refactor: 改进分布式任务锁的清理策略

实现两种锁清理模式：
- 单实例模式（默认）：启动时使用 Lua 脚本原子性清理旧锁，解决 worker 重启时���锁残留问题
- 多实例模式：使用 NX 选项竞争锁，依赖 TTL 处理异常退出

可通过 SINGLE_INSTANCE_MODE 环境变量控制模式选择。

.env.example

@@ -1,8 +1,16 @@
 # ==================== 必须配置（启动前） ====================
 # 以下配置项必须在项目启动前设置
 
-# 数据库密码
+# 数据库配置
+DB_HOST=localhost
+DB_PORT=5432
+DB_USER=postgres
+DB_NAME=aether
 DB_PASSWORD=your_secure_password_here
+
+# Redis 配置
+REDIS_HOST=localhost
+REDIS_PORT=6379
 REDIS_PASSWORD=your_redis_password_here
 
 # JWT密钥（使用 python generate_keys.py 生成）

.github/workflows/docker-publish.yml

@@ -15,6 +15,8 @@ env:
   REGISTRY: ghcr.io
   BASE_IMAGE_NAME: fawney19/aether-base
   APP_IMAGE_NAME: fawney19/aether
+  # Files that affect base image - used for hash calculation
+  BASE_FILES: "Dockerfile.base pyproject.toml frontend/package.json frontend/package-lock.json"
 
 jobs:
   check-base-changes:
@@ -23,8 +25,13 @@ jobs:
       base_changed: ${{ steps.check.outputs.base_changed }}
     steps:
       - uses: actions/checkout@v4
+
+      - name: Log in to Container Registry
+        uses: docker/login-action@v3
         with:
-          fetch-depth: 2
+          registry: ${{ env.REGISTRY }}
+          username: ${{ github.actor }}
+          password: ${{ secrets.GITHUB_TOKEN }}
 
       - name: Check if base image needs rebuild
         id: check
@@ -34,10 +41,26 @@ jobs:
             exit 0
           fi
 
-          # Check if base-related files changed
-          if git diff --name-only HEAD~1 HEAD | grep -qE '^(Dockerfile\.base|pyproject\.toml|frontend/package.*\.json)$'; then
+          # Calculate current hash of base-related files
+          CURRENT_HASH=$(cat ${{ env.BASE_FILES }} 2>/dev/null | sha256sum | cut -d' ' -f1)
+          echo "Current base files hash: $CURRENT_HASH"
+
+          # Try to get hash label from remote image config
+          # Pull the image config and extract labels
+          REMOTE_HASH=""
+          if docker pull ${{ env.REGISTRY }}/${{ env.BASE_IMAGE_NAME }}:latest 2>/dev/null; then
+            REMOTE_HASH=$(docker inspect ${{ env.REGISTRY }}/${{ env.BASE_IMAGE_NAME }}:latest --format '{{ index .Config.Labels "org.opencontainers.image.base.hash" }}' 2>/dev/null) || true
+          fi
+
+          if [ -z "$REMOTE_HASH" ] || [ "$REMOTE_HASH" == "<no value>" ]; then
+            # No remote image or no hash label, need to rebuild
+            echo "No remote base image or hash label found, need rebuild"
+            echo "base_changed=true" >> $GITHUB_OUTPUT
+          elif [ "$CURRENT_HASH" != "$REMOTE_HASH" ]; then
+            echo "Hash mismatch: remote=$REMOTE_HASH, current=$CURRENT_HASH"
             echo "base_changed=true" >> $GITHUB_OUTPUT
           else
+            echo "Hash matches, no rebuild needed"
             echo "base_changed=false" >> $GITHUB_OUTPUT
           fi
 
@@ -61,6 +84,12 @@ jobs:
           username: ${{ github.actor }}
           password: ${{ secrets.GITHUB_TOKEN }}
 
+      - name: Calculate base files hash
+        id: hash
+        run: |
+          HASH=$(cat ${{ env.BASE_FILES }} 2>/dev/null | sha256sum | cut -d' ' -f1)
+          echo "hash=$HASH" >> $GITHUB_OUTPUT
+
       - name: Extract metadata for base image
         id: meta
         uses: docker/metadata-action@v5
@@ -69,6 +98,8 @@ jobs:
           tags: |
             type=raw,value=latest
             type=sha,prefix=
+          labels: |
+            org.opencontainers.image.base.hash=${{ steps.hash.outputs.hash }}
 
       - name: Build and push base image
         uses: docker/build-push-action@v5
@@ -117,7 +148,7 @@ jobs:
 
       - name: Update Dockerfile.app to use registry base image
         run: |
-          sed -i "s|FROM aether-base:latest|FROM ${{ env.REGISTRY }}/${{ env.BASE_IMAGE_NAME }}:latest|g" Dockerfile.app
+          sed -i "s|FROM aether-base:latest AS builder|FROM ${{ env.REGISTRY }}/${{ env.BASE_IMAGE_NAME }}:latest AS builder|g" Dockerfile.app
 
       - name: Build and push app image
         uses: docker/build-push-action@v5

Dockerfile.app

@@ -1,16 +1,134 @@
-# 应用镜像：基于基础镜像，只复制代码（秒级构建）
+# 运行镜像：从 base 提取产物到精简运行时
 # 构建命令: docker build -f Dockerfile.app -t aether-app:latest .
-FROM aether-base:latest
+# 用于 GitHub Actions CI（官方源）
+FROM aether-base:latest AS builder
 
 WORKDIR /app
 
+# 复制前端源码并构建
+COPY frontend/ ./frontend/
+RUN cd frontend && npm run build
+
+# ==================== 运行时镜像 ====================
+FROM python:3.12-slim
+
+WORKDIR /app
+
+# 运行时依赖（无 gcc/nodejs/npm）
+RUN apt-get update && apt-get install -y \
+    nginx \
+    supervisor \
+    libpq5 \
+    curl \
+    && rm -rf /var/lib/apt/lists/*
+
+# 从 base 镜像复制 Python 包
+COPY --from=builder /usr/local/lib/python3.12/site-packages /usr/local/lib/python3.12/site-packages
+
+# 只复制需要的 Python 可执行文件
+COPY --from=builder /usr/local/bin/gunicorn /usr/local/bin/
+COPY --from=builder /usr/local/bin/uvicorn /usr/local/bin/
+COPY --from=builder /usr/local/bin/alembic /usr/local/bin/
+
+# 从 builder 阶段复制前端构建产物
+COPY --from=builder /app/frontend/dist /usr/share/nginx/html
+
 # 复制后端代码
 COPY src/ ./src/
 COPY alembic.ini ./
 COPY alembic/ ./alembic/
 
-# 构建前端（使用基础镜像中已安装的 node_modules）
-COPY frontend/ /tmp/frontend/
-RUN cd /tmp/frontend && npm run build && \
-    cp -r dist/* /usr/share/nginx/html/ && \
-    rm -rf /tmp/frontend
+# Nginx 配置模板
+RUN printf '%s\n' \
+'server {' \
+'    listen 80;' \
+'    server_name _;' \
+'    root /usr/share/nginx/html;' \
+'    index index.html;' \
+'    client_max_body_size 100M;' \
+'' \
+'    location ~* \.(js|css|png|jpg|jpeg|gif|ico|svg|woff|woff2|ttf|eot)$ {' \
+'        expires 1y;' \
+'        add_header Cache-Control "public, no-transform";' \
+'        try_files $uri =404;' \
+'    }' \
+'' \
+'    location ~ ^/(src|node_modules)/ {' \
+'        deny all;' \
+'        return 404;' \
+'    }' \
+'' \
+'    location ~ ^/(dashboard|admin|login)(/|$) {' \
+'        try_files $uri $uri/ /index.html;' \
+'    }' \
+'' \
+'    location / {' \
+'        try_files $uri $uri/ @backend;' \
+'    }' \
+'' \
+'    location @backend {' \
+'        proxy_pass http://127.0.0.1:PORT_PLACEHOLDER;' \
+'        proxy_http_version 1.1;' \
+'        proxy_set_header Host $host;' \
+'        proxy_set_header X-Real-IP $remote_addr;' \
+'        proxy_set_header X-Forwarded-For $proxy_add_x_forwarded_for;' \
+'        proxy_set_header X-Forwarded-Proto $scheme;' \
+'        proxy_set_header Connection "";' \
+'        proxy_set_header Accept $http_accept;' \
+'        proxy_set_header Content-Type $content_type;' \
+'        proxy_set_header Authorization $http_authorization;' \
+'        proxy_set_header X-Api-Key $http_x_api_key;' \
+'        proxy_buffering off;' \
+'        proxy_cache off;' \
+'        proxy_request_buffering off;' \
+'        chunked_transfer_encoding on;' \
+'        gzip off;' \
+'        add_header X-Accel-Buffering no;' \
+'        proxy_connect_timeout 600s;' \
+'        proxy_send_timeout 600s;' \
+'        proxy_read_timeout 600s;' \
+'    }' \
+'}' > /etc/nginx/sites-available/default.template
+
+# Supervisor 配置
+RUN printf '%s\n' \
+'[supervisord]' \
+'nodaemon=true' \
+'logfile=/var/log/supervisor/supervisord.log' \
+'pidfile=/var/run/supervisord.pid' \
+'' \
+'[program:nginx]' \
+'command=/bin/bash -c "sed \"s/PORT_PLACEHOLDER/${PORT:-8084}/g\" /etc/nginx/sites-available/default.template > /etc/nginx/sites-available/default && /usr/sbin/nginx -g \"daemon off;\""' \
+'autostart=true' \
+'autorestart=true' \
+'stdout_logfile=/var/log/nginx/access.log' \
+'stderr_logfile=/var/log/nginx/error.log' \
+'' \
+'[program:app]' \
+'command=gunicorn src.main:app --preload -w %(ENV_GUNICORN_WORKERS)s -k uvicorn.workers.UvicornWorker --bind 0.0.0.0:%(ENV_PORT)s --timeout 120 --access-logfile - --error-logfile - --log-level info' \
+'directory=/app' \
+'autostart=true' \
+'autorestart=true' \
+'stdout_logfile=/dev/stdout' \
+'stdout_logfile_maxbytes=0' \
+'stderr_logfile=/dev/stderr' \
+'stderr_logfile_maxbytes=0' \
+'environment=PYTHONUNBUFFERED=1,PYTHONIOENCODING=utf-8,LANG=C.UTF-8,LC_ALL=C.UTF-8,DOCKER_CONTAINER=true' > /etc/supervisor/conf.d/supervisord.conf
+
+# 创建目录
+RUN mkdir -p /var/log/supervisor /app/logs /app/data
+
+# 环境变量
+ENV PYTHONUNBUFFERED=1 \
+    PYTHONDONTWRITEBYTECODE=1 \
+    PYTHONIOENCODING=utf-8 \
+    LANG=C.UTF-8 \
+    LC_ALL=C.UTF-8 \
+    PORT=8084
+
+EXPOSE 80
+
+HEALTHCHECK --interval=30s --timeout=10s --start-period=5s --retries=3 \
+    CMD curl -f http://localhost/health || exit 1
+
+CMD ["/usr/bin/supervisord", "-c", "/etc/supervisor/conf.d/supervisord.conf"]

Dockerfile.app.local

@@ -0,0 +1,135 @@
+# 运行镜像：从 base 提取产物到精简运行时（国内镜像源版本）
+# 构建命令: docker build -f Dockerfile.app.local -t aether-app:latest .
+# 用于本地/国内服务器部署
+FROM aether-base:latest AS builder
+
+WORKDIR /app
+
+# 复制前端源码并构建
+COPY frontend/ ./frontend/
+RUN cd frontend && npm run build
+
+# ==================== 运行时镜像 ====================
+FROM python:3.12-slim
+
+WORKDIR /app
+
+# 运行时依赖（使用清华镜像源）
+RUN sed -i 's/deb.debian.org/mirrors.tuna.tsinghua.edu.cn/g' /etc/apt/sources.list.d/debian.sources && \
+    apt-get update && apt-get install -y \
+    nginx \
+    supervisor \
+    libpq5 \
+    curl \
+    && rm -rf /var/lib/apt/lists/*
+
+# 从 base 镜像复制 Python 包
+COPY --from=builder /usr/local/lib/python3.12/site-packages /usr/local/lib/python3.12/site-packages
+
+# 只复制需要的 Python 可执行文件
+COPY --from=builder /usr/local/bin/gunicorn /usr/local/bin/
+COPY --from=builder /usr/local/bin/uvicorn /usr/local/bin/
+COPY --from=builder /usr/local/bin/alembic /usr/local/bin/
+
+# 从 builder 阶段复制前端构建产物
+COPY --from=builder /app/frontend/dist /usr/share/nginx/html
+
+# 复制后端代码
+COPY src/ ./src/
+COPY alembic.ini ./
+COPY alembic/ ./alembic/
+
+# Nginx 配置模板
+RUN printf '%s\n' \
+'server {' \
+'    listen 80;' \
+'    server_name _;' \
+'    root /usr/share/nginx/html;' \
+'    index index.html;' \
+'    client_max_body_size 100M;' \
+'' \
+'    location ~* \.(js|css|png|jpg|jpeg|gif|ico|svg|woff|woff2|ttf|eot)$ {' \
+'        expires 1y;' \
+'        add_header Cache-Control "public, no-transform";' \
+'        try_files $uri =404;' \
+'    }' \
+'' \
+'    location ~ ^/(src|node_modules)/ {' \
+'        deny all;' \
+'        return 404;' \
+'    }' \
+'' \
+'    location ~ ^/(dashboard|admin|login)(/|$) {' \
+'        try_files $uri $uri/ /index.html;' \
+'    }' \
+'' \
+'    location / {' \
+'        try_files $uri $uri/ @backend;' \
+'    }' \
+'' \
+'    location @backend {' \
+'        proxy_pass http://127.0.0.1:PORT_PLACEHOLDER;' \
+'        proxy_http_version 1.1;' \
+'        proxy_set_header Host $host;' \
+'        proxy_set_header X-Real-IP $remote_addr;' \
+'        proxy_set_header X-Forwarded-For $proxy_add_x_forwarded_for;' \
+'        proxy_set_header X-Forwarded-Proto $scheme;' \
+'        proxy_set_header Connection "";' \
+'        proxy_set_header Accept $http_accept;' \
+'        proxy_set_header Content-Type $content_type;' \
+'        proxy_set_header Authorization $http_authorization;' \
+'        proxy_set_header X-Api-Key $http_x_api_key;' \
+'        proxy_buffering off;' \
+'        proxy_cache off;' \
+'        proxy_request_buffering off;' \
+'        chunked_transfer_encoding on;' \
+'        gzip off;' \
+'        add_header X-Accel-Buffering no;' \
+'        proxy_connect_timeout 600s;' \
+'        proxy_send_timeout 600s;' \
+'        proxy_read_timeout 600s;' \
+'    }' \
+'}' > /etc/nginx/sites-available/default.template
+
+# Supervisor 配置
+RUN printf '%s\n' \
+'[supervisord]' \
+'nodaemon=true' \
+'logfile=/var/log/supervisor/supervisord.log' \
+'pidfile=/var/run/supervisord.pid' \
+'' \
+'[program:nginx]' \
+'command=/bin/bash -c "sed \"s/PORT_PLACEHOLDER/${PORT:-8084}/g\" /etc/nginx/sites-available/default.template > /etc/nginx/sites-available/default && /usr/sbin/nginx -g \"daemon off;\""' \
+'autostart=true' \
+'autorestart=true' \
+'stdout_logfile=/var/log/nginx/access.log' \
+'stderr_logfile=/var/log/nginx/error.log' \
+'' \
+'[program:app]' \
+'command=gunicorn src.main:app --preload -w %(ENV_GUNICORN_WORKERS)s -k uvicorn.workers.UvicornWorker --bind 0.0.0.0:%(ENV_PORT)s --timeout 120 --access-logfile - --error-logfile - --log-level info' \
+'directory=/app' \
+'autostart=true' \
+'autorestart=true' \
+'stdout_logfile=/dev/stdout' \
+'stdout_logfile_maxbytes=0' \
+'stderr_logfile=/dev/stderr' \
+'stderr_logfile_maxbytes=0' \
+'environment=PYTHONUNBUFFERED=1,PYTHONIOENCODING=utf-8,LANG=C.UTF-8,LC_ALL=C.UTF-8,DOCKER_CONTAINER=true' > /etc/supervisor/conf.d/supervisord.conf
+
+# 创建目录
+RUN mkdir -p /var/log/supervisor /app/logs /app/data
+
+# 环境变量
+ENV PYTHONUNBUFFERED=1 \
+    PYTHONDONTWRITEBYTECODE=1 \
+    PYTHONIOENCODING=utf-8 \
+    LANG=C.UTF-8 \
+    LC_ALL=C.UTF-8 \
+    PORT=8084
+
+EXPOSE 80
+
+HEALTHCHECK --interval=30s --timeout=10s --start-period=5s --retries=3 \
+    CMD curl -f http://localhost/health || exit 1
+
+CMD ["/usr/bin/supervisord", "-c", "/etc/supervisor/conf.d/supervisord.conf"]

Dockerfile.base

@@ -1,124 +1,25 @@
-# 基础镜像：包含所有依赖，只在依赖变化时需要重建
+# 构建镜像：编译环境 + 预编译的依赖
 # 用于 GitHub Actions CI 构建（不使用国内镜像源）
+# 构建命令: docker build -f Dockerfile.base -t aether-base:latest .
+# 只在 pyproject.toml 或 frontend/package*.json 变化时需要重建
 FROM python:3.12-slim
 
 WORKDIR /app
 
-# 系统依赖
+# 构建工具
 RUN apt-get update && apt-get install -y \
-    nginx \
-    supervisor \
     libpq-dev \
     gcc \
-    curl \
-    gettext-base \
     nodejs \
     npm \
     && rm -rf /var/lib/apt/lists/*
 
-# Python 依赖（安装到系统，不用 -e 模式）
+# Python 依赖
 COPY pyproject.toml README.md ./
 RUN mkdir -p src && touch src/__init__.py && \
-    pip install --no-cache-dir .
+    SETUPTOOLS_SCM_PRETEND_VERSION=0.1.0 pip install --no-cache-dir . && \
+    pip cache purge
 
-# 前端依赖
-COPY frontend/package*.json /tmp/frontend/
-WORKDIR /tmp/frontend
-RUN npm ci
-
-# Nginx 配置模板
-RUN printf '%s\n' \
-'server {' \
-'    listen 80;' \
-'    server_name _;' \
-'    root /usr/share/nginx/html;' \
-'    index index.html;' \
-'    client_max_body_size 100M;' \
-'' \
-'    location ~* \.(js|css|png|jpg|jpeg|gif|ico|svg|woff|woff2|ttf|eot)$ {' \
-'        expires 1y;' \
-'        add_header Cache-Control "public, no-transform";' \
-'        try_files $uri =404;' \
-'    }' \
-'' \
-'    location ~ ^/(src|node_modules)/ {' \
-'        deny all;' \
-'        return 404;' \
-'    }' \
-'' \
-'    location ~ ^/(dashboard|admin|login)(/|$) {' \
-'        try_files $uri $uri/ /index.html;' \
-'    }' \
-'' \
-'    location / {' \
-'        try_files $uri $uri/ @backend;' \
-'    }' \
-'' \
-'    location @backend {' \
-'        proxy_pass http://127.0.0.1:PORT_PLACEHOLDER;' \
-'        proxy_http_version 1.1;' \
-'        proxy_set_header Host $host;' \
-'        proxy_set_header X-Real-IP $remote_addr;' \
-'        proxy_set_header X-Forwarded-For $proxy_add_x_forwarded_for;' \
-'        proxy_set_header X-Forwarded-Proto $scheme;' \
-'        proxy_set_header Connection "";' \
-'        proxy_set_header Accept $http_accept;' \
-'        proxy_set_header Content-Type $content_type;' \
-'        proxy_set_header Authorization $http_authorization;' \
-'        proxy_set_header X-Api-Key $http_x_api_key;' \
-'        proxy_buffering off;' \
-'        proxy_cache off;' \
-'        proxy_request_buffering off;' \
-'        chunked_transfer_encoding on;' \
-'        gzip off;' \
-'        add_header X-Accel-Buffering no;' \
-'        proxy_connect_timeout 600s;' \
-'        proxy_send_timeout 600s;' \
-'        proxy_read_timeout 600s;' \
-'    }' \
-'}' > /etc/nginx/sites-available/default.template
-
-# Supervisor 配置
-RUN printf '%s\n' \
-'[supervisord]' \
-'nodaemon=true' \
-'logfile=/var/log/supervisor/supervisord.log' \
-'pidfile=/var/run/supervisord.pid' \
-'' \
-'[program:nginx]' \
-'command=/bin/bash -c "sed \"s/PORT_PLACEHOLDER/${PORT:-8084}/g\" /etc/nginx/sites-available/default.template > /etc/nginx/sites-available/default && /usr/sbin/nginx -g \"daemon off;\""' \
-'autostart=true' \
-'autorestart=true' \
-'stdout_logfile=/var/log/nginx/access.log' \
-'stderr_logfile=/var/log/nginx/error.log' \
-'' \
-'[program:app]' \
-'command=gunicorn src.main:app -w %(ENV_GUNICORN_WORKERS)s -k uvicorn.workers.UvicornWorker --bind 0.0.0.0:%(ENV_PORT)s --timeout 120 --access-logfile - --error-logfile - --log-level info' \
-'directory=/app' \
-'autostart=true' \
-'autorestart=true' \
-'stdout_logfile=/dev/stdout' \
-'stdout_logfile_maxbytes=0' \
-'stderr_logfile=/dev/stderr' \
-'stderr_logfile_maxbytes=0' \
-'environment=PYTHONUNBUFFERED=1,PYTHONIOENCODING=utf-8,LANG=C.UTF-8,LC_ALL=C.UTF-8,DOCKER_CONTAINER=true' > /etc/supervisor/conf.d/supervisord.conf
-
-# 创建目录
-RUN mkdir -p /var/log/supervisor /app/logs /app/data /usr/share/nginx/html
-
-WORKDIR /app
-
-# 环境变量
-ENV PYTHONUNBUFFERED=1 \
-    PYTHONDONTWRITEBYTECODE=1 \
-    PYTHONIOENCODING=utf-8 \
-    LANG=C.UTF-8 \
-    LC_ALL=C.UTF-8 \
-    PORT=8084
-
-EXPOSE 80
-
-HEALTHCHECK --interval=30s --timeout=10s --start-period=5s --retries=3 \
-    CMD curl -f http://localhost/health || exit 1
-
-CMD ["/usr/bin/supervisord", "-c", "/etc/supervisor/conf.d/supervisord.conf"]
+# 前端依赖（只安装，不构建）
+COPY frontend/package*.json ./frontend/
+RUN cd frontend && npm ci

Dockerfile.base.local

@@ -1,18 +1,15 @@
-# 基础镜像：包含所有依赖，只在依赖变化时需要重建
-# 构建命令: docker build -f Dockerfile.base -t aether-base:latest .
+# 构建镜像：编译环境 + 预编译的依赖（国内镜像源版本）
+# 构建命令: docker build -f Dockerfile.base.local -t aether-base:latest .
+# 只在 pyproject.toml 或 frontend/package*.json 变化时需要重建
 FROM python:3.12-slim
 
 WORKDIR /app
 
-# 系统依赖
+# 构建工具（使用清华镜像源）
 RUN sed -i 's/deb.debian.org/mirrors.tuna.tsinghua.edu.cn/g' /etc/apt/sources.list.d/debian.sources && \
     apt-get update && apt-get install -y \
-    nginx \
-    supervisor \
     libpq-dev \
     gcc \
-    curl \
-    gettext-base \
     nodejs \
     npm \
     && rm -rf /var/lib/apt/lists/*
@@ -20,109 +17,12 @@ RUN sed -i 's/deb.debian.org/mirrors.tuna.tsinghua.edu.cn/g' /etc/apt/sources.li
 # pip 镜像源
 RUN pip config set global.index-url https://pypi.tuna.tsinghua.edu.cn/simple
 
-# Python 依赖（安装到系统，不用 -e 模式）
+# Python 依赖
 COPY pyproject.toml README.md ./
 RUN mkdir -p src && touch src/__init__.py && \
-    SETUPTOOLS_SCM_PRETEND_VERSION=0.1.0 pip install --no-cache-dir .
+    SETUPTOOLS_SCM_PRETEND_VERSION=0.1.0 pip install --no-cache-dir . && \
+    pip cache purge
 
-# 前端依赖
-COPY frontend/package*.json /tmp/frontend/
-WORKDIR /tmp/frontend
-RUN npm config set registry https://registry.npmmirror.com && npm ci
-
-# Nginx 配置模板
-RUN printf '%s\n' \
-'server {' \
-'    listen 80;' \
-'    server_name _;' \
-'    root /usr/share/nginx/html;' \
-'    index index.html;' \
-'    client_max_body_size 100M;' \
-'' \
-'    location ~* \.(js|css|png|jpg|jpeg|gif|ico|svg|woff|woff2|ttf|eot)$ {' \
-'        expires 1y;' \
-'        add_header Cache-Control "public, no-transform";' \
-'        try_files $uri =404;' \
-'    }' \
-'' \
-'    location ~ ^/(src|node_modules)/ {' \
-'        deny all;' \
-'        return 404;' \
-'    }' \
-'' \
-'    location ~ ^/(dashboard|admin|login)(/|$) {' \
-'        try_files $uri $uri/ /index.html;' \
-'    }' \
-'' \
-'    location / {' \
-'        try_files $uri $uri/ @backend;' \
-'    }' \
-'' \
-'    location @backend {' \
-'        proxy_pass http://127.0.0.1:PORT_PLACEHOLDER;' \
-'        proxy_http_version 1.1;' \
-'        proxy_set_header Host $host;' \
-'        proxy_set_header X-Real-IP $remote_addr;' \
-'        proxy_set_header X-Forwarded-For $proxy_add_x_forwarded_for;' \
-'        proxy_set_header X-Forwarded-Proto $scheme;' \
-'        proxy_set_header Connection "";' \
-'        proxy_set_header Accept $http_accept;' \
-'        proxy_set_header Content-Type $content_type;' \
-'        proxy_set_header Authorization $http_authorization;' \
-'        proxy_set_header X-Api-Key $http_x_api_key;' \
-'        proxy_buffering off;' \
-'        proxy_cache off;' \
-'        proxy_request_buffering off;' \
-'        chunked_transfer_encoding on;' \
-'        gzip off;' \
-'        add_header X-Accel-Buffering no;' \
-'        proxy_connect_timeout 600s;' \
-'        proxy_send_timeout 600s;' \
-'        proxy_read_timeout 600s;' \
-'    }' \
-'}' > /etc/nginx/sites-available/default.template
-
-# Supervisor 配置
-RUN printf '%s\n' \
-'[supervisord]' \
-'nodaemon=true' \
-'logfile=/var/log/supervisor/supervisord.log' \
-'pidfile=/var/run/supervisord.pid' \
-'' \
-'[program:nginx]' \
-'command=/bin/bash -c "sed \"s/PORT_PLACEHOLDER/${PORT:-8084}/g\" /etc/nginx/sites-available/default.template > /etc/nginx/sites-available/default && /usr/sbin/nginx -g \"daemon off;\""' \
-'autostart=true' \
-'autorestart=true' \
-'stdout_logfile=/var/log/nginx/access.log' \
-'stderr_logfile=/var/log/nginx/error.log' \
-'' \
-'[program:app]' \
-'command=gunicorn src.main:app -w %(ENV_GUNICORN_WORKERS)s -k uvicorn.workers.UvicornWorker --bind 0.0.0.0:%(ENV_PORT)s --timeout 120 --access-logfile - --error-logfile - --log-level info' \
-'directory=/app' \
-'autostart=true' \
-'autorestart=true' \
-'stdout_logfile=/dev/stdout' \
-'stdout_logfile_maxbytes=0' \
-'stderr_logfile=/dev/stderr' \
-'stderr_logfile_maxbytes=0' \
-'environment=PYTHONUNBUFFERED=1,PYTHONIOENCODING=utf-8,LANG=C.UTF-8,LC_ALL=C.UTF-8,DOCKER_CONTAINER=true' > /etc/supervisor/conf.d/supervisord.conf
-
-# 创建目录
-RUN mkdir -p /var/log/supervisor /app/logs /app/data /usr/share/nginx/html
-
-WORKDIR /app
-
-# 环境变量
-ENV PYTHONUNBUFFERED=1 \
-    PYTHONDONTWRITEBYTECODE=1 \
-    PYTHONIOENCODING=utf-8 \
-    LANG=C.UTF-8 \
-    LC_ALL=C.UTF-8 \
-    PORT=8084
-
-EXPOSE 80
-
-HEALTHCHECK --interval=30s --timeout=10s --start-period=5s --retries=3 \
-    CMD curl -f http://localhost/health || exit 1
-
-CMD ["/usr/bin/supervisord", "-c", "/etc/supervisor/conf.d/supervisord.conf"]
+# 前端依赖（只安装，不构建，使用淘宝镜像源）
+COPY frontend/package*.json ./frontend/
+RUN cd frontend && npm config set registry https://registry.npmmirror.com && npm ci

alembic/versions/20251210_baseline.py

@@ -394,6 +394,10 @@ def upgrade() -> None:
             index=True,
         ),
     )
+    # usage 表复合索引（优化常见查询）
+    op.create_index("idx_usage_user_created", "usage", ["user_id", "created_at"])
+    op.create_index("idx_usage_apikey_created", "usage", ["api_key_id", "created_at"])
+    op.create_index("idx_usage_provider_model_created", "usage", ["provider", "model", "created_at"])
 
     # ==================== user_quotas ====================
     op.create_table(

alembic/versions/20251220_1200_add_stats_daily_model_table.py

@@ -0,0 +1,86 @@
+"""add stats_daily_model table and rename provider_model_aliases
+
+Revision ID: a1b2c3d4e5f6
+Revises: f30f9936f6a2
+Create Date: 2025-12-20 12:00:00.000000+00:00
+
+"""
+from alembic import op
+import sqlalchemy as sa
+from sqlalchemy import inspect
+
+# revision identifiers, used by Alembic.
+revision = 'a1b2c3d4e5f6'
+down_revision = 'f30f9936f6a2'
+branch_labels = None
+depends_on = None
+
+
+def table_exists(table_name: str) -> bool:
+    """检查表是否存在"""
+    bind = op.get_bind()
+    inspector = inspect(bind)
+    return table_name in inspector.get_table_names()
+
+
+def column_exists(table_name: str, column_name: str) -> bool:
+    """检查列是否存在"""
+    bind = op.get_bind()
+    inspector = inspect(bind)
+    columns = [col['name'] for col in inspector.get_columns(table_name)]
+    return column_name in columns
+
+
+def upgrade() -> None:
+    """创建 stats_daily_model 表，重命名 provider_model_aliases 为 provider_model_mappings"""
+    # 1. 创建 stats_daily_model 表
+    if not table_exists('stats_daily_model'):
+        op.create_table(
+            'stats_daily_model',
+            sa.Column('id', sa.String(36), primary_key=True),
+            sa.Column('date', sa.DateTime(timezone=True), nullable=False),
+            sa.Column('model', sa.String(100), nullable=False),
+            sa.Column('total_requests', sa.Integer(), nullable=False, default=0),
+            sa.Column('input_tokens', sa.BigInteger(), nullable=False, default=0),
+            sa.Column('output_tokens', sa.BigInteger(), nullable=False, default=0),
+            sa.Column('cache_creation_tokens', sa.BigInteger(), nullable=False, default=0),
+            sa.Column('cache_read_tokens', sa.BigInteger(), nullable=False, default=0),
+            sa.Column('total_cost', sa.Float(), nullable=False, default=0.0),
+            sa.Column('avg_response_time_ms', sa.Float(), nullable=False, default=0.0),
+            sa.Column('created_at', sa.DateTime(timezone=True), nullable=False,
+                      server_default=sa.func.now()),
+            sa.Column('updated_at', sa.DateTime(timezone=True), nullable=False,
+                      server_default=sa.func.now(), onupdate=sa.func.now()),
+            sa.UniqueConstraint('date', 'model', name='uq_stats_daily_model'),
+        )
+
+        # 创建索引
+        op.create_index('idx_stats_daily_model_date', 'stats_daily_model', ['date'])
+        op.create_index('idx_stats_daily_model_date_model', 'stats_daily_model', ['date', 'model'])
+
+    # 2. 重命名 models 表的 provider_model_aliases 为 provider_model_mappings
+    if column_exists('models', 'provider_model_aliases') and not column_exists('models', 'provider_model_mappings'):
+        op.alter_column('models', 'provider_model_aliases', new_column_name='provider_model_mappings')
+
+
+def index_exists(table_name: str, index_name: str) -> bool:
+    """检查索引是否存在"""
+    bind = op.get_bind()
+    inspector = inspect(bind)
+    indexes = [idx['name'] for idx in inspector.get_indexes(table_name)]
+    return index_name in indexes
+
+
+def downgrade() -> None:
+    """删除 stats_daily_model 表，恢复 provider_model_aliases 列名"""
+    # 恢复列名
+    if column_exists('models', 'provider_model_mappings') and not column_exists('models', 'provider_model_aliases'):
+        op.alter_column('models', 'provider_model_mappings', new_column_name='provider_model_aliases')
+
+    # 删除表
+    if table_exists('stats_daily_model'):
+        if index_exists('stats_daily_model', 'idx_stats_daily_model_date_model'):
+            op.drop_index('idx_stats_daily_model_date_model', table_name='stats_daily_model')
+        if index_exists('stats_daily_model', 'idx_stats_daily_model_date'):
+            op.drop_index('idx_stats_daily_model_date', table_name='stats_daily_model')
+        op.drop_table('stats_daily_model')

alembic/versions/20251220_1500_add_usage_composite_indexes.py

@@ -0,0 +1,65 @@
+"""add usage table composite indexes for query optimization
+
+Revision ID: b2c3d4e5f6g7
+Revises: a1b2c3d4e5f6
+Create Date: 2025-12-20 15:00:00.000000+00:00
+
+"""
+from alembic import op
+from sqlalchemy import text
+
+# revision identifiers, used by Alembic.
+revision = 'b2c3d4e5f6g7'
+down_revision = 'a1b2c3d4e5f6'
+branch_labels = None
+depends_on = None
+
+
+def upgrade() -> None:
+    """为 usage 表添加复合索引以优化常见查询
+
+    注意：这些索引已经在 baseline 迁移中创建。
+    此迁移仅用于从旧版本升级的场景，新安装会跳过。
+    """
+    conn = op.get_bind()
+
+    # 检查 usage 表是否存在
+    result = conn.execute(text(
+        "SELECT EXISTS (SELECT FROM information_schema.tables WHERE table_name = 'usage')"
+    ))
+    if not result.scalar():
+        # 表不存在，跳过
+        return
+
+    # 定义需要创建的索引
+    indexes = [
+        ("idx_usage_user_created", "ON usage (user_id, created_at)"),
+        ("idx_usage_apikey_created", "ON usage (api_key_id, created_at)"),
+        ("idx_usage_provider_model_created", "ON usage (provider, model, created_at)"),
+    ]
+
+    # 分别检查并创建每个索引
+    for index_name, index_def in indexes:
+        result = conn.execute(text(
+            f"SELECT EXISTS (SELECT 1 FROM pg_indexes WHERE indexname = '{index_name}')"
+        ))
+        if result.scalar():
+            continue  # 索引已存在，跳过
+
+        conn.execute(text(f"CREATE INDEX {index_name} {index_def}"))
+
+
+def downgrade() -> None:
+    """删除复合索引"""
+    conn = op.get_bind()
+
+    # 使用 IF EXISTS 避免索引不存在时报错
+    conn.execute(text(
+        "DROP INDEX IF EXISTS idx_usage_provider_model_created"
+    ))
+    conn.execute(text(
+        "DROP INDEX IF EXISTS idx_usage_apikey_created"
+    ))
+    conn.execute(text(
+        "DROP INDEX IF EXISTS idx_usage_user_created"
+    ))

deploy.sh

@@ -26,10 +26,13 @@ calc_deps_hash() {
     cat pyproject.toml frontend/package.json frontend/package-lock.json Dockerfile.base.local 2>/dev/null | md5sum | cut -d' ' -f1
 }
 
-# 计算代码文件的哈希值
+# 计算代码文件的哈希值（包含 Dockerfile.app.local）
 calc_code_hash() {
-    find src -type f -name "*.py" 2>/dev/null | sort | xargs cat 2>/dev/null | md5sum | cut -d' ' -f1
-    find frontend/src -type f \( -name "*.vue" -o -name "*.ts" -o -name "*.tsx" -o -name "*.js" \) 2>/dev/null | sort | xargs cat 2>/dev/null | md5sum | cut -d' ' -f1
+    {
+        cat Dockerfile.app.local 2>/dev/null
+        find src -type f -name "*.py" 2>/dev/null | sort | xargs cat 2>/dev/null
+        find frontend/src -type f \( -name "*.vue" -o -name "*.ts" -o -name "*.tsx" -o -name "*.js" \) 2>/dev/null | sort | xargs cat 2>/dev/null
+    } | md5sum | cut -d' ' -f1
 }
 
 # 计算迁移文件的哈希值
@@ -88,7 +91,7 @@ build_base() {
 # 构建应用镜像
 build_app() {
     echo ">>> Building app image (code only)..."
-    docker build -f Dockerfile.app -t aether-app:latest .
+    docker build -f Dockerfile.app.local -t aether-app:latest .
     save_code_hash
 }
 
@@ -179,7 +182,13 @@ else
     echo ">>> Dependencies unchanged."
 fi
 
-# 检查代码是否变化，或者 base 重建了（app 依赖 base）
+# 检查代码或迁移是否变化，或者 base 重建了（app 依赖 base）
+# 注意：迁移文件打包在镜像中，所以迁移变化也需要重建 app 镜像
+MIGRATION_CHANGED=false
+if check_migration_changed; then
+    MIGRATION_CHANGED=true
+fi
+
 if ! docker image inspect aether-app:latest >/dev/null 2>&1; then
     echo ">>> App image not found, building..."
     build_app
@@ -192,6 +201,10 @@ elif check_code_changed; then
     echo ">>> Code changed, rebuilding app image..."
     build_app
     NEED_RESTART=true
+elif [ "$MIGRATION_CHANGED" = true ]; then
+    echo ">>> Migration files changed, rebuilding app image..."
+    build_app
+    NEED_RESTART=true
 else
     echo ">>> Code unchanged."
 fi
@@ -204,9 +217,9 @@ else
     echo ">>> No changes detected, skipping restart."
 fi
 
-# 检查迁移变化
-if check_migration_changed; then
-    echo ">>> Migration files changed, running database migration..."
+# 检查迁移变化（如果前面已经检测到变化并重建了镜像，这里直接运行迁移）
+if [ "$MIGRATION_CHANGED" = true ]; then
+    echo ">>> Running database migration..."
     sleep 3
     run_migration
 else

dev.sh

@@ -8,7 +8,8 @@ source .env
 set +a
 
 # 构建 DATABASE_URL
-export DATABASE_URL="postgresql://postgres:${DB_PASSWORD}@localhost:5432/aether"
+export DATABASE_URL="postgresql://${DB_USER:-postgres}:${DB_PASSWORD}@${DB_HOST:-localhost}:${DB_PORT:-5432}/${DB_NAME:-aether}"
+export REDIS_URL=redis://:${REDIS_PASSWORD}@${REDIS_HOST:-localhost}:${REDIS_PORT:-6379}/0
 
 # 启动 uvicorn（热重载模式）
 echo "🚀 启动本地开发服务器..."

docker-compose.build.yml

@@ -41,7 +41,7 @@ services:
   app:
     build:
       context: .
-      dockerfile: Dockerfile.app
+      dockerfile: Dockerfile.app.local
     image: aether-app:latest
     container_name: aether-app
     environment:

frontend/src/api/admin.ts

@@ -112,7 +112,7 @@ export interface KeyExport {
 export interface ModelExport {
   global_model_name: string | null
   provider_model_name: string
-  provider_model_aliases?: any
+  provider_model_mappings?: any
   price_per_request?: number | null
   tiered_pricing?: any
   supports_vision?: boolean | null

frontend/src/api/cache.ts

@@ -66,6 +66,7 @@ export interface UserAffinity {
   key_name: string | null
   key_prefix: string | null  // Provider Key 脱敏显示（前4...后4）
   rate_multiplier: number
+  global_model_id: string | null  // 原始的 global_model_id（用于删除）
   model_name: string | null  // 模型名称（如 claude-haiku-4-5-20250514）
   model_display_name: string | null  // 模型显示名称（如 Claude Haiku 4.5）
   api_format: string | null  // API 格式 (claude/openai)
@@ -119,6 +120,18 @@ export const cacheApi = {
     await api.delete(`/api/admin/monitoring/cache/users/${userIdentifier}`)
   },
 
+  /**
+   * 清除单条缓存亲和性
+   *
+   * @param affinityKey API Key ID
+   * @param endpointId Endpoint ID
+   * @param modelId GlobalModel ID
+   * @param apiFormat API 格式 (claude/openai)
+   */
+  async clearSingleAffinity(affinityKey: string, endpointId: string, modelId: string, apiFormat: string): Promise<void> {
+    await api.delete(`/api/admin/monitoring/cache/affinity/${affinityKey}/${endpointId}/${modelId}/${apiFormat}`)
+  },
+
   /**
    * 清除所有缓存
    */

frontend/src/api/endpoints/models.ts

@@ -5,6 +5,8 @@ import type {
   ModelUpdate,
   ModelCatalogResponse,
   ProviderAvailableSourceModelsResponse,
+  UpstreamModel,
+  ImportFromUpstreamResponse,
 } from './types'
 
 /**
@@ -119,3 +121,40 @@ export async function batchAssignModelsToProvider(
   )
   return response.data
 }
+
+/**
+ * 查询提供商的上游模型列表
+ */
+export async function queryProviderUpstreamModels(
+  providerId: string
+): Promise<{
+  success: boolean
+  data: {
+    models: UpstreamModel[]
+    error: string | null
+  }
+  provider: {
+    id: string
+    name: string
+    display_name: string
+  }
+}> {
+  const response = await client.post('/api/admin/provider-query/models', {
+    provider_id: providerId,
+  })
+  return response.data
+}
+
+/**
+ * 从上游提供商导入模型
+ */
+export async function importModelsFromUpstream(
+  providerId: string,
+  modelIds: string[]
+): Promise<ImportFromUpstreamResponse> {
+  const response = await client.post(
+    `/api/admin/providers/${providerId}/import-from-upstream`,
+    { model_ids: modelIds }
+  )
+  return response.data
+}

frontend/src/api/endpoints/providers.ts

@@ -58,3 +58,38 @@ export async function deleteProvider(providerId: string): Promise<{ message: str
   return response.data
 }
 
+/**
+ * 测试模型连接性
+ */
+export interface TestModelRequest {
+  provider_id: string
+  model_name: string
+  api_key_id?: string
+  message?: string
+  api_format?: string
+}
+
+export interface TestModelResponse {
+  success: boolean
+  error?: string
+  data?: {
+    response?: {
+      status_code?: number
+      error?: string | { message?: string }
+      choices?: Array<{ message?: { content?: string } }>
+    }
+    content_preview?: string
+  }
+  provider?: {
+    id: string
+    name: string
+    display_name: string
+  }
+  model?: string
+}
+
+export async function testModel(data: TestModelRequest): Promise<TestModelResponse> {
+  const response = await client.post('/api/admin/provider-query/test-model', data)
+  return response.data
+}
+

frontend/src/api/endpoints/types.ts

@@ -110,6 +110,24 @@ export interface EndpointAPIKey {
   request_results_window?: Array<{ ts: number; ok: boolean }>  // 请求结果滑动窗口
 }
 
