Merge pull request #418 from Avilianb/feat/codex-responses-websocket

feat: support Codex Responses WebSocket transport
2026-10-05 00:47:48 +08:00 · 2026-05-15 12:49:45 +08:00 · 2026-05-10 01:59:48 +08:00 · 2026-05-03 02:11:16 +08:00 · 2026-04-30 18:07:00 +08:00 · 2026-04-17 10:37:19 +08:00
658 changed files with 69560 additions and 27960 deletions
@@ -49,6 +49,40 @@ ADMIN_PASSWORD=admin123456
 # max-requests-jitter 会自动设置为 MAX_REQUESTS/20 (5%)
 # MAX_REQUESTS=4000

+# glibc malloc arena 上限（默认 2）
+# 降低 malloc 内存碎片，减少 gunicorn worker RSS
+# MALLOC_ARENA_MAX=2
+
+# HTTP 连接池上限（默认总预算约 200，按 worker 平分）
+# 如果容器内存偏高，可继续下调；例如 2 worker 时设为 80-100
+# HTTP_MAX_CONNECTIONS=100
+
+# HTTP 保活连接数（默认约为 max_connections 的 30%）
+# HTTP_KEEPALIVE_CONNECTIONS=30
+
+# HTTP 代理/Tunnel 客户端空闲清理（默认每 5 分钟扫描，空闲 600 秒即关闭）
+# HTTP_CLIENT_IDLE_CLEANUP_INTERVAL_MINUTES=5
+# HTTP_CLIENT_IDLE_CLEANUP_MAX_SECONDS=600
+
+# curl_cffi session 池上限（默认 20，按 impersonate + proxy 组合缓存）
+# CURL_CFFI_MAX_SESSIONS=20
+
+# 流式响应块缓存上限（单位 MB，默认 2）
+# 说明：
+# - 这是单个流式请求可保留的“解析后响应块”内存上限，不是全局上限
+# - 粗略峰值内存 ≈ 并发流数量 × RESPONSE_CHUNKS_MAX_SIZE_MB
+#   例如：100 并发、2MB 上限，理论峰值约 200MB
+# - 建议：
+#   - 内存敏感环境：1
+#   - 通用生产环境：2（默认）
+#   - 需要更多调试上下文：4
+# RESPONSE_CHUNKS_MAX_SIZE_MB=2
+
+# 流式空闲超时（单位秒，默认 30）
+# 当流已经开始但连续一段时间没有任何新 chunk 时，提前中断并返回 504，
+# 避免一直等到 worker 超时（如 300s）
+# STREAM_IDLE_TIMEOUT_SECONDS=30
+
 # API Key 前缀（默认 sk）
 # API_KEY_PREFIX=sk

@@ -60,6 +94,14 @@ ADMIN_PASSWORD=admin123456
 # 默认: * (允许所有源)
 # CORS_ORIGINS=*

+# 启动预热配置（默认启用，降低首请求冷启动延迟）
+# 是否启用启动期预热任务（默认 true）
+# STARTUP_WARMUP_ENABLED=true
+# /readyz 是否等待预热完成（默认 true）
+# STARTUP_WARMUP_GATE_READINESS=true
+# 预热时优先 bootstrap 的 provider_type 列表（逗号分隔；留空表示自动探测）
+# STARTUP_WARMUP_PROVIDER_TYPES=codex,kiro
+
 # ==================== 计费系统（可选） ====================
 # Video/Image/Audio 缺失 billing_rule 时是否拒绝请求（默认 false：允许请求但 cost=0 并告警）
 # BILLING_REQUIRE_RULE=false
@@ -7,12 +7,6 @@ on:

 permissions:
  contents: write
-  packages: write
-
-env:
-  REGISTRY: ghcr.io
-  GHCR_IMAGE: fawney19/aether-hub
-  DOCKERHUB_IMAGE: fawney19/aether-hub

 jobs:
  build:
@@ -30,7 +24,7 @@ jobs:
            use_cross: true

    steps:
-      - uses: actions/checkout@v4
+      - uses: actions/checkout@v5

      - name: Install Rust toolchain
        uses: dtolnay/rust-toolchain@stable
@@ -65,7 +59,7 @@ jobs:
          tar czf ../../../../aether-hub-${{ matrix.name }}.tar.gz aether-hub

      - name: Upload artifact
-        uses: actions/upload-artifact@v4
+        uses: actions/upload-artifact@v5
        with:
          name: aether-hub-${{ matrix.name }}
          path: aether-hub-*.tar.gz
@@ -77,7 +71,7 @@ jobs:
    if: startsWith(github.ref, 'refs/tags/')
    steps:
      - name: Download all artifacts
-        uses: actions/download-artifact@v4
+        uses: actions/download-artifact@v5
        with:
          merge-multiple: true
          path: artifacts
@@ -95,78 +89,3 @@ jobs:
            artifacts/aether-hub-*
            artifacts/SHA256SUMS.txt
          fail_on_unmatched_files: true
-
-  docker:
-    needs: build
-    runs-on: ubuntu-latest
-    if: startsWith(github.ref, 'refs/tags/')
-    steps:
-      - uses: actions/checkout@v4
-
-      - name: Download Linux artifacts
-        uses: actions/download-artifact@v4
-        with:
-          pattern: aether-hub-linux-*
-          merge-multiple: true
-          path: artifacts
-
-      - name: Prepare binaries
-        run: |
-          mkdir -p aether-hub/build/linux-amd64 aether-hub/build/linux-arm64
-          tar xzf artifacts/aether-hub-linux-amd64.tar.gz -C aether-hub/build/linux-amd64
-          tar xzf artifacts/aether-hub-linux-arm64.tar.gz -C aether-hub/build/linux-arm64
-
-      - name: Generate CI Dockerfile
-        run: |
-          cat > aether-hub/Dockerfile.ci << 'EOF'
-          FROM debian:bookworm-slim
-          ARG TARGETARCH
-          RUN apt-get update && apt-get install -y --no-install-recommends ca-certificates && rm -rf /var/lib/apt/lists/*
-          COPY build/linux-${TARGETARCH}/aether-hub /usr/local/bin/aether-hub
-          EXPOSE 8085
-          ENTRYPOINT ["/usr/local/bin/aether-hub"]
-          CMD ["--bind", "0.0.0.0:8085"]
-          EOF
-
-      - name: Set up QEMU
-        uses: docker/setup-qemu-action@v3
-
-      - name: Set up Docker Buildx
-        uses: docker/setup-buildx-action@v3
-
-      - name: Log in to GHCR
-        uses: docker/login-action@v3
-        with:
-          registry: ${{ env.REGISTRY }}
-          username: ${{ github.actor }}
-          password: ${{ secrets.GITHUB_TOKEN }}
-
-      - name: Log in to Docker Hub
-        uses: docker/login-action@v3
-        with:
-          username: ${{ secrets.DOCKERHUB_USERNAME }}
-          password: ${{ secrets.DOCKERHUB_TOKEN }}
-
-      - name: Extract metadata
-        id: meta
-        uses: docker/metadata-action@v5
-        with:
-          images: |
-            ${{ env.REGISTRY }}/${{ env.GHCR_IMAGE }}
-            docker.io/${{ env.DOCKERHUB_IMAGE }}
-          tags: |
-            type=match,pattern=hub-v(.*),group=1
-            type=match,pattern=hub-v(\d+\.\d+),group=1
-            type=sha,prefix=
-          flavor: |
-            latest=auto
-
-      - name: Build and push
-        uses: docker/build-push-action@v5
-        with:
-          context: ./aether-hub
-          file: ./aether-hub/Dockerfile.ci
-          push: true
-          tags: ${{ steps.meta.outputs.tags }}
-          labels: ${{ steps.meta.outputs.labels }}
-          platforms: linux/amd64,linux/arm64
@@ -44,7 +44,7 @@ jobs:
            use_cross: false

    steps:
-      - uses: actions/checkout@v4
+      - uses: actions/checkout@v5

      - name: Install Rust toolchain
        uses: dtolnay/rust-toolchain@stable
@@ -87,7 +87,7 @@ jobs:
          7z a ../../../../aether-proxy-${{ matrix.name }}.zip aether-proxy.exe

      - name: Upload artifact
-        uses: actions/upload-artifact@v4
+        uses: actions/upload-artifact@v5
        with:
          name: aether-proxy-${{ matrix.name }}
          path: |
@@ -101,7 +101,7 @@ jobs:
    if: startsWith(github.ref, 'refs/tags/')
    steps:
      - name: Download all artifacts
-        uses: actions/download-artifact@v4
+        uses: actions/download-artifact@v5
        with:
          merge-multiple: true
          path: artifacts
@@ -125,10 +125,10 @@ jobs:
    runs-on: ubuntu-latest
    if: startsWith(github.ref, 'refs/tags/')
    steps:
-      - uses: actions/checkout@v4
+      - uses: actions/checkout@v5

      - name: Download Linux artifacts
-        uses: actions/download-artifact@v4
+        uses: actions/download-artifact@v5
        with:
          pattern: aether-proxy-linux-*
          merge-multiple: true
@@ -174,7 +174,7 @@ jobs:
            latest=auto

      - name: Build and push
-        uses: docker/build-push-action@v5
+        uses: docker/build-push-action@v6
        with:
          context: ./aether-proxy
          push: true
@@ -187,7 +187,7 @@ jobs:
    runs-on: ubuntu-latest
    if: startsWith(github.ref, 'refs/tags/')
    steps:
-      - uses: actions/checkout@v4
+      - uses: actions/checkout@v5
        with:
          ref: master

@@ -18,12 +18,12 @@ jobs:
  build:
    runs-on: ubuntu-latest
    steps:
-      - uses: actions/checkout@v4
+      - uses: actions/checkout@v5

      - name: Setup Node.js
-        uses: actions/setup-node@v4
+        uses: actions/setup-node@v5
        with:
-          node-version: '20'
+          node-version: '22'
          cache: 'npm'
          cache-dependency-path: frontend/package-lock.json

@@ -41,7 +41,7 @@ jobs:
        run: cp frontend/dist/index.html frontend/dist/404.html

      - name: Setup Pages
-        uses: actions/configure-pages@v4
+        uses: actions/configure-pages@v5

      - name: Upload artifact
        uses: actions/upload-pages-artifact@v3
@@ -30,7 +30,7 @@ jobs:
    outputs:
      base_changed: ${{ steps.check.outputs.base_changed }}
    steps:
-      - uses: actions/checkout@v4
+      - uses: actions/checkout@v5

      - name: Log in to Container Registry
        uses: docker/login-action@v3
@@ -108,7 +108,7 @@ jobs:
      contents: read
      packages: write
    steps:
-      - uses: actions/checkout@v4
+      - uses: actions/checkout@v5

      - name: Set up Docker Buildx
        uses: docker/setup-buildx-action@v3
@@ -164,7 +164,7 @@ jobs:
            org.opencontainers.image.base.hash=${{ steps.hash.outputs.hash }}

      - name: Build and push base image
-        uses: docker/build-push-action@v5
+        uses: docker/build-push-action@v6
        with:
          context: .
          file: ./Dockerfile.base
@@ -204,7 +204,7 @@ jobs:
      contents: read
      packages: write
    steps:
-      - uses: actions/checkout@v4
+      - uses: actions/checkout@v5

      - name: Set up Docker Buildx
        uses: docker/setup-buildx-action@v3
@@ -271,7 +271,7 @@ jobs:

      - name: Build and push app image (amd64)
        id: build-amd64
-        uses: docker/build-push-action@v5
+        uses: docker/build-push-action@v6
        with:
          context: .
          file: ./Dockerfile.app
@@ -287,7 +287,7 @@ jobs:

      - name: Build and push app image (arm64)
        id: build-arm64
-        uses: docker/build-push-action@v5
+        uses: docker/build-push-action@v6
        with:
          context: .
          file: ./Dockerfile.app
@@ -2,6 +2,7 @@
 # Edit at https://www.toptal.com/developers/gitignore?templates=python

 # AI Assistant Configuration
+.codex/
 .claude/
 .serena/
 .gemini*/
@@ -203,6 +204,7 @@ logs/

 # Git backup
 .git.backup/
+.worktrees/

 # Database backups
 backups/
@@ -25,7 +25,12 @@ RUN --mount=type=cache,target=/var/cache/apt,sharing=locked \
    nginx \
    supervisor \
    libpq5 \
-    curl
+    curl \
+    libjemalloc2
+RUN set -eux; \
+    jemalloc_path="$(find /usr/lib -type f -name 'libjemalloc.so.2' | head -n1)"; \
+    [ -n "$jemalloc_path" ]; \
+    ln -sf "$jemalloc_path" /usr/local/lib/libjemalloc.so.2
 # 从 base 镜像复制 Python 包
 COPY --from=builder /usr/local/lib/python3.13/site-packages /usr/local/lib/python3.13/site-packages
 # 只复制需要的 Python 可执行文件
@@ -85,6 +90,11 @@ RUN printf '%s\n' \
    '    ""      $remote_addr;' \
    '}' \
    '' \
+    'map $http_upgrade $connection_upgrade {' \
+    '    default upgrade;' \
+    '    ""      "";' \
+    '}' \
+    '' \
    'server {' \
    '    listen 80;' \
    '    server_name _;' \
@@ -124,6 +134,25 @@ RUN printf '%s\n' \
    '        proxy_set_header X-Forwarded-Proto $scheme;' \
    '        proxy_set_header Upgrade $http_upgrade;' \
    '        proxy_set_header Connection "upgrade";' \
+    '        # 剥离 CF 头，防止泄露给上游或返回给客户端' \
+    '        proxy_hide_header CF-Connecting-IP;' \
+    '        proxy_hide_header CF-IPCountry;' \
+    '        proxy_hide_header CF-Ray;' \
+    '        proxy_hide_header CF-Visitor;' \
+    '        proxy_hide_header CDN-Loop;' \
+    '        proxy_hide_header True-Client-IP;' \
+    '        proxy_hide_header CF-Worker;' \
+    '        proxy_hide_header CF-EW-Via;' \
+    '        proxy_hide_header CF-Warp-Tag-ID;' \
+    '        proxy_set_header CF-Connecting-IP "";' \
+    '        proxy_set_header CF-IPCountry "";' \
+    '        proxy_set_header CF-Ray "";' \
+    '        proxy_set_header CF-Visitor "";' \
+    '        proxy_set_header CDN-Loop "";' \
+    '        proxy_set_header True-Client-IP "";' \
+    '        proxy_set_header CF-Worker "";' \
+    '        proxy_set_header CF-EW-Via "";' \
+    '        proxy_set_header CF-Warp-Tag-ID "";' \
    '        proxy_read_timeout 86400s;' \
    '        proxy_send_timeout 86400s;' \
    '    }' \
@@ -136,12 +165,22 @@ RUN printf '%s\n' \
    '        proxy_set_header X-Real-IP $real_ip;' \
    '        proxy_set_header X-Forwarded-For $forwarded_for;' \
    '        proxy_set_header X-Forwarded-Proto $scheme;' \
-    '        proxy_set_header Connection "";' \
+    '        proxy_set_header Upgrade $http_upgrade;' \
+    '        proxy_set_header Connection $connection_upgrade;' \
    '        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;' \
-    '        # 剥离 CF 头，防止泄露给上游 AI 提供商' \
+    '        # 剥离 CF 头，防止泄露给上游或返回给客户端' \
+    '        proxy_hide_header CF-Connecting-IP;' \
+    '        proxy_hide_header CF-IPCountry;' \
+    '        proxy_hide_header CF-Ray;' \
+    '        proxy_hide_header CF-Visitor;' \
+    '        proxy_hide_header CDN-Loop;' \
+    '        proxy_hide_header True-Client-IP;' \
+    '        proxy_hide_header CF-Worker;' \
+    '        proxy_hide_header CF-EW-Via;' \
+    '        proxy_hide_header CF-Warp-Tag-ID;' \
    '        proxy_set_header CF-Connecting-IP "";' \
    '        proxy_set_header CF-IPCountry "";' \
    '        proxy_set_header CF-Ray "";' \
@@ -150,6 +189,7 @@ RUN printf '%s\n' \
    '        proxy_set_header True-Client-IP "";' \
    '        proxy_set_header CF-Worker "";' \
    '        proxy_set_header CF-EW-Via "";' \
+    '        proxy_set_header CF-Warp-Tag-ID "";' \
    '        proxy_buffering off;' \
    '        proxy_cache off;' \
    '        proxy_request_buffering off;' \
@@ -169,6 +209,25 @@ RUN printf '%s\n' \
    '        proxy_set_header X-Real-IP $real_ip;' \
    '        proxy_set_header X-Forwarded-For $forwarded_for;' \
    '        proxy_set_header X-Forwarded-Proto $scheme;' \
+    '        # 剥离 CF 头，防止泄露给上游或返回给客户端' \
+    '        proxy_hide_header CF-Connecting-IP;' \
+    '        proxy_hide_header CF-IPCountry;' \
+    '        proxy_hide_header CF-Ray;' \
+    '        proxy_hide_header CF-Visitor;' \
+    '        proxy_hide_header CDN-Loop;' \
+    '        proxy_hide_header True-Client-IP;' \
+    '        proxy_hide_header CF-Worker;' \
+    '        proxy_hide_header CF-EW-Via;' \
+    '        proxy_hide_header CF-Warp-Tag-ID;' \
+    '        proxy_set_header CF-Connecting-IP "";' \
+    '        proxy_set_header CF-IPCountry "";' \
+    '        proxy_set_header CF-Ray "";' \
+    '        proxy_set_header CF-Visitor "";' \
+    '        proxy_set_header CDN-Loop "";' \
+    '        proxy_set_header True-Client-IP "";' \
+    '        proxy_set_header CF-Worker "";' \
+    '        proxy_set_header CF-EW-Via "";' \
+    '        proxy_set_header CF-Warp-Tag-ID "";' \
    '    }' \
    '' \
    '    # 所有其他路由 → 前端 SPA（先尝试静态文件，再回退到 index.html）' \
@@ -199,7 +258,7 @@ RUN printf '%s\n' \
    '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' \
+    'environment=PYTHONUNBUFFERED=1,PYTHONIOENCODING=utf-8,LANG=C.UTF-8,LC_ALL=C.UTF-8,DOCKER_CONTAINER=true,LD_PRELOAD=/usr/local/lib/libjemalloc.so.2,MALLOC_CONF="background_thread:true,dirty_decay_ms:5000,muzzy_decay_ms:5000"' \
    '' \
    '[program:tunnel-hub]' \
    'command=/usr/local/bin/aether-hub --bind 0.0.0.0:8085' \
@@ -220,6 +279,8 @@ ENV PYTHONUNBUFFERED=1 \
    PYTHONIOENCODING=utf-8 \
    LANG=C.UTF-8 \
    LC_ALL=C.UTF-8 \
+    LD_PRELOAD=/usr/local/lib/libjemalloc.so.2 \
+    MALLOC_CONF=background_thread:true,dirty_decay_ms:5000,muzzy_decay_ms:5000 \
    PORT=8084 \
    GUNICORN_WORKERS=2 \
    MAX_REQUESTS=4000
@@ -20,6 +20,11 @@ ARG HUB_RELEASE_REPO=fawney19/Aether
 ARG HUB_TAG
 ARG TARGETARCH
 ARG GITHUB_TOKEN
+# GitHub 下载镜像前缀，国内构建时传入可用的镜像加速地址
+# 用法: --build-arg GITHUB_MIRROR=https://ghfast.top
+#   或: --build-arg GITHUB_MIRROR=https://gh-proxy.com
+#   或: --build-arg GITHUB_MIRROR=https://mirror.ghproxy.com
+ARG GITHUB_MIRROR

 # 运行时依赖（使用清华镜像源 + BuildKit 缓存加速）
 RUN --mount=type=cache,target=/var/cache/apt,sharing=locked \
@@ -29,7 +34,12 @@ RUN --mount=type=cache,target=/var/cache/apt,sharing=locked \
    nginx \
    supervisor \
    libpq5 \
-    curl
+    curl \
+    libjemalloc2
+RUN set -eux; \
+    jemalloc_path="$(find /usr/lib -type f -name 'libjemalloc.so.2' | head -n1)"; \
+    [ -n "$jemalloc_path" ]; \
+    ln -sf "$jemalloc_path" /usr/local/lib/libjemalloc.so.2

 # 从 base 镜像复制 Python 包
 COPY --from=builder /usr/local/lib/python3.13/site-packages /usr/local/lib/python3.13/site-packages
@@ -39,9 +49,20 @@ 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/

-# Hub 预编译二进制（构建时从 GitHub Release 下载）
+# Hub 预编译二进制
+# 国内构建: --build-arg GITHUB_MIRROR=https://ghfast.top 即可走镜像下载
 # GITHUB_TOKEN 可选：未认证 API 限流 60 次/小时，认证后 5000 次/小时
 RUN set -eux; \
+    arch="${TARGETARCH:-}"; \
+    if [ -z "$arch" ]; then \
+        arch="$(dpkg --print-architecture)"; \
+    fi; \
+    case "$arch" in \
+        amd64|arm64) ;; \
+        x86_64) arch="amd64" ;; \
+        aarch64) arch="arm64" ;; \
+        *) echo "Unsupported architecture: $arch"; exit 1 ;; \
+    esac; \
    auth_header=""; \
    if [ -n "${GITHUB_TOKEN:-}" ]; then \
        auth_header="Authorization: token ${GITHUB_TOKEN}"; \
@@ -54,18 +75,14 @@ RUN set -eux; \
        echo "Failed to resolve hub release tag"; \
        exit 1; \
    fi; \
-    arch="${TARGETARCH:-}"; \
-    if [ -z "$arch" ]; then \
-        arch="$(dpkg --print-architecture)"; \
-    fi; \
-    case "$arch" in \
-        amd64|arm64) ;; \
-        x86_64) arch="amd64" ;; \
-        aarch64) arch="arm64" ;; \
-        *) echo "Unsupported architecture: $arch"; exit 1 ;; \
-    esac; \
    echo "Using Hub release tag: $tag"; \
-    url="https://github.com/${HUB_RELEASE_REPO}/releases/download/${tag}/aether-hub-linux-${arch}.tar.gz"; \
+    origin_url="https://github.com/${HUB_RELEASE_REPO}/releases/download/${tag}/aether-hub-linux-${arch}.tar.gz"; \
+    if [ -n "${GITHUB_MIRROR:-}" ]; then \
+        url="${GITHUB_MIRROR}/https://github.com/${HUB_RELEASE_REPO}/releases/download/${tag}/aether-hub-linux-${arch}.tar.gz"; \
+        echo "Using mirror: ${GITHUB_MIRROR}"; \
+    else \
+        url="$origin_url"; \
+    fi; \
    curl -L --fail -o /tmp/aether-hub.tar.gz "$url"; \
    tar xzf /tmp/aether-hub.tar.gz -C /usr/local/bin; \
    chmod +x /usr/local/bin/aether-hub; \
@@ -95,6 +112,11 @@ RUN printf '%s\n' \
    '    ""      $remote_addr;' \
    '}' \
    '' \
+    'map $http_upgrade $connection_upgrade {' \
+    '    default upgrade;' \
+    '    ""      "";' \
+    '}' \
+    '' \
    'server {' \
    '    listen 80;' \
    '    server_name _;' \
@@ -134,6 +156,25 @@ RUN printf '%s\n' \
    '        proxy_set_header X-Forwarded-Proto $scheme;' \
    '        proxy_set_header Upgrade $http_upgrade;' \
    '        proxy_set_header Connection "upgrade";' \
+    '        # 剥离 CF 头，防止泄露给上游或返回给客户端' \
+    '        proxy_hide_header CF-Connecting-IP;' \
+    '        proxy_hide_header CF-IPCountry;' \
+    '        proxy_hide_header CF-Ray;' \
+    '        proxy_hide_header CF-Visitor;' \
+    '        proxy_hide_header CDN-Loop;' \
+    '        proxy_hide_header True-Client-IP;' \
+    '        proxy_hide_header CF-Worker;' \
+    '        proxy_hide_header CF-EW-Via;' \
+    '        proxy_hide_header CF-Warp-Tag-ID;' \
+    '        proxy_set_header CF-Connecting-IP "";' \
+    '        proxy_set_header CF-IPCountry "";' \
+    '        proxy_set_header CF-Ray "";' \
+    '        proxy_set_header CF-Visitor "";' \
+    '        proxy_set_header CDN-Loop "";' \
+    '        proxy_set_header True-Client-IP "";' \
+    '        proxy_set_header CF-Worker "";' \
+    '        proxy_set_header CF-EW-Via "";' \
+    '        proxy_set_header CF-Warp-Tag-ID "";' \
    '        proxy_read_timeout 86400s;' \
    '        proxy_send_timeout 86400s;' \
    '    }' \
@@ -146,12 +187,22 @@ RUN printf '%s\n' \
    '        proxy_set_header X-Real-IP $real_ip;' \
    '        proxy_set_header X-Forwarded-For $forwarded_for;' \
    '        proxy_set_header X-Forwarded-Proto $scheme;' \
-    '        proxy_set_header Connection "";' \
+    '        proxy_set_header Upgrade $http_upgrade;' \
+    '        proxy_set_header Connection $connection_upgrade;' \
    '        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;' \
-    '        # 剥离 CF 头，防止泄露给上游 AI 提供商' \
+    '        # 剥离 CF 头，防止泄露给上游或返回给客户端' \
+    '        proxy_hide_header CF-Connecting-IP;' \
+    '        proxy_hide_header CF-IPCountry;' \
+    '        proxy_hide_header CF-Ray;' \
+    '        proxy_hide_header CF-Visitor;' \
+    '        proxy_hide_header CDN-Loop;' \
+    '        proxy_hide_header True-Client-IP;' \
+    '        proxy_hide_header CF-Worker;' \
+    '        proxy_hide_header CF-EW-Via;' \
+    '        proxy_hide_header CF-Warp-Tag-ID;' \
    '        proxy_set_header CF-Connecting-IP "";' \
    '        proxy_set_header CF-IPCountry "";' \
    '        proxy_set_header CF-Ray "";' \
@@ -160,6 +211,7 @@ RUN printf '%s\n' \
    '        proxy_set_header True-Client-IP "";' \
    '        proxy_set_header CF-Worker "";' \
    '        proxy_set_header CF-EW-Via "";' \
+    '        proxy_set_header CF-Warp-Tag-ID "";' \
    '        proxy_buffering off;' \
    '        proxy_cache off;' \
    '        proxy_request_buffering off;' \
@@ -179,6 +231,25 @@ RUN printf '%s\n' \
    '        proxy_set_header X-Real-IP $real_ip;' \
    '        proxy_set_header X-Forwarded-For $forwarded_for;' \
    '        proxy_set_header X-Forwarded-Proto $scheme;' \
+    '        # 剥离 CF 头，防止泄露给上游或返回给客户端' \
+    '        proxy_hide_header CF-Connecting-IP;' \
+    '        proxy_hide_header CF-IPCountry;' \
+    '        proxy_hide_header CF-Ray;' \
+    '        proxy_hide_header CF-Visitor;' \
+    '        proxy_hide_header CDN-Loop;' \
+    '        proxy_hide_header True-Client-IP;' \
+    '        proxy_hide_header CF-Worker;' \
+    '        proxy_hide_header CF-EW-Via;' \
+    '        proxy_hide_header CF-Warp-Tag-ID;' \
+    '        proxy_set_header CF-Connecting-IP "";' \
+    '        proxy_set_header CF-IPCountry "";' \
+    '        proxy_set_header CF-Ray "";' \
+    '        proxy_set_header CF-Visitor "";' \
+    '        proxy_set_header CDN-Loop "";' \
+    '        proxy_set_header True-Client-IP "";' \
+    '        proxy_set_header CF-Worker "";' \
+    '        proxy_set_header CF-EW-Via "";' \
+    '        proxy_set_header CF-Warp-Tag-ID "";' \
    '    }' \
    '' \
    '    # 所有其他路由 → 前端 SPA（先尝试静态文件，再回退到 index.html）' \
@@ -210,7 +281,7 @@ RUN printf '%s\n' \
    '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' \
+    'environment=PYTHONUNBUFFERED=1,PYTHONIOENCODING=utf-8,LANG=C.UTF-8,LC_ALL=C.UTF-8,DOCKER_CONTAINER=true,LD_PRELOAD=/usr/local/lib/libjemalloc.so.2,MALLOC_CONF="background_thread:true,dirty_decay_ms:5000,muzzy_decay_ms:5000"' \
    '' \
    '[program:tunnel-hub]' \
    'command=/usr/local/bin/aether-hub --bind 0.0.0.0:8085' \
@@ -234,6 +305,8 @@ ENV PYTHONUNBUFFERED=1 \
    PYTHONIOENCODING=utf-8 \
    LANG=C.UTF-8 \
    LC_ALL=C.UTF-8 \
+    LD_PRELOAD=/usr/local/lib/libjemalloc.so.2 \
+    MALLOC_CONF=background_thread:true,dirty_decay_ms:5000,muzzy_decay_ms:5000 \
    PORT=8084 \
    GUNICORN_WORKERS=2 \
    MAX_REQUESTS=4000
@@ -34,7 +34,7 @@ Aether 是一个自托管的 AI API 网关，为团队和个人提供多租户

 ## 部署

-### Docker Compose（推荐：本地构建镜像）
+### Docker Compose（推荐：预构建镜像）

 ```bash
 # 1. 克隆代码
@@ -64,6 +64,7 @@ cp .env.example .env
 python generate_keys.py  # 生成密钥, 并将生成的密钥填入 .env

 # 3. 部署 / 更新（自动构建、启动、迁移）
+git pull
 ./deploy.sh
 ```

