+ * @returns {number} 百分位值
+ *
+ * 边界情况说明：
+ * - percentile=0: 返回最小值 (index=0)
+ * - percentile=100: 返回最大值 (index=len-1)
+ * - percentile=50 且 len=2: 返回第一个元素（nearest-rank 向下取）
+ *
+ * 算法说明（nearest-rank 方法）：
+ * - index = ceil(percentile / 100 * len) - 1
+ * - 示例：len=100, P50 → ceil(50) - 1 = 49（第50个元素，0-indexed）
+ * - 示例：len=100, P99 → ceil(99) - 1 = 98（第99个元素）
+ */
+function getPercentile(sortedArray, percentile) {
+  const len = sortedArray.length
+  if (len === 0) {
+    return 0
+  }
+  if (len === 1) {
+    return sortedArray[0]
+  }
+
+  // 边界处理：percentile <= 0 返回最小值
+  if (percentile <= 0) {
+    return sortedArray[0]
+  }
+  // 边界处理：percentile >= 100 返回最大值
+  if (percentile >= 100) {
+    return sortedArray[len - 1]
+  }
+
+  const index = Math.ceil((percentile / 100) * len) - 1
+  return sortedArray[index]
+}
+
+/**
+ * 计算等待时间分布统计
+ * @param {number[]} waitTimes - 等待时间数组（无需预先排序）
+ * @returns {Object|null} 统计对象，空数组返回 null
+ *
+ * 返回对象包含：
+ * - sampleCount: 样本数量（始终包含，便于调用方判断可靠性）
+ * - count: 样本数量（向后兼容）
+ * - min: 最小值
+ * - max: 最大值
+ * - avg: 平均值（四舍五入）
+ * - p50: 50百分位数（中位数）
+ * - p90: 90百分位数
+ * - p99: 99百分位数
+ * - sampleSizeWarning: 样本量不足时的警告信息（样本 < 10）
+ * - p90Unreliable: P90 统计不可靠标记（样本 < 10）
+ * - p99Unreliable: P99 统计不可靠标记（样本 < 100）
+ *
+ * 可靠性标记说明（详见 design.md Decision 6）：
+ * - 样本 < 10: P90 和 P99 都不可靠
+ * - 样本 < 100: P99 不可靠（P90 需要 10 个样本，P99 需要 100 个样本）
+ * - 即使标记为不可靠，仍返回计算值供参考
+ */
+function calculateWaitTimeStats(waitTimes) {
+  if (!waitTimes || waitTimes.length === 0) {
+    return null
+  }
+
+  const sorted = [...waitTimes].sort((a, b) => a - b)
+  const sum = sorted.reduce((a, b) => a + b, 0)
+  const len = sorted.length
+
+  const stats = {
+    sampleCount: len, // 新增：始终包含样本数
+    count: len, // 向后兼容
+    min: sorted[0],
+    max: sorted[len - 1],
+    avg: Math.round(sum / len),
+    p50: getPercentile(sorted, 50),
+    p90: getPercentile(sorted, 90),
+    p99: getPercentile(sorted, 99)
+  }
+
+  // 渐进式可靠性标记（详见 design.md Decision 6）
+  // 样本 < 10: P90 不可靠（P90 至少需要 ceil(100/10) = 10 个样本）
+  if (len < 10) {
+    stats.sampleSizeWarning = 'Results may be inaccurate due to small sample size'
+    stats.p90Unreliable = true
+  }
+
+  // 样本 < 100: P99 不可靠（P99 至少需要 ceil(100/1) = 100 个样本）
+  if (len < 100) {
+    stats.p99Unreliable = true
+  }
+
+  return stats
+}
+
+module.exports = {
+  getPercentile,
+  calculateWaitTimeStats
+}

src/utils/streamHelper.js

@@ -0,0 +1,36 @@
+/**
+ * Stream Helper Utilities
+ * 流处理辅助工具函数
+ */
+
+/**
+ * 检查响应流是否仍然可写（客户端连接是否有效）
+ * @param {import('http').ServerResponse} stream - HTTP响应流
+ * @returns {boolean} 如果流可写返回true，否则返回false
+ */
+function isStreamWritable(stream) {
+  if (!stream) {
+    return false
+  }
+
+  // 检查流是否已销毁
+  if (stream.destroyed) {
+    return false
+  }
+
+  // 检查底层socket是否已销毁
+  if (stream.socket?.destroyed) {
+    return false
+  }
+
+  // 检查流是否已结束写入
+  if (stream.writableEnded) {
+    return false
+  }
+
+  return true
+}
+
+module.exports = {
+  isStreamWritable
+}

