fix: 修正 README 中的图片文件扩展名

- 本月缓存命中率改为使用本月数据计算（monthly_input_tokens + cache_read_tokens），而非全时数据
- 修复配额显示：有配额时显示实际金额，无配额时显示为 /bin/zsh 并标记为高警告状态

.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

@@ -1,122 +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 $http_x_api_key;' \
-'        proxy_buffering off;' \
-'        proxy_cache off;' \
-'        proxy_request_buffering off;' \
-'        chunked_transfer_encoding on;' \
-'        proxy_connect_timeout 600s;' \
-'        proxy_send_timeout 600s;' \
-'        proxy_read_timeout 600s;' \
-'    }' \
-'}' > /etc/nginx/sites-available/default.template
-
-# Supervisor 配置
-RUN printf '%s\n' \
-'[supervisord]' \
-'nodaemon=true' \
-'logfile=/var/log/supervisor/supervisord.log' \
-'pidfile=/var/run/supervisord.pid' \
-'' \
-'[program:nginx]' \
-'command=/bin/bash -c "sed \"s/PORT_PLACEHOLDER/${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/*
@@ -20,107 +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 $http_x_api_key;' \
-'        proxy_buffering off;' \
-'        proxy_cache off;' \
-'        proxy_request_buffering off;' \
-'        chunked_transfer_encoding on;' \
-'        proxy_connect_timeout 600s;' \
-'        proxy_send_timeout 600s;' \
-'        proxy_read_timeout 600s;' \
-'    }' \
-'}' > /etc/nginx/sites-available/default.template
-
-# Supervisor 配置
-RUN printf '%s\n' \
-'[supervisord]' \
-'nodaemon=true' \
-'logfile=/var/log/supervisor/supervisord.log' \
-'pidfile=/var/run/supervisord.pid' \
-'' \
-'[program:nginx]' \
-'command=/bin/bash -c "sed \"s/PORT_PLACEHOLDER/${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

LICENSE

@@ -5,12 +5,17 @@ Aether 非商业开源许可证
 特此授予任何获得本软件及其相关文档文件（以下简称"软件"）副本的人免费使用、
 复制、修改、合并、发布和分发本软件的权限，但须遵守以下条件：
 
-1. 仅限非商业用途
-   本软件不得用于商业目的。商业目的包括但不限于：
+1. 仅限非盈利用途
+   本软件不得用于盈利目的。盈利目的包括但不限于：
    - 出售本软件或任何衍生作品
    - 使用本软件提供付费服务
-   - 将本软件用于商业产品或服务
-   - 将本软件用于任何旨在获取商业利益或金钱报酬的活动
+   - 将本软件用于以盈利为目的的商业产品或服务
+
+   以下用途被明确允许：
+   - 个人学习和研究
+   - 教育机构的教学和研究
+   - 非盈利组织的内部使用
+   - 企业内部非盈利性质的使用（如内部工具、测试环境等）
 
 2. 署名要求
    上述版权声明和本许可声明应包含在本软件的所有副本或主要部分中。
@@ -22,7 +27,7 @@ Aether 非商业开源许可证
    您不得以不同的条款将本软件再许可给他人。
 
 5. 商业许可
-   如需商业使用，请联系版权持有人以获取单独的商业许可。
+   如需将本软件用于盈利目的，请联系版权持有人以获取单独的商业许可。
 
 本软件按"原样"提供，不提供任何明示或暗示的保证，包括但不限于对适销性、
 特定用途适用性和非侵权性的保证。在任何情况下，作者或版权持有人均不对任何

README.md

@@ -143,7 +143,7 @@ cd frontend && npm install && npm run dev
 - **模型级别**: 在模型管理中针对指定模型开启 1H缓存策略
 - **密钥级别**: 在密钥管理中针对指定密钥使用 1H缓存策略
 
-> **注意**: 若对密钥设置强制 1H缓存, 则该密钥只能调用支持 1H缓存的模型
+> **注意**: 若对密钥设置强制 1H缓存, 则该密钥只能使用支持 1H缓存的模型, 匹配提供商Key, 将会导致这个Key无法同时用于Claude Code、Codex、GeminiCLI, 因为更推荐使用模型开启1H缓存.
 
 ### Q: 如何配置负载均衡？
 
@@ -162,4 +162,16 @@ cd frontend && npm install && npm run dev
 
 ## 许可证
 
-本项目采用 [Aether 非商业开源许可证](LICENSE)。
+本项目采用 [Aether 非商业开源许可证](LICENSE)。允许个人学习、教育研究、非盈利组织及企业内部非盈利性质的使用；禁止用于盈利目的。商业使用请联系获取商业许可。
+
+## 联系作者
+
+<p align="center">
+  <img src="docs/author/qq_qrcode.jpg" width="200" alt="QQ二维码">
+</p>
+
+## Star History
+
+[![Star History Chart](https://api.star-history.com/svg?repos=fawney19/Aether&type=Date)](https://star-history.com/#fawney19/Aether&Date)
+
+

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/20251218_0631_f30f9936f6a2_add_proxy_field_to_provider_endpoints.py

@@ -0,0 +1,57 @@
+"""add proxy field to provider_endpoints
+
+Revision ID: f30f9936f6a2
+Revises: 1cc6942cf06f
+Create Date: 2025-12-18 06:31:58.451112+00:00
+
+"""
+from alembic import op
+import sqlalchemy as sa
+from sqlalchemy.dialects.postgresql import JSONB
+from sqlalchemy import inspect
+
+# revision identifiers, used by Alembic.
+revision = 'f30f9936f6a2'
+down_revision = '1cc6942cf06f'
+branch_labels = None
+depends_on = None
+
+
+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 get_column_type(table_name: str, column_name: str) -> str:
+    """获取列的类型"""
+    bind = op.get_bind()
+    inspector = inspect(bind)
+    for col in inspector.get_columns(table_name):
+        if col['name'] == column_name:
+            return str(col['type']).upper()
+    return ''
+
+
+def upgrade() -> None:
+    """添加 proxy 字段到 provider_endpoints 表"""
+    if not column_exists('provider_endpoints', 'proxy'):
+        # 字段不存在，直接添加 JSONB 类型
+        op.add_column('provider_endpoints', sa.Column('proxy', JSONB(), nullable=True))
+    else:
+        # 字段已存在，检查是否需要转换类型
+        col_type = get_column_type('provider_endpoints', 'proxy')
+        if 'JSONB' not in col_type:
+            # 如果是 JSON 类型，转换为 JSONB
+            op.execute(
+                'ALTER TABLE provider_endpoints '
+                'ALTER COLUMN proxy TYPE JSONB USING proxy::jsonb'
+            )
+
+
+def downgrade() -> None:
+    """移除 proxy 字段"""
+    if column_exists('provider_endpoints', 'proxy'):
+        op.drop_column('provider_endpoints', 'proxy')

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

@@ -21,15 +21,18 @@ HASH_FILE=".deps-hash"
 CODE_HASH_FILE=".code-hash"
 MIGRATION_HASH_FILE=".migration-hash"
 
-# 计算依赖文件的哈希值
+# 计算依赖文件的哈希值（包含 Dockerfile.base.local）
 calc_deps_hash() {
-    cat pyproject.toml frontend/package.json frontend/package-lock.json 2>/dev/null | md5sum | cut -d' ' -f1
+    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
 }
 
@@ -162,29 +165,46 @@ git pull
 
 # 标记是否需要重启
 NEED_RESTART=false
+BASE_REBUILT=false
 
 # 检查基础镜像是否存在，或依赖是否变化
 if ! docker image inspect aether-base:latest >/dev/null 2>&1; then
     echo ">>> Base image not found, building..."
     build_base
+    BASE_REBUILT=true
     NEED_RESTART=true
 elif check_deps_changed; then
     echo ">>> Dependencies changed, rebuilding base image..."
     build_base
+    BASE_REBUILT=true
     NEED_RESTART=true
 else
     echo ">>> Dependencies unchanged."
 fi
 
-# 检查代码是否变化
+# 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
     NEED_RESTART=true
+elif [ "$BASE_REBUILT" = true ]; then
+    echo ">>> Base image rebuilt, rebuilding app image..."
+    build_app
+    NEED_RESTART=true
 elif check_code_changed; then
     echo ">>> Code changed, rebuilding app image..."
     build_app
     NEED_RESTART=true
+elif [ "$MIGRATION_CHANGED" = true ]; then
+    echo ">>> Migration files changed, rebuilding app image..."
+    build_app
+    NEED_RESTART=true
 else
     echo ">>> Code unchanged."
 fi
@@ -197,9 +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/dashboard.ts

@@ -87,6 +87,8 @@ export interface DashboardStatsResponse {
   cache_stats?: CacheStats
   users?: UserStats
   token_breakdown?: TokenBreakdown
+  // 普通用户专用字段
+  monthly_cost?: number
 }
 
 export interface RecentRequestsResponse {

frontend/src/api/endpoints/endpoints.ts

@@ -1,5 +1,5 @@
 import client from '../client'
-import type { ProviderEndpoint } from './types'
+import type { ProviderEndpoint, ProxyConfig } from './types'
 
 /**
  * 获取指定 Provider 的所有 Endpoints
@@ -38,6 +38,7 @@ export async function createEndpoint(
     rate_limit?: number
     is_active?: boolean
     config?: Record<string, any>
+    proxy?: ProxyConfig | null
   }
 ): Promise<ProviderEndpoint> {
   const response = await client.post(`/api/admin/endpoints/providers/${providerId}/endpoints`, data)
@@ -63,6 +64,7 @@ export async function updateEndpoint(
     rate_limit: number
     is_active: boolean
     config: Record<string, any>
+    proxy: ProxyConfig | null
   }>
 ): Promise<ProviderEndpoint> {
   const response = await client.put(`/api/admin/endpoints/${endpointId}`, data)

frontend/src/api/endpoints/keys.ts

@@ -110,6 +110,14 @@ export async function updateEndpointKey(
   return response.data
 }
 
+/**
+ * 获取完整的 API Key（用于查看和复制）
+ */
+export async function revealEndpointKey(keyId: string): Promise<{ api_key: string }> {
+  const response = await client.get(`/api/admin/endpoints/keys/${keyId}/reveal`)
+  return response.data
+}
+
 /**
  * 删除 Endpoint Key
  */

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

@@ -20,6 +20,16 @@ export const API_FORMAT_LABELS: Record<string, string> = {
   [API_FORMATS.GEMINI_CLI]: 'Gemini CLI',
 }
 
+/**
+ * 代理配置类型
+ */
+export interface ProxyConfig {
+  url: string
+  username?: string
+  password?: string
+  enabled?: boolean  // 是否启用代理（false 时保留配置但不使用）
+}
+
 export interface ProviderEndpoint {
   id: string
   provider_id: string
@@ -41,6 +51,7 @@ export interface ProviderEndpoint {
   last_failure_at?: string
   is_active: boolean
   config?: Record<string, any>
+  proxy?: ProxyConfig | null
   total_keys: number
   active_keys: number
   created_at: string
@@ -99,6 +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
@@ -233,18 +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
@@ -274,7 +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
@@ -291,7 +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
@@ -484,3 +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

@@ -22,7 +22,7 @@
         />
       </Transition>
 
-      <div class="relative flex min-h-full items-end justify-center p-4 text-center sm:items-center sm:p-0">
+      <div class="relative flex min-h-full items-end justify-center p-4 text-center sm:items-center sm:p-0 pointer-events-none">
         <!-- 对话框内容 -->
         <Transition
           enter-active-class="duration-300 ease-out"
@@ -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/api-keys/components/StandaloneKeyFormDialog.vue

@@ -132,7 +132,7 @@
               type="number"
               min="1"
               max="10000"
-              placeholder="100"
+              placeholder="留空不限制"
               class="h-10"
               @update:model-value="(v) => form.rate_limit = parseNumberInput(v, { min: 1, max: 10000 })"
             />
@@ -376,7 +376,7 @@ const form = ref<StandaloneKeyFormData>({
   initial_balance_usd: 10,
   expire_days: undefined,
   never_expire: true,
-  rate_limit: 100,
+  rate_limit: undefined,
   auto_delete_on_expiry: false,
   allowed_providers: [],
   allowed_api_formats: [],
@@ -389,7 +389,7 @@ function resetForm() {
     initial_balance_usd: 10,
     expire_days: undefined,
     never_expire: true,
-    rate_limit: 100,
+    rate_limit: undefined,
     auto_delete_on_expiry: false,
     allowed_providers: [],
     allowed_api_formats: [],
@@ -408,7 +408,7 @@ function loadKeyData() {
     initial_balance_usd: props.apiKey.initial_balance_usd,
     expire_days: props.apiKey.expire_days,
     never_expire: props.apiKey.never_expire,
-    rate_limit: props.apiKey.rate_limit || 100,
+    rate_limit: props.apiKey.rate_limit,
     auto_delete_on_expiry: props.apiKey.auto_delete_on_expiry,
     allowed_providers: props.apiKey.allowed_providers || [],
     allowed_api_formats: props.apiKey.allowed_api_formats || [],

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/EndpointFormDialog.vue

@@ -9,7 +9,7 @@
   >
     <form
       class="space-y-6"
-      @submit.prevent="handleSubmit"
+      @submit.prevent="handleSubmit()"
     >
       <!-- API 配置 -->
       <div class="space-y-4">
@@ -132,6 +132,79 @@
           </div>
         </div>
       </div>
+
+      <!-- 代理配置 -->
+      <div class="space-y-4">
+        <div class="flex items-center justify-between">
+          <h3 class="text-sm font-medium">
+            代理配置
+          </h3>
+          <div class="flex items-center gap-2">
+            <Switch v-model="proxyEnabled" />
+            <span class="text-sm text-muted-foreground">启用代理</span>
+          </div>
+        </div>
+
+        <div
+          v-if="proxyEnabled"
+          class="space-y-4 rounded-lg border p-4"
+        >
+          <div class="space-y-2">
+            <Label for="proxy_url">代理 URL *</Label>
+            <Input
+              id="proxy_url"
+              v-model="form.proxy_url"
+              placeholder="http://host:port 或 socks5://host:port"
+              required
+              :class="proxyUrlError ? 'border-red-500' : ''"
+            />
+            <p
+              v-if="proxyUrlError"
+              class="text-xs text-red-500"
+            >
+              {{ proxyUrlError }}
+            </p>
+            <p
+              v-else
+              class="text-xs text-muted-foreground"
+            >
+              支持 HTTP、HTTPS、SOCKS5 代理
+            </p>
+          </div>
+
+          <div class="grid grid-cols-2 gap-4">
+            <div class="space-y-2">
+              <Label for="proxy_user">用户名（可选）</Label>
+              <Input
+                :id="`proxy_user_${formId}`"
+                :name="`proxy_user_${formId}`"
+                v-model="form.proxy_username"
+                placeholder="代理认证用户名"
+                autocomplete="off"
+                data-form-type="other"
+                data-lpignore="true"
+                data-1p-ignore="true"
+              />
+            </div>
+
+            <div class="space-y-2">
+              <Label :for="`proxy_pass_${formId}`">密码（可选）</Label>
+              <Input
+                :id="`proxy_pass_${formId}`"
+                :name="`proxy_pass_${formId}`"
+                v-model="form.proxy_password"
+                type="text"
+                :placeholder="passwordPlaceholder"
+                autocomplete="off"
+                data-form-type="other"
+                data-lpignore="true"
+                data-1p-ignore="true"
+                :style="{ '-webkit-text-security': 'disc', 'text-security': 'disc' }"
+              />
+            </div>
+          </div>
+        </div>
+      </div>
     </form>
 