@@ -1,6 +1,6 @@
 [package]
 name = "aether-hub"
-version = "0.1.4"
+version = "0.2.0"
 edition = "2021"
 description = "Tunnel Hub for Aether - frame router between workers and proxies"

@@ -16,6 +16,10 @@ dashmap = "6"
 parking_lot = "0.12"
 flate2 = "1"
 futures-util = "0.3"
+bytes = "1"
+async-stream = "0.3"
+http-body-util = "0.1"
+reqwest = { version = "0.12", default-features = false, features = ["json", "rustls-tls"] }

 [profile.release]
 lto = true
@@ -28,6 +28,11 @@ cd /path/to/Aether
 - `--load`: 加载到本地 Docker（单平台）
 - `--latest`: 额外打 `latest` tag

+## 运行时参数
+
+- `TUNNEL_HUB_WORKER_IDLE_TIMEOUT`：worker 心跳空闲超时，默认 `60` 秒
+- `TUNNEL_HUB_OUTBOUND_QUEUE_CAPACITY`：单连接出站队列容量，默认 `128`；队列打满时会把连接视为拥塞并主动关闭，避免 Hub 内存无限增长
+
 ## 与部署脚本关系

 - `./deploy.sh`: 本地构建部署（会本地构建 app/base，并在构建 app 时从 GitHub Release 下载 Hub，可用 `--hub-tag` 固定版本）。
@@ -0,0 +1,85 @@
+use reqwest::Client;
+
+#[derive(Clone)]
+pub struct ControlPlaneClient {
+    client: Option<Client>,
+    base_url: String,
+}
+
+impl ControlPlaneClient {
+    pub fn new(base_url: String) -> Self {
+        let client = Client::builder()
+            .timeout(std::time::Duration::from_secs(10))
+            .build()
+            .ok();
+        Self { client, base_url }
+    }
+
+    pub fn disabled() -> Self {
+        Self {
+            client: None,
+            base_url: String::new(),
+        }
+    }
+
+    pub async fn heartbeat_ack(&self, payload: &[u8]) -> Result<Vec<u8>, String> {
+        let Some(client) = &self.client else {
+            return Ok(b"{}".to_vec());
+        };
+        let url = format!(
+            "{}/api/internal/hub/heartbeat",
+            self.base_url.trim_end_matches('/')
+        );
+        let response = client
+            .post(&url)
+            .header("content-type", "application/json")
+            .body(payload.to_vec())
+            .send()
+            .await
+            .map_err(|e| format!("heartbeat callback request failed: {e}"))?;
+        if !response.status().is_success() {
+            return Err(format!(
+                "heartbeat callback failed with status {}",
+                response.status()
+            ));
+        }
+        response
+            .bytes()
+            .await
+            .map(|bytes| bytes.to_vec())
+            .map_err(|e| format!("heartbeat callback body read failed: {e}"))
+    }
+
+    pub async fn push_node_status(
+        &self,
+        node_id: &str,
+        connected: bool,
+        conn_count: usize,
+    ) -> Result<(), String> {
+        let Some(client) = &self.client else {
+            return Ok(());
+        };
+        let url = format!(
+            "{}/api/internal/hub/node-status",
+            self.base_url.trim_end_matches('/')
+        );
+        let response = client
+            .post(&url)
+            .json(&serde_json::json!({
+                "node_id": node_id,
+                "connected": connected,
+                "conn_count": conn_count,
+            }))
+            .send()
+            .await
+            .map_err(|e| format!("node-status callback request failed: {e}"))?;
+        if response.status().is_success() {
+            Ok(())
+        } else {
+            Err(format!(
+                "node-status callback failed with status {}",
+                response.status()
+            ))
+        }
+    }
+}
@@ -0,0 +1,262 @@
+use std::io;
+use std::net::SocketAddr;
+use std::time::Duration;
+
+use async_stream::stream;
+use axum::body::{Body, Bytes};
+use axum::extract::{ConnectInfo, Path, Request, State};
+use axum::http::{HeaderMap, HeaderName, HeaderValue, Response, StatusCode};
+use axum::response::IntoResponse;
+use bytes::BytesMut;
+use futures_util::StreamExt;
+use tracing::warn;
+
+use crate::hub::{LocalBodyEvent, LocalStream};
+use crate::protocol;
+use crate::AppState;
+
+pub const TUNNEL_ERROR_HEADER: &str = "x-aether-tunnel-error";
+const MAX_RELAY_META_LEN: usize = 256 * 1024;
+
+struct StreamGuard {
+    hub: std::sync::Arc<crate::hub::HubRouter>,
+    stream_id: u64,
+    finished: bool,
+}
+
+impl Drop for StreamGuard {
+    fn drop(&mut self) {
+        if !self.finished {
+            self.hub
+                .cancel_local_stream(self.stream_id, "local relay client dropped");
+        }
+    }
+}
+
+pub async fn relay_request(
+    Path(node_id): Path<String>,
+    State(state): State<AppState>,
+    ConnectInfo(addr): ConnectInfo<SocketAddr>,
+    request: Request,
+) -> impl IntoResponse {
+    if !addr.ip().is_loopback() {
+        return tunnel_error_response(
+            StatusCode::FORBIDDEN,
+            "forbidden",
+            "local relay only accepts loopback requests",
+        );
+    }
+
+    let mut body_stream = request.into_body().into_data_stream();
+    let mut envelope_buf = BytesMut::new();
+    let mut meta: Option<protocol::RequestMeta> = None;
+    let mut stream: Option<std::sync::Arc<LocalStream>> = None;
+
+    while let Some(chunk_result) = body_stream.next().await {
+        let chunk = match chunk_result {
+            Ok(chunk) => chunk,
+            Err(error) => {
+                if let Some(active_stream) = &stream {
+                    state
+                        .hub
+                        .cancel_local_stream(active_stream.id, "failed to read relay request body");
+                }
+                warn!(error = %error, "failed to read local relay request body");
+                return tunnel_error_response(
+                    StatusCode::BAD_GATEWAY,
+                    "relay",
+                    "failed to read relay request body",
+                );
+            }
+        };
+
+        if stream.is_none() {
+            envelope_buf.extend_from_slice(&chunk);
+            let Some((parsed_meta, body_offset)) = (match try_decode_envelope_meta(&envelope_buf) {
+                Ok(result) => result,
+                Err(error) => {
+                    return tunnel_error_response(StatusCode::BAD_REQUEST, "bad_request", &error);
+                }
+            }) else {
+                continue;
+            };
+
+            let opened_stream = match state.hub.open_local_stream(&node_id, &parsed_meta) {
+                Ok(stream) => stream,
+                Err(error) => {
+                    return tunnel_error_response(
+                        StatusCode::SERVICE_UNAVAILABLE,
+                        "connect",
+                        &error,
+                    );
+                }
+            };
+
+            if envelope_buf.len() > body_offset {
+                let first_body_chunk = Bytes::copy_from_slice(&envelope_buf[body_offset..]);
+                if let Err(error) =
+                    state
+                        .hub
+                        .push_local_request_body(opened_stream.id, first_body_chunk, false)
+                {
+                    state.hub.cancel_local_stream(opened_stream.id, &error);
+                    return tunnel_error_response(
+                        StatusCode::SERVICE_UNAVAILABLE,
+                        "connect",
+                        &error,
+                    );
+                }
+            }
+
+            envelope_buf.clear();
+            meta = Some(parsed_meta);
+            stream = Some(opened_stream);
+            continue;
+        }
+
+        let Some(active_stream) = &stream else {
+            continue;
+        };
+        if let Err(error) = state
+            .hub
+            .push_local_request_body(active_stream.id, chunk, false)
+        {
+            state.hub.cancel_local_stream(active_stream.id, &error);
+            return tunnel_error_response(StatusCode::SERVICE_UNAVAILABLE, "connect", &error);
+        }
+    }
+
+    let (meta, stream) = match (meta, stream) {
+        (Some(meta), Some(stream)) => (meta, stream),
+        _ => {
+            return tunnel_error_response(
+                StatusCode::BAD_REQUEST,
+                "bad_request",
+                "relay envelope metadata truncated",
+            );
+        }
+    };
+
+    if let Err(error) = state
+        .hub
+        .push_local_request_body(stream.id, Bytes::new(), true)
+    {
+        state.hub.cancel_local_stream(stream.id, &error);
+        return tunnel_error_response(StatusCode::SERVICE_UNAVAILABLE, "connect", &error);
+    }
+
+    let request_guard = StreamGuard {
+        hub: state.hub.clone(),
+        stream_id: stream.id,
+        finished: false,
+    };
+
+    let wait_timeout = Duration::from_secs(meta.timeout.clamp(5, 300));
+    let response_head = match stream.wait_headers(wait_timeout).await {
+        Ok(response) => response,
+        Err(error) => {
+            state.hub.cancel_local_stream(stream.id, &error);
+            return tunnel_error_response(StatusCode::GATEWAY_TIMEOUT, "timeout", &error);
+        }
+    };
+
+    let Some(mut body_rx) = stream.take_body_receiver() else {
+        state
+            .hub
+            .cancel_local_stream(stream.id, "missing relay response body receiver");
+        return tunnel_error_response(
+            StatusCode::BAD_GATEWAY,
+            "relay",
+            "missing relay response body receiver",
+        );
+    };
+
+    let hub = state.hub.clone();
+    let stream_id = stream.id;
+    let body_stream = stream! {
+        let mut guard = request_guard;
+        guard.hub = hub;
+        guard.stream_id = stream_id;
+        while let Some(event) = body_rx.recv().await {
+            match event {
+                LocalBodyEvent::Chunk(chunk) => yield Ok::<Bytes, io::Error>(chunk),
+                LocalBodyEvent::End => {
+                    guard.finished = true;
+                    break;
+                }
+                LocalBodyEvent::Error(error) => {
+                    guard.finished = true;
+                    yield Err(io::Error::other(error));
+                    break;
+                }
+            }
+        }
+        guard.finished = true;
+    };
+
+    let mut builder = Response::builder().status(response_head.status);
+    if let Some(headers) = builder.headers_mut() {
+        append_headers(headers, &response_head.headers);
+    }
+    match builder.body(Body::from_stream(body_stream)) {
+        Ok(response) => response,
+        Err(error) => {
+            warn!(error = %error, "failed to build relay response");
+            tunnel_error_response(
+                StatusCode::BAD_GATEWAY,
+                "relay",
+                "failed to build relay response",
+            )
+        }
+    }
+}
+
+fn try_decode_envelope_meta(
+    buffer: &BytesMut,
+) -> Result<Option<(protocol::RequestMeta, usize)>, String> {
+    if buffer.len() < 4 {
+        return Ok(None);
+    }
+    let meta_len = u32::from_be_bytes([buffer[0], buffer[1], buffer[2], buffer[3]]) as usize;
+    if meta_len > MAX_RELAY_META_LEN {
+        return Err("relay metadata too large".to_string());
+    }
+    let meta_end = 4usize
+        .checked_add(meta_len)
+        .ok_or_else(|| "relay envelope length overflow".to_string())?;
+    if buffer.len() < meta_end {
+        return Ok(None);
+    }
+    let meta = serde_json::from_slice::<protocol::RequestMeta>(&buffer[4..meta_end])
+        .map_err(|e| format!("invalid relay metadata: {e}"))?;
+    Ok(Some((meta, meta_end)))
+}
+
+fn append_headers(target: &mut HeaderMap, headers: &[(String, String)]) {
+    for (name, value) in headers {
+        let Ok(name) = HeaderName::from_bytes(name.as_bytes()) else {
+            continue;
+        };
+        let Ok(value) = HeaderValue::from_str(value) else {
+            continue;
+        };
+        target.append(name, value);
+    }
+}
+
+fn tunnel_error_response(status: StatusCode, kind: &str, message: &str) -> Response<Body> {
+    let mut builder = Response::builder().status(status);
+    if let Some(headers) = builder.headers_mut() {
+        headers.insert(
+            HeaderName::from_static(TUNNEL_ERROR_HEADER),
+            HeaderValue::from_str(kind).unwrap_or_else(|_| HeaderValue::from_static("relay")),
+        );
+        headers.insert(
+            axum::http::header::CONTENT_TYPE,
+            HeaderValue::from_static("text/plain; charset=utf-8"),
+        );
+    }
+    builder
+        .body(Body::from(message.to_string()))
+        .unwrap_or_else(|_| Response::new(Body::from("relay error")))
+}
@@ -1,20 +1,23 @@
+mod control_plane;
 mod hub;
+mod local_relay;
 mod protocol;
 mod proxy_conn;
-mod worker_conn;

-use std::sync::Arc;
+use std::net::SocketAddr;
 use std::time::Duration;

 use axum::extract::ws::WebSocketUpgrade;
 use axum::extract::State;
 use axum::response::{IntoResponse, Json};
-use axum::routing::get;
+use axum::routing::{get, post};
 use axum::Router;
 use clap::Parser;
 use tracing::{info, warn};

-use crate::hub::HubRouter;
+use crate::control_plane::ControlPlaneClient;
+use crate::hub::{ConnConfig, HubRouter};
+use crate::local_relay::relay_request;

 #[derive(Parser, Debug)]
 #[command(name = "aether-hub", about = "Tunnel Hub for Aether")]
@@ -23,14 +26,10 @@ struct Args {
    #[arg(long, default_value = "0.0.0.0:8085", env = "TUNNEL_HUB_BIND")]
    bind: String,

-    /// Proxy-side idle timeout in seconds
-    #[arg(long, default_value_t = 90, env = "TUNNEL_HUB_PROXY_IDLE_TIMEOUT")]
+    /// Proxy-side idle timeout in seconds (0 to disable)
+    #[arg(long, default_value_t = 0, env = "TUNNEL_HUB_PROXY_IDLE_TIMEOUT")]
    proxy_idle_timeout: u64,

-    /// Worker-side idle timeout in seconds
-    #[arg(long, default_value_t = 60, env = "TUNNEL_HUB_WORKER_IDLE_TIMEOUT")]
-    worker_idle_timeout: u64,
-
    /// Ping interval in seconds (for both sides)
    #[arg(long, default_value_t = 15, env = "TUNNEL_HUB_PING_INTERVAL")]
    ping_interval: u64,
@@ -38,15 +37,29 @@ struct Args {
    /// Max concurrent streams per proxy connection
    #[arg(long, default_value_t = 2048, env = "TUNNEL_HUB_MAX_STREAMS")]
    max_streams: usize,
+
+    /// Per-connection outbound queue capacity before treating the socket as congested
+    #[arg(
+        long,
+        default_value_t = 128,
+        env = "TUNNEL_HUB_OUTBOUND_QUEUE_CAPACITY"
+    )]
+    outbound_queue_capacity: usize,
+
+    /// Local Aether app base URL for control-plane callbacks
+    #[arg(
+        long,
+        default_value = "http://127.0.0.1:8084",
+        env = "TUNNEL_HUB_APP_BASE_URL"
+    )]
+    app_base_url: String,
 }

 #[derive(Clone)]
-struct AppState {
-    hub: Arc<HubRouter>,
-    proxy_idle_timeout: Duration,
-    worker_idle_timeout: Duration,
-    ping_interval: Duration,
-    max_streams: usize,
+pub struct AppState {
+    pub hub: std::sync::Arc<HubRouter>,
+    pub proxy_conn_cfg: ConnConfig,
+    pub max_streams: usize,
 }

 #[tokio::main]
@@ -61,12 +74,16 @@ async fn main() -> Result<(), Box<dyn std::error::Error>> {

    let args = Args::parse();

-    let hub = HubRouter::new();
+    let hub = HubRouter::new(ControlPlaneClient::new(args.app_base_url));
+    let outbound_queue_capacity = args.outbound_queue_capacity.clamp(8, 4096);
+    let ping_interval = Duration::from_secs(args.ping_interval);
    let state = AppState {
        hub,
-        proxy_idle_timeout: Duration::from_secs(args.proxy_idle_timeout),
-        worker_idle_timeout: Duration::from_secs(args.worker_idle_timeout),
-        ping_interval: Duration::from_secs(args.ping_interval),
+        proxy_conn_cfg: ConnConfig {
+            ping_interval,
+            idle_timeout: Duration::from_secs(args.proxy_idle_timeout),
+            outbound_queue_capacity,
+        },
        max_streams: args.max_streams,
    };

@@ -74,13 +91,17 @@ async fn main() -> Result<(), Box<dyn std::error::Error>> {
        .route("/health", get(health))
        .route("/stats", get(stats))
        .route("/proxy", get(ws_proxy))
-        .route("/worker", get(ws_worker))
+        .route("/local/relay/{node_id}", post(relay_request))
        .with_state(state);

    let listener = tokio::net::TcpListener::bind(&args.bind).await?;
    info!(bind = %args.bind, "aether-hub started");

-    axum::serve(listener, app).await?;
+    axum::serve(
+        listener,
+        app.into_make_service_with_connect_info::<SocketAddr>(),
+    )
+    .await?;
    Ok(())
 }

@@ -139,21 +160,8 @@ async fn ws_proxy(
                node_id,
                node_name,
                max_streams,
-                state.ping_interval,
-                state.proxy_idle_timeout,
+                state.proxy_conn_cfg,
            )
        })
        .into_response()
 }
-
-async fn ws_worker(ws: WebSocketUpgrade, State(state): State<AppState>) -> impl IntoResponse {
-    ws.max_frame_size(64 * 1024 * 1024)
-        .on_upgrade(move |socket| {
-            worker_conn::handle_worker_connection(
-                socket,
-                state.hub,
-                state.ping_interval,
-                state.worker_idle_timeout,
-            )
-        })
-}
@@ -22,8 +22,6 @@ pub const PONG: u8 = 0x11;
 pub const GOAWAY: u8 = 0x12;
 pub const HEARTBEAT_DATA: u8 = 0x13;
 pub const HEARTBEAT_ACK: u8 = 0x14;
-pub const NODE_STATUS: u8 = 0x15;
-
 // Flags
 pub const FLAG_END_STREAM: u8 = 0x01;
 pub const FLAG_GZIP_COMPRESSED: u8 = 0x02;
@@ -50,162 +48,89 @@ impl FrameHeader {
            payload_len: u32::from_be_bytes([data[6], data[7], data[8], data[9]]),
        })
    }
+}

-    /// Check if this is a stream-terminating frame
-    #[inline]
-    pub fn is_stream_terminal(&self) -> bool {
-        self.msg_type == STREAM_END || self.msg_type == STREAM_ERROR
+#[derive(Debug, Clone, serde::Serialize, serde::Deserialize)]
+pub struct RequestMeta {
+    pub method: String,
+    pub url: String,
+    pub headers: std::collections::HashMap<String, String>,
+    #[serde(default = "default_timeout", deserialize_with = "deserialize_timeout")]
+    pub timeout: u64,
+}
+
+fn default_timeout() -> u64 {
+    60
+}
+
+fn deserialize_timeout<'de, D>(deserializer: D) -> Result<u64, D::Error>
+where
+    D: serde::Deserializer<'de>,
+{
+    #[derive(serde::Deserialize)]
+    #[serde(untagged)]
+    enum TimeoutValue {
+        Int(u64),
+        Float(f64),
+    }
+
+    match <TimeoutValue as serde::Deserialize>::deserialize(deserializer)? {
+        TimeoutValue::Int(v) => Ok(v),
+        TimeoutValue::Float(v) => {
+            if !v.is_finite() || v < 0.0 {
+                return Err(serde::de::Error::custom(
+                    "timeout must be a non-negative finite number",
+                ));
+            }
+            if v.fract() != 0.0 {
+                return Err(serde::de::Error::custom("timeout must be integer seconds"));
+            }
+            if v > (u64::MAX as f64) {
+                return Err(serde::de::Error::custom("timeout is too large"));
+            }
+            Ok(v as u64)
+        }
    }
 }

-#[derive(Debug)]
-pub struct RequestHeadersExtracted {
-    pub node_id: String,
-    pub rebuilt_frame: Vec<u8>,
+#[derive(Debug, Clone, serde::Serialize, serde::Deserialize)]
+pub struct ResponseMeta {
+    pub status: u16,
+    pub headers: Vec<(String, String)>,
+}
+
+pub fn encode_frame(stream_id: u32, msg_type: u8, flags: u8, payload: &[u8]) -> Vec<u8> {
+    let mut buf = Vec::with_capacity(HEADER_SIZE + payload.len());
+    buf.extend_from_slice(&stream_id.to_be_bytes());
+    buf.push(msg_type);
+    buf.push(flags);
+    buf.extend_from_slice(&(payload.len() as u32).to_be_bytes());
+    buf.extend_from_slice(payload);
+    buf
 }

 /// Encode a STREAM_ERROR frame for a given stream_id with an error message
 pub fn encode_stream_error(stream_id: u32, msg: &str) -> Vec<u8> {
-    let payload = msg.as_bytes();
-    let mut buf = Vec::with_capacity(HEADER_SIZE + payload.len());
-    buf.extend_from_slice(&stream_id.to_be_bytes());
-    buf.push(STREAM_ERROR);
-    buf.push(0); // flags
-    buf.extend_from_slice(&(payload.len() as u32).to_be_bytes());
-    buf.extend_from_slice(payload);
-    buf
-}
-
-/// Encode a NODE_STATUS frame (stream_id=0, Hub-generated)
-pub fn encode_node_status(node_id: &str, connected: bool, conn_count: usize) -> Vec<u8> {
-    let payload = serde_json::json!({
-        "node_id": node_id,
-        "connected": connected,
-        "conn_count": conn_count,
-    });
-    let payload_bytes = payload.to_string().into_bytes();
-    let mut buf = Vec::with_capacity(HEADER_SIZE + payload_bytes.len());
-    buf.extend_from_slice(&0u32.to_be_bytes()); // stream_id = 0
-    buf.push(NODE_STATUS);
-    buf.push(0); // flags
-    buf.extend_from_slice(&(payload_bytes.len() as u32).to_be_bytes());
-    buf.extend_from_slice(&payload_bytes);
-    buf
+    encode_frame(stream_id, STREAM_ERROR, 0, msg.as_bytes())
 }

 /// Encode a PING frame (stream_id=0)
 pub fn encode_ping() -> Vec<u8> {
-    let mut buf = Vec::with_capacity(HEADER_SIZE);
-    buf.extend_from_slice(&0u32.to_be_bytes());
-    buf.push(PING);
-    buf.push(0);
-    buf.extend_from_slice(&0u32.to_be_bytes());
-    buf
+    encode_frame(0, PING, 0, &[])
 }

 /// Encode a PONG frame (stream_id=0, echo payload)
 pub fn encode_pong(payload: &[u8]) -> Vec<u8> {
-    let mut buf = Vec::with_capacity(HEADER_SIZE + payload.len());
-    buf.extend_from_slice(&0u32.to_be_bytes());
-    buf.push(PONG);
-    buf.push(0);
-    buf.extend_from_slice(&(payload.len() as u32).to_be_bytes());
-    buf.extend_from_slice(payload);
-    buf
+    encode_frame(0, PONG, 0, payload)
 }

 /// Encode a GOAWAY frame (stream_id=0)
 pub fn encode_goaway() -> Vec<u8> {
-    let mut buf = Vec::with_capacity(HEADER_SIZE);
-    buf.extend_from_slice(&0u32.to_be_bytes());
-    buf.push(GOAWAY);
-    buf.push(0);
-    buf.extend_from_slice(&0u32.to_be_bytes());
-    buf
-}
-
-/// Rewrite the stream_id in raw frame bytes (first 4 bytes) -- near zero-copy
-#[inline]
-pub fn rewrite_stream_id(data: &mut [u8], new_stream_id: u32) {
-    let bytes = new_stream_id.to_be_bytes();
-    data[0] = bytes[0];
-    data[1] = bytes[1];
-    data[2] = bytes[2];
-    data[3] = bytes[3];
-}
-
-/// Get the payload portion of a raw frame (after the 10-byte header)
-#[inline]
-pub fn frame_payload(data: &[u8]) -> &[u8] {
-    if data.len() > HEADER_SIZE {
-        &data[HEADER_SIZE..]
-    } else {
-        &[]
-    }
-}
-
-/// Parse REQUEST_HEADERS payload, extract `node_id`, strip it from JSON,
-/// and rebuild a new REQUEST_HEADERS frame with `new_stream_id`.
-///
-/// If the source frame is gzip-compressed, this function will decode it first,
-/// then try to re-encode with gzip (only keeps compression when payload shrinks).
-pub fn rebuild_request_headers_without_node_id(
-    data: &[u8],
-    new_stream_id: u32,
-) -> Result<RequestHeadersExtracted, String> {
-    let header = FrameHeader::parse(data).ok_or_else(|| "invalid frame header".to_string())?;
-    if header.msg_type != REQUEST_HEADERS {
-        return Err("frame is not REQUEST_HEADERS".to_string());
-    }
-
-    let payload = frame_payload_by_header(data, &header)
-        .ok_or_else(|| "incomplete REQUEST_HEADERS payload".to_string())?;
-
-    let decoded_payload = if header.flags & FLAG_GZIP_COMPRESSED != 0 {
-        let mut decoder = GzDecoder::new(payload);
-        let mut decoded = Vec::new();
-        decoder
-            .read_to_end(&mut decoded)
-            .map_err(|e| format!("failed to decompress REQUEST_HEADERS: {e}"))?;
-        decoded
-    } else {
-        payload.to_vec()
-    };
-
-    let mut meta: serde_json::Value = serde_json::from_slice(&decoded_payload)
-        .map_err(|e| format!("invalid REQUEST_HEADERS JSON: {e}"))?;
-    let obj = meta
-        .as_object_mut()
-        .ok_or_else(|| "REQUEST_HEADERS payload must be a JSON object".to_string())?;
-
-    let node_id = obj
-        .remove("node_id")
-        .and_then(|v| v.as_str().map(|s| s.to_string()))
-        .map(|s| s.trim().to_string())
-        .filter(|s| !s.is_empty())
-        .ok_or_else(|| "missing node_id in REQUEST_HEADERS".to_string())?;
-
-    let stripped_payload = serde_json::to_vec(&meta)
-        .map_err(|e| format!("failed to encode REQUEST_HEADERS payload: {e}"))?;
-    let (final_payload, flags) =
-        maybe_recompress_payload(&stripped_payload, header.flags & FLAG_GZIP_COMPRESSED != 0)
-            .map_err(|e| format!("failed to recompress REQUEST_HEADERS payload: {e}"))?;
-
-    let mut rebuilt = Vec::with_capacity(HEADER_SIZE + final_payload.len());
-    rebuilt.extend_from_slice(&new_stream_id.to_be_bytes());
-    rebuilt.push(REQUEST_HEADERS);
-    rebuilt.push(flags);
-    rebuilt.extend_from_slice(&(final_payload.len() as u32).to_be_bytes());
-    rebuilt.extend_from_slice(&final_payload);
-
-    Ok(RequestHeadersExtracted {
-        node_id,
-        rebuilt_frame: rebuilt,
-    })
+    encode_frame(0, GOAWAY, 0, &[])
 }

 #[inline]
-fn frame_payload_by_header<'a>(data: &'a [u8], header: &FrameHeader) -> Option<&'a [u8]> {
+pub fn frame_payload_by_header<'a>(data: &'a [u8], header: &FrameHeader) -> Option<&'a [u8]> {
    let payload_len = header.payload_len as usize;
    let end = HEADER_SIZE.checked_add(payload_len)?;
    if data.len() < end {
@@ -214,6 +139,25 @@ fn frame_payload_by_header<'a>(data: &'a [u8], header: &FrameHeader) -> Option<&
    Some(&data[HEADER_SIZE..end])
 }

+pub fn decode_payload(data: &[u8], header: &FrameHeader) -> Result<Vec<u8>, String> {
+    let payload = frame_payload_by_header(data, header)
+        .ok_or_else(|| "incomplete frame payload".to_string())?;
+    if header.flags & FLAG_GZIP_COMPRESSED != 0 {
+        let mut decoder = GzDecoder::new(payload);
+        let mut decoded = Vec::new();
+        decoder
+            .read_to_end(&mut decoded)
+            .map_err(|e| format!("failed to decompress payload: {e}"))?;
+        Ok(decoded)
+    } else {
+        Ok(payload.to_vec())
+    }
+}
+
+pub fn compress_payload(payload: &[u8]) -> Result<(Vec<u8>, u8), std::io::Error> {
+    maybe_recompress_payload(payload, true)
+}
+
 fn maybe_recompress_payload(
    payload: &[u8],
    prefer_gzip: bool,
@@ -7,10 +7,10 @@ use std::time::Duration;

 use axum::extract::ws::{Message, WebSocket};
 use futures_util::{SinkExt, StreamExt};
-use tokio::sync::mpsc;
+use tokio::sync::{mpsc, watch};
 use tracing::{debug, info, warn};

-use crate::hub::{HubRouter, ProxyConn};
+use crate::hub::{ConnConfig, HubRouter, ProxyConn, SendStatus};
 use crate::protocol;

 /// Maximum single frame size: 64 MB
@@ -22,69 +22,71 @@ pub async fn handle_proxy_connection(
    node_id: String,
    node_name: String,
    max_streams: usize,
-    ping_interval: Duration,
-    idle_timeout: Duration,
+    cfg: ConnConfig,
 ) {
    let conn_id = hub.alloc_conn_id();
    let (mut ws_tx, ws_rx) = ws.split();

-    // Create channel for outbound messages
-    let (tx, mut rx) = mpsc::unbounded_channel::<Message>();
+    let (tx, mut rx) = mpsc::channel::<Message>(cfg.outbound_queue_capacity);
+    let (close_tx, mut close_rx) = watch::channel(false);

    let conn = Arc::new(ProxyConn::new(
        conn_id,
        node_id.clone(),
        node_name.clone(),
        tx,
+        close_tx,
        max_streams,
    ));

    hub.register_proxy(conn.clone());