tests/concurrencyQueue.integration.test.js

@@ -0,0 +1,860 @@
+/**
+ * 并发请求排队功能集成测试
+ *
+ * 测试分为三个层次：
+ * 1. Mock 测试 - 测试核心逻辑，不需要真实 Redis
+ * 2. Redis 方法测试 - 测试 Redis 操作的原子性和正确性
+ * 3. 端到端场景测试 - 测试完整的排队流程
+ *
+ * 运行方式：
+ * - npm test -- concurrencyQueue.integration  # 运行所有测试（Mock 部分）
+ * - REDIS_TEST=1 npm test -- concurrencyQueue.integration  # 包含真实 Redis 测试
+ */
+
+// Mock logger to avoid console output during tests
+jest.mock('../src/utils/logger', () => ({
+  api: jest.fn(),
+  warn: jest.fn(),
+  error: jest.fn(),
+  info: jest.fn(),
+  database: jest.fn(),
+  security: jest.fn()
+}))
+
+const redis = require('../src/models/redis')
+const claudeRelayConfigService = require('../src/services/claudeRelayConfigService')
+
+// Helper: sleep function
+const sleep = (ms) => new Promise((resolve) => setTimeout(resolve, ms))
+
+// Helper: 创建模拟的 req/res 对象
+function createMockReqRes() {
+  const listeners = {}
+  const req = {
+    destroyed: false,
+    once: jest.fn((event, handler) => {
+      listeners[`req:${event}`] = handler
+    }),
+    removeListener: jest.fn((event) => {
+      delete listeners[`req:${event}`]
+    }),
+    // 触发事件的辅助方法
+    emit: (event) => {
+      const handler = listeners[`req:${event}`]
+      if (handler) {
+        handler()
+      }
+    }
+  }
+
+  const res = {
+    once: jest.fn((event, handler) => {
+      listeners[`res:${event}`] = handler
+    }),
+    removeListener: jest.fn((event) => {
+      delete listeners[`res:${event}`]
+    }),
+    emit: (event) => {
+      const handler = listeners[`res:${event}`]
+      if (handler) {
+        handler()
+      }
+    }
+  }
+
+  return { req, res, listeners }
+}
+
+// ============================================
+// 第一部分：Mock 测试 - waitForConcurrencySlot 核心逻辑
+// ============================================
+describe('ConcurrencyQueue Integration Tests', () => {
+  describe('Part 1: waitForConcurrencySlot Logic (Mocked)', () => {
+    // 导入 auth 模块中的 waitForConcurrencySlot
+    // 由于它是内部函数，我们需要通过测试其行为来验证
+    // 这里我们模拟整个流程
+
+    let mockRedis
+
+    beforeEach(() => {
+      jest.clearAllMocks()
+
+      // 创建 Redis mock
+      mockRedis = {
+        concurrencyCount: {},
+        queueCount: {},
+        stats: {},
+        waitTimes: {},
+        globalWaitTimes: []
+      }
+
+      // Mock Redis 并发方法
+      jest.spyOn(redis, 'incrConcurrency').mockImplementation(async (keyId, requestId, _lease) => {
+        if (!mockRedis.concurrencyCount[keyId]) {
+          mockRedis.concurrencyCount[keyId] = new Set()
+        }
+        mockRedis.concurrencyCount[keyId].add(requestId)
+        return mockRedis.concurrencyCount[keyId].size
+      })
+
+      jest.spyOn(redis, 'decrConcurrency').mockImplementation(async (keyId, requestId) => {
+        if (mockRedis.concurrencyCount[keyId]) {
+          mockRedis.concurrencyCount[keyId].delete(requestId)
+          return mockRedis.concurrencyCount[keyId].size
+        }
+        return 0
+      })
+
+      // Mock 排队计数方法
+      jest.spyOn(redis, 'incrConcurrencyQueue').mockImplementation(async (keyId) => {
+        mockRedis.queueCount[keyId] = (mockRedis.queueCount[keyId] || 0) + 1
+        return mockRedis.queueCount[keyId]
+      })
+
+      jest.spyOn(redis, 'decrConcurrencyQueue').mockImplementation(async (keyId) => {
+        mockRedis.queueCount[keyId] = Math.max(0, (mockRedis.queueCount[keyId] || 0) - 1)
+        return mockRedis.queueCount[keyId]
+      })
+
+      jest
+        .spyOn(redis, 'getConcurrencyQueueCount')
+        .mockImplementation(async (keyId) => mockRedis.queueCount[keyId] || 0)
+
+      // Mock 统计方法
+      jest.spyOn(redis, 'incrConcurrencyQueueStats').mockImplementation(async (keyId, field) => {
+        if (!mockRedis.stats[keyId]) {
+          mockRedis.stats[keyId] = {}
+        }
+        mockRedis.stats[keyId][field] = (mockRedis.stats[keyId][field] || 0) + 1
+        return mockRedis.stats[keyId][field]
+      })
+
+      jest.spyOn(redis, 'recordQueueWaitTime').mockResolvedValue(undefined)
+      jest.spyOn(redis, 'recordGlobalQueueWaitTime').mockResolvedValue(undefined)
+    })
+
+    afterEach(() => {
+      jest.restoreAllMocks()
+    })
+
+    describe('Slot Acquisition Flow', () => {
+      it('should acquire slot immediately when under concurrency limit', async () => {
+        // 模拟 waitForConcurrencySlot 的行为
+        const keyId = 'test-key-1'
+        const requestId = 'req-1'
+        const concurrencyLimit = 5
+
+        // 直接测试 incrConcurrency 的行为
+        const count = await redis.incrConcurrency(keyId, requestId, 300)
+
+        expect(count).toBe(1)
+        expect(count).toBeLessThanOrEqual(concurrencyLimit)
+      })
+
+      it('should track multiple concurrent requests correctly', async () => {
+        const keyId = 'test-key-2'
+        const concurrencyLimit = 3
+
+        // 模拟多个并发请求
+        const results = []
+        for (let i = 1; i <= 5; i++) {
+          const count = await redis.incrConcurrency(keyId, `req-${i}`, 300)
+          results.push({ requestId: `req-${i}`, count, exceeds: count > concurrencyLimit })
+        }
+
+        // 前3个应该在限制内
+        expect(results[0].exceeds).toBe(false)
+        expect(results[1].exceeds).toBe(false)
+        expect(results[2].exceeds).toBe(false)
+        // 后2个超过限制
+        expect(results[3].exceeds).toBe(true)
+        expect(results[4].exceeds).toBe(true)
+      })
+
+      it('should release slot and allow next request', async () => {
+        const keyId = 'test-key-3'
+        const concurrencyLimit = 1
+
+        // 第一个请求获取槽位
+        const count1 = await redis.incrConcurrency(keyId, 'req-1', 300)
+        expect(count1).toBe(1)
+
+        // 第二个请求超限
+        const count2 = await redis.incrConcurrency(keyId, 'req-2', 300)
+        expect(count2).toBe(2)
+        expect(count2).toBeGreaterThan(concurrencyLimit)
+
+        // 释放第二个请求（因为超限）
+        await redis.decrConcurrency(keyId, 'req-2')
+
+        // 释放第一个请求
+        await redis.decrConcurrency(keyId, 'req-1')
+
+        // 现在第三个请求应该能获取
+        const count3 = await redis.incrConcurrency(keyId, 'req-3', 300)
+        expect(count3).toBe(1)
+      })
+    })
+
+    describe('Queue Count Management', () => {
+      it('should increment and decrement queue count atomically', async () => {
+        const keyId = 'test-key-4'
+
+        // 增加排队计数
+        const count1 = await redis.incrConcurrencyQueue(keyId, 60000)
+        expect(count1).toBe(1)
+
+        const count2 = await redis.incrConcurrencyQueue(keyId, 60000)
+        expect(count2).toBe(2)
+
+        // 减少排队计数
+        const count3 = await redis.decrConcurrencyQueue(keyId)
+        expect(count3).toBe(1)
+
+        const count4 = await redis.decrConcurrencyQueue(keyId)
+        expect(count4).toBe(0)
+      })
+
+      it('should not go below zero on decrement', async () => {
+        const keyId = 'test-key-5'
+
+        // 直接减少（没有先增加）
+        const count = await redis.decrConcurrencyQueue(keyId)
+        expect(count).toBe(0)
+      })
+
+      it('should handle concurrent queue operations', async () => {
+        const keyId = 'test-key-6'
+
+        // 并发增加
+        const increments = await Promise.all([
+          redis.incrConcurrencyQueue(keyId, 60000),
+          redis.incrConcurrencyQueue(keyId, 60000),
+          redis.incrConcurrencyQueue(keyId, 60000)
+        ])
+
+        // 所有增量应该是连续的
+        const sortedIncrements = [...increments].sort((a, b) => a - b)
+        expect(sortedIncrements).toEqual([1, 2, 3])
+      })
+    })
+
+    describe('Statistics Tracking', () => {
+      it('should track entered/success/timeout/cancelled stats', async () => {
+        const keyId = 'test-key-7'
+
+        await redis.incrConcurrencyQueueStats(keyId, 'entered')
+        await redis.incrConcurrencyQueueStats(keyId, 'entered')
+        await redis.incrConcurrencyQueueStats(keyId, 'success')
+        await redis.incrConcurrencyQueueStats(keyId, 'timeout')
+        await redis.incrConcurrencyQueueStats(keyId, 'cancelled')
+
+        expect(mockRedis.stats[keyId]).toEqual({
+          entered: 2,
+          success: 1,
+          timeout: 1,
+          cancelled: 1
+        })
+      })
+    })
+
+    describe('Client Disconnection Handling', () => {
+      it('should detect client disconnection via close event', async () => {
+        const { req } = createMockReqRes()
+
+        let clientDisconnected = false
+
+        // 设置监听器
+        req.once('close', () => {
+          clientDisconnected = true
+        })
+
+        // 模拟客户端断开
+        req.emit('close')
+
+        expect(clientDisconnected).toBe(true)
+      })
+
+      it('should detect pre-destroyed request', () => {
+        const { req } = createMockReqRes()
+        req.destroyed = true
+
+        expect(req.destroyed).toBe(true)
+      })
+    })
+
+    describe('Exponential Backoff Simulation', () => {
+      it('should increase poll interval with backoff', () => {
+        const config = {
+          pollIntervalMs: 200,
+          maxPollIntervalMs: 2000,
+          backoffFactor: 1.5,
+          jitterRatio: 0 // 禁用抖动以便测试
+        }
+