     <template #footer>
@@ -145,12 +218,24 @@
       </Button>
       <Button
         :disabled="loading || !form.base_url || (!isEditMode && !form.api_format)"
-        @click="handleSubmit"
+        @click="handleSubmit()"
       >
         {{ loading ? (isEditMode ? '保存中...' : '创建中...') : (isEditMode ? '保存修改' : '创建') }}
       </Button>
     </template>
   </Dialog>
+
+  <!-- 确认清空凭据对话框 -->
+  <AlertDialog
+    v-model="showClearCredentialsDialog"
+    title="清空代理凭据"
+    description="代理 URL 为空，但用户名和密码仍有值。是否清空这些凭据并继续保存？"
+    type="warning"
+    confirm-text="清空并保存"
+    cancel-text="返回编辑"
+    @confirm="confirmClearCredentials"
+    @cancel="showClearCredentialsDialog = false"
+  />
 </template>
 
 <script setup lang="ts">
@@ -165,7 +250,9 @@ import {
   SelectValue,
   SelectContent,
   SelectItem,
+  Switch,
 } from '@/components/ui'
+import AlertDialog from '@/components/common/AlertDialog.vue'
 import { Link, SquarePen } from 'lucide-vue-next'
 import { useToast } from '@/composables/useToast'
 import { useFormDialog } from '@/composables/useFormDialog'
@@ -194,6 +281,11 @@ const emit = defineEmits<{
 const { success, error: showError } = useToast()
 const loading = ref(false)
 const selectOpen = ref(false)
+const proxyEnabled = ref(false)
+const showClearCredentialsDialog = ref(false)  // 确认清空凭据对话框
+
+// 生成随机 ID 防止浏览器自动填充
+const formId = Math.random().toString(36).substring(2, 10)
 
 // 内部状态
 const internalOpen = computed(() => props.modelValue)
@@ -207,7 +299,11 @@ const form = ref({
   max_retries: 3,
   max_concurrent: undefined as number | undefined,
   rate_limit: undefined as number | undefined,
-  is_active: true
+  is_active: true,
+  // 代理配置
+  proxy_url: '',
+  proxy_username: '',
+  proxy_password: '',
 })
 
 // API 格式列表
@@ -237,6 +333,53 @@ const defaultPathPlaceholder = computed(() => {
   return `留空使用默认路径：${defaultPath.value}`
 })
 
+// 检查是否有已保存的密码（后端返回 *** 表示有密码）
+const hasExistingPassword = computed(() => {
+  if (!props.endpoint?.proxy) return false
+  const proxy = props.endpoint.proxy as { password?: string }
+  return proxy?.password === MASKED_PASSWORD
+})
+
+// 密码输入框的 placeholder
+const passwordPlaceholder = computed(() => {
+  if (hasExistingPassword.value) {
+    return '已保存密码，留空保持不变'
+  }
+  return '代理认证密码'
+})
+
+// 代理 URL 验证
+const proxyUrlError = computed(() => {
+  // 只有启用代理且填写了 URL 时才验证
+  if (!proxyEnabled.value || !form.value.proxy_url) {
+    return ''
+  }
+  const url = form.value.proxy_url.trim()
+
+  // 检查禁止的特殊字符
+  if (/[\n\r]/.test(url)) {
+    return '代理 URL 包含非法字符'
+  }
+
+  // 验证协议（不支持 SOCKS4）
+  if (!/^(http|https|socks5):\/\//i.test(url)) {
+    return '代理 URL 必须以 http://, https:// 或 socks5:// 开头'
+  }
+  try {
+    const parsed = new URL(url)
+    if (!parsed.host) {
+      return '代理 URL 必须包含有效的 host'
+    }
+    // 禁止 URL 中内嵌认证信息
+    if (parsed.username || parsed.password) {
+      return '请勿在 URL 中包含用户名和密码，请使用独立的认证字段'
+    }
+  } catch {
+    return '代理 URL 格式无效'
+  }
+  return ''
+})
+
 // 组件挂载时加载API格式
 onMounted(() => {
   loadApiFormats()
@@ -252,14 +395,23 @@ function resetForm() {
     max_retries: 3,
     max_concurrent: undefined,
     rate_limit: undefined,
-    is_active: true
+    is_active: true,
+    proxy_url: '',
+    proxy_username: '',
+    proxy_password: '',
   }
+  proxyEnabled.value = false
 }
 
+// 原始密码占位符（后端返回的脱敏标记）
+const MASKED_PASSWORD = '***'
+
 // 加载端点数据（编辑模式）
 function loadEndpointData() {
   if (!props.endpoint) return
 
+  const proxy = props.endpoint.proxy as { url?: string; username?: string; password?: string; enabled?: boolean } | null
+
   form.value = {
     api_format: props.endpoint.api_format,
     base_url: props.endpoint.base_url,
@@ -268,8 +420,15 @@ function loadEndpointData() {
     max_retries: props.endpoint.max_retries,
     max_concurrent: props.endpoint.max_concurrent || undefined,
     rate_limit: props.endpoint.rate_limit || undefined,
-    is_active: props.endpoint.is_active
+    is_active: props.endpoint.is_active,
+    proxy_url: proxy?.url || '',
+    proxy_username: proxy?.username || '',
+    // 如果密码是脱敏标记，显示为空（让用户知道有密码但看不到）
+    proxy_password: proxy?.password === MASKED_PASSWORD ? '' : (proxy?.password || ''),
   }
+
+  // 根据 enabled 字段或 url 存在判断是否启用代理
+  proxyEnabled.value = proxy?.enabled ?? !!proxy?.url
 }
 
 // 使用 useFormDialog 统一处理对话框逻辑
@@ -282,12 +441,47 @@ const { isEditMode, handleDialogUpdate, handleCancel } = useFormDialog({
   resetForm,
 })
 
+// 构建代理配置
+// - 有 URL 时始终保存配置，通过 enabled 字段控制是否启用
+// - 无 URL 时返回 null
+function buildProxyConfig(): { url: string; username?: string; password?: string; enabled: boolean } | null {
+  if (!form.value.proxy_url) {
+    // 没填 URL，无代理配置
+    return null
+  }
+  return {
+    url: form.value.proxy_url,
+    username: form.value.proxy_username || undefined,
+    password: form.value.proxy_password || undefined,
+    enabled: proxyEnabled.value,  // 开关状态决定是否启用
+  }
+}
+
 // 提交表单
-const handleSubmit = async () => {
+const handleSubmit = async (skipCredentialCheck = false) => {
   if (!props.provider && !props.endpoint) return
 
+  // 只在开关开启且填写了 URL 时验证
+  if (proxyEnabled.value && form.value.proxy_url && proxyUrlError.value) {
+    showError(proxyUrlError.value, '代理配置错误')
+    return
+  }
+
+  // 检查：开关开启但没有 URL，却有用户名或密码
+  const hasOrphanedCredentials = proxyEnabled.value
+    && !form.value.proxy_url
+    && (form.value.proxy_username || form.value.proxy_password)
+
+  if (hasOrphanedCredentials && !skipCredentialCheck) {
+    // 弹出确认对话框
+    showClearCredentialsDialog.value = true
+    return
+  }
+
   loading.value = true
   try {
+    const proxyConfig = buildProxyConfig()
+
     if (isEditMode.value && props.endpoint) {
       // 更新端点
       await updateEndpoint(props.endpoint.id, {
@@ -297,7 +491,8 @@ const handleSubmit = async () => {
         max_retries: form.value.max_retries,
         max_concurrent: form.value.max_concurrent,
         rate_limit: form.value.rate_limit,
-        is_active: form.value.is_active
+        is_active: form.value.is_active,
+        proxy: proxyConfig,
       })
 
       success('端点已更新', '保存成功')
@@ -313,7 +508,8 @@ const handleSubmit = async () => {
         max_retries: form.value.max_retries,
         max_concurrent: form.value.max_concurrent,
         rate_limit: form.value.rate_limit,
-        is_active: form.value.is_active
+        is_active: form.value.is_active,
+        proxy: proxyConfig,
       })
 
       success('端点创建成功', '成功')
@@ -329,4 +525,12 @@ const handleSubmit = async () => {
     loading.value = false
   }
 }
+
+// 确认清空凭据并继续保存
+const confirmClearCredentials = () => {
+  form.value.proxy_username = ''
+  form.value.proxy_password = ''
+  showClearCredentialsDialog.value = false
+  handleSubmit(true)  // 跳过凭据检查，直接提交
+}
 </script>

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

@@ -116,6 +116,19 @@
                 {{ model.global_model_name }}
               </div>
             </div>
+
+            <!-- 测试按钮 -->
+            <Button
+              variant="ghost"
+              size="icon"
+              class="h-7 w-7 shrink-0"
+              title="测试模型连接"
+              :disabled="testingModelName === model.global_model_name"
+              @click.stop="testModelConnection(model)"
+            >
+              <Loader2 v-if="testingModelName === model.global_model_name" class="w-3.5 h-3.5 animate-spin" />
+              <Play v-else class="w-3.5 h-3.5" />
+            </Button>
           </div>
         </div>
       </div>
@@ -148,16 +161,17 @@
 
 <script setup lang="ts">
 import { ref, computed, watch } from 'vue'
-import { Box, Loader2, Settings2 } from 'lucide-vue-next'
+import { Box, Loader2, Settings2, Play } from 'lucide-vue-next'
 import { Dialog } from '@/components/ui'
 import Button from '@/components/ui/button.vue'
 import Badge from '@/components/ui/badge.vue'
 import Checkbox from '@/components/ui/checkbox.vue'
 import { useToast } from '@/composables/useToast'
-import { parseApiError } from '@/utils/errorParser'
+import { parseApiError, parseTestModelError } from '@/utils/errorParser'
 import {
   updateEndpointKey,
   getProviderAvailableSourceModels,
+  testModel,
   type EndpointAPIKey,
   type ProviderAvailableSourceModel
 } from '@/api/endpoints'
@@ -181,6 +195,7 @@ const loadingModels = ref(false)
 const availableModels = ref<ProviderAvailableSourceModel[]>([])
 const selectedModels = ref<string[]>([])
 const initialModels = ref<string[]>([])
+const testingModelName = ref<string | null>(null)
 
 // 监听对话框打开
 watch(() => props.open, (open) => {
@@ -268,6 +283,32 @@ function clearModels() {
   selectedModels.value = []
 }
 
+// 测试模型连接
+async function testModelConnection(model: ProviderAvailableSourceModel) {
+  if (!props.providerId || !props.apiKey || testingModelName.value) return
+
+  testingModelName.value = model.global_model_name
+  try {
+    const result = await testModel({
+      provider_id: props.providerId,
+      model_name: model.provider_model_name,
+      api_key_id: props.apiKey.id,
+      message: "hello"
+    })
+
+    if (result.success) {
+      success(`模型 "${model.display_name}" 测试成功`)
+    } 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 {
+    testingModelName.value = null
+  }
+}
+
 function areArraysEqual(a: string[], b: string[]): boolean {
   if (a.length !== b.length) return false
   const sortedA = [...a].sort()

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

@@ -337,8 +337,40 @@
                                         {{ key.is_active ? '活跃' : '禁用' }}
                                       </Badge>
                                     </div>
-                                    <div class="text-[10px] font-mono text-muted-foreground truncate">
-                                      {{ key.api_key_masked }}
+                                    <div class="flex items-center gap-1">
+                                      <span class="text-[10px] font-mono text-muted-foreground truncate max-w-[180px]">
+                                        {{ revealedKeys.has(key.id) ? revealedKeys.get(key.id) : key.api_key_masked }}
+                                      </span>
+                                      <Button
+                                        variant="ghost"
+                                        size="icon"
+                                        class="h-5 w-5 shrink-0"
+                                        :title="revealedKeys.has(key.id) ? '隐藏密钥' : '显示密钥'"
+                                        :disabled="revealingKeyId === key.id"
+                                        @click.stop="toggleKeyReveal(key)"
+                                      >
+                                        <Loader2
+                                          v-if="revealingKeyId === key.id"
+                                          class="w-3 h-3 animate-spin"
+                                        />
+                                        <EyeOff
+                                          v-else-if="revealedKeys.has(key.id)"
+                                          class="w-3 h-3"
+                                        />
+                                        <Eye
+                                          v-else
+                                          class="w-3 h-3"
+                                        />
+                                      </Button>
+                                      <Button
+                                        variant="ghost"
+                                        size="icon"
+                                        class="h-5 w-5 shrink-0"
+                                        title="复制密钥"
+                                        @click.stop="copyFullKey(key)"
+                                      >
+                                        <Copy class="w-3 h-3" />
+                                      </Button>
                                     </div>
                                   </div>
                                   <div class="flex items-center gap-1.5 ml-auto shrink-0">
@@ -483,9 +515,9 @@
                                   <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: ${key.max_concurrent}`"
                                   >
-                                    {{ key.is_adaptive ? '自适应' : '固定' }}并发: {{ key.learned_max_concurrent || key.max_concurrent || 3 }}
+                                    {{ key.is_adaptive ? '自适应' : '固定' }}并发: {{ key.is_adaptive ? (key.learned_max_concurrent ?? '学习中') : key.max_concurrent }}
                                   </span>
                                 </div>
                               </div>
@@ -531,6 +563,7 @@
               <!-- 模型名称映射 -->
               <ModelAliasesTab
                 v-if="provider"
+                ref="modelAliasesTabRef"
                 :key="`aliases-${provider.id}`"
                 :provider="provider"
                 @refresh="handleRelatedDataRefresh"
@@ -653,12 +686,16 @@ import {
   Power,
   Layers,
   GripVertical,
-  Copy
+  Copy,
+  Eye,
+  EyeOff
 } 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,
@@ -678,6 +715,7 @@ import {
   updateEndpoint,
   updateEndpointKey,
   batchUpdateKeyPriority,
+  revealEndpointKey,
   type ProviderEndpoint,
   type EndpointAPIKey,
   type Model
@@ -704,6 +742,7 @@ const emit = defineEmits<{
 }>()
 
 const { error: showError, success: showSuccess } = useToast()
+const { copyToClipboard } = useClipboard()
 
 const loading = ref(false)
 const provider = ref<any>(null)
@@ -727,6 +766,10 @@ const recoveringEndpointId = ref<string | null>(null)
 const togglingEndpointId = ref<string | null>(null)
 const togglingKeyId = ref<string | null>(null)
 
+// 密钥显示状态：key_id -> 完整密钥
+const revealedKeys = ref<Map<string, string>>(new Map())
+const revealingKeyId = ref<string | null>(null)
+
 // 模型相关状态
 const modelFormDialogOpen = ref(false)
 const editingModel = ref<Model | null>(null)
@@ -734,6 +777,9 @@ 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 +801,9 @@ 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
@@ -791,6 +839,9 @@ watch(() => props.open, (newOpen) => {
     currentEndpoint.value = null
     editingKey.value = null
     keyToDelete.value = null
+
+    // 清除已显示的密钥（安全考虑）
+    revealedKeys.value.clear()
   }
 })
 
@@ -879,6 +930,43 @@ function handleConfigKeyModels(key: EndpointAPIKey) {
   keyAllowedModelsDialogOpen.value = true
 }
 
+// 切换密钥显示/隐藏
+async function toggleKeyReveal(key: EndpointAPIKey) {
+  if (revealedKeys.value.has(key.id)) {
+    // 已显示，隐藏它
+    revealedKeys.value.delete(key.id)
+    return
+  }
+
+  // 未显示，调用 API 获取完整密钥
+  revealingKeyId.value = key.id
+  try {
+    const result = await revealEndpointKey(key.id)
+    revealedKeys.value.set(key.id, result.api_key)
+  } catch (err: any) {
+    showError(err.response?.data?.detail || '获取密钥失败', '错误')
+  } finally {
+    revealingKeyId.value = null
+  }
+}
+
+// 复制完整密钥
+async function copyFullKey(key: EndpointAPIKey) {
+  // 如果已经显示了，直接复制
+  if (revealedKeys.value.has(key.id)) {
+    copyToClipboard(revealedKeys.value.get(key.id)!)
+    return
+  }
+
+  // 否则先获取再复制
+  try {
+    const result = await revealEndpointKey(key.id)
+    copyToClipboard(result.api_key)
+  } catch (err: any) {
+    showError(err.response?.data?.detail || '获取密钥失败', '错误')
+  }
+}
+
 function handleDeleteKey(key: EndpointAPIKey) {
   keyToDelete.value = key
   deleteKeyConfirmOpen.value = true
@@ -1243,16 +1331,6 @@ 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 +1374,16 @@ 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