-    // Spawn writer task: drains channel -> WebSocket
    let writer = tokio::spawn(async move {
-        while let Some(msg) = rx.recv().await {
-            if ws_tx.send(msg).await.is_err() {
-                break;
+        loop {
+            tokio::select! {
+                msg = rx.recv() => match msg {
+                    Some(msg) => {
+                        if ws_tx.send(msg).await.is_err() {
+                            break;
+                        }
+                    }
+                    None => break,
+                },
+                changed = close_rx.changed() => {
+                    if changed.is_err() || *close_rx.borrow() {
+                        break;
+                    }
+                }
            }
        }
        let _ = ws_tx.close().await;
    });

-    // Spawn ping task
-    let ping_tx = conn.tx.clone();
+    let ping_conn = conn.clone();
+    let ping_interval = cfg.ping_interval;
    let ping_task = tokio::spawn(async move {
        loop {
            tokio::time::sleep(ping_interval).await;
            let ping = protocol::encode_ping();
-            if ping_tx.send(Message::Binary(ping.into())).is_err() {
+            if !matches!(
+                ping_conn.send(Message::Binary(ping.into())),
+                SendStatus::Queued
+            ) {
                break;
            }
        }
    });

-    // Spawn reader task
    let reader_hub = hub.clone();
-    let reader_node_id = node_id.clone();
-    let reader_tx = conn.tx.clone();
+    let reader_conn = conn.clone();
    let reader = tokio::spawn(async move {
-        run_proxy_reader(
-            ws_rx,
-            reader_hub,
-            conn_id,
-            reader_node_id,
-            reader_tx,
-            idle_timeout,
-        )
-        .await;
+        run_proxy_reader(ws_rx, reader_hub, reader_conn, cfg.idle_timeout).await;
    });

-    // Wait for reader to end, then cleanup.
    let _ = reader.await;
    ping_task.abort();
+    conn.request_close();
    hub.unregister_proxy(conn_id, &node_id);
-    // conn still holds an Arc<ProxyConn> with a channel sender clone.
-    // Drop it so the writer can drain and exit.
    drop(conn);
    tokio::time::sleep(Duration::from_millis(100)).await;
    writer.abort();
@@ -94,20 +96,24 @@ pub async fn handle_proxy_connection(
 async fn run_proxy_reader(
    mut ws_rx: futures_util::stream::SplitStream<WebSocket>,
    hub: Arc<HubRouter>,
-    conn_id: u64,
-    node_id: String,
-    tx: mpsc::UnboundedSender<Message>,
+    conn: Arc<ProxyConn>,
    idle_timeout: Duration,
 ) {
+    let idle_enabled = !idle_timeout.is_zero();
    let mut oversized_count = 0u32;
    loop {
-        let msg = tokio::select! {
-            msg = ws_rx.next() => msg,
-            _ = tokio::time::sleep(idle_timeout) => {
-                warn!(conn_id = conn_id, node_id = %node_id, "proxy idle timeout");
-                let _ = tx.send(Message::Binary(protocol::encode_goaway().into()));
-                break;
+        let msg = if idle_enabled {
+            tokio::select! {
+                msg = ws_rx.next() => msg,
+                _ = tokio::time::sleep(idle_timeout) => {
+                    warn!(conn_id = conn.id, node_id = %conn.node_id, "proxy idle timeout");
+                    let _ = conn.send(Message::Binary(protocol::encode_goaway().into()));
+                    conn.request_close();
+                    break;
+                }
            }
+        } else {
+            ws_rx.next().await
        };

        match msg {
@@ -116,12 +122,13 @@ async fn run_proxy_reader(
                if data.len() > MAX_FRAME_SIZE {
                    oversized_count += 1;
                    warn!(
-                        conn_id = conn_id,
+                        conn_id = conn.id,
                        size = data.len(),
                        "oversized frame from proxy"
                    );
                    if oversized_count >= 5 {
-                        warn!(conn_id = conn_id, "too many oversized frames, closing");
+                        warn!(conn_id = conn.id, "too many oversized frames, closing");
+                        conn.request_close();
                        break;
                    }
                    continue;
@@ -129,21 +136,21 @@ async fn run_proxy_reader(
                oversized_count = 0;

                if data.len() < protocol::HEADER_SIZE {
-                    debug!(conn_id = conn_id, "frame too small, skipping");
+                    debug!(conn_id = conn.id, "frame too small, skipping");
                    continue;
                }

-                hub.handle_proxy_frame(conn_id, &mut data);
+                hub.handle_proxy_frame(conn.id, &mut data).await;
            }
            Some(Ok(Message::Close(_))) | None => {
-                info!(conn_id = conn_id, node_id = %node_id, "proxy WebSocket closed");
+                info!(conn_id = conn.id, node_id = %conn.node_id, "proxy WebSocket closed");
                break;
            }
            Some(Err(e)) => {
-                warn!(conn_id = conn_id, error = %e, "proxy WebSocket error");
+                warn!(conn_id = conn.id, error = %e, "proxy WebSocket error");
                break;
            }
-            _ => {} // Ignore text/ping/pong at WS level
+            _ => {}
        }
    }
 }
@@ -1,136 +0,0 @@
-/// Worker-side WebSocket connection handler
-///
-/// Handles the lifecycle of a single Gunicorn worker connection:
-/// accept -> read loop (route frames via Hub) -> cleanup
-use std::sync::Arc;
-use std::time::Duration;
-
-use axum::extract::ws::{Message, WebSocket};
-use futures_util::{SinkExt, StreamExt};
-use tokio::sync::mpsc;
-use tracing::{debug, info, warn};
-
-use crate::hub::{HubRouter, WorkerConn};
-use crate::protocol;
-
-pub async fn handle_worker_connection(
-    ws: WebSocket,
-    hub: Arc<HubRouter>,
-    ping_interval: Duration,
-    idle_timeout: Duration,
-) {
-    let conn_id = hub.alloc_conn_id();
-    let (mut ws_tx, ws_rx) = ws.split();
-
-    // Create channel for outbound messages
-    let (tx, mut rx) = mpsc::unbounded_channel::<Message>();
-
-    let conn = Arc::new(WorkerConn::new(conn_id, tx));
-    hub.register_worker(conn.clone());
-
-    // Spawn writer task
-    let writer = tokio::spawn(async move {
-        while let Some(msg) = rx.recv().await {
-            if ws_tx.send(msg).await.is_err() {
-                break;
-            }
-        }
-        let _ = ws_tx.close().await;
-    });
-
-    // Spawn ping task
-    let ping_tx = conn.tx.clone();
-    let ping_task = tokio::spawn(async move {
-        loop {
-            tokio::time::sleep(ping_interval).await;
-            let ping = protocol::encode_ping();
-            if ping_tx.send(Message::Binary(ping.into())).is_err() {
-                break;
-            }
-        }
-    });
-
-    // Spawn reader task
-    let reader_hub = hub.clone();
-    let reader_tx = conn.tx.clone();
-    let reader = tokio::spawn(async move {
-        run_worker_reader(
-            ws_rx,
-            reader_hub,
-            conn_id,
-            conn.clone(),
-            reader_tx,
-            idle_timeout,
-        )
-        .await;
-    });
-
-    // Wait for reader to end, then cleanup.
-    let _ = reader.await;
-    ping_task.abort();
-    // Unregister first so all channel senders are dropped (reader_tx dropped
-    // when reader completes, ping_tx dropped by abort, conn.tx dropped when
-    // the last Arc<WorkerConn> is removed from hub). This lets the writer
-    // drain buffered messages (e.g. GOAWAY) before we force-abort it.
-    hub.unregister_worker(conn_id);
-    tokio::time::sleep(Duration::from_millis(100)).await;
-    writer.abort();
-    let _ = writer.await;
-}
-
-async fn run_worker_reader(
-    mut ws_rx: futures_util::stream::SplitStream<WebSocket>,
-    hub: Arc<HubRouter>,
-    conn_id: u64,
-    conn: Arc<WorkerConn>,
-    tx: mpsc::UnboundedSender<Message>,
-    idle_timeout: Duration,
-) {
-    loop {
-        let msg = tokio::select! {
-            msg = ws_rx.next() => msg,
-            _ = tokio::time::sleep(idle_timeout) => {
-                warn!(worker_id = conn_id, "worker idle timeout");
-                let _ = tx.send(Message::Binary(protocol::encode_goaway().into()));
-                break;
-            }
-        };
-
-        match msg {
-            Some(Ok(Message::Binary(data))) => {
-                let mut data = data.to_vec();
-                if data.len() < protocol::HEADER_SIZE {
-                    debug!(worker_id = conn_id, "frame too small, skipping");
-                    continue;
-                }
-
-                let header = match protocol::FrameHeader::parse(&data) {
-                    Some(h) => h,
-                    None => continue,
-                };
-
-                // HEARTBEAT_ACK from worker -> route back to proxy
-                if header.msg_type == protocol::HEARTBEAT_ACK {
-                    hub.handle_worker_heartbeat_ack(&mut data);
-                    continue;
-                }
-
-                // Regular frames: route via hub
-                if let Some(err_msg) = hub.handle_worker_frame(conn_id, &mut data) {
-                    // Send STREAM_ERROR back to worker
-                    let err_frame = protocol::encode_stream_error(header.stream_id, &err_msg);
-                    let _ = conn.send(Message::Binary(err_frame.into()));
-                }
-            }
-            Some(Ok(Message::Close(_))) | None => {
-                info!(worker_id = conn_id, "worker WebSocket closed");
-                break;
-            }
-            Some(Err(e)) => {
-                warn!(worker_id = conn_id, error = %e, "worker WebSocket error");
-                break;
-            }
-            _ => {} // Ignore text/ping/pong at WS level
-        }
-    }
-}
@@ -10,7 +10,7 @@ checksum = "320119579fcad9c21884f5c4861d16174d0e06250625266f50fe6898340abefa"

 [[package]]
 name = "aether-proxy"
-version = "0.2.3"
+version = "0.2.4"
 dependencies = [
 "anyhow",
 "arc-swap",
@@ -1,6 +1,6 @@
 [package]
 name = "aether-proxy"
-version = "0.2.4"
+version = "0.2.5"
 edition = "2021"
 description = "Tunnel proxy for Aether"

@@ -19,11 +19,11 @@ docker compose up -d
 <!-- DOWNLOAD_TABLE_START -->
 | Platform | Download |
 |----------|----------|
-| Linux x86_64 | [aether-proxy-linux-amd64.tar.gz](https://github.com/fawney19/Aether/releases/download/proxy-v0.2.4/aether-proxy-linux-amd64.tar.gz) |
-| Linux ARM64 | [aether-proxy-linux-arm64.tar.gz](https://github.com/fawney19/Aether/releases/download/proxy-v0.2.4/aether-proxy-linux-arm64.tar.gz) |
-| macOS x86_64 | [aether-proxy-macos-amd64.tar.gz](https://github.com/fawney19/Aether/releases/download/proxy-v0.2.4/aether-proxy-macos-amd64.tar.gz) |
-| macOS ARM64 | [aether-proxy-macos-arm64.tar.gz](https://github.com/fawney19/Aether/releases/download/proxy-v0.2.4/aether-proxy-macos-arm64.tar.gz) |
-| Windows x86_64 | [aether-proxy-windows-amd64.zip](https://github.com/fawney19/Aether/releases/download/proxy-v0.2.4/aether-proxy-windows-amd64.zip) |
+| Linux x86_64 | [aether-proxy-linux-amd64.tar.gz](https://github.com/fawney19/Aether/releases/download/proxy-v0.2.5/aether-proxy-linux-amd64.tar.gz) |
+| Linux ARM64 | [aether-proxy-linux-arm64.tar.gz](https://github.com/fawney19/Aether/releases/download/proxy-v0.2.5/aether-proxy-linux-arm64.tar.gz) |
+| macOS x86_64 | [aether-proxy-macos-amd64.tar.gz](https://github.com/fawney19/Aether/releases/download/proxy-v0.2.5/aether-proxy-macos-amd64.tar.gz) |
+| macOS ARM64 | [aether-proxy-macos-arm64.tar.gz](https://github.com/fawney19/Aether/releases/download/proxy-v0.2.5/aether-proxy-macos-arm64.tar.gz) |
+| Windows x86_64 | [aether-proxy-windows-amd64.zip](https://github.com/fawney19/Aether/releases/download/proxy-v0.2.5/aether-proxy-windows-amd64.zip) |
 <!-- DOWNLOAD_TABLE_END -->

 ## 快速开始
@@ -182,7 +182,9 @@ where

            MsgType::StreamEnd | MsgType::StreamError => {
                // Client-side cancellation or end
-                streams.remove(&frame.stream_id);
+                if let Some(tx) = streams.remove(&frame.stream_id) {
+                    let _ = tx.send(frame).await;
+                }
            }

            MsgType::Ping => {
@@ -3,22 +3,27 @@
 //! Receives request frames, executes the upstream HTTP request,
 //! and sends response frames back through the writer channel.

+use std::io;
+use std::sync::atomic::AtomicUsize;
 use std::sync::atomic::Ordering;
 use std::sync::Arc;
 use std::time::{Duration, Instant};

 use bytes::Bytes;
+use futures_util::stream;
 use futures_util::StreamExt;
 use http_body_util::BodyExt;
+use hyper::body::Frame as BodyFrame;
 use tokio::sync::mpsc;
 use tracing::{debug, warn};

 use crate::state::{AppState, ServerContext};
 use crate::target_filter;
-use crate::upstream_client::{self, UpstreamRequestBody};
+use crate::upstream_client;

 use super::protocol::{
-    compress_payload, decompress_if_gzip, flags, Frame, MsgType, RequestMeta, ResponseMeta,
+    compress_payload, decompress_if_gzip, flags, Frame as TunnelFrame, MsgType, RequestMeta,
+    ResponseMeta,
 };
 use super::writer::FrameSender;

@@ -64,13 +69,13 @@ pub async fn handle_stream(
    server: Arc<ServerContext>,
    stream_id: u32,
    meta: RequestMeta,
-    mut body_rx: mpsc::Receiver<Frame>,
+    body_rx: mpsc::Receiver<TunnelFrame>,
    frame_tx: FrameSender,
 ) {
    server.active_connections.fetch_add(1, Ordering::Release);

    let connect_elapsed =
-        handle_stream_inner(&state, &server, stream_id, meta, &mut body_rx, &frame_tx).await;
+        handle_stream_inner(&state, &server, stream_id, meta, body_rx, &frame_tx).await;

    server.active_connections.fetch_sub(1, Ordering::Release);
    if let Some(d) = connect_elapsed {
@@ -79,7 +84,7 @@ pub async fn handle_stream(
 }

 /// Send a frame to the writer with a timeout. Returns false if send failed.
-async fn send_frame(tx: &FrameSender, frame: Frame) -> bool {
+async fn send_frame(tx: &FrameSender, frame: TunnelFrame) -> bool {
    match tokio::time::timeout(FRAME_SEND_TIMEOUT, tx.send(frame)).await {
        Ok(Ok(())) => true,
        Ok(Err(_)) => {
@@ -102,62 +107,9 @@ async fn handle_stream_inner(
    server: &ServerContext,
    stream_id: u32,
    meta: RequestMeta,
-    body_rx: &mut mpsc::Receiver<Frame>,
+    body_rx: mpsc::Receiver<TunnelFrame>,
    frame_tx: &FrameSender,
 ) -> Option<Duration> {
-    // Collect request body
-    let mut body_parts: Vec<Bytes> = Vec::new();
-    let mut body_done = false;
-
-    // Drain body frames
-    while !body_done {
-        match body_rx.recv().await {
-            Some(frame) => {
-                if frame.msg_type == MsgType::RequestBody {
-                    let payload = match decompress_if_gzip(&frame) {
-                        Ok(d) => d,
-                        Err(e) => {
-                            send_error(
-                                frame_tx,
-                                stream_id,
-                                &format!("gzip decompress failed: {e}"),
-                            )
-                            .await;
-                            return None;
-                        }
-                    };
-                    if !payload.is_empty() {
-                        body_parts.push(payload);
-                    }
-                    if frame.is_end_stream() {
-                        body_done = true;
-                    }
-                } else if frame.msg_type == MsgType::StreamEnd
-                    || frame.msg_type == MsgType::StreamError
-                {
-                    body_done = true;
-                    if frame.msg_type == MsgType::StreamError {
-                        return None; // Client cancelled
-                    }
-                }
-            }
-            None => return None, // Channel closed
-        }
-    }
-
-    let body: Bytes = if body_parts.is_empty() {
-        Bytes::new()
-    } else if body_parts.len() == 1 {
-        body_parts.into_iter().next().unwrap()
-    } else {
-        let total: usize = body_parts.iter().map(|b| b.len()).sum();
-        let mut combined = Vec::with_capacity(total);
-        for part in &body_parts {
-            combined.extend_from_slice(part);
-        }
-        Bytes::from(combined)
-    };
-
    // Validate target
    let target_url = match url::Url::parse(&meta.url) {
        Ok(u) => u,
@@ -207,12 +159,14 @@ async fn handle_stream_inner(
    // Execute upstream request
    let client = &state.upstream_client;
    let timeout = Duration::from_secs(meta.timeout.clamp(MIN_TIMEOUT_SECS, MAX_TIMEOUT_SECS));
+    let request_body_size = Arc::new(AtomicUsize::new(0));
+    let request_body = build_streaming_request_body(body_rx, Arc::clone(&request_body_size));

    let method: hyper::Method = meta.method.parse().unwrap_or(hyper::Method::GET);
    let mut request = match hyper::Request::builder()
        .method(method)
        .uri(meta.url.as_str())
-        .body(UpstreamRequestBody::new(body.clone()))
+        .body(request_body)
    {
        Ok(request) => request,
        Err(e) => {
@@ -240,7 +194,6 @@ async fn handle_stream_inner(
        }
    }

-    let body_size = body.len();
    let mut captured_connection = upstream_client::capture_connection(&mut request);
    let connection_start = Instant::now();
    let connection_capture = tokio::spawn(async move {
@@ -313,7 +266,7 @@ async fn handle_stream_inner(
        "upstream_processing_ms": request_timing.response_wait_ms,
        "timing_source": "instrumented_connector",
        "total_ms": connect_elapsed.as_millis() as u64,
-        "body_size": body_size,
+        "body_size": request_body_size.load(Ordering::Relaxed),
        "mode": "tunnel",
    });
    resp_headers.push(("x-proxy-timing".to_string(), timing.to_string()));
@@ -325,7 +278,7 @@ async fn handle_stream_inner(
    let (meta_payload, meta_flags) = compress_payload(meta_json);
    if !send_frame(
        frame_tx,
-        Frame::new(
+        TunnelFrame::new(
            stream_id,
            MsgType::ResponseHeaders,
            meta_flags,
@@ -350,7 +303,7 @@ async fn handle_stream_inner(
                    let (payload, extra_flags) = compress_payload(chunk);
                    if !send_frame(
                        frame_tx,
-                        Frame::new(stream_id, MsgType::ResponseBody, extra_flags, payload),
+                        TunnelFrame::new(stream_id, MsgType::ResponseBody, extra_flags, payload),
                    )
                    .await
                    {
@@ -365,7 +318,12 @@ async fn handle_stream_inner(
                        let (payload, extra_flags) = compress_payload(slice);
                        if !send_frame(
                            frame_tx,
-                            Frame::new(stream_id, MsgType::ResponseBody, extra_flags, payload),
+                            TunnelFrame::new(
+                                stream_id,
+                                MsgType::ResponseBody,
+                                extra_flags,
+                                payload,
+                            ),
                        )
                        .await
                        {
@@ -387,7 +345,7 @@ async fn handle_stream_inner(
    // Send STREAM_END
    let _ = send_frame(
        frame_tx,
-        Frame::new(
+        TunnelFrame::new(
            stream_id,
            MsgType::StreamEnd,
            flags::END_STREAM,
@@ -404,7 +362,7 @@ async fn send_error(tx: &FrameSender, stream_id: u32, msg: &str) {
    // Error frames use best-effort delivery — don't block if writer is congested
    let _ = send_frame(
        tx,
-        Frame::new(
+        TunnelFrame::new(
            stream_id,
            MsgType::StreamError,
            0,
@@ -413,3 +371,136 @@ async fn send_error(tx: &FrameSender, stream_id: u32, msg: &str) {
    )
    .await;
 }
+
+fn build_streaming_request_body(
+    body_rx: mpsc::Receiver<TunnelFrame>,
+    body_size: Arc<AtomicUsize>,
+) -> upstream_client::UpstreamRequestBody {
+    let body_stream = stream::unfold(
+        (body_rx, body_size, false),
+        |(mut body_rx, body_size, finished)| async move {
+            if finished {
+                return None;
+            }
+
+            loop {
+                let frame = match body_rx.recv().await {
+                    Some(frame) => frame,
+                    None => return None,
+                };
+
+                match frame.msg_type {
+                    MsgType::RequestBody => {
+                        let end_stream = frame.is_end_stream();
+                        let payload = match decompress_if_gzip(&frame) {
+                            Ok(payload) => payload,
+                            Err(error) => {
+                                let err =
+                                    io::Error::other(format!("gzip decompress failed: {error}"));
+                                return Some((Err(err), (body_rx, body_size, true)));
+                            }
+                        };
+
+                        if payload.is_empty() {
+                            if end_stream {
+                                return None;
+                            }
+                            continue;
+                        }
+
+                        body_size.fetch_add(payload.len(), Ordering::Relaxed);
+                        return Some((
+                            Ok(BodyFrame::data(payload)),
+                            (body_rx, body_size, end_stream),
+                        ));
+                    }
+                    MsgType::StreamError => {
+                        let message = String::from_utf8(frame.payload.to_vec())
+                            .unwrap_or_else(|_| "client cancelled request body".to_string());
+                        return Some((Err(io::Error::other(message)), (body_rx, body_size, true)));
+                    }
+                    MsgType::StreamEnd => return None,
+                    _ => continue,
+                }
+            }
+        },
+    );
+
+    upstream_client::stream_request_body(body_stream)
+}
+
+#[cfg(test)]
+mod tests {
+    use super::*;
+
+    #[tokio::test]
+    async fn streaming_request_body_yields_chunks_and_tracks_size() {
+        let (tx, rx) = mpsc::channel(4);
+        let body_size = Arc::new(AtomicUsize::new(0));
+        let mut body = build_streaming_request_body(rx, Arc::clone(&body_size));
+
+        tx.send(TunnelFrame::new(
+            1,
+            MsgType::RequestBody,
+            0,
+            Bytes::from_static(b"abc"),
+        ))
+        .await
+        .expect("send first chunk");
+        tx.send(TunnelFrame::new(
+            1,
+            MsgType::RequestBody,
+            flags::END_STREAM,
+            Bytes::from_static(b"def"),
+        ))
+        .await
+        .expect("send final chunk");
+        drop(tx);
+
+        let first = body
+            .frame()
+            .await
+            .expect("first frame")
+            .expect("first frame ok")
+            .into_data()
+            .expect("first data frame");
+        let second = body
+            .frame()
+            .await
+            .expect("second frame")
+            .expect("second frame ok")
+            .into_data()
+            .expect("second data frame");
+
+        assert_eq!(first, Bytes::from_static(b"abc"));
+        assert_eq!(second, Bytes::from_static(b"def"));
+        assert!(body.frame().await.is_none());
+        assert_eq!(body_size.load(Ordering::Relaxed), 6);
+    }
+
+    #[tokio::test]
+    async fn streaming_request_body_surfaces_client_cancel_as_error() {
+        let (tx, rx) = mpsc::channel(4);
+        let body_size = Arc::new(AtomicUsize::new(0));
+        let mut body = build_streaming_request_body(rx, Arc::clone(&body_size));
+
+        tx.send(TunnelFrame::new(
+            1,
+            MsgType::StreamError,
+            0,
+            Bytes::from_static(b"client cancelled"),
+        ))
+        .await
+        .expect("send cancel frame");
+        drop(tx);
+
+        let err = body
+            .frame()
+            .await
+            .expect("error frame present")
+            .expect_err("body should surface cancellation error");
+        assert!(err.to_string().contains("client cancelled"));
+        assert!(body.frame().await.is_none());
+        assert_eq!(body_size.load(Ordering::Relaxed), 0);
+    }
+}
@@ -7,7 +7,10 @@ use std::task::{Context, Poll};
 use std::time::Duration;

 use bytes::Bytes;
-use http_body_util::Full;
+use futures_util::Stream;
+use http_body_util::combinators::UnsyncBoxBody;
+use http_body_util::{BodyExt, StreamBody};
+use hyper::body::Frame;
 use hyper::rt;
 use hyper::Response;
 use hyper::Uri;
@@ -30,9 +33,16 @@ type BoxError = Box<dyn std::error::Error + Send + Sync>;
 type PlainStream = TokioIo<TcpStream>;
 type TlsStream = TokioIo<tokio_rustls::client::TlsStream<TcpStream>>;

-pub type UpstreamRequestBody = Full<Bytes>;
+pub type UpstreamRequestBody = UnsyncBoxBody<Bytes, io::Error>;
 pub type UpstreamClient = Client<InstrumentedConnector, UpstreamRequestBody>;

+pub fn stream_request_body<S>(stream: S) -> UpstreamRequestBody
+where
+    S: Stream<Item = Result<Frame<Bytes>, io::Error>> + Send + 'static,
+{
+    StreamBody::new(stream).boxed_unsync()
+}
+
 #[derive(Clone, Copy, Debug, Default)]
 pub struct ConnectTiming {
    pub connect_ms: u64,
@@ -14,8 +14,6 @@ from alembic import context

 # 添加项目根目录到 Python 路径
 sys.path.insert(0, os.path.dirname(os.path.dirname(__file__)))
-# 添加 alembic 目录，让迁移文件可以 `from helpers import ...`
-sys.path.insert(0, os.path.dirname(__file__))