+        let interval = config.pollIntervalMs
+        const intervals = [interval]
+
+        for (let i = 0; i < 5; i++) {
+          interval = Math.min(interval * config.backoffFactor, config.maxPollIntervalMs)
+          intervals.push(interval)
+        }
+
+        // 验证指数增长
+        expect(intervals[1]).toBe(300) // 200 * 1.5
+        expect(intervals[2]).toBe(450) // 300 * 1.5
+        expect(intervals[3]).toBe(675) // 450 * 1.5
+        expect(intervals[4]).toBe(1012.5) // 675 * 1.5
+        expect(intervals[5]).toBe(1518.75) // 1012.5 * 1.5
+      })
+
+      it('should cap interval at maximum', () => {
+        const config = {
+          pollIntervalMs: 1000,
+          maxPollIntervalMs: 2000,
+          backoffFactor: 1.5
+        }
+
+        let interval = config.pollIntervalMs
+
+        for (let i = 0; i < 10; i++) {
+          interval = Math.min(interval * config.backoffFactor, config.maxPollIntervalMs)
+        }
+
+        expect(interval).toBe(2000)
+      })
+
+      it('should apply jitter within expected range', () => {
+        const baseInterval = 1000
+        const jitterRatio = 0.2 // ±20%
+        const results = []
+
+        for (let i = 0; i < 100; i++) {
+          const randomValue = Math.random()
+          const jitter = baseInterval * jitterRatio * (randomValue * 2 - 1)
+          const finalInterval = baseInterval + jitter
+          results.push(finalInterval)
+        }
+
+        const min = Math.min(...results)
+        const max = Math.max(...results)
+
+        // 所有结果应该在 [800, 1200] 范围内
+        expect(min).toBeGreaterThanOrEqual(800)
+        expect(max).toBeLessThanOrEqual(1200)
+      })
+    })
+  })
+
+  // ============================================
+  // 第二部分：并发竞争场景测试
+  // ============================================
+  describe('Part 2: Concurrent Race Condition Tests', () => {
+    beforeEach(() => {
+      jest.clearAllMocks()
+    })
+
+    afterEach(() => {
+      jest.restoreAllMocks()
+    })
+
+    describe('Race Condition: Multiple Requests Competing for Same Slot', () => {
+      it('should handle race condition when multiple requests try to acquire last slot', async () => {
+        const keyId = 'race-test-1'
+        const concurrencyLimit = 1
+        const concurrencyState = { count: 0, holders: new Set() }
+
+        // 模拟原子的 incrConcurrency
+        jest.spyOn(redis, 'incrConcurrency').mockImplementation(async (key, reqId) => {
+          // 模拟原子操作
+          concurrencyState.count++
+          concurrencyState.holders.add(reqId)
+          return concurrencyState.count
+        })
+
+        jest.spyOn(redis, 'decrConcurrency').mockImplementation(async (key, reqId) => {
+          if (concurrencyState.holders.has(reqId)) {
+            concurrencyState.count--
+            concurrencyState.holders.delete(reqId)
+          }
+          return concurrencyState.count
+        })
+
+        // 5个请求同时竞争1个槽位
+        const requests = Array.from({ length: 5 }, (_, i) => `req-${i + 1}`)
+
+        const acquireResults = await Promise.all(
+          requests.map(async (reqId) => {
+            const count = await redis.incrConcurrency(keyId, reqId, 300)
+            const acquired = count <= concurrencyLimit
+
+            if (!acquired) {
+              // 超限，释放
+              await redis.decrConcurrency(keyId, reqId)
+            }
+
+            return { reqId, count, acquired }
+          })
+        )
+
+        // 只有一个请求应该成功获取槽位
+        const successfulAcquires = acquireResults.filter((r) => r.acquired)
+        expect(successfulAcquires.length).toBe(1)
+
+        // 最终并发计数应该是1
+        expect(concurrencyState.count).toBe(1)
+      })
+
+      it('should maintain consistency under high contention', async () => {
+        const keyId = 'race-test-2'
+        const concurrencyLimit = 3
+        const requestCount = 20
+        const concurrencyState = { count: 0, maxSeen: 0 }
+
+        jest.spyOn(redis, 'incrConcurrency').mockImplementation(async () => {
+          concurrencyState.count++
+          concurrencyState.maxSeen = Math.max(concurrencyState.maxSeen, concurrencyState.count)
+          return concurrencyState.count
+        })
+
+        jest.spyOn(redis, 'decrConcurrency').mockImplementation(async () => {
+          concurrencyState.count = Math.max(0, concurrencyState.count - 1)
+          return concurrencyState.count
+        })
+
+        // 模拟多轮请求
+        const activeRequests = []
+
+        for (let i = 0; i < requestCount; i++) {
+          const count = await redis.incrConcurrency(keyId, `req-${i}`, 300)
+
+          if (count <= concurrencyLimit) {
+            activeRequests.push(`req-${i}`)
+
+            // 模拟处理时间后释放
+            setTimeout(async () => {
+              await redis.decrConcurrency(keyId, `req-${i}`)
+            }, Math.random() * 50)
+          } else {
+            await redis.decrConcurrency(keyId, `req-${i}`)
+          }
+
+          // 随机延迟
+          await sleep(Math.random() * 10)
+        }
+
+        // 等待所有请求完成
+        await sleep(100)
+
+        // 最大并发不应超过限制
+        expect(concurrencyState.maxSeen).toBeLessThanOrEqual(concurrencyLimit + requestCount) // 允许短暂超限
+      })
+    })
+
+    describe('Queue Overflow Protection', () => {
+      it('should reject requests when queue is full', async () => {
+        const keyId = 'overflow-test-1'
+        const maxQueueSize = 5
+        const queueState = { count: 0 }
+
+        jest.spyOn(redis, 'incrConcurrencyQueue').mockImplementation(async () => {
+          queueState.count++
+          return queueState.count
+        })
+
+        jest.spyOn(redis, 'decrConcurrencyQueue').mockImplementation(async () => {
+          queueState.count = Math.max(0, queueState.count - 1)
+          return queueState.count
+        })
+
+        const results = []
+
+        // 尝试10个请求进入队列
+        for (let i = 0; i < 10; i++) {
+          const queueCount = await redis.incrConcurrencyQueue(keyId, 60000)
+
+          if (queueCount > maxQueueSize) {
+            // 队列满，释放并拒绝
+            await redis.decrConcurrencyQueue(keyId)
+            results.push({ index: i, accepted: false })
+          } else {
+            results.push({ index: i, accepted: true, position: queueCount })
+          }
+        }
+
+        const accepted = results.filter((r) => r.accepted)
+        const rejected = results.filter((r) => !r.accepted)
+
+        expect(accepted.length).toBe(5)
+        expect(rejected.length).toBe(5)
+      })
+    })
+  })
+
+  // ============================================
+  // 第三部分：真实 Redis 集成测试（可选）
+  // ============================================
+  describe('Part 3: Real Redis Integration Tests', () => {
+    const skipRealRedis = !process.env.REDIS_TEST
+
+    // 辅助函数：检查 Redis 连接
+    async function checkRedisConnection() {
+      try {
+        const client = redis.getClient()
+        if (!client) {
+          return false
+        }
+        await client.ping()
+        return true
+      } catch {
+        return false
+      }
+    }
+
+    beforeAll(async () => {
+      if (skipRealRedis) {
+        console.log('⏭️  Skipping real Redis tests (set REDIS_TEST=1 to enable)')
+        return
+      }
+
+      const connected = await checkRedisConnection()
+      if (!connected) {
+        console.log('⚠️  Redis not connected, skipping real Redis tests')
+      }
+    })
+
+    // 清理测试数据
+    afterEach(async () => {
+      if (skipRealRedis) {
+        return
+      }
+
+      try {
+        const client = redis.getClient()
+        if (!client) {
+          return
+        }
+
+        // 清理测试键
+        const testKeys = await client.keys('concurrency:queue:test-*')
+        if (testKeys.length > 0) {
+          await client.del(...testKeys)
+        }
+      } catch {
+        // 忽略清理错误
+      }
+    })
+
+    describe('Redis Queue Operations', () => {
+      const testOrSkip = skipRealRedis ? it.skip : it
+
+      testOrSkip('should atomically increment queue count with TTL', async () => {
+        const keyId = 'test-redis-queue-1'
+        const timeoutMs = 5000
+
+        const count1 = await redis.incrConcurrencyQueue(keyId, timeoutMs)
+        expect(count1).toBe(1)
+
+        const count2 = await redis.incrConcurrencyQueue(keyId, timeoutMs)
+        expect(count2).toBe(2)
+
+        // 验证 TTL 被设置
+        const client = redis.getClient()
+        const ttl = await client.ttl(`concurrency:queue:${keyId}`)
+        expect(ttl).toBeGreaterThan(0)
+        expect(ttl).toBeLessThanOrEqual(Math.ceil(timeoutMs / 1000) + 30)
+      })
+
+      testOrSkip('should atomically decrement and delete when zero', async () => {
+        const keyId = 'test-redis-queue-2'
+
+        await redis.incrConcurrencyQueue(keyId, 60000)
+        const count = await redis.decrConcurrencyQueue(keyId)
+
+        expect(count).toBe(0)
+
+        // 验证键已删除
+        const client = redis.getClient()
+        const exists = await client.exists(`concurrency:queue:${keyId}`)
+        expect(exists).toBe(0)
+      })
+
+      testOrSkip('should handle concurrent increments correctly', async () => {