@@ -213,6 +213,7 @@ import { Box, Edit, Trash2, Layers, Eye, Wrench, Zap, Brain, Power, Copy, Image
 import Card from '@/components/ui/card.vue'
 import Button from '@/components/ui/button.vue'
 import { useToast } from '@/composables/useToast'
+import { useClipboard } from '@/composables/useClipboard'
 import { getProviderModels, type Model } from '@/api/endpoints'
 import { updateModel } from '@/api/endpoints/models'
 
@@ -227,6 +228,7 @@ const emit = defineEmits<{
 }>()
 
 const { error: showError, success: showSuccess } = useToast()
+const { copyToClipboard } = useClipboard()
 
 // 状态
 const loading = ref(false)
@@ -244,12 +246,7 @@ 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

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

@@ -25,7 +25,7 @@
                 </h3>
                 <div class="flex items-center gap-1 text-sm font-mono text-muted-foreground bg-muted px-2 py-0.5 rounded">
                   <span>{{ detail?.model || '-' }}</span>
-                  <template v-if="detail?.target_model">
+                  <template v-if="detail?.target_model && detail.target_model !== detail.model">
                     <svg
                       xmlns="http://www.w3.org/2000/svg"
                       viewBox="0 0 20 20"
@@ -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/data.ts

@@ -5,7 +5,7 @@
 
 import type { User, LoginResponse } from '@/api/auth'
 import type { DashboardStatsResponse, RecentRequest, ProviderStatus, DailyStatsResponse } from '@/api/dashboard'
-import type { User as AdminUser, ApiKey } from '@/api/users'
+import type { User as AdminUser } from '@/api/users'
 import type { AdminApiKeysResponse } from '@/api/admin'
 import type { Profile, UsageResponse } from '@/api/me'
 import type { ProviderWithEndpointsSummary, GlobalModelResponse } from '@/api/endpoints/types'
@@ -185,18 +185,20 @@ export const MOCK_DASHBOARD_STATS: DashboardStatsResponse = {
     output: 700000,
     cache_creation: 50000,
     cache_read: 200000
-  }
+  },
+  // 普通用户专用字段
+  monthly_cost: 45.67
 }
 
 export const MOCK_RECENT_REQUESTS: RecentRequest[] = [
-  { id: 'req-001', user: 'alice', model: 'claude-sonnet-4-20250514', tokens: 15234, time: '2 分钟前' },
-  { id: 'req-002', user: 'bob', model: 'gpt-4o', tokens: 8765, time: '5 分钟前' },
-  { id: 'req-003', user: 'charlie', model: 'claude-opus-4-20250514', tokens: 32100, time: '8 分钟前' },
-  { id: 'req-004', user: 'diana', model: 'gemini-2.0-flash', tokens: 4521, time: '12 分钟前' },
-  { id: 'req-005', user: 'eve', model: 'claude-sonnet-4-20250514', tokens: 9876, time: '15 分钟前' },
-  { id: 'req-006', user: 'frank', model: 'gpt-4o-mini', tokens: 2345, time: '18 分钟前' },
-  { id: 'req-007', user: 'grace', model: 'claude-haiku-3-5-20241022', tokens: 6789, time: '22 分钟前' },
-  { id: 'req-008', user: 'henry', model: 'gemini-2.5-pro', tokens: 12345, time: '25 分钟前' }
+  { id: 'req-001', user: 'alice', model: 'claude-sonnet-4-5-20250929', tokens: 15234, time: '2 分钟前' },
+  { id: 'req-002', user: 'bob', model: 'gpt-5.1', tokens: 8765, time: '5 分钟前' },
+  { id: 'req-003', user: 'charlie', model: 'claude-opus-4-5-20251101', tokens: 32100, time: '8 分钟前' },
+  { id: 'req-004', user: 'diana', model: 'gemini-3-pro-preview', tokens: 4521, time: '12 分钟前' },
+  { id: 'req-005', user: 'eve', model: 'claude-sonnet-4-5-20250929', tokens: 9876, time: '15 分钟前' },
+  { id: 'req-006', user: 'frank', model: 'gpt-5.1-codex-mini', tokens: 2345, time: '18 分钟前' },
+  { id: 'req-007', user: 'grace', model: 'claude-haiku-4-5-20251001', tokens: 6789, time: '22 分钟前' },
+  { id: 'req-008', user: 'henry', model: 'gemini-3-pro-preview', tokens: 12345, time: '25 分钟前' }
 ]
 
 export const MOCK_PROVIDER_STATUS: ProviderStatus[] = [
@@ -231,11 +233,11 @@ function generateDailyStats(): DailyStatsResponse {
       unique_models: 8 + Math.floor(Math.random() * 5),
       unique_providers: 4 + Math.floor(Math.random() * 3),
       model_breakdown: [
-        { model: 'claude-sonnet-4-20250514', requests: Math.floor(baseRequests * 0.35), tokens: Math.floor(baseTokens * 0.35), cost: Number((baseCost * 0.35).toFixed(2)) },
-        { model: 'gpt-4o', requests: Math.floor(baseRequests * 0.25), tokens: Math.floor(baseTokens * 0.25), cost: Number((baseCost * 0.25).toFixed(2)) },
-        { model: 'claude-opus-4-20250514', requests: Math.floor(baseRequests * 0.15), tokens: Math.floor(baseTokens * 0.15), cost: Number((baseCost * 0.20).toFixed(2)) },
-        { model: 'gemini-2.0-flash', requests: Math.floor(baseRequests * 0.15), tokens: Math.floor(baseTokens * 0.15), cost: Number((baseCost * 0.10).toFixed(2)) },
-        { model: 'claude-haiku-3-5-20241022', requests: Math.floor(baseRequests * 0.10), tokens: Math.floor(baseTokens * 0.10), cost: Number((baseCost * 0.10).toFixed(2)) }
+        { model: 'claude-sonnet-4-5-20250929', requests: Math.floor(baseRequests * 0.35), tokens: Math.floor(baseTokens * 0.35), cost: Number((baseCost * 0.35).toFixed(2)) },
+        { model: 'gpt-5.1', requests: Math.floor(baseRequests * 0.25), tokens: Math.floor(baseTokens * 0.25), cost: Number((baseCost * 0.25).toFixed(2)) },
+        { model: 'claude-opus-4-5-20251101', requests: Math.floor(baseRequests * 0.15), tokens: Math.floor(baseTokens * 0.15), cost: Number((baseCost * 0.20).toFixed(2)) },
+        { model: 'gemini-3-pro-preview', requests: Math.floor(baseRequests * 0.15), tokens: Math.floor(baseTokens * 0.15), cost: Number((baseCost * 0.10).toFixed(2)) },
+        { model: 'claude-haiku-4-5-20251001', requests: Math.floor(baseRequests * 0.10), tokens: Math.floor(baseTokens * 0.10), cost: Number((baseCost * 0.10).toFixed(2)) }
       ]
     })
   }
@@ -243,11 +245,11 @@ function generateDailyStats(): DailyStatsResponse {
   return {
     daily_stats: dailyStats,
     model_summary: [
-      { model: 'claude-sonnet-4-20250514', requests: 2456, tokens: 8500000, cost: 125.45, avg_response_time: 1.2, cost_per_request: 0.051, tokens_per_request: 3461 },
-      { model: 'gpt-4o', requests: 1823, tokens: 6200000, cost: 98.32, avg_response_time: 0.9, cost_per_request: 0.054, tokens_per_request: 3401 },
-      { model: 'claude-opus-4-20250514', requests: 987, tokens: 4100000, cost: 156.78, avg_response_time: 2.1, cost_per_request: 0.159, tokens_per_request: 4154 },
-      { model: 'gemini-2.0-flash', requests: 1234, tokens: 3800000, cost: 28.56, avg_response_time: 0.6, cost_per_request: 0.023, tokens_per_request: 3079 },
-      { model: 'claude-haiku-3-5-20241022', requests: 2100, tokens: 5200000, cost: 32.10, avg_response_time: 0.5, cost_per_request: 0.015, tokens_per_request: 2476 }
+      { model: 'claude-sonnet-4-5-20250929', requests: 2456, tokens: 8500000, cost: 125.45, avg_response_time: 1.2, cost_per_request: 0.051, tokens_per_request: 3461 },
+      { model: 'gpt-5.1', requests: 1823, tokens: 6200000, cost: 98.32, avg_response_time: 0.9, cost_per_request: 0.054, tokens_per_request: 3401 },
+      { model: 'claude-opus-4-5-20251101', requests: 987, tokens: 4100000, cost: 156.78, avg_response_time: 2.1, cost_per_request: 0.159, tokens_per_request: 4154 },
+      { model: 'gemini-3-pro-preview', requests: 1234, tokens: 3800000, cost: 28.56, avg_response_time: 0.6, cost_per_request: 0.023, tokens_per_request: 3079 },
+      { model: 'claude-haiku-4-5-20251001', requests: 2100, tokens: 5200000, cost: 32.10, avg_response_time: 0.5, cost_per_request: 0.015, tokens_per_request: 2476 }
     ],
     period: {
       start_date: dailyStats[0].date,
@@ -336,7 +338,7 @@ export const MOCK_ALL_USERS: AdminUser[] = [
 
 // ========== API Key 数据 ==========
 
-export const MOCK_USER_API_KEYS: ApiKey[] = [
+export const MOCK_USER_API_KEYS = [
   {
     id: 'key-uuid-001',
     key_display: 'sk-ae...x7f9',
@@ -346,7 +348,8 @@ export const MOCK_USER_API_KEYS: ApiKey[] = [
     is_active: true,
     is_standalone: false,
     total_requests: 1234,
-    total_cost_usd: 45.67
+    total_cost_usd: 45.67,
+    force_capabilities: null
   },
   {
     id: 'key-uuid-002',
@@ -357,7 +360,8 @@ export const MOCK_USER_API_KEYS: ApiKey[] = [
     is_active: true,
     is_standalone: false,
     total_requests: 5678,
-    total_cost_usd: 123.45
+    total_cost_usd: 123.45,
+    force_capabilities: { cache_1h: true }
   },
   {
     id: 'key-uuid-003',
@@ -367,7 +371,8 @@ export const MOCK_USER_API_KEYS: ApiKey[] = [
     is_active: false,
     is_standalone: false,
     total_requests: 100,
-    total_cost_usd: 2.34
+    total_cost_usd: 2.34,
+    force_capabilities: null
   }
 ]
 
@@ -813,16 +818,16 @@ export const MOCK_USAGE_RESPONSE: UsageResponse = {
   quota_usd: 100,
   used_usd: 45.32,
   summary_by_model: [
-    { model: 'claude-sonnet-4-20250514', requests: 456, input_tokens: 650000, output_tokens: 250000, total_tokens: 900000, total_cost_usd: 18.50, actual_total_cost_usd: 13.50 },
-    { model: 'gpt-4o', requests: 312, input_tokens: 480000, output_tokens: 180000, total_tokens: 660000, total_cost_usd: 12.30, actual_total_cost_usd: 9.20 },
-    { model: 'claude-haiku-3-5-20241022', requests: 289, input_tokens: 420000, output_tokens: 170000, total_tokens: 590000, total_cost_usd: 8.50, actual_total_cost_usd: 6.30 },
-    { model: 'gemini-2.0-flash', requests: 177, input_tokens: 250000, output_tokens: 100000, total_tokens: 350000, total_cost_usd: 6.37, actual_total_cost_usd: 4.33 }
+    { model: 'claude-sonnet-4-5-20250929', requests: 456, input_tokens: 650000, output_tokens: 250000, total_tokens: 900000, total_cost_usd: 18.50, actual_total_cost_usd: 13.50 },
+    { model: 'gpt-5.1', requests: 312, input_tokens: 480000, output_tokens: 180000, total_tokens: 660000, total_cost_usd: 12.30, actual_total_cost_usd: 9.20 },
+    { model: 'claude-haiku-4-5-20251001', requests: 289, input_tokens: 420000, output_tokens: 170000, total_tokens: 590000, total_cost_usd: 8.50, actual_total_cost_usd: 6.30 },
+    { model: 'gemini-3-pro-preview', requests: 177, input_tokens: 250000, output_tokens: 100000, total_tokens: 350000, total_cost_usd: 6.37, actual_total_cost_usd: 4.33 }
   ],
   records: [
     {
       id: 'usage-001',
       provider: 'anthropic',
-      model: 'claude-sonnet-4-20250514',
+      model: 'claude-sonnet-4-5-20250929',
       input_tokens: 1500,
       output_tokens: 800,
       total_tokens: 2300,
@@ -837,7 +842,7 @@ export const MOCK_USAGE_RESPONSE: UsageResponse = {
     {
       id: 'usage-002',
       provider: 'openai',
-      model: 'gpt-4o',
+      model: 'gpt-5.1',
       input_tokens: 2000,
       output_tokens: 500,
       total_tokens: 2500,

frontend/src/mocks/handler.ts

@@ -403,12 +403,12 @@ 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' },
-  { id: 'alias-003', source_model: 'gpt4o', target_global_model_id: 'gm-004', target_global_model_name: 'gpt-4o', target_global_model_display_name: 'GPT-4o', 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-004', source_model: 'gemini-flash', target_global_model_id: 'gm-005', target_global_model_name: 'gemini-2.0-flash', target_global_model_display_name: 'Gemini 2.0 Flash', 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-001', source_model: 'claude-4-sonnet', target_global_model_id: 'gm-003', target_global_model_name: 'claude-sonnet-4-5-20250929', target_global_model_display_name: 'Claude Sonnet 4.5', 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-4-5-20251101', target_global_model_display_name: 'Claude Opus 4.5', 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-003', source_model: 'gpt5', target_global_model_id: 'gm-006', target_global_model_name: 'gpt-5.1', target_global_model_display_name: 'GPT-5.1', 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-004', source_model: 'gemini-pro', target_global_model_id: 'gm-005', target_global_model_name: 'gemini-3-pro-preview', target_global_model_display_name: 'Gemini 3 Pro Preview', 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' }
 ]
 
 // Mock Endpoint Keys
@@ -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）' })
 })
@@ -2172,10 +2172,10 @@ function generateIntervalTimelineData(
 
   // 模型列表（用于按模型区分颜色）
   const models = [
-    'claude-sonnet-4-20250514',
-    'claude-3-5-sonnet-20241022',
-    'claude-3-5-haiku-20241022',
-    'claude-opus-4-20250514'
+    'claude-sonnet-4-5-20250929',
+    'claude-haiku-4-5-20251001',
+    'claude-opus-4-5-20251101',
+    'gpt-5.1'
   ]
 
   // 生成模拟的请求间隔数据

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

@@ -185,32 +185,13 @@
         </div>
       </CardSection>
 
-      <!-- API Key 管理配置 -->
+      <!-- 独立余额 Key 过期管理 -->
       <CardSection
-        title="API Key 管理"
-        description="API Key 相关配置"
+        title="独立余额 Key 过期管理"
+        description="独立余额 Key 的过期处理策略（普通用户 Key 不会过期）"
       >
         <div class="grid grid-cols-1 md:grid-cols-2 gap-6">
-          <div>
-            <Label
-              for="api-key-expire"
-              class="block text-sm font-medium"
-            >
-              API密钥过期天数
-            </Label>
-            <Input
-              id="api-key-expire"
-              v-model.number="systemConfig.api_key_expire_days"
-              type="number"
-              placeholder="0"
-              class="mt-1"
-            />
-            <p class="mt-1 text-xs text-muted-foreground">
-              0 表示永不过期
-            </p>
-          </div>
-
-          <div class="flex items-center h-full pt-6">
+          <div class="flex items-center h-full">
             <div class="flex items-center space-x-2">
               <Checkbox
                 id="auto-delete-expired-keys"
@@ -224,7 +205,7 @@
                   自动删除过期 Key
                 </Label>
                 <p class="text-xs text-muted-foreground">
-                  关闭时仅禁用过期 Key
+                  关闭时仅禁用过期 Key，不会物理删除
                 </p>
               </div>
             </div>
@@ -448,6 +429,25 @@
               避免单次操作过大影响性能
             </p>
           </div>
+
+          <div>
+            <Label
+              for="audit-log-retention-days"
+              class="block text-sm font-medium"
+            >
+              审计日志保留天数
+            </Label>
+            <Input
+              id="audit-log-retention-days"
+              v-model.number="systemConfig.audit_log_retention_days"
+              type="number"
+              placeholder="30"
+              class="mt-1"
+            />
+            <p class="mt-1 text-xs text-muted-foreground">
+              超过后删除审计日志记录
+            </p>
+          </div>
         </div>
 