 # 加载 .env 文件（本地开发时需要）
 try:
@@ -1,171 +0,0 @@
-"""Shared idempotent helpers for Alembic migrations.
-
-All metadata lookups are batched: one query loads an entire table's column info
-or all FK delete rules, then results are cached for the migration's lifetime.
-"""
-
-from __future__ import annotations
-
-from collections import defaultdict
-from collections.abc import Callable
-
-import sqlalchemy as sa
-
-from alembic import op
-
-# ---------------------------------------------------------------------------
-# Batch metadata cache
-# ---------------------------------------------------------------------------
-
-
-class _SchemaCache:
-    """Lazy, per-migration cache for information_schema lookups.
-
-    Call ``load_columns(tables)`` / ``load_fk_rules(tables)`` once at the top
-    of ``upgrade()`` or ``downgrade()`` to prime the cache.  Subsequent
-    ``column_exists`` / ``column_type`` / ``fk_ondelete`` calls are pure
-    dict lookups -- zero extra DB round-trips.
-    """
-
-    def __init__(self) -> None:
-        # {table_name: {column_name: data_type}}
-        self._columns: dict[str, dict[str, str]] = {}
-        # {(table_name, constraint_name): delete_rule}
-        self._fk_rules: dict[tuple[str, str], str] = {}
-        self._fk_loaded_tables: set[str] = set()
-
-    # -- loaders (one query per call) --------------------------------------
-
-    def load_columns(self, tables: list[str]) -> None:
-        """Fetch column names + data types for *tables* in one query."""
-        need = [t for t in tables if t not in self._columns]
-        if not need:
-            return
-        bind = op.get_bind()
-        rows = bind.execute(
-            sa.text(
-                "SELECT table_name, column_name, data_type "
-                "FROM information_schema.columns "
-                "WHERE table_name = ANY(:tables)"
-            ),
-            {"tables": need},
-        ).fetchall()
-        # Initialise even empty tables so we don't re-query
-        for t in need:
-            self._columns.setdefault(t, {})
-        for table, col, dtype in rows:
-            self._columns[table][col] = dtype
-
-    def load_fk_rules(self, tables: list[str]) -> None:
-        """Fetch FK delete rules for *tables* in one query."""
-        need = [t for t in tables if t not in self._fk_loaded_tables]
-        if not need:
-            return
-        bind = op.get_bind()
-        rows = bind.execute(
-            sa.text(
-                "SELECT tc.table_name, tc.constraint_name, rc.delete_rule "
-                "FROM information_schema.referential_constraints rc "
-                "JOIN information_schema.table_constraints tc "
-                "  ON rc.constraint_name = tc.constraint_name "
-                "WHERE tc.table_name = ANY(:tables)"
-            ),
-            {"tables": need},
-        ).fetchall()
-        for table, name, rule in rows:
-            self._fk_rules[(table, name)] = rule
-        self._fk_loaded_tables.update(need)
-
-    # -- lookups (pure dict, zero DB) --------------------------------------
-
-    def column_exists(self, table: str, column: str) -> bool:
-        return column in self._columns.get(table, {})
-
-    def column_type(self, table: str, column: str) -> str | None:
-        return self._columns.get(table, {}).get(column)
-
-    def is_numeric(self, table: str, column: str) -> bool:
-        return self.column_type(table, column) == "numeric"
-
-    def fk_ondelete(self, table: str, constraint: str) -> str | None:
-        return self._fk_rules.get((table, constraint))
-
-    def invalidate_columns(self, table: str) -> None:
-        """Force re-load on next load_columns() for *table* (after ADD/DROP COLUMN)."""
-        self._columns.pop(table, None)
-
-
-def new_cache() -> _SchemaCache:
-    """Create a fresh schema cache for a single migration run."""
-    return _SchemaCache()
-
-
-# ---------------------------------------------------------------------------
-# Idempotent DDL helpers
-# ---------------------------------------------------------------------------
-
-
-def replace_fk_if_needed(
-    cache: _SchemaCache,
-    constraint_name: str,
-    table_name: str,
-    ref_table: str,
-    local_cols: list[str],
-    remote_cols: list[str],
-    desired_ondelete: str,
-) -> None:
-    """Drop and recreate a FK only if the current ON DELETE rule differs."""
-    current = cache.fk_ondelete(table_name, constraint_name)
-    if current and current.upper() == desired_ondelete.upper():
-        return
-    if current:
-        op.drop_constraint(constraint_name, table_name, type_="foreignkey")
-    op.create_foreign_key(
-        constraint_name,
-        table_name,
-        ref_table,
-        local_cols,
-        remote_cols,
-        ondelete=desired_ondelete,
-    )
-
-
-def index_exists(index_name: str) -> bool:
-    bind = op.get_bind()
-    result = bind.execute(
-        sa.text("SELECT 1 FROM pg_indexes WHERE indexname = :name"),
-        {"name": index_name},
-    )
-    return result.scalar() is not None
-
-
-# ---------------------------------------------------------------------------
-# Batch ALTER TYPE helper
-# ---------------------------------------------------------------------------
-
-
-def batch_alter_type(
-    cache: _SchemaCache,
-    columns: list[tuple[str, str, bool, str | None]],
-    cast_suffix: str,
-    type_fn: Callable[[str], str],
-) -> None:
-    """Group columns by table and issue ONE ``ALTER TABLE`` per table.
-
-    Skips columns that don't exist.  Each tuple is
-    ``(table_name, column_name, nullable, server_default)``.
-    """
-    by_table: dict[str, list[tuple[str, str]]] = defaultdict(list)
-    for table, col, _nullable, _default in columns:
-        if not cache.column_exists(table, col):
-            continue
-        by_table[table].append((col, type_fn(col)))
-
-    bind = op.get_bind()
-    for table, col_types in by_table.items():
-        parts = [
-            f"ALTER COLUMN {col} TYPE {target} USING {col}::{cast_suffix}"
-            for col, target in col_types
-        ]
-        if parts:
-            bind.execute(sa.text(f"ALTER TABLE {table} " + ", ".join(parts)))
@@ -53,7 +53,23 @@ def _index_exists(table_name: str, index_name: str) -> bool:
 def _to_decimal(value: object | None) -> Decimal:
    if value is None:
        return Decimal("0")
-    return Decimal(str(value)).quantize(_MONEY_QUANT, rounding=ROUND_HALF_UP)
+    d = Decimal(str(value))
+    if d.is_nan() or d.is_infinite():
+        return Decimal("0")
+    return d.quantize(_MONEY_QUANT, rounding=ROUND_HALF_UP)
+
+
+# NUMERIC(20, 8) -> 整数部分最多 12 位 -> 绝对值上限
+_NUMERIC_20_8_CAP = Decimal("999999999999.99999999")
+
+
+def _clamp(value: Decimal) -> Decimal:
+    """Clamp a Decimal to fit NUMERIC(20, 8) range."""
+    if value > _NUMERIC_20_8_CAP:
+        return _NUMERIC_20_8_CAP
+    if value < -_NUMERIC_20_8_CAP:
+        return -_NUMERIC_20_8_CAP
+    return value


 def _insert_wallet(
@@ -191,6 +207,11 @@ def _backfill_wallets(conn: sa.Connection) -> None:
        total_adjusted = Decimal("0") if is_unlimited else max(quota, Decimal("0"))
        total_consumed = max(total if total > 0 else used, Decimal("0"))

+        recharge_balance = _clamp(recharge_balance)
+        gift_balance = _clamp(gift_balance)
+        total_adjusted = _clamp(total_adjusted)
+        total_consumed = _clamp(total_consumed)
+
        total_balance = recharge_balance + gift_balance
        wallet_id = _insert_wallet(
            conn,
@@ -208,7 +229,7 @@ def _backfill_wallets(conn: sa.Connection) -> None:
        _insert_wallet_migration_tx(
            conn,
            wallet_id=wallet_id,
-            balance=total_balance,
+            balance=_clamp(total_balance),
            recharge_balance=recharge_balance,
            gift_balance=gift_balance,
            created_at=row["created_at"],
@@ -235,7 +256,9 @@ def _backfill_wallets(conn: sa.Connection) -> None:

        current_balance = _to_decimal(row["current_balance_usd"])
        used_balance = _to_decimal(row["balance_used_usd"])
-        recharge_balance = current_balance - used_balance
+        recharge_balance = _clamp(current_balance - used_balance)
+        total_recharged = _clamp(max(current_balance, Decimal("0")))
+        total_consumed = _clamp(max(used_balance, Decimal("0")))

        wallet_id = _insert_wallet(
            conn,
@@ -244,8 +267,8 @@ def _backfill_wallets(conn: sa.Connection) -> None:
            balance=recharge_balance,
            gift_balance=Decimal("0"),
            limit_mode="finite",
-            total_recharged=max(current_balance, Decimal("0")),
-            total_consumed=max(used_balance, Decimal("0")),
+            total_recharged=total_recharged,
+            total_consumed=total_consumed,
            total_adjusted=Decimal("0"),
            created_at=row["created_at"],
            updated_at=row["updated_at"],
@@ -6,8 +6,9 @@ Create Date: 2026-03-08 03:48:49.622091+00:00

 """

+from __future__ import annotations
+
 import sqlalchemy as sa
-from helpers import new_cache, replace_fk_if_needed

 from alembic import op

@@ -20,8 +21,108 @@ depends_on = None
 _TABLES = ["usage", "stats_user_daily", "stats_daily_api_key"]


+# ---------------------------------------------------------------------------
+# Inline helpers
+# ---------------------------------------------------------------------------
+
+
+class _SchemaCache:
+    def __init__(self) -> None:
+        self._columns: dict[str, dict[str, str]] = {}
+        self._fk_rules: dict[tuple[str, str], str] = {}
+        self._fk_loaded_tables: set[str] = set()
+
+    def load_columns(self, tables: list[str]) -> None:
+        need = [t for t in tables if t not in self._columns]
+        if not need:
+            return
+        bind = op.get_bind()
+        rows = bind.execute(
+            sa.text(
+                "SELECT table_name, column_name, data_type "
+                "FROM information_schema.columns "
+                "WHERE table_name = ANY(:tables) "
+                "  AND table_schema = current_schema()"
+            ),
+            {"tables": need},
+        ).fetchall()
+        for t in need:
+            self._columns.setdefault(t, {})
+        for table, col, dtype in rows:
+            self._columns[table][col] = dtype
+
+    def load_fk_rules(self, tables: list[str]) -> None:
+        need = [t for t in tables if t not in self._fk_loaded_tables]
+        if not need:
+            return
+        bind = op.get_bind()
+        rows = bind.execute(
+            sa.text(
+                "SELECT tc.table_name, tc.constraint_name, rc.delete_rule "
+                "FROM information_schema.referential_constraints rc "
+                "JOIN information_schema.table_constraints tc "
+                "  ON rc.constraint_name = tc.constraint_name "
+                " AND rc.constraint_schema = tc.constraint_schema "
+                "WHERE tc.table_name = ANY(:tables) "
+                "  AND tc.table_schema = current_schema()"
+            ),
+            {"tables": need},
+        ).fetchall()
+        for table, name, rule in rows:
+            self._fk_rules[(table, name)] = rule
+        self._fk_loaded_tables.update(need)
+
+    def column_exists(self, table: str, column: str) -> bool:
+        return column in self._columns.get(table, {})
+
+    def fk_ondelete(self, table: str, constraint: str) -> str | None:
+        return self._fk_rules.get((table, constraint))
+
+
+def _fk_exists(constraint_name: str, table_name: str) -> bool:
+    bind = op.get_bind()
+    result = bind.execute(
+        sa.text(
+            "SELECT 1 FROM pg_constraint c "
+            "JOIN pg_class r ON c.conrelid = r.oid "
+            "JOIN pg_namespace n ON r.relnamespace = n.oid "
+            "WHERE c.conname = :name AND r.relname = :table "
+            "  AND n.nspname = current_schema() AND c.contype = 'f'"
+        ),
+        {"name": constraint_name, "table": table_name},
+    )
+    return result.scalar() is not None
+
+
+def _replace_fk_if_needed(
+    cache: _SchemaCache,
+    constraint_name: str,
+    table_name: str,
+    ref_table: str,
+    local_cols: list[str],
+    remote_cols: list[str],
+    desired_ondelete: str,
+) -> None:
+    current = cache.fk_ondelete(table_name, constraint_name)
+    if current and current.upper() == desired_ondelete.upper():
+        return
+    if current or _fk_exists(constraint_name, table_name):
+        op.drop_constraint(constraint_name, table_name, type_="foreignkey")
+    op.create_foreign_key(
+        constraint_name,
+        table_name,
+        ref_table,
+        local_cols,
+        remote_cols,
+        ondelete=desired_ondelete,
+    )
+
+
+# ---------------------------------------------------------------------------
+
+
 def upgrade() -> None:
-    c = new_cache()
+    c = _SchemaCache()
    c.load_columns(_TABLES)
    c.load_fk_rules(["stats_user_daily", "stats_daily_api_key"])

@@ -37,7 +138,7 @@ def upgrade() -> None:
        )

    # --- StatsUserDaily: CASCADE -> SET NULL, add username snapshot ---
-    replace_fk_if_needed(
+    _replace_fk_if_needed(
        c,
        "stats_user_daily_user_id_fkey",
        "stats_user_daily",
@@ -59,7 +160,7 @@ def upgrade() -> None:
        )

    # --- StatsDailyApiKey: CASCADE -> SET NULL, add api_key_name snapshot ---
-    replace_fk_if_needed(
+    _replace_fk_if_needed(
        c,
        "stats_daily_api_key_api_key_id_fkey",
        "stats_daily_api_key",
@@ -82,7 +183,7 @@ def upgrade() -> None:


 def downgrade() -> None:
-    c = new_cache()
+    c = _SchemaCache()
    c.load_columns(["stats_daily_api_key", "stats_user_daily", "usage"])
    c.load_fk_rules(["stats_daily_api_key", "stats_user_daily"])

@@ -97,7 +198,7 @@ def downgrade() -> None:
        op.drop_column("usage", "username")

    # --- StatsDailyApiKey: SET NULL -> CASCADE ---
-    replace_fk_if_needed(
+    _replace_fk_if_needed(
        c,
        "stats_daily_api_key_api_key_id_fkey",
        "stats_daily_api_key",
@@ -111,7 +212,7 @@ def downgrade() -> None:
    )

    # --- StatsUserDaily: SET NULL -> CASCADE ---
-    replace_fk_if_needed(
+    _replace_fk_if_needed(
        c,
        "stats_user_daily_user_id_fkey",
        "stats_user_daily",
@@ -6,8 +6,9 @@ Create Date: 2026-03-08 12:15:00.000000+00:00

 """

+from __future__ import annotations
+
 import sqlalchemy as sa
-from helpers import new_cache, replace_fk_if_needed

 from alembic import op

@@ -20,8 +21,108 @@ depends_on = None
 _TABLES = ["request_candidates", "video_tasks"]


+# ---------------------------------------------------------------------------
+# Inline helpers
+# ---------------------------------------------------------------------------
+
+
+class _SchemaCache:
+    def __init__(self) -> None:
+        self._columns: dict[str, dict[str, str]] = {}
+        self._fk_rules: dict[tuple[str, str], str] = {}
+        self._fk_loaded_tables: set[str] = set()
+
+    def load_columns(self, tables: list[str]) -> None:
+        need = [t for t in tables if t not in self._columns]
+        if not need:
+            return
+        bind = op.get_bind()
+        rows = bind.execute(
+            sa.text(
+                "SELECT table_name, column_name, data_type "
+                "FROM information_schema.columns "
+                "WHERE table_name = ANY(:tables) "
+                "  AND table_schema = current_schema()"
+            ),
+            {"tables": need},
+        ).fetchall()
+        for t in need:
+            self._columns.setdefault(t, {})
+        for table, col, dtype in rows:
+            self._columns[table][col] = dtype
+
+    def load_fk_rules(self, tables: list[str]) -> None:
+        need = [t for t in tables if t not in self._fk_loaded_tables]
+        if not need:
+            return
+        bind = op.get_bind()
+        rows = bind.execute(
+            sa.text(
+                "SELECT tc.table_name, tc.constraint_name, rc.delete_rule "
+                "FROM information_schema.referential_constraints rc "
+                "JOIN information_schema.table_constraints tc "
+                "  ON rc.constraint_name = tc.constraint_name "
+                " AND rc.constraint_schema = tc.constraint_schema "
+                "WHERE tc.table_name = ANY(:tables) "
+                "  AND tc.table_schema = current_schema()"
+            ),
+            {"tables": need},
+        ).fetchall()
+        for table, name, rule in rows:
+            self._fk_rules[(table, name)] = rule
+        self._fk_loaded_tables.update(need)
+
+    def column_exists(self, table: str, column: str) -> bool:
+        return column in self._columns.get(table, {})
+
+    def fk_ondelete(self, table: str, constraint: str) -> str | None:
+        return self._fk_rules.get((table, constraint))
+
+
+def _fk_exists(constraint_name: str, table_name: str) -> bool:
+    bind = op.get_bind()
+    result = bind.execute(
+        sa.text(
+            "SELECT 1 FROM pg_constraint c "
+            "JOIN pg_class r ON c.conrelid = r.oid "
+            "JOIN pg_namespace n ON r.relnamespace = n.oid "
+            "WHERE c.conname = :name AND r.relname = :table "
+            "  AND n.nspname = current_schema() AND c.contype = 'f'"
+        ),
+        {"name": constraint_name, "table": table_name},
+    )
+    return result.scalar() is not None
+
+
+def _replace_fk_if_needed(
+    cache: _SchemaCache,
+    constraint_name: str,
+    table_name: str,
+    ref_table: str,
+    local_cols: list[str],
+    remote_cols: list[str],
+    desired_ondelete: str,
+) -> None:
+    current = cache.fk_ondelete(table_name, constraint_name)
+    if current and current.upper() == desired_ondelete.upper():
+        return
+    if current or _fk_exists(constraint_name, table_name):
+        op.drop_constraint(constraint_name, table_name, type_="foreignkey")
+    op.create_foreign_key(
+        constraint_name,
+        table_name,
+        ref_table,
+        local_cols,
+        remote_cols,
+        ondelete=desired_ondelete,
+    )
+
+
+# ---------------------------------------------------------------------------
+
+
 def upgrade() -> None:
-    c = new_cache()
+    c = _SchemaCache()
    c.load_columns(_TABLES)
    c.load_fk_rules(_TABLES)

@@ -43,7 +144,7 @@ def upgrade() -> None:
        )

    # --- request_candidates: CASCADE -> SET NULL ---
-    replace_fk_if_needed(
+    _replace_fk_if_needed(
        c,
        "request_candidates_user_id_fkey",
        "request_candidates",
@@ -52,7 +153,7 @@ def upgrade() -> None:
        ["id"],
        "SET NULL",
    )
-    replace_fk_if_needed(
+    _replace_fk_if_needed(
        c,
        "request_candidates_api_key_id_fkey",
        "request_candidates",
@@ -81,7 +182,7 @@ def upgrade() -> None:

    # --- video_tasks: CASCADE -> SET NULL, user_id nullable ---
    op.alter_column("video_tasks", "user_id", existing_type=sa.String(length=36), nullable=True)
-    replace_fk_if_needed(
+    _replace_fk_if_needed(
        c,
        "video_tasks_user_id_fkey",
        "video_tasks",
@@ -90,7 +191,7 @@ def upgrade() -> None:
        ["id"],
        "SET NULL",
    )
-    replace_fk_if_needed(
+    _replace_fk_if_needed(
        c,
        "video_tasks_api_key_id_fkey",
        "video_tasks",
@@ -102,12 +203,12 @@ def upgrade() -> None:


 def downgrade() -> None:
-    c = new_cache()
+    c = _SchemaCache()
    c.load_columns(_TABLES)
    c.load_fk_rules(_TABLES)

    # --- video_tasks: SET NULL -> default (no action), restore NOT NULL ---
-    replace_fk_if_needed(
+    _replace_fk_if_needed(
        c,
        "video_tasks_api_key_id_fkey",
        "video_tasks",
@@ -116,7 +217,7 @@ def downgrade() -> None:
        ["id"],
        "NO ACTION",
    )
-    replace_fk_if_needed(
+    _replace_fk_if_needed(
        c,
        "video_tasks_user_id_fkey",
        "video_tasks",
@@ -132,7 +233,7 @@ def downgrade() -> None:
        op.drop_column("video_tasks", "username")

    # --- request_candidates: SET NULL -> CASCADE ---
-    replace_fk_if_needed(
+    _replace_fk_if_needed(
        c,
        "request_candidates_api_key_id_fkey",
        "request_candidates",
@@ -141,7 +242,7 @@ def downgrade() -> None:
        ["id"],
        "CASCADE",
    )
-    replace_fk_if_needed(
+    _replace_fk_if_needed(
        c,
        "request_candidates_user_id_fkey",
        "request_candidates",
@@ -6,8 +6,15 @@ Create Date: 2026-03-08 15:30:00.000000+00:00

 """

+from __future__ import annotations
+
+import re
+from collections import defaultdict
+from collections.abc import Callable
+
+import sqlalchemy as sa
+
 from alembic import op
-from helpers import batch_alter_type, index_exists, new_cache

 # revision identifiers, used by Alembic.
 revision = "2053ab8ed764"
@@ -79,13 +86,106 @@ _COST_COLUMNS: list[tuple[str, str, bool, str | None]] = [
 _ALL_TABLES = list({t for t, *_ in _COST_COLUMNS})


+# ---------------------------------------------------------------------------
+# Inline helpers
+# ---------------------------------------------------------------------------
+
+
+class _SchemaCache:
+    def __init__(self) -> None:
+        self._columns: dict[str, dict[str, str]] = {}
+
+    def load_columns(self, tables: list[str]) -> None:
+        need = [t for t in tables if t not in self._columns]
+        if not need:
+            return
+        bind = op.get_bind()
+        rows = bind.execute(
+            sa.text(
+                "SELECT table_name, column_name, data_type "
+                "FROM information_schema.columns "
+                "WHERE table_name = ANY(:tables) "
+                "  AND table_schema = current_schema()"
+            ),
+            {"tables": need},
+        ).fetchall()
+        for t in need:
+            self._columns.setdefault(t, {})
+        for table, col, dtype in rows:
+            self._columns[table][col] = dtype
+
+    def column_exists(self, table: str, column: str) -> bool:
+        return column in self._columns.get(table, {})
+
+    def column_type(self, table: str, column: str) -> str | None:
+        return self._columns.get(table, {}).get(column)
+
+    def is_numeric(self, table: str, column: str) -> bool:
+        return self.column_type(table, column) == "numeric"
+
+
+def _index_exists(index_name: str) -> bool:
+    bind = op.get_bind()
+    result = bind.execute(
+        sa.text(
+            "SELECT 1 FROM pg_indexes "
+            "WHERE indexname = :name AND schemaname = current_schema()::text"
+        ),
+        {"name": index_name},
+    )
+    return result.scalar() is not None
+
+
+def _numeric_max(type_spec: str) -> float | None:
+    m = re.match(r"NUMERIC\((\d+),(\d+)\)", type_spec, re.IGNORECASE)
+    if not m:
+        return None
+    precision, scale = int(m.group(1)), int(m.group(2))
+    return 10 ** (precision - scale) - 10 ** (-scale)
+
+
+def _batch_alter_type(
+    cache: _SchemaCache,
+    columns: list[tuple[str, str, bool, str | None]],
+    cast_suffix: str,
+    type_fn: Callable[[str], str],
+) -> None:
+    by_table: dict[str, list[tuple[str, str]]] = defaultdict(list)
+    for table, col, _nullable, _default in columns:
+        if not cache.column_exists(table, col):
+            continue
+        by_table[table].append((col, type_fn(col)))
+
+    bind = op.get_bind()
+    for table, col_types in by_table.items():
+        for col, target in col_types:
+            cap = _numeric_max(target)
+            if cap is not None:
+                bind.execute(
+                    sa.text(
+                        f"UPDATE {table} SET {col} = :cap "
+                        f"WHERE {col} IS NOT NULL AND abs({col}) > :cap"
+                    ),
+                    {"cap": cap},
+                )
+        parts = [
+            f"ALTER COLUMN {col} TYPE {target} USING {col}::{cast_suffix}"
+            for col, target in col_types
+        ]
+        if parts:
+            bind.execute(sa.text(f"ALTER TABLE {table} " + ", ".join(parts)))
+
+
+# ---------------------------------------------------------------------------
+
+
 def _type_spec(col: str) -> str:
    """Return the SQL type literal for a given column name."""
    return "NUMERIC(10,6)" if col == "rate_multiplier" else "NUMERIC(20,8)"


 def upgrade() -> None:
-    c = new_cache()
+    c = _SchemaCache()
    c.load_columns(_ALL_TABLES)

    # -- 1. cost fields: Float -> Numeric  (batched per table)
@@ -94,10 +194,10 @@ def upgrade() -> None:
        for t, col, n, d in _COST_COLUMNS
        if c.column_exists(t, col) and not c.is_numeric(t, col)
    ]
-    batch_alter_type(c, cols_to_convert, cast_suffix="numeric", type_fn=_type_spec)
+    _batch_alter_type(c, cols_to_convert, cast_suffix="numeric", type_fn=_type_spec)

    # -- 2. provider_api_keys composite index
-    if not index_exists("idx_provider_api_keys_provider_active"):
+    if not _index_exists("idx_provider_api_keys_provider_active"):
        op.create_index(
            "idx_provider_api_keys_provider_active",
            "provider_api_keys",
@@ -107,14 +207,14 @@ def upgrade() -> None:

 def downgrade() -> None:
    # -- 2. drop composite index
-    if index_exists("idx_provider_api_keys_provider_active"):
+    if _index_exists("idx_provider_api_keys_provider_active"):
        op.drop_index(
            "idx_provider_api_keys_provider_active",
            table_name="provider_api_keys",
        )

    # -- 1. Numeric -> Float  (batched per table)
-    c = new_cache()
+    c = _SchemaCache()
    c.load_columns(_ALL_TABLES)

    cols_to_revert = [
@@ -122,7 +222,7 @@ def downgrade() -> None:
        for t, col, n, d in _COST_COLUMNS
        if c.column_exists(t, col) and c.is_numeric(t, col)
    ]
-    batch_alter_type(
+    _batch_alter_type(
        c,
        cols_to_revert,
        cast_suffix="double precision",
@@ -6,7 +6,11 @@ Create Date: 2026-03-09 01:00:00.000000+00:00

 """

-from helpers import new_cache, replace_fk_if_needed
+from __future__ import annotations
+
+import sqlalchemy as sa
+
+from alembic import op

 # revision identifiers, used by Alembic.
 revision = "d7649c1f8e21"
@@ -18,10 +22,87 @@ _TABLE = "video_tasks"
 _FK_NAME = "video_tasks_key_id_fkey"


+# ---------------------------------------------------------------------------
+# Inline helpers
+# ---------------------------------------------------------------------------
+
+
+class _SchemaCache:
+    def __init__(self) -> None:
+        self._fk_rules: dict[tuple[str, str], str] = {}
+        self._fk_loaded_tables: set[str] = set()
+
+    def load_fk_rules(self, tables: list[str]) -> None:
+        need = [t for t in tables if t not in self._fk_loaded_tables]
+        if not need:
+            return
+        bind = op.get_bind()
+        rows = bind.execute(
+            sa.text(
+                "SELECT tc.table_name, tc.constraint_name, rc.delete_rule "
+                "FROM information_schema.referential_constraints rc "
+                "JOIN information_schema.table_constraints tc "
+                "  ON rc.constraint_name = tc.constraint_name "
+                " AND rc.constraint_schema = tc.constraint_schema "
+                "WHERE tc.table_name = ANY(:tables) "
+                "  AND tc.table_schema = current_schema()"
+            ),
+            {"tables": need},
+        ).fetchall()
+        for table, name, rule in rows:
+            self._fk_rules[(table, name)] = rule
+        self._fk_loaded_tables.update(need)
+
+    def fk_ondelete(self, table: str, constraint: str) -> str | None:
+        return self._fk_rules.get((table, constraint))
+
+
+def _fk_exists(constraint_name: str, table_name: str) -> bool:
+    bind = op.get_bind()
+    result = bind.execute(
+        sa.text(
+            "SELECT 1 FROM pg_constraint c "
+            "JOIN pg_class r ON c.conrelid = r.oid "
+            "JOIN pg_namespace n ON r.relnamespace = n.oid "
+            "WHERE c.conname = :name AND r.relname = :table "
+            "  AND n.nspname = current_schema() AND c.contype = 'f'"
+        ),
+        {"name": constraint_name, "table": table_name},
+    )
+    return result.scalar() is not None
+
+
+def _replace_fk_if_needed(
+    cache: _SchemaCache,
+    constraint_name: str,
+    table_name: str,
+    ref_table: str,
+    local_cols: list[str],
+    remote_cols: list[str],
+    desired_ondelete: str,
+) -> None:
+    current = cache.fk_ondelete(table_name, constraint_name)
+    if current and current.upper() == desired_ondelete.upper():
+        return
+    if current or _fk_exists(constraint_name, table_name):
+        op.drop_constraint(constraint_name, table_name, type_="foreignkey")
+    op.create_foreign_key(
+        constraint_name,
+        table_name,
+        ref_table,
+        local_cols,
+        remote_cols,
+        ondelete=desired_ondelete,
+    )
+
+
+# ---------------------------------------------------------------------------
+
+
 def upgrade() -> None:
-    c = new_cache()
+    c = _SchemaCache()
    c.load_fk_rules([_TABLE])
-    replace_fk_if_needed(
+    _replace_fk_if_needed(
        c,
        _FK_NAME,
        _TABLE,
@@ -33,9 +114,9 @@ def upgrade() -> None:


 def downgrade() -> None:
-    c = new_cache()
+    c = _SchemaCache()
    c.load_fk_rules([_TABLE])
-    replace_fk_if_needed(
+    _replace_fk_if_needed(
        c,
        _FK_NAME,
        _TABLE,
@@ -0,0 +1,42 @@
+"""Strip request_results_window from health_by_format JSON.
+
+This data is now maintained in process memory only, no longer persisted to DB.
+
+Revision ID: a3f1b7c9d2e4
+Revises: d7649c1f8e21
+Create Date: 2026-03-10 12:00:00.000000+00:00
+
+"""
+
+from alembic import op
+
+# revision identifiers, used by Alembic.
+revision = "a3f1b7c9d2e4"
+down_revision = "d7649c1f8e21"
+branch_labels = None
+depends_on = None
+
+
+def upgrade() -> None:
+    op.execute("""
+        UPDATE provider_api_keys
+        SET health_by_format = (
+            SELECT jsonb_object_agg(
+                fmt_key,
+                fmt_value - 'request_results_window'
+            )
+            FROM jsonb_each(health_by_format) AS x(fmt_key, fmt_value)
+        )
+        WHERE health_by_format IS NOT NULL
+          AND health_by_format != '{}'::jsonb
+          AND EXISTS (
+              SELECT 1
+              FROM jsonb_each(health_by_format) AS x(fmt_key, fmt_value)
+              WHERE fmt_value ? 'request_results_window'
+          )
+    """)
+
+
+def downgrade() -> None:
+    # No-op: window data is rebuilt from scratch on process start
+    pass
@@ -0,0 +1,159 @@
+"""tighten usage billing state machine
+
+Revision ID: 9e4f1a2b3c4d
+Revises: a3f1b7c9d2e4
+Create Date: 2026-03-11 19:00:00.000000+00:00
+
+This migration does two things:
+1. Change new `usage.billing_status` default from `settled` to `pending`.
+2. Repair only the clearly-safe inconsistent historical rows for production:
+   - failed/cancelled zero-cost rows that were marked settled are converted to void
+   - terminal rows missing finalized_at are backfilled from created_at
+
+Ambiguous positive-cost settled rows are intentionally left untouched for manual audit.
+
+All data updates are batched (10000 rows per iteration) to avoid long-held locks
+and excessive WAL generation on large usage tables.
+"""
+
+from __future__ import annotations
+
+from collections.abc import Sequence
+
+import sqlalchemy as sa
+from sqlalchemy import inspect
+
+from alembic import op
+
+# revision identifiers, used by Alembic.
+revision: str = "9e4f1a2b3c4d"
+down_revision: str | None = "a3f1b7c9d2e4"
+branch_labels: str | Sequence[str] | None = None
+depends_on: str | Sequence[str] | None = None
+
+BATCH_SIZE = 10000
+
+
+def _table_exists(table_name: str) -> bool:
+    bind = op.get_bind()
+    insp = inspect(bind)
+    insp.clear_cache()
+    return table_name in insp.get_table_names()
+
+
+def _column_exists(table_name: str, column_name: str) -> bool:
+    bind = op.get_bind()
+    insp = inspect(bind)
+    insp.clear_cache()
+    return column_name in [col["name"] for col in insp.get_columns(table_name)]
+
+
+def upgrade() -> None:
+    if not _table_exists("usage"):
+        return
+
+    if _column_exists("usage", "billing_status"):
+        op.alter_column(
+            "usage",
+            "billing_status",
+            existing_type=sa.String(length=20),
+            server_default="pending",
+            existing_nullable=False,
+        )
+
+    required_columns = {
+        "billing_status",
+        "status",
+        "total_cost_usd",
+        "request_cost_usd",
+        "actual_total_cost_usd",
+        "actual_request_cost_usd",
+        "wallet_balance_after",
+        "finalized_at",
+        "created_at",
+    }
+    if not required_columns.issubset(
+        {col for col in required_columns if _column_exists("usage", col)}
+    ):
+        return
+
+    conn = op.get_bind()
+
+    # Step 1: billing_status IS NULL -> 'pending' (batched)
+    while True:
+        result = conn.execute(
+            sa.text("""
+                WITH batch AS (
+                    SELECT id FROM usage
+                    WHERE billing_status IS NULL
+                    LIMIT :batch_size
+                    FOR UPDATE SKIP LOCKED
+                )
+                UPDATE usage
+                SET billing_status = 'pending'
+                FROM batch WHERE usage.id = batch.id
+                """),
+            {"batch_size": BATCH_SIZE},
+        )
+        if result.rowcount < BATCH_SIZE:
+            break
+
+    # Step 2: failed/cancelled zero-cost settled -> void (batched)
+    while True:
+        result = conn.execute(
+            sa.text("""
+                WITH batch AS (
+                    SELECT id FROM usage
+                    WHERE billing_status = 'settled'
+                      AND status IN ('failed', 'cancelled')
+                      AND COALESCE(total_cost_usd, 0) = 0
+                      AND wallet_balance_after IS NULL
+                    LIMIT :batch_size
+                    FOR UPDATE SKIP LOCKED
+                )
+                UPDATE usage
+                SET billing_status = 'void',
+                    finalized_at = COALESCE(usage.finalized_at, usage.created_at),
+                    total_cost_usd = 0,
+                    request_cost_usd = 0,
+                    actual_total_cost_usd = 0,
+                    actual_request_cost_usd = 0
+                FROM batch WHERE usage.id = batch.id
+                """),
+            {"batch_size": BATCH_SIZE},
+        )
+        if result.rowcount < BATCH_SIZE:
+            break
+
+    # Step 3: backfill finalized_at for terminal rows (batched)
+    while True:
+        result = conn.execute(
+            sa.text("""
+                WITH batch AS (
+                    SELECT id FROM usage
+                    WHERE billing_status IN ('settled', 'void')
+                      AND finalized_at IS NULL
+                    LIMIT :batch_size
+                    FOR UPDATE SKIP LOCKED
+                )
+                UPDATE usage
+                SET finalized_at = COALESCE(usage.finalized_at, usage.created_at)
+                FROM batch WHERE usage.id = batch.id
+                """),
+            {"batch_size": BATCH_SIZE},
+        )
+        if result.rowcount < BATCH_SIZE:
+            break
+
+
+def downgrade() -> None:
+    if not _table_exists("usage") or not _column_exists("usage", "billing_status"):
+        return
+
+    op.alter_column(
+        "usage",
+        "billing_status",
+        existing_type=sa.String(length=20),
+        server_default="settled",
+        existing_nullable=False,
+    )
@@ -0,0 +1,113 @@
+"""add wallet daily usage ledgers
+
+Revision ID: d4e5f6a7b8c9
+Revises: 9e4f1a2b3c4d
+Create Date: 2026-03-11 21:00:00.000000+00:00
+"""
+
+from __future__ import annotations
+
+from collections.abc import Sequence
+
+import sqlalchemy as sa
+from sqlalchemy import inspect
+
+from alembic import op
+
+# revision identifiers, used by Alembic.
+revision: str = "d4e5f6a7b8c9"
+down_revision: str | None = "9e4f1a2b3c4d"
+branch_labels: str | Sequence[str] | None = None
+depends_on: str | Sequence[str] | None = None
+
+
+def _table_exists(table_name: str) -> bool:
+    bind = op.get_bind()
+    insp = inspect(bind)
+    insp.clear_cache()
+    return table_name in insp.get_table_names()
+
+
+def _column_exists(table_name: str, column_name: str) -> bool:
+    bind = op.get_bind()
+    insp = inspect(bind)
+    insp.clear_cache()
+    return column_name in [col["name"] for col in insp.get_columns(table_name)]
+
+
+def _index_exists(table_name: str, index_name: str) -> bool:
+    bind = op.get_bind()
+    insp = inspect(bind)
+    insp.clear_cache()
+    return any(idx["name"] == index_name for idx in insp.get_indexes(table_name))
+
+
+def upgrade() -> None:
+    if not _table_exists("wallet_daily_usage_ledgers"):
+        op.create_table(
+            "wallet_daily_usage_ledgers",
+            sa.Column("id", sa.String(length=36), nullable=False),
+            sa.Column("wallet_id", sa.String(length=36), nullable=False),
+            sa.Column("billing_date", sa.Date(), nullable=False),
+            sa.Column("billing_timezone", sa.String(length=64), nullable=False),
+            sa.Column("total_cost_usd", sa.Numeric(20, 8), nullable=False, server_default="0"),
+            sa.Column("total_requests", sa.Integer(), nullable=False, server_default="0"),
+            sa.Column("input_tokens", sa.BigInteger(), nullable=False, server_default="0"),
+            sa.Column("output_tokens", sa.BigInteger(), nullable=False, server_default="0"),
+            sa.Column("cache_creation_tokens", sa.BigInteger(), nullable=False, server_default="0"),
+            sa.Column("cache_read_tokens", sa.BigInteger(), nullable=False, server_default="0"),
+            sa.Column("first_finalized_at", sa.DateTime(timezone=True), nullable=True),
+            sa.Column("last_finalized_at", sa.DateTime(timezone=True), nullable=True),
+            sa.Column("aggregated_at", sa.DateTime(timezone=True), nullable=False),
+            sa.Column("created_at", sa.DateTime(timezone=True), nullable=False),
+            sa.Column("updated_at", sa.DateTime(timezone=True), nullable=False),
+            sa.ForeignKeyConstraint(["wallet_id"], ["wallets.id"], ondelete="CASCADE"),
+            sa.PrimaryKeyConstraint("id"),
+            sa.UniqueConstraint(
+                "wallet_id",
+                "billing_date",
+                "billing_timezone",
+                name="uq_wallet_daily_usage_ledgers_wallet_date_tz",
+            ),
+        )
+
+    if not _index_exists("wallet_daily_usage_ledgers", "idx_wallet_daily_usage_wallet_date"):
+        op.create_index(
+            "idx_wallet_daily_usage_wallet_date",
+            "wallet_daily_usage_ledgers",
+            ["wallet_id", "billing_date"],
+        )
+    if not _index_exists("wallet_daily_usage_ledgers", "idx_wallet_daily_usage_date"):
+        op.create_index(
+            "idx_wallet_daily_usage_date",
+            "wallet_daily_usage_ledgers",
+            ["billing_date"],
+        )
+
+    if (
+        _table_exists("usage")
+        and all(
+            _column_exists("usage", col) for col in ["billing_status", "finalized_at", "wallet_id"]
+        )
+        and not _index_exists("usage", "idx_usage_billing_finalized_wallet")
+    ):
+        op.create_index(
+            "idx_usage_billing_finalized_wallet",
+            "usage",
+            ["billing_status", "finalized_at", "wallet_id"],
+        )
+
+
+def downgrade() -> None:
+    if _table_exists("usage") and _index_exists("usage", "idx_usage_billing_finalized_wallet"):
+        op.drop_index("idx_usage_billing_finalized_wallet", table_name="usage")
+
+    if _table_exists("wallet_daily_usage_ledgers"):
+        if _index_exists("wallet_daily_usage_ledgers", "idx_wallet_daily_usage_date"):
+            op.drop_index("idx_wallet_daily_usage_date", table_name="wallet_daily_usage_ledgers")
+        if _index_exists("wallet_daily_usage_ledgers", "idx_wallet_daily_usage_wallet_date"):
+            op.drop_index(
+                "idx_wallet_daily_usage_wallet_date",
+                table_name="wallet_daily_usage_ledgers",
+            )
+        op.drop_table("wallet_daily_usage_ledgers")
@@ -0,0 +1,59 @@
+"""add provider_api_keys usage total columns
+
+Revision ID: 9b7c6d5e4f3a
+Revises: d4e5f6a7b8c9
+Create Date: 2026-03-11 22:00:00.000000+00:00
+"""
+
+from __future__ import annotations
+
+from collections.abc import Sequence
+
+import sqlalchemy as sa
+from sqlalchemy import inspect
+
+from alembic import op
+
+# revision identifiers, used by Alembic.
+revision: str = "9b7c6d5e4f3a"
+down_revision: str | None = "d4e5f6a7b8c9"
+branch_labels: str | Sequence[str] | None = None
+depends_on: str | Sequence[str] | None = None
+
+
+def column_exists(table_name: str, column_name: str) -> bool:
+    bind = op.get_bind()
+    inspector = inspect(bind)
+    columns = [c["name"] for c in inspector.get_columns(table_name)]
+    return column_name in columns
+
+
+def upgrade() -> None:
+    if not column_exists("provider_api_keys", "total_tokens"):
+        op.add_column(
+            "provider_api_keys",
+            sa.Column("total_tokens", sa.BigInteger(), nullable=False, server_default="0"),
+        )
+    if column_exists("provider_api_keys", "total_tokens"):
+        op.alter_column("provider_api_keys", "total_tokens", server_default=None)
+
+    if not column_exists("provider_api_keys", "total_cost_usd"):
+        op.add_column(
+            "provider_api_keys",
+            sa.Column(
+                "total_cost_usd",
+                sa.Numeric(20, 8),
+                nullable=False,
+                server_default="0.0",
+            ),
+        )
+    if column_exists("provider_api_keys", "total_cost_usd"):
+        op.alter_column("provider_api_keys", "total_cost_usd", server_default=None)
+
+
+def downgrade() -> None:
+    if column_exists("provider_api_keys", "total_cost_usd"):
+        op.drop_column("provider_api_keys", "total_cost_usd")
+
+    if column_exists("provider_api_keys", "total_tokens"):
+        op.drop_column("provider_api_keys", "total_tokens")
@@ -0,0 +1,116 @@
+"""cleanup stale provider references after provider deletion
+
+Revision ID: c1d2e3f4a5b6
+Revises: 9b7c6d5e4f3a
+Create Date: 2026-03-11 23:00:00.000000+00:00
+
+"""
+
+from __future__ import annotations
+
+import sqlalchemy as sa
+
+from alembic import op
+
+# revision identifiers, used by Alembic.
+revision = "c1d2e3f4a5b6"
+down_revision = "9b7c6d5e4f3a"
+branch_labels = None
+depends_on = None
+
+_users = sa.table(
+    "users",
+    sa.column("id", sa.String(36)),
+    sa.column("allowed_providers", sa.JSON()),
+)
+_api_keys = sa.table(
+    "api_keys",
+    sa.column("id", sa.String(36)),
+    sa.column("allowed_providers", sa.JSON()),
+)
+_user_preferences = sa.table(
+    "user_preferences",
+    sa.column("id", sa.String(36)),
+    sa.column("default_provider_id", sa.String(36)),
+)
+_video_tasks = sa.table(
+    "video_tasks",
+    sa.column("id", sa.String(36)),
+    sa.column("provider_id", sa.String(36)),
+    sa.column("endpoint_id", sa.String(36)),
+)
+_providers = sa.table("providers", sa.column("id", sa.String(36)))
+_provider_endpoints = sa.table("provider_endpoints", sa.column("id", sa.String(36)))
+
+
+def _load_valid_ids(conn: sa.Connection, table: sa.Table) -> set[str]:
+    return {str(row[0]) for row in conn.execute(sa.select(table.c.id)).fetchall() if row[0]}
+
+
+def _cleanup_allowed_providers(
+    conn: sa.Connection,
+    table: sa.Table,
+    valid_provider_ids: set[str],
+) -> None:
+    rows = conn.execute(
+        sa.select(table.c.id, table.c.allowed_providers).where(
+            table.c.allowed_providers.isnot(None)
+        )
+    ).fetchall()
+    for row_id, allowed_providers in rows:
+        if not isinstance(allowed_providers, list):
+            continue
+        filtered = [
+            provider_id for provider_id in allowed_providers if provider_id in valid_provider_ids
+        ]
+        if filtered == allowed_providers:
+            continue
+        conn.execute(table.update().where(table.c.id == row_id).values(allowed_providers=filtered))
+
+
+def _nullify_missing_fk(
+    conn: sa.Connection,
+    table: sa.Table,
+    id_column: sa.ColumnElement[str],
+    fk_column: sa.ColumnElement[str],
+    valid_ids: set[str],
+) -> None:
+    rows = conn.execute(sa.select(id_column, fk_column).where(fk_column.isnot(None))).fetchall()
+    invalid_row_ids = [row_id for row_id, fk_value in rows if fk_value not in valid_ids]
+    if not invalid_row_ids:
+        return
+    conn.execute(table.update().where(id_column.in_(invalid_row_ids)).values({fk_column.key: None}))
+
+
+def upgrade() -> None:
+    conn = op.get_bind()
+    valid_provider_ids = _load_valid_ids(conn, _providers)
+    valid_endpoint_ids = _load_valid_ids(conn, _provider_endpoints)
+
+    _cleanup_allowed_providers(conn, _users, valid_provider_ids)
+    _cleanup_allowed_providers(conn, _api_keys, valid_provider_ids)
+    _nullify_missing_fk(
+        conn,
+        _user_preferences,
+        _user_preferences.c.id,
+        _user_preferences.c.default_provider_id,
+        valid_provider_ids,
+    )
+    _nullify_missing_fk(
+        conn,
+        _video_tasks,
+        _video_tasks.c.id,
+        _video_tasks.c.provider_id,
+        valid_provider_ids,
+    )
+    _nullify_missing_fk(
+        conn,
+        _video_tasks,
+        _video_tasks.c.id,
+        _video_tasks.c.endpoint_id,
+        valid_endpoint_ids,
+    )
+
+
+def downgrade() -> None:
+    pass
@@ -0,0 +1,64 @@
+"""decouple request_candidates.key_id foreign key from provider_api_keys lifecycle
+
+Revision ID: b7c8d9e0f1a2
+Revises: c1d2e3f4a5b6
+Create Date: 2026-03-12 19:15:00.000000+00:00
+
+"""
+
+from __future__ import annotations
+
+import sqlalchemy as sa
+
+from alembic import op
+
+# revision identifiers, used by Alembic.
+revision = "b7c8d9e0f1a2"
+down_revision = "c1d2e3f4a5b6"
+branch_labels = None
+depends_on = None
+
+
+def _fk_exists(constraint_name: str, table_name: str) -> bool:
+    bind = op.get_bind()
+    result = bind.execute(
+        sa.text(
+            "SELECT 1 FROM pg_constraint c "
+            "JOIN pg_class r ON c.conrelid = r.oid "
+            "JOIN pg_namespace n ON r.relnamespace = n.oid "
+            "WHERE c.conname = :name AND r.relname = :table "
+            "  AND n.nspname = current_schema() AND c.contype = 'f'"
+        ),
+        {"name": constraint_name, "table": table_name},
+    )
+    return result.scalar() is not None
+
+
+def upgrade() -> None:
+    if _fk_exists("request_candidates_key_id_fkey", "request_candidates"):
+        op.drop_constraint(
+            "request_candidates_key_id_fkey", "request_candidates", type_="foreignkey"
+        )
+
+
+def downgrade() -> None:
+    bind = op.get_bind()
+    bind.execute(
+        sa.text(
+            "UPDATE request_candidates rc "
+            "SET key_id = NULL "
+            "WHERE key_id IS NOT NULL "
+            "  AND NOT EXISTS ("
+            "    SELECT 1 FROM provider_api_keys pak WHERE pak.id = rc.key_id"
+            "  )"
+        )
+    )
+    if not _fk_exists("request_candidates_key_id_fkey", "request_candidates"):
+        op.create_foreign_key(
+            "request_candidates_key_id_fkey",
+            "request_candidates",
+            "provider_api_keys",
+            ["key_id"],
+            ["id"],
+            ondelete="CASCADE",
+        )
@@ -0,0 +1,54 @@
+"""add user rate_limit and backfill normal api key limits
+
+Revision ID: b7e8f9a0c1d2
+Revises: b7c8d9e0f1a2
+Create Date: 2026-03-13 12:00:00.000000+00:00
+"""
+
+from __future__ import annotations
+
+from collections.abc import Sequence
+
+import sqlalchemy as sa
+from sqlalchemy import inspect
+
+from alembic import op
+
+# revision identifiers, used by Alembic.
+revision: str = "b7e8f9a0c1d2"
+down_revision: str | None = "b7c8d9e0f1a2"
+branch_labels: str | Sequence[str] | None = None
+depends_on: str | Sequence[str] | None = None
+
+
+def column_exists(table_name: str, column_name: str) -> bool:
+    bind = op.get_bind()
+    inspector = inspect(bind)
+    columns = [c["name"] for c in inspector.get_columns(table_name)]
+    return column_name in columns
+
+
+def upgrade() -> None:
+    if not column_exists("users", "rate_limit"):
+        op.add_column("users", sa.Column("rate_limit", sa.Integer(), nullable=True))
+
+    # 普通 Key 新语义不再允许 NULL；存量 NULL 统一回填为 0（不限制）。
+    op.execute(sa.text("""
+            UPDATE api_keys
+            SET rate_limit = 0
+            WHERE is_standalone = FALSE
+              AND rate_limit IS NULL
+            """))
+
+
+def downgrade() -> None:
+    # 恢复普通 Key 的 rate_limit 为 NULL（与 upgrade 中回填 0 对应）
+    op.execute(sa.text("""
+            UPDATE api_keys
+            SET rate_limit = NULL
+            WHERE is_standalone = FALSE
+              AND rate_limit = 0
+            """))
+
+    if column_exists("users", "rate_limit"):
+        op.drop_column("users", "rate_limit")
@@ -0,0 +1,87 @@
+"""add user sessions table for device-level auth
+
+Revision ID: f6e7d8c9b0a1
+Revises: b7e8f9a0c1d2
+Create Date: 2026-03-15 12:00:00.000000+00:00
+
+"""
+
+from __future__ import annotations
+
+import sqlalchemy as sa
+
+from alembic import op
+
+# revision identifiers, used by Alembic.
+revision = "f6e7d8c9b0a1"
+down_revision = "b7e8f9a0c1d2"
+branch_labels = None
+depends_on = None
+
+
+def upgrade() -> None:
+    bind = op.get_bind()
+    inspector = sa.inspect(bind)
+    if "user_sessions" in inspector.get_table_names():
+        return
+
+    op.create_table(
+        "user_sessions",
+        sa.Column("id", sa.String(length=36), nullable=False),
+        sa.Column("user_id", sa.String(length=36), nullable=False),
+        sa.Column("client_device_id", sa.String(length=128), nullable=False),
+        sa.Column("device_label", sa.String(length=120), nullable=True),
+        sa.Column("device_type", sa.String(length=20), nullable=False, server_default="unknown"),
+        sa.Column("browser_name", sa.String(length=50), nullable=True),
+        sa.Column("browser_version", sa.String(length=50), nullable=True),
+        sa.Column("os_name", sa.String(length=50), nullable=True),
+        sa.Column("os_version", sa.String(length=50), nullable=True),
+        sa.Column("device_model", sa.String(length=100), nullable=True),
+        sa.Column("ip_address", sa.String(length=45), nullable=True),
+        sa.Column("user_agent", sa.String(length=1000), nullable=True),
+        sa.Column("client_hints", sa.JSON(), nullable=True),
+        sa.Column("refresh_token_hash", sa.String(length=64), nullable=False),
+        sa.Column("prev_refresh_token_hash", sa.String(length=64), nullable=True),
+        sa.Column("rotated_at", sa.DateTime(timezone=True), nullable=True),
+        sa.Column(
+            "last_seen_at", sa.DateTime(timezone=True), nullable=False, server_default=sa.func.now()
+        ),
+        sa.Column("expires_at", sa.DateTime(timezone=True), nullable=False),
+        sa.Column("revoked_at", sa.DateTime(timezone=True), nullable=True),
+        sa.Column("revoke_reason", sa.String(length=100), nullable=True),
+        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()
+        ),
+        sa.ForeignKeyConstraint(["user_id"], ["users.id"], ondelete="CASCADE"),
+        sa.PrimaryKeyConstraint("id"),
+    )
+    op.create_index("ix_user_sessions_user_id", "user_sessions", ["user_id"], unique=False)
+    op.create_index(
+        "ix_user_sessions_client_device_id",
+        "user_sessions",
+        ["client_device_id"],
+        unique=False,
+    )
+    op.create_index(
+        "idx_user_sessions_user_active",
+        "user_sessions",
+        ["user_id", "revoked_at", "expires_at"],
+        unique=False,
+    )
+    op.create_index(
+        "idx_user_sessions_user_device",
+        "user_sessions",
+        ["user_id", "client_device_id"],
+        unique=False,
+    )
+
+
+def downgrade() -> None:
+    op.drop_index("idx_user_sessions_user_device", table_name="user_sessions")
+    op.drop_index("idx_user_sessions_user_active", table_name="user_sessions")
+    op.drop_index("ix_user_sessions_client_device_id", table_name="user_sessions")
+    op.drop_index("ix_user_sessions_user_id", table_name="user_sessions")
+    op.drop_table("user_sessions")
@@ -0,0 +1,41 @@
+"""add status_snapshot column to provider_api_keys
+
+Revision ID: c9d8e7f6a5b4
+Revises: f6e7d8c9b0a1
+Create Date: 2026-03-20 12:00:00.000000
+"""
+
+from __future__ import annotations
+
+from collections.abc import Sequence
+
+import sqlalchemy as sa
+from sqlalchemy import inspect
+
+from alembic import op
+
+# revision identifiers, used by Alembic.
+revision: str = "c9d8e7f6a5b4"
+down_revision: str | None = "f6e7d8c9b0a1"
+branch_labels: str | Sequence[str] | None = None
+depends_on: str | Sequence[str] | None = None
+
+
+def column_exists(table_name: str, column_name: str) -> bool:
+    bind = op.get_bind()
+    inspector = inspect(bind)
+    columns = [c["name"] for c in inspector.get_columns(table_name)]
+    return column_name in columns
+
+
+def upgrade() -> None:
+    if not column_exists("provider_api_keys", "status_snapshot"):
+        op.add_column(
+            "provider_api_keys",
+            sa.Column("status_snapshot", sa.JSON(), nullable=True),
+        )
+
+
+def downgrade() -> None:
+    if column_exists("provider_api_keys", "status_snapshot"):
+        op.drop_column("provider_api_keys", "status_snapshot")
@@ -0,0 +1,317 @@
+"""usage token semantics v2
+
+Revision ID: c3d4e5f6a7b8
+Revises: c9d8e7f6a5b4
+Create Date: 2026-03-24 14:00:00.000000+00:00
+"""
+
+from __future__ import annotations
+
+from collections.abc import Sequence
+
+import sqlalchemy as sa
+from sqlalchemy import inspect
+
+from alembic import op
+
+# revision identifiers, used by Alembic.
+revision: str = "c3d4e5f6a7b8"
+down_revision: str | None = "c9d8e7f6a5b4"
+branch_labels: str | Sequence[str] | None = None
+depends_on: str | Sequence[str] | None = None
+
+BACKFILL_BATCH_SIZE = 2000
+
+# 最大批次数，防止因数据异常导致死循环（2000 * 500000 = 10亿行上限）
+_MAX_BATCHES = 500000
+
+# 使用子查询中间层展开 input_output_total_tokens 的计算，
+# 确保 total_tokens 引用的是本次 SET 后的新值而非旧值。
+_UPGRADE_BACKFILL_SQL = sa.text(
+    """
+    UPDATE usage
+    SET
+        input_output_total_tokens = src.new_iot,
+        input_context_tokens      = src.new_ict,
+        total_tokens              = src.new_total,
+        cache_creation_cost_usd_5m        = src.new_cc5m,
+        cache_creation_cost_usd_1h        = src.new_cc1h,
+        actual_cache_creation_cost_usd_5m = src.new_acc5m,
+        actual_cache_creation_cost_usd_1h = src.new_acc1h,
+        actual_cache_cost_usd             = src.new_accu,
+        cache_creation_price_per_1m_5m    = src.new_cp5m,
+        cache_creation_price_per_1m_1h    = src.new_cp1h,
+        cache_cost_usd                    = src.new_ccu
+    FROM (
+        SELECT
+            id,
+            COALESCE(input_output_total_tokens,
+                     COALESCE(input_tokens, 0) + COALESCE(output_tokens, 0))
+                AS new_iot,
+            COALESCE(input_tokens, 0) + COALESCE(cache_read_input_tokens, 0)
+                AS new_ict,
+            /* total_tokens 引用本行计算出的 new_iot，避免依赖 SET 顺序 */
+            COALESCE(input_output_total_tokens,
+                     COALESCE(input_tokens, 0) + COALESCE(output_tokens, 0))
+                + COALESCE(cache_creation_input_tokens, 0)
+                + COALESCE(cache_read_input_tokens, 0)
+                AS new_total,
+            CASE
+                WHEN COALESCE(cache_creation_input_tokens_5m, 0) > 0
+                     AND COALESCE(cache_creation_input_tokens_1h, 0) = 0
+                THEN COALESCE(cache_creation_cost_usd, 0)
+                WHEN COALESCE(cache_creation_input_tokens_5m, 0) > 0
+                     AND COALESCE(cache_creation_input_tokens, 0) > 0
+                THEN COALESCE(cache_creation_cost_usd, 0)
+                     * (COALESCE(cache_creation_input_tokens_5m, 0) * 1.0
+                        / GREATEST(COALESCE(cache_creation_input_tokens, 0), 1))
+                ELSE 0
+            END AS new_cc5m,
+            CASE
+                WHEN COALESCE(cache_creation_input_tokens_1h, 0) > 0
+                     AND COALESCE(cache_creation_input_tokens_5m, 0) = 0
+                THEN COALESCE(cache_creation_cost_usd, 0)
+                WHEN COALESCE(cache_creation_input_tokens_1h, 0) > 0
+                     AND COALESCE(cache_creation_input_tokens, 0) > 0
+                THEN COALESCE(cache_creation_cost_usd, 0)
+                     * (COALESCE(cache_creation_input_tokens_1h, 0) * 1.0
+                        / GREATEST(COALESCE(cache_creation_input_tokens, 0), 1))
+                ELSE 0
+            END AS new_cc1h,
+            CASE
+                WHEN COALESCE(cache_creation_input_tokens_5m, 0) > 0
+                     AND COALESCE(cache_creation_input_tokens_1h, 0) = 0
+                THEN COALESCE(actual_cache_creation_cost_usd, 0)
+                WHEN COALESCE(cache_creation_input_tokens_5m, 0) > 0
+                     AND COALESCE(cache_creation_input_tokens, 0) > 0
+                THEN COALESCE(actual_cache_creation_cost_usd, 0)
+                     * (COALESCE(cache_creation_input_tokens_5m, 0) * 1.0
+                        / GREATEST(COALESCE(cache_creation_input_tokens, 0), 1))
+                ELSE 0
+            END AS new_acc5m,
+            CASE
+                WHEN COALESCE(cache_creation_input_tokens_1h, 0) > 0
+                     AND COALESCE(cache_creation_input_tokens_5m, 0) = 0
+                THEN COALESCE(actual_cache_creation_cost_usd, 0)
+                WHEN COALESCE(cache_creation_input_tokens_1h, 0) > 0
+                     AND COALESCE(cache_creation_input_tokens, 0) > 0
+                THEN COALESCE(actual_cache_creation_cost_usd, 0)
+                     * (COALESCE(cache_creation_input_tokens_1h, 0) * 1.0
+                        / GREATEST(COALESCE(cache_creation_input_tokens, 0), 1))
+                ELSE 0
+            END AS new_acc1h,
+            COALESCE(actual_cache_creation_cost_usd, 0)
+                + COALESCE(actual_cache_read_cost_usd, 0) AS new_accu,
+            CASE
+                WHEN COALESCE(cache_creation_input_tokens_5m, 0) > 0
+                     AND COALESCE(cache_creation_input_tokens_1h, 0) = 0
+                THEN cache_creation_price_per_1m
+                ELSE NULL
+            END AS new_cp5m,
+            CASE
+                WHEN COALESCE(cache_creation_input_tokens_1h, 0) > 0
+                     AND COALESCE(cache_creation_input_tokens_5m, 0) = 0
+                THEN cache_creation_price_per_1m
+                ELSE NULL
+            END AS new_cp1h,
+            COALESCE(cache_creation_cost_usd, 0)
+                + COALESCE(cache_read_cost_usd, 0) AS new_ccu
+        FROM usage
+        WHERE id IN (
+            SELECT id FROM usage
+            WHERE
+                input_output_total_tokens IS DISTINCT FROM
+                    COALESCE(input_output_total_tokens,
+                             COALESCE(input_tokens, 0) + COALESCE(output_tokens, 0))
+                OR input_context_tokens IS DISTINCT FROM
+                    COALESCE(input_tokens, 0) + COALESCE(cache_read_input_tokens, 0)
+                OR total_tokens IS DISTINCT FROM (
+                    COALESCE(input_output_total_tokens,
+                             COALESCE(input_tokens, 0) + COALESCE(output_tokens, 0))
+                    + COALESCE(cache_creation_input_tokens, 0)
+                    + COALESCE(cache_read_input_tokens, 0)
+                )
+                OR cache_creation_cost_usd_5m IS DISTINCT FROM (
+                    CASE
+                        WHEN COALESCE(cache_creation_input_tokens_5m, 0) > 0
+                             AND COALESCE(cache_creation_input_tokens_1h, 0) = 0
+                        THEN COALESCE(cache_creation_cost_usd, 0)
+                        WHEN COALESCE(cache_creation_input_tokens_5m, 0) > 0
+                             AND COALESCE(cache_creation_input_tokens, 0) > 0
+                        THEN COALESCE(cache_creation_cost_usd, 0)
+                             * (COALESCE(cache_creation_input_tokens_5m, 0) * 1.0
+                                / GREATEST(COALESCE(cache_creation_input_tokens, 0), 1))
+                        ELSE 0
+                    END
+                )
+                OR cache_creation_cost_usd_1h IS DISTINCT FROM (
+                    CASE
+                        WHEN COALESCE(cache_creation_input_tokens_1h, 0) > 0
+                             AND COALESCE(cache_creation_input_tokens_5m, 0) = 0
+                        THEN COALESCE(cache_creation_cost_usd, 0)
+                        WHEN COALESCE(cache_creation_input_tokens_1h, 0) > 0
+                             AND COALESCE(cache_creation_input_tokens, 0) > 0
+                        THEN COALESCE(cache_creation_cost_usd, 0)
+                             * (COALESCE(cache_creation_input_tokens_1h, 0) * 1.0
+                                / GREATEST(COALESCE(cache_creation_input_tokens, 0), 1))
+                        ELSE 0
+                    END
+                )
+                OR actual_cache_creation_cost_usd_5m IS DISTINCT FROM (
+                    CASE
+                        WHEN COALESCE(cache_creation_input_tokens_5m, 0) > 0
+                             AND COALESCE(cache_creation_input_tokens_1h, 0) = 0
+                        THEN COALESCE(actual_cache_creation_cost_usd, 0)
+                        WHEN COALESCE(cache_creation_input_tokens_5m, 0) > 0
+                             AND COALESCE(cache_creation_input_tokens, 0) > 0
+                        THEN COALESCE(actual_cache_creation_cost_usd, 0)
+                             * (COALESCE(cache_creation_input_tokens_5m, 0) * 1.0
+                                / GREATEST(COALESCE(cache_creation_input_tokens, 0), 1))
+                        ELSE 0
+                    END
+                )
+                OR actual_cache_creation_cost_usd_1h IS DISTINCT FROM (
+                    CASE
+                        WHEN COALESCE(cache_creation_input_tokens_1h, 0) > 0
+                             AND COALESCE(cache_creation_input_tokens_5m, 0) = 0
+                        THEN COALESCE(actual_cache_creation_cost_usd, 0)
+                        WHEN COALESCE(cache_creation_input_tokens_1h, 0) > 0
+                             AND COALESCE(cache_creation_input_tokens, 0) > 0
+                        THEN COALESCE(actual_cache_creation_cost_usd, 0)
+                             * (COALESCE(cache_creation_input_tokens_1h, 0) * 1.0
+                                / GREATEST(COALESCE(cache_creation_input_tokens, 0), 1))
+                        ELSE 0
+                    END
+                )
+                OR actual_cache_cost_usd IS DISTINCT FROM (
+                    COALESCE(actual_cache_creation_cost_usd, 0)
+                    + COALESCE(actual_cache_read_cost_usd, 0)
+                )
+                OR cache_creation_price_per_1m_5m IS DISTINCT FROM (
+                    CASE
+                        WHEN COALESCE(cache_creation_input_tokens_5m, 0) > 0
+                             AND COALESCE(cache_creation_input_tokens_1h, 0) = 0
+                        THEN cache_creation_price_per_1m
+                        ELSE NULL
+                    END
+                )
+                OR cache_creation_price_per_1m_1h IS DISTINCT FROM (
+                    CASE
+                        WHEN COALESCE(cache_creation_input_tokens_1h, 0) > 0
+                             AND COALESCE(cache_creation_input_tokens_5m, 0) = 0
+                        THEN cache_creation_price_per_1m
+                        ELSE NULL
+                    END
+                )
+                OR cache_cost_usd IS DISTINCT FROM (
+                    COALESCE(cache_creation_cost_usd, 0) + COALESCE(cache_read_cost_usd, 0)
+                )
+            ORDER BY id
+            LIMIT :batch_size
+        )
+    ) AS src
+    WHERE usage.id = src.id
+    """
+)
+
+_DOWNGRADE_BACKFILL_SQL = sa.text(
+    """
+    UPDATE usage
+    SET total_tokens = COALESCE(input_output_total_tokens, COALESCE(input_tokens, 0) + COALESCE(output_tokens, 0))
+    WHERE id IN (
+        SELECT id
+        FROM usage
+        WHERE total_tokens IS DISTINCT FROM
+            COALESCE(input_output_total_tokens, COALESCE(input_tokens, 0) + COALESCE(output_tokens, 0))
+        ORDER BY id
+        LIMIT :batch_size
+    )
+    """
+)
+
+
+def column_exists(table_name: str, column_name: str) -> bool:
+    bind = op.get_bind()
+    inspector = inspect(bind)
+    columns = [c["name"] for c in inspector.get_columns(table_name)]
+    return column_name in columns
+
+
+def run_backfill_in_batches(sql: sa.TextClause, batch_size: int = BACKFILL_BATCH_SIZE) -> None:
+    context = op.get_context()
+    for _ in range(_MAX_BATCHES):
+        # Commit the preceding schema transaction before each batch so PostgreSQL
+        # does not keep ALTER TABLE locks for the entire data backfill.
+        with context.autocommit_block():
+            rowcount = op.get_bind().execute(sql, {"batch_size": batch_size}).rowcount
+        if rowcount == 0:
+            break
+    else:
+        raise RuntimeError(
+            f"Backfill did not converge after {_MAX_BATCHES} batches "
+            f"(batch_size={batch_size}). Possible infinite loop due to data anomaly."
+        )
+
+
+def upgrade() -> None:
+    if column_exists("usage", "total_tokens") and not column_exists("usage", "input_output_total_tokens"):
+        with op.batch_alter_table("usage") as batch_op:
+            batch_op.alter_column(
+                "total_tokens",
+                new_column_name="input_output_total_tokens",
+                existing_type=sa.Integer(),
+                existing_nullable=True,
+            )
+
+    with op.batch_alter_table("usage") as batch_op:
+        if not column_exists("usage", "input_context_tokens"):
+            batch_op.add_column(sa.Column("input_context_tokens", sa.Integer(), nullable=False, server_default="0"))
+        if not column_exists("usage", "total_tokens"):
+            batch_op.add_column(sa.Column("total_tokens", sa.Integer(), nullable=False, server_default="0"))
+        if not column_exists("usage", "cache_creation_cost_usd_5m"):
+            batch_op.add_column(sa.Column("cache_creation_cost_usd_5m", sa.Numeric(20, 8), nullable=False, server_default="0"))
+        if not column_exists("usage", "cache_creation_cost_usd_1h"):
+            batch_op.add_column(sa.Column("cache_creation_cost_usd_1h", sa.Numeric(20, 8), nullable=False, server_default="0"))
+        if not column_exists("usage", "actual_cache_creation_cost_usd_5m"):
+            batch_op.add_column(sa.Column("actual_cache_creation_cost_usd_5m", sa.Numeric(20, 8), nullable=False, server_default="0"))
+        if not column_exists("usage", "actual_cache_creation_cost_usd_1h"):
+            batch_op.add_column(sa.Column("actual_cache_creation_cost_usd_1h", sa.Numeric(20, 8), nullable=False, server_default="0"))
+        if not column_exists("usage", "actual_cache_cost_usd"):
+            batch_op.add_column(sa.Column("actual_cache_cost_usd", sa.Numeric(20, 8), nullable=False, server_default="0"))
+        if not column_exists("usage", "cache_creation_price_per_1m_5m"):
+            batch_op.add_column(sa.Column("cache_creation_price_per_1m_5m", sa.Numeric(20, 8), nullable=True))
+        if not column_exists("usage", "cache_creation_price_per_1m_1h"):
+            batch_op.add_column(sa.Column("cache_creation_price_per_1m_1h", sa.Numeric(20, 8), nullable=True))
+
+    run_backfill_in_batches(_UPGRADE_BACKFILL_SQL)
+
+
+def downgrade() -> None:
+    run_backfill_in_batches(_DOWNGRADE_BACKFILL_SQL)
+
+    with op.batch_alter_table("usage") as batch_op:
+        if column_exists("usage", "cache_creation_price_per_1m_1h"):
+            batch_op.drop_column("cache_creation_price_per_1m_1h")
+        if column_exists("usage", "cache_creation_price_per_1m_5m"):
+            batch_op.drop_column("cache_creation_price_per_1m_5m")
+        if column_exists("usage", "actual_cache_cost_usd"):
+            batch_op.drop_column("actual_cache_cost_usd")
+        if column_exists("usage", "actual_cache_creation_cost_usd_1h"):
+            batch_op.drop_column("actual_cache_creation_cost_usd_1h")
+        if column_exists("usage", "actual_cache_creation_cost_usd_5m"):
+            batch_op.drop_column("actual_cache_creation_cost_usd_5m")
+        if column_exists("usage", "cache_creation_cost_usd_1h"):
+            batch_op.drop_column("cache_creation_cost_usd_1h")
+        if column_exists("usage", "cache_creation_cost_usd_5m"):
+            batch_op.drop_column("cache_creation_cost_usd_5m")
+        if column_exists("usage", "input_context_tokens"):
+            batch_op.drop_column("input_context_tokens")
+        if column_exists("usage", "total_tokens"):
+            batch_op.drop_column("total_tokens")
+        if column_exists("usage", "input_output_total_tokens"):
+            batch_op.alter_column(
+                "input_output_total_tokens",
+                new_column_name="total_tokens",
+                existing_type=sa.Integer(),
+                existing_nullable=True,
+            )
@@ -5,6 +5,7 @@
 #   部署/更新:     ./deploy.sh                    (自动检测所有变化)
 #   指定 Hub 版本: ./deploy.sh --hub-tag hub-v0.1.0
 #   更新 Hub:      ./deploy.sh --update-hub
+#   GitHub 镜像:   ./deploy.sh --mirror https://ghfast.top
 #   强制重建:      ./deploy.sh --rebuild-base
 #   强制全部重建:  ./deploy.sh --force