+        const keyId = 'test-redis-queue-3'
+        const numRequests = 10
+
+        // 并发增加
+        const results = await Promise.all(
+          Array.from({ length: numRequests }, () => redis.incrConcurrencyQueue(keyId, 60000))
+        )
+
+        // 所有结果应该是 1 到 numRequests
+        const sorted = [...results].sort((a, b) => a - b)
+        expect(sorted).toEqual(Array.from({ length: numRequests }, (_, i) => i + 1))
+      })
+    })
+
+    describe('Redis Stats Operations', () => {
+      const testOrSkip = skipRealRedis ? it.skip : it
+
+      testOrSkip('should track queue statistics correctly', async () => {
+        const keyId = 'test-redis-stats-1'
+
+        await redis.incrConcurrencyQueueStats(keyId, 'entered')
+        await redis.incrConcurrencyQueueStats(keyId, 'entered')
+        await redis.incrConcurrencyQueueStats(keyId, 'success')
+        await redis.incrConcurrencyQueueStats(keyId, 'timeout')
+
+        const stats = await redis.getConcurrencyQueueStats(keyId)
+
+        expect(stats.entered).toBe(2)
+        expect(stats.success).toBe(1)
+        expect(stats.timeout).toBe(1)
+        expect(stats.cancelled).toBe(0)
+      })
+
+      testOrSkip('should record and retrieve wait times', async () => {
+        const keyId = 'test-redis-wait-1'
+        const waitTimes = [100, 200, 150, 300, 250]
+
+        for (const wt of waitTimes) {
+          await redis.recordQueueWaitTime(keyId, wt)
+        }
+
+        const recorded = await redis.getQueueWaitTimes(keyId)
+
+        // 应该按 LIFO 顺序存储
+        expect(recorded.length).toBe(5)
+        expect(recorded[0]).toBe(250) // 最后插入的在前面
+      })
+
+      testOrSkip('should record global wait times', async () => {
+        const waitTimes = [500, 600, 700]
+
+        for (const wt of waitTimes) {
+          await redis.recordGlobalQueueWaitTime(wt)
+        }
+
+        const recorded = await redis.getGlobalQueueWaitTimes()
+
+        expect(recorded.length).toBeGreaterThanOrEqual(3)
+      })
+    })
+
+    describe('Redis Cleanup Operations', () => {
+      const testOrSkip = skipRealRedis ? it.skip : it
+
+      testOrSkip('should clear specific queue', async () => {
+        const keyId = 'test-redis-clear-1'
+
+        await redis.incrConcurrencyQueue(keyId, 60000)
+        await redis.incrConcurrencyQueue(keyId, 60000)
+
+        const cleared = await redis.clearConcurrencyQueue(keyId)
+        expect(cleared).toBe(true)
+
+        const count = await redis.getConcurrencyQueueCount(keyId)
+        expect(count).toBe(0)
+      })
+
+      testOrSkip('should clear all queues but preserve stats', async () => {
+        const keyId1 = 'test-redis-clearall-1'
+        const keyId2 = 'test-redis-clearall-2'
+
+        // 创建队列和统计
+        await redis.incrConcurrencyQueue(keyId1, 60000)
+        await redis.incrConcurrencyQueue(keyId2, 60000)
+        await redis.incrConcurrencyQueueStats(keyId1, 'entered')
+
+        // 清理所有队列
+        const cleared = await redis.clearAllConcurrencyQueues()
+        expect(cleared).toBeGreaterThanOrEqual(2)
+
+        // 验证队列已清理
+        const count1 = await redis.getConcurrencyQueueCount(keyId1)
+        const count2 = await redis.getConcurrencyQueueCount(keyId2)
+        expect(count1).toBe(0)
+        expect(count2).toBe(0)
+
+        // 统计应该保留
+        const stats = await redis.getConcurrencyQueueStats(keyId1)
+        expect(stats.entered).toBe(1)
+      })
+    })
+  })
+
+  // ============================================
+  // 第四部分：配置服务集成测试
+  // ============================================
+  describe('Part 4: Configuration Service Integration', () => {
+    beforeEach(() => {
+      // 清除配置缓存
+      claudeRelayConfigService.clearCache()
+    })
+
+    afterEach(() => {
+      jest.restoreAllMocks()
+    })
+
+    describe('Queue Configuration', () => {
+      it('should return default queue configuration', async () => {
+        jest.spyOn(redis, 'getClient').mockReturnValue(null)
+
+        const config = await claudeRelayConfigService.getConfig()
+
+        expect(config.concurrentRequestQueueEnabled).toBe(false)
+        expect(config.concurrentRequestQueueMaxSize).toBe(3)
+        expect(config.concurrentRequestQueueMaxSizeMultiplier).toBe(0)
+        expect(config.concurrentRequestQueueTimeoutMs).toBe(10000)
+      })
+
+      it('should calculate max queue size correctly', async () => {
+        const testCases = [
+          { concurrencyLimit: 5, multiplier: 2, fixedMin: 3, expected: 10 }, // 5*2=10 > 3
+          { concurrencyLimit: 1, multiplier: 1, fixedMin: 5, expected: 5 }, // 1*1=1 < 5
+          { concurrencyLimit: 10, multiplier: 0.5, fixedMin: 3, expected: 5 }, // 10*0.5=5 > 3
+          { concurrencyLimit: 2, multiplier: 1, fixedMin: 10, expected: 10 } // 2*1=2 < 10
+        ]
+
+        for (const tc of testCases) {
+          const maxQueueSize = Math.max(tc.concurrencyLimit * tc.multiplier, tc.fixedMin)
+          expect(maxQueueSize).toBe(tc.expected)
+        }
+      })
+    })
+  })
+
+  // ============================================
+  // 第五部分：端到端场景测试
+  // ============================================
+  describe('Part 5: End-to-End Scenario Tests', () => {
+    describe('Scenario: Claude Code Agent Parallel Tool Calls', () => {
+      it('should handle burst of parallel tool results', async () => {
+        // 模拟 Claude Code Agent 发送多个并行工具结果的场景
+        const concurrencyLimit = 2
+        const maxQueueSize = 5
+
+        const state = {
+          concurrency: 0,
+          queue: 0,
+          completed: 0,
+          rejected: 0
+        }
+
+        // 模拟 8 个并行工具结果请求
+        const requests = Array.from({ length: 8 }, (_, i) => ({
+          id: `tool-result-${i + 1}`,
+          startTime: Date.now()
+        }))
+
+        // 模拟处理逻辑
+        async function processRequest(req) {
+          // 尝试获取并发槽位
+          state.concurrency++
+
+          if (state.concurrency > concurrencyLimit) {
+            // 超限，进入队列
+            state.concurrency--
+            state.queue++
+
+            if (state.queue > maxQueueSize) {
+              // 队列满，拒绝
+              state.queue--
+              state.rejected++
+              return { ...req, status: 'rejected', reason: 'queue_full' }
+            }
+
+            // 等待槽位（模拟）
+            await sleep(Math.random() * 100)
+            state.queue--
+            state.concurrency++
+          }
+
+          // 处理请求
+          await sleep(50) // 模拟处理时间
+          state.concurrency--
+          state.completed++
+
+          return { ...req, status: 'completed', duration: Date.now() - req.startTime }
+        }
+
+        const results = await Promise.all(requests.map(processRequest))
+
+        const completed = results.filter((r) => r.status === 'completed')
+        const rejected = results.filter((r) => r.status === 'rejected')
+
+        // 大部分请求应该完成
+        expect(completed.length).toBeGreaterThan(0)
+        // 可能有一些被拒绝
+        expect(state.rejected).toBe(rejected.length)
+
+        console.log(
+          `  ✓ Completed: ${completed.length}, Rejected: ${rejected.length}, Max concurrent: ${concurrencyLimit}`
+        )
+      })
+    })
+
+    describe('Scenario: Graceful Degradation', () => {
+      it('should fallback when Redis fails', async () => {
+        jest
+          .spyOn(redis, 'incrConcurrencyQueue')
+          .mockRejectedValue(new Error('Redis connection lost'))
+
+        // 模拟降级行为：Redis 失败时直接拒绝而不是崩溃
+        let result
+        try {
+          await redis.incrConcurrencyQueue('fallback-test', 60000)
+          result = { success: true }
+        } catch (error) {
+          // 优雅降级：返回 429 而不是 500
+          result = { success: false, fallback: true, error: error.message }
+        }
+
+        expect(result.fallback).toBe(true)
+        expect(result.error).toContain('Redis')
+      })
+    })
+
+    describe('Scenario: Timeout Behavior', () => {
+      it('should respect queue timeout', async () => {
+        const timeoutMs = 100
+        const startTime = Date.now()
+
+        // 模拟等待超时
+        await new Promise((resolve) => setTimeout(resolve, timeoutMs))
+
+        const elapsed = Date.now() - startTime
+        expect(elapsed).toBeGreaterThanOrEqual(timeoutMs - 10) // 允许 10ms 误差
+      })
+
+      it('should track timeout statistics', async () => {
+        const stats = { entered: 0, success: 0, timeout: 0, cancelled: 0 }
+
+        // 模拟多个请求，部分超时
+        const requests = [
+          { id: 'req-1', willTimeout: false },
+          { id: 'req-2', willTimeout: true },
+          { id: 'req-3', willTimeout: false },
+          { id: 'req-4', willTimeout: true }
+        ]
+
+        for (const req of requests) {
+          stats.entered++
+          if (req.willTimeout) {
+            stats.timeout++
+          } else {
+            stats.success++
+          }
+        }
+
+        expect(stats.entered).toBe(4)
+        expect(stats.success).toBe(2)
+        expect(stats.timeout).toBe(2)
+
+        // 成功率应该是 50%
+        const successRate = (stats.success / stats.entered) * 100
+        expect(successRate).toBe(50)
+      })
+    })
+  })
+})