         <!-- 清理策略说明 -->
@@ -460,9 +460,11 @@
             <p>2. <strong>压缩日志阶段</strong>: body 字段被压缩存储，节省空间</p>
             <p>3. <strong>统计阶段</strong>: 仅保留 tokens、成本等统计信息</p>
             <p>4. <strong>归档删除</strong>: 超过保留期限后完全删除记录</p>
+            <p>5. <strong>审计日志</strong>: 独立清理，记录用户登录、操作等安全事件</p>
           </div>
         </div>
       </CardSection>
+
     </div>
 
     <!-- 导入配置对话框 -->
@@ -796,8 +798,7 @@ interface SystemConfig {
   // 用户注册
   enable_registration: boolean
   require_email_verification: boolean
-  // API Key 管理
-  api_key_expire_days: number
+  // 独立余额 Key 过期管理
   auto_delete_expired_keys: boolean
   // 日志记录
   request_log_level: string
@@ -811,6 +812,7 @@ interface SystemConfig {
   header_retention_days: number
   log_retention_days: number
   cleanup_batch_size: number
+  audit_log_retention_days: number
 }
 
 const loading = ref(false)
@@ -845,8 +847,7 @@ const systemConfig = ref<SystemConfig>({
   // 用户注册
   enable_registration: false,
   require_email_verification: false,
-  // API Key 管理
-  api_key_expire_days: 0,
+  // 独立余额 Key 过期管理
   auto_delete_expired_keys: false,
   // 日志记录
   request_log_level: 'basic',
@@ -860,6 +861,7 @@ const systemConfig = ref<SystemConfig>({
   header_retention_days: 90,
   log_retention_days: 365,
   cleanup_batch_size: 1000,
+  audit_log_retention_days: 30,
 })
 
 // 计算属性：KB 和 字节 之间的转换
@@ -901,8 +903,7 @@ async function loadSystemConfig() {
       // 用户注册
       'enable_registration',
       'require_email_verification',
-      // API Key 管理
-      'api_key_expire_days',
+      // 独立余额 Key 过期管理
       'auto_delete_expired_keys',
       // 日志记录
       'request_log_level',
@@ -916,6 +917,7 @@ async function loadSystemConfig() {
       'header_retention_days',
       'log_retention_days',
       'cleanup_batch_size',
+      'audit_log_retention_days',
     ]
 
     for (const key of configs) {
@@ -960,12 +962,7 @@ async function saveSystemConfig() {
         value: systemConfig.value.require_email_verification,
         description: '是否需要邮箱验证'
       },
-      // API Key 管理
-      {
-        key: 'api_key_expire_days',
-        value: systemConfig.value.api_key_expire_days,
-        description: 'API密钥过期天数'
-      },
+      // 独立余额 Key 过期管理
       {
         key: 'auto_delete_expired_keys',
         value: systemConfig.value.auto_delete_expired_keys,
@@ -1023,6 +1020,11 @@ async function saveSystemConfig() {
         value: systemConfig.value.cleanup_batch_size,
         description: '每批次清理的记录数'
       },
+      {
+        key: 'audit_log_retention_days',
+        value: systemConfig.value.audit_log_retention_days,
+        description: '审计日志保留天数'
+      },
     ]
 
     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/Dashboard.vue

@@ -145,10 +145,10 @@
               <DollarSign class="absolute top-3 right-3 h-3.5 w-3.5 sm:h-4 sm:w-4 text-muted-foreground" />
               <div class="pr-6">
                 <p class="text-[9px] sm:text-[10px] font-semibold uppercase tracking-[0.2em] sm:tracking-[0.3em] text-muted-foreground">
-                  实际成本
+                  本月费用
                 </p>
                 <p class="mt-1.5 sm:mt-2 text-lg sm:text-xl font-semibold text-foreground">
-                  {{ formatCurrency(costStats.total_actual_cost) }}
+                  {{ formatCurrency(costStats.total_cost) }}
                 </p>
                 <Badge
                   v-if="costStats.cost_savings > 0"
@@ -162,14 +162,14 @@
           </div>
         </div>
 
-        <!-- 普通用户：缓存统计 -->
+        <!-- 普通用户：月度统计 -->
         <div
-          v-else-if="!isAdmin && cacheStats && cacheStats.total_cache_tokens > 0"
+          v-else-if="!isAdmin && (hasCacheData || (userMonthlyCost !== null && userMonthlyCost > 0))"
           class="mt-6"
         >
           <div class="mb-3 flex items-center justify-between">
             <h3 class="text-sm font-medium text-foreground">
-              本月缓存使用
+              本月统计
             </h3>
             <Badge
               variant="outline"
@@ -178,8 +178,16 @@
               Monthly
             </Badge>
           </div>
-          <div class="grid grid-cols-2 gap-2 sm:gap-3 xl:grid-cols-4">
-            <Card class="relative p-3 sm:p-4 border-book-cloth/30">
+          <div
+            :class="[
+              'grid gap-2 sm:gap-3',
+              hasCacheData ? 'grid-cols-2 xl:grid-cols-4' : 'grid-cols-1 max-w-xs'
+            ]"
+          >
+            <Card
+              v-if="cacheStats"
+              class="relative p-3 sm:p-4 border-book-cloth/30"
+            >
               <Database class="absolute top-3 right-3 h-3.5 w-3.5 sm:h-4 sm:w-4 text-muted-foreground" />
               <div class="pr-6">
                 <p class="text-[9px] sm:text-[10px] font-semibold uppercase tracking-[0.2em] sm:tracking-[0.3em] text-muted-foreground">
@@ -190,7 +198,10 @@
                 </p>
               </div>
             </Card>
-            <Card class="relative p-3 sm:p-4 border-kraft/30">
+            <Card
+              v-if="cacheStats"
+              class="relative p-3 sm:p-4 border-kraft/30"
+            >
               <Hash class="absolute top-3 right-3 h-3.5 w-3.5 sm:h-4 sm:w-4 text-muted-foreground" />
               <div class="pr-6">
                 <p class="text-[9px] sm:text-[10px] font-semibold uppercase tracking-[0.2em] sm:tracking-[0.3em] text-muted-foreground">
@@ -201,7 +212,10 @@
                 </p>
               </div>
             </Card>
-            <Card class="relative p-3 sm:p-4 border-book-cloth/25">
+            <Card
+              v-if="cacheStats"
+              class="relative p-3 sm:p-4 border-book-cloth/25"
+            >
               <Database class="absolute top-3 right-3 h-3.5 w-3.5 sm:h-4 sm:w-4 text-muted-foreground" />
               <div class="pr-6">
                 <p class="text-[9px] sm:text-[10px] font-semibold uppercase tracking-[0.2em] sm:tracking-[0.3em] text-muted-foreground">
@@ -213,19 +227,16 @@
               </div>
             </Card>
             <Card
-              v-if="tokenBreakdown"
+              v-if="userMonthlyCost !== null"
               class="relative p-3 sm:p-4 border-manilla/40"
             >
-              <Hash class="absolute top-3 right-3 h-3.5 w-3.5 sm:h-4 sm:w-4 text-muted-foreground" />
+              <DollarSign class="absolute top-3 right-3 h-3.5 w-3.5 sm:h-4 sm:w-4 text-muted-foreground" />
               <div class="pr-6">
                 <p class="text-[9px] sm:text-[10px] font-semibold uppercase tracking-[0.2em] sm:tracking-[0.3em] text-muted-foreground">
-                  总Token
+                  本月费用
                 </p>
                 <p class="mt-1.5 sm:mt-2 text-lg sm:text-xl font-semibold text-foreground">
-                  {{ formatTokens((tokenBreakdown.input || 0) + (tokenBreakdown.output || 0)) }}
-                </p>
-                <p class="mt-0.5 sm:mt-1 text-[9px] sm:text-[10px] text-muted-foreground">
-                  输入 {{ formatTokens(tokenBreakdown.input || 0) }} / 输出 {{ formatTokens(tokenBreakdown.output || 0) }}
+                  {{ formatCurrency(userMonthlyCost) }}
                 </p>
               </div>
             </Card>
@@ -831,6 +842,12 @@ const cacheStats = ref<{
   total_cache_tokens: number
 } | null>(null)
 
+const userMonthlyCost = ref<number | null>(null)
+
+const hasCacheData = computed(() =>
+  cacheStats.value && cacheStats.value.total_cache_tokens > 0
+)
+
 const tokenBreakdown = ref<{
   input: number
   output: number
@@ -1086,6 +1103,7 @@ async function loadDashboardData() {
     } else {
       if (statsData.cache_stats) cacheStats.value = statsData.cache_stats
       if (statsData.token_breakdown) tokenBreakdown.value = statsData.token_breakdown
+      if (statsData.monthly_cost !== undefined) userMonthlyCost.value = statsData.monthly_cost
     }
   } finally {
     loading.value = false

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>

pyproject.toml

@@ -13,7 +13,7 @@ authors = [
 classifiers = [
     "Development Status :: 4 - Beta",
     "Intended Audience :: Developers",
-    "License :: OSI Approved :: MIT License",
+    "License :: Other/Proprietary License",
     "Programming Language :: Python :: 3",
     "Programming Language :: Python :: 3.8",
     "Programming Language :: Python :: 3.9",

src/__init__.py

@@ -3,10 +3,8 @@
 A proxy server that enables AI models to work with multiple API providers.
 """
 
-from dotenv import load_dotenv
-
-# Load environment variables from .env file
-load_dotenv()
+# 注意: dotenv 加载已统一移至 src/config/settings.py
+# 不要在此处重复加载
 
 try:
     from src._version import __version__

src/api/admin/api_keys/routes.py

@@ -223,7 +223,7 @@ class AdminCreateStandaloneKeyAdapter(AdminApiAdapter):
             allowed_providers=self.key_data.allowed_providers,
             allowed_api_formats=self.key_data.allowed_api_formats,
             allowed_models=self.key_data.allowed_models,
-            rate_limit=self.key_data.rate_limit or 100,
+            rate_limit=self.key_data.rate_limit,  # None 表示不限制
             expire_days=self.key_data.expire_days,
             initial_balance_usd=self.key_data.initial_balance_usd,
             is_standalone=True,  # 标记为独立Key

src/api/admin/endpoints/keys.py

@@ -80,6 +80,17 @@ async def get_keys_grouped_by_format(
     return await pipeline.run(adapter=adapter, http_request=request, db=db, mode=adapter.mode)
 
 
+@router.get("/keys/{key_id}/reveal")
+async def reveal_endpoint_key(
+    key_id: str,
+    request: Request,
+    db: Session = Depends(get_db),
+) -> dict:
+    """获取完整的 API Key（用于查看和复制）"""
+    adapter = AdminRevealEndpointKeyAdapter(key_id=key_id)
+    return await pipeline.run(adapter=adapter, http_request=request, db=db, mode=adapter.mode)
+
+
 @router.delete("/keys/{key_id}")
 async def delete_endpoint_key(
     key_id: str,
@@ -246,6 +257,15 @@ 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 +273,7 @@ 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)
@@ -284,6 +304,30 @@ class AdminUpdateEndpointKeyAdapter(AdminApiAdapter):
         return EndpointAPIKeyResponse(**response_dict)
 
 
+@dataclass
+class AdminRevealEndpointKeyAdapter(AdminApiAdapter):
+    """获取完整的 API Key（用于查看和复制）"""
+
+    key_id: str
+
+    async def handle(self, context):  # type: ignore[override]
+        db = context.db
+        key = db.query(ProviderAPIKey).filter(ProviderAPIKey.id == self.key_id).first()
+        if not key:
+            raise NotFoundException(f"Key {self.key_id} 不存在")
+
+        try:
+            decrypted_key = crypto_service.decrypt(key.api_key)
+        except Exception as e:
+            logger.error(f"解密 Key 失败: ID={self.key_id}, Error={e}")
+            raise InvalidRequestException(
+                "无法解密 API Key，可能是加密密钥已更改。请重新添加该密钥。"
+            )
+
+        logger.info(f"[REVEAL] 查看完整 Key: ID={self.key_id}, Name={key.name}")
+        return {"api_key": decrypted_key}
+
+
 @dataclass
 class AdminDeleteEndpointKeyAdapter(AdminApiAdapter):
     key_id: str

src/api/admin/endpoints/routes.py

@@ -5,7 +5,7 @@ ProviderEndpoint CRUD 管理 API
 import uuid
 from dataclasses import dataclass
 from datetime import datetime, timezone
-from typing import List
+from typing import List, Optional
 
 from fastapi import APIRouter, Depends, Query, Request
 from sqlalchemy import and_, func
@@ -27,6 +27,16 @@ router = APIRouter(tags=["Endpoint Management"])
 pipeline = ApiRequestPipeline()
 
 
+def mask_proxy_password(proxy_config: Optional[dict]) -> Optional[dict]:
+    """对代理配置中的密码进行脱敏处理"""
+    if not proxy_config:
+        return None
+    masked = dict(proxy_config)
+    if masked.get("password"):
+        masked["password"] = "***"
+    return masked
+
+
 @router.get("/providers/{provider_id}/endpoints", response_model=List[ProviderEndpointResponse])
 async def list_provider_endpoints(
     provider_id: str,
@@ -153,6 +163,7 @@ class AdminListProviderEndpointsAdapter(AdminApiAdapter):
                 "api_format": endpoint.api_format,
                 "total_keys": total_keys_map.get(endpoint.id, 0),
                 "active_keys": active_keys_map.get(endpoint.id, 0),
+                "proxy": mask_proxy_password(endpoint.proxy),
             }
             endpoint_dict.pop("_sa_instance_state", None)
             result.append(ProviderEndpointResponse(**endpoint_dict))
@@ -202,6 +213,7 @@ class AdminCreateProviderEndpointAdapter(AdminApiAdapter):
             rate_limit=self.endpoint_data.rate_limit,
             is_active=True,
             config=self.endpoint_data.config,
+            proxy=self.endpoint_data.proxy.model_dump() if self.endpoint_data.proxy else None,
             created_at=now,
             updated_at=now,
         )
@@ -215,12 +227,13 @@ class AdminCreateProviderEndpointAdapter(AdminApiAdapter):
         endpoint_dict = {
             k: v
             for k, v in new_endpoint.__dict__.items()
-            if k not in {"api_format", "_sa_instance_state"}
+            if k not in {"api_format", "_sa_instance_state", "proxy"}
         }
         return ProviderEndpointResponse(
             **endpoint_dict,
             provider_name=provider.name,
             api_format=new_endpoint.api_format,
+            proxy=mask_proxy_password(new_endpoint.proxy),
             total_keys=0,
             active_keys=0,
         )
@@ -259,12 +272,13 @@ class AdminGetProviderEndpointAdapter(AdminApiAdapter):
         endpoint_dict = {
             k: v
             for k, v in endpoint_obj.__dict__.items()
-            if k not in {"api_format", "_sa_instance_state"}
+            if k not in {"api_format", "_sa_instance_state", "proxy"}
         }
         return ProviderEndpointResponse(
             **endpoint_dict,
             provider_name=provider.name,
             api_format=endpoint_obj.api_format,
+            proxy=mask_proxy_password(endpoint_obj.proxy),
             total_keys=total_keys,
             active_keys=active_keys,
         )
@@ -284,6 +298,17 @@ class AdminUpdateProviderEndpointAdapter(AdminApiAdapter):
             raise NotFoundException(f"Endpoint {self.endpoint_id} 不存在")
 
         update_data = self.endpoint_data.model_dump(exclude_unset=True)
+        # 把 proxy 转换为 dict 存储，支持显式设置为 None 清除代理
+        if "proxy" in update_data:
+            if update_data["proxy"] is not None:
+                new_proxy = dict(update_data["proxy"])
+                # 只有当密码字段未提供时才保留原密码（空字符串视为显式清除）
+                if "password" not in new_proxy and endpoint.proxy:
+                    old_password = endpoint.proxy.get("password")
+                    if old_password:
+                        new_proxy["password"] = old_password
+                update_data["proxy"] = new_proxy
+            # proxy 为 None 时保留，用于清除代理配置
         for field, value in update_data.items():
             setattr(endpoint, field, value)
         endpoint.updated_at = datetime.now(timezone.utc)
@@ -311,12 +336,13 @@ class AdminUpdateProviderEndpointAdapter(AdminApiAdapter):
         endpoint_dict = {
             k: v
             for k, v in endpoint.__dict__.items()
-            if k not in {"api_format", "_sa_instance_state"}
+            if k not in {"api_format", "_sa_instance_state", "proxy"}
         }
         return ProviderEndpointResponse(
             **endpoint_dict,
             provider_name=provider.name if provider else "Unknown",
             api_format=endpoint.api_format,
+            proxy=mask_proxy_password(endpoint.proxy),
             total_keys=total_keys,
             active_keys=active_keys,
         )