@@ -44,6 +45,7 @@ Usage: ./deploy.sh [options]
 Options:
  --hub-tag <hub-vX.Y.Z>  指定 Hub Release tag（例如 hub-v0.1.0）
  --update-hub            强制刷新 Hub 版本标记（下次构建会重新下载）
+  --mirror <url>          GitHub 下载镜像（例如 https://ghfast.top）
  --rebuild-base, -r      仅重建 base 镜像
  --force, -f             强制重建全部（hub/base/app）并重启
  -h, --help              显示帮助
@@ -54,6 +56,7 @@ FORCE_REBUILD_ALL=false
 REBUILD_BASE_ONLY=false
 FORCE_UPDATE_HUB=false
 HUB_TAG="${HUB_TAG:-}"
+GITHUB_MIRROR="${GITHUB_MIRROR:-}"
 RESOLVED_HUB_TAG=""

 while [ $# -gt 0 ]; do
@@ -70,6 +73,14 @@ while [ $# -gt 0 ]; do
            FORCE_UPDATE_HUB=true
            shift
            ;;
+        --mirror)
+            if [ $# -lt 2 ]; then
+                echo "ERROR: --mirror needs a URL, e.g. https://ghfast.top"
+                exit 1
+            fi
+            GITHUB_MIRROR="$2"
+            shift 2
+            ;;
        --rebuild-base|-r)
            REBUILD_BASE_ONLY=true
            shift
@@ -97,27 +108,26 @@ if [ -n "$HUB_TAG" ]; then
    esac
 fi

-# 提取 pyproject.toml 中"会影响运行时依赖安装"的最小指纹（与 CI 保持一致）：
-# - [build-system] requires / build-backend
-# - [project] requires-python / dependencies
-# 使用 Python tomllib 解析，不受 TOML 格式变化影响。
+# 提取 pyproject.toml 中会影响运行时依赖安装的字段指纹（纯 shell，无需 Python）
+# 用 sed 提取 dependencies / requires 数组块和单值字段，排序后输出稳定文本
 pyproject_deps_fingerprint() {
-    python3 - <<'PY'
-import json, pathlib, tomllib
-
-data = tomllib.loads(pathlib.Path("pyproject.toml").read_text("utf-8"))
-project = data.get("project") or {}
-build = data.get("build-system") or {}
-
-fingerprint = {
-    "requires-python": project.get("requires-python"),
-    "dependencies": sorted(project.get("dependencies") or []),
-    "build-backend": build.get("build-backend"),
-    "build-requires": sorted(build.get("requires") or []),
-}
-
-print(json.dumps(fingerprint, sort_keys=True, separators=(",", ":")))
-PY
+    local file="pyproject.toml"
+    # 提取 "key = [..." 多行数组块（从 key 行到 ] 行）
+    extract_array() {
+        sed -n "/^$1[[:space:]]*=[[:space:]]*\[/,/\]/p" "$file" | grep '"' | sed 's/.*"\(.*\)".*/\1/' | sort
+    }
+    # 提取 "key = "value"" 单行值
+    extract_value() {
+        grep -m1 "^$1[[:space:]]*=" "$file" 2>/dev/null | sed 's/.*"\(.*\)".*/\1/'
+    }
+    {
+        echo "requires-python=$(extract_value requires-python)"
+        echo "build-backend=$(extract_value build-backend)"
+        echo "dependencies:"
+        extract_array dependencies
+        echo "build-requires:"
+        extract_array requires
+    }
 }

 # 计算依赖文件的哈希值（包含 Dockerfile.base.local）
@@ -286,10 +296,15 @@ build_app() {
    if [ -n "${GITHUB_TOKEN:-}" ]; then
        token_args=(--build-arg "GITHUB_TOKEN=${GITHUB_TOKEN}")
    fi
+    local mirror_args=()
+    if [ -n "${GITHUB_MIRROR:-}" ]; then
+        mirror_args=(--build-arg "GITHUB_MIRROR=${GITHUB_MIRROR}")
+    fi
    docker build --pull=false \
        --build-arg HUB_RELEASE_REPO="$GITHUB_REPO" \
        --build-arg HUB_TAG="$RESOLVED_HUB_TAG" \
        "${token_args[@]}" \
+        "${mirror_args[@]}" \
        -f Dockerfile.app.local \
        -t aether-app:latest .
    save_code_hash
@@ -372,10 +387,6 @@ if [ "$FORCE_UPDATE_HUB" = true ]; then
    exit 0
 fi

-# 拉取最新代码
-echo ">>> Pulling latest code..."
-git pull
-
 # 标记是否需要重启
 NEED_RESTART=false
 BASE_REBUILT=false
@@ -11,10 +11,16 @@ set +a
 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..."
-echo "📍 后端地址: http://localhost:8084"
-echo "📊 数据库: ${DATABASE_URL}"
+# 开发环境连接池低配（节省内存）
+export DB_POOL_SIZE=${DB_POOL_SIZE:-5}
+export DB_MAX_OVERFLOW=${DB_MAX_OVERFLOW:-5}
+export HTTP_MAX_CONNECTIONS=${HTTP_MAX_CONNECTIONS:-20}
+export HTTP_KEEPALIVE_CONNECTIONS=${HTTP_KEEPALIVE_CONNECTIONS:-5}
+
+# 启动 uvicorn（热重载模式，只监视 src 目录）
+echo "=> 启动本地开发服务器..."
+echo "=> 后端地址: http://localhost:8084"
+echo "=> 数据库: ${DATABASE_URL}"
 echo ""

-uv run uvicorn src.main:app --reload --port 8084
+uv run uvicorn src.main:app --reload --reload-dir src --port 8084
@@ -1,6 +1,26 @@
 #!/bin/bash
 set -e

+# Wait for PostgreSQL to be ready
+MAX_ATTEMPTS=30
+ATTEMPT=0
+until python -c "
+from sqlalchemy import create_engine, text
+import os
+engine = create_engine(os.environ['DATABASE_URL'])
+with engine.connect() as conn:
+    conn.execute(text('SELECT 1'))
+" 2>/dev/null; do
+    ATTEMPT=$((ATTEMPT + 1))
+    if [ "$ATTEMPT" -ge "$MAX_ATTEMPTS" ]; then
+        echo "Database not ready after $MAX_ATTEMPTS attempts, exiting."
+        exit 1
+    fi
+    echo "Waiting for database... (attempt $ATTEMPT/$MAX_ATTEMPTS)"
+    sleep 2
+done
+echo "Database is ready."
+
 echo "Running database migrations..."
 alembic upgrade head

@@ -24,6 +24,7 @@
        "marked": "^16.0.0",
        "otpauth": "^9.5.0",
        "pinia": "^3.0.3",
+        "pinyin-pro": "^3.28.0",
        "radix-vue": "^1.9.17",
        "tailwind-merge": "^3.3.1",
        "three": "^0.180.0",
@@ -4718,6 +4719,12 @@
        }
      }
    },
+    "node_modules/pinyin-pro": {
+      "version": "3.28.0",
+      "resolved": "https://registry.npmmirror.com/pinyin-pro/-/pinyin-pro-3.28.0.tgz",
+      "integrity": "sha512-mMRty6RisoyYNphJrTo3pnvp3w8OMZBrXm9YSWkxhAfxKj1KZk2y8T2PDIZlDDRsvZ0No+Hz6FI4sZpA6Ey25g==",
+      "license": "MIT"
+    },
    "node_modules/pirates": {
      "version": "4.0.7",
      "resolved": "https://registry.npmjs.org/pirates/-/pirates-4.0.7.tgz",
@@ -32,6 +32,7 @@
    "marked": "^16.0.0",
    "otpauth": "^9.5.0",
    "pinia": "^3.0.3",
+    "pinyin-pro": "^3.28.0",
    "radix-vue": "^1.9.17",
    "tailwind-merge": "^3.3.1",
    "three": "^0.180.0",
@@ -5,12 +5,14 @@
 </template>

 <script setup lang="ts">
-import { onMounted, onErrorCaptured } from 'vue'
+import { onMounted, onErrorCaptured, onUnmounted } from 'vue'
 import { useAuthStore } from '@/stores/auth'
 import ToastContainer from '@/components/ToastContainer.vue'
 import ConfirmContainer from '@/components/ConfirmContainer.vue'
-import apiClient from '@/api/client'
+import apiClient, { AUTH_STATE_CHANGE_EVENT } from '@/api/client'
 import { NETWORK_CONFIG, AUTH_CONFIG } from '@/config/constants'
+import router from '@/router'
+import { hasAuthIdentityChanged } from '@/utils/authToken'
 import { log } from '@/utils/logger'

 const authStore = useAuthStore()
@@ -86,7 +88,59 @@ if (typeof window !== 'undefined') {
  })
 }

+async function syncExternalAuthState(nextToken: string | null): Promise<void> {
+  const previousToken = authStore.token
+  const previousUser = authStore.user
+    ? {
+        id: authStore.user.id,
+        role: authStore.user.role,
+      }
+    : null
+
+  authStore.syncToken()
+
+  if (!nextToken) {
+    if (previousToken || previousUser) {
+      authStore.applyExternalLogout()
+      await router.replace('/')
+    }
+    return
+  }
+
+  const identityChanged = hasAuthIdentityChanged(previousToken, nextToken, previousUser)
+  if (!identityChanged && previousUser) {
+    return
+  }
+
+  const user = await authStore.fetchCurrentUser()
+  if (!user) {
+    return
+  }
+
+  if (router.currentRoute.value.path.startsWith('/admin') && user.role !== 'admin') {
+    await router.replace('/dashboard')
+  }
+}
+
+function handleAuthStorageChange(event: StorageEvent): void {
+  if (event.key !== 'access_token') {
+    return
+  }
+
+  syncExternalAuthState(event.newValue).catch((err) => log.error('syncExternalAuthState failed', err))
+}
+
+function handleLocalAuthStateChange(event: Event): void {
+  const authEvent = event as CustomEvent<{ token: string | null }>
+  syncExternalAuthState(authEvent.detail?.token ?? apiClient.getToken()).catch((err) => log.error('syncExternalAuthState failed', err))
+}
+
 onMounted(async () => {
+  if (typeof window !== 'undefined') {
+    window.addEventListener('storage', handleAuthStorageChange)
+    window.addEventListener(AUTH_STATE_CHANGE_EVENT, handleLocalAuthStateChange as (event: Event) => void)
+  }
+
  // 延迟检查认证状态,让页面先加载
  setTimeout(async () => {
    try {
@@ -97,4 +151,14 @@ onMounted(async () => {
    }
  }, AUTH_CONFIG.TOKEN_REFRESH_INTERVAL)
 })
+
+onUnmounted(() => {
+  if (typeof window !== 'undefined') {
+    window.removeEventListener('storage', handleAuthStorageChange)
+    window.removeEventListener(
+      AUTH_STATE_CHANGE_EVENT,
+      handleLocalAuthStateChange as (event: Event) => void,
+    )
+  }
+})
 </script>
@@ -0,0 +1,31 @@
+import { afterEach, beforeEach, describe, expect, it, vi } from 'vitest'
+
+import apiClient, { AUTH_STATE_CHANGE_EVENT } from '@/api/client'
+
+describe('apiClient auth state change event', () => {
+  beforeEach(() => {
+    localStorage.clear()
+    apiClient.clearAuth()
+  })
+
+  afterEach(() => {
+    localStorage.clear()
+    apiClient.clearAuth()
+  })
+
+  it('dispatches a same-tab auth change event when clearing auth', () => {
+    const handler = vi.fn()
+    window.addEventListener(AUTH_STATE_CHANGE_EVENT, handler as EventListener)
+
+    apiClient.setToken('access-token')
+    apiClient.clearAuth()
+
+    expect(localStorage.getItem('access_token')).toBeNull()
+    expect(handler).toHaveBeenCalledTimes(1)
+
+    const event = handler.mock.calls[0][0] as CustomEvent<{ token: string | null }>
+    expect(event.detail).toEqual({ token: null })
+
+    window.removeEventListener(AUTH_STATE_CHANGE_EVENT, handler as EventListener)
+  })
+})
@@ -86,6 +86,7 @@ export interface UserExport {
  allowed_providers?: string[] | null
  allowed_api_formats?: string[] | null
  allowed_models?: string[] | null
+  rate_limit?: number | null  // null = 跟随系统默认，0 = 不限制
  model_capability_settings?: Record<string, Record<string, boolean>>
  unlimited?: boolean
  wallet?: BillingSummary | null
@@ -102,7 +103,7 @@ export interface UserApiKeyExport {
  allowed_providers?: string[] | null
  allowed_api_formats?: string[] | null
  allowed_models?: string[] | null
-  rate_limit?: number | null  // null = 无限制
+  rate_limit?: number | null  // legacy/null 兼容；1.3+ standalone null = 跟随系统默认
  concurrent_limit?: number | null
  force_capabilities?: Record<string, boolean>
  is_active: boolean
@@ -349,7 +350,7 @@ export interface AdminApiKey {
  total_requests?: number
  total_tokens?: number
  total_cost_usd?: number
-  rate_limit?: number | null  // null = 无限制
+  rate_limit?: number | null  // null = 跟随系统默认，0 = 不限制
  allowed_providers?: string[] | null  // 允许的提供商列表
  allowed_api_formats?: string[] | null  // 允许的 API 格式列表
  allowed_models?: string[] | null  // 允许的模型列表
@@ -365,7 +366,7 @@ export interface CreateStandaloneApiKeyRequest {
  allowed_providers?: string[] | null
  allowed_api_formats?: string[] | null
  allowed_models?: string[] | null
-  rate_limit?: number | null  // null = 无限制
+  rate_limit?: number | null  // null = 跟随系统默认，0 = 不限制
  expires_at?: string | null  // ISO 日期字符串，如 "2025-12-31"，null = 永不过期
  initial_balance_usd: number | null  // 初始余额，null = 无限制
  unlimited_balance?: boolean | null  // 编辑时仅切换额度模式，不调整余额数值
@@ -379,102 +380,6 @@ export interface AdminApiKeysResponse {
  skip: number
 }

-export interface LeaderboardItem {
-  rank: number
-  id: string
-  name: string
-  value: number
-  requests: number
-  tokens: number
-  cost: number
-}
-
-export interface LeaderboardResponse {
-  items: LeaderboardItem[]
-  total: number
-  metric: string
-  start_date?: string | null
-  end_date?: string | null
-}
-
-export interface CostForecastResponse {
-  history: Array<{ date: string; total_cost: number }>
-  forecast: Array<{ date: string; total_cost: number }>
-  slope: number
-  intercept: number
-  start_date: string
-  end_date: string
-}
-
-export interface CostSavingsResponse {
-  cache_read_tokens: number
-  cache_read_cost: number
-  cache_creation_cost: number
-  estimated_full_cost: number
-  cache_savings: number
-}
-
-export interface QuotaUsageProvider {
-  id: string
-  name: string
-  quota_usd: number
-  used_usd: number
-  remaining_usd: number
-  usage_percent: number
-  quota_expires_at?: string | null
-  estimated_exhaust_at?: string | null
-}
-
-export interface QuotaUsageResponse {
-  providers: QuotaUsageProvider[]
-}
-
-export interface PercentileItem {
-  date: string
-  p50_response_time_ms?: number | null
-  p90_response_time_ms?: number | null
-  p99_response_time_ms?: number | null
-  p50_first_byte_time_ms?: number | null
-  p90_first_byte_time_ms?: number | null
-  p99_first_byte_time_ms?: number | null
-}
-
-export interface ErrorDistributionItem {
-  category: string
-  count: number
-}
-
-export interface ErrorTrendItem {
-  date: string
-  total: number
-  categories: Record<string, number>
-}
-
-export interface ErrorDistributionResponse {
-  distribution: ErrorDistributionItem[]
-  trend: ErrorTrendItem[]
-}
-
-export interface ComparisonMetric {
-  total_requests: number
-  total_tokens: number
-  total_cost: number
-  actual_total_cost: number
-  avg_response_time_ms: number
-  error_requests: number
-}
-
-export interface ComparisonResponse {
-  current: ComparisonMetric
-  comparison: ComparisonMetric
-  change_percent: Record<string, number | null>
-  current_start: string
-  current_end: string
-  comparison_start: string
-  comparison_end: string
-}
-
-
 export interface ApiKeyToggleResponse {
  id: string // UUID
  is_active: boolean
@@ -752,197 +657,6 @@ export const adminApi = {
    return response.data
  },

-  // Stats / Leaderboards
-  async getLeaderboardUsers(params?: {
-    start_date?: string
-    end_date?: string
-    preset?: string
-    timezone?: string
-    tz_offset_minutes?: number
-    metric?: 'requests' | 'tokens' | 'cost'
-    order?: 'asc' | 'desc'
-    limit?: number
-    offset?: number
-    provider_name?: string
-    model?: string
-    include_inactive?: boolean
-    exclude_admin?: boolean
-  }): Promise<LeaderboardResponse> {
-    const cacheKey = buildCacheKey('admin:stats:leaderboard:users', params)
-    return cachedRequest(
-      cacheKey,
-      async () => {
-        const response = await apiClient.get<LeaderboardResponse>('/api/admin/stats/leaderboard/users', {
-          params
-        })
-        return response.data
-      },
-      20 * 1000
-    )
-  },
-
-  async getLeaderboardApiKeys(params?: {
-    start_date?: string
-    end_date?: string
-    preset?: string
-    timezone?: string
-    tz_offset_minutes?: number
-    metric?: 'requests' | 'tokens' | 'cost'
-    order?: 'asc' | 'desc'
-    limit?: number
-    offset?: number
-    provider_name?: string
-    model?: string
-    include_inactive?: boolean
-    exclude_admin?: boolean
-  }): Promise<LeaderboardResponse> {
-    const cacheKey = buildCacheKey('admin:stats:leaderboard:api-keys', params)
-    return cachedRequest(
-      cacheKey,
-      async () => {
-        const response = await apiClient.get<LeaderboardResponse>('/api/admin/stats/leaderboard/api-keys', {
-          params
-        })
-        return response.data
-      },
-      20 * 1000
-    )
-  },
-
-  async getLeaderboardModels(params?: {
-    start_date?: string
-    end_date?: string
-    preset?: string
-    timezone?: string
-    tz_offset_minutes?: number
-    metric?: 'requests' | 'tokens' | 'cost'
-    order?: 'asc' | 'desc'
-    limit?: number
-    offset?: number
-    provider_name?: string
-    model?: string
-  }): Promise<LeaderboardResponse> {
-    const cacheKey = buildCacheKey('admin:stats:leaderboard:models', params)
-    return cachedRequest(
-      cacheKey,
-      async () => {
-        const response = await apiClient.get<LeaderboardResponse>('/api/admin/stats/leaderboard/models', {
-          params
-        })
-        return response.data
-      },
-      20 * 1000
-    )
-  },
-
-  async getCostForecast(params?: {
-    start_date?: string
-    end_date?: string
-    preset?: string
-    timezone?: string
-    tz_offset_minutes?: number
-    days?: number
-    forecast_days?: number
-  }): Promise<CostForecastResponse> {
-    const cacheKey = buildCacheKey('admin:stats:cost:forecast', params)
-    return cachedRequest(
-      cacheKey,
-      async () => {
-        const response = await apiClient.get<CostForecastResponse>('/api/admin/stats/cost/forecast', {
-          params
-        })
-        return response.data
-      },
-      30 * 1000
-    )
-  },
-
-  async getCostSavings(params?: {
-    start_date?: string
-    end_date?: string
-    preset?: string
-    timezone?: string
-    tz_offset_minutes?: number
-    provider_name?: string
-    model?: string
-  }): Promise<CostSavingsResponse> {
-    const cacheKey = buildCacheKey('admin:stats:cost:savings', params)
-    return cachedRequest(
-      cacheKey,
-      async () => {
-        const response = await apiClient.get<CostSavingsResponse>('/api/admin/stats/cost/savings', {
-          params
-        })
-        return response.data
-      },
-      30 * 1000
-    )
-  },
-
-  async getQuotaUsage(): Promise<QuotaUsageResponse> {
-    return cachedRequest(
-      'admin:stats:providers:quota-usage',
-      async () => {
-        const response = await apiClient.get<QuotaUsageResponse>('/api/admin/stats/providers/quota-usage')
-        return response.data
-      },
-      30 * 1000
-    )
-  },
-
-  async getPercentiles(params?: {
-    start_date?: string
-    end_date?: string
-    preset?: string
-    timezone?: string
-    tz_offset_minutes?: number
-  }): Promise<PercentileItem[]> {
-    const cacheKey = buildCacheKey('admin:stats:performance:percentiles', params)
-    return cachedRequest(
-      cacheKey,
-      async () => {
-        const response = await apiClient.get<PercentileItem[]>('/api/admin/stats/performance/percentiles', {
-          params
-        })
-        return response.data
-      },
-      20 * 1000
-    )
-  },
-
-  async getErrorDistribution(params?: {
-    start_date?: string
-    end_date?: string
-    preset?: string
-    timezone?: string
-    tz_offset_minutes?: number
-  }): Promise<ErrorDistributionResponse> {
-    const cacheKey = buildCacheKey('admin:stats:errors:distribution', params)
-    return cachedRequest(
-      cacheKey,
-      async () => {
-        const response = await apiClient.get<ErrorDistributionResponse>('/api/admin/stats/errors/distribution', {
-          params
-        })
-        return response.data
-      },
-      20 * 1000
-    )
-  },
-
-  async getComparison(params: {
-    current_start: string
-    current_end: string
-    comparison_type?: 'period' | 'year'
-    timezone?: string
-    tz_offset_minutes?: number
-  }): Promise<ComparisonResponse> {
-    const response = await apiClient.get<ComparisonResponse>('/api/admin/stats/comparison', {
-      params
-    })
-    return response.data
-  },
-
  // 数据清空
  purgeConfig: () => purge<{ message: string; deleted: Record<string, number> }>('config'),
  purgeUsers: () => purge<{ message: string; deleted: Record<string, number> }>('users'),
@@ -951,26 +665,4 @@ export const adminApi = {
  purgeRequestBodies: () => purge<{ message: string; cleaned: Record<string, number> }>('request-bodies'),
  purgeStats: () => purge<{ message: string }>('stats'),