tests/concurrencyQueue.test.js

@@ -0,0 +1,278 @@
+/**
+ * 并发请求排队功能测试
+ * 测试排队逻辑中的核心算法：百分位数计算、等待时间统计、指数退避等
+ *
+ * 注意：Redis 方法的测试需要集成测试环境，这里主要测试纯算法逻辑
+ */
+
+// Mock logger to avoid console output during tests
+jest.mock('../src/utils/logger', () => ({
+  api: jest.fn(),
+  warn: jest.fn(),
+  error: jest.fn(),
+  info: jest.fn(),
+  database: jest.fn(),
+  security: jest.fn()
+}))
+
+// 使用共享的统计工具函数（与生产代码一致）
+const { getPercentile, calculateWaitTimeStats } = require('../src/utils/statsHelper')
+
+describe('ConcurrencyQueue', () => {
+  describe('Percentile Calculation (nearest-rank method)', () => {
+    // 直接测试共享工具函数，确保与生产代码行为一致
+    it('should return 0 for empty array', () => {
+      expect(getPercentile([], 50)).toBe(0)
+    })
+
+    it('should return single element for single-element array', () => {
+      expect(getPercentile([100], 50)).toBe(100)
+      expect(getPercentile([100], 99)).toBe(100)
+    })
+
+    it('should return min for percentile 0', () => {
+      expect(getPercentile([10, 20, 30, 40, 50], 0)).toBe(10)
+    })
+
+    it('should return max for percentile 100', () => {
+      expect(getPercentile([10, 20, 30, 40, 50], 100)).toBe(50)
+    })
+
+    it('should calculate P50 correctly for len=10', () => {
+      // For [10, 20, 30, 40, 50, 60, 70, 80, 90, 100] (len=10)
+      // P50: ceil(50/100 * 10) - 1 = ceil(5) - 1 = 4 → value at index 4 = 50
+      const arr = [10, 20, 30, 40, 50, 60, 70, 80, 90, 100]
+      expect(getPercentile(arr, 50)).toBe(50)
+    })
+
+    it('should calculate P90 correctly for len=10', () => {
+      // For len=10, P90: ceil(90/100 * 10) - 1 = ceil(9) - 1 = 8 → value at index 8 = 90
+      const arr = [10, 20, 30, 40, 50, 60, 70, 80, 90, 100]
+      expect(getPercentile(arr, 90)).toBe(90)
+    })
+
+    it('should calculate P99 correctly for len=100', () => {
+      // For len=100, P99: ceil(99/100 * 100) - 1 = ceil(99) - 1 = 98
+      const arr = Array.from({ length: 100 }, (_, i) => i + 1)
+      expect(getPercentile(arr, 99)).toBe(99)
+    })
+
+    it('should handle two-element array correctly', () => {
+      // For [10, 20] (len=2)
+      // P50: ceil(50/100 * 2) - 1 = ceil(1) - 1 = 0 → value = 10
+      expect(getPercentile([10, 20], 50)).toBe(10)
+    })
+
+    it('should handle negative percentile as 0', () => {
+      expect(getPercentile([10, 20, 30], -10)).toBe(10)
+    })
+
+    it('should handle percentile > 100 as 100', () => {
+      expect(getPercentile([10, 20, 30], 150)).toBe(30)
+    })
+  })
+
+  describe('Wait Time Stats Calculation', () => {
+    // 直接测试共享工具函数
+    it('should return null for empty array', () => {
+      expect(calculateWaitTimeStats([])).toBeNull()
+    })
+
+    it('should return null for null input', () => {
+      expect(calculateWaitTimeStats(null)).toBeNull()
+    })
+
+    it('should return null for undefined input', () => {
+      expect(calculateWaitTimeStats(undefined)).toBeNull()
+    })
+
+    it('should calculate stats correctly for typical data', () => {
+      const waitTimes = [100, 200, 150, 300, 250, 180, 220, 280, 190, 210]
+      const stats = calculateWaitTimeStats(waitTimes)
+
+      expect(stats.count).toBe(10)
+      expect(stats.min).toBe(100)
+      expect(stats.max).toBe(300)
+      // Sum: 100+150+180+190+200+210+220+250+280+300 = 2080
+      expect(stats.avg).toBe(208)
+      expect(stats.sampleSizeWarning).toBeUndefined()
+    })
+
+    it('should add warning for small sample size (< 10)', () => {
+      const waitTimes = [100, 200, 300]
+      const stats = calculateWaitTimeStats(waitTimes)
+
+      expect(stats.count).toBe(3)
+      expect(stats.sampleSizeWarning).toBe('Results may be inaccurate due to small sample size')
+    })
+
+    it('should handle single value', () => {
+      const stats = calculateWaitTimeStats([500])
+
+      expect(stats.count).toBe(1)
+      expect(stats.min).toBe(500)
+      expect(stats.max).toBe(500)
+      expect(stats.avg).toBe(500)
+      expect(stats.p50).toBe(500)
+      expect(stats.p90).toBe(500)
+      expect(stats.p99).toBe(500)
+    })
+
+    it('should sort input array before calculating', () => {
+      const waitTimes = [500, 100, 300, 200, 400]
+      const stats = calculateWaitTimeStats(waitTimes)
+
+      expect(stats.min).toBe(100)
+      expect(stats.max).toBe(500)
+    })
+
+    it('should not modify original array', () => {
+      const waitTimes = [500, 100, 300]
+      calculateWaitTimeStats(waitTimes)
+
+      expect(waitTimes).toEqual([500, 100, 300])
+    })
+  })
+
+  describe('Exponential Backoff with Jitter', () => {
+    /**
+     * 指数退避计算函数（与 auth.js 中的实现一致）
+     * @param {number} currentInterval - 当前轮询间隔
+     * @param {number} backoffFactor - 退避系数
+     * @param {number} jitterRatio - 抖动比例
+     * @param {number} maxInterval - 最大间隔
+     * @param {number} randomValue - 随机值 [0, 1)，用于确定性测试
+     */
+    function calculateNextInterval(
+      currentInterval,
+      backoffFactor,
+      jitterRatio,
+      maxInterval,
+      randomValue
+    ) {
+      let nextInterval = currentInterval * backoffFactor
+      // 抖动范围：[-jitterRatio, +jitterRatio]
+      const jitter = nextInterval * jitterRatio * (randomValue * 2 - 1)
+      nextInterval = nextInterval + jitter
+      return Math.max(1, Math.min(nextInterval, maxInterval))
+    }
+
+    it('should apply exponential backoff without jitter (randomValue=0.5)', () => {
+      // randomValue = 0.5 gives jitter = 0
+      const next = calculateNextInterval(100, 1.5, 0.2, 1000, 0.5)
+      expect(next).toBe(150) // 100 * 1.5 = 150
+    })
+
+    it('should apply maximum positive jitter (randomValue=1.0)', () => {
+      // randomValue = 1.0 gives maximum positive jitter (+20%)
+      const next = calculateNextInterval(100, 1.5, 0.2, 1000, 1.0)
+      // 100 * 1.5 = 150, jitter = 150 * 0.2 * 1 = 30
+      expect(next).toBe(180) // 150 + 30
+    })
+
+    it('should apply maximum negative jitter (randomValue=0.0)', () => {
+      // randomValue = 0.0 gives maximum negative jitter (-20%)
+      const next = calculateNextInterval(100, 1.5, 0.2, 1000, 0.0)
+      // 100 * 1.5 = 150, jitter = 150 * 0.2 * -1 = -30
+      expect(next).toBe(120) // 150 - 30
+    })
+
+    it('should respect maximum interval', () => {
+      const next = calculateNextInterval(800, 1.5, 0.2, 1000, 1.0)
+      // 800 * 1.5 = 1200, with +20% jitter = 1440, capped at 1000
+      expect(next).toBe(1000)
+    })
+
+    it('should never go below 1ms even with extreme negative jitter', () => {
+      const next = calculateNextInterval(1, 1.0, 0.9, 1000, 0.0)
+      // 1 * 1.0 = 1, jitter = 1 * 0.9 * -1 = -0.9
+      // 1 - 0.9 = 0.1, but Math.max(1, ...) ensures minimum is 1
+      expect(next).toBe(1)
+    })
+
+    it('should handle zero jitter ratio', () => {
+      const next = calculateNextInterval(100, 2.0, 0, 1000, 0.0)
+      expect(next).toBe(200) // Pure exponential, no jitter
+    })
+
+    it('should handle large backoff factor', () => {
+      const next = calculateNextInterval(100, 3.0, 0.1, 1000, 0.5)
+      expect(next).toBe(300) // 100 * 3.0 = 300
+    })
+
+    describe('jitter distribution', () => {
+      it('should produce values in expected range', () => {
+        const results = []
+        // Test with various random values
+        for (let r = 0; r <= 1; r += 0.1) {
+          results.push(calculateNextInterval(100, 1.5, 0.2, 1000, r))
+        }
+        // All values should be between 120 (150 - 30) and 180 (150 + 30)
+        expect(Math.min(...results)).toBeGreaterThanOrEqual(120)
+        expect(Math.max(...results)).toBeLessThanOrEqual(180)
+      })
+    })
+  })
+
+  describe('Queue Size Calculation', () => {
+    /**
+     * 最大排队数计算（与 auth.js 中的实现一致）
+     */
+    function calculateMaxQueueSize(concurrencyLimit, multiplier, fixedMin) {
+      return Math.max(concurrencyLimit * multiplier, fixedMin)
+    }
+
+    it('should use multiplier when result is larger', () => {
+      // concurrencyLimit=10, multiplier=2, fixedMin=5
+      // max(10*2, 5) = max(20, 5) = 20
+      expect(calculateMaxQueueSize(10, 2, 5)).toBe(20)
+    })
+
+    it('should use fixed minimum when multiplier result is smaller', () => {
+      // concurrencyLimit=2, multiplier=1, fixedMin=5
+      // max(2*1, 5) = max(2, 5) = 5
+      expect(calculateMaxQueueSize(2, 1, 5)).toBe(5)
+    })
+
+    it('should handle zero multiplier', () => {
+      // concurrencyLimit=10, multiplier=0, fixedMin=3
+      // max(10*0, 3) = max(0, 3) = 3
+      expect(calculateMaxQueueSize(10, 0, 3)).toBe(3)
+    })
+
+    it('should handle fractional multiplier', () => {
+      // concurrencyLimit=10, multiplier=1.5, fixedMin=5
+      // max(10*1.5, 5) = max(15, 5) = 15
+      expect(calculateMaxQueueSize(10, 1.5, 5)).toBe(15)
+    })
+  })
+
+  describe('TTL Calculation', () => {
+    /**
+     * 排队计数器 TTL 计算（与 redis.js 中的实现一致）
+     */
+    function calculateQueueTtl(timeoutMs, bufferSeconds = 30) {
+      return Math.ceil(timeoutMs / 1000) + bufferSeconds
+    }
+
+    it('should calculate TTL with default buffer', () => {
+      // 60000ms = 60s + 30s buffer = 90s
+      expect(calculateQueueTtl(60000)).toBe(90)
+    })
+
+    it('should round up milliseconds to seconds', () => {
+      // 61500ms = ceil(61.5) = 62s + 30s = 92s
+      expect(calculateQueueTtl(61500)).toBe(92)
+    })
+
+    it('should handle custom buffer', () => {
+      // 30000ms = 30s + 60s buffer = 90s
+      expect(calculateQueueTtl(30000, 60)).toBe(90)
+    })
+
+    it('should handle very short timeout', () => {
+      // 1000ms = 1s + 30s = 31s
+      expect(calculateQueueTtl(1000)).toBe(31)
+    })
+  })
+})