+export interface EndpointAPIKeyUpdate {
+  name?: string
+  api_key?: string  // 仅在需要更新时提供
+  rate_multiplier?: number
+  internal_priority?: number
+  global_priority?: number | null
+  max_concurrent?: number | null  // null 表示切换为自适应模式
+  rate_limit?: number
+  daily_limit?: number
+  monthly_limit?: number
+  allowed_models?: string[] | null
+  capabilities?: Record<string, boolean> | null
+  cache_ttl_minutes?: number
+  max_probe_interval_minutes?: number
+  note?: string
+  is_active?: boolean
+}
+
 export interface EndpointHealthDetail {
   api_format: string
   health_score: number
@@ -244,18 +262,21 @@ export interface ConcurrencyStatus {
   key_max_concurrent?: number
 }
 
-export interface ProviderModelAlias {
+export interface ProviderModelMapping {
   name: string
   priority: number  // 优先级（数字越小优先级越高）
   api_formats?: string[]  // 作用域（适用的 API 格式），为空表示对所有格式生效
 }
 
+// 保留别名以保持向后兼容
+export type ProviderModelAlias = ProviderModelMapping
+
 export interface Model {
   id: string
   provider_id: string
   global_model_id?: string  // 关联的 GlobalModel ID
   provider_model_name: string  // Provider 侧的主模型名称
-  provider_model_aliases?: ProviderModelAlias[] | null  // 模型名称别名列表（带优先级）
+  provider_model_mappings?: ProviderModelMapping[] | null  // 模型名称映射列表（带优先级）
   // 原始配置值（可能为空，为空时使用 GlobalModel 默认值）
   price_per_request?: number | null  // 按次计费价格
   tiered_pricing?: TieredPricingConfig | null  // 阶梯计费配置
@@ -285,7 +306,7 @@ export interface Model {
 
 export interface ModelCreate {
   provider_model_name: string  // Provider 侧的主模型名称
-  provider_model_aliases?: ProviderModelAlias[]  // 模型名称别名列表（带优先级）
+  provider_model_mappings?: ProviderModelMapping[]  // 模型名称映射列表（带优先级）
   global_model_id: string  // 关联的 GlobalModel ID（必填）
   // 计费配置（可选，为空时使用 GlobalModel 默认值）
   price_per_request?: number  // 按次计费价格
@@ -302,7 +323,7 @@ export interface ModelCreate {
 
 export interface ModelUpdate {
   provider_model_name?: string
-  provider_model_aliases?: ProviderModelAlias[] | null  // 模型名称别名列表（带优先级）
+  provider_model_mappings?: ProviderModelMapping[] | null  // 模型名称映射列表（带优先级）
   global_model_id?: string
   price_per_request?: number | null  // 按次计费价格（null 表示清空/使用默认值）
   tiered_pricing?: TieredPricingConfig | null  // 阶梯计费配置
@@ -495,3 +516,42 @@ export interface GlobalModelListResponse {
   models: GlobalModelResponse[]
   total: number
 }
+
+// ==================== 上游模型导入相关 ====================
+
+/**
+ * 上游模型（从提供商 API 获取的原始模型）
+ */
+export interface UpstreamModel {
+  id: string
+  owned_by?: string
+  display_name?: string
+  api_format?: string
+}
+
+/**
+ * 导入成功的模型信息
+ */
+export interface ImportFromUpstreamSuccessItem {
+  model_id: string
+  global_model_id: string
+  global_model_name: string
+  provider_model_id: string
+  created_global_model: boolean
+}
+
+/**
+ * 导入失败的模型信息
+ */
+export interface ImportFromUpstreamErrorItem {
+  model_id: string
+  error: string
+}
+
+/**
+ * 从上游提供商导入模型响应
+ */
+export interface ImportFromUpstreamResponse {
+  success: ImportFromUpstreamSuccessItem[]
+  errors: ImportFromUpstreamErrorItem[]
+}

frontend/src/components/ui/dialog/Dialog.vue

@@ -92,6 +92,7 @@
 
 <script setup lang="ts">
 import { computed, useSlots, type Component } from 'vue'
+import { useEscapeKey } from '@/composables/useEscapeKey'
 
 // Props 定义
 const props = defineProps<{
@@ -157,4 +158,16 @@ const maxWidthClass = computed(() => {
 const containerZIndex = computed(() => props.zIndex || 60)
 const backdropZIndex = computed(() => props.zIndex || 60)
 const contentZIndex = computed(() => (props.zIndex || 60) + 10)
+
+// 添加 ESC 键监听
+useEscapeKey(() => {
+  if (isOpen.value) {
+    handleClose()
+    return true  // 阻止其他监听器（如父级抽屉的 ESC 监听器）
+  }
+  return false
+}, {
+  disableOnInput: true,
+  once: false
+})
 </script>

frontend/src/composables/useClipboard.ts

@@ -4,11 +4,11 @@ import { log } from '@/utils/logger'
 export function useClipboard() {
   const { success, error: showError } = useToast()
 
-  async function copyToClipboard(text: string): Promise<boolean> {
+  async function copyToClipboard(text: string, showToast = true): Promise<boolean> {
     try {
       if (navigator.clipboard && window.isSecureContext) {
         await navigator.clipboard.writeText(text)
-        success('已复制到剪贴板')
+        if (showToast) success('已复制到剪贴板')
         return true
       }
 
@@ -25,17 +25,17 @@ export function useClipboard() {
       try {
         const successful = document.execCommand('copy')
         if (successful) {
-          success('已复制到剪贴板')
+          if (showToast) success('已复制到剪贴板')
           return true
         }
-        showError('复制失败，请手动复制')
+        if (showToast) showError('复制失败，请手动复制')
         return false
       } finally {
         document.body.removeChild(textArea)
       }
     } catch (err) {
       log.error('复制失败:', err)
-      showError('复制失败，请手动选择文本进行复制')
+      if (showToast) showError('复制失败，请手动选择文本进行复制')
       return false
     }
   }

frontend/src/composables/useConfirm.ts

@@ -47,11 +47,11 @@ export function useConfirm() {
   /**
    * 便捷方法：危险操作确认（红色主题）
    */
-  const confirmDanger = (message: string, title?: string): Promise<boolean> => {
+  const confirmDanger = (message: string, title?: string, confirmText?: string): Promise<boolean> => {
     return confirm({
       message,
       title: title || '危险操作',
-      confirmText: '删除',
+      confirmText: confirmText || '删除',
       variant: 'danger'
     })
   }

frontend/src/composables/useEscapeKey.ts

@@ -0,0 +1,83 @@
+import { onMounted, onUnmounted, ref } from 'vue'
+
+/**
+ * ESC 键监听 Composable（简化版本，直接使用独立监听器）
+ * 用于按 ESC 键关闭弹窗或其他可关闭的组件
+ *
+ * @param callback - 按 ESC 键时执行的回调函数，返回 true 表示已处理事件，阻止其他监听器执行
+ * @param options - 配置选项
+ */
+export function useEscapeKey(
+  callback: () => void | boolean,
+  options: {
+    /** 是否在输入框获得焦点时禁用 ESC 键，默认 true */
+    disableOnInput?: boolean
+    /** 是否只监听一次，默认 false */
+    once?: boolean
+  } = {}
+) {
+  const { disableOnInput = true, once = false } = options
+  const isActive = ref(true)
+
+  function handleKeyDown(event: KeyboardEvent) {
+    // 只处理 ESC 键
+    if (event.key !== 'Escape') return
+
+    // 检查组件是否还活跃
+    if (!isActive.value) return
+
+    // 如果配置了在输入框获得焦点时禁用，则检查当前焦点元素
+    if (disableOnInput) {
+      const activeElement = document.activeElement
+      const isInputElement = activeElement && (
+        activeElement.tagName === 'INPUT' ||
+        activeElement.tagName === 'TEXTAREA' ||
+        activeElement.tagName === 'SELECT' ||
+        activeElement.contentEditable === 'true' ||
+        activeElement.getAttribute('role') === 'textbox' ||
+        activeElement.getAttribute('role') === 'combobox'
+      )
+
+      // 如果焦点在输入框中，不处理 ESC 键
+      if (isInputElement) return
+    }
+
+    // 执行回调，如果返回 true 则阻止其他监听器
+    const handled = callback()
+    if (handled === true) {
+      event.stopImmediatePropagation()
+    }
+
+    // 移除当前元素的焦点，避免残留样式
+    if (document.activeElement instanceof HTMLElement) {
+      document.activeElement.blur()
+    }
+
+    // 如果只监听一次，则移除监听器
+    if (once) {
+      removeEventListener()
+    }
+  }
+
+  function addEventListener() {
+    document.addEventListener('keydown', handleKeyDown)
+  }
+
+  function removeEventListener() {
+    document.removeEventListener('keydown', handleKeyDown)
+  }
+
+  onMounted(() => {
+    addEventListener()
+  })
+
+  onUnmounted(() => {
+    isActive.value = false
+    removeEventListener()
+  })
+
+  return {
+    addEventListener,
+    removeEventListener
+  }
+}

frontend/src/features/models/components/ModelDetailDrawer.vue

@@ -698,7 +698,9 @@ import {
   Layers,
   BarChart3
 } from 'lucide-vue-next'
+import { useEscapeKey } from '@/composables/useEscapeKey'
 import { useToast } from '@/composables/useToast'
+import { useClipboard } from '@/composables/useClipboard'
 import Card from '@/components/ui/card.vue'
 import Badge from '@/components/ui/badge.vue'
 import Button from '@/components/ui/button.vue'
@@ -730,6 +732,7 @@ const emit = defineEmits<{
   'refreshProviders': []
 }>()
 const { success: showSuccess, error: showError } = useToast()
+const { copyToClipboard } = useClipboard()
 
 interface Props {
   model: GlobalModelResponse | null
@@ -762,16 +765,6 @@ function handleClose() {
   }
 }
 
-// 复制到剪贴板
-async function copyToClipboard(text: string) {
-  try {
-    await navigator.clipboard.writeText(text)
-    showSuccess('已复制')
-  } catch {
-    showError('复制失败')
-  }
-}
-
 // 格式化日期
 function formatDate(dateStr: string): string {
   if (!dateStr) return '-'
@@ -833,6 +826,16 @@ watch(() => props.open, (newOpen) => {
     detailTab.value = 'basic'
   }
 })
+
+// 添加 ESC 键监听
+useEscapeKey(() => {
+  if (props.open) {
+    handleClose()
+  }
+}, {
+  disableOnInput: true,
+  once: false
+})
 </script>
 
 <style scoped>

frontend/src/features/providers/components/BatchAssignModelsDialog.vue

@@ -31,29 +31,46 @@
 
         <!-- 左右对比布局 -->
         <div class="flex gap-2 items-stretch">
-          <!-- 左侧：可添加的模型 -->
+          <!-- 左侧：可添加的模型（分组折叠） -->
           <div class="flex-1 space-y-2">
-            <div class="flex items-center justify-between">
-              <div class="flex items-center gap-2">
-                <p class="text-sm font-medium">
-                  可添加
-                </p>
-                <Button
-                  v-if="availableModels.length > 0"
-                  variant="ghost"
-                  size="sm"
-                  class="h-6 px-2 text-xs"
-                  @click="toggleSelectAllLeft"
-                >
-                  {{ isAllLeftSelected ? '取消全选' : '全选' }}
-                </Button>
+            <div class="flex items-center justify-between gap-2">
+              <p class="text-sm font-medium shrink-0">
+                可添加
+              </p>
+              <div class="flex-1 relative">
+                <Search class="absolute left-2 top-1/2 -translate-y-1/2 w-3.5 h-3.5 text-muted-foreground" />
+                <Input
+                  v-model="searchQuery"
+                  placeholder="搜索模型..."
+                  class="pl-7 h-7 text-xs"
+                />
               </div>
-              <Badge
-                variant="secondary"
-                class="text-xs"
+              <button
+                v-if="upstreamModelsLoaded"
+                type="button"
+                class="p-1.5 hover:bg-muted rounded-md transition-colors shrink-0"
+                title="刷新上游模型"
+                :disabled="fetchingUpstreamModels"
+                @click="fetchUpstreamModels(true)"
               >
-                {{ availableModels.length }} 个
-              </Badge>
+                <RefreshCw
+                  class="w-3.5 h-3.5"
+                  :class="{ 'animate-spin': fetchingUpstreamModels }"
+                />
+              </button>
+              <button
+                v-else-if="!fetchingUpstreamModels"
+                type="button"
+                class="p-1.5 hover:bg-muted rounded-md transition-colors shrink-0"
+                title="从提供商获取模型"
+                @click="fetchUpstreamModels"
+              >
+                <Zap class="w-3.5 h-3.5" />
+              </button>
+              <Loader2
+                v-else
+                class="w-3.5 h-3.5 animate-spin text-muted-foreground shrink-0"
+              />
             </div>
             <div class="border rounded-lg h-80 overflow-y-auto">
               <div
@@ -63,7 +80,7 @@
                 <Loader2 class="w-6 h-6 animate-spin text-primary" />
               </div>
               <div
-                v-else-if="availableModels.length === 0"
+                v-else-if="totalAvailableCount === 0 && !upstreamModelsLoaded"
                 class="flex flex-col items-center justify-center h-full text-muted-foreground"
               >
                 <Layers class="w-10 h-10 mb-2 opacity-30" />
@@ -73,37 +90,142 @@
               </div>
               <div
                 v-else
-                class="p-2 space-y-1"
+                class="p-2 space-y-2"
               >
+                <!-- 全局模型折叠组 -->
                 <div
-                  v-for="model in availableModels"
-                  :key="model.id"
-                  class="flex items-center gap-2 p-2 rounded-lg border transition-colors"
-                  :class="selectedLeftIds.includes(model.id)
-                    ? 'border-primary bg-primary/10'
-                    : 'hover:bg-muted/50 cursor-pointer'"
-                  @click="toggleLeftSelection(model.id)"
+                  v-if="availableGlobalModels.length > 0 || !upstreamModelsLoaded"
+                  class="border rounded-lg overflow-hidden"
                 >
-                  <Checkbox
-                    :checked="selectedLeftIds.includes(model.id)"
-                    @update:checked="toggleLeftSelection(model.id)"
-                    @click.stop
-                  />
-                  <div class="flex-1 min-w-0">
-                    <p class="font-medium text-sm truncate">
-                      {{ model.display_name }}
-                    </p>
-                    <p class="text-xs text-muted-foreground truncate font-mono">
-                      {{ model.name }}
-                    </p>
+                  <div class="flex items-center gap-2 px-3 py-2 bg-muted/30">
+                    <button
+                      type="button"
+                      class="flex items-center gap-2 flex-1 hover:bg-muted/50 -mx-1 px-1 rounded transition-colors"
+                      @click="toggleGroupCollapse('global')"
+                    >
+                      <ChevronDown
+                        class="w-4 h-4 transition-transform shrink-0"
+                        :class="collapsedGroups.has('global') ? '-rotate-90' : ''"
+                      />
+                      <span class="text-xs font-medium">
+                        全局模型
+                      </span>
+                      <span class="text-xs text-muted-foreground">
+                        ({{ availableGlobalModels.length }})
+                      </span>
+                    </button>
+                    <button
+                      v-if="availableGlobalModels.length > 0"
+                      type="button"
+                      class="text-xs text-primary hover:underline shrink-0"
+                      @click.stop="selectAllGlobalModels"
+                    >
+                      {{ isAllGlobalModelsSelected ? '取消' : '全选' }}
+                    </button>
                   </div>
-                  <Badge
-                    :variant="model.is_active ? 'outline' : 'secondary'"
-                    :class="model.is_active ? 'text-green-600 border-green-500/60' : ''"
-                    class="text-xs shrink-0"
+                  <div
+                    v-show="!collapsedGroups.has('global')"
+                    class="p-2 space-y-1 border-t"
                   >
-                    {{ model.is_active ? '活跃' : '停用' }}
-                  </Badge>
+                    <div
+                      v-if="availableGlobalModels.length === 0"
+                      class="py-4 text-center text-xs text-muted-foreground"
+                    >
+                      所有全局模型均已关联
+                    </div>
+                    <div
+                      v-for="model in availableGlobalModels"
+                      v-else
+                      :key="model.id"
+                      class="flex items-center gap-2 p-2 rounded-lg border transition-colors cursor-pointer"
+                      :class="selectedGlobalModelIds.includes(model.id)
+                        ? 'border-primary bg-primary/10'
+                        : 'hover:bg-muted/50'"
+                      @click="toggleGlobalModelSelection(model.id)"
+                    >
+                      <Checkbox
+                        :checked="selectedGlobalModelIds.includes(model.id)"
+                        @update:checked="toggleGlobalModelSelection(model.id)"
+                        @click.stop
+                      />
+                      <div class="flex-1 min-w-0">
+                        <p class="font-medium text-sm truncate">
+                          {{ model.display_name }}
+                        </p>
+                        <p class="text-xs text-muted-foreground truncate font-mono">
+                          {{ model.name }}
+                        </p>
+                      </div>
+                      <Badge
+                        :variant="model.is_active ? 'outline' : 'secondary'"
+                        :class="model.is_active ? 'text-green-600 border-green-500/60' : ''"
+                        class="text-xs shrink-0"
+                      >
+                        {{ model.is_active ? '活跃' : '停用' }}
+                      </Badge>
+                    </div>
+                  </div>
+                </div>
+
+                <!-- 从提供商获取的模型折叠组 -->
+                <div
+                  v-for="group in upstreamModelGroups"
+                  :key="group.api_format"
+                  class="border rounded-lg overflow-hidden"
+                >
+                  <div class="flex items-center gap-2 px-3 py-2 bg-muted/30">
+                    <button
+                      type="button"
+                      class="flex items-center gap-2 flex-1 hover:bg-muted/50 -mx-1 px-1 rounded transition-colors"
+                      @click="toggleGroupCollapse(group.api_format)"
+                    >
+                      <ChevronDown
+                        class="w-4 h-4 transition-transform shrink-0"
+                        :class="collapsedGroups.has(group.api_format) ? '-rotate-90' : ''"
+                      />
+                      <span class="text-xs font-medium">
+                        {{ API_FORMAT_LABELS[group.api_format] || group.api_format }}
+                      </span>
+                      <span class="text-xs text-muted-foreground">
+                        ({{ group.models.length }})
+                      </span>
+                    </button>
+                    <button
+                      type="button"
+                      class="text-xs text-primary hover:underline shrink-0"
+                      @click.stop="selectAllUpstreamModels(group.api_format)"
+                    >
+                      {{ isUpstreamGroupAllSelected(group.api_format) ? '取消' : '全选' }}
+                    </button>
+                  </div>
+                  <div
+                    v-show="!collapsedGroups.has(group.api_format)"
+                    class="p-2 space-y-1 border-t"
+                  >
+                    <div
+                      v-for="model in group.models"
+                      :key="model.id"
+                      class="flex items-center gap-2 p-2 rounded-lg border transition-colors cursor-pointer"
+                      :class="selectedUpstreamModelIds.includes(model.id)
+                        ? 'border-primary bg-primary/10'
+                        : 'hover:bg-muted/50'"
+                      @click="toggleUpstreamModelSelection(model.id)"
+                    >
+                      <Checkbox
+                        :checked="selectedUpstreamModelIds.includes(model.id)"
+                        @update:checked="toggleUpstreamModelSelection(model.id)"
+                        @click.stop
+                      />
+                      <div class="flex-1 min-w-0">
+                        <p class="font-medium text-sm truncate">
+                          {{ model.id }}
+                        </p>
+                        <p class="text-xs text-muted-foreground truncate font-mono">
+                          {{ model.owned_by || model.id }}
+                        </p>
+                      </div>
+                    </div>
+                  </div>
                 </div>
               </div>
             </div>
@@ -115,8 +237,8 @@
               variant="outline"
               size="sm"
               class="w-9 h-8"
-              :class="selectedLeftIds.length > 0 && !submittingAdd ? 'border-primary' : ''"
-              :disabled="selectedLeftIds.length === 0 || submittingAdd"
+              :class="totalSelectedCount > 0 && !submittingAdd ? 'border-primary' : ''"
+              :disabled="totalSelectedCount === 0 || submittingAdd"
               title="添加选中"
               @click="batchAddSelected"
             >
@@ -127,7 +249,7 @@
               <ChevronRight
                 v-else
                 class="w-6 h-6 stroke-[3]"
-                :class="selectedLeftIds.length > 0 && !submittingAdd ? 'text-primary' : ''"
+                :class="totalSelectedCount > 0 && !submittingAdd ? 'text-primary' : ''"
               />
             </Button>
             <Button
@@ -154,26 +276,18 @@
           <!-- 右侧：已添加的模型 -->
           <div class="flex-1 space-y-2">
             <div class="flex items-center justify-between">
-              <div class="flex items-center gap-2">
-                <p class="text-sm font-medium">
-                  已添加
-                </p>
-                <Button
-                  v-if="existingModels.length > 0"
-                  variant="ghost"
-                  size="sm"
-                  class="h-6 px-2 text-xs"
-                  @click="toggleSelectAllRight"
-                >
-                  {{ isAllRightSelected ? '取消全选' : '全选' }}
-                </Button>
-              </div>
-              <Badge
-                variant="secondary"
-                class="text-xs"
+              <p class="text-sm font-medium">
+                已添加
+              </p>
+              <Button
+                v-if="existingModels.length > 0"
+                variant="ghost"
+                size="sm"
+                class="h-6 px-2 text-xs"
+                @click="toggleSelectAllRight"
               >
-                {{ existingModels.length }} 个
-              </Badge>
+                {{ isAllRightSelected ? '取消' : '全选' }}
+              </Button>
             </div>
             <div class="border rounded-lg h-80 overflow-y-auto">
               <div
@@ -238,11 +352,12 @@
 
 <script setup lang="ts">
 import { ref, computed, watch } from 'vue'
-import { Layers, Loader2, ChevronRight, ChevronLeft } from 'lucide-vue-next'
+import { Layers, Loader2, ChevronRight, ChevronLeft, ChevronDown, Zap, RefreshCw, Search } from 'lucide-vue-next'
 import Dialog from '@/components/ui/dialog/Dialog.vue'
 import Button from '@/components/ui/button.vue'
 import Badge from '@/components/ui/badge.vue'
 import Checkbox from '@/components/ui/checkbox.vue'
+import Input from '@/components/ui/input.vue'
 import { useToast } from '@/composables/useToast'
 import { parseApiError } from '@/utils/errorParser'
 import {
@@ -253,8 +368,13 @@ import {
   getProviderModels,
   batchAssignModelsToProvider,
   deleteModel,
+  importModelsFromUpstream,
+  API_FORMAT_LABELS,
   type Model
 } from '@/api/endpoints'
+import { useUpstreamModelsCache, type UpstreamModel } from '../composables/useUpstreamModelsCache'
+
+const { fetchModels: fetchCachedModels, clearCache, getCachedModels } = useUpstreamModelsCache()
 
 const props = defineProps<{
   open: boolean
@@ -274,17 +394,27 @@ const { error: showError, success } = useToast()
 const loadingGlobalModels = ref(false)
 const submittingAdd = ref(false)
 const submittingRemove = ref(false)
+const fetchingUpstreamModels = ref(false)
+const upstreamModelsLoaded = ref(false)
 
 // 数据
 const allGlobalModels = ref<GlobalModelResponse[]>([])
 const existingModels = ref<Model[]>([])
+const upstreamModels = ref<UpstreamModel[]>([])
 
 // 选择状态
-const selectedLeftIds = ref<string[]>([])
+const selectedGlobalModelIds = ref<string[]>([])
+const selectedUpstreamModelIds = ref<string[]>([])
 const selectedRightIds = ref<string[]>([])
 
-// 计算可添加的模型（排除已关联的）
-const availableModels = computed(() => {
+// 折叠状态
+const collapsedGroups = ref<Set<string>>(new Set())
+
+// 搜索状态
+const searchQuery = ref('')
+
+// 计算可添加的全局模型（排除已关联的）
+const availableGlobalModelsBase = computed(() => {
   const existingGlobalModelIds = new Set(
     existingModels.value
       .filter(m => m.global_model_id)
@@ -293,31 +423,129 @@ const availableModels = computed(() => {
   return allGlobalModels.value.filter(m => !existingGlobalModelIds.has(m.id))
 })
 
-// 全选状态
-const isAllLeftSelected = computed(() =>
-  availableModels.value.length > 0 &&
-  selectedLeftIds.value.length === availableModels.value.length
-)
+// 搜索过滤后的全局模型
+const availableGlobalModels = computed(() => {
+  if (!searchQuery.value.trim()) return availableGlobalModelsBase.value
+  const query = searchQuery.value.toLowerCase()
+  return availableGlobalModelsBase.value.filter(m =>
+    m.name.toLowerCase().includes(query) ||
+    m.display_name.toLowerCase().includes(query)
+  )
+})
 
+// 计算可添加的上游模型（排除已关联的，包括主模型名和映射名称）
+const availableUpstreamModelsBase = computed(() => {
+  const existingModelNames = new Set<string>()
+  for (const m of existingModels.value) {
+    // 主模型名
+    existingModelNames.add(m.provider_model_name)
+    // 映射名称
+    for (const mapping of m.provider_model_mappings ?? []) {
+      if (mapping.name) existingModelNames.add(mapping.name)
+    }
+  }
+  return upstreamModels.value.filter(m => !existingModelNames.has(m.id))
+})
+
+// 搜索过滤后的上游模型
+const availableUpstreamModels = computed(() => {
+  if (!searchQuery.value.trim()) return availableUpstreamModelsBase.value
+  const query = searchQuery.value.toLowerCase()
+  return availableUpstreamModelsBase.value.filter(m =>
+    m.id.toLowerCase().includes(query) ||
+    (m.owned_by && m.owned_by.toLowerCase().includes(query))
+  )
+})
+
+// 按 API 格式分组的上游模型
+const upstreamModelGroups = computed(() => {
+  const groups: Record<string, UpstreamModel[]> = {}
+
+  for (const model of availableUpstreamModels.value) {
+    const format = model.api_format || 'unknown'
+    if (!groups[format]) {
+      groups[format] = []
+    }
+    groups[format].push(model)
+  }
+
+  // 按 API_FORMAT_LABELS 的顺序排序
+  const order = Object.keys(API_FORMAT_LABELS)
+  return Object.entries(groups)
+    .map(([api_format, models]) => ({ api_format, models }))
+    .sort((a, b) => {
+      const aIndex = order.indexOf(a.api_format)
+      const bIndex = order.indexOf(b.api_format)
+      if (aIndex === -1 && bIndex === -1) return a.api_format.localeCompare(b.api_format)
+      if (aIndex === -1) return 1
+      if (bIndex === -1) return -1
+      return aIndex - bIndex
+    })
+})
+
+// 总可添加数量
+const totalAvailableCount = computed(() => {
+  return availableGlobalModels.value.length + availableUpstreamModels.value.length
+})
+
+// 总选中数量
+const totalSelectedCount = computed(() => {
+  return selectedGlobalModelIds.value.length + selectedUpstreamModelIds.value.length
+})
+
+// 全选状态
 const isAllRightSelected = computed(() =>
   existingModels.value.length > 0 &&
   selectedRightIds.value.length === existingModels.value.length
 )
 
+// 全局模型是否全选
+const isAllGlobalModelsSelected = computed(() => {
+  if (availableGlobalModels.value.length === 0) return false
+  return availableGlobalModels.value.every(m => selectedGlobalModelIds.value.includes(m.id))
+})
+
+// 检查某个上游组是否全选
+function isUpstreamGroupAllSelected(apiFormat: string): boolean {
+  const group = upstreamModelGroups.value.find(g => g.api_format === apiFormat)
+  if (!group || group.models.length === 0) return false
+  return group.models.every(m => selectedUpstreamModelIds.value.includes(m.id))
+}
+
 // 监听打开状态
 watch(() => props.open, async (isOpen) => {
   if (isOpen && props.providerId) {
     await loadData()
   } else {
     // 重置状态
-    selectedLeftIds.value = []
+    selectedGlobalModelIds.value = []
+    selectedUpstreamModelIds.value = []
     selectedRightIds.value = []
+    upstreamModels.value = []
+    upstreamModelsLoaded.value = false
+    collapsedGroups.value = new Set()
+    searchQuery.value = ''
   }
 })
 
 // 加载数据
 async function loadData() {
   await Promise.all([loadGlobalModels(), loadExistingModels()])
+  // 默认折叠全局模型组
+  collapsedGroups.value = new Set(['global'])
+
+  // 检查缓存，如果有缓存数据则直接使用
+  const cachedModels = getCachedModels(props.providerId)
+  if (cachedModels) {
+    upstreamModels.value = cachedModels
+    upstreamModelsLoaded.value = true
+    // 折叠所有上游模型组
+    for (const model of cachedModels) {
+      if (model.api_format) {
+        collapsedGroups.value.add(model.api_format)
+      }
+    }
+  }
 }
 
 // 加载全局模型列表
@@ -342,13 +570,91 @@ async function loadExistingModels() {
   }
 }
 
-// 切换左侧选择
-function toggleLeftSelection(id: string) {
-  const index = selectedLeftIds.value.indexOf(id)
-  if (index === -1) {
-    selectedLeftIds.value.push(id)
+// 从提供商获取模型
+async function fetchUpstreamModels(forceRefresh = false) {
+  if (forceRefresh) {
+    clearCache(props.providerId)
+  }
+
+  try {
+    fetchingUpstreamModels.value = true
+    const result = await fetchCachedModels(props.providerId, forceRefresh)
+    if (result) {
+      if (result.error) {
+        showError(result.error, '错误')
+      } else {
+        upstreamModels.value = result.models
+        upstreamModelsLoaded.value = true
+        // 折叠所有上游模型组
+        const allGroups = new Set(collapsedGroups.value)
+        for (const model of result.models) {
+          if (model.api_format) {
+            allGroups.add(model.api_format)
+          }
+        }
+        collapsedGroups.value = allGroups
+      }
+    }
+  } finally {
+    fetchingUpstreamModels.value = false
+  }
+}
+
+// 切换折叠状态
+function toggleGroupCollapse(group: string) {
+  if (collapsedGroups.value.has(group)) {
+    collapsedGroups.value.delete(group)
   } else {
-    selectedLeftIds.value.splice(index, 1)
+    collapsedGroups.value.add(group)
+  }
+  // 触发响应式更新
+  collapsedGroups.value = new Set(collapsedGroups.value)
+}
+
+// 切换全局模型选择
+function toggleGlobalModelSelection(id: string) {
+  const index = selectedGlobalModelIds.value.indexOf(id)
+  if (index === -1) {
+    selectedGlobalModelIds.value.push(id)
+  } else {
+    selectedGlobalModelIds.value.splice(index, 1)
+  }
+}
+
+// 切换上游模型选择
+function toggleUpstreamModelSelection(id: string) {
+  const index = selectedUpstreamModelIds.value.indexOf(id)
+  if (index === -1) {
+    selectedUpstreamModelIds.value.push(id)
+  } else {
+    selectedUpstreamModelIds.value.splice(index, 1)
+  }
+}
+
+// 全选全局模型
+function selectAllGlobalModels() {
+  const allIds = availableGlobalModels.value.map(m => m.id)
+  const allSelected = allIds.every(id => selectedGlobalModelIds.value.includes(id))
+  if (allSelected) {
+    selectedGlobalModelIds.value = selectedGlobalModelIds.value.filter(id => !allIds.includes(id))
+  } else {
+    const newIds = allIds.filter(id => !selectedGlobalModelIds.value.includes(id))
+    selectedGlobalModelIds.value.push(...newIds)
+  }
+}
+
+// 全选某个 API 格式的上游模型
+function selectAllUpstreamModels(apiFormat: string) {
+  const group = upstreamModelGroups.value.find(g => g.api_format === apiFormat)
+  if (!group) return
+
+  const allIds = group.models.map(m => m.id)
+  const allSelected = allIds.every(id => selectedUpstreamModelIds.value.includes(id))
+  if (allSelected) {
+    selectedUpstreamModelIds.value = selectedUpstreamModelIds.value.filter(id => !allIds.includes(id))
+  } else {
+    const newIds = allIds.filter(id => !selectedUpstreamModelIds.value.includes(id))
+    selectedUpstreamModelIds.value.push(...newIds)
   }
 }
 
@@ -362,15 +668,6 @@ function toggleRightSelection(id: string) {
   }
 }
 
-// 全选/取消全选左侧
-function toggleSelectAllLeft() {
-  if (isAllLeftSelected.value) {
-    selectedLeftIds.value = []
-  } else {
-    selectedLeftIds.value = availableModels.value.map(m => m.id)
-  }
-}
-
 // 全选/取消全选右侧
 function toggleSelectAllRight() {
   if (isAllRightSelected.value) {
@@ -382,22 +679,41 @@ function toggleSelectAllRight() {
 
 // 批量添加选中的模型
 async function batchAddSelected() {
-  if (selectedLeftIds.value.length === 0) return
+  if (totalSelectedCount.value === 0) return
 
   try {
     submittingAdd.value = true
-    const result = await batchAssignModelsToProvider(props.providerId, selectedLeftIds.value)
+    let totalSuccess = 0
+    const allErrors: string[] = []
 
-    if (result.success.length > 0) {
-      success(`成功添加 ${result.success.length} 个模型`)
+    // 处理全局模型
+    if (selectedGlobalModelIds.value.length > 0) {
+      const result = await batchAssignModelsToProvider(props.providerId, selectedGlobalModelIds.value)
+      totalSuccess += result.success.length
+      if (result.errors.length > 0) {
+        allErrors.push(...result.errors.map(e => e.error))
+      }
     }
 
-    if (result.errors.length > 0) {
-      const errorMessages = result.errors.map(e => e.error).join(', ')
-      showError(`部分模型添加失败: ${errorMessages}`, '警告')
+    // 处理上游模型（调用 import-from-upstream API）
+    if (selectedUpstreamModelIds.value.length > 0) {
+      const result = await importModelsFromUpstream(props.providerId, selectedUpstreamModelIds.value)
+      totalSuccess += result.success.length
+      if (result.errors.length > 0) {
+        allErrors.push(...result.errors.map(e => e.error))
+      }
     }
 
-    selectedLeftIds.value = []
+    if (totalSuccess > 0) {
+      success(`成功添加 ${totalSuccess} 个模型`)
+    }
+
+    if (allErrors.length > 0) {
+      showError(`部分模型添加失败: ${allErrors.slice(0, 3).join(', ')}${allErrors.length > 3 ? '...' : ''}`, '警告')
+    }
+
+    selectedGlobalModelIds.value = []
+    selectedUpstreamModelIds.value = []
     await loadExistingModels()
     emit('changed')
   } catch (err: any) {

frontend/src/features/providers/components/KeyFormDialog.vue

@@ -260,6 +260,7 @@ import {
   updateEndpointKey,
   getAllCapabilities,
   type EndpointAPIKey,
+  type EndpointAPIKeyUpdate,
   type ProviderEndpoint,
   type CapabilityDefinition
 } from '@/api/endpoints'
@@ -386,10 +387,11 @@ function loadKeyData() {
     api_key: '',
     rate_multiplier: props.editingKey.rate_multiplier || 1.0,
     internal_priority: props.editingKey.internal_priority ?? 50,
-    max_concurrent: props.editingKey.max_concurrent || undefined,
-    rate_limit: props.editingKey.rate_limit || undefined,
-    daily_limit: props.editingKey.daily_limit || undefined,
-    monthly_limit: props.editingKey.monthly_limit || undefined,
+    // 保留原始的 null/undefined 状态，null 表示自适应模式
+    max_concurrent: props.editingKey.max_concurrent ?? undefined,
+    rate_limit: props.editingKey.rate_limit ?? undefined,
+    daily_limit: props.editingKey.daily_limit ?? undefined,
+    monthly_limit: props.editingKey.monthly_limit ?? undefined,
     cache_ttl_minutes: props.editingKey.cache_ttl_minutes ?? 5,
     max_probe_interval_minutes: props.editingKey.max_probe_interval_minutes ?? 32,
     note: props.editingKey.note || '',
@@ -439,12 +441,17 @@ async function handleSave() {
   saving.value = true
   try {
     if (props.editingKey) {
-      // 更新
-      const updateData: any = {
+      // 更新模式
+      // 注意：max_concurrent 需要显式发送 null 来切换到自适应模式
+      // undefined 会在 JSON 中被忽略，所以用 null 表示"清空/自适应"
+      const updateData: EndpointAPIKeyUpdate = {
         name: form.value.name,
         rate_multiplier: form.value.rate_multiplier,
         internal_priority: form.value.internal_priority,
-        max_concurrent: form.value.max_concurrent,
+        // 显式使用 null 表示自适应模式，这样后端能区分"未提供"和"设置为 null"
+        // 注意：只有 max_concurrent 需要这种处理，因为它有"自适应模式"的概念
+        // 其他限制字段（rate_limit 等）不支持"清空"操作，undefined 会被 JSON 忽略即不更新
+        max_concurrent: form.value.max_concurrent === undefined ? null : form.value.max_concurrent,
         rate_limit: form.value.rate_limit,
         daily_limit: form.value.daily_limit,
         monthly_limit: form.value.monthly_limit,

frontend/src/features/providers/components/ModelAliasDialog.vue

@@ -18,7 +18,7 @@
         </p>
       </div>
 
-      <!-- 别名列表 -->
+      <!-- 映射列表 -->
       <div class="space-y-3">
         <div class="flex items-center justify-between">
           <Label class="text-sm font-medium">名称映射</Label>
@@ -92,7 +92,7 @@
               </div>
             </div>
 
-            <!-- 别名输入框 -->
+            <!-- 映射输入框 -->
             <Input
               v-model="alias.name"
               placeholder="映射名称，如 Claude-Sonnet-4.5"
@@ -184,9 +184,9 @@ const editingPriorityIndex = ref<number | null>(null)
 // 监听 open 变化
 watch(() => props.open, (newOpen) => {
   if (newOpen && props.model) {
-    // 加载现有别名配置
-    if (props.model.provider_model_aliases && Array.isArray(props.model.provider_model_aliases)) {
-      aliases.value = JSON.parse(JSON.stringify(props.model.provider_model_aliases))
+    // 加载现有映射配置
+    if (props.model.provider_model_mappings && Array.isArray(props.model.provider_model_mappings)) {
+      aliases.value = JSON.parse(JSON.stringify(props.model.provider_model_mappings))
     } else {
       aliases.value = []
     }
@@ -197,16 +197,16 @@ watch(() => props.open, (newOpen) => {
   }
 })
 
-// 添加别名
+// 添加映射
 function addAlias() {
-  // 新别名优先级为当前最大优先级 + 1，或者默认为 1
+  // 新映射优先级为当前最大优先级 + 1，或者默认为 1
   const maxPriority = aliases.value.length > 0
     ? Math.max(...aliases.value.map(a => a.priority))
     : 0
   aliases.value.push({ name: '', priority: maxPriority + 1 })
 }
 
-// 移除别名
+// 移除映射
 function removeAlias(index: number) {
   aliases.value.splice(index, 1)
 }
@@ -244,7 +244,7 @@ function handleDrop(targetIndex: number) {
   const items = [...aliases.value]
   const draggedItem = items[dragIndex]
 
-  // 记录每个别名的原始优先级（在修改前）
+  // 记录每个映射的原始优先级（在修改前）
   const originalPriorityMap = new Map<number, number>()
   items.forEach((alias, idx) => {
     originalPriorityMap.set(idx, alias.priority)
@@ -255,7 +255,7 @@ function handleDrop(targetIndex: number) {
   items.splice(targetIndex, 0, draggedItem)
 
   // 按新顺序为每个组分配新的优先级
-  // 同组的别名保持相同的优先级（被拖动的别名单独成组）
+  // 同组的映射保持相同的优先级（被拖动的映射单独成组）
   const groupNewPriority = new Map<number, number>() // 原优先级 -> 新优先级
   let currentPriority = 1
 
@@ -263,12 +263,12 @@ function handleDrop(targetIndex: number) {
   const draggedOriginalPriority = originalPriorityMap.get(dragIndex)!
 