-  async getTimeSeries(params?: {
-    start_date?: string
-    end_date?: string
-    preset?: string
-    granularity?: 'hour' | 'day' | 'week' | 'month'
-    timezone?: string
-    tz_offset_minutes?: number
-    user_id?: string
-    model?: string
-    provider_name?: string
-  }): Promise<Array<Record<string, unknown>>> {
-    const cacheKey = buildCacheKey('admin:stats:time-series', params)
-    return cachedRequest(
-      cacheKey,
-      async () => {
-        const response = await apiClient.get<Array<Record<string, unknown>>>('/api/admin/stats/time-series', { params })
-        return response.data
-      },
-      20 * 1000
-    )
-  },
-
 }
@@ -0,0 +1,396 @@
+import apiClient from './client'
+import { cachedRequest, buildCacheKey } from '@/utils/cache'
+import type { ActivityHeatmap } from '@/types/activity'
+
+export type AnalyticsScopeKind = 'global' | 'me' | 'user' | 'api_key'
+export type AnalyticsGranularity = 'hour' | 'day' | 'week' | 'month'
+export type AnalyticsBreakdownDimension = 'model' | 'provider' | 'api_format' | 'api_key' | 'user'
+export type AnalyticsBreakdownMetric = 'requests_total' | 'total_tokens' | 'total_cost_usd' | 'actual_total_cost_usd'
+export type AnalyticsLeaderboardEntity = 'user' | 'api_key'
+export type AnalyticsLeaderboardMetric = 'requests_total' | 'total_tokens' | 'total_cost_usd'
+
+export interface AnalyticsScope {
+  kind: AnalyticsScopeKind
+  user_id?: string | null
+  api_key_id?: string | null
+}
+
+export interface AnalyticsTimeRange {
+  start_date?: string
+  end_date?: string
+  preset?: string
+  granularity?: AnalyticsGranularity
+  timezone?: string | null
+  tz_offset_minutes?: number
+}
+
+export interface AnalyticsFilters {
+  user_ids?: string[]
+  provider_names?: string[]
+  models?: string[]
+  target_models?: string[]
+  api_key_ids?: string[]
+  api_formats?: string[]
+  request_types?: string[]
+  statuses?: string[]
+  error_categories?: string[]
+  is_stream?: boolean | null
+  has_format_conversion?: boolean | null
+}
+
+export interface AnalyticsBaseRequest {
+  scope: AnalyticsScope
+  time_range: AnalyticsTimeRange
+  filters?: AnalyticsFilters
+}
+
+export interface AnalyticsSummary {
+  requests_total: number
+  requests_success: number
+  requests_error: number
+  requests_stream: number
+  success_rate: number
+  input_tokens: number
+  output_tokens: number
+  input_output_total_tokens: number
+  cache_creation_input_tokens: number
+  cache_creation_input_tokens_5m: number
+  cache_creation_input_tokens_1h: number
+  cache_read_input_tokens: number
+  input_context_tokens: number
+  total_tokens: number
+  cache_hit_rate: number
+  input_cost_usd: number
+  output_cost_usd: number
+  cache_creation_cost_usd: number
+  cache_creation_cost_usd_5m: number
+  cache_creation_cost_usd_1h: number
+  cache_read_cost_usd: number
+  cache_cost_usd: number
+  request_cost_usd: number
+  total_cost_usd: number
+  actual_total_cost_usd: number
+  actual_cache_cost_usd: number
+  avg_response_time_ms: number
+  avg_first_byte_time_ms: number
+  format_conversion_count: number
+  models_used_count: number
+}
+
+export interface AnalyticsCompositionSegment {
+  key: string
+  value: number
+  percentage: number
+}
+
+export interface AnalyticsOverviewResponse {
+  query_context: {
+    scope: AnalyticsScope
+    time_range: AnalyticsTimeRange
+  }
+  summary: AnalyticsSummary
+  composition: {
+    token_segments: AnalyticsCompositionSegment[]
+    cost_segments: AnalyticsCompositionSegment[]
+  }
+}
+
+export interface AnalyticsTimeseriesBucket extends AnalyticsSummary {
+  bucket_start: string
+  bucket_end: string
+}
+
+export interface AnalyticsTimeseriesResponse {
+  buckets: AnalyticsTimeseriesBucket[]
+}
+
+export interface AnalyticsBreakdownRequest extends AnalyticsBaseRequest {
+  dimension: AnalyticsBreakdownDimension
+  metric?: AnalyticsBreakdownMetric
+  limit?: number
+}
+
+export interface AnalyticsBreakdownRow extends AnalyticsSummary {
+  key: string
+  label: string
+  share_of_total_cost: number
+  share_of_total_tokens: number
+}
+
+export interface AnalyticsBreakdownResponse {
+  dimension: AnalyticsBreakdownDimension
+  metric: AnalyticsBreakdownMetric
+  rows: AnalyticsBreakdownRow[]
+}
+
+export interface AnalyticsRecordsRequest extends AnalyticsBaseRequest {
+  search?: {
+    text?: string | null
+    request_id?: string | null
+  }
+  pagination?: {
+    limit?: number
+    offset?: number
+  }
+}
+
+export interface AnalyticsRecord {
+  id: string
+  request_id: string
+  created_at: string | null
+  user_id: string | null
+  username: string | null
+  api_key_id: string | null
+  api_key_name: string | null
+  provider_api_key_name: string | null
+  provider_name: string | null
+  model: string
+  target_model: string | null
+  api_format: string | null
+  request_type: string | null
+  status: string
+  billing_status: string
+  is_stream: boolean
+  has_format_conversion: boolean | null
+  has_fallback?: boolean
+  has_retry?: boolean
+  status_code: number | null
+  error_message: string | null
+  error_category: string | null
+  response_time_ms: number | null
+  first_byte_time_ms: number | null
+  input_tokens: number
+  output_tokens: number
+  input_output_total_tokens: number
+  cache_creation_input_tokens: number
+  cache_creation_input_tokens_5m: number
+  cache_creation_input_tokens_1h: number
+  cache_read_input_tokens: number
+  input_context_tokens: number
+  total_tokens: number
+  input_cost_usd: number
+  output_cost_usd: number
+  cache_creation_cost_usd: number
+  cache_creation_cost_usd_5m: number
+  cache_creation_cost_usd_1h: number
+  cache_read_cost_usd: number
+  cache_cost_usd: number
+  request_cost_usd: number
+  total_cost_usd: number
+  actual_total_cost_usd: number
+  actual_cache_cost_usd: number
+  rate_multiplier: number
+}
+
+export interface AnalyticsRecordsResponse {
+  total: number
+  limit: number
+  offset: number
+  records: AnalyticsRecord[]
+}
+
+export interface AnalyticsFilterOption {
+  value: string
+  label: string
+}
+
+export interface AnalyticsFilterOptionsResponse {
+  providers: AnalyticsFilterOption[]
+  models: AnalyticsFilterOption[]
+  target_models: AnalyticsFilterOption[]
+  api_formats: AnalyticsFilterOption[]
+  request_types: AnalyticsFilterOption[]
+  error_categories: AnalyticsFilterOption[]
+  statuses: AnalyticsFilterOption[]
+  users?: AnalyticsFilterOption[]
+  api_keys?: AnalyticsFilterOption[]
+}
+
+export interface AnalyticsLeaderboardRequest extends AnalyticsBaseRequest {
+  entity: AnalyticsLeaderboardEntity
+  metric: AnalyticsLeaderboardMetric
+  limit?: number
+}
+
+export interface AnalyticsLeaderboardItem {
+  rank: number
+  id: string
+  label: string
+  requests_total: number
+  total_tokens: number
+  total_cost_usd: number
+  actual_total_cost_usd: number
+  metric_value: number
+}
+
+export interface AnalyticsLeaderboardResponse {
+  entity: AnalyticsLeaderboardEntity
+  metric: AnalyticsLeaderboardMetric
+  items: AnalyticsLeaderboardItem[]
+}
+
+export interface AnalyticsPercentilePoint {
+  date: string
+  p50_response_time_ms: number | null
+  p90_response_time_ms: number | null
+  p99_response_time_ms: number | null
+  p50_first_byte_time_ms: number | null
+  p90_first_byte_time_ms: number | null
+  p99_first_byte_time_ms: number | null
+}
+
+export interface AnalyticsErrorCategory {
+  category: string
+  label: string
+  count: number
+}
+
+export interface AnalyticsErrorTrendItem {
+  date: string
+  total: number
+}
+
+export interface AnalyticsProviderHealthItem {
+  provider_name: string
+  requests_total: number
+  success_rate: number
+  error_rate: number
+  avg_response_time_ms: number
+  avg_first_byte_time_ms: number
+}
+
+export interface AnalyticsPerformanceResponse {
+  latency: {
+    response_time_ms: {
+      avg: number
+      p50: number | null
+      p90: number | null
+      p99: number | null
+    }
+    first_byte_time_ms: {
+      avg: number
+      p50: number | null
+      p90: number | null
+      p99: number | null
+    }
+  }
+  percentiles: AnalyticsPercentilePoint[]
+  errors: {
+    total: number
+    rate: number
+    categories: AnalyticsErrorCategory[]
+    trend: AnalyticsErrorTrendItem[]
+  }
+  provider_health: AnalyticsProviderHealthItem[]
+}
+
+export interface AnalyticsActiveRequest {
+  id: string
+  status: 'pending' | 'streaming' | 'completed' | 'failed' | 'cancelled'
+  input_tokens: number
+  output_tokens: number
+  cache_creation_input_tokens?: number | null
+  cache_read_input_tokens?: number | null
+  cost?: number
+  total_cost_usd?: number
+  actual_cost?: number | null
+  actual_total_cost_usd?: number | null
+  rate_multiplier?: number | null
+  response_time_ms: number | null
+  first_byte_time_ms: number | null
+  provider?: string | null
+  provider_name?: string | null
+  api_key_name?: string | null
+  api_format?: string | null
+  endpoint_api_format?: string | null
+  has_format_conversion?: boolean | null
+  target_model?: string | null
+}
+
+export interface AnalyticsActiveRequestsResponse {
+  requests: AnalyticsActiveRequest[]
+}
+
+export const analyticsApi = {
+  async getOverview(payload: AnalyticsBaseRequest): Promise<AnalyticsOverviewResponse> {
+    const response = await apiClient.post<AnalyticsOverviewResponse>('/api/analytics/overview', payload)
+    return response.data
+  },
+
+  async getTimeseries(payload: AnalyticsBaseRequest): Promise<AnalyticsTimeseriesResponse> {
+    const response = await apiClient.post<AnalyticsTimeseriesResponse>('/api/analytics/timeseries', payload)
+    return response.data
+  },
+
+  async getBreakdown(payload: AnalyticsBreakdownRequest): Promise<AnalyticsBreakdownResponse> {
+    const response = await apiClient.post<AnalyticsBreakdownResponse>('/api/analytics/breakdown', payload)
+    return response.data
+  },
+
+  async getRecords(payload: AnalyticsRecordsRequest): Promise<AnalyticsRecordsResponse> {
+    const response = await apiClient.post<AnalyticsRecordsResponse>('/api/analytics/records', payload)
+    return response.data
+  },
+
+  async getFilterOptions(payload: AnalyticsBaseRequest): Promise<AnalyticsFilterOptionsResponse> {
+    const response = await apiClient.post<AnalyticsFilterOptionsResponse>('/api/analytics/filter-options', payload)
+    return response.data
+  },
+
+  async getLeaderboard(payload: AnalyticsLeaderboardRequest): Promise<AnalyticsLeaderboardResponse> {
+    const response = await apiClient.post<AnalyticsLeaderboardResponse>('/api/analytics/leaderboard', payload)
+    return response.data
+  },
+
+  async getPerformance(payload: AnalyticsBaseRequest): Promise<AnalyticsPerformanceResponse> {
+    const response = await apiClient.post<AnalyticsPerformanceResponse>('/api/analytics/performance', payload)
+    return response.data
+  },
+
+  async getHeatmap(payload: {
+    scope: AnalyticsScope
+    user_id?: string | null
+    api_key_id?: string | null
+  }): Promise<ActivityHeatmap> {
+    const cacheKey = buildCacheKey('analytics:heatmap', payload as Record<string, unknown>)
+    return cachedRequest(
+      cacheKey,
+      async () => {
+        const response = await apiClient.post<ActivityHeatmap>('/api/analytics/heatmap', payload)
+        return response.data
+      },
+      60000,
+    )
+  },
+
+  async getActiveRequests(payload: {
+    scope: AnalyticsScope
+    ids?: string[]
+  }): Promise<AnalyticsActiveRequestsResponse> {
+    const response = await apiClient.post<AnalyticsActiveRequestsResponse>('/api/analytics/active-requests', {
+      ...payload,
+      ids: payload.ids ?? [],
+    })
+    return response.data
+  },
+
+  async analyzeCacheAffinityTTL(payload: {
+    scope: AnalyticsScope
+    user_id?: string | null
+    api_key_id?: string | null
+    hours?: number
+  }): Promise<Record<string, unknown>> {
+    const response = await apiClient.post('/api/analytics/cache-affinity/ttl-analysis', payload)
+    return response.data
+  },
+
+  async analyzeCacheAffinityHit(payload: {
+    scope: AnalyticsScope
+    user_id?: string | null
+    api_key_id?: string | null
+    hours?: number
+  }): Promise<Record<string, unknown>> {
+    const response = await apiClient.post('/api/analytics/cache-affinity/hit-analysis', payload)
+    return response.data
+  },
+}
@@ -33,6 +33,203 @@ export interface AuditFilters {
  offset?: number
 }

+export type MonitoringMetricStatus = 'ok' | 'warning' | 'danger' | 'degraded' | 'error' | 'unknown'
+
+export interface MonitoringCpuMetric {
+  status: MonitoringMetricStatus
+  label: string
+  usage_percent: number | null
+  load_percent: number | null
+  core_count: number
+  message?: string | null
+}
+
+export interface MonitoringMemoryMetric {
+  status: MonitoringMetricStatus
+  label: string
+  used_percent: number | null
+  used_bytes: number | null
+  available_bytes: number | null
+  total_bytes: number | null
+  message?: string | null
+}
+
+export interface MonitoringServiceMetric {
+  status: MonitoringMetricStatus
+  label: string
+  latency_ms: number | null
+  memory_status?: MonitoringMetricStatus
+  memory_label?: string | null
+  used_memory_bytes?: number | null
+  peak_memory_bytes?: number | null
+  maxmemory_bytes?: number | null
+  memory_ceiling_bytes?: number | null
+  memory_source?: 'configured' | 'maxmemory' | 'system' | 'unknown'
+  available_memory_bytes?: number | null
+  memory_percent?: number | null
+  message?: string | null
+}
+
+export interface MonitoringPostgresMetric {
+  status: MonitoringMetricStatus
+  label: string
+  usage_percent: number | null
+  pool_usage_percent?: number | null
+  checked_out: number
+  pool_size: number
+  overflow: number
+  max_capacity: number
+  pool_timeout: number
+  server_connections?: number | null
+  server_max_connections?: number | null
+  server_usage_percent?: number | null
+  storage_status?: MonitoringMetricStatus
+  storage_label?: string | null
+  storage_total_bytes?: number | null
+  storage_free_bytes?: number | null
+  storage_free_percent?: number | null
+  database_size_bytes?: number | null
+  storage_message?: string | null
+  message?: string | null
+}
+
+export interface MonitoringSystemStatus {
+  timestamp: string
+  users: {
+    total: number
+    active: number
+  }
+  providers: {
+    total: number
+    active: number
+  }
+  api_keys: {
+    total: number
+    active: number
+  }
+  today_stats: {
+    requests: number
+    tokens: number
+    cost_usd: number
+  }
+  recent_errors: number
+  system_metrics?: {
+    cpu: MonitoringCpuMetric
+    memory: MonitoringMemoryMetric
+    redis: MonitoringServiceMetric
+    postgres: MonitoringPostgresMetric
+  }
+}
+
+function parseNumericLike(value: unknown): number {
+  if (typeof value === 'number') {
+    return Number.isFinite(value) ? value : 0
+  }
+  if (typeof value === 'string') {
+    const normalized = value.trim().replace(/[$,\s]/g, '')
+    if (!normalized) return 0
+    const parsed = Number(normalized)
+    return Number.isFinite(parsed) ? parsed : 0
+  }
+  return 0
+}
+
+function parseNullableNumericLike(value: unknown): number | null {
+  if (value == null) {
+    return null
+  }
+  if (typeof value === 'number') {
+    return Number.isFinite(value) ? value : null
+  }
+  if (typeof value === 'string') {
+    const normalized = value.trim().replace(/[$,\s]/g, '')
+    if (!normalized) return null
+    const parsed = Number(normalized)
+    return Number.isFinite(parsed) ? parsed : null
+  }
+  return null
+}
+
+function normalizeMonitoringSystemStatus(data: MonitoringSystemStatus): MonitoringSystemStatus {
+  return {
+    ...data,
+    users: {
+      total: parseNumericLike(data.users?.total),
+      active: parseNumericLike(data.users?.active),
+    },
+    providers: {
+      total: parseNumericLike(data.providers?.total),
+      active: parseNumericLike(data.providers?.active),
+    },
+    api_keys: {
+      total: parseNumericLike(data.api_keys?.total),
+      active: parseNumericLike(data.api_keys?.active),
+    },
+    today_stats: {
+      requests: parseNumericLike(data.today_stats?.requests),
+      tokens: parseNumericLike(data.today_stats?.tokens),
+      cost_usd: parseNumericLike(data.today_stats?.cost_usd),
+    },
+    recent_errors: parseNumericLike(data.recent_errors),
+    system_metrics: data.system_metrics ? {
+      cpu: {
+        status: data.system_metrics.cpu?.status ?? 'unknown',
+        label: data.system_metrics.cpu?.label ?? '未知',
+        usage_percent: parseNullableNumericLike(data.system_metrics.cpu?.usage_percent),
+        load_percent: parseNullableNumericLike(data.system_metrics.cpu?.load_percent),
+        core_count: parseNumericLike(data.system_metrics.cpu?.core_count),
+        message: data.system_metrics.cpu?.message ?? null,
+      },
+      memory: {
+        status: data.system_metrics.memory?.status ?? 'unknown',
+        label: data.system_metrics.memory?.label ?? '未知',
+        used_percent: parseNullableNumericLike(data.system_metrics.memory?.used_percent),
+        used_bytes: parseNullableNumericLike(data.system_metrics.memory?.used_bytes),
+        available_bytes: parseNullableNumericLike(data.system_metrics.memory?.available_bytes),
+        total_bytes: parseNullableNumericLike(data.system_metrics.memory?.total_bytes),
+        message: data.system_metrics.memory?.message ?? null,
+      },
+      redis: {
+        status: data.system_metrics.redis?.status ?? 'unknown',
+        label: data.system_metrics.redis?.label ?? '未知',
+        latency_ms: parseNullableNumericLike(data.system_metrics.redis?.latency_ms),
+        memory_status: data.system_metrics.redis?.memory_status ?? 'unknown',
+        memory_label: data.system_metrics.redis?.memory_label ?? '未知',
+        used_memory_bytes: parseNullableNumericLike(data.system_metrics.redis?.used_memory_bytes),
+        peak_memory_bytes: parseNullableNumericLike(data.system_metrics.redis?.peak_memory_bytes),
+        maxmemory_bytes: parseNullableNumericLike(data.system_metrics.redis?.maxmemory_bytes),
+        memory_ceiling_bytes: parseNullableNumericLike(data.system_metrics.redis?.memory_ceiling_bytes),
+        memory_source: (data.system_metrics.redis?.memory_source as MonitoringServiceMetric['memory_source']) ?? 'unknown',
+        available_memory_bytes: parseNullableNumericLike(data.system_metrics.redis?.available_memory_bytes),
+        memory_percent: parseNullableNumericLike(data.system_metrics.redis?.memory_percent),
+        message: data.system_metrics.redis?.message ?? null,
+      },
+      postgres: {
+        status: data.system_metrics.postgres?.status ?? 'unknown',
+        label: data.system_metrics.postgres?.label ?? '未知',
+        usage_percent: parseNullableNumericLike(data.system_metrics.postgres?.usage_percent),
+        pool_usage_percent: parseNullableNumericLike(data.system_metrics.postgres?.pool_usage_percent),
+        checked_out: parseNumericLike(data.system_metrics.postgres?.checked_out),
+        pool_size: parseNumericLike(data.system_metrics.postgres?.pool_size),
+        overflow: parseNumericLike(data.system_metrics.postgres?.overflow),
+        max_capacity: parseNumericLike(data.system_metrics.postgres?.max_capacity),
+        pool_timeout: parseNumericLike(data.system_metrics.postgres?.pool_timeout),
+        server_connections: parseNullableNumericLike(data.system_metrics.postgres?.server_connections),
+        server_max_connections: parseNullableNumericLike(data.system_metrics.postgres?.server_max_connections),
+        server_usage_percent: parseNullableNumericLike(data.system_metrics.postgres?.server_usage_percent),
+        storage_status: data.system_metrics.postgres?.storage_status ?? 'unknown',
+        storage_label: data.system_metrics.postgres?.storage_label ?? '未知',
+        storage_total_bytes: parseNullableNumericLike(data.system_metrics.postgres?.storage_total_bytes),
+        storage_free_bytes: parseNullableNumericLike(data.system_metrics.postgres?.storage_free_bytes),
+        storage_free_percent: parseNullableNumericLike(data.system_metrics.postgres?.storage_free_percent),
+        database_size_bytes: parseNullableNumericLike(data.system_metrics.postgres?.database_size_bytes),
+        storage_message: data.system_metrics.postgres?.storage_message ?? null,
+        message: data.system_metrics.postgres?.message ?? null,
+      },
+    } : undefined,
+  }
+}
+
 function normalizeAuditResponse(data: Record<string, unknown>): AuditLogsResponse {
  const items: AuditLog[] = (data.items ?? data.logs ?? []) as AuditLog[]
  const meta: PaginationMeta = (data.meta as PaginationMeta) ?? {
@@ -78,6 +275,11 @@ export const auditApi = {
    return response.data
  },

+  async getSystemStatus(): Promise<MonitoringSystemStatus> {
+    const response = await apiClient.get<MonitoringSystemStatus>('/api/admin/monitoring/system-status')
+    return normalizeMonitoringSystemStatus(response.data)
+  },
+
  // 分析用户行为 (管理员)
  async analyzeUserBehavior(userId: number, days: number = 7): Promise<{
    analysis: Record<string, unknown>
@@ -9,7 +9,6 @@ export interface LoginRequest {

 export interface LoginResponse {
  access_token: string
-  refresh_token?: string
  token_type?: string
  expires_in?: number
  user_id?: string // UUID
@@ -25,13 +24,6 @@ export interface UserPreferences {
  [key: string]: unknown // 允许扩展其他偏好设置
 }

-export interface UserStats {
-  total_requests?: number
-  total_cost?: number
-  last_request_at?: string
-  [key: string]: unknown // 允许扩展其他统计数据
-}
-
 export interface SendVerificationCodeRequest {
  email: string
 }
@@ -81,6 +73,7 @@ export interface RegistrationSettingsResponse {
  enable_registration: boolean
  require_email_verification: boolean
  email_configured: boolean
+  password_policy_level: string
 }

 export interface AuthSettingsResponse {
@@ -113,23 +106,15 @@ export interface User {
  role: string  // 'admin' or 'user'
  is_active: boolean
  billing?: BillingSummary
-  allowed_providers?: string[] | null  // 允许使用的提供商 ID 列表
-  allowed_api_formats?: string[] | null  // 允许使用的 API 格式列表
-  allowed_models?: string[] | null  // 允许使用的模型名称列表
  created_at: string
  last_login_at?: string
  preferences?: UserPreferences
-  stats?: UserStats
 }

 export const authApi = {
  async login(credentials: LoginRequest): Promise<LoginResponse> {
    const response = await apiClient.post<LoginResponse>('/api/auth/login', credentials)
    apiClient.setToken(response.data.access_token)
-    // 后端暂时没有返回 refresh_token
-    if (response.data.refresh_token) {
-      localStorage.setItem('refresh_token', response.data.refresh_token)
-    }
    return response.data
  },

@@ -151,14 +136,9 @@ export const authApi = {
    return response.data
  },

-  async refreshToken(refreshToken: string): Promise<LoginResponse> {
-    const response = await apiClient.post<LoginResponse>('/api/auth/refresh', {
-      refresh_token: refreshToken
-    })
+  async refreshToken(): Promise<LoginResponse> {
+    const response = await apiClient.post<LoginResponse>('/api/auth/refresh', {})
    apiClient.setToken(response.data.access_token)
-    if (response.data.refresh_token) {
-      localStorage.setItem('refresh_token', response.data.refresh_token)
-    }
    return response.data
  },

@@ -288,7 +288,12 @@ export const cacheAnalysisApi = {
    api_key_id?: string
    hours?: number
  }): Promise<TTLAnalysisResponse> {
-    const response = await api.get('/api/admin/usage/cache-affinity/ttl-analysis', { params })
+    const response = await api.post('/api/analytics/cache-affinity/ttl-analysis', {
+      scope: { kind: 'global' },
+      user_id: params?.user_id ?? null,
+      api_key_id: params?.api_key_id ?? null,
+      hours: params?.hours,
+    })
    return response.data
  },

@@ -300,7 +305,12 @@ export const cacheAnalysisApi = {
    api_key_id?: string
    hours?: number
  }): Promise<CacheHitAnalysisResponse> {
-    const response = await api.get('/api/admin/usage/cache-affinity/hit-analysis', { params })
+    const response = await api.post('/api/analytics/cache-affinity/hit-analysis', {
+      scope: { kind: 'global' },
+      user_id: params?.user_id ?? null,
+      api_key_id: params?.api_key_id ?? null,
+      hours: params?.hours,
+    })
    return response.data
  },

@@ -319,7 +329,13 @@ export const cacheAnalysisApi = {
    return cachedRequest(
      cacheKey,
      async () => {
-        const response = await api.get('/api/admin/usage/cache-affinity/interval-timeline', { params })
+        const response = await api.post('/api/analytics/interval-timeline', {
+          scope: params?.user_id ? { kind: 'user', user_id: params.user_id } : { kind: 'global' },
+          user_id: params?.user_id ?? null,
+          hours: params?.hours,
+          limit: params?.limit,
+          include_user_info: params?.include_user_info ?? false,
+        })
        return response.data
      },
      30000
@@ -3,10 +3,13 @@ import type { AxiosInstance, AxiosRequestConfig, AxiosResponse, InternalAxiosReq
 import { NETWORK_CONFIG, AUTH_CONFIG } from '@/config/constants'
 import { isDemoMode } from '@/config/demo'
 import { handleMockRequest, setMockUserToken } from '@/mocks'
+import { getClientDeviceId } from '@/utils/deviceId'
+import { CrossTabRefreshCoordinator } from '@/utils/crossTabRefresh'
 import { log } from '@/utils/logger'