tests/userMessageQueue.test.js

@@ -0,0 +1,434 @@
+/**
+ * 用户消息队列服务测试
+ * 测试消息类型检测、队列串行行为、延迟间隔、超时处理和功能开关
+ */
+
+const redis = require('../src/models/redis')
+const userMessageQueueService = require('../src/services/userMessageQueueService')
+
+describe('UserMessageQueueService', () => {
+  describe('isUserMessageRequest', () => {
+    it('should return true when last message role is user', () => {
+      const requestBody = {
+        messages: [
+          { role: 'user', content: 'Hello' },
+          { role: 'assistant', content: 'Hi there' },
+          { role: 'user', content: 'How are you?' }
+        ]
+      }
+      expect(userMessageQueueService.isUserMessageRequest(requestBody)).toBe(true)
+    })
+
+    it('should return false when last message role is assistant', () => {
+      const requestBody = {
+        messages: [
+          { role: 'user', content: 'Hello' },
+          { role: 'assistant', content: 'Hi there' }
+        ]
+      }
+      expect(userMessageQueueService.isUserMessageRequest(requestBody)).toBe(false)
+    })
+
+    it('should return false when last message contains tool_result', () => {
+      const requestBody = {
+        messages: [
+          { role: 'user', content: 'Hello' },
+          { role: 'assistant', content: 'Let me check that' },
+          {
+            role: 'user',
+            content: [
+              {
+                type: 'tool_result',
+                tool_use_id: 'test-id',
+                content: 'Tool result'
+              }
+            ]
+          }
+        ]
+      }
+      // tool_result 消息虽然 role 是 user，但不是真正的用户消息
+      // 应该返回 false，不进入用户消息队列
+      expect(userMessageQueueService.isUserMessageRequest(requestBody)).toBe(false)
+    })
+
+    it('should return false when last message contains multiple tool_results', () => {
+      const requestBody = {
+        messages: [
+          { role: 'user', content: 'Run multiple tools' },
+          {
+            role: 'user',
+            content: [
+              {
+                type: 'tool_result',
+                tool_use_id: 'tool-1',
+                content: 'Result 1'
+              },
+              {
+                type: 'tool_result',
+                tool_use_id: 'tool-2',
+                content: 'Result 2'
+              }
+            ]
+          }
+        ]
+      }
+      expect(userMessageQueueService.isUserMessageRequest(requestBody)).toBe(false)
+    })
+
+    it('should return true when user message has array content with text type', () => {
+      const requestBody = {
+        messages: [
+          {
+            role: 'user',
+            content: [
+              {
+                type: 'text',
+                text: 'Hello, this is a user message'
+              }
+            ]
+          }
+        ]
+      }
+      expect(userMessageQueueService.isUserMessageRequest(requestBody)).toBe(true)
+    })
+
+    it('should return true when user message has mixed text and image content', () => {
+      const requestBody = {
+        messages: [
+          {
+            role: 'user',
+            content: [
+              {
+                type: 'text',
+                text: 'What is in this image?'
+              },
+              {
+                type: 'image',
+                source: { type: 'base64', media_type: 'image/png', data: '...' }
+              }
+            ]
+          }
+        ]
+      }
+      expect(userMessageQueueService.isUserMessageRequest(requestBody)).toBe(true)
+    })
+
+    it('should return false when messages is empty', () => {
+      const requestBody = { messages: [] }
+      expect(userMessageQueueService.isUserMessageRequest(requestBody)).toBe(false)
+    })
+
+    it('should return false when messages is not an array', () => {
+      const requestBody = { messages: 'not an array' }
+      expect(userMessageQueueService.isUserMessageRequest(requestBody)).toBe(false)
+    })
+
+    it('should return false when messages is undefined', () => {
+      const requestBody = {}
+      expect(userMessageQueueService.isUserMessageRequest(requestBody)).toBe(false)
+    })
+
+    it('should return false when requestBody is null', () => {
+      expect(userMessageQueueService.isUserMessageRequest(null)).toBe(false)
+    })
+
+    it('should return false when requestBody is undefined', () => {
+      expect(userMessageQueueService.isUserMessageRequest(undefined)).toBe(false)
+    })
+
+    it('should return false when last message has no role', () => {
+      const requestBody = {
+        messages: [{ content: 'Hello' }]
+      }
+      expect(userMessageQueueService.isUserMessageRequest(requestBody)).toBe(false)
+    })
+
+    it('should handle single user message', () => {
+      const requestBody = {
+        messages: [{ role: 'user', content: 'Hello' }]
+      }
+      expect(userMessageQueueService.isUserMessageRequest(requestBody)).toBe(true)
+    })
+
+    it('should handle single assistant message', () => {
+      const requestBody = {
+        messages: [{ role: 'assistant', content: 'Hello' }]
+      }
+      expect(userMessageQueueService.isUserMessageRequest(requestBody)).toBe(false)
+    })
+  })
+
+  describe('getConfig', () => {
+    it('should return config with expected properties', async () => {
+      const config = await userMessageQueueService.getConfig()
+      expect(config).toHaveProperty('enabled')
+      expect(config).toHaveProperty('delayMs')
+      expect(config).toHaveProperty('timeoutMs')
+      expect(config).toHaveProperty('lockTtlMs')
+      expect(typeof config.enabled).toBe('boolean')
+      expect(typeof config.delayMs).toBe('number')
+      expect(typeof config.timeoutMs).toBe('number')
+      expect(typeof config.lockTtlMs).toBe('number')
+    })
+  })
+
+  describe('isEnabled', () => {
+    it('should return boolean', async () => {
+      const enabled = await userMessageQueueService.isEnabled()
+      expect(typeof enabled).toBe('boolean')
+    })
+  })
+
+  describe('acquireQueueLock', () => {
+    afterEach(() => {
+      jest.restoreAllMocks()
+    })
+
+    it('should acquire lock immediately when no lock exists', async () => {
+      jest.spyOn(userMessageQueueService, 'getConfig').mockResolvedValue({
+        enabled: true,
+        delayMs: 200,
+        timeoutMs: 30000,
+        lockTtlMs: 120000
+      })
+      jest.spyOn(redis, 'acquireUserMessageLock').mockResolvedValue({
+        acquired: true,
+        waitMs: 0
+      })
+
+      const result = await userMessageQueueService.acquireQueueLock('acct-1', 'req-1')
+
+      expect(result.acquired).toBe(true)
+      expect(result.requestId).toBe('req-1')
+      expect(result.error).toBeUndefined()
+    })
+
+    it('should skip lock acquisition when queue disabled', async () => {
+      jest.spyOn(userMessageQueueService, 'getConfig').mockResolvedValue({
+        enabled: false,
+        delayMs: 200,
+        timeoutMs: 30000,
+        lockTtlMs: 120000
+      })
+      const acquireSpy = jest.spyOn(redis, 'acquireUserMessageLock')
+
+      const result = await userMessageQueueService.acquireQueueLock('acct-1')
+
+      expect(result.acquired).toBe(true)
+      expect(result.skipped).toBe(true)
+      expect(acquireSpy).not.toHaveBeenCalled()
+    })
+
+    it('should generate requestId when not provided', async () => {
+      jest.spyOn(userMessageQueueService, 'getConfig').mockResolvedValue({
+        enabled: true,
+        delayMs: 200,
+        timeoutMs: 30000,
+        lockTtlMs: 120000
+      })
+      jest.spyOn(redis, 'acquireUserMessageLock').mockResolvedValue({
+        acquired: true,
+        waitMs: 0
+      })
+
+      const result = await userMessageQueueService.acquireQueueLock('acct-1')
+
+      expect(result.acquired).toBe(true)
+      expect(result.requestId).toBeDefined()
+      expect(result.requestId.length).toBeGreaterThan(0)
+    })
+
+    it('should wait and retry when lock is held by another request', async () => {
+      jest.spyOn(userMessageQueueService, 'getConfig').mockResolvedValue({
+        enabled: true,
+        delayMs: 200,
+        timeoutMs: 1000,
+        lockTtlMs: 120000
+      })
+
+      let callCount = 0
+      jest.spyOn(redis, 'acquireUserMessageLock').mockImplementation(async () => {
+        callCount++
+        if (callCount < 3) {
+          return { acquired: false, waitMs: -1 } // lock held
+        }
+        return { acquired: true, waitMs: 0 }
+      })
+
+      // Mock sleep to speed up test
+      jest.spyOn(userMessageQueueService, '_sleep').mockResolvedValue(undefined)
+
+      const result = await userMessageQueueService.acquireQueueLock('acct-1', 'req-1')
+
+      expect(result.acquired).toBe(true)
+      expect(callCount).toBe(3)
+    })
+
+    it('should respect delay when previous request just completed', async () => {
+      jest.spyOn(userMessageQueueService, 'getConfig').mockResolvedValue({
+        enabled: true,
+        delayMs: 200,
+        timeoutMs: 1000,
+        lockTtlMs: 120000
+      })
+
+      let callCount = 0
+      jest.spyOn(redis, 'acquireUserMessageLock').mockImplementation(async () => {
+        callCount++
+        if (callCount === 1) {
+          return { acquired: false, waitMs: 150 } // need to wait 150ms for delay
+        }
+        return { acquired: true, waitMs: 0 }
+      })
+
+      const sleepSpy = jest.spyOn(userMessageQueueService, '_sleep').mockResolvedValue(undefined)
+
+      const result = await userMessageQueueService.acquireQueueLock('acct-1', 'req-1')
+
+      expect(result.acquired).toBe(true)
+      expect(sleepSpy).toHaveBeenCalledWith(150) // Should wait for delay
+    })
+
+    it('should timeout and return error when wait exceeds timeout', async () => {
+      jest.spyOn(userMessageQueueService, 'getConfig').mockResolvedValue({
+        enabled: true,
+        delayMs: 200,
+        timeoutMs: 100, // very short timeout
+        lockTtlMs: 120000
+      })
+
+      jest.spyOn(redis, 'acquireUserMessageLock').mockResolvedValue({
+        acquired: false,
+        waitMs: -1 // always held
+      })
+
+      // Use real timers for timeout test but mock sleep to be instant
+      jest.spyOn(userMessageQueueService, '_sleep').mockImplementation(async () => {
+        // Simulate time passing
+        await new Promise((resolve) => setTimeout(resolve, 60))
+      })
+
+      const result = await userMessageQueueService.acquireQueueLock('acct-1', 'req-1', 100)
+
+      expect(result.acquired).toBe(false)
+      expect(result.error).toBe('queue_timeout')
+    })
+  })
+
+  describe('releaseQueueLock', () => {
+    afterEach(() => {
+      jest.restoreAllMocks()
+    })
+
+    it('should release lock successfully when holding the lock', async () => {
+      jest.spyOn(redis, 'releaseUserMessageLock').mockResolvedValue(true)
+
+      const result = await userMessageQueueService.releaseQueueLock('acct-1', 'req-1')
+
+      expect(result).toBe(true)
+      expect(redis.releaseUserMessageLock).toHaveBeenCalledWith('acct-1', 'req-1')
+    })
+
+    it('should return false when not holding the lock', async () => {
+      jest.spyOn(redis, 'releaseUserMessageLock').mockResolvedValue(false)
+
+      const result = await userMessageQueueService.releaseQueueLock('acct-1', 'req-1')
+
+      expect(result).toBe(false)
+    })
+
+    it('should return false when accountId is missing', async () => {
+      const releaseSpy = jest.spyOn(redis, 'releaseUserMessageLock')
+
+      const result = await userMessageQueueService.releaseQueueLock(null, 'req-1')
+
+      expect(result).toBe(false)
+      expect(releaseSpy).not.toHaveBeenCalled()
+    })
+
+    it('should return false when requestId is missing', async () => {
+      const releaseSpy = jest.spyOn(redis, 'releaseUserMessageLock')
+
+      const result = await userMessageQueueService.releaseQueueLock('acct-1', null)
+
+      expect(result).toBe(false)
+      expect(releaseSpy).not.toHaveBeenCalled()
+    })
+  })
+
+  describe('queue serialization behavior', () => {
+    afterEach(() => {
+      jest.restoreAllMocks()
+    })
+
+    it('should allow different accounts to acquire locks simultaneously', async () => {
+      jest.spyOn(userMessageQueueService, 'getConfig').mockResolvedValue({
+        enabled: true,
+        delayMs: 200,
+        timeoutMs: 30000,
+        lockTtlMs: 120000
+      })
+      jest.spyOn(redis, 'acquireUserMessageLock').mockResolvedValue({
+        acquired: true,
+        waitMs: 0
+      })
+
+      const [result1, result2] = await Promise.all([
+        userMessageQueueService.acquireQueueLock('acct-1', 'req-1'),
+        userMessageQueueService.acquireQueueLock('acct-2', 'req-2')
+      ])
+
+      expect(result1.acquired).toBe(true)
+      expect(result2.acquired).toBe(true)
+    })
+
+    it('should serialize requests for same account', async () => {
+      jest.spyOn(userMessageQueueService, 'getConfig').mockResolvedValue({
+        enabled: true,
+        delayMs: 50,
+        timeoutMs: 5000,
+        lockTtlMs: 120000
+      })
+
+      const lockState = { held: false, holderId: null }
+
+      jest
+        .spyOn(redis, 'acquireUserMessageLock')
+        .mockImplementation(async (accountId, requestId) => {
+          if (!lockState.held) {
+            lockState.held = true
+            lockState.holderId = requestId
+            return { acquired: true, waitMs: 0 }
+          }
+          return { acquired: false, waitMs: -1 }
+        })
+
+      jest
+        .spyOn(redis, 'releaseUserMessageLock')
+        .mockImplementation(async (accountId, requestId) => {
+          if (lockState.holderId === requestId) {
+            lockState.held = false
+            lockState.holderId = null
+            return true
+          }
+          return false
+        })
+
+      jest.spyOn(userMessageQueueService, '_sleep').mockResolvedValue(undefined)
+
+      // First request acquires lock
+      const result1 = await userMessageQueueService.acquireQueueLock('acct-1', 'req-1')
+      expect(result1.acquired).toBe(true)
+
+      // Second request should fail to acquire (lock held)
+      const acquirePromise = userMessageQueueService.acquireQueueLock('acct-1', 'req-2', 200)
+
+      // Release first lock
+      await userMessageQueueService.releaseQueueLock('acct-1', 'req-1')
+
+      // Now second request should acquire
+      const result2 = await acquirePromise
+      expect(result2.acquired).toBe(true)
+    })
+  })
+})