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/monitoring/trace.py

@@ -7,7 +7,7 @@ from datetime import datetime
 from typing import List, Optional
 
 from fastapi import APIRouter, Depends, HTTPException, Query, Request
-from pydantic import BaseModel
+from pydantic import BaseModel, ConfigDict
 from sqlalchemy.orm import Session
 
 from src.api.base.admin_adapter import AdminApiAdapter
@@ -52,8 +52,7 @@ class CandidateResponse(BaseModel):
     started_at: Optional[datetime] = None
     finished_at: Optional[datetime] = None
 
-    class Config:
-        from_attributes = True
+    model_config = ConfigDict(from_attributes=True)
 
 
 class RequestTraceResponse(BaseModel):

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

@@ -9,6 +9,7 @@ from fastapi import APIRouter, Depends, Request
 from sqlalchemy.orm import Session, joinedload
 
 from src.api.base.admin_adapter import AdminApiAdapter
+from src.api.base.models_service import invalidate_models_list_cache
 from src.api.base.pipeline import ApiRequestPipeline
 from src.core.exceptions import InvalidRequestException, NotFoundException
 from src.core.logger import logger
@@ -21,16 +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"])
@@ -157,6 +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 --------
 
 
@@ -419,4 +444,135 @@ class AdminBatchAssignModelsToProviderAdapter(AdminApiAdapter):
             f"Batch assigned {len(success)} GlobalModels to provider {provider.name} by {context.user.username}"
         )
 
+        # 清除 /v1/models 列表缓存
+        if success:
+            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/announcements/routes.py

@@ -140,9 +140,9 @@ class AnnouncementOptionalAuthAdapter(ApiAdapter):
         if not authorization or not authorization.lower().startswith("bearer "):
             return None
 
-        token = authorization.replace("Bearer ", "").strip()
+        token = authorization[7:].strip()
         try:
-            payload = await AuthService.verify_token(token)
+            payload = await AuthService.verify_token(token, token_type="access")
             user_id = payload.get("user_id")
             if not user_id:
                 return None

src/api/auth/routes.py

@@ -211,7 +211,7 @@ class AuthRefreshAdapter(AuthPublicAdapter):
 
 class AuthRegisterAdapter(AuthPublicAdapter):
     async def handle(self, context):  # type: ignore[override]
-        from ..models.database import SystemConfig
+        from src.models.database import SystemConfig
 
         db = context.db
         payload = context.ensure_json_body()

src/api/base/models_service.py

@@ -55,6 +55,23 @@ async def _set_cached_models(api_formats: list[str], models: list["ModelInfo"])
         logger.warning(f"[ModelsService] 缓存写入失败: {e}")
 
 
+async def invalidate_models_list_cache() -> None:
+    """
+    清除所有 /v1/models 列表缓存
+
+    在模型创建、更新、删除时调用，确保模型列表实时更新
+    """
+    # 清除所有格式的缓存
+    all_formats = ["CLAUDE", "OPENAI", "GEMINI"]
+    for fmt in all_formats:
+        cache_key = f"{_CACHE_KEY_PREFIX}:{fmt}"
+        try:
+            await CacheService.delete(cache_key)
+            logger.debug(f"[ModelsService] 已清除缓存: {cache_key}")
+        except Exception as e:
+            logger.warning(f"[ModelsService] 清除缓存失败 {cache_key}: {e}")
+
+
 @dataclass
 class ModelInfo:
     """统一的模型信息结构"""

src/api/base/pipeline.py

@@ -5,13 +5,12 @@ from enum import Enum
 from typing import Any, Optional, Tuple
 
 from fastapi import HTTPException, Request
-from sqlalchemy.orm import Session, sessionmaker
+from sqlalchemy.orm import Session
 
 from src.core.exceptions import QuotaExceededException
 from src.core.logger import logger
 from src.models.database import ApiKey, AuditEventType, User, UserRole
 from src.services.auth.service import AuthService
-from src.services.cache.user_cache import UserCacheService
 from src.services.system.audit import AuditService
 from src.services.usage.service import UsageService
 
@@ -178,9 +177,9 @@ class ApiRequestPipeline:
         if not authorization or not authorization.lower().startswith("bearer "):
             raise HTTPException(status_code=401, detail="缺少管理员凭证")
 
-        token = authorization.replace("Bearer ", "").strip()
+        token = authorization[7:].strip()
         try:
-            payload = await self.auth_service.verify_token(token)
+            payload = await self.auth_service.verify_token(token, token_type="access")
         except HTTPException:
             raise
         except Exception as exc:
@@ -191,8 +190,8 @@ class ApiRequestPipeline:
         if not user_id:
             raise HTTPException(status_code=401, detail="无效的管理员令牌")
 
-        # 使用缓存查询用户
-        user = await UserCacheService.get_user_by_id(db, user_id)
+        # 直接查询数据库，确保返回的是当前 Session 绑定的对象
+        user = db.query(User).filter(User.id == user_id).first()
         if not user or not user.is_active:
             raise HTTPException(status_code=403, detail="用户不存在或已禁用")
 
@@ -205,9 +204,9 @@ class ApiRequestPipeline:
         if not authorization or not authorization.lower().startswith("bearer "):
             raise HTTPException(status_code=401, detail="缺少用户凭证")
 
-        token = authorization.replace("Bearer ", "").strip()
+        token = authorization[7:].strip()
         try:
-            payload = await self.auth_service.verify_token(token)
+            payload = await self.auth_service.verify_token(token, token_type="access")
         except HTTPException:
             raise
         except Exception as exc:
@@ -218,8 +217,8 @@ class ApiRequestPipeline:
         if not user_id:
             raise HTTPException(status_code=401, detail="无效的用户令牌")
 
-        # 使用缓存查询用户
-        user = await UserCacheService.get_user_by_id(db, user_id)
+        # 直接查询数据库，确保返回的是当前 Session 绑定的对象
+        user = db.query(User).filter(User.id == user_id).first()
         if not user or not user.is_active:
             raise HTTPException(status_code=403, detail="用户不存在或已禁用")
 
@@ -242,11 +241,15 @@ class ApiRequestPipeline:
         status_code: Optional[int] = None,
         error: Optional[str] = None,
     ) -> None:
+        """记录审计事件
+
+        事务策略：复用请求级 Session，不单独提交。
+        审计记录随主事务一起提交，由中间件统一管理。
+        """
         if not getattr(adapter, "audit_log_enabled", True):
             return
 
-        bind = context.db.get_bind()
-        if bind is None:
+        if context.db is None:
             return
 
         event_type = adapter.audit_success_event if success else adapter.audit_failure_event
@@ -266,11 +269,11 @@ class ApiRequestPipeline:
             error=error,
         )
 
-        SessionMaker = sessionmaker(bind=bind)
-        audit_session = SessionMaker()
         try:
+            # 复用请求级 Session，不创建新的连接
+            # 审计记录随主事务一起提交，由中间件统一管理
             self.audit_service.log_event(
-                db=audit_session,
+                db=context.db,
                 event_type=event_type,
                 description=f"{context.request.method} {context.request.url.path} via {adapter.name}",
                 user_id=context.user.id if context.user else None,
@@ -282,12 +285,9 @@ class ApiRequestPipeline:
                 error_message=error,
                 metadata=metadata,
             )
-            audit_session.commit()
         except Exception as exc:
-            audit_session.rollback()
+            # 审计失败不应影响主请求，仅记录警告
             logger.warning(f"[Audit] Failed to record event for adapter={adapter.name}: {exc}")
-        finally:
-            audit_session.close()
 
     def _build_audit_metadata(
         self,

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
@@ -118,7 +118,9 @@ class AdminDashboardStatsAdapter(AdminApiAdapter):
         # 转换为 UTC 用于与 stats_daily.date 比较（存储的是业务日期对应的 UTC 开始时间）
         today = today_local.astimezone(timezone.utc)
         yesterday = (today_local - timedelta(days=1)).astimezone(timezone.utc)
-        last_month = (today_local - timedelta(days=30)).astimezone(timezone.utc)
+        # 本月第一天（自然月）
+        month_start_local = today_local.replace(day=1)
+        month_start = month_start_local.astimezone(timezone.utc)
 
         # ==================== 使用预聚合数据 ====================
         # 从 stats_summary + 今日实时数据获取全局统计
@@ -208,7 +210,7 @@ class AdminDashboardStatsAdapter(AdminApiAdapter):
                 func.sum(StatsDaily.cache_read_cost).label("cache_read_cost"),
                 func.sum(StatsDaily.fallback_count).label("fallback_count"),
             )
-            .filter(StatsDaily.date >= last_month, StatsDaily.date < today)
+            .filter(StatsDaily.date >= month_start, StatsDaily.date < today)
             .first()
         )
 
@@ -227,24 +229,24 @@ class AdminDashboardStatsAdapter(AdminApiAdapter):
         else:
             # 回退到实时查询（没有预聚合数据时）
             total_requests = (
-                db.query(func.count(Usage.id)).filter(Usage.created_at >= last_month).scalar() or 0
+                db.query(func.count(Usage.id)).filter(Usage.created_at >= month_start).scalar() or 0
             )
             total_cost = (
-                db.query(func.sum(Usage.total_cost_usd)).filter(Usage.created_at >= last_month).scalar() or 0
+                db.query(func.sum(Usage.total_cost_usd)).filter(Usage.created_at >= month_start).scalar() or 0
             )
             total_actual_cost = (
                 db.query(func.sum(Usage.actual_total_cost_usd))
-                .filter(Usage.created_at >= last_month).scalar() or 0
+                .filter(Usage.created_at >= month_start).scalar() or 0
             )
             error_requests = (
                 db.query(func.count(Usage.id))
                 .filter(
-                    Usage.created_at >= last_month,
+                    Usage.created_at >= month_start,
                     (Usage.status_code >= 400) | (Usage.error_message.isnot(None)),
                 ).scalar() or 0
             )
             total_tokens = (
-                db.query(func.sum(Usage.total_tokens)).filter(Usage.created_at >= last_month).scalar() or 0
+                db.query(func.sum(Usage.total_tokens)).filter(Usage.created_at >= month_start).scalar() or 0
             )
             cache_stats = (
                 db.query(
@@ -253,7 +255,7 @@ class AdminDashboardStatsAdapter(AdminApiAdapter):
                     func.sum(Usage.cache_creation_cost_usd).label("cache_creation_cost"),
                     func.sum(Usage.cache_read_cost_usd).label("cache_read_cost"),
                 )
-                .filter(Usage.created_at >= last_month)
+                .filter(Usage.created_at >= month_start)
                 .first()
             )
             cache_creation_tokens = int(cache_stats.cache_creation_tokens or 0) if cache_stats else 0
@@ -267,7 +269,7 @@ class AdminDashboardStatsAdapter(AdminApiAdapter):
                     RequestCandidate.request_id, func.count(RequestCandidate.id).label("executed_count")
                 )
                 .filter(
-                    RequestCandidate.created_at >= last_month,
+                    RequestCandidate.created_at >= month_start,
                     RequestCandidate.status.in_(["success", "failed"]),
                 )
                 .group_by(RequestCandidate.request_id)
@@ -447,7 +449,9 @@ class UserDashboardStatsAdapter(DashboardAdapter):
         # 转换为 UTC 用于数据库查询
         today = today_local.astimezone(timezone.utc)
         yesterday = (today_local - timedelta(days=1)).astimezone(timezone.utc)
-        last_month = (today_local - timedelta(days=30)).astimezone(timezone.utc)
+        # 本月第一天（自然月）
+        month_start_local = today_local.replace(day=1)
+        month_start = month_start_local.astimezone(timezone.utc)
 
         user_api_keys = db.query(func.count(ApiKey.id)).filter(ApiKey.user_id == user.id).scalar()
         active_keys = (
@@ -483,12 +487,12 @@ class UserDashboardStatsAdapter(DashboardAdapter):
         # 本月请求统计
         user_requests = (
             db.query(func.count(Usage.id))
-            .filter(and_(Usage.user_id == user.id, Usage.created_at >= last_month))
+            .filter(and_(Usage.user_id == user.id, Usage.created_at >= month_start))
             .scalar()
         )
         user_cost = (
             db.query(func.sum(Usage.total_cost_usd))
-            .filter(and_(Usage.user_id == user.id, Usage.created_at >= last_month))
+            .filter(and_(Usage.user_id == user.id, Usage.created_at >= month_start))
             .scalar()
             or 0
         )
@@ -532,18 +536,19 @@ class UserDashboardStatsAdapter(DashboardAdapter):
                 func.sum(Usage.cache_read_input_tokens).label("cache_read_tokens"),
                 func.sum(Usage.input_tokens).label("total_input_tokens"),
             )
-            .filter(and_(Usage.user_id == user.id, Usage.created_at >= last_month))
+            .filter(and_(Usage.user_id == user.id, Usage.created_at >= month_start))
             .first()
         )
         cache_creation_tokens = int(cache_stats.cache_creation_tokens or 0) if cache_stats else 0
         cache_read_tokens = int(cache_stats.cache_read_tokens or 0) if cache_stats else 0
+        monthly_input_tokens = int(cache_stats.total_input_tokens or 0) if cache_stats else 0
 