   items.forEach((alias, newIdx) => {
-    // 找到这个别名在原数组中的索引
+    // 找到这个映射在原数组中的索引
     const originalIdx = aliases.value.findIndex(a => a === alias)
     const originalPriority = originalIdx >= 0 ? originalPriorityMap.get(originalIdx)! : alias.priority
 
     if (alias === draggedItem) {
-      // 被拖动的别名是独立的新组，获得当前优先级
+      // 被拖动的映射是独立的新组，获得当前优先级
       alias.priority = currentPriority
       currentPriority++
     } else {
@@ -318,11 +318,11 @@ async function handleSubmit() {
 
   submitting.value = true
   try {
-    // 过滤掉空的别名
+    // 过滤掉空的映射
     const validAliases = aliases.value.filter(a => a.name.trim())
 
     await updateModel(props.providerId, props.model.id, {
-      provider_model_aliases: validAliases.length > 0 ? validAliases : null
+      provider_model_mappings: validAliases.length > 0 ? validAliases : null
     })
 
     showSuccess('映射配置已保存')

frontend/src/features/providers/components/ModelMappingDialog.vue

@@ -0,0 +1,796 @@
+<template>
+  <Dialog
+    :model-value="open"
+    :title="editingGroup ? '编辑模型映射' : '添加模型映射'"
+    :description="editingGroup ? '修改映射配置' : '为模型添加新的名称映射'"
+    :icon="Tag"
+    size="4xl"
+    @update:model-value="$emit('update:open', $event)"
+  >
+    <div class="space-y-4">
+      <!-- 第一行：目标模型 | 作用域 -->
+      <div class="flex gap-4">
+        <!-- 目标模型 -->
+        <div class="flex-1 space-y-1.5">
+          <Label class="text-xs">目标模型</Label>
+          <Select
+            v-model:open="modelSelectOpen"
+            :model-value="formData.modelId"
+            :disabled="!!editingGroup"
+            @update:model-value="handleModelChange"
+          >
+            <SelectTrigger class="h-9">
+              <SelectValue placeholder="请选择模型" />
+            </SelectTrigger>
+            <SelectContent>
+              <SelectItem
+                v-for="model in models"
+                :key="model.id"
+                :value="model.id"
+              >
+                {{ model.global_model_display_name || model.provider_model_name }}
+              </SelectItem>
+            </SelectContent>
+          </Select>
+        </div>
+
+        <!-- 作用域 -->
+        <div class="flex-1 space-y-1.5">
+          <Label class="text-xs">作用域 <span class="text-muted-foreground font-normal">(不选则适用全部)</span></Label>
+          <div
+            v-if="providerApiFormats.length > 0"
+            class="flex flex-wrap gap-1.5 p-2 rounded-md border bg-muted/30 min-h-[36px]"
+          >
+            <button
+              v-for="format in providerApiFormats"
+              :key="format"
+              type="button"
+              class="px-2.5 py-0.5 rounded text-xs font-medium transition-colors"
+              :class="[
+                formData.apiFormats.includes(format)
+                  ? 'bg-primary text-primary-foreground'
+                  : 'bg-background border border-border hover:bg-muted'
+              ]"
+              @click="toggleApiFormat(format)"
+            >
+              {{ API_FORMAT_LABELS[format] || format }}
+            </button>
+          </div>
+          <div
+            v-else
+            class="h-9 flex items-center text-xs text-muted-foreground"
+          >
+            无可用格式
+          </div>
+        </div>
+      </div>
+
+      <!-- 第二行：两栏布局 -->
+      <div class="flex gap-4 items-stretch">
+        <!-- 左侧：上游模型列表 -->
+        <div class="flex-1 space-y-2">
+          <div class="flex items-center justify-between gap-2">
+            <span class="text-sm font-medium shrink-0">
+              上游模型
+            </span>
+            <div class="flex-1 relative">
+              <Search class="absolute left-2 top-1/2 -translate-y-1/2 w-3.5 h-3.5 text-muted-foreground" />
+              <Input
+                v-model="upstreamModelSearch"
+                placeholder="搜索模型..."
+                class="pl-7 h-7 text-xs"
+              />
+            </div>
+            <button
+              v-if="upstreamModelsLoaded"
+              type="button"
+              class="p-1.5 hover:bg-muted rounded-md transition-colors shrink-0"
+              title="刷新列表"
+              :disabled="refreshingUpstreamModels"
+              @click="refreshUpstreamModels"
+            >
+              <RefreshCw
+                class="w-3.5 h-3.5"
+                :class="{ 'animate-spin': refreshingUpstreamModels }"
+              />
+            </button>
+            <button
+              v-else-if="!fetchingUpstreamModels"
+              type="button"
+              class="p-1.5 hover:bg-muted rounded-md transition-colors shrink-0"
+              title="获取上游模型列表"
+              @click="fetchUpstreamModels"
+            >
+              <Zap class="w-3.5 h-3.5" />
+            </button>
+            <Loader2
+              v-else
+              class="w-3.5 h-3.5 animate-spin text-muted-foreground shrink-0"
+            />
+          </div>
+          <div class="border rounded-lg h-80 overflow-y-auto">
+            <template v-if="upstreamModelsLoaded">
+              <div
+                v-if="groupedAvailableUpstreamModels.length === 0"
+                class="flex flex-col items-center justify-center h-full text-muted-foreground"
+              >
+                <Zap class="w-10 h-10 mb-2 opacity-30" />
+                <p class="text-sm">
+                  {{ upstreamModelSearch ? '没有匹配的模型' : '所有模型已添加' }}
+                </p>
+              </div>
+              <div
+                v-else
+                class="p-2 space-y-2"
+              >
+                <!-- 按分组显示（可折叠） -->
+                <div
+                  v-for="group in groupedAvailableUpstreamModels"
+                  :key="group.api_format"
+                  class="border rounded-lg overflow-hidden"
+                >
+                  <div class="flex items-center gap-2 px-3 py-2 bg-muted/30">
+                    <button
+                      type="button"
+                      class="flex items-center gap-2 flex-1 hover:bg-muted/50 -mx-1 px-1 rounded transition-colors"
+                      @click="toggleGroupCollapse(group.api_format)"
+                    >
+                      <ChevronDown
+                        class="w-4 h-4 transition-transform shrink-0"
+                        :class="collapsedGroups.has(group.api_format) ? '-rotate-90' : ''"
+                      />
+                      <span class="text-xs font-medium">
+                        {{ API_FORMAT_LABELS[group.api_format] || group.api_format }}
+                      </span>
+                      <span class="text-xs text-muted-foreground">
+                        ({{ group.models.length }})
+                      </span>
+                    </button>
+                  </div>
+                  <div
+                    v-show="!collapsedGroups.has(group.api_format)"
+                    class="p-2 space-y-1 border-t"
+                  >
+                    <div
+                      v-for="model in group.models"
+                      :key="model.id"
+                      class="flex items-center gap-2 p-2 rounded-lg border transition-colors hover:bg-muted/30"
+                      :title="model.id"
+                    >
+                      <div class="flex-1 min-w-0">
+                        <p class="font-medium text-sm truncate">
+                          {{ model.id }}
+                        </p>
+                        <p class="text-xs text-muted-foreground truncate font-mono">
+                          {{ model.owned_by || model.id }}
+                        </p>
+                      </div>
+                      <button
+                        type="button"
+                        class="p-1 hover:bg-primary/10 rounded transition-colors shrink-0"
+                        title="添加到映射"
+                        @click="addUpstreamModel(model.id)"
+                      >
+                        <ChevronRight class="w-4 h-4 text-muted-foreground hover:text-primary" />
+                      </button>
+                    </div>
+                  </div>
+                </div>
+              </div>
+            </template>
+
+            <!-- 未加载状态 -->
+            <div
+              v-else
+              class="flex flex-col items-center justify-center h-full text-muted-foreground"
+            >
+              <Zap class="w-10 h-10 mb-2 opacity-30" />
+              <p class="text-sm">
+                点击右上角按钮
+              </p>
+              <p class="text-xs mt-1">
+                从上游获取可用模型
+              </p>
+            </div>
+          </div>
+        </div>
+
+        <!-- 右侧：映射名称列表 -->
+        <div class="flex-1 space-y-2">
+          <div class="flex items-center justify-between">
+            <p class="text-sm font-medium">
+              映射名称
+            </p>
+            <button
+              type="button"
+              class="p-1.5 hover:bg-muted rounded-md transition-colors"
+              title="手动添加"
+              @click="addAliasItem"
+            >
+              <Plus class="w-3.5 h-3.5" />
+            </button>
+          </div>
+          <div class="border rounded-lg h-80 overflow-y-auto">
+            <div
+              v-if="formData.aliases.length === 0"
+              class="flex flex-col items-center justify-center h-full text-muted-foreground"
+            >
+              <Tag class="w-10 h-10 mb-2 opacity-30" />
+              <p class="text-sm">
+                从左侧选择模型
+              </p>
+              <p class="text-xs mt-1">
+                或点击上方"手动添加"
+              </p>
+            </div>
+            <div
+              v-else
+              class="p-2 space-y-1"
+            >
+              <div
+                v-for="(alias, index) in formData.aliases"
+                :key="`alias-${index}`"
+                class="group flex items-center gap-2 p-2 rounded-lg border transition-colors hover:bg-muted/30"
+                :class="[
+                  draggedIndex === index ? 'bg-primary/5' : '',
+                  dragOverIndex === index ? 'bg-primary/10 border-primary' : ''
+                ]"
+                draggable="true"
+                @dragstart="handleDragStart(index, $event)"
+                @dragend="handleDragEnd"
+                @dragover.prevent="handleDragOver(index)"
+                @dragleave="handleDragLeave"
+                @drop="handleDrop(index)"
+              >
+                <!-- 删除按钮 -->
+                <button
+                  type="button"
+                  class="p-1 hover:bg-destructive/10 rounded transition-colors shrink-0"
+                  title="移除"
+                  @click="removeAliasItem(index)"
+                >
+                  <ChevronLeft class="w-4 h-4 text-muted-foreground hover:text-destructive" />
+                </button>
+
+                <!-- 优先级 -->
+                <div class="shrink-0">
+                  <input
+                    v-if="editingPriorityIndex === index"
+                    type="number"
+                    min="1"
+                    :value="alias.priority"
+                    class="w-7 h-6 rounded bg-background border border-primary text-xs text-center focus:outline-none [appearance:textfield] [&::-webkit-outer-spin-button]:appearance-none [&::-webkit-inner-spin-button]:appearance-none"
+                    autofocus
+                    @blur="finishEditPriority(index, $event)"
+                    @keydown.enter="($event.target as HTMLInputElement).blur()"
+                    @keydown.escape="cancelEditPriority"
+                  >
+                  <div
+                    v-else
+                    class="w-6 h-6 rounded bg-muted/50 flex items-center justify-center text-xs text-muted-foreground cursor-pointer hover:bg-primary/10 hover:text-primary"
+                    title="点击编辑优先级"
+                    @click.stop="startEditPriority(index)"
+                  >
+                    {{ alias.priority }}
+                  </div>
+                </div>
+
+                <!-- 名称显示/编辑 -->
+                <div class="flex-1 min-w-0">
+                  <Input
+                    v-if="alias.isEditing"
+                    v-model="alias.name"
+                    placeholder="输入映射名称"
+                    class="h-7 text-xs"
+                    autofocus
+                    @blur="alias.isEditing = false"
+                    @keydown.enter="alias.isEditing = false"
+                  />
+                  <p
+                    v-else
+                    class="font-medium text-sm truncate cursor-pointer hover:text-primary"
+                    title="点击编辑"
+                    @click="alias.isEditing = true"
+                  >
+                    {{ alias.name || '点击输入名称' }}
+                  </p>
+                </div>
+
+                <!-- 拖拽手柄 -->
+                <div class="cursor-grab active:cursor-grabbing text-muted-foreground/30 group-hover:text-muted-foreground shrink-0">
+                  <GripVertical class="w-4 h-4" />
+                </div>
+              </div>
+            </div>
+            <!-- 拖拽提示 -->
+            <div
+              v-if="formData.aliases.length > 1"
+              class="px-3 py-1.5 bg-muted/30 border-t text-xs text-muted-foreground text-center"
+            >
+              拖拽调整优先级顺序
+            </div>
+          </div>
+        </div>
+      </div>
+    </div>
+
+    <template #footer>
+      <Button
+        variant="outline"
+        @click="$emit('update:open', false)"
+      >
+        取消
+      </Button>
+      <Button
+        :disabled="submitting || !formData.modelId || formData.aliases.length === 0 || !hasValidAliases"
+        @click="handleSubmit"
+      >
+        <Loader2
+          v-if="submitting"
+          class="w-4 h-4 mr-2 animate-spin"
+        />
+        {{ editingGroup ? '保存' : '添加' }}
+      </Button>
+    </template>
+  </Dialog>
+</template>
+
+<script setup lang="ts">
+import { ref, computed, watch } from 'vue'
+import { Tag, Loader2, GripVertical, Zap, Search, RefreshCw, ChevronDown, ChevronRight, ChevronLeft, Plus } from 'lucide-vue-next'
+import {
+  Button,
+  Input,
+  Label,
+  Dialog,
+  Select,
+  SelectContent,
+  SelectItem,
+  SelectTrigger,
+  SelectValue,
+} from '@/components/ui'
+import { useToast } from '@/composables/useToast'
+import {
+  API_FORMAT_LABELS,
+  type Model,
+  type ProviderModelAlias
+} from '@/api/endpoints'
+import { updateModel } from '@/api/endpoints/models'
+import { useUpstreamModelsCache, type UpstreamModel } from '../composables/useUpstreamModelsCache'
+
+interface FormAlias {
+  name: string
+  priority: number
+  isEditing?: boolean
+}
+
+export interface AliasGroup {
+  model: Model
+  apiFormatsKey: string
+  apiFormats: string[]
+  aliases: ProviderModelAlias[]
+}
+
+const props = defineProps<{
+  open: boolean
+  providerId: string
+  providerApiFormats: string[]
+  models: Model[]
+  editingGroup?: AliasGroup | null
+}>()
+
+const emit = defineEmits<{
+  'update:open': [value: boolean]
+  'saved': []
+}>()
+
+const { error: showError, success: showSuccess } = useToast()
+const { fetchModels: fetchCachedModels, clearCache, getCachedModels } = useUpstreamModelsCache()
+
+// 状态
+const submitting = ref(false)
+const modelSelectOpen = ref(false)
+
+// 拖拽状态
+const draggedIndex = ref<number | null>(null)
+const dragOverIndex = ref<number | null>(null)
+
+// 优先级编辑状态
+const editingPriorityIndex = ref<number | null>(null)
+
+// 快速添加（上游模型）状态
+const fetchingUpstreamModels = ref(false)
+const refreshingUpstreamModels = ref(false)
+const upstreamModelsLoaded = ref(false)
+const upstreamModels = ref<UpstreamModel[]>([])
+const upstreamModelSearch = ref('')
+
+// 分组折叠状态
+const collapsedGroups = ref<Set<string>>(new Set())
+
+// 表单数据
+const formData = ref<{
+  modelId: string
+  apiFormats: string[]
+  aliases: FormAlias[]
+}>({
+  modelId: '',
+  apiFormats: [],
+  aliases: []
+})
+
+// 检查是否有有效的映射
+const hasValidAliases = computed(() => {
+  return formData.value.aliases.some(a => a.name.trim())
+})
+
+// 过滤和排序后的上游模型列表
+const filteredUpstreamModels = computed(() => {
+  const searchText = upstreamModelSearch.value.toLowerCase().trim()
+  let result = [...upstreamModels.value]
+
+  result.sort((a, b) => a.id.localeCompare(b.id))
+
+  if (searchText) {
+    const keywords = searchText.split(/\s+/).filter(k => k.length > 0)
+    result = result.filter(m => {
+      const searchableText = `${m.id} ${m.owned_by || ''} ${m.api_format || ''}`.toLowerCase()
+      return keywords.every(keyword => searchableText.includes(keyword))
+    })
+  }
+
+  return result
+})
+
+// 按 API 格式分组的上游模型列表
+interface UpstreamModelGroup {
+  api_format: string
+  models: Array<{ id: string; owned_by?: string; api_format?: string }>
+}
+
+const groupedAvailableUpstreamModels = computed<UpstreamModelGroup[]>(() => {
+  // 收集当前表单已添加的名称
+  const addedNames = new Set(formData.value.aliases.map(a => a.name.trim()))
+
+  // 收集所有已存在的映射名称（包括主模型名和映射名称）
+  for (const m of props.models) {
+    addedNames.add(m.provider_model_name)
+    for (const mapping of m.provider_model_mappings ?? []) {
+      if (mapping.name) addedNames.add(mapping.name)
+    }
+  }
+
+  const availableModels = filteredUpstreamModels.value.filter(m => !addedNames.has(m.id))
+
+  const groups = new Map<string, UpstreamModelGroup>()
+
+  for (const model of availableModels) {
+    const format = model.api_format || 'UNKNOWN'
+    if (!groups.has(format)) {
+      groups.set(format, { api_format: format, models: [] })
+    }
+    groups.get(format)!.models.push(model)
+  }
+
+  const order = Object.keys(API_FORMAT_LABELS)
+  return Array.from(groups.values()).sort((a, b) => {
+    const aIndex = order.indexOf(a.api_format)
+    const bIndex = order.indexOf(b.api_format)
+    if (aIndex === -1 && bIndex === -1) return a.api_format.localeCompare(b.api_format)
+    if (aIndex === -1) return 1
+    if (bIndex === -1) return -1
+    return aIndex - bIndex
+  })
+})
+
+// 监听打开状态
+watch(() => props.open, (isOpen) => {
+  if (isOpen) {
+    initForm()
+  }
+})
+
+// 初始化表单
+function initForm() {
+  if (props.editingGroup) {
+    formData.value = {
+      modelId: props.editingGroup.model.id,
+      apiFormats: [...props.editingGroup.apiFormats],
+      aliases: props.editingGroup.aliases.map(a => ({ name: a.name, priority: a.priority }))
+    }
+  } else {
+    formData.value = {
+      modelId: '',
+      apiFormats: [],
+      aliases: []
+    }
+  }
+  // 重置状态
+  editingPriorityIndex.value = null
+  draggedIndex.value = null
+  dragOverIndex.value = null
+  upstreamModelSearch.value = ''
+  collapsedGroups.value = new Set()
+
+  // 检查缓存，如果有缓存数据则直接使用
+  const cachedModels = getCachedModels(props.providerId)
+  if (cachedModels) {
+    upstreamModels.value = cachedModels
+    upstreamModelsLoaded.value = true
+    // 默认折叠所有分组
+    for (const model of cachedModels) {
+      if (model.api_format) {
+        collapsedGroups.value.add(model.api_format)
+      }
+    }
+  } else {
+    upstreamModelsLoaded.value = false
+    upstreamModels.value = []
+  }
+}
+
+// 处理模型选择变更
+function handleModelChange(value: string) {
+  formData.value.modelId = value
+  const selectedModel = props.models.find(m => m.id === value)
+  if (selectedModel) {
+    upstreamModelSearch.value = selectedModel.provider_model_name
+  }
+}
+
+// 切换 API 格式
+function toggleApiFormat(format: string) {
+  const index = formData.value.apiFormats.indexOf(format)
+  if (index >= 0) {
+    formData.value.apiFormats.splice(index, 1)
+  } else {
+    formData.value.apiFormats.push(format)
+  }
+}
+
+// 切换分组折叠状态
+function toggleGroupCollapse(apiFormat: string) {
+  if (collapsedGroups.value.has(apiFormat)) {
+    collapsedGroups.value.delete(apiFormat)
+  } else {
+    collapsedGroups.value.add(apiFormat)
+  }
+}
+
+// 添加映射项
+function addAliasItem() {
+  const maxPriority = formData.value.aliases.length > 0
+    ? Math.max(...formData.value.aliases.map(a => a.priority))
+    : 0
+  formData.value.aliases.push({ name: '', priority: maxPriority + 1, isEditing: true })
+}
+
+// 删除映射项
+function removeAliasItem(index: number) {
+  formData.value.aliases.splice(index, 1)
+}
+
+// ===== 拖拽排序 =====
+function handleDragStart(index: number, event: DragEvent) {
+  draggedIndex.value = index
+  if (event.dataTransfer) {
+    event.dataTransfer.effectAllowed = 'move'
+  }
+}
+
+function handleDragEnd() {
+  draggedIndex.value = null
+  dragOverIndex.value = null
+}
+
+function handleDragOver(index: number) {
+  if (draggedIndex.value !== null && draggedIndex.value !== index) {
+    dragOverIndex.value = index
+  }
+}
+
+function handleDragLeave() {
+  dragOverIndex.value = null
+}
+
+function handleDrop(targetIndex: number) {
+  const dragIndex = draggedIndex.value
+  if (dragIndex === null || dragIndex === targetIndex) {
+    dragOverIndex.value = null
+    return
+  }
+
+  const items = [...formData.value.aliases]
+  const draggedItem = items[dragIndex]
+
+  const originalPriorityMap = new Map<number, number>()
+  items.forEach((alias, idx) => {
+    originalPriorityMap.set(idx, alias.priority)
+  })
+
+  items.splice(dragIndex, 1)
+  items.splice(targetIndex, 0, draggedItem)
+
+  const groupNewPriority = new Map<number, number>()
+  let currentPriority = 1
+
+  items.forEach((alias) => {
+    const originalIdx = formData.value.aliases.findIndex(a => a === alias)
+    const originalPriority = originalIdx >= 0 ? originalPriorityMap.get(originalIdx)! : alias.priority
+
+    if (alias === draggedItem) {
+      alias.priority = currentPriority
+      currentPriority++
+    } else {
+      if (groupNewPriority.has(originalPriority)) {
+        alias.priority = groupNewPriority.get(originalPriority)!
+      } else {
+        groupNewPriority.set(originalPriority, currentPriority)
+        alias.priority = currentPriority
+        currentPriority++
+      }
+    }
+  })
+
+  formData.value.aliases = items
+  draggedIndex.value = null
+  dragOverIndex.value = null
+}
+
+// ===== 优先级编辑 =====
+function startEditPriority(index: number) {
+  editingPriorityIndex.value = index
+}
+
+function finishEditPriority(index: number, event: FocusEvent) {
+  const input = event.target as HTMLInputElement
+  const newPriority = parseInt(input.value) || 1
+  formData.value.aliases[index].priority = Math.max(1, newPriority)
+  editingPriorityIndex.value = null
+}
+
+function cancelEditPriority() {
+  editingPriorityIndex.value = null
+}
+
+// ===== 快速添加（上游模型）=====
+async function fetchUpstreamModels() {
+  if (!props.providerId) return
+
+  upstreamModelSearch.value = ''
+  fetchingUpstreamModels.value = true
+
+  try {
+    const result = await fetchCachedModels(props.providerId)
+    if (result) {
+      if (result.error) {
+        showError(result.error, '错误')
+      } else {
+        upstreamModels.value = result.models
+        upstreamModelsLoaded.value = true
+        // 默认折叠所有分组
+        for (const model of result.models) {
+          if (model.api_format) {
+            collapsedGroups.value.add(model.api_format)
+          }
+        }
+      }
+    }
+  } finally {
+    fetchingUpstreamModels.value = false
+  }
+}
+
+function addUpstreamModel(modelId: string) {
+  if (formData.value.aliases.some(a => a.name === modelId)) {
+    return
+  }
+
+  const maxPriority = formData.value.aliases.length > 0
+    ? Math.max(...formData.value.aliases.map(a => a.priority))
+    : 0
+
+  formData.value.aliases.push({ name: modelId, priority: maxPriority + 1 })
+}
+
+async function refreshUpstreamModels() {
+  if (!props.providerId || refreshingUpstreamModels.value) return
+
+  refreshingUpstreamModels.value = true
+  clearCache(props.providerId)
+
+  try {
+    const result = await fetchCachedModels(props.providerId, true)
+    if (result) {
+      if (result.error) {
+        showError(result.error, '错误')
+      } else {
+        upstreamModels.value = result.models
+      }
+    }
+  } finally {
+    refreshingUpstreamModels.value = false
+  }
+}
+
+// 生成作用域唯一键
+function getApiFormatsKey(formats: string[] | undefined): string {
+  if (!formats || formats.length === 0) return ''
+  return [...formats].sort().join(',')
+}
+
+// 提交表单
+async function handleSubmit() {
+  if (submitting.value) return
+  if (!formData.value.modelId || formData.value.aliases.length === 0) return
+
+  const validAliases = formData.value.aliases.filter(a => a.name.trim())
+  if (validAliases.length === 0) {
+    showError('请至少添加一个有效的映射名称', '错误')
+    return
+  }
+
+  submitting.value = true
+  try {
+    const targetModel = props.models.find(m => m.id === formData.value.modelId)
+    if (!targetModel) {
+      showError('模型不存在', '错误')
+      return
+    }
+
+    const currentAliases = targetModel.provider_model_mappings || []
+    let newAliases: ProviderModelAlias[]
+
+    const buildAlias = (a: FormAlias): ProviderModelAlias => ({
+      name: a.name.trim(),
+      priority: a.priority,
+      ...(formData.value.apiFormats.length > 0 ? { api_formats: formData.value.apiFormats } : {})
+    })
+
+    if (props.editingGroup) {
+      const oldApiFormatsKey = props.editingGroup.apiFormatsKey
+      const oldAliasNames = new Set(props.editingGroup.aliases.map(a => a.name))
+
+      const filteredAliases = currentAliases.filter((a: ProviderModelAlias) => {
+        const currentKey = getApiFormatsKey(a.api_formats)
+        return !(currentKey === oldApiFormatsKey && oldAliasNames.has(a.name))
+      })
+
+      const existingNames = new Set(filteredAliases.map((a: ProviderModelAlias) => a.name))
+      const duplicates = validAliases.filter(a => existingNames.has(a.name.trim()))
+      if (duplicates.length > 0) {
+        showError(`以下映射名称已存在：${duplicates.map(d => d.name).join(', ')}`, '错误')
+        return
+      }
+
+      newAliases = [
+        ...filteredAliases,
+        ...validAliases.map(buildAlias)
+      ]
+    } else {
+      const existingNames = new Set(currentAliases.map((a: ProviderModelAlias) => a.name))
+      const duplicates = validAliases.filter(a => existingNames.has(a.name.trim()))
+      if (duplicates.length > 0) {
+        showError(`以下映射名称已存在：${duplicates.map(d => d.name).join(', ')}`, '错误')
+        return
+      }
+      newAliases = [
+        ...currentAliases,
+        ...validAliases.map(buildAlias)
+      ]
+    }
+
+    await updateModel(props.providerId, targetModel.id, {
+      provider_model_mappings: newAliases
+    })
+
+    showSuccess(props.editingGroup ? '映射组已更新' : '映射已添加')
+    emit('update:open', false)
+    emit('saved')
+  } catch (err: any) {
+    showError(err.response?.data?.detail || '操作失败', '错误')
+  } finally {
+    submitting.value = false
+  }
+}
+</script>

frontend/src/features/providers/components/ProviderDetailDrawer.vue

@@ -483,9 +483,9 @@
                                   <span
                                     v-if="key.max_concurrent || key.is_adaptive"
                                     class="text-muted-foreground"
-                                    :title="key.is_adaptive ? `自适应并发限制（学习值: ${key.learned_max_concurrent ?? '未学习'}）` : '固定并发限制'"
+                                    :title="key.is_adaptive ? `自适应并发限制（学习值: ${key.learned_max_concurrent ?? '未学习'}）` : `固定并发限制: ${key.max_concurrent}`"
                                   >
-                                    {{ key.is_adaptive ? '自适应' : '固定' }}并发: {{ key.learned_max_concurrent || key.max_concurrent || 3 }}
+                                    {{ key.is_adaptive ? '自适应' : '固定' }}并发: {{ key.is_adaptive ? (key.learned_max_concurrent ?? '学习中') : key.max_concurrent }}
                                   </span>
                                 </div>
                               </div>
@@ -531,6 +531,7 @@
               <!-- 模型名称映射 -->
               <ModelAliasesTab
                 v-if="provider"
+                ref="modelAliasesTabRef"
                 :key="`aliases-${provider.id}`"
                 :provider="provider"
                 @refresh="handleRelatedDataRefresh"
@@ -655,10 +656,12 @@ import {
   GripVertical,
   Copy
 } from 'lucide-vue-next'
+import { useEscapeKey } from '@/composables/useEscapeKey'
 import Button from '@/components/ui/button.vue'
 import Badge from '@/components/ui/badge.vue'
 import Card from '@/components/ui/card.vue'
 import { useToast } from '@/composables/useToast'
+import { useClipboard } from '@/composables/useClipboard'
 import { getProvider, getProviderEndpoints } from '@/api/endpoints'
 import {
   KeyFormDialog,
@@ -704,6 +707,7 @@ const emit = defineEmits<{
 }>()
 
 const { error: showError, success: showSuccess } = useToast()
+const { copyToClipboard } = useClipboard()
 
 const loading = ref(false)
 const provider = ref<any>(null)
@@ -734,6 +738,9 @@ const deleteModelConfirmOpen = ref(false)
 const modelToDelete = ref<Model | null>(null)
 const batchAssignDialogOpen = ref(false)
 
+// ModelAliasesTab 组件引用
+const modelAliasesTabRef = ref<InstanceType<typeof ModelAliasesTab> | null>(null)
+
 // 拖动排序相关状态
 const dragState = ref({
   isDragging: false,
@@ -755,7 +762,9 @@ const hasBlockingDialogOpen = computed(() =>
   deleteKeyConfirmOpen.value ||
   modelFormDialogOpen.value ||
   deleteModelConfirmOpen.value ||
-  batchAssignDialogOpen.value
+  batchAssignDialogOpen.value ||
+  // 检测 ModelAliasesTab 子组件的 Dialog 是否打开
+  modelAliasesTabRef.value?.dialogOpen
 )
 
 // 监听 providerId 变化
@@ -1243,16 +1252,6 @@ function getHealthScoreBarColor(score: number): string {
   return 'bg-red-500 dark:bg-red-400'
 }
 
-// 复制到剪贴板
-async function copyToClipboard(text: string) {
-  try {
-    await navigator.clipboard.writeText(text)
-    showSuccess('已复制到剪贴板')
-  } catch {
-    showError('复制失败', '错误')
-  }
-}
-
 // 加载 Provider 信息
 async function loadProvider() {
   if (!props.providerId) return
@@ -1296,6 +1295,16 @@ async function loadEndpoints() {
     showError(err.response?.data?.detail || '加载端点失败', '错误')
   }
 }
+
+// 添加 ESC 键监听
+useEscapeKey(() => {
+  if (props.open) {
+    handleClose()
+  }
+}, {
+  disableOnInput: true,
+  once: false
+})
 </script>
 
 <style scoped>

frontend/src/features/providers/components/provider-tabs/ModelsTab.vue

@@ -156,6 +156,17 @@
             </td>
             <td class="align-top px-4 py-3">
               <div class="flex justify-center gap-1.5">
+                <Button
+                  variant="ghost"
+                  size="icon"
+                  class="h-8 w-8"
+                  title="测试模型"
+                  :disabled="testingModelId === model.id"
+                  @click="testModelConnection(model)"
+                >
+                  <Loader2 v-if="testingModelId === model.id" class="w-3.5 h-3.5 animate-spin" />
+                  <Play v-else class="w-3.5 h-3.5" />
+                </Button>
                 <Button
                   variant="ghost"
                   size="icon"
@@ -209,12 +220,14 @@
 
 <script setup lang="ts">
 import { ref, computed, onMounted } from 'vue'
-import { Box, Edit, Trash2, Layers, Eye, Wrench, Zap, Brain, Power, Copy, Image } from 'lucide-vue-next'
+import { Box, Edit, Trash2, Layers, Eye, Wrench, Zap, Brain, Power, Copy, Image, Loader2, Play } from 'lucide-vue-next'
 import Card from '@/components/ui/card.vue'
 import Button from '@/components/ui/button.vue'
 import { useToast } from '@/composables/useToast'
-import { getProviderModels, type Model } from '@/api/endpoints'
+import { useClipboard } from '@/composables/useClipboard'
+import { getProviderModels, testModel, type Model } from '@/api/endpoints'
 import { updateModel } from '@/api/endpoints/models'
+import { parseTestModelError } from '@/utils/errorParser'
 
 const props = defineProps<{
   provider: any
@@ -227,11 +240,13 @@ const emit = defineEmits<{
 }>()
 
 const { error: showError, success: showSuccess } = useToast()
+const { copyToClipboard } = useClipboard()
 
 // 状态
 const loading = ref(false)
 const models = ref<Model[]>([])
 const togglingModelId = ref<string | null>(null)
+const testingModelId = ref<string | null>(null)
 
 // 按名称排序的模型列表
 const sortedModels = computed(() => {
@@ -244,12 +259,7 @@ const sortedModels = computed(() => {
 
 // 复制模型 ID 到剪贴板
 async function copyModelId(modelId: string) {
-  try {
-    await navigator.clipboard.writeText(modelId)
-    showSuccess('已复制到剪贴板')
-  } catch {
-    showError('复制失败', '错误')
-  }
+  await copyToClipboard(modelId)
 }
 
 // 加载模型
@@ -380,6 +390,39 @@ async function toggleModelActive(model: Model) {
   }
 }
 
+// 测试模型连接性
+async function testModelConnection(model: Model) {
+  if (testingModelId.value) return
+
+  testingModelId.value = model.id
+  try {
+    const result = await testModel({
+      provider_id: props.provider.id,
+      model_name: model.provider_model_name,
+      message: "hello"
+    })
+
+    if (result.success) {
+      showSuccess(`模型 "${model.provider_model_name}" 测试成功`)
+
+      // 如果有响应内容，可以显示更多信息
+      if (result.data?.response?.choices?.[0]?.message?.content) {
+        const content = result.data.response.choices[0].message.content
+        showSuccess(`测试成功，响应: ${content.substring(0, 100)}${content.length > 100 ? '...' : ''}`)
+      } else if (result.data?.content_preview) {
+        showSuccess(`流式测试成功，预览: ${result.data.content_preview}`)
+      }
+    } else {
+      showError(`模型测试失败: ${parseTestModelError(result)}`)
+    }
+  } catch (err: any) {
+    const errorMsg = err.response?.data?.detail || err.message || '测试请求失败'
+    showError(`模型测试失败: ${errorMsg}`)
+  } finally {
+    testingModelId.value = null
+  }
+}
+
 onMounted(() => {
   loadModels()
 })

frontend/src/features/providers/composables/useUpstreamModelsCache.ts

@@ -0,0 +1,112 @@
+/**
+ * 上游模型缓存 - 共享缓存，避免重复请求
+ */
+import { ref } from 'vue'
+import { adminApi } from '@/api/admin'
+import type { UpstreamModel } from '@/api/endpoints/types'
+
+// 扩展类型，包含可能的额外字段
+export type { UpstreamModel }
+
+interface CacheEntry {
+  models: UpstreamModel[]
+  timestamp: number
+}
+
+type FetchResult = { models: UpstreamModel[]; error?: string }
+
+// 全局缓存（模块级别，所有组件共享）
+const cache = new Map<string, CacheEntry>()
+const CACHE_TTL = 5 * 60 * 1000 // 5分钟
+
+// 进行中的请求（用于去重并发请求）
+const pendingRequests = new Map<string, Promise<FetchResult>>()
+
+// 请求状态
+const loadingMap = ref<Map<string, boolean>>(new Map())
+
+export function useUpstreamModelsCache() {
+  /**
+   * 获取上游模型列表
+   * @param providerId 提供商ID
+   * @param forceRefresh 是否强制刷新
+   * @returns 模型列表或 null（如果请求失败）
+   */
+  async function fetchModels(
+    providerId: string,
+    forceRefresh = false
+  ): Promise<FetchResult> {
+    // 检查缓存
+    if (!forceRefresh) {
+      const cached = cache.get(providerId)
+      if (cached && Date.now() - cached.timestamp < CACHE_TTL) {
+        return { models: cached.models }
+      }
+    }
+
+    // 检查是否有进行中的请求（非强制刷新时复用）
+    if (!forceRefresh && pendingRequests.has(providerId)) {
+      return pendingRequests.get(providerId)!
+    }
+
+    // 创建新请求
+    const requestPromise = (async (): Promise<FetchResult> => {
+      try {
+        loadingMap.value.set(providerId, true)
+        const response = await adminApi.queryProviderModels(providerId)
+
+        if (response.success && response.data?.models) {
+          // 存入缓存
+          cache.set(providerId, {
+            models: response.data.models,
+            timestamp: Date.now()
+          })
+          return { models: response.data.models }
+        } else {
+          return { models: [], error: response.data?.error || '获取上游模型失败' }
+        }
+      } catch (err: any) {
+        return { models: [], error: err.response?.data?.detail || '获取上游模型失败' }
+      } finally {
+        loadingMap.value.set(providerId, false)
+        pendingRequests.delete(providerId)
+      }
+    })()
+
+    pendingRequests.set(providerId, requestPromise)
+    return requestPromise
+  }
+
+  /**
+   * 获取缓存的模型（不发起请求）
+   */
+  function getCachedModels(providerId: string): UpstreamModel[] | null {
+    const cached = cache.get(providerId)
+    if (cached && Date.now() - cached.timestamp < CACHE_TTL) {
+      return cached.models
+    }
+    return null
+  }
+
+  /**
+   * 清除指定提供商的缓存
+   */
+  function clearCache(providerId: string) {
+    cache.delete(providerId)
+  }
+
+  /**
+   * 检查是否正在加载
+   */
+  function isLoading(providerId: string): boolean {
+    return loadingMap.value.get(providerId) || false
+  }
+
+  return {
+    fetchModels,
+    getCachedModels,
+    clearCache,
+    isLoading,
+    loadingMap
+  }
+}

frontend/src/features/usage/components/RequestDetailDrawer.vue

@@ -472,6 +472,8 @@
 <script setup lang="ts">
 import { ref, watch, computed } from 'vue'
 import Button from '@/components/ui/button.vue'
+import { useEscapeKey } from '@/composables/useEscapeKey'
+import { useClipboard } from '@/composables/useClipboard'
 import Card from '@/components/ui/card.vue'
 import Badge from '@/components/ui/badge.vue'
 import Separator from '@/components/ui/separator.vue'
@@ -504,6 +506,7 @@ const copiedStates = ref<Record<string, boolean>>({})
 const viewMode = ref<'compare' | 'formatted' | 'raw'>('compare')
 const currentExpandDepth = ref(1)
 const dataSource = ref<'client' | 'provider'>('client')
+const { copyToClipboard } = useClipboard()
 const historicalPricing = ref<{
   input_price: string
   output_price: string
@@ -783,7 +786,7 @@ function copyJsonToClipboard(tabName: string) {
   }
 
   if (data) {
-    navigator.clipboard.writeText(JSON.stringify(data, null, 2))
+    copyToClipboard(JSON.stringify(data, null, 2), false)
     copiedStates.value[tabName] = true
     setTimeout(() => {
       copiedStates.value[tabName] = false
@@ -897,6 +900,16 @@ const providerHeadersWithDiff = computed(() => {
 
   return result
 })
+
+// 添加 ESC 键监听
+useEscapeKey(() => {
+  if (props.isOpen) {
+    handleClose()
+  }
+}, {
+  disableOnInput: true,
+  once: false
+})
 </script>
 
 <style scoped>

frontend/src/features/usage/components/UsageRecordsTable.vue

@@ -136,11 +136,20 @@
       <!-- 分隔线 -->
       <div class="hidden sm:block h-4 w-px bg-border" />
 
-      <!-- 刷新按钮 -->
-      <RefreshButton
-        :loading="loading"
-        @click="$emit('refresh')"
-      />
+      <!-- 自动刷新按钮 -->
+      <Button
+        variant="ghost"
+        size="icon"
+        class="h-8 w-8"
+        :class="autoRefresh ? 'text-primary' : ''"
+        :title="autoRefresh ? '点击关闭自动刷新' : '点击开启自动刷新（每10秒刷新）'"
+        @click="$emit('update:autoRefresh', !autoRefresh)"
+      >
+        <RefreshCcw
+          class="w-3.5 h-3.5"
+          :class="autoRefresh ? 'animate-spin' : ''"
+        />
+      </Button>
     </template>
 
     <Table>
@@ -357,14 +366,34 @@
             </div>
           </TableCell>
           <TableCell class="text-right py-4 w-[70px]">
+            <!-- pending 状态：只显示增长的总时间 -->
             <div
-              v-if="record.status === 'pending' || record.status === 'streaming'"
+              v-if="record.status === 'pending'"
               class="flex flex-col items-end text-xs gap-0.5"
             >
+              <span class="text-muted-foreground">-</span>
               <span class="text-primary tabular-nums">
                 {{ getElapsedTime(record) }}
               </span>
             </div>
+            <!-- streaming 状态：首字固定 + 总时间增长 -->
+            <div
+              v-else-if="record.status === 'streaming'"
+              class="flex flex-col items-end text-xs gap-0.5"
+            >
+              <span
+                v-if="record.first_byte_time_ms != null"
+                class="tabular-nums"
+              >{{ (record.first_byte_time_ms / 1000).toFixed(2) }}s</span>
+              <span
+                v-else
+                class="text-muted-foreground"
+              >-</span>
+              <span class="text-primary tabular-nums">
+                {{ getElapsedTime(record) }}
+              </span>
+            </div>
+            <!-- 已完成状态：首字 + 总耗时 -->
             <div
               v-else-if="record.response_time_ms != null"
               class="flex flex-col items-end text-xs gap-0.5"
@@ -408,6 +437,7 @@ import { ref, computed, onUnmounted, watch } from 'vue'
 import {
   TableCard,
   Badge,
+  Button,
   Select,
   SelectTrigger,
   SelectValue,
@@ -420,8 +450,8 @@ import {
   TableHead,
   TableCell,
   Pagination,
-  RefreshButton,
 } from '@/components/ui'
+import { RefreshCcw } from 'lucide-vue-next'
 import { formatTokens, formatCurrency } from '@/utils/format'
 import { formatDateTime } from '../composables'
 import { useRowClick } from '@/composables/useRowClick'
@@ -453,6 +483,8 @@ const props = defineProps<{
   pageSize: number
   totalRecords: number
   pageSizeOptions: number[]
+  // 自动刷新
+  autoRefresh: boolean
 }>()
 
 const emit = defineEmits<{
@@ -463,6 +495,7 @@ const emit = defineEmits<{
   'update:filterStatus': [value: string]
   'update:currentPage': [value: number]
   'update:pageSize': [value: number]
+  'update:autoRefresh': [value: boolean]
   'refresh': []
   'showDetail': [id: string]
 }>()

frontend/src/features/users/components/UserFormDialog.vue

@@ -86,6 +86,34 @@
             </p>
           </div>
 
+          <div
+            v-if="isEditMode && form.password.length > 0"
+            class="space-y-2"
+          >
+            <Label class="text-sm font-medium">
+              确认新密码 <span class="text-muted-foreground">*</span>
+            </Label>
+            <Input
+              :id="`pwd-confirm-${formNonce}`"
+              v-model="form.confirmPassword"
+              type="password"
+              autocomplete="new-password"
+              data-form-type="other"
+              data-lpignore="true"
+              :name="`confirm-${formNonce}`"
+              required
+              minlength="6"
+              placeholder="再次输入新密码"
+              class="h-10"
+            />
+            <p
+              v-if="form.confirmPassword.length > 0 && form.password !== form.confirmPassword"
+              class="text-xs text-destructive"
+            >
+              两次输入的密码不一致
+            </p>
+          </div>
+
           <div class="space-y-2">
             <Label
               for="form-email"
@@ -423,6 +451,7 @@ const apiFormats = ref<Array<{ value: string; label: string }>>([])
 const form = ref({
   username: '',
   password: '',
+  confirmPassword: '',
   email: '',
   quota: 10,
   role: 'user' as 'admin' | 'user',
@@ -443,6 +472,7 @@ function resetForm() {
   form.value = {
     username: '',
     password: '',
+    confirmPassword: '',
     email: '',
     quota: 10,
     role: 'user',
@@ -461,6 +491,7 @@ function loadUserData() {
   form.value = {
     username: props.user.username,
     password: '',
+    confirmPassword: '',
     email: props.user.email || '',
     quota: props.user.quota_usd == null ? 10 : props.user.quota_usd,
     role: props.user.role,
@@ -486,7 +517,9 @@ const isFormValid = computed(() => {
   const hasUsername = form.value.username.trim().length > 0
   const hasEmail = form.value.email.trim().length > 0
   const hasPassword = isEditMode.value || form.value.password.length >= 6
-  return hasUsername && hasEmail && hasPassword
+  // 编辑模式下如果填写了密码，必须确认密码一致
+  const passwordConfirmed = !isEditMode.value || form.value.password.length === 0 || form.value.password === form.value.confirmPassword
+  return hasUsername && hasEmail && hasPassword && passwordConfirmed
 })
 