web/admin-spa/src/components/accounts/AccountForm.vue

@@ -1843,7 +1843,7 @@
                     >
                       请从已登录 Gemini CLI 的机器上获取
                       <code class="rounded bg-blue-100 px-1 py-0.5 font-mono dark:bg-blue-900/50"
-                        >~/.config/gemini/credentials.json</code
+                        >~/.config/.gemini/oauth_creds.json</code
                       >
                       文件中的凭证。
                     </p>

web/admin-spa/src/views/AccountsView.vue

@@ -2049,7 +2049,7 @@ const bindingCounts = ref({}) // 轻量级绑定计数，用于显示"绑定: X
 const accountGroups = ref([])
 const groupFilter = ref('all')
 const platformFilter = ref('all')
-const statusFilter = ref('normal') // 状态过滤 (normal/rateLimited/other/all)
+const statusFilter = ref('all') // 状态过滤 (normal/rateLimited/other/all)
 const searchKeyword = ref('')
 const PAGE_SIZE_STORAGE_KEY = 'accountsPageSize'
 const getInitialPageSize = () => {
@@ -2804,11 +2804,12 @@ const loadAccounts = async (forceReload = false) => {
     let openaiResponsesRaw = []
 
     const appendAccounts = (platform, data) => {
-      if (!data || data.length === 0) return
+      const list = Array.isArray(data) ? data : []
+      if (list.length === 0) return
 
       switch (platform) {
         case 'claude': {
-          const items = data.map((acc) => {
+          const items = list.map((acc) => {
             const boundApiKeysCount = counts.claudeAccountId?.[acc.id] || 0
             return { ...acc, platform: 'claude', boundApiKeysCount }
           })
@@ -2816,7 +2817,7 @@ const loadAccounts = async (forceReload = false) => {
           break
         }
         case 'claude-console': {
-          const items = data.map((acc) => {
+          const items = list.map((acc) => {
             const boundApiKeysCount = counts.claudeConsoleAccountId?.[acc.id] || 0
             return { ...acc, platform: 'claude-console', boundApiKeysCount }
           })
@@ -2824,12 +2825,12 @@ const loadAccounts = async (forceReload = false) => {
           break
         }
         case 'bedrock': {
-          const items = data.map((acc) => ({ ...acc, platform: 'bedrock', boundApiKeysCount: 0 }))
+          const items = list.map((acc) => ({ ...acc, platform: 'bedrock', boundApiKeysCount: 0 }))
           allAccounts.push(...items)
           break
         }
         case 'gemini': {
-          const items = data.map((acc) => {
+          const items = list.map((acc) => {
             const boundApiKeysCount = counts.geminiAccountId?.[acc.id] || 0
             return { ...acc, platform: 'gemini', boundApiKeysCount }
           })
@@ -2837,7 +2838,7 @@ const loadAccounts = async (forceReload = false) => {
           break
         }
         case 'openai': {
-          const items = data.map((acc) => {
+          const items = list.map((acc) => {
             const boundApiKeysCount = counts.openaiAccountId?.[acc.id] || 0
             return { ...acc, platform: 'openai', boundApiKeysCount }
           })
@@ -2845,7 +2846,7 @@ const loadAccounts = async (forceReload = false) => {
           break
         }
         case 'azure_openai': {
-          const items = data.map((acc) => {
+          const items = list.map((acc) => {
             const boundApiKeysCount = counts.azureOpenaiAccountId?.[acc.id] || 0
             return { ...acc, platform: 'azure_openai', boundApiKeysCount }
           })
@@ -2853,16 +2854,16 @@ const loadAccounts = async (forceReload = false) => {
           break
         }
         case 'openai-responses': {
-          openaiResponsesRaw = data
+          openaiResponsesRaw = list
           break
         }
         case 'ccr': {
-          const items = data.map((acc) => ({ ...acc, platform: 'ccr', boundApiKeysCount: 0 }))
+          const items = list.map((acc) => ({ ...acc, platform: 'ccr', boundApiKeysCount: 0 }))
           allAccounts.push(...items)
           break
         }
         case 'droid': {
-          const items = data.map((acc) => {
+          const items = list.map((acc) => {
             const boundApiKeysCount = counts.droidAccountId?.[acc.id] || acc.boundApiKeysCount || 0
             return { ...acc, platform: 'droid', boundApiKeysCount }
           })
@@ -2870,7 +2871,7 @@ const loadAccounts = async (forceReload = false) => {
           break
         }
         case 'gemini-api': {
-          const items = data.map((acc) => {
+          const items = list.map((acc) => {
             const boundApiKeysCount = counts.geminiAccountId?.[`api:${acc.id}`] || 0
             return { ...acc, platform: 'gemini-api', boundApiKeysCount }
           })