-        # 计算缓存命中率：cache_read / (input_tokens + cache_read)
+        # 计算本月缓存命中率：cache_read / (input_tokens + cache_read)
         # input_tokens 是实际发送给模型的输入（不含缓存读取），cache_read 是从缓存读取的
         # 总输入 = input_tokens + cache_read，缓存命中率 = cache_read / 总输入
-        total_input_with_cache = all_time_input_tokens + all_time_cache_read
+        total_input_with_cache = monthly_input_tokens + cache_read_tokens
         cache_hit_rate = (
-            round((all_time_cache_read / total_input_with_cache) * 100, 1)
+            round((cache_read_tokens / total_input_with_cache) * 100, 1)
             if total_input_with_cache > 0
             else 0
         )
@@ -569,15 +574,15 @@ class UserDashboardStatsAdapter(DashboardAdapter):
             quota_value = "无限制"
             quota_change = f"已用 ${user.used_usd:.2f}"
             quota_high = False
-        elif user.quota_usd and user.quota_usd > 0:
+        elif user.quota_usd > 0:
             percent = min(100, int((user.used_usd / user.quota_usd) * 100))
-            quota_value = "无限制"
+            quota_value = f"${user.quota_usd:.0f}"
             quota_change = f"已用 ${user.used_usd:.2f}"
             quota_high = percent > 80
         else:
-            quota_value = "0%"
+            quota_value = "$0"
             quota_change = f"已用 ${user.used_usd:.2f}"
-            quota_high = False
+            quota_high = True
 
         return {
             "stats": [
@@ -605,9 +610,15 @@ class UserDashboardStatsAdapter(DashboardAdapter):
                     "icon": "TrendingUp",
                 },
                 {
-                    "name": "本月费用",
-                    "value": f"${user_cost:.2f}",
-                    "icon": "DollarSign",
+                    "name": "总Token",
+                    "value": format_tokens(
+                        all_time_input_tokens
+                        + all_time_output_tokens
+                        + all_time_cache_creation
+                        + all_time_cache_read
+                    ),
+                    "subValue": f"输入 {format_tokens(all_time_input_tokens)} / 输出 {format_tokens(all_time_output_tokens)}",
+                    "icon": "Hash",
                 },
             ],
             "today": {
@@ -631,6 +642,8 @@ class UserDashboardStatsAdapter(DashboardAdapter):
                 "cache_hit_rate": cache_hit_rate,
                 "total_cache_tokens": cache_creation_tokens + cache_read_tokens,
             },
+            # 本月费用（用于下方缓存区域显示）
+            "monthly_cost": float(user_cost),
         }
 
 
@@ -893,69 +906,172 @@ 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

@@ -28,7 +28,7 @@
 from __future__ import annotations
 
 import time
-from typing import Any, Callable, Dict, Optional, Protocol, runtime_checkable
+from typing import TYPE_CHECKING, Any, Callable, Dict, Optional, Protocol, runtime_checkable
 
 from fastapi import Request
 from fastapi.responses import JSONResponse, StreamingResponse
@@ -43,6 +43,9 @@ from src.services.provider.format import normalize_api_format
 from src.services.system.audit import audit_service
 from src.services.usage.service import UsageService
 
+if TYPE_CHECKING:
+    from src.api.handlers.base.stream_context import StreamContext
+
 
 
 class MessageTelemetry:
@@ -373,6 +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
         """
@@ -399,6 +405,44 @@ class BaseMessageHandler:
         # 创建后台任务，不阻塞当前流
         asyncio.create_task(_do_update())
 
+    def _update_usage_to_streaming_with_ctx(self, ctx: "StreamContext") -> None:
+        """更新 Usage 状态为 streaming，同时更新 provider 和 target_model
+
+        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
+        """
+        import asyncio
+        from src.database.database import get_db
+
+        target_request_id = self.request_id
+        provider = ctx.provider_name
+        target_model = ctx.mapped_model
+
+        async def _do_update() -> None:
+            try:
+                db_gen = get_db()
+                db = next(db_gen)
+                try:
+                    UsageService.update_usage_status(
+                        db=db,
+                        request_id=target_request_id,
+                        status="streaming",
+                        provider=provider,
+                        target_model=target_model,
+                    )
+                finally:
+                    db.close()
+            except Exception as e:
+                logger.warning(f"[{target_request_id}] 更新 Usage 状态为 streaming 失败: {e}")
+
+        # 创建后台任务，不阻塞当前流
+        asyncio.create_task(_do_update())
+
     def _log_request_error(self, message: str, error: Exception) -> None:
         """记录请求错误日志，对业务异常不打印堆栈
 
@@ -411,9 +455,10 @@ class BaseMessageHandler:
             QuotaExceededException,
             RateLimitException,
             ModelNotSupportedException,
+            UpstreamClientException,
         )
 
-        if isinstance(error, (ProviderException, QuotaExceededException, RateLimitException, ModelNotSupportedException)):
+        if isinstance(error, (ProviderException, QuotaExceededException, RateLimitException, ModelNotSupportedException, UpstreamClientException)):
             # 业务异常：简洁日志，不打印堆栈
             logger.error(f"{message}: [{type(error).__name__}] {error}")
         else:

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
 
@@ -62,20 +63,36 @@ 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]
-        self.response_normalizer = None
-        # 可选启用响应规范化
-        self._init_response_normalizer()
-
-    def _init_response_normalizer(self):
-        """初始化响应规范化器 - 子类可覆盖"""
-        try:
-            from src.services.provider.response_normalizer import ResponseNormalizer
-
-            self.response_normalizer = ResponseNormalizer()
-        except ImportError:
-            pass
 
     async def handle(self, context: ApiRequestContext):
         """处理 Chat API 请求"""
@@ -228,8 +245,6 @@ class ChatAdapterBase(ApiAdapter):
             user_agent=user_agent,
             start_time=start_time,
             allowed_api_formats=self.allowed_api_formats,
-            response_normalizer=self.response_normalizer,
-            enable_response_normalization=self.response_normalizer is not None,
             adapter_detector=self.detect_capability_requirements,
         )
 
@@ -634,6 +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

@@ -88,8 +88,6 @@ class ChatHandlerBase(BaseMessageHandler, ABC):
         user_agent: str,
         start_time: float,
         allowed_api_formats: Optional[list] = None,
-        response_normalizer: Optional[Any] = None,
-        enable_response_normalization: bool = False,
         adapter_detector: Optional[Callable[[Dict[str, str], Optional[Dict[str, Any]]], Dict[str, bool]]] = None,
     ):
         allowed = allowed_api_formats or [self.FORMAT_ID]
@@ -106,8 +104,6 @@ class ChatHandlerBase(BaseMessageHandler, ABC):
         )
         self._parser: Optional[ResponseParser] = None
         self._request_builder = PassthroughRequestBuilder()
-        self.response_normalizer = response_normalizer
-        self.enable_response_normalization = enable_response_normalization
 
     @property
     def parser(self) -> ResponseParser:
@@ -264,9 +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
@@ -297,11 +293,15 @@ class ChatHandlerBase(BaseMessageHandler, ABC):
         # 创建类型安全的流式上下文
         ctx = StreamContext(model=model, api_format=api_format)
 
+        # 创建更新状态的回调闭包（可以访问 ctx）
+        def update_streaming_status() -> None:
+            self._update_usage_to_streaming_with_ctx(ctx)
+
         # 创建流处理器
         stream_processor = StreamProcessor(
             request_id=self.request_id,
             default_parser=self.parser,
-            on_streaming_start=self._update_usage_to_streaming,
+            on_streaming_start=update_streaming_status,
         )
 
         # 定义请求函数
@@ -466,7 +466,13 @@ class ChatHandlerBase(BaseMessageHandler, ABC):
             pool=config.http_pool_timeout,
         )
 
-        http_client = httpx.AsyncClient(timeout=timeout_config, follow_redirects=True)
+        # 创建 HTTP 客户端（支持代理配置）
+        from src.clients.http_client import HTTPClientPool
+
+        http_client = HTTPClientPool.create_client_with_proxy(
+            proxy_config=endpoint.proxy,
+            timeout=timeout_config,
+        )
         try:
             response_ctx = http_client.stream(
                 "POST", url, json=provider_payload, headers=provider_headers
@@ -478,9 +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(
@@ -633,11 +638,17 @@ 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')}")
 
-            async with httpx.AsyncClient(
-                timeout=float(endpoint.timeout),
-                follow_redirects=True,
-            ) as http_client:
+            # 创建 HTTP 客户端（支持代理配置）
+            from src.clients.http_client import HTTPClientPool
+
+            http_client = HTTPClientPool.create_client_with_proxy(
+                proxy_config=endpoint.proxy,
+                timeout=httpx.Timeout(float(endpoint.timeout)),
+            )
+            async with http_client:
                 resp = await http_client.post(url, json=provider_payload, headers=provider_hdrs)
 
                 status_code = resp.status_code
@@ -652,10 +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),
@@ -454,7 +456,13 @@ class CliMessageHandlerBase(BaseMessageHandler):
             f"Key=***{key.api_key[-4:]}, "
             f"原始模型={ctx.model}, 映射后={mapped_model or '无映射'}, URL模型={url_model}")
 
-        http_client = httpx.AsyncClient(timeout=timeout_config, follow_redirects=True)
+        # 创建 HTTP 客户端（支持代理配置）
+        from src.clients.http_client import HTTPClientPool
+
+        http_client = HTTPClientPool.create_client_with_proxy(
+            proxy_config=endpoint.proxy,
+            timeout=timeout_config,
+        )
         try:
             response_ctx = http_client.stream(
                 "POST", url, json=provider_payload, headers=provider_headers
@@ -468,8 +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(
@@ -523,10 +531,10 @@ 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(ctx.request_id)
+                    self._update_usage_to_streaming_with_ctx(ctx)
                     streaming_status_updated = True
 
                 buffer += chunk
@@ -666,6 +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
@@ -678,6 +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
@@ -698,7 +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
 
@@ -779,12 +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
 
@@ -810,7 +842,7 @@ class CliMessageHandlerBase(BaseMessageHandler):
 
             # 在第一次输出数据前更新状态为 streaming
             if prefetched_chunks:
-                self._update_usage_to_streaming(ctx.request_id)
+                self._update_usage_to_streaming_with_ctx(ctx)
 
             # 先处理预读的字节块
             for chunk in prefetched_chunks:
@@ -1108,8 +1140,10 @@ class CliMessageHandlerBase(BaseMessageHandler):
             async for chunk in stream_generator:
                 yield chunk
         except asyncio.CancelledError:
-            ctx.status_code = 499
-            ctx.error_message = "Client disconnected"
+            # 如果响应已完成，不标记为失败
+            if not ctx.has_completion:
+                ctx.status_code = 499
+                ctx.error_message = "Client disconnected"
             raise
         except httpx.TimeoutException as e:
             ctx.status_code = 504
@@ -1374,7 +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),
@@ -1419,10 +1453,14 @@ class CliMessageHandlerBase(BaseMessageHandler):
                 f"Key=***{key.api_key[-4:]}, "
                 f"原始模型={model}, 映射后={mapped_model or '无映射'}, URL模型={url_model}")
 
-            async with httpx.AsyncClient(
-                timeout=float(endpoint.timeout),
-                follow_redirects=True,
-            ) as http_client:
+            # 创建 HTTP 客户端（支持代理配置）
+            from src.clients.http_client import HTTPClientPool
+
+            http_client = HTTPClientPool.create_client_with_proxy(
+                proxy_config=endpoint.proxy,
+                timeout=httpx.Timeout(float(endpoint.timeout)),
+            )
+            async with http_client:
                 resp = await http_client.post(url, json=provider_payload, headers=provider_headers)
 
                 status_code = resp.status_code

src/api/handlers/base/content_extractors.py

@@ -0,0 +1,274 @@
+"""
+流式内容提取器 - 策略模式实现
+
+为不同 API 格式（OpenAI、Claude、Gemini）提供内容提取和 chunk 构造的抽象。
+StreamSmoother 使用这些提取器来处理不同格式的 SSE 事件。
+"""
+
+import copy
+import json
+from abc import ABC, abstractmethod
+from typing import Optional
+
+
+class ContentExtractor(ABC):
+    """
+    流式内容提取器抽象基类
+
+    定义从 SSE 事件中提取文本内容和构造新 chunk 的接口。
+    每种 API 格式（OpenAI、Claude、Gemini）需要实现自己的提取器。
+    """
+
+    @abstractmethod
+    def extract_content(self, data: dict) -> Optional[str]:
+        """
+        从 SSE 数据中提取可拆分的文本内容
+
+        Args:
+            data: 解析后的 JSON 数据
+
+        Returns:
+            提取的文本内容，如果无法提取则返回 None
+        """
+        pass
+
+    @abstractmethod
+    def create_chunk(
+        self,
+        original_data: dict,
+        new_content: str,
+        event_type: str = "",
+        is_first: bool = False,
+    ) -> bytes:
+        """
+        使用新内容构造 SSE chunk
+
+        Args:
+            original_data: 原始 JSON 数据
+            new_content: 新的文本内容
+            event_type: SSE 事件类型（某些格式需要）
+            is_first: 是否是第一个 chunk（用于保留 role 等字段）
+
+        Returns:
+            编码后的 SSE 字节数据
+        """
+        pass
+
+
+class OpenAIContentExtractor(ContentExtractor):
+    """
+    OpenAI 格式内容提取器
+
+    处理 OpenAI Chat Completions API 的流式响应格式：
+    - 数据结构: choices[0].delta.content
+    - 只在 delta 仅包含 role/content 时允许拆分，避免破坏 tool_calls 等结构
+    """
+
+    def extract_content(self, data: dict) -> Optional[str]:
+        if not isinstance(data, dict):
+            return None
+
+        choices = data.get("choices")
+        if not isinstance(choices, list) or len(choices) != 1:
+            return None
+
+        first_choice = choices[0]
+        if not isinstance(first_choice, dict):
+            return None
+
+        delta = first_choice.get("delta")
+        if not isinstance(delta, dict):
+            return None
+
+        content = delta.get("content")
+        if not isinstance(content, str):
+            return None
+
+        # 只有 delta 仅包含 role/content 时才允许拆分
+        # 避免破坏 tool_calls、function_call 等复杂结构
+        allowed_keys = {"role", "content"}
+        if not all(key in allowed_keys for key in delta.keys()):
+            return None
+
+        return content
+
+    def create_chunk(
+        self,
+        original_data: dict,
+        new_content: str,
+        event_type: str = "",
+        is_first: bool = False,
+    ) -> bytes:
+        new_data = original_data.copy()
+
+        if "choices" in new_data and new_data["choices"]:
+            new_choices = []
+            for choice in new_data["choices"]:
+                new_choice = choice.copy()
+                if "delta" in new_choice:
+                    new_delta = {}
+                    # 只有第一个 chunk 保留 role
+                    if is_first and "role" in new_choice["delta"]:
+                        new_delta["role"] = new_choice["delta"]["role"]
+                    new_delta["content"] = new_content
+                    new_choice["delta"] = new_delta
+                new_choices.append(new_choice)
+            new_data["choices"] = new_choices
+
+        return f"data: {json.dumps(new_data, ensure_ascii=False)}\n\n".encode("utf-8")
+
+
+class ClaudeContentExtractor(ContentExtractor):
+    """
+    Claude 格式内容提取器
+
+    处理 Claude Messages API 的流式响应格式：
+    - 事件类型: content_block_delta
+    - 数据结构: delta.type=text_delta, delta.text
+    """
+
+    def extract_content(self, data: dict) -> Optional[str]:
+        if not isinstance(data, dict):
+            return None
+
+        # 检查事件类型
+        if data.get("type") != "content_block_delta":
+            return None
+
+        delta = data.get("delta", {})
+        if not isinstance(delta, dict):
+            return None
+
+        # 检查 delta 类型
+        if delta.get("type") != "text_delta":
+            return None
+
+        text = delta.get("text")
+        if not isinstance(text, str):
+            return None
+
+        return text
+
+    def create_chunk(
+        self,
+        original_data: dict,
+        new_content: str,
+        event_type: str = "",
+        is_first: bool = False,
+    ) -> bytes:
+        new_data = original_data.copy()
+
+        if "delta" in new_data:
+            new_delta = new_data["delta"].copy()
+            new_delta["text"] = new_content
+            new_data["delta"] = new_delta
+
+        # Claude 格式需要 event: 前缀
+        event_name = event_type or "content_block_delta"
+        return f"event: {event_name}\ndata: {json.dumps(new_data, ensure_ascii=False)}\n\n".encode(
+            "utf-8"
+        )
+
+
+class GeminiContentExtractor(ContentExtractor):
+    """
+    Gemini 格式内容提取器
+
+    处理 Gemini API 的流式响应格式：
+    - 数据结构: candidates[0].content.parts[0].text
+    - 只有纯文本块才拆分
+    """
+
+    def extract_content(self, data: dict) -> Optional[str]:
+        if not isinstance(data, dict):
+            return None
+
+        candidates = data.get("candidates")
+        if not isinstance(candidates, list) or len(candidates) != 1:
+            return None
+
+        first_candidate = candidates[0]
+        if not isinstance(first_candidate, dict):
+            return None
+
+        content = first_candidate.get("content", {})
+        if not isinstance(content, dict):
+            return None
+
+        parts = content.get("parts", [])
+        if not isinstance(parts, list) or len(parts) != 1:
+            return None
+
+        first_part = parts[0]
+        if not isinstance(first_part, dict):
+            return None
+
+        text = first_part.get("text")
+        # 只有纯文本块（只有 text 字段）才拆分
+        if not isinstance(text, str) or len(first_part) != 1:
+            return None
+
+        return text
+
+    def create_chunk(
+        self,
+        original_data: dict,
+        new_content: str,
+        event_type: str = "",
+        is_first: bool = False,
+    ) -> bytes:
+        new_data = copy.deepcopy(original_data)
+
+        if "candidates" in new_data and new_data["candidates"]:
+            first_candidate = new_data["candidates"][0]
+            if "content" in first_candidate:
+                content = first_candidate["content"]
+                if "parts" in content and content["parts"]:
+                    content["parts"][0]["text"] = new_content
+
+        return f"data: {json.dumps(new_data, ensure_ascii=False)}\n\n".encode("utf-8")
+
+
+# 提取器注册表
+_EXTRACTORS: dict[str, type[ContentExtractor]] = {
+    "openai": OpenAIContentExtractor,
+    "claude": ClaudeContentExtractor,
+    "gemini": GeminiContentExtractor,
+}
+
+
+def get_extractor(format_name: str) -> Optional[ContentExtractor]:
+    """
+    根据格式名获取对应的内容提取器实例
+
+    Args:
+        format_name: 格式名称（openai, claude, gemini）
+
+    Returns:
+        对应的提取器实例，如果格式不支持则返回 None
+    """
+    extractor_class = _EXTRACTORS.get(format_name.lower())
+    if extractor_class:
+        return extractor_class()
+    return None
+
+
+def register_extractor(format_name: str, extractor_class: type[ContentExtractor]) -> None:
+    """
+    注册新的内容提取器
+
+    Args:
+        format_name: 格式名称
+        extractor_class: 提取器类
+    """
+    _EXTRACTORS[format_name.lower()] = extractor_class
+
+
+def get_extractor_formats() -> list[str]:
+    """
+    获取所有已注册的格式名称列表
+
+    Returns:
+        格式名称列表
+    """
+    return list(_EXTRACTORS.keys())