 // 加载访问控制选项

frontend/src/mocks/handler.ts

@@ -403,7 +403,7 @@ function getUsageRecords() {
   return cachedUsageRecords
 }
 
-// Mock 别名数据
+// Mock 映射数据
 const MOCK_ALIASES = [
   { id: 'alias-001', source_model: 'claude-4-sonnet', target_global_model_id: 'gm-001', target_global_model_name: 'claude-sonnet-4-20250514', target_global_model_display_name: 'Claude Sonnet 4', provider_id: null, provider_name: null, scope: 'global', mapping_type: 'alias', is_active: true, created_at: '2024-01-01T00:00:00Z', updated_at: '2024-01-01T00:00:00Z' },
   { id: 'alias-002', source_model: 'claude-4-opus', target_global_model_id: 'gm-002', target_global_model_name: 'claude-opus-4-20250514', target_global_model_display_name: 'Claude Opus 4', provider_id: null, provider_name: null, scope: 'global', mapping_type: 'alias', is_active: true, created_at: '2024-01-01T00:00:00Z', updated_at: '2024-01-01T00:00:00Z' },
@@ -1682,7 +1682,7 @@ registerDynamicRoute('GET', '/api/admin/models/mappings/:mappingId', async (_con
   requireAdmin()
   const alias = MOCK_ALIASES.find(a => a.id === params.mappingId)
   if (!alias) {
-    throw { response: createMockResponse({ detail: '别名不存在' }, 404) }
+    throw { response: createMockResponse({ detail: '映射不存在' }, 404) }
   }
   return createMockResponse(alias)
 })
@@ -1693,7 +1693,7 @@ registerDynamicRoute('PATCH', '/api/admin/models/mappings/:mappingId', async (co
   requireAdmin()
   const alias = MOCK_ALIASES.find(a => a.id === params.mappingId)
   if (!alias) {
-    throw { response: createMockResponse({ detail: '别名不存在' }, 404) }
+    throw { response: createMockResponse({ detail: '映射不存在' }, 404) }
   }
   const body = JSON.parse(config.data || '{}')
   return createMockResponse({ ...alias, ...body, updated_at: new Date().toISOString() })
@@ -1705,7 +1705,7 @@ registerDynamicRoute('DELETE', '/api/admin/models/mappings/:mappingId', async (_
   requireAdmin()
   const alias = MOCK_ALIASES.find(a => a.id === params.mappingId)
   if (!alias) {
-    throw { response: createMockResponse({ detail: '别名不存在' }, 404) }
+    throw { response: createMockResponse({ detail: '映射不存在' }, 404) }
   }
   return createMockResponse({ message: '删除成功（演示模式）' })
 })

frontend/src/stores/users.ts

@@ -14,7 +14,7 @@ export const useUsersStore = defineStore('users', () => {
     try {
       users.value = await usersApi.getAllUsers()
     } catch (err: any) {
-      error.value = err.response?.data?.detail || '获取用户列表失败'
+      error.value = err.response?.data?.error?.message || err.response?.data?.detail || '获取用户列表失败'
     } finally {
       loading.value = false
     }
@@ -29,7 +29,7 @@ export const useUsersStore = defineStore('users', () => {
       users.value.push(newUser)
       return newUser
     } catch (err: any) {
-      error.value = err.response?.data?.detail || '创建用户失败'
+      error.value = err.response?.data?.error?.message || err.response?.data?.detail || '创建用户失败'
       throw err
     } finally {
       loading.value = false
@@ -52,7 +52,7 @@ export const useUsersStore = defineStore('users', () => {
       }
       return updatedUser
     } catch (err: any) {
-      error.value = err.response?.data?.detail || '更新用户失败'
+      error.value = err.response?.data?.error?.message || err.response?.data?.detail || '更新用户失败'
       throw err
     } finally {
       loading.value = false
@@ -67,7 +67,7 @@ export const useUsersStore = defineStore('users', () => {
       await usersApi.deleteUser(userId)
       users.value = users.value.filter(u => u.id !== userId)
     } catch (err: any) {
-      error.value = err.response?.data?.detail || '删除用户失败'
+      error.value = err.response?.data?.error?.message || err.response?.data?.detail || '删除用户失败'
       throw err
     } finally {
       loading.value = false
@@ -78,7 +78,7 @@ export const useUsersStore = defineStore('users', () => {
     try {
       return await usersApi.getUserApiKeys(userId)
     } catch (err: any) {
-      error.value = err.response?.data?.detail || '获取 API Keys 失败'
+      error.value = err.response?.data?.error?.message || err.response?.data?.detail || '获取 API Keys 失败'
       throw err
     }
   }
@@ -87,7 +87,7 @@ export const useUsersStore = defineStore('users', () => {
     try {
       return await usersApi.createApiKey(userId, name)
     } catch (err: any) {
-      error.value = err.response?.data?.detail || '创建 API Key 失败'
+      error.value = err.response?.data?.error?.message || err.response?.data?.detail || '创建 API Key 失败'
       throw err
     }
   }
@@ -96,7 +96,7 @@ export const useUsersStore = defineStore('users', () => {
     try {
       await usersApi.deleteApiKey(userId, keyId)
     } catch (err: any) {
-      error.value = err.response?.data?.detail || '删除 API Key 失败'
+      error.value = err.response?.data?.error?.message || err.response?.data?.detail || '删除 API Key 失败'
       throw err
     }
   }
@@ -110,7 +110,7 @@ export const useUsersStore = defineStore('users', () => {
       // 刷新用户列表以获取最新数据
       await fetchUsers()
     } catch (err: any) {
-      error.value = err.response?.data?.detail || '重置配额失败'
+      error.value = err.response?.data?.error?.message || err.response?.data?.detail || '重置配额失败'
       throw err
     } finally {
       loading.value = false

frontend/src/utils/errorParser.ts

@@ -198,3 +198,49 @@ export function parseApiErrorShort(err: unknown, defaultMessage: string = '操
   const lines = fullError.split('\n')
   return lines[0] || defaultMessage
 }
+
+/**
+ * 解析模型测试响应的错误信息
+ * @param result 测试响应结果
+ * @returns 格式化的错误信息
+ */
+export function parseTestModelError(result: {
+  error?: string
+  data?: {
+    response?: {
+      status_code?: number
+      error?: string | { message?: string }
+    }
+  }
+}): string {
+  let errorMsg = result.error || '测试失败'
+
+  // 检查HTTP状态码错误
+  if (result.data?.response?.status_code) {
+    const status = result.data.response.status_code
+    if (status === 403) {
+      errorMsg = '认证失败: API密钥无效或客户端类型不被允许'
+    } else if (status === 401) {
+      errorMsg = '认证失败: API密钥无效或已过期'
+    } else if (status === 404) {
+      errorMsg = '模型不存在: 请检查模型名称是否正确'
+    } else if (status === 429) {
+      errorMsg = '请求频率过高: 请稍后重试'
+    } else if (status >= 500) {
+      errorMsg = `服务器错误: HTTP ${status}`
+    } else {
+      errorMsg = `请求失败: HTTP ${status}`
+    }
+  }
+
+  // 尝试从错误响应中提取更多信息
+  if (result.data?.response?.error) {
+    if (typeof result.data.response.error === 'string') {
+      errorMsg = result.data.response.error
+    } else if (result.data.response.error?.message) {
+      errorMsg = result.data.response.error.message
+    }
+  }
+
+  return errorMsg
+}

frontend/src/views/admin/ApiKeys.vue

@@ -650,6 +650,7 @@
 import { ref, computed, onMounted } from 'vue'
 import { useToast } from '@/composables/useToast'
 import { useConfirm } from '@/composables/useConfirm'
+import { useClipboard } from '@/composables/useClipboard'
 import { adminApi, type AdminApiKey, type CreateStandaloneApiKeyRequest } from '@/api/admin'
 
 import {
@@ -693,6 +694,7 @@ import { log } from '@/utils/logger'
 
 const { success, error } = useToast()
 const { confirmDanger } = useConfirm()
+const { copyToClipboard } = useClipboard()
 
 const apiKeys = ref<AdminApiKey[]>([])
 const loading = ref(false)
@@ -927,20 +929,14 @@ function selectKey() {
 }
 
 async function copyKey() {
-  try {
-    await navigator.clipboard.writeText(newKeyValue.value)
-    success('API Key 已复制到剪贴板')
-  } catch {
-    error('复制失败，请手动复制')
-  }
+  await copyToClipboard(newKeyValue.value)
 }
 
 async function copyKeyPrefix(apiKey: AdminApiKey) {
   try {
     // 调用后端 API 获取完整密钥
     const response = await adminApi.getFullApiKey(apiKey.id)
-    await navigator.clipboard.writeText(response.key)
-    success('完整密钥已复制到剪贴板')
+    await copyToClipboard(response.key)
   } catch (err) {
     log.error('复制密钥失败:', err)
     error('复制失败，请重试')
@@ -1046,9 +1042,10 @@ async function handleKeyFormSubmit(data: StandaloneKeyFormData) {
         rate_limit: data.rate_limit,
         expire_days: data.never_expire ? null : (data.expire_days || null),
         auto_delete_on_expiry: data.auto_delete_on_expiry,
-        allowed_providers: data.allowed_providers.length > 0 ? data.allowed_providers : undefined,
-        allowed_api_formats: data.allowed_api_formats.length > 0 ? data.allowed_api_formats : undefined,
-        allowed_models: data.allowed_models.length > 0 ? data.allowed_models : undefined
+        // 空数组表示清除限制（允许全部），后端会将空数组存为 NULL
+        allowed_providers: data.allowed_providers,
+        allowed_api_formats: data.allowed_api_formats,
+        allowed_models: data.allowed_models
       }
       await adminApi.updateApiKey(data.id, updateData)
       success('API Key 更新成功')
@@ -1064,9 +1061,10 @@ async function handleKeyFormSubmit(data: StandaloneKeyFormData) {
         rate_limit: data.rate_limit,
         expire_days: data.never_expire ? null : (data.expire_days || null),
         auto_delete_on_expiry: data.auto_delete_on_expiry,
-        allowed_providers: data.allowed_providers.length > 0 ? data.allowed_providers : undefined,
-        allowed_api_formats: data.allowed_api_formats.length > 0 ? data.allowed_api_formats : undefined,
-        allowed_models: data.allowed_models.length > 0 ? data.allowed_models : undefined
+        // 空数组表示不设置限制（允许全部），后端会将空数组存为 NULL
+        allowed_providers: data.allowed_providers,
+        allowed_api_formats: data.allowed_api_formats,
+        allowed_models: data.allowed_models
       }
       const response = await adminApi.createStandaloneApiKey(createData)
       newKeyValue.value = response.key

frontend/src/views/admin/CacheMonitoring.vue

@@ -46,6 +46,7 @@ const clearingRowAffinityKey = ref<string | null>(null)
 const currentPage = ref(1)
 const pageSize = ref(20)
 const currentTime = ref(Math.floor(Date.now() / 1000))
+const analysisHoursSelectOpen = ref(false)
 
 // ==================== 模型映射缓存 ====================
 
@@ -142,32 +143,37 @@ async function resetAffinitySearch() {
   await fetchAffinityList()
 }
 
-async function clearUserCache(identifier: string, displayName?: string) {
-  const target = identifier?.trim()
-  if (!target) {
-    showError('无法识别标识符')
+async function clearSingleAffinity(item: UserAffinity) {
+  const affinityKey = item.affinity_key?.trim()
+  const endpointId = item.endpoint_id?.trim()
+  const modelId = item.global_model_id?.trim()
+  const apiFormat = item.api_format?.trim()
+
+  if (!affinityKey || !endpointId || !modelId || !apiFormat) {
+    showError('缓存记录信息不完整，无法删除')
     return
   }
 
-  const label = displayName || target
+  const label = item.user_api_key_name || affinityKey
+  const modelLabel = item.model_display_name || item.model_name || modelId
   const confirmed = await showConfirm({
     title: '确认清除',
-    message: `确定要清除 ${label} 的缓存吗？`,
+    message: `确定要清除 ${label} 在模型 ${modelLabel} 上的缓存亲和性吗？`,
     confirmText: '确认清除',
     variant: 'destructive'
   })
 
   if (!confirmed) return
 
-  clearingRowAffinityKey.value = target
+  clearingRowAffinityKey.value = affinityKey
   try {
-    await cacheApi.clearUserCache(target)
+    await cacheApi.clearSingleAffinity(affinityKey, endpointId, modelId, apiFormat)
     showSuccess('清除成功')
     await fetchCacheStats()
     await fetchAffinityList(tableKeyword.value.trim() || undefined)
   } catch (error) {
     showError('清除失败')
-    log.error('清除用户缓存失败', error)
+    log.error('清除单条缓存失败', error)
   } finally {
     clearingRowAffinityKey.value = null
   }
@@ -618,7 +624,7 @@ onBeforeUnmount(() => {
                 class="h-7 w-7 text-muted-foreground/70 hover:text-destructive"
                 :disabled="clearingRowAffinityKey === item.affinity_key"
                 title="清除缓存"
-                @click="clearUserCache(item.affinity_key, item.user_api_key_name || item.affinity_key)"
+                @click="clearSingleAffinity(item)"
               >
                 <Trash2 class="h-3.5 w-3.5" />
               </Button>
@@ -668,7 +674,7 @@ onBeforeUnmount(() => {
               variant="ghost"
               class="h-7 w-7 text-muted-foreground/70 hover:text-destructive shrink-0"
               :disabled="clearingRowAffinityKey === item.affinity_key"
-              @click="clearUserCache(item.affinity_key, item.user_api_key_name || item.affinity_key)"
+              @click="clearSingleAffinity(item)"
             >
               <Trash2 class="h-3.5 w-3.5" />
             </Button>
@@ -1051,7 +1057,7 @@ onBeforeUnmount(() => {
             <span class="text-xs text-muted-foreground hidden sm:inline">分析用户请求间隔，推荐合适的缓存 TTL</span>
           </div>
           <div class="flex flex-wrap items-center gap-2">
-            <Select v-model="analysisHours">
+            <Select v-model="analysisHours" v-model:open="analysisHoursSelectOpen">
               <SelectTrigger class="w-24 sm:w-28 h-8">
                 <SelectValue placeholder="时间段" />
               </SelectTrigger>

frontend/src/views/admin/ModelManagement.vue

@@ -713,6 +713,7 @@ import ProviderModelFormDialog from '@/features/providers/components/ProviderMod
 import type { Model } from '@/api/endpoints'
 import { useToast } from '@/composables/useToast'
 import { useConfirm } from '@/composables/useConfirm'
+import { useClipboard } from '@/composables/useClipboard'
 import { useRowClick } from '@/composables/useRowClick'
 import { parseApiError } from '@/utils/errorParser'
 import {
@@ -743,6 +744,7 @@ import { getProvidersSummary } from '@/api/endpoints/providers'
 import { getAllCapabilities, type CapabilityDefinition } from '@/api/endpoints'
 
 const { success, error: showError } = useToast()
+const { copyToClipboard } = useClipboard()
 
 // 状态
 const loading = ref(false)
@@ -1066,16 +1068,6 @@ function handleRowClick(event: MouseEvent, model: GlobalModelResponse) {
   selectModel(model)
 }
 
-// 复制到剪贴板
-async function copyToClipboard(text: string) {
-  try {
-    await navigator.clipboard.writeText(text)
-    success('已复制')
-  } catch {
-    showError('复制失败')
-  }
-}
-
 async function selectModel(model: GlobalModelResponse) {
   selectedModel.value = model
   detailTab.value = 'basic'

frontend/src/views/admin/ProviderManagement.vue

@@ -723,9 +723,19 @@ async function handleDeleteProvider(provider: ProviderWithEndpointsSummary) {
 // 切换提供商状态
 async function toggleProviderStatus(provider: ProviderWithEndpointsSummary) {
   try {
-    await updateProvider(provider.id, { is_active: !provider.is_active })
-    provider.is_active = !provider.is_active
-    showSuccess(provider.is_active ? '提供商已启用' : '提供商已停用')
+    const newStatus = !provider.is_active
+    await updateProvider(provider.id, { is_active: newStatus })
+
+    // 更新抽屉内部的 provider 对象
+    provider.is_active = newStatus
+
+    // 同时更新主页面 providers 数组中的对象，实现无感更新
+    const targetProvider = providers.value.find(p => p.id === provider.id)
+    if (targetProvider) {
+      targetProvider.is_active = newStatus
+    }
+
+    showSuccess(newStatus ? '提供商已启用' : '提供商已停用')
   } catch (err: any) {
     showError(err.response?.data?.detail || '操作失败', '错误')
   }

frontend/src/views/admin/SystemSettings.vue

@@ -465,77 +465,6 @@
         </div>
       </CardSection>
 
-      <!-- 流式输出配置 -->
-      <CardSection
-        title="流式输出"
-        description="配置流式响应的输出效果"
-      >
-        <div class="grid grid-cols-1 md:grid-cols-2 gap-6">
-          <div class="md:col-span-2">
-            <div class="flex items-center space-x-2">
-              <Checkbox
-                id="stream-smoothing-enabled"
-                v-model:checked="systemConfig.stream_smoothing_enabled"
-              />
-              <div>
-                <Label
-                  for="stream-smoothing-enabled"
-                  class="cursor-pointer"
-                >
-                  启用平滑输出
-                </Label>
-                <p class="text-xs text-muted-foreground">
-                  将上游返回的大块内容拆分成小块，模拟打字效果
-                </p>
-              </div>
-            </div>
-          </div>
-
-          <div>
-            <Label
-              for="stream-smoothing-chunk-size"
-              class="block text-sm font-medium"
-            >
-              每块字符数
-            </Label>
-            <Input
-              id="stream-smoothing-chunk-size"
-              v-model.number="systemConfig.stream_smoothing_chunk_size"
-              type="number"
-              min="1"
-              max="100"
-              placeholder="20"
-              class="mt-1"
-              :disabled="!systemConfig.stream_smoothing_enabled"
-            />
-            <p class="mt-1 text-xs text-muted-foreground">
-              每次输出的字符数量（1-100）
-            </p>
-          </div>
-
-          <div>
-            <Label
-              for="stream-smoothing-delay-ms"
-              class="block text-sm font-medium"
-            >
-              输出间隔 (毫秒)
-            </Label>
-            <Input
-              id="stream-smoothing-delay-ms"
-              v-model.number="systemConfig.stream_smoothing_delay_ms"
-              type="number"
-              min="1"
-              max="100"
-              placeholder="8"
-              class="mt-1"
-              :disabled="!systemConfig.stream_smoothing_enabled"
-            />
-            <p class="mt-1 text-xs text-muted-foreground">
-              每块之间的延迟毫秒数（1-100）
-            </p>
-          </div>
-        </div>
-      </CardSection>
     </div>
 
     <!-- 导入配置对话框 -->
@@ -884,10 +813,6 @@ interface SystemConfig {
   log_retention_days: number
   cleanup_batch_size: number
   audit_log_retention_days: number
-  // 流式输出
-  stream_smoothing_enabled: boolean
-  stream_smoothing_chunk_size: number
-  stream_smoothing_delay_ms: number
 }
 
 const loading = ref(false)
@@ -937,10 +862,6 @@ const systemConfig = ref<SystemConfig>({
   log_retention_days: 365,
   cleanup_batch_size: 1000,
   audit_log_retention_days: 30,
-  // 流式输出
-  stream_smoothing_enabled: false,
-  stream_smoothing_chunk_size: 20,
-  stream_smoothing_delay_ms: 8,
 })
 
 // 计算属性：KB 和 字节 之间的转换
@@ -997,10 +918,6 @@ async function loadSystemConfig() {
       'log_retention_days',
       'cleanup_batch_size',
       'audit_log_retention_days',
-      // 流式输出
-      'stream_smoothing_enabled',
-      'stream_smoothing_chunk_size',
-      'stream_smoothing_delay_ms',
     ]
 
     for (const key of configs) {
@@ -1108,22 +1025,6 @@ async function saveSystemConfig() {
         value: systemConfig.value.audit_log_retention_days,
         description: '审计日志保留天数'
       },
-      // 流式输出
-      {
-        key: 'stream_smoothing_enabled',
-        value: systemConfig.value.stream_smoothing_enabled,
-        description: '是否启用流式平滑输出'
-      },
-      {
-        key: 'stream_smoothing_chunk_size',
-        value: systemConfig.value.stream_smoothing_chunk_size,
-        description: '流式平滑输出每个小块的字符数'
-      },
-      {
-        key: 'stream_smoothing_delay_ms',
-        value: systemConfig.value.stream_smoothing_delay_ms,
-        description: '流式平滑输出每个小块之间的延迟毫秒数'
-      },
     ]
 
     const promises = configItems.map(item =>

frontend/src/views/admin/Users.vue

@@ -701,6 +701,7 @@ import { ref, computed, onMounted, watch } from 'vue'
 import { useUsersStore } from '@/stores/users'
 import { useToast } from '@/composables/useToast'
 import { useConfirm } from '@/composables/useConfirm'
+import { useClipboard } from '@/composables/useClipboard'
 import { usageApi, type UsageByUser } from '@/api/usage'
 import { adminApi } from '@/api/admin'
 
@@ -748,6 +749,7 @@ import { log } from '@/utils/logger'
 
 const { success, error } = useToast()
 const { confirmDanger, confirmWarning } = useConfirm()
+const { copyToClipboard } = useClipboard()
 const usersStore = useUsersStore()
 
 // 用户表单对话框状态
@@ -875,7 +877,8 @@ async function toggleUserStatus(user: any) {
   const action = user.is_active ? '禁用' : '启用'
   const confirmed = await confirmDanger(
     `确定要${action}用户 ${user.username} 吗？`,
-    `${action}用户`
+    `${action}用户`,
+    action
   )
 
   if (!confirmed) return
@@ -884,7 +887,7 @@ async function toggleUserStatus(user: any) {
     await usersStore.updateUser(user.id, { is_active: !user.is_active })
     success(`用户已${action}`)
   } catch (err: any) {
-    error(err.response?.data?.detail || '未知错误', `${action}用户失败`)
+    error(err.response?.data?.error?.message || err.response?.data?.detail || '未知错误', `${action}用户失败`)
   }
 }
 
@@ -955,7 +958,7 @@ async function handleUserFormSubmit(data: UserFormData & { password?: string })
     closeUserFormDialog()
   } catch (err: any) {
     const title = data.id ? '更新用户失败' : '创建用户失败'
-    error(err.response?.data?.detail || '未知错误', title)
+    error(err.response?.data?.error?.message || err.response?.data?.detail || '未知错误', title)
   } finally {
     userFormDialogRef.value?.setSaving(false)
   }
@@ -989,7 +992,7 @@ async function createApiKey() {
     showNewApiKeyDialog.value = true
     await loadUserApiKeys(selectedUser.value.id)
   } catch (err: any) {
-    error(err.response?.data?.detail || '未知错误', '创建 API Key 失败')
+    error(err.response?.data?.error?.message || err.response?.data?.detail || '未知错误', '创建 API Key 失败')
   } finally {
     creatingApiKey.value = false
   }
@@ -1000,12 +1003,7 @@ function selectApiKey() {
 }
 
 async function copyApiKey() {
-  try {
-    await navigator.clipboard.writeText(newApiKey.value)
-    success('API Key已复制到剪贴板')
-  } catch {
-    error('复制失败，请手动复制')
-  }
+  await copyToClipboard(newApiKey.value)
 }
 
 async function closeNewApiKeyDialog() {
@@ -1026,7 +1024,7 @@ async function deleteApiKey(apiKey: any) {
     await loadUserApiKeys(selectedUser.value.id)
     success('API Key已删除')
   } catch (err: any) {
-    error(err.response?.data?.detail || '未知错误', '删除 API Key 失败')
+    error(err.response?.data?.error?.message || err.response?.data?.detail || '未知错误', '删除 API Key 失败')
   }
 }
 
@@ -1034,11 +1032,10 @@ async function copyFullKey(apiKey: any) {
   try {
     // 调用后端 API 获取完整密钥
     const response = await adminApi.getFullApiKey(apiKey.id)
-    await navigator.clipboard.writeText(response.key)
-    success('完整密钥已复制到剪贴板')
+    await copyToClipboard(response.key)
   } catch (err: any) {
     log.error('复制密钥失败:', err)
-    error(err.response?.data?.detail || '未知错误', '复制密钥失败')
+    error(err.response?.data?.error?.message || err.response?.data?.detail || '未知错误', '复制密钥失败')
   }
 }
 
@@ -1054,7 +1051,7 @@ async function resetQuota(user: any) {
     await usersStore.resetUserQuota(user.id)
     success('配额已重置')
   } catch (err: any) {
-    error(err.response?.data?.detail || '未知错误', '重置配额失败')
+    error(err.response?.data?.error?.message || err.response?.data?.detail || '未知错误', '重置配额失败')
   }
 }
 
@@ -1070,7 +1067,7 @@ async function deleteUser(user: any) {
     await usersStore.deleteUser(user.id)
     success('用户已删除')
   } catch (err: any) {
-    error(err.response?.data?.detail || '未知错误', '删除用户失败')
+    error(err.response?.data?.error?.message || err.response?.data?.detail || '未知错误', '删除用户失败')
   }
 }
 </script>

frontend/src/views/public/Home.vue

@@ -102,9 +102,9 @@
     <!-- Main Content -->
     <main class="relative z-10">
       <!-- Fixed Logo Container -->
-      <div class="fixed inset-0 z-20 pointer-events-none flex items-center justify-center overflow-hidden">
+      <div class="mt-4 fixed inset-0 z-20 pointer-events-none flex items-center justify-center overflow-hidden">
         <div
-          class="transform-gpu logo-container"
+          class="mt-16 transform-gpu logo-container"
           :class="[currentSection === SECTIONS.HOME ? 'home-section' : '', `logo-transition-${scrollDirection}`]"
           :style="fixedLogoStyle"
         >
@@ -151,7 +151,7 @@
         class="min-h-screen snap-start flex items-center justify-center px-16 lg:px-20 py-20"
       >
         <div class="max-w-4xl mx-auto text-center">
-          <div class="h-80 w-full mb-16" />
+          <div class="h-80 w-full mb-16 mt-8" />
           <h1
             class="mb-6 text-5xl md:text-7xl font-bold text-[#191919] dark:text-white leading-tight transition-all duration-700"
             :style="getTitleStyle(SECTIONS.HOME)"
@@ -166,7 +166,7 @@
             整合 Claude Code、Codex CLI、Gemini CLI 等多个 AI 编程助手
           </p>
           <button
-            class="mt-16 transition-all duration-700 cursor-pointer hover:scale-110"
+            class="mt-8 transition-all duration-700 cursor-pointer hover:scale-110"
             :style="getScrollIndicatorStyle(SECTIONS.HOME)"
             @click="scrollToSection(SECTIONS.CLAUDE)"
           >

frontend/src/views/shared/Usage.vue

@@ -65,6 +65,7 @@
       :page-size="pageSize"
       :total-records="totalRecords"
       :page-size-options="pageSizeOptions"
+      :auto-refresh="globalAutoRefresh"
       @update:selected-period="handlePeriodChange"
       @update:filter-user="handleFilterUserChange"
       @update:filter-model="handleFilterModelChange"
@@ -72,6 +73,7 @@
       @update:filter-status="handleFilterStatusChange"
       @update:current-page="handlePageChange"
       @update:page-size="handlePageSizeChange"
+      @update:auto-refresh="handleAutoRefreshChange"
       @refresh="refreshData"
       @export="exportData"
       @show-detail="showRequestDetail"
@@ -214,7 +216,10 @@ const hasActiveRequests = computed(() => activeRequestIds.value.length > 0)
 
 // 自动刷新定时器
 let autoRefreshTimer: ReturnType<typeof setInterval> | null = null
-const AUTO_REFRESH_INTERVAL = 1000 // 1秒刷新一次
+let globalAutoRefreshTimer: ReturnType<typeof setInterval> | null = null
+const AUTO_REFRESH_INTERVAL = 1000 // 1秒刷新一次（用于活跃请求）
+const GLOBAL_AUTO_REFRESH_INTERVAL = 10000 // 10秒刷新一次（全局自动刷新）
+const globalAutoRefresh = ref(false) // 全局自动刷新开关
 
 // 轮询活跃请求状态（轻量级，只更新状态变化的记录）
 async function pollActiveRequests() {
@@ -278,9 +283,35 @@ watch(hasActiveRequests, (hasActive) => {
   }
 }, { immediate: true })
 
+// 启动全局自动刷新
+function startGlobalAutoRefresh() {
+  if (globalAutoRefreshTimer) return
+  globalAutoRefreshTimer = setInterval(refreshData, GLOBAL_AUTO_REFRESH_INTERVAL)
+}
+
+// 停止全局自动刷新
+function stopGlobalAutoRefresh() {
+  if (globalAutoRefreshTimer) {
+    clearInterval(globalAutoRefreshTimer)
+    globalAutoRefreshTimer = null
+  }
+}
+
+// 处理自动刷新开关变化
+function handleAutoRefreshChange(value: boolean) {
+  globalAutoRefresh.value = value
+  if (value) {
+    refreshData() // 立即刷新一次
+    startGlobalAutoRefresh()
+  } else {
+    stopGlobalAutoRefresh()
+  }
+}
+
 // 组件卸载时清理定时器
 onUnmounted(() => {
   stopAutoRefresh()
+  stopGlobalAutoRefresh()
 })
 
 // 用户页面的前端分页

frontend/src/views/user/ModelCatalog.vue

@@ -342,6 +342,7 @@ import {
   Plus,
 } from 'lucide-vue-next'
 import { useToast } from '@/composables/useToast'
+import { useClipboard } from '@/composables/useClipboard'
 import {
   Card,
   Table,
@@ -370,6 +371,7 @@ import { useRowClick } from '@/composables/useRowClick'
 import { log } from '@/utils/logger'
 
 const { success, error: showError } = useToast()
+const { copyToClipboard } = useClipboard()
 
 // 状态
 const loading = ref(false)
@@ -565,16 +567,6 @@ function hasTieredPricing(model: PublicGlobalModel): boolean {
   return (tiered?.tiers?.length || 0) > 1
 }
 
-async function copyToClipboard(text: string) {
-  try {
-    await navigator.clipboard.writeText(text)
-    success('已复制')
-  } catch (err) {
-    log.error('复制失败:', err)
-    showError('复制失败')
-  }
-}
-
 onMounted(() => {
   refreshData()
 })

frontend/src/views/user/components/UserModelDetailDrawer.vue

@@ -350,7 +350,9 @@ import {
   Layers,
   Image as ImageIcon
 } from 'lucide-vue-next'
+import { useEscapeKey } from '@/composables/useEscapeKey'
 import { useToast } from '@/composables/useToast'
+import { useClipboard } from '@/composables/useClipboard'
 import Card from '@/components/ui/card.vue'
 import Badge from '@/components/ui/badge.vue'
 import Button from '@/components/ui/button.vue'
@@ -374,6 +376,7 @@ const emit = defineEmits<{
 }>()
 
 const { success: showSuccess, error: showError } = useToast()
+const { copyToClipboard } = useClipboard()
 
 interface Props {
   model: PublicGlobalModel | null
@@ -407,15 +410,6 @@ function handleClose() {
   emit('update:open', false)
 }
 
-async function copyToClipboard(text: string) {
-  try {
-    await navigator.clipboard.writeText(text)
-    showSuccess('已复制')
-  } catch {
-    showError('复制失败')
-  }
-}
-
 function getFirstTierPrice(
   tieredPricing: TieredPricingConfig | undefined | null,
   priceKey: 'input_price_per_1m' | 'output_price_per_1m' | 'cache_creation_price_per_1m' | 'cache_read_price_per_1m'
@@ -453,6 +447,16 @@ function getFirst1hCachePrice(tieredPricing: TieredPricingConfig | undefined | n
   if (!tieredPricing?.tiers?.length) return '-'
   return get1hCachePrice(tieredPricing.tiers[0])
 }
+
+// 添加 ESC 键监听
+useEscapeKey(() => {
+  if (props.open) {
+    handleClose()
+  }
+}, {
+  disableOnInput: true,
+  once: false
+})
 </script>
 
 <style scoped>

src/api/admin/endpoints/keys.py

@@ -246,6 +246,15 @@ class AdminUpdateEndpointKeyAdapter(AdminApiAdapter):
         if "api_key" in update_data:
             update_data["api_key"] = crypto_service.encrypt(update_data["api_key"])
 
+        # 特殊处理 max_concurrent：需要区分"未提供"和"显式设置为 null"
+        # 当 max_concurrent 被显式设置时（在 model_fields_set 中），即使值为 None 也应该更新
+        if "max_concurrent" in self.key_data.model_fields_set:
+            update_data["max_concurrent"] = self.key_data.max_concurrent
+            # 切换到自适应模式时，清空学习到的并发限制，让系统重新学习
+            if self.key_data.max_concurrent is None:
+                update_data["learned_max_concurrent"] = None
+                logger.info("Key %s 切换为自适应并发模式", self.key_id)
+
         for field, value in update_data.items():
             setattr(key, field, value)
         key.updated_at = datetime.now(timezone.utc)
@@ -253,7 +262,7 @@ class AdminUpdateEndpointKeyAdapter(AdminApiAdapter):
         db.commit()
         db.refresh(key)
 
-        logger.info(f"[OK] 更新 Key: ID={self.key_id}, Updates={list(update_data.keys())}")
+        logger.info("[OK] 更新 Key: ID=%s, Updates=%s", self.key_id, list(update_data.keys()))
 
         try:
             decrypted_key = crypto_service.decrypt(key.api_key)

src/api/admin/monitoring/cache.py

@@ -186,6 +186,30 @@ async def clear_user_cache(
     return await pipeline.run(adapter=adapter, http_request=request, db=db, mode=adapter.mode)
 
 
+@router.delete("/affinity/{affinity_key}/{endpoint_id}/{model_id}/{api_format}")
+async def clear_single_affinity(
+    affinity_key: str,
+    endpoint_id: str,
+    model_id: str,
+    api_format: str,
+    request: Request,
+    db: Session = Depends(get_db),
+) -> Any:
+    """
+    Clear a single cache affinity entry
+
+    Parameters:
+    - affinity_key: API Key ID
+    - endpoint_id: Endpoint ID
+    - model_id: Model ID (GlobalModel ID)
+    - api_format: API format (claude/openai)
+    """
+    adapter = AdminClearSingleAffinityAdapter(
+        affinity_key=affinity_key, endpoint_id=endpoint_id, model_id=model_id, api_format=api_format
+    )
+    return await pipeline.run(adapter=adapter, http_request=request, db=db, mode=adapter.mode)
+
+
 @router.delete("")
 async def clear_all_cache(
     request: Request,
@@ -655,6 +679,7 @@ class AdminListAffinitiesAdapter(AdminApiAdapter):
                 "key_name": key.name if key else None,
                 "key_prefix": provider_key_masked,
                 "rate_multiplier": key.rate_multiplier if key else 1.0,
+                "global_model_id": affinity.get("model_name"),  # 原始的 global_model_id
                 "model_name": (
                     global_model_map.get(affinity.get("model_name")).name
                     if affinity.get("model_name") and global_model_map.get(affinity.get("model_name"))
@@ -817,6 +842,65 @@ class AdminClearUserCacheAdapter(AdminApiAdapter):
             raise HTTPException(status_code=500, detail=f"清除失败: {exc}")
 
 
+@dataclass
+class AdminClearSingleAffinityAdapter(AdminApiAdapter):
+    affinity_key: str
+    endpoint_id: str
+    model_id: str
+    api_format: str
+
+    async def handle(self, context: ApiRequestContext) -> Dict[str, Any]:  # type: ignore[override]
+        db = context.db
+        try:
+            redis_client = get_redis_client_sync()
+            affinity_mgr = await get_affinity_manager(redis_client)
+
+            # 直接获取指定的亲和性记录（无需遍历全部）
+            existing_affinity = await affinity_mgr.get_affinity(
+                self.affinity_key, self.api_format, self.model_id
+            )
+
+            if not existing_affinity:
+                raise HTTPException(status_code=404, detail="未找到指定的缓存亲和性记录")
+
+            # 验证 endpoint_id 是否匹配
+            if existing_affinity.endpoint_id != self.endpoint_id:
+                raise HTTPException(status_code=404, detail="未找到指定的缓存亲和性记录")
+
+            # 失效单条记录
+            await affinity_mgr.invalidate_affinity(
+                self.affinity_key, self.api_format, self.model_id, endpoint_id=self.endpoint_id
+            )
+
+            # 获取用于日志的信息
+            api_key = db.query(ApiKey).filter(ApiKey.id == self.affinity_key).first()
+            api_key_name = api_key.name if api_key else None
+
+            logger.info(
+                f"已清除单条缓存亲和性: affinity_key={self.affinity_key[:8]}..., "
+                f"endpoint_id={self.endpoint_id[:8]}..., model_id={self.model_id[:8]}..."
+            )
+
+            context.add_audit_metadata(
+                action="cache_clear_single",
+                affinity_key=self.affinity_key,
+                endpoint_id=self.endpoint_id,
+                model_id=self.model_id,
+            )
+            return {
+                "status": "ok",
+                "message": f"已清除缓存亲和性: {api_key_name or self.affinity_key[:8]}",
+                "affinity_key": self.affinity_key,
+                "endpoint_id": self.endpoint_id,
+                "model_id": self.model_id,
+            }
+        except HTTPException:
+            raise
+        except Exception as exc:
+            logger.exception(f"清除单条缓存亲和性失败: {exc}")
+            raise HTTPException(status_code=500, detail=f"清除失败: {exc}")
+
+
 class AdminClearAllCacheAdapter(AdminApiAdapter):
     async def handle(self, context: ApiRequestContext) -> Dict[str, Any]:  # type: ignore[override]
         try:
@@ -863,7 +947,7 @@ class AdminClearProviderCacheAdapter(AdminApiAdapter):
 class AdminCacheConfigAdapter(AdminApiAdapter):
     async def handle(self, context: ApiRequestContext) -> Dict[str, Any]:  # type: ignore[override]
         from src.services.cache.affinity_manager import CacheAffinityManager
-        from src.services.cache.aware_scheduler import CacheAwareScheduler
+        from src.config.constants import ConcurrencyDefaults
         from src.services.rate_limit.adaptive_reservation import get_adaptive_reservation_manager
 
         # 获取动态预留管理器的配置
@@ -874,7 +958,7 @@ class AdminCacheConfigAdapter(AdminApiAdapter):
             "status": "ok",
             "data": {
                 "cache_ttl_seconds": CacheAffinityManager.DEFAULT_CACHE_TTL,
-                "cache_reservation_ratio": CacheAwareScheduler.CACHE_RESERVATION_RATIO,
+                "cache_reservation_ratio": ConcurrencyDefaults.CACHE_RESERVATION_RATIO,
                 "dynamic_reservation": {
                     "enabled": True,
                     "config": reservation_stats["config"],
@@ -897,7 +981,7 @@ class AdminCacheConfigAdapter(AdminApiAdapter):
         context.add_audit_metadata(
             action="cache_config",
             cache_ttl_seconds=CacheAffinityManager.DEFAULT_CACHE_TTL,
-            cache_reservation_ratio=CacheAwareScheduler.CACHE_RESERVATION_RATIO,
+            cache_reservation_ratio=ConcurrencyDefaults.CACHE_RESERVATION_RATIO,
             dynamic_reservation_enabled=True,
         )
         return response
@@ -1083,14 +1167,14 @@ class AdminModelMappingCacheStatsAdapter(AdminApiAdapter):
                                                 provider.display_name or provider.name
                                             )
                                             continue
-                                        # 检查是否在别名列表中
-                                        if model.provider_model_aliases:
-                                            alias_names = [
+                                        # 检查是否在映射列表中
+                                        if model.provider_model_mappings:
+                                            mapping_list = [
                                                 a.get("name")
-                                                for a in model.provider_model_aliases
+                                                for a in model.provider_model_mappings
                                                 if isinstance(a, dict)
                                             ]
-                                            if mapping_name in alias_names:
+                                            if mapping_name in mapping_list:
                                                 provider_names.append(
                                                     provider.display_name or provider.name
                                                 )
@@ -1152,19 +1236,19 @@ class AdminModelMappingCacheStatsAdapter(AdminApiAdapter):
                         try:
                             cached_data = json.loads(cached_str)
                             provider_model_name = cached_data.get("provider_model_name")
-                            provider_model_aliases = cached_data.get("provider_model_aliases", [])
+                            cached_model_mappings = cached_data.get("provider_model_mappings", [])
 
                             # 获取 Provider 和 GlobalModel 信息
                             provider = provider_map.get(provider_id)
                             global_model = global_model_map.get(global_model_id)
 
                             if provider and global_model:
-                                # 提取别名名称
-                                alias_names = []
-                                if provider_model_aliases:
-                                    for alias_entry in provider_model_aliases:
-                                        if isinstance(alias_entry, dict) and alias_entry.get("name"):
-                                            alias_names.append(alias_entry["name"])
+                                # 提取映射名称
+                                mapping_names = []
+                                if cached_model_mappings:
+                                    for mapping_entry in cached_model_mappings:
+                                        if isinstance(mapping_entry, dict) and mapping_entry.get("name"):
+                                            mapping_names.append(mapping_entry["name"])
 
                                 # provider_model_name 为空时跳过
                                 if not provider_model_name:
@@ -1172,14 +1256,14 @@ class AdminModelMappingCacheStatsAdapter(AdminApiAdapter):
 
                                 # 只显示有实际映射的条目：
                                 # 1. 全局模型名 != Provider 模型名（模型名称映射）
-                                # 2. 或者有别名配置
+                                # 2. 或者有映射配置
                                 has_name_mapping = global_model.name != provider_model_name
-                                has_aliases = len(alias_names) > 0
+                                has_mappings = len(mapping_names) > 0
 
-                                if has_name_mapping or has_aliases:
-                                    # 构建用于展示的别名列表
-                                    # 如果只有名称映射没有别名，则用 global_model_name 作为"请求名称"
-                                    display_aliases = alias_names if alias_names else [global_model.name]
+                                if has_name_mapping or has_mappings:
+                                    # 构建用于展示的映射列表
+                                    # 如果只有名称映射没有额外映射，则用 global_model_name 作为"请求名称"
+                                    display_mappings = mapping_names if mapping_names else [global_model.name]
 
                                     provider_model_mappings.append({
                                         "provider_id": provider_id,
@@ -1188,7 +1272,7 @@ class AdminModelMappingCacheStatsAdapter(AdminApiAdapter):
                                         "global_model_name": global_model.name,
                                         "global_model_display_name": global_model.display_name,
                                         "provider_model_name": provider_model_name,
-                                        "aliases": display_aliases,
+                                        "aliases": display_mappings,
                                         "ttl": ttl if ttl > 0 else None,
                                         "hit_count": hit_count,
                                     })

src/api/admin/provider_query.py

@@ -11,6 +11,8 @@ from fastapi import APIRouter, Depends, HTTPException
 from pydantic import BaseModel
 from sqlalchemy.orm import Session, joinedload
 
+from src.api.handlers.base.chat_adapter_base import get_adapter_class
+from src.api.handlers.base.cli_adapter_base import get_cli_adapter_class
 from src.core.crypto import crypto_service
 from src.core.logger import logger
 from src.database.database import get_db
@@ -30,145 +32,33 @@ class ModelsQueryRequest(BaseModel):
     api_key_id: Optional[str] = None
 
 
+class TestModelRequest(BaseModel):
+    """模型测试请求"""
+
+    provider_id: str
+    model_name: str
+    api_key_id: Optional[str] = None
+    stream: bool = False
+    message: Optional[str] = "你好"
+    api_format: Optional[str] = None  # 指定使用的API格式，如果不指定则使用端点的默认格式
+
+
 # ============ API Endpoints ============
 
 
-async def _fetch_openai_models(
-    client: httpx.AsyncClient,
-    base_url: str,
-    api_key: str,
-    api_format: str,
-    extra_headers: Optional[dict] = None,
-) -> tuple[list, Optional[str]]:
-    """获取 OpenAI 格式的模型列表
+def _get_adapter_for_format(api_format: str):
+    """根据 API 格式获取对应的 Adapter 类"""
+    # 先检查 Chat Adapter 注册表
+    adapter_class = get_adapter_class(api_format)
+    if adapter_class:
+        return adapter_class
 
-    Returns:
-        tuple[list, Optional[str]]: (模型列表, 错误信息)
-    """
-    headers = {"Authorization": f"Bearer {api_key}"}
-    if extra_headers:
-        # 防止 extra_headers 覆盖 Authorization
-        safe_headers = {k: v for k, v in extra_headers.items() if k.lower() != "authorization"}
-        headers.update(safe_headers)
+    # 再检查 CLI Adapter 注册表
+    cli_adapter_class = get_cli_adapter_class(api_format)
+    if cli_adapter_class:
+        return cli_adapter_class
 