web/admin-spa/src/views/SettingsView.vue

@@ -804,6 +804,214 @@
               </div>
             </div>
 
+            <!-- 用户消息串行队列 -->
+            <div
+              class="mb-6 rounded-lg bg-white/80 p-6 shadow-lg backdrop-blur-sm dark:bg-gray-800/80"
+            >
+              <div class="flex items-center justify-between">
+                <div>
+                  <div class="flex items-center">
+                    <div
+                      class="mr-3 flex h-10 w-10 items-center justify-center rounded-full bg-gradient-to-br from-cyan-500 to-teal-600 text-white shadow-lg"
+                    >
+                      <i class="fas fa-list-ol"></i>
+                    </div>
+                    <div>
+                      <h2 class="text-lg font-semibold text-gray-800 dark:text-gray-200">
+                        用户消息串行队列
+                      </h2>
+                      <p class="mt-1 text-sm text-gray-600 dark:text-gray-400">
+                        启用后，同一账户的用户消息请求将串行执行，并在请求之间添加延迟，防止触发上游限流
+                      </p>
+                    </div>
+                  </div>
+                </div>
+                <label class="relative inline-flex cursor-pointer items-center">
+                  <input
+                    v-model="claudeConfig.userMessageQueueEnabled"
+                    class="peer sr-only"
+                    type="checkbox"
+                    @change="saveClaudeConfig"
+                  />
+                  <div
+                    class="peer h-6 w-11 rounded-full bg-gray-200 after:absolute after:left-[2px] after:top-[2px] after:h-5 after:w-5 after:rounded-full after:border after:border-gray-300 after:bg-white after:transition-all after:content-[''] peer-checked:bg-teal-500 peer-checked:after:translate-x-full peer-checked:after:border-white peer-focus:outline-none peer-focus:ring-4 peer-focus:ring-teal-300 dark:border-gray-600 dark:bg-gray-700 dark:peer-focus:ring-teal-800"
+                  ></div>
+                </label>
+              </div>
+
+              <!-- 队列配置详情（仅在启用时显示） -->
+              <div v-if="claudeConfig.userMessageQueueEnabled" class="mt-6 space-y-4">
+                <!-- 请求间隔 -->
+                <div>
+                  <label class="block text-sm font-medium text-gray-700 dark:text-gray-300">
+                    <i class="fas fa-hourglass-half mr-2 text-gray-400"></i>
+                    请求间隔（毫秒）
+                  </label>
+                  <input
+                    v-model.number="claudeConfig.userMessageQueueDelayMs"
+                    class="mt-1 block w-full max-w-xs rounded-lg border border-gray-300 bg-white px-3 py-2 shadow-sm focus:border-teal-500 focus:outline-none focus:ring-2 focus:ring-teal-500/20 dark:border-gray-500 dark:bg-gray-700 dark:text-white sm:text-sm"
+                    max="10000"
+                    min="0"
+                    placeholder="200"
+                    type="number"
+                    @change="saveClaudeConfig"
+                  />
+                  <p class="mt-1 text-xs text-gray-500 dark:text-gray-400">
+                    同一账户的用户消息请求之间的最小间隔时间（0-10000毫秒）
+                  </p>
+                </div>
+
+                <!-- 队列超时 -->
+                <div>
+                  <label class="block text-sm font-medium text-gray-700 dark:text-gray-300">
+                    <i class="fas fa-stopwatch mr-2 text-gray-400"></i>
+                    队列超时（毫秒）
+                  </label>
+                  <input
+                    v-model.number="claudeConfig.userMessageQueueTimeoutMs"
+                    class="mt-1 block w-full max-w-xs rounded-lg border border-gray-300 bg-white px-3 py-2 shadow-sm focus:border-teal-500 focus:outline-none focus:ring-2 focus:ring-teal-500/20 dark:border-gray-500 dark:bg-gray-700 dark:text-white sm:text-sm"
+                    max="300000"
+                    min="1000"
+                    placeholder="30000"
+                    type="number"
+                    @change="saveClaudeConfig"
+                  />
+                  <p class="mt-1 text-xs text-gray-500 dark:text-gray-400">
+                    请求在队列中等待的最大时间，超时将返回 503 错误（1000-300000毫秒）
+                  </p>
+                </div>
+              </div>
+
+              <div class="mt-4 rounded-lg bg-teal-50 p-4 dark:bg-teal-900/20">
+                <div class="flex">
+                  <i class="fas fa-info-circle mt-0.5 text-teal-500"></i>
+                  <div class="ml-3">
+                    <p class="text-sm text-teal-700 dark:text-teal-300">
+                      <strong>工作原理：</strong>系统检测请求中最后一条消息的
+                      <code class="rounded bg-teal-100 px-1 dark:bg-teal-800">role</code>
+                      是否为
+                      <code class="rounded bg-teal-100 px-1 dark:bg-teal-800">user</code
+                      >。用户消息请求需要排队串行执行，而工具调用结果、助手消息续传等不受此限制。
+                    </p>
+                  </div>
+                </div>
+              </div>
+            </div>
+
+            <!-- 并发请求排队 -->
+            <div
+              class="mb-6 rounded-lg bg-white/80 p-6 shadow-lg backdrop-blur-sm dark:bg-gray-800/80"
+            >
+              <div class="flex items-center justify-between">
+                <div class="flex items-center">
+                  <div
+                    class="flex h-12 w-12 items-center justify-center rounded-lg bg-gradient-to-r from-blue-500 to-cyan-500 text-white shadow-lg"
+                  >
+                    <i class="fas fa-layer-group text-xl"></i>
+                  </div>
+                  <div class="ml-4">
+                    <h4 class="text-lg font-semibold text-gray-900 dark:text-white">
+                      并发请求排队
+                    </h4>
+                    <p class="text-sm text-gray-500 dark:text-gray-400">
+                      当 API Key 并发请求超限时进入队列等待，而非直接拒绝
+                    </p>
+                  </div>
+                </div>
+                <label class="relative inline-flex cursor-pointer items-center">
+                  <input
+                    v-model="claudeConfig.concurrentRequestQueueEnabled"
+                    class="peer sr-only"
+                    type="checkbox"
+                    @change="saveClaudeConfig"
+                  />
+                  <div
+                    class="peer h-6 w-11 rounded-full bg-gray-200 after:absolute after:left-[2px] after:top-[2px] after:h-5 after:w-5 after:rounded-full after:border after:border-gray-300 after:bg-white after:transition-all after:content-[''] peer-checked:bg-blue-500 peer-checked:after:translate-x-full peer-checked:after:border-white peer-focus:outline-none peer-focus:ring-4 peer-focus:ring-blue-300 dark:border-gray-600 dark:bg-gray-700 dark:peer-focus:ring-blue-800"
+                  ></div>
+                </label>
+              </div>
+
+              <!-- 排队配置详情（仅在启用时显示） -->
+              <div v-if="claudeConfig.concurrentRequestQueueEnabled" class="mt-6 space-y-4">
+                <!-- 固定最小排队数 -->
+                <div>
+                  <label class="block text-sm font-medium text-gray-700 dark:text-gray-300">
+                    <i class="fas fa-list-ol mr-2 text-gray-400"></i>
+                    固定最小排队数
+                  </label>
+                  <input
+                    v-model.number="claudeConfig.concurrentRequestQueueMaxSize"
+                    class="mt-1 block w-full max-w-xs rounded-lg border border-gray-300 bg-white px-3 py-2 shadow-sm focus:border-blue-500 focus:outline-none focus:ring-2 focus:ring-blue-500/20 dark:border-gray-500 dark:bg-gray-700 dark:text-white sm:text-sm"
+                    max="100"
+                    min="1"
+                    placeholder="3"
+                    type="number"
+                    @change="saveClaudeConfig"
+                  />
+                  <p class="mt-1 text-xs text-gray-500 dark:text-gray-400">
+                    最大排队数的固定最小值（1-100）
+                  </p>
+                </div>
+
+                <!-- 排队数倍数 -->
+                <div>
+                  <label class="block text-sm font-medium text-gray-700 dark:text-gray-300">
+                    <i class="fas fa-times mr-2 text-gray-400"></i>
+                    排队数倍数
+                  </label>
+                  <input
+                    v-model.number="claudeConfig.concurrentRequestQueueMaxSizeMultiplier"
+                    class="mt-1 block w-full max-w-xs rounded-lg border border-gray-300 bg-white px-3 py-2 shadow-sm focus:border-blue-500 focus:outline-none focus:ring-2 focus:ring-blue-500/20 dark:border-gray-500 dark:bg-gray-700 dark:text-white sm:text-sm"
+                    max="10"
+                    min="0"
+                    placeholder="1"
+                    step="0.5"
+                    type="number"
+                    @change="saveClaudeConfig"
+                  />
+                  <p class="mt-1 text-xs text-gray-500 dark:text-gray-400">
+                    最大排队数 = MAX(倍数 × 并发限制, 固定值)，设为 0 则仅使用固定值
+                  </p>
+                </div>
+
+                <!-- 排队超时时间 -->
+                <div>
+                  <label class="block text-sm font-medium text-gray-700 dark:text-gray-300">
+                    <i class="fas fa-stopwatch mr-2 text-gray-400"></i>
+                    排队超时时间（毫秒）
+                  </label>
+                  <input
+                    v-model.number="claudeConfig.concurrentRequestQueueTimeoutMs"
+                    class="mt-1 block w-full max-w-xs rounded-lg border border-gray-300 bg-white px-3 py-2 shadow-sm focus:border-blue-500 focus:outline-none focus:ring-2 focus:ring-blue-500/20 dark:border-gray-500 dark:bg-gray-700 dark:text-white sm:text-sm"
+                    max="300000"
+                    min="5000"
+                    placeholder="10000"
+                    type="number"
+                    @change="saveClaudeConfig"
+                  />
+                  <p class="mt-1 text-xs text-gray-500 dark:text-gray-400">
+                    请求在排队中等待的最大时间，超时将返回 429 错误（5秒-5分钟，默认10秒）
+                  </p>
+                </div>
+              </div>
+
+              <div class="mt-4 rounded-lg bg-blue-50 p-4 dark:bg-blue-900/20">
+                <div class="flex">
+                  <i class="fas fa-info-circle mt-0.5 text-blue-500"></i>
+                  <div class="ml-3">
+                    <p class="text-sm text-blue-700 dark:text-blue-300">
+                      <strong>工作原理：</strong>当 API Key 的并发请求超过
+                      <code class="rounded bg-blue-100 px-1 dark:bg-blue-800"
+                        >concurrencyLimit</code
+                      >
+                      时，超限请求会进入队列等待而非直接返回 429。适合 Claude Code Agent
+                      并行工具调用场景。
+                    </p>
+                  </div>
+                </div>
+              </div>
+            </div>
+
             <!-- 配置更新信息 -->
             <div
               v-if="claudeConfig.updatedAt"
@@ -1469,6 +1677,13 @@ const claudeConfig = ref({
   globalSessionBindingEnabled: false,
   sessionBindingErrorMessage: '你的本地session已污染，请清理后使用。',
   sessionBindingTtlDays: 30,
+  userMessageQueueEnabled: false, // 与后端默认值保持一致
+  userMessageQueueDelayMs: 200,
+  userMessageQueueTimeoutMs: 5000, // 与后端默认值保持一致（优化后锁持有时间短无需长等待）
+  concurrentRequestQueueEnabled: false,
+  concurrentRequestQueueMaxSize: 3,
+  concurrentRequestQueueMaxSizeMultiplier: 0,
+  concurrentRequestQueueTimeoutMs: 10000,
   updatedAt: null,
   updatedBy: null
 })
@@ -1738,6 +1953,14 @@ const loadClaudeConfig = async () => {
         sessionBindingErrorMessage:
           response.config?.sessionBindingErrorMessage || '你的本地session已污染，请清理后使用。',
         sessionBindingTtlDays: response.config?.sessionBindingTtlDays ?? 30,
+        userMessageQueueEnabled: response.config?.userMessageQueueEnabled ?? false, // 与后端默认值保持一致
+        userMessageQueueDelayMs: response.config?.userMessageQueueDelayMs ?? 200,
+        userMessageQueueTimeoutMs: response.config?.userMessageQueueTimeoutMs ?? 5000, // 与后端默认值保持一致
+        concurrentRequestQueueEnabled: response.config?.concurrentRequestQueueEnabled ?? false,
+        concurrentRequestQueueMaxSize: response.config?.concurrentRequestQueueMaxSize ?? 3,
+        concurrentRequestQueueMaxSizeMultiplier:
+          response.config?.concurrentRequestQueueMaxSizeMultiplier ?? 0,
+        concurrentRequestQueueTimeoutMs: response.config?.concurrentRequestQueueTimeoutMs ?? 10000,
         updatedAt: response.config?.updatedAt || null,
         updatedBy: response.config?.updatedBy || null
       }
@@ -1762,7 +1985,15 @@ const saveClaudeConfig = async () => {
       claudeCodeOnlyEnabled: claudeConfig.value.claudeCodeOnlyEnabled,
       globalSessionBindingEnabled: claudeConfig.value.globalSessionBindingEnabled,
       sessionBindingErrorMessage: claudeConfig.value.sessionBindingErrorMessage,
-      sessionBindingTtlDays: claudeConfig.value.sessionBindingTtlDays
+      sessionBindingTtlDays: claudeConfig.value.sessionBindingTtlDays,
+      userMessageQueueEnabled: claudeConfig.value.userMessageQueueEnabled,
+      userMessageQueueDelayMs: claudeConfig.value.userMessageQueueDelayMs,
+      userMessageQueueTimeoutMs: claudeConfig.value.userMessageQueueTimeoutMs,
+      concurrentRequestQueueEnabled: claudeConfig.value.concurrentRequestQueueEnabled,
+      concurrentRequestQueueMaxSize: claudeConfig.value.concurrentRequestQueueMaxSize,
+      concurrentRequestQueueMaxSizeMultiplier:
+        claudeConfig.value.concurrentRequestQueueMaxSizeMultiplier,
+      concurrentRequestQueueTimeoutMs: claudeConfig.value.concurrentRequestQueueTimeoutMs
     }
 
     const response = await apiClient.put('/admin/claude-relay-config', payload, {