src/api/handlers/base/stream_processor.py

@@ -6,30 +6,47 @@
 2. 响应流生成
 3. 预读和嵌套错误检测
 4. 客户端断开检测
+5. 流式平滑输出
 """
 
 import asyncio
 import codecs
 import json
-import time
+from dataclasses import dataclass
 from typing import Any, AsyncGenerator, Callable, Optional
 
 import httpx
 
+from src.api.handlers.base.content_extractors import (
+    ContentExtractor,
+    get_extractor,
+    get_extractor_formats,
+)
 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
+class StreamSmoothingConfig:
+    """流式平滑输出配置"""
+
+    enabled: bool = False
+    chunk_size: int = 20
+    delay_ms: int = 8
 
 
 class StreamProcessor:
     """
     流式响应处理器
 
-    负责处理 SSE 流的解析、错误检测和响应生成。
+    负责处理 SSE 流的解析、错误检测、响应生成和平滑输出。
     从 ChatHandlerBase 中提取，使其职责更加单一。
     """
 
@@ -40,6 +57,7 @@ class StreamProcessor:
         on_streaming_start: Optional[Callable[[], None]] = None,
         *,
         collect_text: bool = False,
+        smoothing_config: Optional[StreamSmoothingConfig] = None,
     ):
         """
         初始化流处理器
@@ -48,11 +66,17 @@ class StreamProcessor:
             request_id: 请求 ID（用于日志）
             default_parser: 默认响应解析器
             on_streaming_start: 流开始时的回调（用于更新状态）
+            collect_text: 是否收集文本内容
+            smoothing_config: 流式平滑输出配置
         """
         self.request_id = request_id
         self.default_parser = default_parser
         self.on_streaming_start = on_streaming_start
         self.collect_text = collect_text
+        self.smoothing_config = smoothing_config or StreamSmoothingConfig()
+
+        # 内容提取器缓存
+        self._extractors: dict[str, ContentExtractor] = {}
 
     def get_parser_for_provider(self, ctx: StreamContext) -> ResponseParser:
         """
@@ -127,6 +151,13 @@ class StreamProcessor:
         if event_type in ("response.completed", "message_stop"):
             ctx.has_completion = True
 
+        # 检查 OpenAI 格式的 finish_reason
+        choices = data.get("choices", [])
+        if choices and isinstance(choices, list) and len(choices) > 0:
+            finish_reason = choices[0].get("finish_reason")
+            if finish_reason is not None:
+                ctx.has_completion = True
+
     async def prefetch_and_check_error(
         self,
         byte_iterator: Any,
@@ -141,6 +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
@@ -153,6 +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)
@@ -163,7 +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
 
@@ -233,10 +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
 
@@ -369,7 +426,7 @@ class StreamProcessor:
             sse_parser: SSE 解析器
             line: 原始行数据
         """
-        # SSEEventParser 以“去掉换行符”的单行文本作为输入；这里统一剔除 CR/LF，
+        # SSEEventParser 以"去掉换行符"的单行文本作为输入；这里统一剔除 CR/LF，
         # 避免把空行误判成 "\n" 并导致事件边界解析错误。
         normalized_line = line.rstrip("\r\n")
         events = sse_parser.feed_line(normalized_line)
@@ -400,32 +457,201 @@ class StreamProcessor:
             响应数据块
         """
         try:
-            # 断连检查频率：每次 await 都会引入调度开销，过于频繁会让流式"发一段停一段"
-            # 这里按时间间隔节流，兼顾及时停止上游读取与吞吐平滑性。
-            next_disconnect_check_at = 0.0
-            disconnect_check_interval_s = 0.25
+            # 使用后台任务检查断连，完全不阻塞流式传输
+            disconnected = False
 
-            async for chunk in stream_generator:
-                now = time.monotonic()
-                if now >= next_disconnect_check_at:
-                    next_disconnect_check_at = now + disconnect_check_interval_s
+            async def check_disconnect_background() -> None:
+                nonlocal disconnected
+                while not disconnected and not ctx.has_completion:
+                    await asyncio.sleep(0.5)
                     if await is_disconnected():
-                        logger.warning(f"ID:{self.request_id} | Client disconnected")
-                        ctx.status_code = 499  # Client Closed Request
-                        ctx.error_message = "client_disconnected"
-
+                        disconnected = True
                         break
-                yield chunk
-        except asyncio.CancelledError:
-            ctx.status_code = 499
-            ctx.error_message = "client_disconnected"
 
+            # 启动后台检查任务
+            check_task = asyncio.create_task(check_disconnect_background())
+
+            try:
+                async for chunk in stream_generator:
+                    if disconnected:
+                        # 如果响应已完成，客户端断开不算失败
+                        if ctx.has_completion:
+                            logger.info(
+                                f"ID:{self.request_id} | Client disconnected after completion"
+                            )
+                        else:
+                            logger.warning(f"ID:{self.request_id} | Client disconnected")
+                            ctx.status_code = 499
+                            ctx.error_message = "client_disconnected"
+                        break
+                    yield chunk
+            finally:
+                check_task.cancel()
+                try:
+                    await check_task
+                except asyncio.CancelledError:
+                    pass
+        except asyncio.CancelledError:
+            # 如果响应已完成，不标记为失败
+            if not ctx.has_completion:
+                ctx.status_code = 499
+                ctx.error_message = "client_disconnected"
             raise
         except Exception as e:
             ctx.status_code = 500
             ctx.error_message = str(e)
             raise
 
+    async def create_smoothed_stream(
+        self,
+        stream_generator: AsyncGenerator[bytes, None],
+    ) -> AsyncGenerator[bytes, None]:
+        """
+        创建平滑输出的流生成器
+
+        如果启用了平滑输出，将大 chunk 拆分成小块并添加微小延迟。
+        否则直接透传原始流。
+
+        Args:
+            stream_generator: 原始流生成器
+
+        Yields:
+            平滑处理后的响应数据块
+        """
+        if not self.smoothing_config.enabled:
+            # 未启用平滑输出，直接透传
+            async for chunk in stream_generator:
+                yield chunk
+            return
+
+        # 启用平滑输出
+        buffer = b""
+        is_first_content = True
+
+        async for chunk in stream_generator:
+            buffer += chunk
+
+            # 按双换行分割 SSE 事件（标准 SSE 格式）
+            while b"\n\n" in buffer:
+                event_block, buffer = buffer.split(b"\n\n", 1)
+                event_str = event_block.decode("utf-8", errors="replace")
+
+                # 解析事件块
+                lines = event_str.strip().split("\n")
+                data_str = None
+                event_type = ""
+
+                for line in lines:
+                    line = line.rstrip("\r")
+                    if line.startswith("event: "):
+                        event_type = line[7:].strip()
+                    elif line.startswith("data: "):
+                        data_str = line[6:]
+
+                # 没有 data 行，直接透传
+                if data_str is None:
+                    yield event_block + b"\n\n"
+                    continue
+
+                # [DONE] 直接透传
+                if data_str.strip() == "[DONE]":
+                    yield event_block + b"\n\n"
+                    continue
+
+                # 尝试解析 JSON
+                try:
+                    data = json.loads(data_str)
+                except json.JSONDecodeError:
+                    yield event_block + b"\n\n"
+                    continue
+
+                # 检测格式并提取内容
+                content, extractor = self._detect_format_and_extract(data)
+
+                # 只有内容长度大于 1 才需要平滑处理
+                if content and len(content) > 1 and extractor:
+                    # 获取配置的延迟
+                    delay_seconds = self._calculate_delay()
+
+                    # 拆分内容
+                    content_chunks = self._split_content(content)
+
+                    for i, sub_content in enumerate(content_chunks):
+                        is_first = is_first_content and i == 0
+
+                        # 使用提取器创建新 chunk
+                        sse_chunk = extractor.create_chunk(
+                            data,
+                            sub_content,
+                            event_type=event_type,
+                            is_first=is_first,
+                        )
+
+                        yield sse_chunk
+
+                        # 除了最后一个块，其他块之间加延迟
+                        if i < len(content_chunks) - 1:
+                            await asyncio.sleep(delay_seconds)
+
+                    is_first_content = False
+                else:
+                    # 不需要拆分，直接透传
+                    yield event_block + b"\n\n"
+                    if content:
+                        is_first_content = False
+
+        # 处理剩余数据
+        if buffer:
+            yield buffer
+
+    def _get_extractor(self, format_name: str) -> Optional[ContentExtractor]:
+        """获取或创建格式对应的提取器（带缓存）"""
+        if format_name not in self._extractors:
+            extractor = get_extractor(format_name)
+            if extractor:
+                self._extractors[format_name] = extractor
+        return self._extractors.get(format_name)
+
+    def _detect_format_and_extract(
+        self, data: dict
+    ) -> tuple[Optional[str], Optional[ContentExtractor]]:
+        """
+        检测数据格式并提取内容
+
+        依次尝试各格式的提取器，返回第一个成功提取内容的结果。
+
+        Returns:
+            (content, extractor): 提取的内容和对应的提取器
+        """
+        for format_name in get_extractor_formats():
+            extractor = self._get_extractor(format_name)
+            if extractor:
+                content = extractor.extract_content(data)
+                if content is not None:
+                    return content, extractor
+
+        return None, None
+
+    def _calculate_delay(self) -> float:
+        """获取配置的延迟（秒）"""
+        return self.smoothing_config.delay_ms / 1000.0
+
+    def _split_content(self, content: str) -> list[str]:
+        """
+        按块拆分文本
+        """
+        chunk_size = self.smoothing_config.chunk_size
+        text_length = len(content)
+
+        if text_length <= chunk_size:
+            return [content]
+
+        # 按块拆分
+        chunks = []
+        for i in range(0, text_length, chunk_size):
+            chunks.append(content[i : i + chunk_size])
+        return chunks
+
     async def _cleanup(
         self,
         response_ctx: Any,
@@ -440,3 +666,128 @@ class StreamProcessor:
             await http_client.aclose()
         except Exception:
             pass
+
+
+async def create_smoothed_stream(
+    stream_generator: AsyncGenerator[bytes, None],
+    chunk_size: int = 20,
+    delay_ms: int = 8,
+) -> AsyncGenerator[bytes, None]:
+    """
+    独立的平滑流生成函数
+
+    供 CLI handler 等场景使用，无需创建完整的 StreamProcessor 实例。
+
+    Args:
+        stream_generator: 原始流生成器
+        chunk_size: 每块字符数
+        delay_ms: 每块之间的延迟毫秒数
+
+    Yields:
+        平滑处理后的响应数据块
+    """
+    processor = _LightweightSmoother(chunk_size=chunk_size, delay_ms=delay_ms)
+    async for chunk in processor.smooth(stream_generator):
+        yield chunk
+
+
+class _LightweightSmoother:
+    """
+    轻量级平滑处理器
+
+    只包含平滑输出所需的最小逻辑，不依赖 StreamProcessor 的其他功能。
+    """
+
+    def __init__(self, chunk_size: int = 20, delay_ms: int = 8) -> None:
+        self.chunk_size = chunk_size
+        self.delay_ms = delay_ms
+        self._extractors: dict[str, ContentExtractor] = {}
+
+    def _get_extractor(self, format_name: str) -> Optional[ContentExtractor]:
+        if format_name not in self._extractors:
+            extractor = get_extractor(format_name)
+            if extractor:
+                self._extractors[format_name] = extractor
+        return self._extractors.get(format_name)
+
+    def _detect_format_and_extract(
+        self, data: dict
+    ) -> tuple[Optional[str], Optional[ContentExtractor]]:
+        for format_name in get_extractor_formats():
+            extractor = self._get_extractor(format_name)
+            if extractor:
+                content = extractor.extract_content(data)
+                if content is not None:
+                    return content, extractor
+        return None, None
+
+    def _calculate_delay(self) -> float:
+        return self.delay_ms / 1000.0
+
+    def _split_content(self, content: str) -> list[str]:
+        text_length = len(content)
+        if text_length <= self.chunk_size:
+            return [content]
+        return [content[i : i + self.chunk_size] for i in range(0, text_length, self.chunk_size)]
+
+    async def smooth(
+        self, stream_generator: AsyncGenerator[bytes, None]
+    ) -> AsyncGenerator[bytes, None]:
+        buffer = b""
+        is_first_content = True
+
+        async for chunk in stream_generator:
+            buffer += chunk
+
+            while b"\n\n" in buffer:
+                event_block, buffer = buffer.split(b"\n\n", 1)
+                event_str = event_block.decode("utf-8", errors="replace")
+
+                lines = event_str.strip().split("\n")
+                data_str = None
+                event_type = ""
+
+                for line in lines:
+                    line = line.rstrip("\r")
+                    if line.startswith("event: "):
+                        event_type = line[7:].strip()
+                    elif line.startswith("data: "):
+                        data_str = line[6:]
+
+                if data_str is None:
+                    yield event_block + b"\n\n"
+                    continue
+
+                if data_str.strip() == "[DONE]":
+                    yield event_block + b"\n\n"
+                    continue
+
+                try:
+                    data = json.loads(data_str)
+                except json.JSONDecodeError:
+                    yield event_block + b"\n\n"
+                    continue
+
+                content, extractor = self._detect_format_and_extract(data)
+
+                if content and len(content) > 1 and extractor:
+                    delay_seconds = self._calculate_delay()
+                    content_chunks = self._split_content(content)
+
+                    for i, sub_content in enumerate(content_chunks):
+                        is_first = is_first_content and i == 0
+                        sse_chunk = extractor.create_chunk(
+                            data, sub_content, event_type=event_type, is_first=is_first
+                        )
+                        yield sse_chunk
+                        if i < len(content_chunks) - 1:
+                            await asyncio.sleep(delay_seconds)
+
+                    is_first_content = False
+                else:
+                    yield event_block + b"\n\n"
+                    if content:
+                        is_first_content = False
+
+        if buffer:
+            yield buffer