-    # 构建 /v1/models URL
-    if base_url.endswith("/v1"):
-        models_url = f"{base_url}/models"
-    else:
-        models_url = f"{base_url}/v1/models"
-
-    try:
-        response = await client.get(models_url, headers=headers)
-        logger.debug(f"OpenAI models request to {models_url}: status={response.status_code}")
-        if response.status_code == 200:
-            data = response.json()
-            models = []
-            if "data" in data:
-                models = data["data"]
-            elif isinstance(data, list):
-                models = data
-            # 为每个模型添加 api_format 字段
-            for m in models:
-                m["api_format"] = api_format
-            return models, None
-        else:
-            # 记录详细的错误信息
-            error_body = response.text[:500] if response.text else "(empty)"
-            error_msg = f"HTTP {response.status_code}: {error_body}"
-            logger.warning(f"OpenAI models request to {models_url} failed: {error_msg}")
-            return [], error_msg
-    except Exception as e:
-        error_msg = f"Request error: {str(e)}"
-        logger.warning(f"Failed to fetch models from {models_url}: {e}")
-        return [], error_msg
-
-
-async def _fetch_claude_models(
-    client: httpx.AsyncClient, base_url: str, api_key: str, api_format: str
-) -> tuple[list, Optional[str]]:
-    """获取 Claude 格式的模型列表
-
-    Returns:
-        tuple[list, Optional[str]]: (模型列表, 错误信息)
-    """
-    headers = {
-        "x-api-key": api_key,
-        "Authorization": f"Bearer {api_key}",
-        "anthropic-version": "2023-06-01",
-    }
-
-    # 构建 /v1/models URL
-    if base_url.endswith("/v1"):
-        models_url = f"{base_url}/models"
-    else:
-        models_url = f"{base_url}/v1/models"
-
-    try:
-        response = await client.get(models_url, headers=headers)
-        logger.debug(f"Claude models request to {models_url}: status={response.status_code}")
-        if response.status_code == 200:
-            data = response.json()
-            models = []
-            if "data" in data:
-                models = data["data"]
-            elif isinstance(data, list):
-                models = data
-            # 为每个模型添加 api_format 字段
-            for m in models:
-                m["api_format"] = api_format
-            return models, None
-        else:
-            error_body = response.text[:500] if response.text else "(empty)"
-            error_msg = f"HTTP {response.status_code}: {error_body}"
-            logger.warning(f"Claude models request to {models_url} failed: {error_msg}")
-            return [], error_msg
-    except Exception as e:
-        error_msg = f"Request error: {str(e)}"
-        logger.warning(f"Failed to fetch Claude models from {models_url}: {e}")
-        return [], error_msg
-
-
-async def _fetch_gemini_models(
-    client: httpx.AsyncClient, base_url: str, api_key: str, api_format: str
-) -> tuple[list, Optional[str]]:
-    """获取 Gemini 格式的模型列表
-
-    Returns:
-        tuple[list, Optional[str]]: (模型列表, 错误信息)
-    """
-    # 兼容 base_url 已包含 /v1beta 的情况
-    base_url_clean = base_url.rstrip("/")
-    if base_url_clean.endswith("/v1beta"):
-        models_url = f"{base_url_clean}/models?key={api_key}"
-    else:
-        models_url = f"{base_url_clean}/v1beta/models?key={api_key}"
-
-    try:
-        response = await client.get(models_url)
-        logger.debug(f"Gemini models request to {models_url}: status={response.status_code}")
-        if response.status_code == 200:
-            data = response.json()
-            if "models" in data:
-                # 转换为统一格式
-                return [
-                    {
-                        "id": m.get("name", "").replace("models/", ""),
-                        "owned_by": "google",
-                        "display_name": m.get("displayName", ""),
-                        "api_format": api_format,
-                    }
-                    for m in data["models"]
-                ], None
-            return [], None
-        else:
-            error_body = response.text[:500] if response.text else "(empty)"
-            error_msg = f"HTTP {response.status_code}: {error_body}"
-            logger.warning(f"Gemini models request to {models_url} failed: {error_msg}")
-            return [], error_msg
-    except Exception as e:
-        error_msg = f"Request error: {str(e)}"
-        logger.warning(f"Failed to fetch Gemini models from {models_url}: {e}")
-        return [], error_msg
+    return None
 
 
 @router.post("/models")
@@ -180,10 +70,10 @@ async def query_available_models(
     """
     查询提供商可用模型
 
-    遍历所有活跃端点，根据端点的 API 格式选择正确的请求方式：
-    - OPENAI/OPENAI_CLI: /v1/models (Bearer token)
-    - CLAUDE/CLAUDE_CLI: /v1/models (x-api-key)
-    - GEMINI/GEMINI_CLI: /v1beta/models (URL key parameter)
+    遍历所有活跃端点，根据端点的 API 格式选择正确的 Adapter 进行请求：
+    - OPENAI/OPENAI_CLI: 使用 OpenAIChatAdapter.fetch_models
+    - CLAUDE/CLAUDE_CLI: 使用 ClaudeChatAdapter.fetch_models
+    - GEMINI/GEMINI_CLI: 使用 GeminiChatAdapter.fetch_models
 
     Args:
         request: 查询请求
@@ -265,37 +155,53 @@ async def query_available_models(
         base_url = base_url.rstrip("/")
         api_format = config["api_format"]
         api_key_value = config["api_key"]
-        extra_headers = config["extra_headers"]
+        extra_headers = config.get("extra_headers")
 
         try:
-            if api_format in ["CLAUDE", "CLAUDE_CLI"]:
-                return await _fetch_claude_models(client, base_url, api_key_value, api_format)
-            elif api_format in ["GEMINI", "GEMINI_CLI"]:
-                return await _fetch_gemini_models(client, base_url, api_key_value, api_format)
-            else:
-                return await _fetch_openai_models(
-                    client, base_url, api_key_value, api_format, extra_headers
-                )
+            # 获取对应的 Adapter 类并调用 fetch_models
+            adapter_class = _get_adapter_for_format(api_format)
+            if not adapter_class:
+                return [], f"Unknown API format: {api_format}"
+            models, error = await adapter_class.fetch_models(
+                client, base_url, api_key_value, extra_headers
+            )
+            # 确保所有模型都有 api_format 字段
+            for m in models:
+                if "api_format" not in m:
+                    m["api_format"] = api_format
+            return models, error
         except Exception as e:
             logger.error(f"Error fetching models from {api_format} endpoint: {e}")
             return [], f"{api_format}: {str(e)}"
 
+    # 限制并发请求数量，避免触发上游速率限制
+    MAX_CONCURRENT_REQUESTS = 5
+    semaphore = asyncio.Semaphore(MAX_CONCURRENT_REQUESTS)
+
+    async def fetch_with_semaphore(
+        client: httpx.AsyncClient, config: dict
+    ) -> tuple[list, Optional[str]]:
+        async with semaphore:
+            return await fetch_endpoint_models(client, config)
+
     async with httpx.AsyncClient(timeout=30.0) as client:
         results = await asyncio.gather(
-            *[fetch_endpoint_models(client, c) for c in endpoint_configs]
+            *[fetch_with_semaphore(client, c) for c in endpoint_configs]
         )
         for models, error in results:
             all_models.extend(models)
             if error:
                 errors.append(error)
 
-    # 按 model id 去重（保留第一个）
-    seen_ids: set[str] = set()
+    # 按 model id + api_format 去重（保留第一个）
+    seen_keys: set[str] = set()
     unique_models: list = []
     for model in all_models:
         model_id = model.get("id")
-        if model_id and model_id not in seen_ids:
-            seen_ids.add(model_id)
+        api_format = model.get("api_format", "")
+        unique_key = f"{model_id}:{api_format}"
+        if model_id and unique_key not in seen_keys:
+            seen_keys.add(unique_key)
             unique_models.append(model)
 
     error = "; ".join(errors) if errors else None
@@ -311,3 +217,228 @@ async def query_available_models(
             "display_name": provider.display_name,
         },
     }
+
+
+@router.post("/test-model")
+async def test_model(
+    request: TestModelRequest,
+    db: Session = Depends(get_db),
+    current_user: User = Depends(get_current_user),
+):
+    """
+    测试模型连接性
+
+    向指定提供商的指定模型发送测试请求，验证模型是否可用
+
+    Args:
+        request: 测试请求
+
+    Returns:
+        测试结果
+    """
+    # 获取提供商及其端点
+    provider = (
+        db.query(Provider)
+        .options(joinedload(Provider.endpoints).joinedload(ProviderEndpoint.api_keys))
+        .filter(Provider.id == request.provider_id)
+        .first()
+    )
+
+    if not provider:
+        raise HTTPException(status_code=404, detail="Provider not found")
+
+    # 找到合适的端点和API Key
+    endpoint_config = None
+    endpoint = None
+    api_key = None
+
+    if request.api_key_id:
+        # 使用指定的API Key
+        for ep in provider.endpoints:
+            for key in ep.api_keys:
+                if key.id == request.api_key_id and key.is_active and ep.is_active:
+                    endpoint = ep
+                    api_key = key
+                    break
+            if endpoint:
+                break
+    else:
+        # 使用第一个可用的端点和密钥
+        for ep in provider.endpoints:
+            if not ep.is_active or not ep.api_keys:
+                continue
+            for key in ep.api_keys:
+                if key.is_active:
+                    endpoint = ep
+                    api_key = key
+                    break
+            if endpoint:
+                break
+
+    if not endpoint or not api_key:
+        raise HTTPException(status_code=404, detail="No active endpoint or API key found")
+
+    try:
+        api_key_value = crypto_service.decrypt(api_key.api_key)
+    except Exception as e:
+        logger.error(f"[test-model] Failed to decrypt API key: {e}")
+        raise HTTPException(status_code=500, detail="Failed to decrypt API key")
+
+    # 构建请求配置
+    endpoint_config = {
+        "api_key": api_key_value,
+        "api_key_id": api_key.id,  # 添加API Key ID用于用量记录
+        "base_url": endpoint.base_url,
+        "api_format": endpoint.api_format,
+        "extra_headers": endpoint.headers,
+        "timeout": endpoint.timeout or 30.0,
+    }
+
+    try:
+        # 获取对应的 Adapter 类
+        adapter_class = _get_adapter_for_format(endpoint.api_format)
+        if not adapter_class:
+            return {
+                "success": False,
+                "error": f"Unknown API format: {endpoint.api_format}",
+                "provider": {
+                    "id": provider.id,
+                    "name": provider.name,
+                    "display_name": provider.display_name,
+                },
+                "model": request.model_name,
+            }
+
+        logger.debug(f"[test-model] 使用 Adapter: {adapter_class.__name__}")
+        logger.debug(f"[test-model] 端点 API Format: {endpoint.api_format}")
+
+        # 如果请求指定了 api_format，优先使用它
+        target_api_format = request.api_format or endpoint.api_format
+        if request.api_format and request.api_format != endpoint.api_format:
+            logger.debug(f"[test-model] 请求指定 API Format: {request.api_format}")
+            # 重新获取适配器
+            adapter_class = _get_adapter_for_format(request.api_format)
+            if not adapter_class:
+                return {
+                    "success": False,
+                    "error": f"Unknown API format: {request.api_format}",
+                    "provider": {
+                        "id": provider.id,
+                        "name": provider.name,
+                        "display_name": provider.display_name,
+                    },
+                    "model": request.model_name,
+                }
+            logger.debug(f"[test-model] 重新选择 Adapter: {adapter_class.__name__}")
+
+        # 准备测试请求数据
+        check_request = {
+            "model": request.model_name,
+            "messages": [
+                {"role": "user", "content": request.message or "Hello! This is a test message."}
+            ],
+            "max_tokens": 30,
+            "temperature": 0.7,
+        }
+
+        # 发送测试请求
+        async with httpx.AsyncClient(timeout=endpoint_config["timeout"]) as client:
+            # 非流式测试
+            logger.debug(f"[test-model] 开始非流式测试...")
+
+            response = await adapter_class.check_endpoint(
+                client,
+                endpoint_config["base_url"],
+                endpoint_config["api_key"],
+                check_request,
+                endpoint_config.get("extra_headers"),
+                # 用量计算参数（现在强制记录）
+                db=db,
+                user=current_user,
+                provider_name=provider.name,
+                provider_id=provider.id,
+                api_key_id=endpoint_config.get("api_key_id"),
+                model_name=request.model_name,
+            )
+
+            # 记录提供商返回信息
+            logger.debug(f"[test-model] 非流式测试结果:")
+            logger.debug(f"[test-model] Status Code: {response.get('status_code')}")
+            logger.debug(f"[test-model] Response Headers: {response.get('headers', {})}")
+            response_data = response.get('response', {})
+            response_body = response_data.get('response_body', {})
+            logger.debug(f"[test-model] Response Data: {response_data}")
+            logger.debug(f"[test-model] Response Body: {response_body}")
+            # 尝试解析 response_body (通常是 JSON 字符串)
+            parsed_body = response_body
+            import json
+            if isinstance(response_body, str):
+                try:
+                    parsed_body = json.loads(response_body)
+                except json.JSONDecodeError:
+                    pass
+
+            if isinstance(parsed_body, dict) and 'error' in parsed_body:
+                error_obj = parsed_body['error']
+                # 兼容 error 可能是字典或字符串的情况
+                if isinstance(error_obj, dict):
+                    logger.debug(f"[test-model] Error Message: {error_obj.get('message')}")
+                    raise HTTPException(status_code=500, detail=error_obj.get('message'))
+                else:
+                    logger.debug(f"[test-model] Error: {error_obj}")
+                    raise HTTPException(status_code=500, detail=error_obj)
+            elif 'error' in response:
+                logger.debug(f"[test-model] Error: {response['error']}")
+                raise HTTPException(status_code=500, detail=response['error'])
+            else:
+                # 如果有选择或消息，记录内容预览
+                if isinstance(response_data, dict):
+                    if 'choices' in response_data and response_data['choices']:
+                        choice = response_data['choices'][0]
+                        if 'message' in choice:
+                            content = choice['message'].get('content', '')
+                            logger.debug(f"[test-model] Content Preview: {content[:200]}...")
+                    elif 'content' in response_data and response_data['content']:
+                        content = str(response_data['content'])
+                        logger.debug(f"[test-model] Content Preview: {content[:200]}...")
+
+            # 检查测试是否成功（基于HTTP状态码）
+            status_code = response.get('status_code', 0)
+            is_success = status_code == 200 and 'error' not in response
+
+            return {
+                "success": is_success,
+                "data": {
+                    "stream": False,
+                    "response": response,
+                },
+                "provider": {
+                    "id": provider.id,
+                    "name": provider.name,
+                    "display_name": provider.display_name,
+                },
+                "model": request.model_name,
+                "endpoint": {
+                    "id": endpoint.id,
+                    "api_format": endpoint.api_format,
+                    "base_url": endpoint.base_url,
+                },
+            }
+
+    except Exception as e:
+        logger.error(f"[test-model] Error testing model {request.model_name}: {e}")
+        return {
+            "success": False,
+            "error": str(e),
+            "provider": {
+                "id": provider.id,
+                "name": provider.name,
+                "display_name": provider.display_name,
+            },
+            "model": request.model_name,
+            "endpoint": {
+                "id": endpoint.id,
+                "api_format": endpoint.api_format,
+                "base_url": endpoint.base_url,
+            } if endpoint else None,
+        }

src/api/admin/providers/models.py

@@ -22,16 +22,18 @@ from src.models.api import (
 from src.models.pydantic_models import (
     BatchAssignModelsToProviderRequest,
     BatchAssignModelsToProviderResponse,
+    ImportFromUpstreamRequest,
+    ImportFromUpstreamResponse,
+    ImportFromUpstreamSuccessItem,
+    ImportFromUpstreamErrorItem,
+    ProviderAvailableSourceModel,
+    ProviderAvailableSourceModelsResponse,
 )
 from src.models.database import (
     GlobalModel,
     Model,
     Provider,
 )
-from src.models.pydantic_models import (
-    ProviderAvailableSourceModel,
-    ProviderAvailableSourceModelsResponse,
-)
 from src.services.model.service import ModelService
 
 router = APIRouter(tags=["Model Management"])
@@ -158,6 +160,28 @@ async def batch_assign_global_models_to_provider(
     return await pipeline.run(adapter=adapter, http_request=request, db=db, mode=adapter.mode)
 
 
+@router.post(
+    "/{provider_id}/import-from-upstream",
+    response_model=ImportFromUpstreamResponse,
+)
+async def import_models_from_upstream(
+    provider_id: str,
+    payload: ImportFromUpstreamRequest,
+    request: Request,
+    db: Session = Depends(get_db),
+) -> ImportFromUpstreamResponse:
+    """
+    从上游提供商导入模型
+
+    流程：
+    1. 根据 model_ids 检查全局模型是否存在（按 name 匹配）
+    2. 如不存在，自动创建新的 GlobalModel（使用默认配置）
+    3. 创建 Model 关联到当前 Provider
+    """
+    adapter = AdminImportFromUpstreamAdapter(provider_id=provider_id, payload=payload)
+    return await pipeline.run(adapter=adapter, http_request=request, db=db, mode=adapter.mode)
+
+
 # -------- Adapters --------
 
 
@@ -425,3 +449,130 @@ class AdminBatchAssignModelsToProviderAdapter(AdminApiAdapter):
             await invalidate_models_list_cache()
 
         return BatchAssignModelsToProviderResponse(success=success, errors=errors)
+
+
+@dataclass
+class AdminImportFromUpstreamAdapter(AdminApiAdapter):
+    """从上游提供商导入模型"""
+
+    provider_id: str
+    payload: ImportFromUpstreamRequest
+
+    async def handle(self, context):  # type: ignore[override]
+        db = context.db
+        provider = db.query(Provider).filter(Provider.id == self.provider_id).first()
+        if not provider:
+            raise NotFoundException("Provider not found", "provider")
+
+        success: list[ImportFromUpstreamSuccessItem] = []
+        errors: list[ImportFromUpstreamErrorItem] = []
+
+        # 默认阶梯计费配置（免费）
+        default_tiered_pricing = {
+            "tiers": [
+                {
+                    "up_to": None,
+                    "input_price_per_1m": 0.0,
+                    "output_price_per_1m": 0.0,
+                }
+            ]
+        }
+
+        for model_id in self.payload.model_ids:
+            # 输入验证：检查 model_id 长度
+            if not model_id or len(model_id) > 100:
+                errors.append(
+                    ImportFromUpstreamErrorItem(
+                        model_id=model_id[:50] + "..." if model_id and len(model_id) > 50 else model_id or "<empty>",
+                        error="Invalid model_id: must be 1-100 characters",
+                    )
+                )
+                continue
+
+            try:
+                # 使用 savepoint 确保单个模型导入的原子性
+                savepoint = db.begin_nested()
+                try:
+                    # 1. 检查是否已存在同名的 GlobalModel
+                    global_model = (
+                        db.query(GlobalModel).filter(GlobalModel.name == model_id).first()
+                    )
+                    created_global_model = False
+
+                    if not global_model:
+                        # 2. 创建新的 GlobalModel
+                        global_model = GlobalModel(
+                            name=model_id,
+                            display_name=model_id,
+                            default_tiered_pricing=default_tiered_pricing,
+                            is_active=True,
+                        )
+                        db.add(global_model)
+                        db.flush()
+                        created_global_model = True
+                        logger.info(
+                            f"Created new GlobalModel: {model_id} during upstream import"
+                        )
+
+                    # 3. 检查是否已存在关联
+                    existing = (
+                        db.query(Model)
+                        .filter(
+                            Model.provider_id == self.provider_id,
+                            Model.global_model_id == global_model.id,
+                        )
+                        .first()
+                    )
+                    if existing:
+                        # 已存在关联，提交 savepoint 并记录成功
+                        savepoint.commit()
+                        success.append(
+                            ImportFromUpstreamSuccessItem(
+                                model_id=model_id,
+                                global_model_id=global_model.id,
+                                global_model_name=global_model.name,
+                                provider_model_id=existing.id,
+                                created_global_model=created_global_model,
+                            )
+                        )
+                        continue
+
+                    # 4. 创建新的 Model 记录
+                    new_model = Model(
+                        provider_id=self.provider_id,
+                        global_model_id=global_model.id,
+                        provider_model_name=global_model.name,
+                        is_active=True,
+                    )
+                    db.add(new_model)
+                    db.flush()
+
+                    # 提交 savepoint
+                    savepoint.commit()
+                    success.append(
+                        ImportFromUpstreamSuccessItem(
+                            model_id=model_id,
+                            global_model_id=global_model.id,
+                            global_model_name=global_model.name,
+                            provider_model_id=new_model.id,
+                            created_global_model=created_global_model,
+                        )
+                    )
+                except Exception as e:
+                    # 回滚到 savepoint
+                    savepoint.rollback()
+                    raise e
+            except Exception as e:
+                logger.error(f"Error importing model {model_id}: {e}")
+                errors.append(ImportFromUpstreamErrorItem(model_id=model_id, error=str(e)))
+
+        db.commit()
+        logger.info(
+            f"Imported {len(success)} models from upstream to provider {provider.name} by {context.user.username}"
+        )
+
+        # 清除 /v1/models 列表缓存
+        if success:
+            await invalidate_models_list_cache()
+
+        return ImportFromUpstreamResponse(success=success, errors=errors)

src/api/admin/system.py

@@ -436,7 +436,7 @@ class AdminExportConfigAdapter(AdminApiAdapter):
                     {
                         "global_model_name": global_model.name if global_model else None,
                         "provider_model_name": model.provider_model_name,
-                        "provider_model_aliases": model.provider_model_aliases,
+                        "provider_model_mappings": model.provider_model_mappings,
                         "price_per_request": model.price_per_request,
                         "tiered_pricing": model.tiered_pricing,
                         "supports_vision": model.supports_vision,
@@ -790,8 +790,8 @@ class AdminImportConfigAdapter(AdminApiAdapter):
                             )
                         elif merge_mode == "overwrite":
                             existing_model.global_model_id = global_model_id
-                            existing_model.provider_model_aliases = model_data.get(
-                                "provider_model_aliases"
+                            existing_model.provider_model_mappings = model_data.get(
+                                "provider_model_mappings"
                             )
                             existing_model.price_per_request = model_data.get(
                                 "price_per_request"
@@ -824,8 +824,8 @@ class AdminImportConfigAdapter(AdminApiAdapter):
                             provider_id=provider_id,
                             global_model_id=global_model_id,
                             provider_model_name=model_data["provider_model_name"],
-                            provider_model_aliases=model_data.get(
-                                "provider_model_aliases"
+                            provider_model_mappings=model_data.get(
+                                "provider_model_mappings"
                             ),
                             price_per_request=model_data.get("price_per_request"),
                             tiered_pricing=model_data.get("tiered_pricing"),

src/api/dashboard/routes.py

@@ -13,7 +13,7 @@ from src.api.base.admin_adapter import AdminApiAdapter
 from src.api.base.pipeline import ApiRequestPipeline
 from src.core.enums import UserRole
 from src.database import get_db
-from src.models.database import ApiKey, Provider, RequestCandidate, StatsDaily, Usage
+from src.models.database import ApiKey, Provider, RequestCandidate, StatsDaily, StatsDailyModel, Usage
 from src.models.database import User as DBUser
 from src.services.system.stats_aggregator import StatsAggregatorService
 from src.utils.cache_decorator import cache_result
@@ -893,69 +893,172 @@ class DashboardDailyStatsAdapter(DashboardAdapter):
                 })
             current_date += timedelta(days=1)
 
-        # ==================== 模型统计（仍需实时查询）====================
-        model_query = db.query(Usage)
-        if not is_admin:
-            model_query = model_query.filter(Usage.user_id == user.id)
-        model_query = model_query.filter(
-            and_(Usage.created_at >= start_date, Usage.created_at <= end_date)
-        )
-
-        model_stats = (
-            model_query.with_entities(
-                Usage.model,
-                func.count(Usage.id).label("requests"),
-                func.sum(Usage.total_tokens).label("tokens"),
-                func.sum(Usage.total_cost_usd).label("cost"),
-                func.avg(Usage.response_time_ms).label("avg_response_time"),
+        # ==================== 模型统计 ====================
+        if is_admin:
+            # 管理员：使用预聚合数据 + 今日实时数据
+            # 历史数据从 stats_daily_model 获取
+            historical_model_stats = (
+                db.query(StatsDailyModel)
+                .filter(and_(StatsDailyModel.date >= start_date, StatsDailyModel.date < today))
+                .all()
             )
-            .group_by(Usage.model)
-            .order_by(func.sum(Usage.total_cost_usd).desc())
-            .all()
-        )
 
-        model_summary = [
-            {
-                "model": stat.model,
-                "requests": stat.requests or 0,
-                "tokens": int(stat.tokens or 0),
-                "cost": float(stat.cost or 0),
-                "avg_response_time": (
-                    float(stat.avg_response_time or 0) / 1000.0 if stat.avg_response_time else 0
-                ),
-                "cost_per_request": float(stat.cost or 0) / max(stat.requests or 1, 1),
-                "tokens_per_request": int(stat.tokens or 0) / max(stat.requests or 1, 1),
-            }
-            for stat in model_stats
-        ]
+            # 按模型汇总历史数据
+            model_agg: dict = {}
+            daily_breakdown: dict = {}
 
-        daily_model_stats = (
-            model_query.with_entities(
-                func.date(Usage.created_at).label("date"),
-                Usage.model,
-                func.count(Usage.id).label("requests"),
-                func.sum(Usage.total_tokens).label("tokens"),
-                func.sum(Usage.total_cost_usd).label("cost"),
+            for stat in historical_model_stats:
+                model = stat.model
+                if model not in model_agg:
+                    model_agg[model] = {
+                        "requests": 0, "tokens": 0, "cost": 0.0,
+                        "total_response_time": 0.0, "response_count": 0
+                    }
+                model_agg[model]["requests"] += stat.total_requests
+                tokens = (stat.input_tokens + stat.output_tokens +
+                          stat.cache_creation_tokens + stat.cache_read_tokens)
+                model_agg[model]["tokens"] += tokens
+                model_agg[model]["cost"] += stat.total_cost
+                if stat.avg_response_time_ms is not None:
+                    model_agg[model]["total_response_time"] += stat.avg_response_time_ms * stat.total_requests
+                    model_agg[model]["response_count"] += stat.total_requests
+
+                # 按日期分组
+                if stat.date.tzinfo is None:
+                    date_utc = stat.date.replace(tzinfo=timezone.utc)
+                else:
+                    date_utc = stat.date.astimezone(timezone.utc)
+                date_str = date_utc.astimezone(app_tz).date().isoformat()
+
+                daily_breakdown.setdefault(date_str, []).append({
+                    "model": model,
+                    "requests": stat.total_requests,
+                    "tokens": tokens,
+                    "cost": stat.total_cost,
+                })
+
+            # 今日实时模型统计
+            today_model_stats = (
+                db.query(
+                    Usage.model,
+                    func.count(Usage.id).label("requests"),
+                    func.sum(Usage.total_tokens).label("tokens"),
+                    func.sum(Usage.total_cost_usd).label("cost"),
+                    func.avg(Usage.response_time_ms).label("avg_response_time"),
+                )
+                .filter(Usage.created_at >= today)
+                .group_by(Usage.model)
+                .all()
             )
-            .group_by(func.date(Usage.created_at), Usage.model)
-            .order_by(func.date(Usage.created_at).desc(), func.sum(Usage.total_cost_usd).desc())
-            .all()
-        )
 
-        breakdown = {}
-        for stat in daily_model_stats:
-            date_str = stat.date.isoformat()
-            breakdown.setdefault(date_str, []).append(
+            today_str = today_local.date().isoformat()
+            for stat in today_model_stats:
+                model = stat.model
+                if model not in model_agg:
+                    model_agg[model] = {
+                        "requests": 0, "tokens": 0, "cost": 0.0,
+                        "total_response_time": 0.0, "response_count": 0
+                    }
+                model_agg[model]["requests"] += stat.requests or 0
+                model_agg[model]["tokens"] += int(stat.tokens or 0)
+                model_agg[model]["cost"] += float(stat.cost or 0)
+                if stat.avg_response_time is not None:
+                    model_agg[model]["total_response_time"] += float(stat.avg_response_time) * (stat.requests or 0)
+                    model_agg[model]["response_count"] += stat.requests or 0
+
+                # 今日 breakdown
+                daily_breakdown.setdefault(today_str, []).append({
+                    "model": model,
+                    "requests": stat.requests or 0,
+                    "tokens": int(stat.tokens or 0),
+                    "cost": float(stat.cost or 0),
+                })
+
+            # 构建 model_summary
+            model_summary = []
+            for model, agg in model_agg.items():
+                avg_rt = (agg["total_response_time"] / agg["response_count"] / 1000.0
+                          if agg["response_count"] > 0 else 0)
+                model_summary.append({
+                    "model": model,
+                    "requests": agg["requests"],
+                    "tokens": agg["tokens"],
+                    "cost": agg["cost"],
+                    "avg_response_time": avg_rt,
+                    "cost_per_request": agg["cost"] / max(agg["requests"], 1),
+                    "tokens_per_request": agg["tokens"] / max(agg["requests"], 1),
+                })
+            model_summary.sort(key=lambda x: x["cost"], reverse=True)
+
+            # 填充 model_breakdown
+            for item in formatted:
+                item["model_breakdown"] = daily_breakdown.get(item["date"], [])
+
+        else:
+            # 普通用户：实时查询（数据量较小）
+            model_query = db.query(Usage).filter(
+                and_(
+                    Usage.user_id == user.id,
+                    Usage.created_at >= start_date,
+                    Usage.created_at <= end_date
+                )
+            )
+
+            model_stats = (
+                model_query.with_entities(
+                    Usage.model,
+                    func.count(Usage.id).label("requests"),
+                    func.sum(Usage.total_tokens).label("tokens"),
+                    func.sum(Usage.total_cost_usd).label("cost"),
+                    func.avg(Usage.response_time_ms).label("avg_response_time"),
+                )
+                .group_by(Usage.model)
+                .order_by(func.sum(Usage.total_cost_usd).desc())
+                .all()
+            )
+
+            model_summary = [
                 {
                     "model": stat.model,
                     "requests": stat.requests or 0,
                     "tokens": int(stat.tokens or 0),
                     "cost": float(stat.cost or 0),
+                    "avg_response_time": (
+                        float(stat.avg_response_time or 0) / 1000.0 if stat.avg_response_time else 0
+                    ),
+                    "cost_per_request": float(stat.cost or 0) / max(stat.requests or 1, 1),
+                    "tokens_per_request": int(stat.tokens or 0) / max(stat.requests or 1, 1),
                 }
+                for stat in model_stats
+            ]
+
+            daily_model_stats = (
+                model_query.with_entities(
+                    func.date(Usage.created_at).label("date"),
+                    Usage.model,
+                    func.count(Usage.id).label("requests"),
+                    func.sum(Usage.total_tokens).label("tokens"),
+                    func.sum(Usage.total_cost_usd).label("cost"),
+                )
+                .group_by(func.date(Usage.created_at), Usage.model)
+                .order_by(func.date(Usage.created_at).desc(), func.sum(Usage.total_cost_usd).desc())
+                .all()
             )
 
-        for item in formatted:
-            item["model_breakdown"] = breakdown.get(item["date"], [])
+            breakdown = {}
+            for stat in daily_model_stats:
+                date_str = stat.date.isoformat()
+                breakdown.setdefault(date_str, []).append(
+                    {
+                        "model": stat.model,
+                        "requests": stat.requests or 0,
+                        "tokens": int(stat.tokens or 0),
+                        "cost": float(stat.cost or 0),
+                    }
+                )
+
+            for item in formatted:
+                item["model_breakdown"] = breakdown.get(item["date"], [])
 
         return {
             "daily_stats": formatted,

src/api/handlers/base/base_handler.py

@@ -376,6 +376,9 @@ class BaseMessageHandler:
 
         使用 asyncio 后台任务执行数据库更新，避免阻塞流式传输
 
+        注意：TTFB（首字节时间）由 StreamContext.record_first_byte_time() 记录，
+        并在最终 record_success 时传递到数据库，避免重复记录导致数据不一致。
+
         Args:
             request_id: 请求 ID，如果不传则使用 self.request_id
         """
@@ -407,6 +410,9 @@ class BaseMessageHandler:
 
         使用 asyncio 后台任务执行数据库更新，避免阻塞流式传输
 
+        注意：TTFB（首字节时间）由 StreamContext.record_first_byte_time() 记录，
+        并在最终 record_success 时传递到数据库，避免重复记录导致数据不一致。
+
         Args:
             ctx: 流式上下文，包含 provider_name 和 mapped_model
         """

src/api/handlers/base/chat_adapter_base.py

@@ -19,8 +19,9 @@ Chat Adapter 通用基类
 import time
 import traceback
 from abc import abstractmethod
-from typing import Any, Dict, Optional, Type
+from typing import Any, Dict, Optional, Tuple, Type
 
+import httpx
 from fastapi import HTTPException
 from fastapi.responses import JSONResponse
 
@@ -62,6 +63,34 @@ class ChatAdapterBase(ApiAdapter):
     name: str = "chat.base"
     mode = ApiMode.STANDARD
 
+    # 子类可以配置的特殊方法（用于check_endpoint）
+    @classmethod
+    def build_endpoint_url(cls, base_url: str) -> str:
+        """构建端点URL，子类可以覆盖以自定义URL构建逻辑"""
+        # 默认实现：在base_url后添加特定路径
+        return base_url
+
+    @classmethod
+    def build_base_headers(cls, api_key: str) -> Dict[str, str]:
+        """构建基础请求头，子类可以覆盖以自定义认证头"""
+        # 默认实现：Bearer token认证
+        return {
+            "Authorization": f"Bearer {api_key}",
+            "Content-Type": "application/json",
+        }
+
+    @classmethod
+    def get_protected_header_keys(cls) -> tuple:
+        """返回不应被extra_headers覆盖的头部key，子类可以覆盖"""
+        # 默认保护认证相关头部
+        return ("authorization", "content-type")
+
+    @classmethod
+    def build_request_body(cls, request_data: Dict[str, Any]) -> Dict[str, Any]:
+        """构建请求体，子类可以覆盖以自定义请求格式转换"""
+        # 默认实现：直接使用请求数据
+        return request_data.copy()
+
     def __init__(self, allowed_api_formats: Optional[list[str]] = None):
         self.allowed_api_formats = allowed_api_formats or [self.FORMAT_ID]
 
@@ -620,6 +649,98 @@ class ChatAdapterBase(ApiAdapter):
         # 如果所有阶梯都有上限且都超过了，返回最后一个阶梯
         return tiers[-1] if tiers else None
 
+    # =========================================================================
+    # 模型列表查询 - 子类应覆盖此方法
+    # =========================================================================
+
+    @classmethod
+    async def fetch_models(
+        cls,
+        client: httpx.AsyncClient,
+        base_url: str,
+        api_key: str,
+        extra_headers: Optional[Dict[str, str]] = None,
+    ) -> Tuple[list, Optional[str]]:
+        """
+        查询上游 API 支持的模型列表
+
+        这是 Aether 内部发起的请求（非用户透传），用于：
+        - 管理后台查询提供商支持的模型
+        - 自动发现可用模型
+
+        Args:
+            client: httpx 异步客户端
+            base_url: API 基础 URL
+            api_key: API 密钥（已解密）
+            extra_headers: 端点配置的额外请求头
+
+        Returns:
+            (models, error): 模型列表和错误信息
+            - models: 模型信息列表，每个模型至少包含 id 字段
+            - error: 错误信息，成功时为 None
+        """
+        # 默认实现返回空列表，子类应覆盖
+        return [], f"{cls.FORMAT_ID} adapter does not implement fetch_models"
+
+    @classmethod
+    async def check_endpoint(
+        cls,
+        client: httpx.AsyncClient,
+        base_url: str,
+        api_key: str,
+        request_data: Dict[str, Any],
+        extra_headers: Optional[Dict[str, str]] = None,
+        # 用量计算参数（现在强制记录）
+        db: Optional[Any] = None,
+        user: Optional[Any] = None,
+        provider_name: Optional[str] = None,
+        provider_id: Optional[str] = None,
+        api_key_id: Optional[str] = None,
+        model_name: Optional[str] = None,
+    ) -> Dict[str, Any]:
+        """
+        测试模型连接性（非流式）
+
+        Args:
+            client: httpx 异步客户端
+            base_url: API 基础 URL
+            api_key: API 密钥（已解密）
+            request_data: 请求数据
+            extra_headers: 端点配置的额外请求头
+            db: 数据库会话
+            user: 用户对象
+            provider_name: 提供商名称
+            provider_id: 提供商ID
+            api_key_id: API Key ID
+            model_name: 模型名称
+
+        Returns:
+            测试响应数据
+        """
+        from src.api.handlers.base.endpoint_checker import build_safe_headers, run_endpoint_check
+
+        # 使用子类配置方法构建请求组件
+        url = cls.build_endpoint_url(base_url)
+        base_headers = cls.build_base_headers(api_key)
+        protected_keys = cls.get_protected_header_keys()
+        headers = build_safe_headers(base_headers, extra_headers, protected_keys)
+        body = cls.build_request_body(request_data)
+
+        # 使用通用的endpoint checker执行请求
+        return await run_endpoint_check(
+            client=client,
+            url=url,
+            headers=headers,
+            json_body=body,
+            api_format=cls.name,
+            # 用量计算参数（现在强制记录）
+            db=db,
+            user=user,
+            provider_name=provider_name,
+            provider_id=provider_id,
+            api_key_id=api_key_id,
+            model_name=model_name or request_data.get("model"),
+        )
 
 # =========================================================================
 # Adapter 注册表 - 用于根据 API format 获取 Adapter 实例

src/api/handlers/base/chat_handler_base.py

@@ -260,9 +260,9 @@ class ChatHandlerBase(BaseMessageHandler, ABC):
         mapping = await mapper.get_mapping(source_model, provider_id)
 
         if mapping and mapping.model:
-            # 使用 select_provider_model_name 支持别名功能
-            # 传入 api_key.id 作为 affinity_key，实现相同用户稳定选择同一别名
-            # 传入 api_format 用于过滤适用的别名作用域
+            # 使用 select_provider_model_name 支持映射功能
+            # 传入 api_key.id 作为 affinity_key，实现相同用户稳定选择同一映射
+            # 传入 api_format 用于过滤适用的映射作用域
             affinity_key = self.api_key.id if self.api_key else None
             mapped_name = mapping.model.select_provider_model_name(
                 affinity_key, api_format=self.FORMAT_ID
@@ -484,9 +484,8 @@ class ChatHandlerBase(BaseMessageHandler, ABC):
 
             stream_response.raise_for_status()
 
-            # 使用字节流迭代器（避免 aiter_lines 的性能问题）
-            # aiter_raw() 返回原始数据块，无缓冲，实现真正的流式传输
-            byte_iterator = stream_response.aiter_raw()
+            # 使用字节流迭代器（避免 aiter_lines 的性能问题, aiter_bytes 会自动解压 gzip/deflate）
+            byte_iterator = stream_response.aiter_bytes()
 
             # 预读检测嵌套错误
             prefetched_chunks = await stream_processor.prefetch_and_check_error(
@@ -639,6 +638,8 @@ class ChatHandlerBase(BaseMessageHandler, ABC):
 
             logger.info(f"  [{self.request_id}] 发送非流式请求: Provider={provider.name}, "
                 f"模型={model} -> {mapped_model or '无映射'}")
+            logger.debug(f"  [{self.request_id}] 请求URL: {url}")
+            logger.debug(f"  [{self.request_id}] 请求体stream字段: {provider_payload.get('stream', 'N/A')}")
 
             # 创建 HTTP 客户端（支持代理配置）
             from src.clients.http_client import HTTPClientPool
@@ -662,10 +663,32 @@ class ChatHandlerBase(BaseMessageHandler, ABC):
                         response_headers=response_headers,
                     )
                 elif resp.status_code >= 500:
-                    raise ProviderNotAvailableException(f"提供商服务不可用: {provider.name}")
-                elif resp.status_code != 200:
+                    # 记录响应体以便调试
+                    error_body = ""
+                    try:
+                        error_body = resp.text[:1000]
+                        logger.error(f"  [{self.request_id}] 上游返回5xx错误: status={resp.status_code}, body={error_body[:500]}")
+                    except Exception:
+                        pass
                     raise ProviderNotAvailableException(
-                        f"提供商返回错误: {provider.name}, 状态: {resp.status_code}"
+                        f"提供商服务不可用: {provider.name}",
+                        provider_name=str(provider.name),
+                        upstream_status=resp.status_code,
+                        upstream_response=error_body,
+                    )
+                elif resp.status_code != 200:
+                    # 记录非200响应以便调试
+                    error_body = ""
+                    try:
+                        error_body = resp.text[:1000]
+                        logger.warning(f"  [{self.request_id}] 上游返回非200: status={resp.status_code}, body={error_body[:500]}")
+                    except Exception:
+                        pass
+                    raise ProviderNotAvailableException(
+                        f"提供商返回错误: {provider.name}, 状态: {resp.status_code}",
+                        provider_name=str(provider.name),
+                        upstream_status=resp.status_code,
+                        upstream_response=error_body,
                     )
 
                 response_json = resp.json()

src/api/handlers/base/cli_adapter_base.py

@@ -17,8 +17,9 @@ CLI Adapter 通用基类
 
 import time
 import traceback
-from typing import Any, Dict, Optional, Type
+from typing import Any, Dict, Optional, Tuple, Type
 
+import httpx
 from fastapi import HTTPException
 from fastapi.responses import JSONResponse
 
@@ -580,6 +581,179 @@ class CliAdapterBase(ApiAdapter):
 
         return tiers[-1] if tiers else None
 
+    # =========================================================================
+    # 模型列表查询 - 子类应覆盖此方法
+    # =========================================================================
+
+    @classmethod
+    async def fetch_models(
+        cls,
+        client: httpx.AsyncClient,
+        base_url: str,
+        api_key: str,
+        extra_headers: Optional[Dict[str, str]] = None,
+    ) -> Tuple[list, Optional[str]]:
+        """
+        查询上游 API 支持的模型列表
+
+        这是 Aether 内部发起的请求（非用户透传），用于：
+        - 管理后台查询提供商支持的模型
+        - 自动发现可用模型
+
+        Args:
+            client: httpx 异步客户端
+            base_url: API 基础 URL
+            api_key: API 密钥（已解密）
+            extra_headers: 端点配置的额外请求头
+
+        Returns:
+            (models, error): 模型列表和错误信息
+            - models: 模型信息列表，每个模型至少包含 id 字段
+            - error: 错误信息，成功时为 None
+        """
+        # 默认实现返回空列表，子类应覆盖
+        return [], f"{cls.FORMAT_ID} adapter does not implement fetch_models"
+
+    @classmethod
+    async def check_endpoint(
+        cls,
+        client: httpx.AsyncClient,
+        base_url: str,
+        api_key: str,
+        request_data: Dict[str, Any],
+        extra_headers: Optional[Dict[str, str]] = None,
+        # 用量计算参数
+        db: Optional[Any] = None,
+        user: Optional[Any] = None,
+        provider_name: Optional[str] = None,
+        provider_id: Optional[str] = None,
+        api_key_id: Optional[str] = None,
+        model_name: Optional[str] = None,
+    ) -> Dict[str, Any]:
+        """
+        测试模型连接性（非流式）
+
+        通用的CLI endpoint测试方法，使用配置方法模式：
+        - build_endpoint_url(): 构建请求URL
+        - build_base_headers(): 构建基础认证头
+        - get_protected_header_keys(): 获取受保护的头部key
+        - build_request_body(): 构建请求体
+        - get_cli_user_agent(): 获取CLI User-Agent（子类可覆盖）
+
+        Args:
+            client: httpx 异步客户端
+            base_url: API 基础 URL
+            api_key: API 密钥（已解密）
+            request_data: 请求数据
+            extra_headers: 端点配置的额外请求头
+            db: 数据库会话
+            user: 用户对象
+            provider_name: 提供商名称
+            provider_id: 提供商ID
+            api_key_id: API密钥ID
+            model_name: 模型名称
+
+        Returns:
+            测试响应数据
+        """
+        from src.api.handlers.base.endpoint_checker import build_safe_headers, run_endpoint_check
+
+        # 构建请求组件
+        url = cls.build_endpoint_url(base_url, request_data, model_name)
+        base_headers = cls.build_base_headers(api_key)
+        protected_keys = cls.get_protected_header_keys()
+
+        # 添加CLI User-Agent
+        cli_user_agent = cls.get_cli_user_agent()
+        if cli_user_agent:
+            base_headers["User-Agent"] = cli_user_agent
+            protected_keys = tuple(list(protected_keys) + ["user-agent"])
+
+        headers = build_safe_headers(base_headers, extra_headers, protected_keys)
+        body = cls.build_request_body(request_data)
+
+        # 获取有效的模型名称
+        effective_model_name = model_name or request_data.get("model")
+
+        return await run_endpoint_check(
+            client=client,
+            url=url,
+            headers=headers,
+            json_body=body,
+            api_format=cls.name,
+            # 用量计算参数（现在强制记录）
+            db=db,
+            user=user,
+            provider_name=provider_name,
+            provider_id=provider_id,
+            api_key_id=api_key_id,
+            model_name=effective_model_name,
+        )
+
+    # =========================================================================
+    # CLI Adapter 配置方法 - 子类应覆盖这些方法而不是整个 check_endpoint
+    # =========================================================================
+
+    @classmethod
+    def build_endpoint_url(cls, base_url: str, request_data: Dict[str, Any], model_name: Optional[str] = None) -> str:
+        """
+        构建CLI API端点URL - 子类应覆盖
+
+        Args:
+            base_url: API基础URL
+            request_data: 请求数据
+            model_name: 模型名称（某些API需要，如Gemini）
+
+        Returns:
+            完整的端点URL
+        """
+        raise NotImplementedError(f"{cls.FORMAT_ID} adapter must implement build_endpoint_url")
+
+    @classmethod
+    def build_base_headers(cls, api_key: str) -> Dict[str, str]:
+        """
+        构建CLI API认证头 - 子类应覆盖
+
+        Args:
+            api_key: API密钥
+
+        Returns:
+            基础认证头部字典
+        """
+        raise NotImplementedError(f"{cls.FORMAT_ID} adapter must implement build_base_headers")
+
+    @classmethod
+    def get_protected_header_keys(cls) -> tuple:
+        """
+        返回CLI API的保护头部key - 子类应覆盖
+
+        Returns:
+            保护头部key的元组
+        """
+        raise NotImplementedError(f"{cls.FORMAT_ID} adapter must implement get_protected_header_keys")
+
+    @classmethod
+    def build_request_body(cls, request_data: Dict[str, Any]) -> Dict[str, Any]:
+        """
+        构建CLI API请求体 - 子类应覆盖
+
+        Args:
+            request_data: 请求数据
+
+        Returns:
+            请求体字典
+        """
+        raise NotImplementedError(f"{cls.FORMAT_ID} adapter must implement build_request_body")
+
+    @classmethod
+    def get_cli_user_agent(cls) -> Optional[str]:
+        """
+        获取CLI User-Agent - 子类可覆盖
+
+        Returns:
+            CLI User-Agent字符串，如果不需要则为None
+        """
+        return None
 