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/handler.py

@@ -131,10 +131,5 @@ class ClaudeChatHandler(ChatHandlerBase):
         Returns:
             规范化后的响应
         """
-        if self.response_normalizer and self.response_normalizer.should_normalize(response):
-            result: Dict[str, Any] = self.response_normalizer.normalize_claude_response(
-                response_data=response,
-                request_id=self.request_id,
-            )
-            return result
+        # 作为中转站，直接透传响应，不做标准化处理
         return response

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/handler.py

@@ -148,17 +148,6 @@ class GeminiChatHandler(ChatHandlerBase):
 
         Returns:
             规范化后的响应
-
-        TODO: 如果需要，实现响应规范化逻辑
         """
-        # 可选：使用 response_normalizer 进行规范化
-        # if (
-        #     self.response_normalizer
-        #     and self.response_normalizer.should_normalize(response)
-        # ):
-        #     return self.response_normalizer.normalize_gemini_response(
-        #         response_data=response,
-        #         request_id=self.request_id,
-        #         strict=False,
-        #     )
+        # 作为中转站，直接透传响应，不做标准化处理
         return response

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/handler.py

@@ -128,10 +128,5 @@ class OpenAIChatHandler(ChatHandlerBase):
         Returns:
             规范化后的响应
         """
-        if self.response_normalizer and self.response_normalizer.should_normalize(response):
-            return self.response_normalizer.normalize_openai_response(
-                response_data=response,
-                request_id=self.request_id,
-                strict=False,
-            )
+        # 作为中转站，直接透传响应，不做标准化处理
         return response

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/clients/http_client.py

@@ -5,12 +5,55 @@
 
 from contextlib import asynccontextmanager
 from typing import Any, Dict, Optional
+from urllib.parse import quote, urlparse
 
 import httpx
 
 from src.core.logger import logger
 
 
+def build_proxy_url(proxy_config: Dict[str, Any]) -> Optional[str]:
+    """
+    根据代理配置构建完整的代理 URL
+
+    Args:
+        proxy_config: 代理配置字典，包含 url, username, password, enabled
+
+    Returns:
+        完整的代理 URL，如 socks5://user:pass@host:port
+        如果 enabled=False 或无配置，返回 None
+    """
+    if not proxy_config:
+        return None
+
+    # 检查 enabled 字段，默认为 True（兼容旧数据）
+    if not proxy_config.get("enabled", True):
+        return None
+
+    proxy_url = proxy_config.get("url")
+    if not proxy_url:
+        return None
+
+    username = proxy_config.get("username")
+    password = proxy_config.get("password")
+
+    # 只要有用户名就添加认证信息（密码可以为空）
+    if username:
+        parsed = urlparse(proxy_url)
+        # URL 编码用户名和密码，处理特殊字符（如 @, :, /）
+        encoded_username = quote(username, safe="")
+        encoded_password = quote(password, safe="") if password else ""
+        # 重新构建带认证的代理 URL
+        if encoded_password:
+            auth_proxy = f"{parsed.scheme}://{encoded_username}:{encoded_password}@{parsed.netloc}"
+        else:
+            auth_proxy = f"{parsed.scheme}://{encoded_username}@{parsed.netloc}"
+        if parsed.path:
+            auth_proxy += parsed.path
+        return auth_proxy
+
+    return proxy_url
+
 
 class HTTPClientPool:
     """
@@ -121,6 +164,44 @@ class HTTPClientPool:
         finally:
             await client.aclose()
 
+    @classmethod
+    def create_client_with_proxy(
+        cls,
+        proxy_config: Optional[Dict[str, Any]] = None,
+        timeout: Optional[httpx.Timeout] = None,
+        **kwargs: Any,
+    ) -> httpx.AsyncClient:
+        """
+        创建带代理配置的HTTP客户端
+
+        Args:
+            proxy_config: 代理配置字典，包含 url, username, password
+            timeout: 超时配置
+            **kwargs: 其他 httpx.AsyncClient 配置参数
+
+        Returns:
+            配置好的 httpx.AsyncClient 实例
+        """
+        config: Dict[str, Any] = {
+            "http2": False,
+            "verify": True,
+            "follow_redirects": True,
+        }
+
+        if timeout:
+            config["timeout"] = timeout
+        else:
+            config["timeout"] = httpx.Timeout(10.0, read=300.0)
+
+        # 添加代理配置
+        proxy_url = build_proxy_url(proxy_config) if proxy_config else None
+        if proxy_url:
+            config["proxy"] = proxy_url
+            logger.debug(f"创建带代理的HTTP客户端: {proxy_config.get('url', 'unknown')}")
+
+        config.update(kwargs)
+        return httpx.AsyncClient(**config)
+
 
 # 便捷访问函数
 def get_http_client() -> httpx.AsyncClient:

src/clients/redis_client.py

@@ -120,7 +120,7 @@ class RedisClientManager:
         if self._circuit_open_until and time.time() < self._circuit_open_until:
             remaining = self._circuit_open_until - time.time()
             logger.warning(
-                "Redis 客户端处于熔断状态，跳过初始化，剩余 %.1f 秒 (last_error: %s)",
+                "Redis 客户端处于熔断状态，跳过初始化，剩余 {:.1f} 秒 (last_error: {})",
                 remaining,
                 self._last_error,
             )
@@ -200,7 +200,7 @@ class RedisClientManager:
             if self._consecutive_failures >= self._circuit_threshold:
                 self._circuit_open_until = time.time() + self._circuit_reset_seconds
                 logger.warning(
-                    "Redis 初始化连续失败 %s 次，开启熔断 %s 秒。"
+                    "Redis 初始化连续失败 {} 次，开启熔断 {} 秒。"
                     "熔断期间以下功能将降级: 缓存亲和性、分布式并发控制、RPM限流。"
                     "可通过管理 API /api/admin/system/redis/reset-circuit 手动重置。",
                     self._consecutive_failures,
@@ -267,6 +267,9 @@ async def get_redis_client(require_redis: bool = False) -> Optional[aioredis.Red
 
     if _redis_manager is None:
         _redis_manager = RedisClientManager()
+    # 如果尚未连接（例如启动时降级、或 close() 后），尝试重新初始化。
+    # initialize() 内部包含熔断器逻辑，避免频繁重试导致抖动。
+    if _redis_manager.get_client() is None:
         await _redis_manager.initialize(require_redis=require_redis)
 
     return _redis_manager.get_client()

src/config/constants.py

@@ -41,8 +41,8 @@ class CacheSize:
 class ConcurrencyDefaults:
     """并发控制默认值"""
 
-    # 自适应并发初始限制（保守值）
-    INITIAL_LIMIT = 3
+    # 自适应并发初始限制（宽松起步，遇到 429 再降低）
+    INITIAL_LIMIT = 50
 
     # 429错误后的冷却时间（分钟）- 在此期间不会增加并发限制
     COOLDOWN_AFTER_429_MINUTES = 5
@@ -67,16 +67,20 @@ class ConcurrencyDefaults:
     MIN_SAMPLES_FOR_DECISION = 5
 
     # 扩容步长 - 每次扩容增加的并发数
-    INCREASE_STEP = 1
+    INCREASE_STEP = 2
 
-    # 缩容乘数 - 遇到 429 时的缩容比例
-    DECREASE_MULTIPLIER = 0.7
+    # 缩容乘数 - 遇到 429 时基于当前并发数的缩容比例
+    # 0.85 表示降到触发 429 时并发数的 85%
+    DECREASE_MULTIPLIER = 0.85
 
     # 最大并发限制上限
-    MAX_CONCURRENT_LIMIT = 100
+    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
@@ -85,6 +89,11 @@ class ConcurrencyDefaults:
     # 探测性扩容最小请求数 - 在探测间隔内至少需要这么多请求
     PROBE_INCREASE_MIN_REQUESTS = 10
 
+    # === 缓存用户预留比例 ===
+    # 缓存用户槽位预留比例（新用户可用 1 - 此值）
+    # 0.1 表示缓存用户预留 10%，新用户可用 90%
+    CACHE_RESERVATION_RATIO = 0.1
+
 
 class CircuitBreakerDefaults:
     """熔断器配置默认值（滑动窗口 + 半开状态模式）

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"
@@ -105,6 +106,13 @@ class Config:
         self.llm_api_rate_limit = int(os.getenv("LLM_API_RATE_LIMIT", "100"))
         self.public_api_rate_limit = int(os.getenv("PUBLIC_API_RATE_LIMIT", "60"))
 
+        # 可信代理配置
+        # TRUSTED_PROXY_COUNT: 信任的代理层数（默认 1，即信任最近一层代理）
+        # 设置为 0 表示不信任任何代理头，直接使用连接 IP
+        # 当服务部署在 Nginx/CloudFlare 等反向代理后面时，设置为对应的代理层数
+        # 如果服务直接暴露公网，应设置为 0 以防止 IP 伪造
+        self.trusted_proxy_count = int(os.getenv("TRUSTED_PROXY_COUNT", "1"))
+
         # 异常处理配置
         # 设置为 True 时，ProxyException 会传播到路由层以便记录 provider_request_headers
         # 设置为 False 时，使用全局异常处理器统一处理
@@ -122,9 +130,21 @@ class Config:
 
         # 并发控制配置
         # CONCURRENCY_SLOT_TTL: 并发槽位 TTL（秒），防止死锁
-        # CACHE_RESERVATION_RATIO: 缓存用户预留比例（默认 30%）
+        # CACHE_RESERVATION_RATIO: 缓存用户预留比例（默认 10%，新用户可用 90%）
         self.concurrency_slot_ttl = int(os.getenv("CONCURRENCY_SLOT_TTL", "600"))
-        self.cache_reservation_ratio = float(os.getenv("CACHE_RESERVATION_RATIO", "0.3"))
+        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"))
@@ -134,8 +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"))
+        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()
@@ -158,6 +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
@@ -205,6 +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/batch_committer.py

@@ -46,6 +46,11 @@ class BatchCommitter:
 
     def mark_dirty(self, session: Session):
         """标记 Session 有待提交的更改"""
+        # 请求级事务由中间件统一 commit/rollback；避免后台任务在请求中途误提交。
+        if session is None:
+            return
+        if session.info.get("managed_by_middleware"):
+            return
         self._pending_sessions.add(session)
 
     async def _batch_commit_loop(self):

src/core/context.py

@@ -1,168 +0,0 @@
-"""
-统一的请求上下文
-
-RequestContext 贯穿整个请求生命周期，包含所有请求相关信息。
-这确保了数据在各层之间传递时不会丢失。
-
-使用方式：
-1. Pipeline 层创建 RequestContext
-2. 各层通过 context 访问和更新信息
-3. Adapter 层使用 context 记录 Usage
-"""
-
-import time
-import uuid
-from dataclasses import dataclass, field
-from typing import Any, Dict, Optional
-
-
-@dataclass
-class RequestContext:
-    """
-    请求上下文 - 贯穿整个请求生命周期
-
-    设计原则：
-    1. 在请求开始时创建，包含所有已知信息
-    2. 在请求执行过程中逐步填充 Provider 信息
-    3. 在请求结束时用于记录 Usage
-    """
-
-    # ==================== 请求标识 ====================
-    request_id: str
-
-    # ==================== 认证信息 ====================
-    user: Any  # User model
-    api_key: Any  # ApiKey model
-    db: Any  # Database session
-
-    # ==================== 请求信息 ====================
-    api_format: str  # CLAUDE, OPENAI, GEMINI, etc.
-    model: str  # 用户请求的模型名
-    is_stream: bool = False
-
-    # ==================== 原始请求 ====================
-    original_headers: Dict[str, str] = field(default_factory=dict)
-    original_body: Dict[str, Any] = field(default_factory=dict)
-
-    # ==================== 客户端信息 ====================
-    client_ip: str = "unknown"
-    user_agent: str = ""
-
-    # ==================== 计时 ====================
-    start_time: float = field(default_factory=time.time)
-
-    # ==================== Provider 信息（请求执行后填充）====================
-    provider_name: Optional[str] = None
-    provider_id: Optional[str] = None
-    endpoint_id: Optional[str] = None
-    provider_api_key_id: Optional[str] = None
-
-    # ==================== 模型映射信息 ====================
-    resolved_model: Optional[str] = None  # 映射后的模型名
-    original_model: Optional[str] = None  # 原始模型名（用于价格计算）
-
-    # ==================== 请求/响应头 ====================
-    provider_request_headers: Dict[str, str] = field(default_factory=dict)
-    provider_response_headers: Dict[str, str] = field(default_factory=dict)
-
-    # ==================== 追踪信息 ====================
-    attempt_id: Optional[str] = None
-
-    # ==================== 能力需求 ====================
-    capability_requirements: Dict[str, bool] = field(default_factory=dict)
-    # 运行时计算的能力需求，来源于:
-    # 1. 用户 model_capability_settings
-    # 2. 用户 ApiKey.force_capabilities
-    # 3. 请求头 X-Require-Capability
-    # 4. 失败重试时动态添加
-
-    @classmethod
-    def create(
-        cls,
-        *,
-        db: Any,
-        user: Any,
-        api_key: Any,
-        api_format: str,
-        model: str,
-        is_stream: bool = False,
-        original_headers: Optional[Dict[str, str]] = None,
-        original_body: Optional[Dict[str, Any]] = None,
-        client_ip: str = "unknown",
-        user_agent: str = "",
-        request_id: Optional[str] = None,
-    ) -> "RequestContext":
-        """创建请求上下文"""
-        return cls(
-            request_id=request_id or str(uuid.uuid4()),
-            db=db,
-            user=user,
-            api_key=api_key,
-            api_format=api_format,
-            model=model,
-            is_stream=is_stream,
-            original_headers=original_headers or {},
-            original_body=original_body or {},
-            client_ip=client_ip,
-            user_agent=user_agent,
-            original_model=model,  # 初始时原始模型等于请求模型
-        )
-
-    def update_provider_info(
-        self,
-        *,
-        provider_name: str,
-        provider_id: str,
-        endpoint_id: str,
-        provider_api_key_id: str,
-        resolved_model: Optional[str] = None,
-    ) -> None:
-        """更新 Provider 信息（请求执行后调用）"""
-        self.provider_name = provider_name
-        self.provider_id = provider_id
-        self.endpoint_id = endpoint_id
-        self.provider_api_key_id = provider_api_key_id
-        if resolved_model:
-            self.resolved_model = resolved_model
-
-    def update_headers(
-        self,
-        *,
-        request_headers: Optional[Dict[str, str]] = None,
-        response_headers: Optional[Dict[str, str]] = None,
-    ) -> None:
-        """更新请求/响应头"""
-        if request_headers:
-            self.provider_request_headers = request_headers
-        if response_headers:
-            self.provider_response_headers = response_headers
-
-    @property
-    def elapsed_ms(self) -> int:
-        """计算已经过的时间（毫秒）"""
-        return int((time.time() - self.start_time) * 1000)
-
-    @property
-    def effective_model(self) -> str:
-        """获取有效的模型名（映射后优先）"""
-        return self.resolved_model or self.model
-
-    @property
-    def billing_model(self) -> str:
-        """获取计费模型名（原始模型优先）"""
-        return self.original_model or self.model
-
-    def to_metadata_dict(self) -> Dict[str, Any]:
-        """转换为元数据字典（用于 Usage 记录）"""
-        return {
-            "api_format": self.api_format,
-            "provider": self.provider_name or "unknown",
-            "model": self.effective_model,
-            "original_model": self.billing_model,
-            "provider_id": self.provider_id,
-            "provider_endpoint_id": self.endpoint_id,
-            "provider_api_key_id": self.provider_api_key_id,
-            "provider_request_headers": self.provider_request_headers,
-            "provider_response_headers": self.provider_response_headers,
-            "attempt_id": self.attempt_id,
-        }

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