 # =========================================================================
 # CLI Adapter 注册表 - 用于根据 API format 获取 CLI Adapter 实例

src/api/handlers/base/cli_handler_base.py

@@ -57,8 +57,10 @@ from src.models.database import (
     ProviderEndpoint,
     User,
 )
+from src.config.settings import config
 from src.services.provider.transport import build_provider_url
 from src.utils.sse_parser import SSEEventParser
+from src.utils.timeout import read_first_chunk_with_ttfb_timeout
 
 
 class CliMessageHandlerBase(BaseMessageHandler):
@@ -136,7 +138,7 @@ class CliMessageHandlerBase(BaseMessageHandler):
         查找逻辑：
         1. 直接通过 GlobalModel.name 匹配
         2. 查找该 Provider 的 Model 实现
-        3. 使用 provider_model_name / provider_model_aliases 选择最终名称
+        3. 使用 provider_model_name / provider_model_mappings 选择最终名称
 
         Args:
             source_model: 用户请求的模型名（必须是 GlobalModel.name）
@@ -153,9 +155,9 @@ class CliMessageHandlerBase(BaseMessageHandler):
         logger.debug(f"[CLI] _get_mapped_model: source={source_model}, provider={provider_id[:8]}..., mapping={mapping}")
 
         if mapping and mapping.model:
-            # 使用 select_provider_model_name 支持别名功能
-            # 传入 api_key.id 作为 affinity_key，实现相同用户稳定选择同一别名
-            # 传入 api_format 用于过滤适用的别名作用域
+            # 使用 select_provider_model_name 支持模型映射功能
+            # 传入 api_key.id 作为 affinity_key，实现相同用户稳定选择同一映射
+            # 传入 api_format 用于过滤适用的映射作用域
             affinity_key = self.api_key.id if self.api_key else None
             mapped_name = mapping.model.select_provider_model_name(
                 affinity_key, api_format=self.FORMAT_ID
@@ -400,7 +402,7 @@ class CliMessageHandlerBase(BaseMessageHandler):
         ctx.provider_api_format = str(endpoint.api_format) if endpoint.api_format else ""
         ctx.client_api_format = ctx.api_format  # 已在 process_stream 中设置
 
-        # 获取模型映射（别名/映射 → 实际模型名）
+        # 获取模型映射（映射名称 → 实际模型名）
         mapped_model = await self._get_mapped_model(
             source_model=ctx.model,
             provider_id=str(provider.id),
@@ -474,8 +476,8 @@ class CliMessageHandlerBase(BaseMessageHandler):
 
             stream_response.raise_for_status()
 
-            # 使用字节流迭代器（避免 aiter_lines 的性能问题）
-            byte_iterator = stream_response.aiter_raw()
+            # 使用字节流迭代器（避免 aiter_lines 的性能问题, aiter_bytes 会自动解压 gzip/deflate）
+            byte_iterator = stream_response.aiter_bytes()
 
             # 预读第一个数据块，检测嵌套错误（HTTP 200 但响应体包含错误）
             prefetched_chunks = await self._prefetch_and_check_embedded_error(
@@ -529,7 +531,7 @@ class CliMessageHandlerBase(BaseMessageHandler):
             # 检查是否需要格式转换
             needs_conversion = self._needs_format_conversion(ctx)
 
-            async for chunk in stream_response.aiter_raw():
+            async for chunk in stream_response.aiter_bytes():
                 # 在第一次输出数据前更新状态为 streaming
                 if not streaming_status_updated:
                     self._update_usage_to_streaming_with_ctx(ctx)
@@ -672,6 +674,8 @@ class CliMessageHandlerBase(BaseMessageHandler):
 
         同时检测 HTML 响应（通常是 base_url 配置错误导致返回网页）。
 
+        首次读取时会应用 TTFB（首字节超时）检测，超时则触发故障转移。
+
         Args:
             byte_iterator: 字节流迭代器
             provider: Provider 对象
@@ -684,6 +688,7 @@ class CliMessageHandlerBase(BaseMessageHandler):
         Raises:
             EmbeddedErrorException: 如果检测到嵌套错误
             ProviderNotAvailableException: 如果检测到 HTML 响应（配置错误）
+            ProviderTimeoutException: 如果首字节超时（TTFB timeout）
         """
         prefetched_chunks: list = []
         max_prefetch_lines = 5  # 最多预读5行来检测错误
@@ -704,7 +709,19 @@ class CliMessageHandlerBase(BaseMessageHandler):
             else:
                 provider_parser = self.parser
 
-            async for chunk in byte_iterator:
+            # 使用共享的 TTFB 超时函数读取首字节
+            ttfb_timeout = config.stream_first_byte_timeout
+            first_chunk, aiter = await read_first_chunk_with_ttfb_timeout(
+                byte_iterator,
+                timeout=ttfb_timeout,
+                request_id=self.request_id,
+                provider_name=str(provider.name),
+            )
+            prefetched_chunks.append(first_chunk)
+            buffer += first_chunk
+
+            # 继续读取剩余的预读数据
+            async for chunk in aiter:
                 prefetched_chunks.append(chunk)
                 buffer += chunk
 
@@ -785,12 +802,21 @@ class CliMessageHandlerBase(BaseMessageHandler):
                 if should_stop or line_count >= max_prefetch_lines:
                     break
 
-        except EmbeddedErrorException:
-            # 重新抛出嵌套错误
+        except (EmbeddedErrorException, ProviderTimeoutException, ProviderNotAvailableException):
+            # 重新抛出可重试的 Provider 异常，触发故障转移
             raise
+        except (OSError, IOError) as e:
+            # 网络 I/O 异常：记录警告，可能需要重试
+            logger.warning(
+                f"  [{self.request_id}] 预读流时发生网络异常: {type(e).__name__}: {e}"
+            )
         except Exception as e:
-            # 其他异常（如网络错误）在预读阶段发生，记录日志但不中断
-            logger.debug(f"  [{self.request_id}] 预读流时发生异常: {e}")
+            # 未预期的严重异常：记录错误并重新抛出，避免掩盖问题
+            logger.error(
+                f"  [{self.request_id}] 预读流时发生严重异常: {type(e).__name__}: {e}",
+                exc_info=True
+            )
+            raise
 
         return prefetched_chunks
 
@@ -1382,7 +1408,7 @@ class CliMessageHandlerBase(BaseMessageHandler):
             provider_name = str(provider.name)
             provider_api_format = str(endpoint.api_format) if endpoint.api_format else ""
 
-            # 获取模型映射（别名/映射 → 实际模型名）
+            # 获取模型映射（映射名称 → 实际模型名）
             mapped_model = await self._get_mapped_model(
                 source_model=model,
                 provider_id=str(provider.id),

src/api/handlers/base/stream_processor.py

@@ -25,10 +25,12 @@ from src.api.handlers.base.content_extractors import (
 from src.api.handlers.base.parsers import get_parser_for_format
 from src.api.handlers.base.response_parser import ResponseParser
 from src.api.handlers.base.stream_context import StreamContext
-from src.core.exceptions import EmbeddedErrorException
+from src.config.settings import config
+from src.core.exceptions import EmbeddedErrorException, ProviderTimeoutException
 from src.core.logger import logger
 from src.models.database import Provider, ProviderEndpoint
 from src.utils.sse_parser import SSEEventParser
+from src.utils.timeout import read_first_chunk_with_ttfb_timeout
 
 
 @dataclass
@@ -170,6 +172,8 @@ class StreamProcessor:
         某些 Provider（如 Gemini）可能返回 HTTP 200，但在响应体中包含错误信息。
         这种情况需要在流开始输出之前检测，以便触发重试逻辑。
 
+        首次读取时会应用 TTFB（首字节超时）检测，超时则触发故障转移。
+
         Args:
             byte_iterator: 字节流迭代器
             provider: Provider 对象
@@ -182,6 +186,7 @@ class StreamProcessor:
 
         Raises:
             EmbeddedErrorException: 如果检测到嵌套错误
+            ProviderTimeoutException: 如果首字节超时（TTFB timeout）
         """
         prefetched_chunks: list = []
         parser = self.get_parser_for_provider(ctx)
@@ -192,7 +197,19 @@ class StreamProcessor:
         decoder = codecs.getincrementaldecoder("utf-8")(errors="replace")
 
         try:
-            async for chunk in byte_iterator:
+            # 使用共享的 TTFB 超时函数读取首字节
+            ttfb_timeout = config.stream_first_byte_timeout
+            first_chunk, aiter = await read_first_chunk_with_ttfb_timeout(
+                byte_iterator,
+                timeout=ttfb_timeout,
+                request_id=self.request_id,
+                provider_name=str(provider.name),
+            )
+            prefetched_chunks.append(first_chunk)
+            buffer += first_chunk
+
+            # 继续读取剩余的预读数据
+            async for chunk in aiter:
                 prefetched_chunks.append(chunk)
                 buffer += chunk
 
@@ -262,10 +279,21 @@ class StreamProcessor:
                 if should_stop or line_count >= max_prefetch_lines:
                     break
 
-        except EmbeddedErrorException:
+        except (EmbeddedErrorException, ProviderTimeoutException):
+            # 重新抛出可重试的 Provider 异常，触发故障转移
             raise
+        except (OSError, IOError) as e:
+            # 网络 I/O <20><><EFBFBD>常：记录警告，可能需要重试
+            logger.warning(
+                f"  [{self.request_id}] 预读流时发生网络异常: {type(e).__name__}: {e}"
+            )
         except Exception as e:
-            logger.debug(f"  [{self.request_id}] 预读流时发生异常: {e}")
+            # 未预期的严重异常：记录错误并重新抛出，避免掩盖问题
+            logger.error(
+                f"  [{self.request_id}] 预读流时发生严重异常: {type(e).__name__}: {e}",
+                exc_info=True
+            )
+            raise
 
         return prefetched_chunks
 

src/api/handlers/base/utils.py

@@ -4,17 +4,28 @@ Handler 基础工具函数
 
 from typing import Any, Dict, Optional
 
+from src.core.logger import logger
+
 
 def extract_cache_creation_tokens(usage: Dict[str, Any]) -> int:
     """
-    提取缓存创建 tokens（兼容新旧格式）
+    提取缓存创建 tokens（兼容三种格式）
 
-    Claude API 在不同版本中使用了不同的字段名来表示缓存创建 tokens：
-    - 新格式（2024年后）：使用 claude_cache_creation_5_m_tokens 和
-      claude_cache_creation_1_h_tokens 分别表示 5 分钟和 1 小时缓存
-    - 旧格式：使用 cache_creation_input_tokens 表示总的缓存创建 tokens
+    根据 Anthropic API 文档，支持三种格式（按优先级）：
 
-    此函数自动检测并适配两种格式，优先使用新格式。
+    1. **嵌套格式（优先级最高）**：
+       usage.cache_creation.ephemeral_5m_input_tokens
+       usage.cache_creation.ephemeral_1h_input_tokens
+
+    2. **扁平新格式（优先级第二）**：
+       usage.claude_cache_creation_5_m_tokens
+       usage.claude_cache_creation_1_h_tokens
+
+    3. **旧格式（优先级第三）**：
+       usage.cache_creation_input_tokens
+
+    优先使用嵌套格式，如果嵌套格式字段存在但值为 0，则智能 fallback 到旧格式。
+    扁平格式和嵌套格式互斥，按顺序检查。
 
     Args:
         usage: API 响应中的 usage 字典
@@ -22,20 +33,63 @@ def extract_cache_creation_tokens(usage: Dict[str, Any]) -> int:
     Returns:
         缓存创建 tokens 总数
     """
-    # 检查新格式字段是否存在（而非值是否为 0）
-    # 如果字段存在，即使值为 0 也是合法的，不应 fallback 到旧格式
-    has_new_format = (
+    # 1. 检查嵌套格式（最新格式）
+    cache_creation = usage.get("cache_creation")
+    if isinstance(cache_creation, dict):
+        cache_5m = int(cache_creation.get("ephemeral_5m_input_tokens", 0))
+        cache_1h = int(cache_creation.get("ephemeral_1h_input_tokens", 0))
+        total = cache_5m + cache_1h
+
+        if total > 0:
+            logger.debug(
+                f"Using nested cache_creation: 5m={cache_5m}, 1h={cache_1h}, total={total}"
+            )
+            return total
+
+        # 嵌套格式存在但为 0，fallback 到旧格式
+        old_format = int(usage.get("cache_creation_input_tokens", 0))
+        if old_format > 0:
+            logger.debug(
+                f"Nested cache_creation is 0, using old format: {old_format}"
+            )
+            return old_format
+
+        # 都是 0，返回 0
+        return 0
+
+    # 2. 检查扁平新格式
+    has_flat_format = (
         "claude_cache_creation_5_m_tokens" in usage
         or "claude_cache_creation_1_h_tokens" in usage
     )
 
-    if has_new_format:
-        cache_5m = usage.get("claude_cache_creation_5_m_tokens", 0)
-        cache_1h = usage.get("claude_cache_creation_1_h_tokens", 0)
-        return int(cache_5m) + int(cache_1h)
+    if has_flat_format:
+        cache_5m = int(usage.get("claude_cache_creation_5_m_tokens", 0))
+        cache_1h = int(usage.get("claude_cache_creation_1_h_tokens", 0))
+        total = cache_5m + cache_1h
 
-    # 回退到旧格式
-    return int(usage.get("cache_creation_input_tokens", 0))
+        if total > 0:
+            logger.debug(
+                f"Using flat new format: 5m={cache_5m}, 1h={cache_1h}, total={total}"
+            )
+            return total
+
+        # 扁平格式存在但为 0，fallback 到旧格式
+        old_format = int(usage.get("cache_creation_input_tokens", 0))
+        if old_format > 0:
+            logger.debug(
+                f"Flat cache_creation is 0, using old format: {old_format}"
+            )
+            return old_format
+
+        # 都是 0，返回 0
+        return 0
+
+    # 3. 回退到旧格式
+    old_format = int(usage.get("cache_creation_input_tokens", 0))
+    if old_format > 0:
+        logger.debug(f"Using old format: cache_creation_input_tokens={old_format}")
+    return old_format
 
 
 def build_sse_headers(extra_headers: Optional[Dict[str, str]] = None) -> Dict[str, str]:

src/api/handlers/claude/adapter.py

@@ -4,8 +4,9 @@ Claude Chat Adapter - 基于 ChatAdapterBase 的 Claude Chat API 适配器
 处理 /v1/messages 端点的 Claude Chat 格式请求。
 """
 
-from typing import Any, Dict, Optional, Type
+from typing import Any, Dict, Optional, Tuple, Type
 
+import httpx
 from fastapi import HTTPException, Request
 from fastapi.responses import JSONResponse
 
@@ -155,6 +156,91 @@ class ClaudeChatAdapter(ChatAdapterBase):
             "thinking_enabled": bool(request_obj.thinking),
         }
 
+    @classmethod
+    async def fetch_models(
+        cls,
+        client: httpx.AsyncClient,
+        base_url: str,
+        api_key: str,
+        extra_headers: Optional[Dict[str, str]] = None,
+    ) -> Tuple[list, Optional[str]]:
+        """查询 Claude API 支持的模型列表"""
+        headers = {
+            "x-api-key": api_key,
+            "Authorization": f"Bearer {api_key}",
+            "anthropic-version": "2023-06-01",
+        }
+        if extra_headers:
+            # 防止 extra_headers 覆盖认证头
+            safe_headers = {
+                k: v for k, v in extra_headers.items()
+                if k.lower() not in ("x-api-key", "authorization", "anthropic-version")
+            }
+            headers.update(safe_headers)
+
+        # 构建 /v1/models URL
+        base_url = base_url.rstrip("/")
+        if base_url.endswith("/v1"):
+            models_url = f"{base_url}/models"
+        else:
+            models_url = f"{base_url}/v1/models"
+
+        try:
+            response = await client.get(models_url, headers=headers)
+            logger.debug(f"Claude models request to {models_url}: status={response.status_code}")
+            if response.status_code == 200:
+                data = response.json()
+                models = []
+                if "data" in data:
+                    models = data["data"]
+                elif isinstance(data, list):
+                    models = data
+                # 为每个模型添加 api_format 字段
+                for m in models:
+                    m["api_format"] = cls.FORMAT_ID
+                return models, None
+            else:
+                error_body = response.text[:500] if response.text else "(empty)"
+                error_msg = f"HTTP {response.status_code}: {error_body}"
+                logger.warning(f"Claude models request to {models_url} failed: {error_msg}")
+                return [], error_msg
+        except Exception as e:
+            error_msg = f"Request error: {str(e)}"
+            logger.warning(f"Failed to fetch Claude models from {models_url}: {e}")
+            return [], error_msg
+
+    @classmethod
+    def build_endpoint_url(cls, base_url: str) -> str:
+        """构建Claude API端点URL"""
+        base_url = base_url.rstrip("/")
+        if base_url.endswith("/v1"):
+            return f"{base_url}/messages"
+        else:
+            return f"{base_url}/v1/messages"
+
+    @classmethod
+    def build_base_headers(cls, api_key: str) -> Dict[str, str]:
+        """构建Claude API认证头"""
+        return {
+            "x-api-key": api_key,
+            "Content-Type": "application/json",
+            "anthropic-version": "2023-06-01",
+        }
+
+    @classmethod
+    def get_protected_header_keys(cls) -> tuple:
+        """返回Claude API的保护头部key"""
+        return ("x-api-key", "content-type", "anthropic-version")
+
+    @classmethod
+    def build_request_body(cls, request_data: Dict[str, Any]) -> Dict[str, Any]:
+        """构建Claude API请求体"""
+        return {
+            "model": request_data.get("model"),
+            "max_tokens": request_data.get("max_tokens", 100),
+            "messages": request_data.get("messages", []),
+        }
+
 
 def build_claude_adapter(x_app_header: Optional[str]):
     """根据 x-app 头部构造 Chat 或 Claude Code 适配器。"""

src/api/handlers/claude_cli/adapter.py

@@ -4,13 +4,15 @@ Claude CLI Adapter - 基于通用 CLI Adapter 基类的简化实现
 继承 CliAdapterBase，只需配置 FORMAT_ID 和 HANDLER_CLASS。
 """
 
-from typing import Any, Dict, Optional, Type
+from typing import Any, AsyncIterator, Dict, Optional, Tuple, Type, Union
 
+import httpx
 from fastapi import Request
 
 from src.api.handlers.base.cli_adapter_base import CliAdapterBase, register_cli_adapter
 from src.api.handlers.base.cli_handler_base import CliMessageHandlerBase
-from src.api.handlers.claude.adapter import ClaudeCapabilityDetector
+from src.api.handlers.claude.adapter import ClaudeCapabilityDetector, ClaudeChatAdapter
+from src.config.settings import config
 
 
 @register_cli_adapter
@@ -99,5 +101,66 @@ class ClaudeCliAdapter(CliAdapterBase):
             "system_present": bool(payload.get("system")),
         }
 
+    # =========================================================================
+    # 模型列表查询
+    # =========================================================================
+
+    @classmethod
+    async def fetch_models(
+        cls,
+        client: httpx.AsyncClient,
+        base_url: str,
+        api_key: str,
+        extra_headers: Optional[Dict[str, str]] = None,
+    ) -> Tuple[list, Optional[str]]:
+        """查询 Claude API 支持的模型列表（带 CLI User-Agent）"""
+        # 复用 ClaudeChatAdapter 的实现，添加 CLI User-Agent
+        cli_headers = {"User-Agent": config.internal_user_agent_claude_cli}
+        if extra_headers:
+            cli_headers.update(extra_headers)
+        models, error = await ClaudeChatAdapter.fetch_models(
+            client, base_url, api_key, cli_headers
+        )
+        # 更新 api_format 为 CLI 格式
+        for m in models:
+            m["api_format"] = cls.FORMAT_ID
+        return models, error
+
+    @classmethod
+    def build_endpoint_url(cls, base_url: str, request_data: Dict[str, Any], model_name: Optional[str] = None) -> str:
+        """构建Claude CLI API端点URL"""
+        base_url = base_url.rstrip("/")
+        if base_url.endswith("/v1"):
+            return f"{base_url}/messages"
+        else:
+            return f"{base_url}/v1/messages"
+
+    @classmethod
+    def build_base_headers(cls, api_key: str) -> Dict[str, str]:
+        """构建Claude CLI API认证头"""
+        return {
+            "Authorization": f"Bearer {api_key}",
+            "Content-Type": "application/json",
+        }
+
+    @classmethod
+    def get_protected_header_keys(cls) -> tuple:
+        """返回Claude CLI API的保护头部key"""
+        return ("authorization", "content-type")
+
+    @classmethod
+    def build_request_body(cls, request_data: Dict[str, Any]) -> Dict[str, Any]:
+        """构建Claude CLI API请求体"""
+        return {
+            "model": request_data.get("model"),
+            "max_tokens": request_data.get("max_tokens", 100),
+            "messages": request_data.get("messages", []),
+        }
+
+    @classmethod
+    def get_cli_user_agent(cls) -> Optional[str]:
+        """获取Claude CLI User-Agent"""
+        return config.internal_user_agent_claude_cli
+
 
 __all__ = ["ClaudeCliAdapter"]

src/api/handlers/gemini/adapter.py

@@ -4,13 +4,15 @@ Gemini Chat Adapter
 处理 Gemini API 格式的请求适配
 """
 
-from typing import Any, Dict, Optional, Type
+from typing import Any, AsyncIterator, Dict, Optional, Tuple, Type, Union
 
+import httpx
 from fastapi import HTTPException, Request
 from fastapi.responses import JSONResponse
 
 from src.api.handlers.base.chat_adapter_base import ChatAdapterBase, register_adapter
 from src.api.handlers.base.chat_handler_base import ChatHandlerBase
+from src.api.handlers.base.endpoint_checker import build_safe_headers, run_endpoint_check
 from src.core.logger import logger
 from src.models.gemini import GeminiRequest
 
@@ -151,6 +153,141 @@ class GeminiChatAdapter(ChatAdapterBase):
             },
         )
 
+    @classmethod
+    async def fetch_models(
+        cls,
+        client: httpx.AsyncClient,
+        base_url: str,
+        api_key: str,
+        extra_headers: Optional[Dict[str, str]] = None,
+    ) -> Tuple[list, Optional[str]]:
+        """查询 Gemini API 支持的模型列表"""
+        # 兼容 base_url 已包含 /v1beta 的情况
+        base_url_clean = base_url.rstrip("/")
+        if base_url_clean.endswith("/v1beta"):
+            models_url = f"{base_url_clean}/models?key={api_key}"
+        else:
+            models_url = f"{base_url_clean}/v1beta/models?key={api_key}"
+
+        headers: Dict[str, str] = {}
+        if extra_headers:
+            headers.update(extra_headers)
+
+        try:
+            response = await client.get(models_url, headers=headers)
+            logger.debug(f"Gemini models request to {models_url}: status={response.status_code}")
+            if response.status_code == 200:
+                data = response.json()
+                if "models" in data:
+                    # 转换为统一格式
+                    return [
+                        {
+                            "id": m.get("name", "").replace("models/", ""),
+                            "owned_by": "google",
+                            "display_name": m.get("displayName", ""),
+                            "api_format": cls.FORMAT_ID,
+                        }
+                        for m in data["models"]
+                    ], None
+                return [], None
+            else:
+                error_body = response.text[:500] if response.text else "(empty)"
+                error_msg = f"HTTP {response.status_code}: {error_body}"
+                logger.warning(f"Gemini models request to {models_url} failed: {error_msg}")
+                return [], error_msg
+        except Exception as e:
+            error_msg = f"Request error: {str(e)}"
+            logger.warning(f"Failed to fetch Gemini models from {models_url}: {e}")
+            return [], error_msg
+
+    @classmethod
+    def build_endpoint_url(cls, base_url: str) -> str:
+        """构建Gemini API端点URL"""
+        base_url = base_url.rstrip("/")
+        if base_url.endswith("/v1beta"):
+            return base_url  # 子类需要处理model参数
+        else:
+            return f"{base_url}/v1beta"
+
+    @classmethod
+    def build_base_headers(cls, api_key: str) -> Dict[str, str]:
+        """构建Gemini API认证头"""
+        return {
+            "x-goog-api-key": api_key,
+            "Content-Type": "application/json",
+        }
+
+    @classmethod
+    def get_protected_header_keys(cls) -> tuple:
+        """返回Gemini API的保护头部key"""
+        return ("x-goog-api-key", "content-type")
+
+    @classmethod
+    def build_request_body(cls, request_data: Dict[str, Any]) -> Dict[str, Any]:
+        """构建Gemini API请求体"""
+        return {
+            "contents": request_data.get("messages", []),
+            "generationConfig": {
+                "maxOutputTokens": request_data.get("max_tokens", 100),
+                "temperature": request_data.get("temperature", 0.7),
+            },
+            "safetySettings": [
+                {"category": "HARM_CATEGORY_HARASSMENT", "threshold": "BLOCK_NONE"}
+            ],
+        }
+
+    @classmethod
+    async def check_endpoint(
+        cls,
+        client: httpx.AsyncClient,
+        base_url: str,
+        api_key: str,
+        request_data: Dict[str, Any],
+        extra_headers: Optional[Dict[str, str]] = None,
+        # 用量计算参数
+        db: Optional[Any] = None,
+        user: Optional[Any] = None,
+        provider_name: Optional[str] = None,
+        provider_id: Optional[str] = None,
+        api_key_id: Optional[str] = None,
+        model_name: Optional[str] = None,
+    ) -> Dict[str, Any]:
+        """测试 Gemini API 模型连接性（非流式）"""
+        # Gemini需要从request_data或model_name参数获取model名称
+        effective_model_name = model_name or request_data.get("model", "")
+        if not effective_model_name:
+            return {
+                "error": "Model name is required for Gemini API",
+                "status_code": 400,
+            }
+
+        # 使用基类配置方法，但重写URL构建逻辑
+        base_url = cls.build_endpoint_url(base_url)
+        url = f"{base_url}/models/{effective_model_name}:generateContent"
+
+        # 构建请求组件
+        base_headers = cls.build_base_headers(api_key)
+        protected_keys = cls.get_protected_header_keys()
+        headers = build_safe_headers(base_headers, extra_headers, protected_keys)
+        body = cls.build_request_body(request_data)
+
+        # 使用基类的通用endpoint checker
+        from src.api.handlers.base.endpoint_checker import run_endpoint_check
+        return await run_endpoint_check(
+            client=client,
+            url=url,
+            headers=headers,
+            json_body=body,
+            api_format=cls.name,
+            # 用量计算参数（现在强制记录）
+            db=db,
+            user=user,
+            provider_name=provider_name,
+            provider_id=provider_id,
+            api_key_id=api_key_id,
+            model_name=effective_model_name,
+        )
+
 
 def build_gemini_adapter(x_app_header: str = "") -> GeminiChatAdapter:
     """

src/api/handlers/gemini_cli/adapter.py

@@ -4,12 +4,15 @@ Gemini CLI Adapter - 基于通用 CLI Adapter 基类的实现
 继承 CliAdapterBase，处理 Gemini CLI 格式的请求。
 """
 
-from typing import Any, Dict, Optional, Type
+from typing import Any, AsyncIterator, Dict, Optional, Tuple, Type, Union
 
+import httpx
 from fastapi import Request
 
 from src.api.handlers.base.cli_adapter_base import CliAdapterBase, register_cli_adapter
 from src.api.handlers.base.cli_handler_base import CliMessageHandlerBase
+from src.api.handlers.gemini.adapter import GeminiChatAdapter
+from src.config.settings import config
 
 
 @register_cli_adapter
@@ -95,6 +98,77 @@ class GeminiCliAdapter(CliAdapterBase):
             "safety_settings_count": len(payload.get("safety_settings") or []),
         }
 
+    # =========================================================================
+    # 模型列表查询
+    # =========================================================================
+
+    @classmethod
+    async def fetch_models(
+        cls,
+        client: httpx.AsyncClient,
+        base_url: str,
+        api_key: str,
+        extra_headers: Optional[Dict[str, str]] = None,
+    ) -> Tuple[list, Optional[str]]:
+        """查询 Gemini API 支持的模型列表（带 CLI User-Agent）"""
+        # 复用 GeminiChatAdapter 的实现，添加 CLI User-Agent
+        cli_headers = {"User-Agent": config.internal_user_agent_gemini_cli}
+        if extra_headers:
+            cli_headers.update(extra_headers)
+        models, error = await GeminiChatAdapter.fetch_models(
+            client, base_url, api_key, cli_headers
+        )
+        # 更新 api_format 为 CLI 格式
+        for m in models:
+            m["api_format"] = cls.FORMAT_ID
+        return models, error
+
+    @classmethod
+    def build_endpoint_url(cls, base_url: str, request_data: Dict[str, Any], model_name: Optional[str] = None) -> str:
+        """构建Gemini CLI API端点URL"""
+        effective_model_name = model_name or request_data.get("model", "")
+        if not effective_model_name:
+            raise ValueError("Model name is required for Gemini API")
+
+        base_url = base_url.rstrip("/")
+        if base_url.endswith("/v1beta"):
+            prefix = base_url
+        else:
+            prefix = f"{base_url}/v1beta"
+        return f"{prefix}/models/{effective_model_name}:generateContent"
+
+    @classmethod
+    def build_base_headers(cls, api_key: str) -> Dict[str, str]:
+        """构建Gemini CLI API认证头"""
+        return {
+            "x-goog-api-key": api_key,
+            "Content-Type": "application/json",
+        }
+
+    @classmethod
+    def get_protected_header_keys(cls) -> tuple:
+        """返回Gemini CLI API的保护头部key"""
+        return ("x-goog-api-key", "content-type")
+
+    @classmethod
+    def build_request_body(cls, request_data: Dict[str, Any]) -> Dict[str, Any]:
+        """构建Gemini CLI API请求体"""
+        return {
+            "contents": request_data.get("messages", []),
+            "generationConfig": {
+                "maxOutputTokens": request_data.get("max_tokens", 100),
+                "temperature": request_data.get("temperature", 0.7),
+            },
+            "safetySettings": [
+                {"category": "HARM_CATEGORY_HARASSMENT", "threshold": "BLOCK_NONE"}
+            ],
+        }
+
+    @classmethod
+    def get_cli_user_agent(cls) -> Optional[str]:
+        """获取Gemini CLI User-Agent"""
+        return config.internal_user_agent_gemini_cli
+
 
 def build_gemini_cli_adapter(x_app_header: str = "") -> GeminiCliAdapter:
     """

src/api/handlers/openai/adapter.py

@@ -4,12 +4,14 @@ OpenAI Chat Adapter - 基于 ChatAdapterBase 的 OpenAI Chat API 适配器
 处理 /v1/chat/completions 端点的 OpenAI Chat 格式请求。
 """
 
-from typing import Any, Dict, Optional, Type
+from typing import Any, AsyncIterator, Dict, Optional, Tuple, Type, Union
 
+import httpx
 from fastapi import Request
 from fastapi.responses import JSONResponse
 
 from src.api.handlers.base.chat_adapter_base import ChatAdapterBase, register_adapter
+from src.api.handlers.base.endpoint_checker import build_safe_headers, run_endpoint_check
 from src.api.handlers.base.chat_handler_base import ChatHandlerBase
 from src.core.logger import logger
 from src.models.openai import OpenAIRequest
@@ -105,5 +107,80 @@ class OpenAIChatAdapter(ChatAdapterBase):
             },
         )
 
+    @classmethod
+    async def fetch_models(
+        cls,
+        client: httpx.AsyncClient,
+        base_url: str,
+        api_key: str,
+        extra_headers: Optional[Dict[str, str]] = None,
+    ) -> Tuple[list, Optional[str]]:
+        """查询 OpenAI 兼容 API 支持的模型列表"""
+        headers = {
+            "Authorization": f"Bearer {api_key}",
+        }
+        if extra_headers:
+            # 防止 extra_headers 覆盖 Authorization
+            safe_headers = {k: v for k, v in extra_headers.items() if k.lower() != "authorization"}
+            headers.update(safe_headers)
+
+        # 构建 /v1/models URL
+        base_url = base_url.rstrip("/")
+        if base_url.endswith("/v1"):
+            models_url = f"{base_url}/models"
+        else:
+            models_url = f"{base_url}/v1/models"
+
+        try:
+            response = await client.get(models_url, headers=headers)
+            logger.debug(f"OpenAI models request to {models_url}: status={response.status_code}")
+            if response.status_code == 200:
+                data = response.json()
+                models = []
+                if "data" in data:
+                    models = data["data"]
+                elif isinstance(data, list):
+                    models = data
+                # 为每个模型添加 api_format 字段
+                for m in models:
+                    m["api_format"] = cls.FORMAT_ID
+                return models, None
+            else:
+                error_body = response.text[:500] if response.text else "(empty)"
+                error_msg = f"HTTP {response.status_code}: {error_body}"
+                logger.warning(f"OpenAI models request to {models_url} failed: {error_msg}")
+                return [], error_msg
+        except Exception as e:
+            error_msg = f"Request error: {str(e)}"
+            logger.warning(f"Failed to fetch models from {models_url}: {e}")
+            return [], error_msg
+
+    @classmethod
+    def build_endpoint_url(cls, base_url: str) -> str:
+        """构建OpenAI API端点URL"""
+        base_url = base_url.rstrip("/")
+        if base_url.endswith("/v1"):
+            return f"{base_url}/chat/completions"
+        else:
+            return f"{base_url}/v1/chat/completions"
+
+    @classmethod
+    def build_base_headers(cls, api_key: str) -> Dict[str, str]:
+        """构建OpenAI API认证头"""
+        return {
+            "Authorization": f"Bearer {api_key}",
+            "Content-Type": "application/json",
+        }
+
+    @classmethod
+    def get_protected_header_keys(cls) -> tuple:
+        """返回OpenAI API的保护头部key"""
+        return ("authorization", "content-type")
+
+    @classmethod
+    def build_request_body(cls, request_data: Dict[str, Any]) -> Dict[str, Any]:
+        """构建OpenAI API请求体"""
+        return request_data.copy()
+
 
 __all__ = ["OpenAIChatAdapter"]

src/api/handlers/openai_cli/adapter.py

@@ -4,12 +4,15 @@ OpenAI CLI Adapter - 基于通用 CLI Adapter 基类的简化实现
 继承 CliAdapterBase，只需配置 FORMAT_ID 和 HANDLER_CLASS。
 """
 
-from typing import Optional, Type
+from typing import Any, AsyncIterator, Dict, Optional, Tuple, Type, Union
 
+import httpx
 from fastapi import Request
 
 from src.api.handlers.base.cli_adapter_base import CliAdapterBase, register_cli_adapter
 from src.api.handlers.base.cli_handler_base import CliMessageHandlerBase
+from src.api.handlers.openai.adapter import OpenAIChatAdapter
+from src.config.settings import config
 
 
 @register_cli_adapter
@@ -40,5 +43,62 @@ class OpenAICliAdapter(CliAdapterBase):
             return authorization.replace("Bearer ", "")
         return None
 
+    # =========================================================================
+    # 模型列表查询
+    # =========================================================================
+
+    @classmethod
+    async def fetch_models(
+        cls,
+        client: httpx.AsyncClient,
+        base_url: str,
+        api_key: str,
+        extra_headers: Optional[Dict[str, str]] = None,
+    ) -> Tuple[list, Optional[str]]:
+        """查询 OpenAI 兼容 API 支持的模型列表（带 CLI User-Agent）"""
+        # 复用 OpenAIChatAdapter 的实现，添加 CLI User-Agent
+        cli_headers = {"User-Agent": config.internal_user_agent_openai_cli}
+        if extra_headers:
+            cli_headers.update(extra_headers)
+        models, error = await OpenAIChatAdapter.fetch_models(
+            client, base_url, api_key, cli_headers
+        )
+        # 更新 api_format 为 CLI 格式
+        for m in models:
+            m["api_format"] = cls.FORMAT_ID
+        return models, error
+
+    @classmethod
+    def build_endpoint_url(cls, base_url: str, request_data: Dict[str, Any], model_name: Optional[str] = None) -> str:
+        """构建OpenAI CLI API端点URL"""
+        base_url = base_url.rstrip("/")
+        if base_url.endswith("/v1"):
+            return f"{base_url}/chat/completions"
+        else:
+            return f"{base_url}/v1/chat/completions"
+
+    @classmethod
+    def build_base_headers(cls, api_key: str) -> Dict[str, str]:
+        """构建OpenAI CLI API认证头"""
+        return {
+            "Authorization": f"Bearer {api_key}",
+            "Content-Type": "application/json",
+        }
+
+    @classmethod
+    def get_protected_header_keys(cls) -> tuple:
+        """返回OpenAI CLI API的保护头部key"""
+        return ("authorization", "content-type")
+
+    @classmethod
+    def build_request_body(cls, request_data: Dict[str, Any]) -> Dict[str, Any]:
+        """构建OpenAI CLI API请求体"""
+        return request_data.copy()
+
+    @classmethod
+    def get_cli_user_agent(cls) -> Optional[str]:
+        """获取OpenAI CLI User-Agent"""
+        return config.internal_user_agent_openai_cli
+
 
 __all__ = ["OpenAICliAdapter"]

src/config/constants.py

@@ -77,7 +77,10 @@ class ConcurrencyDefaults:
     MAX_CONCURRENT_LIMIT = 200
 
     # 最小并发限制下限
-    MIN_CONCURRENT_LIMIT = 1
+    # 设置为 3 而不是 1，因为预留机制（10%预留给缓存用户）会导致
+    # 当 learned_max_concurrent=1 时新用户实际可用槽位为 0，永远无法命中
+    # 注意：当 limit < 10 时，预留机制实际不生效（预留槽位 = 0），这是可接受的
+    MIN_CONCURRENT_LIMIT = 3
 
     # === 探测性扩容参数 ===
     # 探测性扩容间隔（分钟）- 长时间无 429 且有流量时尝试扩容

src/config/settings.py

@@ -56,10 +56,11 @@ class Config:
 
         # Redis 依赖策略（生产默认必需，开发默认可选，可通过 REDIS_REQUIRED 覆盖）
         redis_required_env = os.getenv("REDIS_REQUIRED")
-        if redis_required_env is None:
-            self.require_redis = self.environment not in {"development", "test", "testing"}
-        else:
+        if redis_required_env is not None:
             self.require_redis = redis_required_env.lower() == "true"
+        else:
+            # 保持向后兼容：开发环境可选，生产环境必需
+            self.require_redis = self.environment not in {"development", "test", "testing"}
 
         # CORS配置 - 使用环境变量配置允许的源
         # 格式: 逗号分隔的域名列表,如 "http://localhost:3000,https://example.com"
@@ -133,6 +134,18 @@ class Config:
         self.concurrency_slot_ttl = int(os.getenv("CONCURRENCY_SLOT_TTL", "600"))
         self.cache_reservation_ratio = float(os.getenv("CACHE_RESERVATION_RATIO", "0.1"))
 
+        # 限流降级策略配置
+        # RATE_LIMIT_FAIL_OPEN: 当限流服务（Redis）异常时的行为
+        #
+        # True (默认): fail-open - 放行请求（优先可用性）
+        #   风险：Redis 故障期间无法限流，可能被滥用
+        #   适用：API 网关作为关键基础设施，必须保持高可用
+        #
+        # False: fail-close - 拒绝所有请求（优先安全性）
+        #   风险：Redis 故障会导致 API 网关不可用
+        #   适用：有严格速率限制要求的安全敏感场景
+        self.rate_limit_fail_open = os.getenv("RATE_LIMIT_FAIL_OPEN", "true").lower() == "true"
+
         # HTTP 请求超时配置（秒）
         self.http_connect_timeout = float(os.getenv("HTTP_CONNECT_TIMEOUT", "10.0"))
         self.http_write_timeout = float(os.getenv("HTTP_WRITE_TIMEOUT", "60.0"))
@@ -141,9 +154,23 @@ class Config:
         # 流式处理配置
         # STREAM_PREFETCH_LINES: 预读行数，用于检测嵌套错误
         # STREAM_STATS_DELAY: 统计记录延迟（秒），等待流完全关闭
+        # STREAM_FIRST_BYTE_TIMEOUT: 首字节超时（秒），等待首字节超过此时间触发故障转移
+        #   范围: 10-120 秒，默认 30 秒（必须小于 http_write_timeout 避免竞态）
         self.stream_prefetch_lines = int(os.getenv("STREAM_PREFETCH_LINES", "5"))
         self.stream_stats_delay = float(os.getenv("STREAM_STATS_DELAY", "0.1"))
-        # 注：流式平滑输出配置已移至数据库系统设置（stream_smoothing_*）
+        self.stream_first_byte_timeout = self._parse_ttfb_timeout()
+
+        # 内部请求 User-Agent 配置（用于查询上游模型列表等）
+        # 可通过环境变量覆盖默认值，模拟对应 CLI 客户端
+        self.internal_user_agent_claude_cli = os.getenv(
+            "CLAUDE_CLI_USER_AGENT", "claude-code/1.0.1"
+        )
+        self.internal_user_agent_openai_cli = os.getenv(
+            "OPENAI_CLI_USER_AGENT", "openai-codex/1.0"
+        )
+        self.internal_user_agent_gemini_cli = os.getenv(
+            "GEMINI_CLI_USER_AGENT", "gemini-cli/0.1.0"
+        )
 
         # 验证连接池配置
         self._validate_pool_config()
@@ -166,6 +193,39 @@ class Config:
         """智能计算最大溢出连接数 - 与 pool_size 相同"""
         return self.db_pool_size
 
+    def _parse_ttfb_timeout(self) -> float:
+        """
+        解析 TTFB 超时配置，带错误处理和范围限制
+
+        TTFB (Time To First Byte) 用于检测慢响应的 Provider，超时触发故障转移。
+        此值必须小于 http_write_timeout，避免竞态条件。
+
+        Returns:
+            超时时间（秒），范围 10-120，默认 30
+        """
+        default_timeout = 30.0
+        min_timeout = 10.0
+        max_timeout = 120.0  # 必须小于 http_write_timeout (默认 60s) 的 2 倍
+
+        raw_value = os.getenv("STREAM_FIRST_BYTE_TIMEOUT", str(default_timeout))
+        try:
+            timeout = float(raw_value)
+        except ValueError:
+            # 延迟导入，避免循环依赖（Config 初始化时 logger 可能未就绪）
+            self._ttfb_config_warning = (
+                f"无效的 STREAM_FIRST_BYTE_TIMEOUT 配置 '{raw_value}'，使用默认值 {default_timeout}秒"
+            )
+            return default_timeout
+
+        # 范围限制
+        clamped = max(min_timeout, min(max_timeout, timeout))
+        if clamped != timeout:
+            self._ttfb_config_warning = (
+                f"STREAM_FIRST_BYTE_TIMEOUT={timeout}秒超出范围 [{min_timeout}-{max_timeout}]，"
+                f"已调整为 {clamped}秒"
+            )
+        return clamped
+
     def _validate_pool_config(self) -> None:
         """验证连接池配置是否安全"""
         total_per_worker = self.db_pool_size + self.db_max_overflow
@@ -213,6 +273,10 @@ class Config:
         if hasattr(self, "_pool_config_warning") and self._pool_config_warning:
             logger.warning(self._pool_config_warning)
 
+        # TTFB 超时配置警告
+        if hasattr(self, "_ttfb_config_warning") and self._ttfb_config_warning:
+            logger.warning(self._ttfb_config_warning)
+
         # 管理员密码检查（必须在环境变量中设置）
         if hasattr(self, "_missing_admin_password") and self._missing_admin_password:
             logger.error("必须设置 ADMIN_PASSWORD 环境变量！")

src/core/enums.py

@@ -10,8 +10,8 @@ class APIFormat(Enum):
     """API格式枚举 - 决定请求/响应的处理方式"""
 
     CLAUDE = "CLAUDE"  # Claude API 格式
-    OPENAI = "OPENAI"  # OpenAI API 格式
     CLAUDE_CLI = "CLAUDE_CLI"  # Claude CLI API 格式（使用 authorization: Bearer）
+    OPENAI = "OPENAI"  # OpenAI API 格式
     OPENAI_CLI = "OPENAI_CLI"  # OpenAI CLI/Responses API 格式（用于 Claude Code 等客户端）
     GEMINI = "GEMINI"  # Google Gemini API 格式
     GEMINI_CLI = "GEMINI_CLI"  # Gemini CLI API 格式

src/core/exceptions.py

@@ -188,12 +188,16 @@ class ProviderNotAvailableException(ProviderException):
         message: str,
         provider_name: Optional[str] = None,
         request_metadata: Optional[Any] = None,
+        upstream_status: Optional[int] = None,
+        upstream_response: Optional[str] = None,
     ):
         super().__init__(
             message=message,
             provider_name=provider_name,
             request_metadata=request_metadata,
         )
+        self.upstream_status = upstream_status
+        self.upstream_response = upstream_response
 
 
 class ProviderTimeoutException(ProviderException):
@@ -442,6 +446,36 @@ class EmbeddedErrorException(ProviderException):
         self.error_status = error_status
 
 
+class ProviderCompatibilityException(ProviderException):
+    """Provider 兼容性错误异常 - 应该触发故障转移
+
+    用于处理因 Provider 不支持某些参数或功能导致的错误。
+    这类错误不是用户请求本身的问题，换一个 Provider 可能就能成功，应该触发故障转移。
+
+    常见场景：
+    - Unsupported parameter（不支持的参数）
+    - Unsupported model（不支持的模型）
+    - Unsupported feature（不支持的功能）
+    """
+
+    def __init__(
+        self,
+        message: str,
+        provider_name: Optional[str] = None,
+        status_code: int = 400,
+        upstream_error: Optional[str] = None,
+        request_metadata: Optional[Any] = None,
+    ):
+        self.upstream_error = upstream_error
+        super().__init__(
+            message=message,
+            provider_name=provider_name,
+            request_metadata=request_metadata,
+        )
+        # 覆盖状态码为 400（保持与上游一致）
+        self.status_code = status_code
+
+
 class UpstreamClientException(ProxyException):
     """上游返回的客户端错误异常 - HTTP 4xx 错误，不应该重试
 

src/core/metrics.py

@@ -50,7 +50,7 @@ model_mapping_resolution_total = Counter(
     "model_mapping_resolution_total",
     "Total number of model mapping resolutions",
     ["method", "cache_hit"],
-    # method: direct_match, provider_model_name, alias, not_found
+    # method: direct_match, provider_model_name, mapping, not_found
     # cache_hit: true, false
 )
 

src/main.py

@@ -4,13 +4,10 @@
 """
 
 from contextlib import asynccontextmanager
-from pathlib import Path
 
 import uvicorn
-from fastapi import FastAPI, HTTPException, Request
+from fastapi import FastAPI, HTTPException
 from fastapi.middleware.cors import CORSMiddleware
-from fastapi.responses import FileResponse
-from fastapi.staticfiles import StaticFiles
 
 from src.api.admin import router as admin_router
 from src.api.announcements import router as announcement_router
@@ -299,33 +296,6 @@ app.include_router(dashboard_router)  # 仪表盘端点
 app.include_router(public_router)  # 公开API端点（用户可查看提供商和模型）
 app.include_router(monitoring_router)  # 监控端点
 
-# 静态文件服务（前端构建产物）
-# 检查前端构建目录是否存在
-frontend_dist = Path(__file__).parent.parent / "frontend" / "dist"
-if frontend_dist.exists():
-    # 挂载静态资源目录
-    app.mount("/assets", StaticFiles(directory=str(frontend_dist / "assets")), name="assets")
-
-    # SPA catch-all路由 - 必须放在最后
-    @app.get("/{full_path:path}")
-    async def serve_spa(request: Request, full_path: str):
-        """
-        处理所有未匹配的GET请求，返回index.html供前端路由处理
-        仅对非API路径生效
-        """
-        # 如果是API路径，不处理
-        if full_path in {"api", "v1"} or full_path.startswith(("api/", "v1/")):
-            raise HTTPException(status_code=404, detail="Not Found")
-
-        # 返回index.html，让前端路由处理
-        index_file = frontend_dist / "index.html"
-        if index_file.exists():
-            return FileResponse(str(index_file))
-        else:
-            raise HTTPException(status_code=404, detail="Frontend not built")
-
-else:
-    logger.warning("前端构建目录不存在，前端路由将无法使用")
 
 
 def main():

src/middleware/plugin_middleware.py

@@ -336,10 +336,44 @@ class PluginMiddleware:
                     )
                 return result
             return None
+        except ConnectionError as e:
+            # Redis 连接错误：根据配置决定
+            logger.warning(f"Rate limit connection error: {e}")
+            if config.rate_limit_fail_open:
+                return None
+            else:
+                return RateLimitResult(
+                    allowed=False,
+                    remaining=0,
+                    retry_after=30,
+                    message="Rate limit service unavailable"
+                )
+        except TimeoutError as e:
+            # 超时错误：可能是负载过高，根据配置决定
+            logger.warning(f"Rate limit timeout: {e}")
+            if config.rate_limit_fail_open:
+                return None
+            else:
+                return RateLimitResult(
+                    allowed=False,
+                    remaining=0,
+                    retry_after=30,
+                    message="Rate limit service timeout"
+                )
         except Exception as e:
-            logger.error(f"Rate limit error: {e}")
-            # 发生错误时允许请求通过
-            return None
+            logger.error(f"Rate limit error: {type(e).__name__}: {e}")
+            # 其他异常：根据配置决定
+            if config.rate_limit_fail_open:
+                # fail-open: 异常时放行请求（优先可用性）
+                return None
+            else:
+                # fail-close: 异常时拒绝请求（优先安全性）
+                return RateLimitResult(
+                    allowed=False,
+                    remaining=0,
+                    retry_after=60,
+                    message="Rate limit service error"
+                )
 
     async def _call_pre_request_plugins(self, request: Request) -> None:
         """调用请求前的插件（当前保留扩展点）"""

src/models/admin_requests.py

@@ -317,6 +317,7 @@ class UpdateUserRequest(BaseModel):
 
     username: Optional[str] = Field(None, min_length=1, max_length=50)
     email: Optional[str] = Field(None, max_length=100)
+    password: Optional[str] = Field(None, min_length=6, max_length=128, description="新密码（留空保持不变）")
     quota_usd: Optional[float] = Field(None, ge=0)
     is_active: Optional[bool] = None
     role: Optional[str] = None

src/models/api.py

@@ -346,9 +346,9 @@ class ModelCreate(BaseModel):
     provider_model_name: str = Field(
         ..., min_length=1, max_length=200, description="Provider 侧的主模型名称"
     )
-    provider_model_aliases: Optional[List[dict]] = Field(
+    provider_model_mappings: Optional[List[dict]] = Field(
         None,
-        description="模型名称别名列表，格式: [{'name': 'alias1', 'priority': 1}, ...]",
+        description="模型名称映射列表，格式: [{'name': 'alias1', 'priority': 1}, ...]",
     )
     global_model_id: str = Field(..., description="关联的 GlobalModel ID（必填）")
     # 按次计费配置 - 可选，为空时使用 GlobalModel 默认值
@@ -376,9 +376,9 @@ class ModelUpdate(BaseModel):
     """更新模型请求"""
 
     provider_model_name: Optional[str] = Field(None, min_length=1, max_length=200)
-    provider_model_aliases: Optional[List[dict]] = Field(
+    provider_model_mappings: Optional[List[dict]] = Field(
         None,
-        description="模型名称别名列表，格式: [{'name': 'alias1', 'priority': 1}, ...]",
+        description="模型名称映射列表，格式: [{'name': 'alias1', 'priority': 1}, ...]",
     )
     global_model_id: Optional[str] = None
     # 按次计费配置
@@ -404,7 +404,7 @@ class ModelResponse(BaseModel):
     provider_id: str
     global_model_id: Optional[str]
     provider_model_name: str
-    provider_model_aliases: Optional[List[dict]] = None
+    provider_model_mappings: Optional[List[dict]] = None
 
     # 按次计费配置
     price_per_request: Optional[float] = None

src/models/database.py

@@ -671,10 +671,10 @@ class Model(Base):
 
     # Provider 映射配置
     provider_model_name = Column(String(200), nullable=False)  # Provider 侧的主模型名称
-    # 模型名称别名列表（带优先级），用于同一模型在 Provider 侧有多个名称变体的场景
+    # 模型名称映射列表（带优先级），用于同一模型在 Provider 侧有多个名称变体的场景
     # 格式: [{"name": "Claude-Sonnet-4.5", "priority": 1}, {"name": "Claude-Sonnet-4-5", "priority": 2}]
     # 为空时只使用 provider_model_name
-    provider_model_aliases = Column(JSON, nullable=True, default=None)
+    provider_model_mappings = Column(JSON, nullable=True, default=None)
 
     # 按次计费配置（每次请求的固定费用，美元）- 可为空，为空时使用 GlobalModel 的默认值
     price_per_request = Column(Float, nullable=True)  # 每次请求固定费用
@@ -820,25 +820,25 @@ class Model(Base):
     ) -> str:
         """按优先级选择要使用的 Provider 模型名称
 
-        如果配置了 provider_model_aliases，按优先级选择（数字越小越优先）；
-        相同优先级的别名通过哈希分散实现负载均衡（与 Key 调度策略一致）；
+        如果配置了 provider_model_mappings，按优先级选择（数字越小越优先）；
+        相同优先级的映射通过哈希分散实现负载均衡（与 Key 调度策略一致）；
         否则返回 provider_model_name。
 
         Args:
-            affinity_key: 用于哈希分散的亲和键（如用户 API Key 哈希），确保同一用户稳定选择同一别名
-            api_format: 当前请求的 API 格式（如 CLAUDE、OPENAI 等），用于过滤适用的别名
+            affinity_key: 用于哈希分散的亲和键（如用户 API Key 哈希），确保同一用户稳定选择同一映射
+            api_format: 当前请求的 API 格式（如 CLAUDE、OPENAI 等），用于过滤适用的映射
         """
         import hashlib
 
-        if not self.provider_model_aliases:
+        if not self.provider_model_mappings:
             return self.provider_model_name
 
-        raw_aliases = self.provider_model_aliases
-        if not isinstance(raw_aliases, list) or len(raw_aliases) == 0:
+        raw_mappings = self.provider_model_mappings
+        if not isinstance(raw_mappings, list) or len(raw_mappings) == 0:
             return self.provider_model_name
 
-        aliases: list[dict] = []
-        for raw in raw_aliases:
+        mappings: list[dict] = []
+        for raw in raw_mappings:
             if not isinstance(raw, dict):
                 continue
             name = raw.get("name")
@@ -846,10 +846,10 @@ class Model(Base):
                 continue
 
             # 检查 api_formats 作用域（如果配置了且当前有 api_format）
-            alias_api_formats = raw.get("api_formats")
-            if api_format and alias_api_formats:
+            mapping_api_formats = raw.get("api_formats")
+            if api_format and mapping_api_formats:
                 # 如果配置了作用域，只有匹配时才生效
-                if isinstance(alias_api_formats, list) and api_format not in alias_api_formats:
+                if isinstance(mapping_api_formats, list) and api_format not in mapping_api_formats:
                     continue
 
             raw_priority = raw.get("priority", 1)
@@ -860,47 +860,47 @@ class Model(Base):
             if priority < 1:
                 priority = 1
 
-            aliases.append({"name": name.strip(), "priority": priority})
+            mappings.append({"name": name.strip(), "priority": priority})
 
-        if not aliases:
+        if not mappings:
             return self.provider_model_name
 
         # 按优先级排序（数字越小越优先）
-        sorted_aliases = sorted(aliases, key=lambda x: x["priority"])
+        sorted_mappings = sorted(mappings, key=lambda x: x["priority"])
 
         # 获取最高优先级（最小数字）
-        highest_priority = sorted_aliases[0]["priority"]
+        highest_priority = sorted_mappings[0]["priority"]
 
-        # 获取所有最高优先级的别名
-        top_priority_aliases = [
-            alias for alias in sorted_aliases
-            if alias["priority"] == highest_priority
+        # 获取所有最高优先级的映射
+        top_priority_mappings = [
+            mapping for mapping in sorted_mappings
+            if mapping["priority"] == highest_priority
         ]
 
-        # 如果有多个相同优先级的别名，通过哈希分散选择
-        if len(top_priority_aliases) > 1 and affinity_key:
-            # 为每个别名计算哈希得分，选择得分最小的
-            def hash_score(alias: dict) -> int:
-                combined = f"{affinity_key}:{alias['name']}"
+        # 如果有多个相同优先级的映射，通过哈希分散选择
+        if len(top_priority_mappings) > 1 and affinity_key:
+            # 为每个映射计算哈希得分，选择得分最小的
+            def hash_score(mapping: dict) -> int:
+                combined = f"{affinity_key}:{mapping['name']}"
                 return int(hashlib.md5(combined.encode()).hexdigest(), 16)
 
-            selected = min(top_priority_aliases, key=hash_score)
-        elif len(top_priority_aliases) > 1:
+            selected = min(top_priority_mappings, key=hash_score)
+        elif len(top_priority_mappings) > 1:
             # 没有 affinity_key 时，使用确定性选择（按名称排序后取第一个）
             # 避免随机选择导致同一请求重试时选择不同的模型名称
-            selected = min(top_priority_aliases, key=lambda x: x["name"])
+            selected = min(top_priority_mappings, key=lambda x: x["name"])
         else:
-            selected = top_priority_aliases[0]
+            selected = top_priority_mappings[0]
 
         return selected["name"]
 
     def get_all_provider_model_names(self) -> list[str]:
-        """获取所有可用的 Provider 模型名称（主名称 + 别名）"""
+        """获取所有可用的 Provider 模型名称（主名称 + 映射名称）"""
         names = [self.provider_model_name]
-        if self.provider_model_aliases:
-            for alias in self.provider_model_aliases:
-                if isinstance(alias, dict) and alias.get("name"):
-                    names.append(alias["name"])
+        if self.provider_model_mappings:
+            for mapping in self.provider_model_mappings:
+                if isinstance(mapping, dict) and mapping.get("name"):
+                    names.append(mapping["name"])
         return names
 
 
@@ -1308,6 +1308,53 @@ class StatsDaily(Base):
     )
 
 
+class StatsDailyModel(Base):
+    """每日模型统计快照 - 用于快速查询每日模型维度数据"""
+
+    __tablename__ = "stats_daily_model"
+
+    id = Column(String(36), primary_key=True, default=lambda: str(uuid.uuid4()))
+
+    # 统计日期 (UTC)
+    date = Column(DateTime(timezone=True), nullable=False, index=True)
+
+    # 模型名称
+    model = Column(String(100), nullable=False)
+
+    # 请求统计
+    total_requests = Column(Integer, default=0, nullable=False)
+
+    # Token 统计
+    input_tokens = Column(BigInteger, default=0, nullable=False)
+    output_tokens = Column(BigInteger, default=0, nullable=False)
+    cache_creation_tokens = Column(BigInteger, default=0, nullable=False)
+    cache_read_tokens = Column(BigInteger, default=0, nullable=False)
+
+    # 成本统计 (USD)
+    total_cost = Column(Float, default=0.0, nullable=False)
+
+    # 性能统计
+    avg_response_time_ms = Column(Float, default=0.0, nullable=False)
+
+    # 时间戳
+    created_at = Column(
+        DateTime(timezone=True), default=lambda: datetime.now(timezone.utc), nullable=False
+    )
+    updated_at = Column(
+        DateTime(timezone=True),
+        default=lambda: datetime.now(timezone.utc),
+        onupdate=lambda: datetime.now(timezone.utc),
+        nullable=False,
+    )
+
+    # 唯一约束：每个模型每天只有一条记录
+    __table_args__ = (
+        UniqueConstraint("date", "model", name="uq_stats_daily_model"),
+        Index("idx_stats_daily_model_date", "date"),
+        Index("idx_stats_daily_model_date_model", "date", "model"),
+    )
+
+
 class StatsSummary(Base):
     """全局统计汇总 - 单行记录，存储截止到昨天的累计数据"""
 

src/models/endpoint_models.py

@@ -226,8 +226,11 @@ class EndpointAPIKeyUpdate(BaseModel):
     global_priority: Optional[int] = Field(
         default=None, description="全局 Key 优先级（全局 Key 优先模式，数字越小越优先）"
     )
-    # 注意：max_concurrent=None 表示不更新，要切换为自适应模式请使用专用 API
-    max_concurrent: Optional[int] = Field(default=None, ge=1, description="最大并发数")
+    # max_concurrent: 使用特殊标记区分"未提供"和"设置为 null（自适应模式）"
+    # - 不提供字段：不更新
+    # - 提供 null：切换为自适应模式
+    # - 提供数字：设置固定并发限制
+    max_concurrent: Optional[int] = Field(default=None, ge=1, description="最大并发数（null=自适应模式）")
     rate_limit: Optional[int] = Field(default=None, ge=1, description="速率限制")
     daily_limit: Optional[int] = Field(default=None, ge=1, description="每日限制")
     monthly_limit: Optional[int] = Field(default=None, ge=1, description="每月限制")

src/models/pydantic_models.py

@@ -301,6 +301,36 @@ class BatchAssignModelsToProviderResponse(BaseModel):
     errors: List[dict]
 
 
+class ImportFromUpstreamRequest(BaseModel):
+    """从上游提供商导入模型请求"""
+
+    model_ids: List[str] = Field(..., min_length=1, description="上游模型 ID 列表")
+
+
+class ImportFromUpstreamSuccessItem(BaseModel):
+    """导入成功的模型信息"""
+
+    model_id: str = Field(..., description="上游模型 ID")
+    global_model_id: str = Field(..., description="GlobalModel ID")
+    global_model_name: str = Field(..., description="GlobalModel 名称")
+    provider_model_id: str = Field(..., description="Provider Model ID")
+    created_global_model: bool = Field(..., description="是否新创建了 GlobalModel")
+
+
+class ImportFromUpstreamErrorItem(BaseModel):
+    """导入失败的模型信息"""
+
+    model_id: str = Field(..., description="上游模型 ID")
+    error: str = Field(..., description="错误信息")
+
+
+class ImportFromUpstreamResponse(BaseModel):
+    """从上游提供商导入模型响应"""
+
+    success: List[ImportFromUpstreamSuccessItem]
+    errors: List[ImportFromUpstreamErrorItem]
+
+
 __all__ = [
     "BatchAssignModelsToProviderRequest",
     "BatchAssignModelsToProviderResponse",
@@ -311,6 +341,10 @@ __all__ = [
     "GlobalModelResponse",
     "GlobalModelUpdate",
     "GlobalModelWithStats",
+    "ImportFromUpstreamErrorItem",
+    "ImportFromUpstreamRequest",
+    "ImportFromUpstreamResponse",
+    "ImportFromUpstreamSuccessItem",
     "ModelCapabilities",
     "ModelCatalogItem",
     "ModelCatalogProviderDetail",

src/services/auth/service.py

@@ -27,7 +27,7 @@ if not config.jwt_secret_key:
     if config.environment == "production":
         raise ValueError("JWT_SECRET_KEY must be set in production environment!")
     config.jwt_secret_key = secrets.token_urlsafe(32)
-    logger.warning(f"JWT_SECRET_KEY未在环境变量中找到，已生成随机密钥用于开发: {config.jwt_secret_key[:10]}...")
+    logger.warning("JWT_SECRET_KEY未在环境变量中找到，已生成随机密钥用于开发")
     logger.warning("生产环境请设置JWT_SECRET_KEY环境变量!")
 
 JWT_SECRET_KEY = config.jwt_secret_key

src/services/cache/aware_scheduler.py

@@ -30,6 +30,8 @@
 
 from __future__ import annotations
 
+import hashlib
+import random
 import time
 from dataclasses import dataclass
 from typing import TYPE_CHECKING, Dict, List, Optional, Tuple, Union
@@ -589,14 +591,14 @@ class CacheAwareScheduler:
 
         target_format = normalize_api_format(api_format)
 
-        # 0. 解析 model_name 到 GlobalModel（支持直接匹配和别名匹配，使用 ModelCacheService）
+        # 0. 解析 model_name 到 GlobalModel（支持直接匹配和映射名匹配，使用 ModelCacheService）
         global_model = await ModelCacheService.resolve_global_model_by_name_or_alias(db, model_name)
 
         if not global_model:
             logger.warning(f"GlobalModel not found: {model_name}")
             raise ModelNotSupportedException(model=model_name)
 
-        # 使用 GlobalModel.id 作为缓存亲和性的模型标识，确保别名和规范名都能命中同一个缓存
+        # 使用 GlobalModel.id 作为缓存亲和性的模型标识，确保映射名和规范名都能命中同一个缓存
         global_model_id: str = str(global_model.id)
         requested_model_name = model_name
         resolved_model_name = str(global_model.name)
@@ -751,19 +753,19 @@ class CacheAwareScheduler:
 
         支持两种匹配方式：
         1. 直接匹配 GlobalModel.name
-        2. 通过 ModelCacheService 匹配别名（全局查找）
+        2. 通过 ModelCacheService 匹配映射名（全局查找）
 
         Args:
             db: 数据库会话
             provider: Provider 对象
-            model_name: 模型名称（可以是 GlobalModel.name 或别名）
+            model_name: 模型名称（可以是 GlobalModel.name 或映射名）
             is_stream: 是否是流式请求，如果为 True 则同时检查流式支持
             capability_requirements: 能力需求（可选），用于检查模型是否支持所需能力
 
         Returns:
             (is_supported, skip_reason, supported_capabilities) - 是否支持、跳过原因、模型支持的能力列表
         """
-        # 使用 ModelCacheService 解析模型名称（支持别名）
+        # 使用 ModelCacheService 解析模型名称（支持映射名）
         global_model = await ModelCacheService.resolve_global_model_by_name_or_alias(db, model_name)
 
         if not global_model:
@@ -914,7 +916,7 @@ class CacheAwareScheduler:
             db: 数据库会话
             providers: Provider 列表
             target_format: 目标 API 格式
-            model_name: 模型名称（用户请求的名称，可能是别名）
+            model_name: 模型名称（用户请求的名称，可能是映射名）
             affinity_key: 亲和性标识符（通常为API Key ID）
             resolved_model_name: 解析后的 GlobalModel.name（用于 Key.allowed_models 校验）
             max_candidates: 最大候选数
@@ -956,7 +958,16 @@ class CacheAwareScheduler:
 
                 # 获取活跃的 Key 并按 internal_priority + 负载均衡排序
                 active_keys = [key for key in endpoint.api_keys if key.is_active]
-                keys = self._shuffle_keys_by_internal_priority(active_keys, affinity_key)
+                # 检查是否所有 Key 都是 TTL=0（轮换模式）
+                # 如果所有 Key 的 cache_ttl_minutes 都是 0 或 None，则使用随机排序
+                use_random = all(
+                    (key.cache_ttl_minutes or 0) == 0 for key in active_keys
+                ) if active_keys else False
+                if use_random and len(active_keys) > 1:
+                    logger.debug(
+                        f"  Endpoint {endpoint.id[:8]}... 启用 Key 轮换模式 (TTL=0, {len(active_keys)} keys)"
+                    )
+                keys = self._shuffle_keys_by_internal_priority(active_keys, affinity_key, use_random)
 
                 for key in keys:
                     # Key 级别的能力检查（模型级别的能力检查已在上面完成）
@@ -1170,6 +1181,7 @@ class CacheAwareScheduler:
         self,
         keys: List[ProviderAPIKey],
         affinity_key: Optional[str] = None,
+        use_random: bool = False,
     ) -> List[ProviderAPIKey]:
         """
         对 API Key 按 internal_priority 分组，同优先级内部基于 affinity_key 进行确定性打乱
@@ -1178,10 +1190,12 @@ class CacheAwareScheduler:
         - 数字越小越优先使用
         - 同优先级 Key 之间实现负载均衡
         - 使用 affinity_key 哈希确保同一请求 Key 的请求稳定（避免破坏缓存亲和性）
+        - 当 use_random=True 时，使用随机排序实现轮换（用于 TTL=0 的场景）
 
         Args:
             keys: API Key 列表
             affinity_key: 亲和性标识符（通常为 API Key ID，用于确定性打乱）
+            use_random: 是否使用随机排序（TTL=0 时为 True）
 
         Returns:
             排序后的 Key 列表
@@ -1198,28 +1212,35 @@ class CacheAwareScheduler:
             priority = key.internal_priority if key.internal_priority is not None else 999999
             priority_groups[priority].append(key)
 
-        # 对每个优先级组内的 Key 进行确定性打乱
+        # 对每个优先级组内的 Key 进行打乱
         result = []
         for priority in sorted(priority_groups.keys()):  # 数字小的优先级高，排前面
             group_keys = priority_groups[priority]
 
-            if len(group_keys) > 1 and affinity_key:
-                # 改进的哈希策略：为每个 key 计算独立的哈希值
-                import hashlib
+            if len(group_keys) > 1:
+                if use_random:
+                    # TTL=0 模式：使用随机排序实现 Key 轮换
+                    shuffled = list(group_keys)
+                    random.shuffle(shuffled)
+                    result.extend(shuffled)
+                elif affinity_key:
+                    # 正常模式：使用哈希确定性打乱（保持缓存亲和性）
+                    key_scores = []
+                    for key in group_keys:
+                        # 使用 affinity_key + key.id 的组合哈希
+                        hash_input = f"{affinity_key}:{key.id}"
+                        hash_value = int(hashlib.sha256(hash_input.encode()).hexdigest()[:16], 16)
+                        key_scores.append((hash_value, key))
 
-                key_scores = []
-                for key in group_keys:
-                    # 使用 affinity_key + key.id 的组合哈希
-                    hash_input = f"{affinity_key}:{key.id}"
-                    hash_value = int(hashlib.sha256(hash_input.encode()).hexdigest()[:16], 16)
-                    key_scores.append((hash_value, key))
-
-                # 按哈希值排序
-                sorted_group = [key for _, key in sorted(key_scores)]
-                result.extend(sorted_group)
+                    # 按哈希值排序
+                    sorted_group = [key for _, key in sorted(key_scores)]
+                    result.extend(sorted_group)
+                else:
+                    # 没有 affinity_key 时按 ID 排序保持稳定性
+                    result.extend(sorted(group_keys, key=lambda k: k.id))
             else:
-                # 单个 Key 或没有 affinity_key 时保持原顺序
-                result.extend(sorted(group_keys, key=lambda k: k.id))
+                # 单个 Key 直接添加
+                result.extend(group_keys)
 
         return result
 

src/services/cache/model_cache.py

@@ -198,7 +198,7 @@ class ModelCacheService:
         provider_id: Optional[str] = None,
         global_model_id: Optional[str] = None,
         provider_model_name: Optional[str] = None,
-        provider_model_aliases: Optional[list] = None,
+        provider_model_mappings: Optional[list] = None,
     ) -> None:
         """清除 Model 缓存
 
@@ -207,7 +207,7 @@ class ModelCacheService:
             provider_id: Provider ID（用于清除 provider_global 缓存）
             global_model_id: GlobalModel ID（用于清除 provider_global 缓存）
             provider_model_name: provider_model_name（用于清除 resolve 缓存）
-            provider_model_aliases: 映射名称列表（用于清除 resolve 缓存）
+            provider_model_mappings: 映射名称列表（用于清除 resolve 缓存）
         """
         # 清除 model:id 缓存
         await CacheService.delete(f"model:id:{model_id}")
@@ -222,16 +222,16 @@ class ModelCacheService:
         else:
             logger.debug(f"Model 缓存已清除: {model_id}")
 
-        # 清除 resolve 缓存（provider_model_name 和 aliases 可能都被用作解析 key）
+        # 清除 resolve 缓存（provider_model_name 和 mappings 可能都被用作解析 key）
         resolve_keys_to_clear = []
         if provider_model_name:
             resolve_keys_to_clear.append(provider_model_name)
-        if provider_model_aliases:
-            for alias_entry in provider_model_aliases:
-                if isinstance(alias_entry, dict):
-                    alias_name = alias_entry.get("name", "").strip()
-                    if alias_name:
-                        resolve_keys_to_clear.append(alias_name)
+        if provider_model_mappings:
+            for mapping_entry in provider_model_mappings:
+                if isinstance(mapping_entry, dict):
+                    mapping_name = mapping_entry.get("name", "").strip()
+                    if mapping_name:
+                        resolve_keys_to_clear.append(mapping_name)
 
         for key in resolve_keys_to_clear:
             await CacheService.delete(f"global_model:resolve:{key}")
@@ -261,8 +261,8 @@ class ModelCacheService:
         2. 通过 provider_model_name 匹配（查询 Model 表）
         3. 直接匹配 GlobalModel.name（兜底）
 
-        注意：此方法不使用 provider_model_aliases 进行全局解析。
-        provider_model_aliases 是 Provider 级别的别名配置，只在特定 Provider 上下文中生效，
+        注意：此方法不使用 provider_model_mappings 进行全局解析。
+        provider_model_mappings 是 Provider 级别的映射配置，只在特定 Provider 上下文中生效，
         由 resolve_provider_model() 处理。
 
         Args:
@@ -301,9 +301,9 @@ class ModelCacheService:
                     logger.debug(f"GlobalModel 缓存命中(映射解析): {normalized_name}")
                     return ModelCacheService._dict_to_global_model(cached_data)
 
-            # 2. 通过 provider_model_name 匹配（不考虑 provider_model_aliases）
-            # 重要：provider_model_aliases 是 Provider 级别的别名配置，只在特定 Provider 上下文中生效
-            # 全局解析不应该受到某个 Provider 别名配置的影响
+            # 2. 通过 provider_model_name 匹配（不考虑 provider_model_mappings）
+            # 重要：provider_model_mappings 是 Provider 级别的映射配置，只在特定 Provider 上下文中生效
+            # 全局解析不应该受到某个 Provider 映射配置的影响
             # 例如：Provider A 把 "haiku" 映射到 "sonnet"，不应该影响 Provider B 的 "haiku" 解析
             from src.models.database import Provider
 
@@ -401,7 +401,7 @@ class ModelCacheService:
             "provider_id": model.provider_id,
             "global_model_id": model.global_model_id,
             "provider_model_name": model.provider_model_name,
-            "provider_model_aliases": getattr(model, "provider_model_aliases", None),
+            "provider_model_mappings": getattr(model, "provider_model_mappings", None),
             "is_active": model.is_active,
             "is_available": model.is_available if hasattr(model, "is_available") else True,
             "price_per_request": (
@@ -424,7 +424,7 @@ class ModelCacheService:
             provider_id=model_dict["provider_id"],
             global_model_id=model_dict["global_model_id"],
             provider_model_name=model_dict["provider_model_name"],
-            provider_model_aliases=model_dict.get("provider_model_aliases"),
+            provider_model_mappings=model_dict.get("provider_model_mappings"),
             is_active=model_dict["is_active"],
             is_available=model_dict.get("is_available", True),
             price_per_request=model_dict.get("price_per_request"),

src/services/health/endpoint.py

@@ -234,8 +234,15 @@ class EndpointHealthService:
                 for api_format in format_key_mapping.keys()
             }
 
+        # 参数校验（API 层已通过 Query(ge=1) 保证，这里做防御性检查）
+        if lookback_hours <= 0 or segments <= 0:
+            raise ValueError(
+                f"lookback_hours and segments must be positive, "
+                f"got lookback_hours={lookback_hours}, segments={segments}"
+            )
+
         # 计算时间范围
-        interval_minutes = (lookback_hours * 60) // segments
+        segment_seconds = (lookback_hours * 3600) / segments
         start_time = now - timedelta(hours=lookback_hours)
 
         # 使用 RequestCandidate 表查询所有尝试记录
@@ -243,7 +250,7 @@ class EndpointHealthService:
         final_statuses = ["success", "failed", "skipped"]
 
         segment_expr = func.floor(
-            func.extract('epoch', RequestCandidate.created_at - start_time) / (interval_minutes * 60)
+            func.extract('epoch', RequestCandidate.created_at - start_time) / segment_seconds
         ).label('segment_idx')
 
         candidate_stats = (

src/services/model/cost.py

@@ -443,7 +443,7 @@ class ModelCostService:
 
         Args:
             provider: Provider 对象或提供商名称
-            model: 用户请求的模型名（可能是 GlobalModel.name 或别名）
+            model: 用户请求的模型名（可能是 GlobalModel.name 或映射名）
 
         Returns:
             按次计费价格，如果没有配置则返回 None

src/services/model/mapper.py

@@ -84,11 +84,11 @@ class ModelMapperMiddleware:
         获取模型映射
 
         简化后的逻辑:
-        1. 通过 GlobalModel.name 或别名解析 GlobalModel
+        1. 通过 GlobalModel.name 或映射名解析 GlobalModel
         2. 找到 GlobalModel 后，查找该 Provider 的 Model 实现
 
         Args:
-            source_model: 用户请求的模型名（可以是 GlobalModel.name 或别名）
+            source_model: 用户请求的模型名（可以是 GlobalModel.name 或映射名）
             provider_id: 提供商ID (UUID)
 
         Returns:
@@ -101,7 +101,7 @@ class ModelMapperMiddleware:
 
         mapping = None
 
-        # 步骤 1: 解析 GlobalModel（支持别名）
+        # 步骤 1: 解析 GlobalModel（支持映射名）
         global_model = await ModelCacheService.resolve_global_model_by_name_or_alias(
             self.db, source_model
         )

src/services/model/service.py

@@ -51,7 +51,7 @@ class ModelService:
                 provider_id=provider_id,
                 global_model_id=model_data.global_model_id,
                 provider_model_name=model_data.provider_model_name,
-                provider_model_aliases=model_data.provider_model_aliases,
+                provider_model_mappings=model_data.provider_model_mappings,
                 price_per_request=model_data.price_per_request,
                 tiered_pricing=model_data.tiered_pricing,
                 supports_vision=model_data.supports_vision,
@@ -153,9 +153,9 @@ class ModelService:
         if not model:
             raise NotFoundException(f"模型 {model_id} 不存在")
 
-        # 保存旧的别名，用于清除缓存
+        # 保存旧的映射，用于清除缓存
         old_provider_model_name = model.provider_model_name
-        old_provider_model_aliases = model.provider_model_aliases
+        old_provider_model_mappings = model.provider_model_mappings
 
         # 更新字段
         update_data = model_data.model_dump(exclude_unset=True)
@@ -174,26 +174,26 @@ class ModelService:
             db.refresh(model)
 
             # 清除 Redis 缓存（异步执行，不阻塞返回）
-            # 先清除旧的别名缓存
+            # 先清除旧的映射缓存
             asyncio.create_task(
                 ModelCacheService.invalidate_model_cache(
                     model_id=model.id,
                     provider_id=model.provider_id,
                     global_model_id=model.global_model_id,
                     provider_model_name=old_provider_model_name,
-                    provider_model_aliases=old_provider_model_aliases,
+                    provider_model_mappings=old_provider_model_mappings,
                 )
             )
-            # 再清除新的别名缓存（如果有变化）
+            # 再清除新的映射缓存（如果有变化）
             if (model.provider_model_name != old_provider_model_name or
-                model.provider_model_aliases != old_provider_model_aliases):
+                model.provider_model_mappings != old_provider_model_mappings):
                 asyncio.create_task(
                     ModelCacheService.invalidate_model_cache(
                         model_id=model.id,
                         provider_id=model.provider_id,
                         global_model_id=model.global_model_id,
                         provider_model_name=model.provider_model_name,
-                        provider_model_aliases=model.provider_model_aliases,
+                        provider_model_mappings=model.provider_model_mappings,
                     )
                 )
 
@@ -246,7 +246,7 @@ class ModelService:
             "provider_id": model.provider_id,
             "global_model_id": model.global_model_id,
             "provider_model_name": model.provider_model_name,
-            "provider_model_aliases": model.provider_model_aliases,
+            "provider_model_mappings": model.provider_model_mappings,
         }
 
         try:
@@ -260,7 +260,7 @@ class ModelService:
                     provider_id=cache_info["provider_id"],
                     global_model_id=cache_info["global_model_id"],
                     provider_model_name=cache_info["provider_model_name"],
-                    provider_model_aliases=cache_info["provider_model_aliases"],
+                    provider_model_mappings=cache_info["provider_model_mappings"],
                 )
             )
 
@@ -297,7 +297,7 @@ class ModelService:
                 provider_id=model.provider_id,
                 global_model_id=model.global_model_id,
                 provider_model_name=model.provider_model_name,
-                provider_model_aliases=model.provider_model_aliases,
+                provider_model_mappings=model.provider_model_mappings,
             )
         )
 
@@ -390,7 +390,7 @@ class ModelService:
             provider_id=model.provider_id,
             global_model_id=model.global_model_id,
             provider_model_name=model.provider_model_name,
-            provider_model_aliases=model.provider_model_aliases,
+            provider_model_mappings=model.provider_model_mappings,
             # 原始配置值（可能为空）
             price_per_request=model.price_per_request,
             tiered_pricing=model.tiered_pricing,

src/services/orchestration/error_classifier.py

@@ -15,6 +15,7 @@ from src.core.enums import APIFormat
 from src.core.exceptions import (
     ConcurrencyLimitError,
     ProviderAuthException,
+    ProviderCompatibilityException,
     ProviderException,
     ProviderNotAvailableException,
     ProviderRateLimitException,
@@ -81,7 +82,9 @@ class ErrorClassifier:
         "context_length_exceeded",  # 上下文长度超限
         "content_length_limit",  # 请求内容长度超限 (Claude API)
         "content_length_exceeds",  # 内容长度超限变体 (AWS CodeWhisperer)
-        "max_tokens",  # token 数超限
+        # 注意：移除了 "max_tokens"，因为 max_tokens 相关错误可能是 Provider 兼容性问题
+        # 如 "Unsupported parameter: 'max_tokens' is not supported with this model"
+        # 这类错误应由 COMPATIBILITY_ERROR_PATTERNS 处理
         "invalid_prompt",  # 无效的提示词
         "content too long",  # 内容过长
         "input is too long",  # 输入过长 (AWS)
@@ -136,6 +139,19 @@ class ErrorClassifier:
         "CONTENT_POLICY_VIOLATION",
     )
 
+    # Provider 兼容性错误模式 - 这类错误应该触发故障转移
+    # 因为换一个 Provider 可能就能成功
+    COMPATIBILITY_ERROR_PATTERNS: Tuple[str, ...] = (
+        "unsupported parameter",  # 不支持的参数
+        "unsupported model",  # 不支持的模型
+        "unsupported feature",  # 不支持的功能
+        "not supported with this model",  # 此模型不支持
+        "model does not support",  # 模型不支持
+        "parameter is not supported",  # 参数不支持
+        "feature is not supported",  # 功能不支持
+        "not available for this model",  # 此模型不可用
+    )
+
     def _parse_error_response(self, error_text: Optional[str]) -> Dict[str, Any]:
         """
         解析错误响应为结构化数据
@@ -261,6 +277,25 @@ class ErrorClassifier:
         search_text = f"{parsed['message']} {parsed['raw']}".lower()
         return any(pattern.lower() in search_text for pattern in self.CLIENT_ERROR_PATTERNS)
 
+    def _is_compatibility_error(self, error_text: Optional[str]) -> bool:
+        """
+        检测错误响应是否为 Provider 兼容性错误（应触发故障转移）
+
+        这类错误是因为 Provider 不支持某些参数或功能导致的，
+        换一个 Provider 可能就能成功。
+
+        Args:
+            error_text: 错误响应文本
+
+        Returns:
+            是否为兼容性错误
+        """
+        if not error_text:
+            return False
+
+        search_text = error_text.lower()
+        return any(pattern.lower() in search_text for pattern in self.COMPATIBILITY_ERROR_PATTERNS)
+
     def _extract_error_message(self, error_text: Optional[str]) -> Optional[str]:
         """
         从错误响应中提取错误消息
@@ -425,6 +460,16 @@ class ErrorClassifier:
                 ),
             )
 
+        # 400 错误：先检查是否为 Provider 兼容性错误（应触发故障转移）
+        if status == 400 and self._is_compatibility_error(error_response_text):
+            logger.info(f"检测到 Provider 兼容性错误，将触发故障转移: {extracted_message}")
+            return ProviderCompatibilityException(
+                message=extracted_message or "Provider 不支持此请求",
+                provider_name=provider_name,
+                status_code=400,
+                upstream_error=error_response_text,
+            )
+
         # 400 错误：检查是否为客户端请求错误（不应重试）
         if status == 400 and self._is_client_error(error_response_text):
             logger.info(f"检测到客户端请求错误，不进行重试: {extracted_message}")

src/services/orchestration/fallback_orchestrator.py

@@ -427,6 +427,9 @@ class FallbackOrchestrator:
             )
             # str(cause) 可能为空（如 httpx 超时异常），使用 repr() 作为备用
             error_msg = str(cause) or repr(cause)
+            # 如果是 ProviderNotAvailableException，附加上游响应
+            if hasattr(cause, "upstream_response") and cause.upstream_response:
+                error_msg = f"{error_msg} | 上游响应: {cause.upstream_response[:500]}"
             RequestCandidateService.mark_candidate_failed(
                 db=self.db,
                 candidate_id=candidate_record_id,
@@ -439,6 +442,9 @@ class FallbackOrchestrator:
 
         # 未知错误：记录失败并抛出
         error_msg = str(cause) or repr(cause)
+        # 如果是 ProviderNotAvailableException，附加上游响应
+        if hasattr(cause, "upstream_response") and cause.upstream_response:
+            error_msg = f"{error_msg} | 上游响应: {cause.upstream_response[:500]}"
         RequestCandidateService.mark_candidate_failed(
             db=self.db,
             candidate_id=candidate_record_id,

src/services/request/result.py

@@ -289,11 +289,17 @@ class RequestResult:
             status_code = 500
             error_type = "internal_error"
 
+        # 构建错误消息，包含上游响应信息
+        error_message = str(exception)
+        if isinstance(exception, ProviderNotAvailableException):
+            if exception.upstream_response:
+                error_message = f"{error_message} | 上游响应: {exception.upstream_response[:500]}"
+
         return cls(
             status=RequestStatus.FAILED,
             metadata=metadata,
             status_code=status_code,
-            error_message=str(exception),
+            error_message=error_message,
             error_type=error_type,
             response_time_ms=response_time_ms,
             is_stream=is_stream,

src/services/system/cleanup_scheduler.py

@@ -208,89 +208,127 @@ class CleanupScheduler:
                         return
 
                     # 非首次运行，检查最近是否有缺失的日期需要回填
-                    latest_stat = db.query(StatsDaily).order_by(StatsDaily.date.desc()).first()
+                    from src.models.database import StatsDailyModel
 
-                    if latest_stat:
-                        latest_date_utc = latest_stat.date
-                        if latest_date_utc.tzinfo is None:
-                            latest_date_utc = latest_date_utc.replace(tzinfo=timezone.utc)
-                        else:
-                            latest_date_utc = latest_date_utc.astimezone(timezone.utc)
+                    yesterday_business_date = today_local.date() - timedelta(days=1)
+                    max_backfill_days: int = SystemConfigService.get_config(
+                        db, "max_stats_backfill_days", 30
+                    ) or 30
 
-                        # 使用业务日期计算缺失区间（避免用 UTC 年月日导致日期偏移，且对 DST 更安全）
-                        latest_business_date = latest_date_utc.astimezone(app_tz).date()
-                        yesterday_business_date = today_local.date() - timedelta(days=1)
-                        missing_start_date = latest_business_date + timedelta(days=1)
+                    # 计算回填检查的起始日期
+                    check_start_date = yesterday_business_date - timedelta(
+                        days=max_backfill_days - 1
+                    )
 
-                        if missing_start_date <= yesterday_business_date:
-                            missing_days = (
-                                yesterday_business_date - missing_start_date
-                            ).days + 1
+                    # 获取 StatsDaily 和 StatsDailyModel 中已有数据的日期集合
+                    existing_daily_dates = set()
+                    existing_model_dates = set()
 
-                            # 限制最大回填天数，防止停机很久后一次性回填太多
-                            max_backfill_days: int = SystemConfigService.get_config(
-                                db, "max_stats_backfill_days", 30
-                            ) or 30
-                            if missing_days > max_backfill_days:
-                                logger.warning(
-                                    f"缺失 {missing_days} 天数据超过最大回填限制 "
-                                    f"{max_backfill_days} 天，只回填最近 {max_backfill_days} 天"
+                    daily_stats = (
+                        db.query(StatsDaily.date)
+                        .filter(StatsDaily.date >= check_start_date.isoformat())
+                        .all()
+                    )
+                    for (stat_date,) in daily_stats:
+                        if stat_date.tzinfo is None:
+                            stat_date = stat_date.replace(tzinfo=timezone.utc)
+                        existing_daily_dates.add(stat_date.astimezone(app_tz).date())
+
+                    model_stats = (
+                        db.query(StatsDailyModel.date)
+                        .filter(StatsDailyModel.date >= check_start_date.isoformat())
+                        .distinct()
+                        .all()
+                    )
+                    for (stat_date,) in model_stats:
+                        if stat_date.tzinfo is None:
+                            stat_date = stat_date.replace(tzinfo=timezone.utc)
+                        existing_model_dates.add(stat_date.astimezone(app_tz).date())
+
+                    # 找出需要回填的日期
+                    all_dates = set()
+                    current = check_start_date
+                    while current <= yesterday_business_date:
+                        all_dates.add(current)
+                        current += timedelta(days=1)
+
+                    # 需要回填 StatsDaily 的日期
+                    missing_daily_dates = all_dates - existing_daily_dates
+                    # 需要回填 StatsDailyModel 的日期
+                    missing_model_dates = all_dates - existing_model_dates
+                    # 合并所有需要处理的日期
+                    dates_to_process = missing_daily_dates | missing_model_dates
+
+                    if dates_to_process:
+                        sorted_dates = sorted(dates_to_process)
+                        logger.info(
+                            f"检测到 {len(dates_to_process)} 天的统计数据需要回填 "
+                            f"(StatsDaily 缺失 {len(missing_daily_dates)} 天, "
+                            f"StatsDailyModel 缺失 {len(missing_model_dates)} 天)"
+                        )
+
+                        users = (
+                            db.query(DBUser.id).filter(DBUser.is_active.is_(True)).all()
+                        )
+
+                        failed_dates = 0
+                        failed_users = 0
+
+                        for current_date in sorted_dates:
+                            try:
+                                current_date_local = datetime.combine(
+                                    current_date, datetime.min.time(), tzinfo=app_tz
                                 )
-                                missing_start_date = yesterday_business_date - timedelta(
-                                    days=max_backfill_days - 1
-                                )
-                                missing_days = max_backfill_days
-
-                            logger.info(
-                                f"检测到缺失 {missing_days} 天的统计数据 "
-                                f"({missing_start_date} ~ {yesterday_business_date})，开始回填..."
-                            )
-
-                            current_date = missing_start_date
-                            users = (
-                                db.query(DBUser.id).filter(DBUser.is_active.is_(True)).all()
-                            )
-
-                            while current_date <= yesterday_business_date:
-                                try:
-                                    current_date_local = datetime.combine(
-                                        current_date, datetime.min.time(), tzinfo=app_tz
-                                    )
+                                # 只在缺失时才聚合对应的表
+                                if current_date in missing_daily_dates:
                                     StatsAggregatorService.aggregate_daily_stats(
                                         db, current_date_local
                                     )
-                                    for (user_id,) in users:
-                                        try:
-                                            StatsAggregatorService.aggregate_user_daily_stats(
-                                                db, user_id, current_date_local
-                                            )
-                                        except Exception as e:
-                                            logger.warning(
-                                                f"回填用户 {user_id} 日期 {current_date} 失败: {e}"
-                                            )
-                                            try:
-                                                db.rollback()
-                                            except Exception:
-                                                pass
-                                except Exception as e:
-                                    logger.warning(f"回填日期 {current_date} 失败: {e}")
+                                if current_date in missing_model_dates:
+                                    StatsAggregatorService.aggregate_daily_model_stats(
+                                        db, current_date_local
+                                    )
+                                # 用户统计在任一缺失时都回填
+                                for (user_id,) in users:
                                     try:
-                                        db.rollback()
-                                    except Exception:
-                                        pass
+                                        StatsAggregatorService.aggregate_user_daily_stats(
+                                            db, user_id, current_date_local
+                                        )
+                                    except Exception as e:
+                                        failed_users += 1
+                                        logger.warning(
+                                            f"回填用户 {user_id} 日期 {current_date} 失败: {e}"
+                                        )
+                                        try:
+                                            db.rollback()
+                                        except Exception as rollback_err:
+                                            logger.error(f"回滚失败: {rollback_err}")
+                            except Exception as e:
+                                failed_dates += 1
+                                logger.warning(f"回填日期 {current_date} 失败: {e}")
+                                try:
+                                    db.rollback()
+                                except Exception as rollback_err:
+                                    logger.error(f"回滚失败: {rollback_err}")
 
-                                current_date += timedelta(days=1)
+                        StatsAggregatorService.update_summary(db)
 
-                            StatsAggregatorService.update_summary(db)
-                            logger.info(f"缺失数据回填完成，共 {missing_days} 天")
+                        if failed_dates > 0 or failed_users > 0:
+                            logger.warning(
+                                f"回填完成，共处理 {len(dates_to_process)} 天，"
+                                f"失败: {failed_dates} 天, {failed_users} 个用户记录"
+                            )
                         else:
-                            logger.info("统计数据已是最新，无需回填")
+                            logger.info(f"缺失数据回填完成，共处理 {len(dates_to_process)} 天")
+                    else:
+                        logger.info("统计数据已是最新，无需回填")
                     return
 
                 # 定时任务：聚合昨天的数据
                 yesterday_local = today_local - timedelta(days=1)
 
                 StatsAggregatorService.aggregate_daily_stats(db, yesterday_local)
+                StatsAggregatorService.aggregate_daily_model_stats(db, yesterday_local)
 
                 users = db.query(DBUser.id).filter(DBUser.is_active.is_(True)).all()
                 for (user_id,) in users:

src/services/system/config.py

@@ -78,19 +78,6 @@ class SystemConfigService:
             "value": False,
             "description": "是否自动删除过期的API Key（True=物理删除，False=仅禁用），仅管理员可配置",
         },
-        # 流式平滑输出配置
-        "stream_smoothing_enabled": {
-            "value": False,
-            "description": "是否启用流式平滑输出，自动根据文本长度调整输出速度",
-        },
-        "stream_smoothing_chunk_size": {
-            "value": 20,
-            "description": "流式平滑输出每个小块的字符数",
-        },
-        "stream_smoothing_delay_ms": {
-            "value": 8,
-            "description": "流式平滑输出每个小块之间的延迟毫秒数",
-        },
     }
 
     @classmethod

src/services/system/stats_aggregator.py

@@ -16,6 +16,7 @@ from src.models.database import (
     ApiKey,
     RequestCandidate,
     StatsDaily,
+    StatsDailyModel,
     StatsSummary,
     StatsUserDaily,
     Usage,
@@ -219,6 +220,120 @@ class StatsAggregatorService:
         logger.info(f"[StatsAggregator] 聚合日期 {date.date()} 完成: {computed['total_requests']} 请求")
         return stats
 
+    @staticmethod
+    def aggregate_daily_model_stats(db: Session, date: datetime) -> list[StatsDailyModel]:
+        """聚合指定日期的模型维度统计数据
+
+        Args:
+            db: 数据库会话
+            date: 要聚合的业务日期
+
+        Returns:
+            StatsDailyModel 记录列表
+        """
+        day_start, day_end = _get_business_day_range(date)
+
+        # 按模型分组统计
+        model_stats = (
+            db.query(
+                Usage.model,
+                func.count(Usage.id).label("total_requests"),
+                func.sum(Usage.input_tokens).label("input_tokens"),
+                func.sum(Usage.output_tokens).label("output_tokens"),
+                func.sum(Usage.cache_creation_input_tokens).label("cache_creation_tokens"),
+                func.sum(Usage.cache_read_input_tokens).label("cache_read_tokens"),
+                func.sum(Usage.total_cost_usd).label("total_cost"),
+                func.avg(Usage.response_time_ms).label("avg_response_time"),
+            )
+            .filter(and_(Usage.created_at >= day_start, Usage.created_at < day_end))
+            .group_by(Usage.model)
+            .all()
+        )
+
+        results = []
+        for stat in model_stats:
+            if not stat.model:
+                continue
+
+            existing = (
+                db.query(StatsDailyModel)
+                .filter(and_(StatsDailyModel.date == day_start, StatsDailyModel.model == stat.model))
+                .first()
+            )
+
+            if existing:
+                record = existing
+            else:
+                record = StatsDailyModel(
+                    id=str(uuid.uuid4()), date=day_start, model=stat.model
+                )
+
+            record.total_requests = stat.total_requests or 0
+            record.input_tokens = int(stat.input_tokens or 0)
+            record.output_tokens = int(stat.output_tokens or 0)
+            record.cache_creation_tokens = int(stat.cache_creation_tokens or 0)
+            record.cache_read_tokens = int(stat.cache_read_tokens or 0)
+            record.total_cost = float(stat.total_cost or 0)
+            record.avg_response_time_ms = float(stat.avg_response_time or 0)
+
+            if not existing:
+                db.add(record)
+            results.append(record)
+
+        db.commit()
+        logger.info(
+            f"[StatsAggregator] 聚合日期 {date.date()} 模型统计完成: {len(results)} 个模型"
+        )
+        return results
+
+    @staticmethod
+    def get_daily_model_stats(db: Session, start_date: datetime, end_date: datetime) -> list[dict]:
+        """获取日期范围内的模型统计数据（优先使用预聚合）
+
+        Args:
+            db: 数据库会话
+            start_date: 开始日期 (UTC)
+            end_date: 结束日期 (UTC)
+
+        Returns:
+            模型统计数据列表
+        """
+        from zoneinfo import ZoneInfo
+
+        app_tz = ZoneInfo(APP_TIMEZONE)
+
+        # 从预聚合表获取历史数据
+        stats = (
+            db.query(StatsDailyModel)
+            .filter(and_(StatsDailyModel.date >= start_date, StatsDailyModel.date < end_date))
+            .order_by(StatsDailyModel.date.asc(), StatsDailyModel.total_cost.desc())
+            .all()
+        )
+
+        # 转换为字典格式，按日期分组
+        result = []
+        for stat in stats:
+            # 转换日期为业务时区
+            if stat.date.tzinfo is None:
+                date_utc = stat.date.replace(tzinfo=timezone.utc)
+            else:
+                date_utc = stat.date.astimezone(timezone.utc)
+            date_str = date_utc.astimezone(app_tz).date().isoformat()
+
+            result.append({
+                "date": date_str,
+                "model": stat.model,
+                "requests": stat.total_requests,
+                "tokens": (
+                    stat.input_tokens + stat.output_tokens +
+                    stat.cache_creation_tokens + stat.cache_read_tokens
+                ),
+                "cost": stat.total_cost,
+                "avg_response_time": stat.avg_response_time_ms / 1000.0 if stat.avg_response_time_ms else 0,
+            })
+
+        return result
+
     @staticmethod
     def aggregate_user_daily_stats(
         db: Session, user_id: str, date: datetime
@@ -497,6 +612,7 @@ class StatsAggregatorService:
         current_date = start_date
         while current_date < today_local:
             StatsAggregatorService.aggregate_daily_stats(db, current_date)
+            StatsAggregatorService.aggregate_daily_model_stats(db, current_date)
             count += 1
             current_date += timedelta(days=1)
 

src/services/usage/service.py

@@ -3,6 +3,7 @@
 """
 
 import uuid
+from dataclasses import dataclass
 from datetime import datetime, timedelta, timezone
 from typing import Any, Dict, List, Optional, Tuple
 
@@ -16,6 +17,71 @@ from src.services.model.cost import ModelCostService
 from src.services.system.config import SystemConfigService
 
 
+@dataclass
+class UsageRecordParams:
+    """用量记录参数数据类，用于在内部方法间传递数据"""
+    db: Session
+    user: Optional[User]
+    api_key: Optional[ApiKey]
+    provider: str
+    model: str
+    input_tokens: int
+    output_tokens: int
+    cache_creation_input_tokens: int
+    cache_read_input_tokens: int
+    request_type: str
+    api_format: Optional[str]
+    is_stream: bool
+    response_time_ms: Optional[int]
+    first_byte_time_ms: Optional[int]
+    status_code: int
+    error_message: Optional[str]
+    metadata: Optional[Dict[str, Any]]
+    request_headers: Optional[Dict[str, Any]]
+    request_body: Optional[Any]
+    provider_request_headers: Optional[Dict[str, Any]]
+    response_headers: Optional[Dict[str, Any]]
+    response_body: Optional[Any]
+    request_id: str
+    provider_id: Optional[str]
+    provider_endpoint_id: Optional[str]
+    provider_api_key_id: Optional[str]
+    status: str
+    cache_ttl_minutes: Optional[int]
+    use_tiered_pricing: bool
+    target_model: Optional[str]
+
+    def __post_init__(self) -> None:
+        """验证关键字段，确保数据完整性"""
+        # Token 数量不能为负数
+        if self.input_tokens < 0:
+            raise ValueError(f"input_tokens 不能为负数: {self.input_tokens}")
+        if self.output_tokens < 0:
+            raise ValueError(f"output_tokens 不能为负数: {self.output_tokens}")
+        if self.cache_creation_input_tokens < 0:
+            raise ValueError(
+                f"cache_creation_input_tokens 不能为负数: {self.cache_creation_input_tokens}"
+            )
+        if self.cache_read_input_tokens < 0:
+            raise ValueError(
+                f"cache_read_input_tokens 不能为负数: {self.cache_read_input_tokens}"
+            )
+
+        # 响应时间不能为负数
+        if self.response_time_ms is not None and self.response_time_ms < 0:
+            raise ValueError(f"response_time_ms 不能为负数: {self.response_time_ms}")
+        if self.first_byte_time_ms is not None and self.first_byte_time_ms < 0:
+            raise ValueError(f"first_byte_time_ms 不能为负数: {self.first_byte_time_ms}")
+
+        # HTTP 状态码范围校验
+        if not (100 <= self.status_code <= 599):
+            raise ValueError(f"无效的 HTTP 状态码: {self.status_code}")
+
+        # 状态值校验
+        valid_statuses = {"pending", "streaming", "completed", "failed"}
+        if self.status not in valid_statuses:
+            raise ValueError(f"无效的状态值: {self.status}，有效值: {valid_statuses}")
+
 
 class UsageService:
     """用量统计服务"""
@@ -471,6 +537,97 @@ class UsageService:
             cache_ttl_minutes=cache_ttl_minutes,
         )
 
+    @classmethod
+    async def _prepare_usage_record(
+        cls,
+        params: UsageRecordParams,
+    ) -> Tuple[Dict[str, Any], float]:
+        """准备用量记录的共享逻辑
+
+        此方法提取了 record_usage 和 record_usage_async 的公共处理逻辑：
+        - 获取费率倍数
+        - 计算成本
+        - 构建 Usage 参数
+
+        Args:
+            params: 用量记录参数数据类
+
+        Returns:
+            (usage_params 字典, total_cost 总成本)
+        """
+        # 获取费率倍数和是否免费套餐
+        actual_rate_multiplier, is_free_tier = await cls._get_rate_multiplier_and_free_tier(
+            params.db, params.provider_api_key_id, params.provider_id
+        )
+
+        # 计算成本
+        is_failed_request = params.status_code >= 400 or params.error_message is not None
+        (
+            input_price, output_price, cache_creation_price, cache_read_price, request_price,
+            input_cost, output_cost, cache_creation_cost, cache_read_cost, cache_cost,
+            request_cost, total_cost, _tier_index
+        ) = await cls._calculate_costs(
+            db=params.db,
+            provider=params.provider,
+            model=params.model,
+            input_tokens=params.input_tokens,
+            output_tokens=params.output_tokens,
+            cache_creation_input_tokens=params.cache_creation_input_tokens,
+            cache_read_input_tokens=params.cache_read_input_tokens,
+            api_format=params.api_format,
+            cache_ttl_minutes=params.cache_ttl_minutes,
+            use_tiered_pricing=params.use_tiered_pricing,
+            is_failed_request=is_failed_request,
+        )
+
+        # 构建 Usage 参数
+        usage_params = cls._build_usage_params(
+            db=params.db,
+            user=params.user,
+            api_key=params.api_key,
+            provider=params.provider,
+            model=params.model,
+            input_tokens=params.input_tokens,
+            output_tokens=params.output_tokens,
+            cache_creation_input_tokens=params.cache_creation_input_tokens,
+            cache_read_input_tokens=params.cache_read_input_tokens,
+            request_type=params.request_type,
+            api_format=params.api_format,
+            is_stream=params.is_stream,
+            response_time_ms=params.response_time_ms,
+            first_byte_time_ms=params.first_byte_time_ms,
+            status_code=params.status_code,
+            error_message=params.error_message,
+            metadata=params.metadata,
+            request_headers=params.request_headers,
+            request_body=params.request_body,
+            provider_request_headers=params.provider_request_headers,
+            response_headers=params.response_headers,
+            response_body=params.response_body,
+            request_id=params.request_id,
+            provider_id=params.provider_id,
+            provider_endpoint_id=params.provider_endpoint_id,
+            provider_api_key_id=params.provider_api_key_id,
+            status=params.status,
+            target_model=params.target_model,
+            input_cost=input_cost,
+            output_cost=output_cost,
+            cache_creation_cost=cache_creation_cost,
+            cache_read_cost=cache_read_cost,
+            cache_cost=cache_cost,
+            request_cost=request_cost,
+            total_cost=total_cost,
+            input_price=input_price,
+            output_price=output_price,
+            cache_creation_price=cache_creation_price,
+            cache_read_price=cache_read_price,
+            request_price=request_price,
+            actual_rate_multiplier=actual_rate_multiplier,
+            is_free_tier=is_free_tier,
+        )
+
+        return usage_params, total_cost
+
     @classmethod
     async def record_usage_async(
         cls,
@@ -516,76 +673,25 @@ class UsageService:
         if request_id is None:
             request_id = str(uuid.uuid4())[:8]
 
-        # 获取费率倍数和是否免费套餐
-        actual_rate_multiplier, is_free_tier = await cls._get_rate_multiplier_and_free_tier(
-            db, provider_api_key_id, provider_id
-        )
-
-        # 计算成本
-        is_failed_request = status_code >= 400 or error_message is not None
-        (
-            input_price, output_price, cache_creation_price, cache_read_price, request_price,
-            input_cost, output_cost, cache_creation_cost, cache_read_cost, cache_cost,
-            request_cost, total_cost, tier_index
-        ) = await cls._calculate_costs(
-            db=db,
-            provider=provider,
-            model=model,
-            input_tokens=input_tokens,
-            output_tokens=output_tokens,
+        # 使用共享逻辑准备记录参数
+        params = UsageRecordParams(
+            db=db, user=user, api_key=api_key, provider=provider, model=model,
+            input_tokens=input_tokens, output_tokens=output_tokens,
             cache_creation_input_tokens=cache_creation_input_tokens,
             cache_read_input_tokens=cache_read_input_tokens,
-            api_format=api_format,
-            cache_ttl_minutes=cache_ttl_minutes,
-            use_tiered_pricing=use_tiered_pricing,
-            is_failed_request=is_failed_request,
-        )
-
-        # 构建 Usage 参数
-        usage_params = cls._build_usage_params(
-            db=db,
-            user=user,
-            api_key=api_key,
-            provider=provider,
-            model=model,
-            input_tokens=input_tokens,
-            output_tokens=output_tokens,
-            cache_creation_input_tokens=cache_creation_input_tokens,
-            cache_read_input_tokens=cache_read_input_tokens,
-            request_type=request_type,
-            api_format=api_format,
-            is_stream=is_stream,
-            response_time_ms=response_time_ms,
-            first_byte_time_ms=first_byte_time_ms,
-            status_code=status_code,
-            error_message=error_message,
-            metadata=metadata,
-            request_headers=request_headers,
-            request_body=request_body,
+            request_type=request_type, api_format=api_format, is_stream=is_stream,
+            response_time_ms=response_time_ms, first_byte_time_ms=first_byte_time_ms,
+            status_code=status_code, error_message=error_message, metadata=metadata,
+            request_headers=request_headers, request_body=request_body,
             provider_request_headers=provider_request_headers,
-            response_headers=response_headers,
-            response_body=response_body,
-            request_id=request_id,
-            provider_id=provider_id,
+            response_headers=response_headers, response_body=response_body,
+            request_id=request_id, provider_id=provider_id,
             provider_endpoint_id=provider_endpoint_id,
-            provider_api_key_id=provider_api_key_id,
-            status=status,
+            provider_api_key_id=provider_api_key_id, status=status,
+            cache_ttl_minutes=cache_ttl_minutes, use_tiered_pricing=use_tiered_pricing,
             target_model=target_model,
-            input_cost=input_cost,
-            output_cost=output_cost,
-            cache_creation_cost=cache_creation_cost,
-            cache_read_cost=cache_read_cost,
-            cache_cost=cache_cost,
-            request_cost=request_cost,
-            total_cost=total_cost,
-            input_price=input_price,
-            output_price=output_price,
-            cache_creation_price=cache_creation_price,
-            cache_read_price=cache_read_price,
-            request_price=request_price,
-            actual_rate_multiplier=actual_rate_multiplier,
-            is_free_tier=is_free_tier,
         )
+        usage_params, _ = await cls._prepare_usage_record(params)
 
         # 创建 Usage 记录
         usage = Usage(**usage_params)
@@ -660,76 +766,25 @@ class UsageService:
         if request_id is None:
             request_id = str(uuid.uuid4())[:8]
 
-        # 获取费率倍数和是否免费套餐
-        actual_rate_multiplier, is_free_tier = await cls._get_rate_multiplier_and_free_tier(
-            db, provider_api_key_id, provider_id
-        )
-
-        # 计算成本
-        is_failed_request = status_code >= 400 or error_message is not None
-        (
-            input_price, output_price, cache_creation_price, cache_read_price, request_price,
-            input_cost, output_cost, cache_creation_cost, cache_read_cost, cache_cost,
-            request_cost, total_cost, _tier_index
-        ) = await cls._calculate_costs(
-            db=db,
-            provider=provider,
-            model=model,
-            input_tokens=input_tokens,
-            output_tokens=output_tokens,
+        # 使用共享逻辑准备记录参数
+        params = UsageRecordParams(
+            db=db, user=user, api_key=api_key, provider=provider, model=model,
+            input_tokens=input_tokens, output_tokens=output_tokens,
             cache_creation_input_tokens=cache_creation_input_tokens,
             cache_read_input_tokens=cache_read_input_tokens,
-            api_format=api_format,
-            cache_ttl_minutes=cache_ttl_minutes,
-            use_tiered_pricing=use_tiered_pricing,
-            is_failed_request=is_failed_request,
-        )
-
-        # 构建 Usage 参数
-        usage_params = cls._build_usage_params(
-            db=db,
-            user=user,
-            api_key=api_key,
-            provider=provider,
-            model=model,
-            input_tokens=input_tokens,
-            output_tokens=output_tokens,
-            cache_creation_input_tokens=cache_creation_input_tokens,
-            cache_read_input_tokens=cache_read_input_tokens,
-            request_type=request_type,
-            api_format=api_format,
-            is_stream=is_stream,
-            response_time_ms=response_time_ms,
-            first_byte_time_ms=first_byte_time_ms,
-            status_code=status_code,
-            error_message=error_message,
-            metadata=metadata,
-            request_headers=request_headers,
-            request_body=request_body,
+            request_type=request_type, api_format=api_format, is_stream=is_stream,
+            response_time_ms=response_time_ms, first_byte_time_ms=first_byte_time_ms,
+            status_code=status_code, error_message=error_message, metadata=metadata,
+            request_headers=request_headers, request_body=request_body,
             provider_request_headers=provider_request_headers,
-            response_headers=response_headers,
-            response_body=response_body,
-            request_id=request_id,
-            provider_id=provider_id,
+            response_headers=response_headers, response_body=response_body,
+            request_id=request_id, provider_id=provider_id,
             provider_endpoint_id=provider_endpoint_id,
-            provider_api_key_id=provider_api_key_id,
-            status=status,
+            provider_api_key_id=provider_api_key_id, status=status,
+            cache_ttl_minutes=cache_ttl_minutes, use_tiered_pricing=use_tiered_pricing,
             target_model=target_model,
-            input_cost=input_cost,
-            output_cost=output_cost,
-            cache_creation_cost=cache_creation_cost,
-            cache_read_cost=cache_read_cost,
-            cache_cost=cache_cost,
-            request_cost=request_cost,
-            total_cost=total_cost,
-            input_price=input_price,
-            output_price=output_price,
-            cache_creation_price=cache_creation_price,
-            cache_read_price=cache_read_price,
-            request_price=request_price,
-            actual_rate_multiplier=actual_rate_multiplier,
-            is_free_tier=is_free_tier,
         )
+        usage_params, total_cost = await cls._prepare_usage_record(params)
 
         # 检查是否已存在相同 request_id 的记录
         existing_usage = db.query(Usage).filter(Usage.request_id == request_id).first()
@@ -751,7 +806,7 @@ class UsageService:
             api_key = db.merge(api_key)
 
         # 使用原子更新避免并发竞态条件
-        from sqlalchemy import func, update
+        from sqlalchemy import func as sql_func, update
         from src.models.database import ApiKey as ApiKeyModel, User as UserModel, GlobalModel
 
         # 更新用户使用量（独立 Key 不计入创建者的使用记录）
@@ -762,7 +817,7 @@ class UsageService:
                 .values(
                     used_usd=UserModel.used_usd + total_cost,
                     total_usd=UserModel.total_usd + total_cost,
-                    updated_at=func.now(),
+                    updated_at=sql_func.now(),
                 )
             )
 
@@ -776,8 +831,8 @@ class UsageService:
                         total_requests=ApiKeyModel.total_requests + 1,
                         total_cost_usd=ApiKeyModel.total_cost_usd + total_cost,
                         balance_used_usd=ApiKeyModel.balance_used_usd + total_cost,
-                        last_used_at=func.now(),
-                        updated_at=func.now(),
+                        last_used_at=sql_func.now(),
+                        updated_at=sql_func.now(),
                     )
                 )
             else:
@@ -787,8 +842,8 @@ class UsageService:
                     .values(
                         total_requests=ApiKeyModel.total_requests + 1,
                         total_cost_usd=ApiKeyModel.total_cost_usd + total_cost,
-                        last_used_at=func.now(),
-                        updated_at=func.now(),
+                        last_used_at=sql_func.now(),
+                        updated_at=sql_func.now(),
                     )
                 )
 
@@ -1121,19 +1176,48 @@ class UsageService:
         ]
 
     @staticmethod
-    def cleanup_old_usage_records(db: Session, days_to_keep: int = 90) -> int:
-        """清理旧的使用记录"""
+    def cleanup_old_usage_records(
+        db: Session, days_to_keep: int = 90, batch_size: int = 1000
+    ) -> int:
+        """清理旧的使用记录（分批删除避免长事务锁定）
 
+        Args:
+            db: 数据库会话
+            days_to_keep: 保留天数，默认 90 天
+            batch_size: 每批删除数量，默认 1000 条
+
+        Returns:
+            删除的总记录数
+        """
         cutoff_date = datetime.now(timezone.utc) - timedelta(days=days_to_keep)
+        total_deleted = 0
 
-        # 删除旧记录
-        deleted = db.query(Usage).filter(Usage.created_at < cutoff_date).delete()
+        while True:
+            # 查询待删除的 ID（使用新索引 idx_usage_user_created）
+            batch_ids = (
+                db.query(Usage.id)
+                .filter(Usage.created_at < cutoff_date)
+                .limit(batch_size)
+                .all()
+            )
 
-        db.commit()
+            if not batch_ids:
+                break
 
-        logger.info(f"清理使用记录: 删除 {deleted} 条超过 {days_to_keep} 天的记录")
+            # 批量删除
+            deleted_count = (
+                db.query(Usage)
+                .filter(Usage.id.in_([row.id for row in batch_ids]))
+                .delete(synchronize_session=False)
+            )
+            db.commit()
+            total_deleted += deleted_count
 
-        return deleted
+            logger.debug(f"清理使用记录: 本批删除 {deleted_count} 条")
+
+        logger.info(f"清理使用记录: 共删除 {total_deleted} 条超过 {days_to_keep} 天的记录")
+
+        return total_deleted
 
     # ========== 请求状态追踪方法 ==========
 
@@ -1219,6 +1303,7 @@ class UsageService:
         error_message: Optional[str] = None,
         provider: Optional[str] = None,
         target_model: Optional[str] = None,
+        first_byte_time_ms: Optional[int] = None,
     ) -> Optional[Usage]:
         """
         快速更新使用记录状态
@@ -1230,6 +1315,7 @@ class UsageService:
             error_message: 错误消息（仅在 failed 状态时使用）
             provider: 提供商名称（可选，streaming 状态时更新）
             target_model: 映射后的目标模型名（可选）
+            first_byte_time_ms: 首字时间/TTFB（可选，streaming 状态时更新）
 
         Returns:
             更新后的 Usage 记录，如果未找到则返回 None
@@ -1247,6 +1333,8 @@ class UsageService:
             usage.provider = provider
         if target_model:
             usage.target_model = target_model
+        if first_byte_time_ms is not None:
+            usage.first_byte_time_ms = first_byte_time_ms
 
         db.commit()
 

src/services/usage/stream.py

@@ -5,6 +5,7 @@
 
 import json
 import re
+import time
 from typing import Any, AsyncIterator, Dict, Optional, Tuple
 
 from sqlalchemy.orm import Session
@@ -457,26 +458,32 @@ class StreamUsageTracker:
 
         logger.debug(f"ID:{self.request_id} | 开始跟踪流式响应 | 估算输入tokens:{self.input_tokens}")
 
-        # 更新状态为 streaming，同时更新 provider
-        if self.request_id:
-            try:
-                from src.services.usage.service import UsageService
-                UsageService.update_usage_status(
-                    db=self.db,
-                    request_id=self.request_id,
-                    status="streaming",
-                    provider=self.provider,
-                )
-            except Exception as e:
-                logger.warning(f"更新使用记录状态为 streaming 失败: {e}")
-
         chunk_count = 0
+        first_chunk_received = False
         try:
             async for chunk in stream:
                 chunk_count += 1
                 # 保存原始字节流（用于错误诊断）
                 self.raw_chunks.append(chunk)
 
+                # 第一个 chunk 收到时，更新状态为 streaming 并记录 TTFB
+                if not first_chunk_received:
+                    first_chunk_received = True
+                    if self.request_id:
+                        try:
+                            # 计算 TTFB（使用请求原始开始时间或 track_stream 开始时间）
+                            base_time = self.request_start_time or self.start_time
+                            first_byte_time_ms = int((time.time() - base_time) * 1000) if base_time else None
+                            UsageService.update_usage_status(
+                                db=self.db,
+                                request_id=self.request_id,
+                                status="streaming",
+                                provider=self.provider,
+                                first_byte_time_ms=first_byte_time_ms,
+                            )
+                        except Exception as e:
+                            logger.warning(f"更新使用记录状态为 streaming 失败: {e}")
+
                 # 返回原始块给客户端
                 yield chunk
 

src/services/user/apikey.py

@@ -59,14 +59,15 @@ class ApiKeyService:
         if expire_days:
             expires_at = datetime.now(timezone.utc) + timedelta(days=expire_days)
 
+        # 空数组转为 None（表示不限制）
         api_key = ApiKey(
             user_id=user_id,
             key_hash=key_hash,
             key_encrypted=key_encrypted,
             name=name or f"API Key {datetime.now(timezone.utc).strftime('%Y%m%d%H%M%S')}",
-            allowed_providers=allowed_providers,
-            allowed_api_formats=allowed_api_formats,
-            allowed_models=allowed_models,
+            allowed_providers=allowed_providers or None,
+            allowed_api_formats=allowed_api_formats or None,
+            allowed_models=allowed_models or None,
             rate_limit=rate_limit,
             concurrent_limit=concurrent_limit,
             expires_at=expires_at,
@@ -141,8 +142,18 @@ class ApiKeyService:
             "auto_delete_on_expiry",
         ]
 
+        # 允许显式设置为空数组/None 的字段（空数组会转为 None，表示"全部"）
+        nullable_list_fields = {"allowed_providers", "allowed_api_formats", "allowed_models"}
+
         for field, value in kwargs.items():
-            if field in updatable_fields and value is not None:
+            if field not in updatable_fields:
+                continue
+            # 对于 nullable_list_fields，空数组应该转为 None（表示不限制）
+            if field in nullable_list_fields:
+                if value is not None:
+                    # 空数组转为 None（表示允许全部）
+                    setattr(api_key, field, value if value else None)
+            elif value is not None:
                 setattr(api_key, field, value)
 
         api_key.updated_at = datetime.now(timezone.utc)

src/utils/task_coordinator.py

@@ -1,8 +1,16 @@
-"""分布式任务协调器，确保仅有一个 worker 执行特定任务"""
+"""分布式任务协调器，确保仅有一个 worker 执行特定任务
+
+锁清理策略：
+- 单实例模式（默认）：启动时使用原子操作清理旧锁并获取新锁
+- 多实例模式：使用 NX 选项竞争锁，依赖 TTL 处理异常退出
+
+使用方式：
+- 默认行为：启动时清理旧锁（适用于单机部署）
+- 多实例部署：设置 SINGLE_INSTANCE_MODE=false 禁用启动清理
+"""
 
 from __future__ import annotations
 
-import asyncio
 import os
 import pathlib
 import uuid
@@ -19,6 +27,10 @@ except ImportError:  # pragma: no cover - Windows 环境
 class StartupTaskCoordinator:
     """利用 Redis 或文件锁，保证任务只在单个进程/实例中运行"""
 
+    # 类级别标记：在当前进程中是否已尝试过启动清理
+    # 注意：这在 fork 模式下每个 worker 都是独立的
+    _startup_cleanup_attempted = False
+
     def __init__(self, redis_client=None, lock_dir: Optional[str] = None):
         self.redis = redis_client
         self._tokens: Dict[str, str] = {}
@@ -26,6 +38,8 @@ class StartupTaskCoordinator:
         self._lock_dir = pathlib.Path(lock_dir or os.getenv("TASK_LOCK_DIR", "./.locks"))
         if not self._lock_dir.exists():
             self._lock_dir.mkdir(parents=True, exist_ok=True)
+        # 单实例模式：启动时清理旧锁（适用于单机部署，避免残留锁问题）
+        self._single_instance_mode = os.getenv("SINGLE_INSTANCE_MODE", "true").lower() == "true"
 
     def _redis_key(self, name: str) -> str:
         return f"task_lock:{name}"
@@ -36,12 +50,51 @@ class StartupTaskCoordinator:
         if self.redis:
             token = str(uuid.uuid4())
             try:
-                acquired = await self.redis.set(self._redis_key(name), token, nx=True, ex=ttl)
-                if acquired:
-                    self._tokens[name] = token
-                    logger.info(f"任务 {name} 通过 Redis 锁独占执行")
-                    return True
-                return False
+                if self._single_instance_mode:
+                    # 单实例模式：使用 Lua 脚本原子性地"清理旧锁 + 竞争获取"
+                    # 只有当锁不存在或成功获取时才返回 1
+                    # 这样第一个执行的 worker 会清理旧锁并获取，后续 worker 会正常竞争
+                    script = """
+                    local key = KEYS[1]
+                    local token = ARGV[1]
+                    local ttl = tonumber(ARGV[2])
+                    local startup_key = KEYS[1] .. ':startup'
+
+                    -- 检查是否已有 worker 执行过启动清理
+                    local cleaned = redis.call('GET', startup_key)
+                    if not cleaned then
+                        -- 第一个 worker：删除旧锁，标记已清理
+                        redis.call('DEL', key)
+                        redis.call('SET', startup_key, '1', 'EX', 60)
+                    end
+
+                    -- 尝试获取锁（NX 模式）
+                    local result = redis.call('SET', key, token, 'NX', 'EX', ttl)
+                    if result then
+                        return 1
+                    end
+                    return 0
+                    """
+                    result = await self.redis.eval(
+                        script, 2,
+                        self._redis_key(name), self._redis_key(name),
+                        token, ttl
+                    )
+                    if result == 1:
+                        self._tokens[name] = token
+                        logger.info(f"任务 {name} 通过 Redis 锁独占执行")
+                        return True
+                    return False
+                else:
+                    # 多实例模式：直接使用 NX 选项竞争锁
+                    acquired = await self.redis.set(
+                        self._redis_key(name), token, nx=True, ex=ttl
+                    )
+                    if acquired:
+                        self._tokens[name] = token
+                        logger.info(f"任务 {name} 通过 Redis 锁独占执行")
+                        return True
+                    return False
             except Exception as exc:  # pragma: no cover - Redis 异常回退
                 logger.warning(f"Redis 锁获取失败，回退到文件锁: {exc}")
 

src/utils/timeout.py

@@ -139,3 +139,83 @@ async def with_timeout_context(timeout: float, operation_name: str = "operation"
         # Python 3.10 及以下版本的兼容实现
         # 注意：这个简单实现不支持嵌套取消
         pass
+
+
+async def read_first_chunk_with_ttfb_timeout(
+    byte_iterator: Any,
+    timeout: float,
+    request_id: str,
+    provider_name: str,
+) -> tuple[bytes, Any]:
+    """
+    读取流的首字节并应用 TTFB 超时检测
+
+    首字节超时（Time To First Byte）用于检测慢响应的 Provider，
+    超时时触发故障转移到其他可用的 Provider。
+
+    Args:
+        byte_iterator: 异步字节流迭代器
+        timeout: TTFB 超时时间（秒）
+        request_id: 请求 ID（用于日志）
+        provider_name: Provider 名称（用于日志和异常）
+
+    Returns:
+        (first_chunk, aiter): 首个字节块和异步迭代器
+
+    Raises:
+        ProviderTimeoutException: 如果首字节超时
+    """
+    from src.core.exceptions import ProviderTimeoutException
+
+    aiter = byte_iterator.__aiter__()
+
+    try:
+        first_chunk = await asyncio.wait_for(aiter.__anext__(), timeout=timeout)
+        return first_chunk, aiter
+    except asyncio.TimeoutError:
+        # 完整的资源清理：先关闭迭代器，再关闭底层响应
+        await _cleanup_iterator_resources(aiter, request_id)
+        logger.warning(
+            f"  [{request_id}] 流首字节超时 (TTFB): "
+            f"Provider={provider_name}, timeout={timeout}s"
+        )
+        raise ProviderTimeoutException(
+            provider_name=provider_name,
+            timeout=int(timeout),
+        )
+
+
+async def _cleanup_iterator_resources(aiter: Any, request_id: str) -> None:
+    """
+    清理异步迭代器及其底层资源
+
+    确保在 TTFB 超时后正确释放 HTTP 连接，避免连接泄漏。
+
+    Args:
+        aiter: 异步迭代器
+        request_id: 请求 ID（用于日志）
+    """
+    # 1. 关闭迭代器本身
+    if hasattr(aiter, "aclose"):
+        try:
+            await aiter.aclose()
+        except Exception as e:
+            logger.debug(f"  [{request_id}] 关闭迭代器失败: {e}")
+
+    # 2. 关闭底层响应对象（httpx.Response）
+    # 迭代器可能持有 _response 属性指向底层响应
+    response = getattr(aiter, "_response", None)
+    if response is not None and hasattr(response, "aclose"):
+        try:
+            await response.aclose()
+        except Exception as e:
+            logger.debug(f"  [{request_id}] 关闭底层响应失败: {e}")
+
+    # 3. 尝试关闭 httpx 流（如果迭代器是 httpx 的 aiter_bytes）
+    # httpx 的 Response.aiter_bytes() 返回的生成器可能有 _stream 属性
+    stream = getattr(aiter, "_stream", None)
+    if stream is not None and hasattr(stream, "aclose"):
+        try:
+            await stream.aclose()
+        except Exception as e:
+            logger.debug(f"  [{request_id}] 关闭流对象失败: {e}")