 // 在开发环境下使用代理,生产环境使用环境变量
 const API_BASE_URL = import.meta.env.VITE_API_URL || ''
+export const AUTH_STATE_CHANGE_EVENT = 'aether-auth-state-change'

 /**
 * 判断请求是否为公共端点
@@ -31,17 +34,15 @@ function isAuthRequest(url?: string): boolean {
 }

 /**
- * 判断是否为可刷新的认证错误
+ * 判断 403 错误是否表示用户账号级别的问题（需要清除认证并跳转）
 */
-function isRefreshableAuthError(errorDetail: string): boolean {
-  const nonRefreshableErrors = [
+function isAccountLevelForbidden(status: number, errorDetail: string): boolean {
+  if (status !== 403) return false
+  const accountErrors = [
    '用户不存在或已禁用',
-    '需要管理员权限',
-    '权限不足',
    '用户已禁用',
  ]
-
-  return !nonRefreshableErrors.some((msg) => errorDetail.includes(msg))
+  return accountErrors.some((msg) => errorDetail.includes(msg))
 }

 /**
@@ -83,12 +84,21 @@ class ApiClient {
  private client: AxiosInstance
  private token: string | null = null
  private isRefreshing = false
-  private refreshPromise: Promise<AxiosResponse> | null = null
+  private refreshPromise: Promise<string> | null = null
+  private readonly refreshCoordinator = new CrossTabRefreshCoordinator()
+
+  private readonly onStorageSync = (event: StorageEvent): void => {
+    if (event.key !== 'access_token') {
+      return
+    }
+    this.syncTokenState(event.newValue)
+  }

  constructor() {
    this.client = axios.create({
      baseURL: API_BASE_URL,
      timeout: NETWORK_CONFIG.API_TIMEOUT,
+      withCredentials: true,
      headers: {
        'Content-Type': 'application/json',
      },
@@ -99,6 +109,7 @@ class ApiClient {
    this.client.defaults.adapter = createDemoAdapter(defaultAdapter)

    this.setupInterceptors()
+    this.setupCrossTabAuthSync()
  }

  /**
@@ -108,6 +119,10 @@ class ApiClient {
    // 请求拦截器 - 仅处理认证
    this.client.interceptors.request.use(
      (config) => {
+        if (config.url?.includes('/api/')) {
+          config.headers['X-Client-Device-Id'] = getClientDeviceId()
+        }
+
        const requiresAuth = !isPublicEndpoint(config.url, config.method) &&
                           config.url?.includes('/api/')

@@ -129,6 +144,23 @@ class ApiClient {
    )
  }

+  private setupCrossTabAuthSync(): void {
+    if (typeof window !== 'undefined') {
+      window.addEventListener('storage', this.onStorageSync)
+    }
+  }
+
+  private emitAuthStateChange(token: string | null): void {
+    if (typeof window === 'undefined') {
+      return
+    }
+    window.dispatchEvent(
+      new CustomEvent<{ token: string | null }>(AUTH_STATE_CHANGE_EVENT, {
+        detail: { token },
+      })
+    )
+  }
+
  /**
   * 处理响应错误
   */
@@ -155,8 +187,22 @@ class ApiClient {
      return Promise.reject(error)
    }

+    const status = error.response?.status ?? 0
+
+    // 处理 403 用户账号级别错误（被禁用/删除）
+    if (status === 403) {
+      const rawDetail = (error.response?.data as Record<string, unknown>)?.detail
+      const errorDetail = typeof rawDetail === 'string' ? rawDetail : ''
+      if (isAccountLevelForbidden(status, errorDetail)) {
+        log.info('User account issue detected, clearing auth', { errorDetail })
+        this.clearAuth()
+        window.location.href = '/'
+        return Promise.reject(error)
+      }
+    }
+
    // 处理401错误
-    if (error.response?.status === 401) {
+    if (status === 401) {
      return this.handle401Error(error, originalRequest)
    }

@@ -177,25 +223,7 @@ class ApiClient {
      return Promise.reject(error)
    }

-    const errorDetail = (error.response?.data as Record<string, unknown>)?.detail as string || ''
-    log.debug('Got 401 error, attempting token refresh', { errorDetail })
-
-    // 检查是否为业务相关的401错误（用户被禁用/删除等）
-    if (!isRefreshableAuthError(errorDetail)) {
-      log.info('User account issue detected, logging out and redirecting to home', { errorDetail })
-      this.clearAuth()
-      // 跳转到首页
-      window.location.href = '/'
-      return Promise.reject(error)
-    }
-
-    // 获取refresh token
-    const refreshToken = localStorage.getItem('refresh_token')
-    if (!refreshToken) {
-      log.info('No refresh token available, clearing invalid token')
-      this.clearAuth()
-      return Promise.reject(error)
-    }
+    log.debug('Got 401 error, attempting token refresh')

    // 标记为已重试
    originalRequest._retry = true
@@ -210,8 +238,8 @@ class ApiClient {
    // 如果正在刷新,等待刷新完成
    if (this.isRefreshing) {
      try {
-        await this.refreshPromise
-        originalRequest.headers.Authorization = `Bearer ${this.getToken()}`
+        const accessToken = await this.refreshPromise
+        originalRequest.headers.Authorization = `Bearer ${accessToken}`
        return this.client.request(originalRequest)
      } catch {
        return Promise.reject(error)
@@ -219,29 +247,27 @@ class ApiClient {
    }

    // 开始刷新token
-    return this.refreshTokenAndRetry(refreshToken, originalRequest, error)
+    return this.refreshTokenAndRetry(originalRequest, error)
  }

  /**
   * 刷新token并重试原始请求
   */
  private async refreshTokenAndRetry(
-    refreshToken: string,
    originalRequest: InternalAxiosRequestConfig,
    originalError: import('axios').AxiosError
  ): Promise<AxiosResponse> {
    this.isRefreshing = true
-    this.refreshPromise = this.refreshToken(refreshToken)
+    this.refreshPromise = this.coordinatedRefresh()

    try {
-      const response = await this.refreshPromise
-      this.setToken(response.data.access_token)
-      localStorage.setItem('refresh_token', response.data.refresh_token)
+      const accessToken = await this.refreshPromise
+      this.setToken(accessToken)
      this.isRefreshing = false
      this.refreshPromise = null

      // 重试原始请求
-      originalRequest.headers.Authorization = `Bearer ${response.data.access_token}`
+      originalRequest.headers.Authorization = `Bearer ${accessToken}`
      return this.client.request(originalRequest)
    } catch (refreshError: unknown) {
      log.error('Token refresh failed', refreshError instanceof Error ? refreshError.message : String(refreshError))
@@ -252,15 +278,29 @@ class ApiClient {
    }
  }

-  setToken(token: string): void {
+  private async coordinatedRefresh(): Promise<string> {
+    return this.refreshCoordinator.run(async () => {
+      const response = await this.refreshToken()
+      const accessToken = response.data.access_token
+      if (!accessToken) {
+        throw new Error('Refresh response missing access token')
+      }
+      return accessToken
+    })
+  }
+
+  private syncTokenState(token: string | null): void {
    this.token = token
-    localStorage.setItem('access_token', token)
-    // 同步到 mock handler
    if (isDemoMode()) {
      setMockUserToken(token)
    }
  }

+  setToken(token: string): void {
+    this.syncTokenState(token)
+    localStorage.setItem('access_token', token)
+  }
+
  getToken(): string | null {
    if (!this.token) {
      this.token = localStorage.getItem('access_token')
@@ -273,18 +313,17 @@ class ApiClient {
  }

  clearAuth(): void {
-    this.token = null
+    const hadAuth = this.token !== null || localStorage.getItem('access_token') !== null
+    this.syncTokenState(null)
    localStorage.removeItem('access_token')
-    localStorage.removeItem('refresh_token')
-    // 同步清除 mock token
-    if (isDemoMode()) {
-      setMockUserToken(null)
+    // 同标签页内清理认证状态时不会触发 storage 事件，这里主动广播一次。
+    if (hadAuth) {
+      this.emitAuthStateChange(null)
    }
  }

-  async refreshToken(refreshToken: string): Promise<AxiosResponse> {
-    // refreshToken 会通过 adapter 处理 Demo 模式
-    return this.client.post('/api/auth/refresh', { refresh_token: refreshToken })
+  async refreshToken(): Promise<AxiosResponse> {
+    return this.client.post('/api/auth/refresh', {})
  }

  // 以下方法直接委托给 axios client，Demo 模式由 adapter 统一处理
@@ -1,370 +0,0 @@
-import apiClient from './client'
-import { cachedRequest, buildCacheKey } from '@/utils/cache'
-
-export interface DashboardStat {
-  name: string
-  value: string
-  subValue?: string
-  change?: string
-  changeType?: 'increase' | 'decrease' | 'neutral'
-  extraBadge?: string
-  icon: string
-}
-
-export interface RecentRequest {
-  id: string // UUID
-  user: string
-  model: string
-  tokens: number
-  time: string
-}
-
-export interface ProviderStatus {
-  name: string
-  status: 'active' | 'inactive'
-  requests: number
-}
-
-// 系统健康指标（管理员专用）
-export interface SystemHealth {
-  avg_response_time: number
-  error_rate: number
-  error_requests: number
-  fallback_count: number
-  total_requests: number
-}
-
-// 成本统计（管理员专用）
-export interface CostStats {
-  total_cost: number
-  total_actual_cost: number
-  cost_savings: number
-}
-
-// 缓存统计
-export interface CacheStats {
-  cache_creation_tokens: number
-  cache_read_tokens: number
-  cache_creation_cost?: number
-  cache_read_cost?: number
-  cache_hit_rate?: number
-  total_cache_tokens: number
-}
-
-// 用户统计（管理员专用）
-export interface UserStats {
-  total: number
-  active: number
-}
-
-// Token 详细分类
-export interface TokenBreakdown {
-  input: number
-  output: number
-  cache_creation: number
-  cache_read: number
-}
-
-export interface DashboardStatsResponse {
-  stats: DashboardStat[]
-  today?: {
-    requests: number
-    tokens: number
-    cost: number
-    actual_cost?: number
-    cache_creation_tokens?: number
-    cache_read_tokens?: number
-  }
-  api_keys?: {
-    total: number
-    active: number
-  }
-  tokens?: {
-    month: number
-  }
-  // 管理员专用字段
-  system_health?: SystemHealth
-  cost_stats?: CostStats
-  cache_stats?: CacheStats
-  users?: UserStats
-  token_breakdown?: TokenBreakdown
-  // 普通用户专用字段
-  monthly_cost?: number
-}
-
-export interface RecentRequestsResponse {
-  requests: RecentRequest[]
-}
-
-export interface ProviderStatusResponse {
-  providers: ProviderStatus[]
-}
-
-// 视频/图像/音频计费信息
-export interface VideoBilling {
-  task_type: 'video' | 'image' | 'audio'
-  duration_seconds?: number  // 视频时长（秒）
-  resolution?: string        // 分辨率
-  video_price_per_second?: number  // 每秒单价
-  video_cost?: number        // 视频费用
-  cost?: number              // 总费用
-  rule_name?: string         // 计费规则名称
-  expression?: string        // 计费公式
-  status?: string            // 计费状态
-}
-
-export interface RequestDetail {
-  id: string // UUID
-  request_id: string
-  user: {
-    id: string // UUID
-    username: string
-    email: string
-  }
-  api_key: {
-    id: string // UUID
-    name: string
-    display: string
-  }
-  provider: string
-  api_format?: string
-  model: string
-  target_model?: string | null  // 映射后的目标模型名
-  tokens: {
-    input: number
-    output: number
-    total: number
-  }
-  cost: {
-    input: number
-    output: number
-    total: number
-  }
-  // Additional token fields
-  input_tokens?: number
-  output_tokens?: number
-  total_tokens?: number
-  cache_creation_input_tokens?: number
-  cache_creation_input_tokens_5m?: number
-  cache_creation_input_tokens_1h?: number
-  cache_read_input_tokens?: number
-  // Additional cost fields
-  input_cost?: number
-  output_cost?: number
-  total_cost?: number
-  cache_creation_cost?: number
-  cache_read_cost?: number
-  request_cost?: number  // 按次计费费用
-  // Historical pricing fields (per 1M tokens)
-  input_price_per_1m?: number
-  output_price_per_1m?: number
-  cache_creation_price_per_1m?: number
-  cache_read_price_per_1m?: number
-  price_per_request?: number  // 按次计费价格
-  request_type: string
-  is_stream: boolean
-  status_code: number
-  status?: string  // pending, streaming, completed, failed, cancelled
-  error_message?: string
-  response_time_ms: number
-  created_at: string
-  request_headers?: Record<string, unknown>
-  request_body?: Record<string, unknown>
-  provider_request_headers?: Record<string, unknown>
-  provider_request_body?: Record<string, unknown>
-  response_headers?: Record<string, unknown>
-  client_response_headers?: Record<string, unknown>
-  response_body?: Record<string, unknown>
-  client_response_body?: Record<string, unknown>
-  has_request_body?: boolean
-  has_provider_request_body?: boolean
-  has_response_body?: boolean
-  has_client_response_body?: boolean
-  metadata?: Record<string, unknown>
-  // 阶梯计费信息
-  tiered_pricing?: {
-    total_input_context: number  // 总输入上下文 (input + cache_read)
-    tier_index: number  // 命中的阶梯索引 (0-based)
-    tier_count: number  // 阶梯总数
-    source?: 'provider' | 'global'  // 定价来源: 提供商或全局
-    current_tier: {  // 当前命中的阶梯配置
-      up_to?: number | null
-      input_price_per_1m: number
-      output_price_per_1m: number
-      cache_creation_price_per_1m?: number
-      cache_read_price_per_1m?: number
-      cache_ttl_pricing?: Array<{
-        ttl_minutes: number
-        cache_creation_price_per_1m?: number
-        cache_read_price_per_1m?: number
-      }>
-    }
-    tiers: Array<{  // 完整阶梯配置列表
-      up_to?: number | null
-      input_price_per_1m: number
-      output_price_per_1m: number
-      cache_creation_price_per_1m?: number
-      cache_read_price_per_1m?: number
-      cache_ttl_pricing?: Array<{
-        ttl_minutes: number
-        cache_creation_price_per_1m?: number
-        cache_read_price_per_1m?: number
-      }>
-    }>
-  } | null
-  // 视频/图像/音频计费信息
-  video_billing?: VideoBilling | null
-}
-
-export interface CurlData {
-  url: string
-  method: string
-  headers: Record<string, string>
-  body: Record<string, unknown>
-  curl: string
-}
-
-export interface ReplayRequest {
-  provider_id?: string
-  endpoint_id?: string
-  api_key_id?: string
-  body_override?: Record<string, unknown>
-}
-
-export interface ReplayResponse {
-  url: string
-  provider: string
-  status_code: number
-  response_headers: Record<string, string>
-  response_body: Record<string, unknown>
-  response_time_ms: number
-}
-
-export interface ModelBreakdown {
-  model: string
-  requests: number
-  tokens: number
-  cost: number
-}
-
-export interface ModelSummary {
-  model: string
-  requests: number
-  tokens: number
-  cost: number
-  avg_response_time: number
-  cost_per_request: number
-  tokens_per_request: number
-}
-
-export interface ProviderSummary {
-  provider: string
-  requests: number
-  tokens: number
-  cost: number
-}
-
-export interface DailyStat {
-  date: string // ISO date string
-  requests: number
-  tokens: number
-  cost: number
-  avg_response_time: number // in seconds
-  unique_models: number
-  unique_providers?: number // 仅管理员返回
-  model_breakdown: ModelBreakdown[]
-}
-
-export interface DailyStatsResponse {
-  daily_stats: DailyStat[]
-  model_summary: ModelSummary[]
-  provider_summary?: ProviderSummary[] // 仅管理员返回
-  period: {
-    start_date: string
-    end_date: string
-    days: number
-  }
-}
-
-export interface TimeRangeParams {
-  start_date?: string
-  end_date?: string
-  preset?: string
-  granularity?: 'hour' | 'day' | 'week' | 'month'
-  timezone?: string
-  tz_offset_minutes?: number
-}
-
-export const dashboardApi = {
-  // 获取仪表盘统计数据
-  async getStats(params?: TimeRangeParams): Promise<DashboardStatsResponse> {
-    const cacheKey = buildCacheKey('dashboard:stats', params)
-    return cachedRequest(
-      cacheKey,
-      async () => {
-        const response = await apiClient.get<DashboardStatsResponse>('/api/dashboard/stats', { params })
-        return response.data
-      },
-      10 * 1000
-    )
-  },
-
-  // 获取最近的请求记录
-  async getRecentRequests(limit: number = 10): Promise<RecentRequest[]> {
-    const response = await apiClient.get<RecentRequestsResponse>('/api/dashboard/recent-requests', {
-      params: { limit }
-    })
-    return response.data.requests
-  },
-
-  // 获取提供商状态
-  async getProviderStatus(): Promise<ProviderStatus[]> {
-    return cachedRequest(
-      'dashboard:provider-status',
-      async () => {
-        const response = await apiClient.get<ProviderStatusResponse>('/api/dashboard/provider-status')
-        return response.data.providers
-      },
-      20 * 1000
-    )
-  },
-
-  // 获取请求详情
-  // NOTE: This method now calls the new RESTful API at /api/admin/usage/{id}
-  async getRequestDetail(requestId: string, options: { includeBodies?: boolean } = {}): Promise<RequestDetail> {
-    const response = await apiClient.get<RequestDetail>(`/api/admin/usage/${requestId}`, {
-      params: { include_bodies: options.includeBodies ?? true },
-    })
-    return response.data
-  },
-
-  // 获取每日统计数据
-  async getDailyStats(params?: TimeRangeParams & { days?: number }): Promise<DailyStatsResponse> {
-    const cacheKey = buildCacheKey('dashboard:daily-stats', params)
-    return cachedRequest(
-      cacheKey,
-      async () => {
-        const response = await apiClient.get<DailyStatsResponse>('/api/dashboard/daily-stats', {
-          params
-        })
-        return response.data
-      },
-      20 * 1000
-    )
-  },
-
-  // 获取 cURL 命令数据（含明文 API Key）
-  async getCurlData(requestId: string): Promise<CurlData> {
-    const response = await apiClient.get<CurlData>(`/api/admin/usage/${requestId}/curl`)
-    return response.data
-  },
-
-  // 回放请求到提供商
-  async replayRequest(requestId: string, params?: ReplayRequest): Promise<ReplayResponse> {
-    const response = await apiClient.post<ReplayResponse>(
-      `/api/admin/usage/${requestId}/replay`,
-      params || {}
-    )
-    return response.data
-  }
-}
@@ -4,6 +4,34 @@ import type { EndpointAPIKey, AllowedModels } from './types'
 // Re-export types for convenience
 export type { EndpointAPIKey, AllowedModels }

+export interface GroupedFormatKey {
+  id: string
+  provider_id: string
+  name: string
+  auth_type?: string
+  api_key_masked: string
+  internal_priority: number
+  global_priority_by_format: Record<string, number> | null
+  format_priority: number | null
+  rate_multipliers: Record<string, number> | null
+  is_active: boolean
+  provider_active: boolean
+  pool_enabled: boolean
+  circuit_breaker_open: boolean
+  provider_name: string
+  api_format: string
+  api_formats: string[]
+  capabilities: string[]
+  health_score: number | null
+  success_rate: number | null
+  avg_response_time_ms: number | null
+  request_count: number
+}
+
+function toNumberOrNull(value: unknown): number | null {
+  return typeof value === 'number' && Number.isFinite(value) ? value : null
+}
+
 /**
 * 能力定义类型
 */
@@ -125,6 +153,51 @@ export async function getProviderKeys(providerId: string): Promise<EndpointAPIKe
  return allKeys
 }

+/**
+ * 获取按 API 格式分组的 Key 列表
+ */
+export async function getKeysGroupedByFormat(): Promise<Record<string, GroupedFormatKey[]>> {
+  const response = await client.get('/api/admin/endpoints/keys/grouped-by-format')
+  const grouped = response.data as Record<string, Array<Record<string, unknown>>>
+  const result: Record<string, GroupedFormatKey[]> = {}
+
+  for (const [apiFormat, keys] of Object.entries(grouped || {})) {
+    if (!Array.isArray(keys)) continue
+
+    result[apiFormat] = keys.map((key) => ({
+      id: String(key.id || ''),
+      provider_id: String(key.provider_id || ''),
+      name: String(key.name || 'Unnamed Key'),
+      auth_type: typeof key.auth_type === 'string' ? key.auth_type : undefined,
+      api_key_masked: String(key.api_key_masked || '***'),
+      internal_priority: toNumberOrNull(key.internal_priority) ?? 0,
+      global_priority_by_format:
+        key.global_priority_by_format && typeof key.global_priority_by_format === 'object'
+          ? (key.global_priority_by_format as Record<string, number>)
+          : null,
+      format_priority: toNumberOrNull(key.format_priority),
+      rate_multipliers:
+        key.rate_multipliers && typeof key.rate_multipliers === 'object'
+          ? (key.rate_multipliers as Record<string, number>)
+          : null,
+      is_active: key.is_active !== false,
+      provider_active: key.provider_active !== false,
+      pool_enabled: key.pool_enabled === true,
+      circuit_breaker_open: key.circuit_breaker_open === true,
+      provider_name: String(key.provider_name || 'Unknown Provider'),
+      api_format: typeof key.api_format === 'string' ? key.api_format : apiFormat,
+      api_formats: Array.isArray(key.api_formats) ? key.api_formats.map(item => String(item)) : [apiFormat],
+      capabilities: Array.isArray(key.capabilities) ? key.capabilities.map(item => String(item)) : [],
+      health_score: toNumberOrNull(key.health_score),
+      success_rate: toNumberOrNull(key.success_rate),
+      avg_response_time_ms: toNumberOrNull(key.avg_response_time_ms),
+      request_count: toNumberOrNull(key.request_count) ?? 0,
+    }))
+  }
+
+  return result
+}
+
 /**
 * 为 Provider 添加 Key
 */
@@ -1,6 +1,7 @@
 import client from '../client'
 import { dedupedRequest } from '@/utils/cache'
-import type { AllowedModels, ProxyConfig } from './types/provider'
+import type { AllowedModels, OAuthOrganizationInfo, ProxyConfig } from './types/provider'
+import type { ProviderKeyStatusSnapshot } from './types/statusSnapshot'

 const POOL_BATCH_ACTION_TIMEOUT_MS = 5 * 60 * 1000

@@ -29,7 +30,7 @@ export interface PoolStatusResponse {
 * 获取 Provider 的号池状态
 */
 export async function getPoolStatus(providerId: string): Promise<PoolStatusResponse> {
-  const response = await client.get(`/api/admin/providers/${providerId}/pool-status`)
+  const response = await client.get<PoolStatusResponse>(`/api/admin/providers/${providerId}/pool-status`)
  return response.data
 }

@@ -40,7 +41,7 @@ export async function clearPoolCooldown(
  providerId: string,
  keyId: string,
 ): Promise<{ message: string }> {
-  const response = await client.post(
+  const response = await client.post<{ message: string }>(
    `/api/admin/providers/${providerId}/pool/clear-cooldown/${keyId}`,
  )
  return response.data
@@ -53,7 +54,7 @@ export async function resetPoolCost(
  providerId: string,
  keyId: string,
 ): Promise<{ message: string }> {
-  const response = await client.post(
+  const response = await client.post<{ message: string }>(
    `/api/admin/providers/${providerId}/pool/reset-cost/${keyId}`,
  )
  return response.data
@@ -99,9 +100,20 @@ export interface PoolKeyDetail {
  is_active: boolean
  auth_type: string
  oauth_expires_at?: number | null
-  oauth_invalid_at?: number | null
-  oauth_invalid_reason?: string | null
+  oauth_invalid_at?: number | null  // 兼容字段；优先使用 status_snapshot.oauth
+  oauth_invalid_reason?: string | null  // 兼容字段；优先使用 status_snapshot.oauth
  oauth_plan_type?: string | null
+  oauth_account_id?: string | null
+  oauth_account_user_id?: string | null
+  oauth_account_name?: string | null
+  oauth_organizations?: OAuthOrganizationInfo[] | null
+  account_status_code?: string | null  // 兼容字段；优先使用 status_snapshot.account
+  account_status_label?: string | null  // 兼容字段；优先使用 status_snapshot.account
+  account_status_reason?: string | null  // 兼容字段；优先使用 status_snapshot.account
+  account_status_blocked?: boolean  // 兼容字段；优先使用 status_snapshot.account
+  account_status_recoverable?: boolean  // 兼容字段；优先使用 status_snapshot.account
+  account_status_source?: string | null  // 兼容字段；优先使用 status_snapshot.account
+  status_snapshot?: ProviderKeyStatusSnapshot | null
  quota_updated_at?: number | null
  health_score?: number
  circuit_breaker_open?: boolean
@@ -125,6 +137,8 @@ export interface PoolKeyDetail {
  cost_window_usage: number
  cost_limit: number | null
  request_count: number
+  total_tokens: number
+  total_cost_usd: string
  sticky_sessions: number
  lru_score: number | null
  created_at: string | null
@@ -167,6 +181,24 @@ export interface PoolKeysQuery {
  page_size?: number
  search?: string
  status?: 'all' | 'active' | 'cooldown' | 'inactive'
+  quick_selectors?: string[]
+  search_scope?: 'name' | 'full'
+}
+
+export interface PoolKeySelectionRequest {
+  search?: string
+  quick_selectors?: string[]
+}
+
+export interface PoolKeySelectionItem {
+  key_id: string
+  key_name: string
+  auth_type: string
+}
+
+export interface PoolKeySelectionResponse {
+  total: number
+  items: PoolKeySelectionItem[]
 }

 export interface PoolBatchAction {
@@ -201,13 +233,29 @@ export async function listPoolKeys(
  providerId: string,
  params: PoolKeysQuery = {},
 ): Promise<PoolKeysPageResponse> {
-  const key = `pool:keys:${providerId}|${params.page ?? ''}|${params.page_size ?? ''}|${params.search ?? ''}|${params.status ?? ''}`
+  const normalizedParams = {
+    ...params,
+    quick_selectors: params.quick_selectors?.length ? params.quick_selectors.join(',') : undefined,
+  }
+  const key = `pool:keys:${providerId}|${normalizedParams.page ?? ''}|${normalizedParams.page_size ?? ''}|${normalizedParams.search ?? ''}|${normalizedParams.status ?? ''}|${normalizedParams.quick_selectors ?? ''}|${normalizedParams.search_scope ?? ''}`
  return dedupedRequest(key, async () => {
-    const response = await client.get<PoolKeysPageResponse>(`/api/admin/pool/${providerId}/keys`, { params })
+    const response = await client.get<PoolKeysPageResponse>(`/api/admin/pool/${providerId}/keys`, { params: normalizedParams })
    return response.data
  })
 }

+export async function resolvePoolKeySelection(
+  providerId: string,
+  body: PoolKeySelectionRequest,
+): Promise<PoolKeySelectionResponse> {
+  const response = await client.post<PoolKeySelectionResponse>(
+    `/api/admin/pool/${providerId}/keys/resolve-selection`,
+    body,
+    { timeout: POOL_BATCH_ACTION_TIMEOUT_MS },
+  )
+  return response.data
+}
+
 export async function batchActionPoolKeys(
  providerId: string,
  body: PoolBatchAction,
@@ -18,6 +18,8 @@ export interface ProviderOAuthCompleteResponse {
  expires_at?: number | null
  has_refresh_token: boolean
  email?: string | null
+  account_state_recheck_attempted?: boolean
+  account_state_recheck_error?: string | null
 }

 export interface ProviderOAuthCompleteResponseWithKey {
@@ -92,7 +94,7 @@ export async function completeProviderLevelOAuth(

 export async function importProviderRefreshToken(
  providerId: string,
-  data: { refresh_token: string; name?: string; proxy_node_id?: string }
+  data: { refresh_token?: string; access_token?: string; name?: string; proxy_node_id?: string }
 ): Promise<ProviderOAuthCompleteResponseWithKey> {
  const resp = await client.post(`/api/admin/provider-oauth/providers/${providerId}/import-refresh-token`, data)
  return resp.data
@@ -114,8 +114,36 @@ export async function createProvider(
 /**
 * 删除 Provider
 */
-export async function deleteProvider(providerId: string): Promise<{ message: string }> {
-  const response = await client.delete(`/api/admin/providers/${providerId}`)
+export interface ProviderDeleteSubmitResponse {
+  task_id: string
+  status: string
+  message: string
+}
+
+export interface ProviderDeleteTaskResponse {
+  task_id: string
+  provider_id: string
+  status: string
+  stage: string
+  total_keys: number
+  deleted_keys: number
+  total_endpoints: number
+  deleted_endpoints: number
+  message: string
+}
+
+export async function deleteProvider(providerId: string): Promise<ProviderDeleteSubmitResponse> {
+  const response = await client.delete<ProviderDeleteSubmitResponse>(`/api/admin/providers/${providerId}`)
+  return response.data
+}
+
+export async function getProviderDeleteTask(
+  providerId: string,
+  taskId: string,
+): Promise<ProviderDeleteTaskResponse> {
+  const response = await client.get<ProviderDeleteTaskResponse>(
+    `/api/admin/providers/${providerId}/delete-task/${taskId}`,
+  )
  return response.data
 }

@@ -166,6 +194,8 @@ export interface TestModelFailoverRequest {
  api_format?: string
  endpoint_id?: string
  message?: string
+  request_headers?: Record<string, unknown>
+  request_body?: Record<string, unknown>
  request_id?: string
  concurrency?: number
 }
@@ -184,6 +214,11 @@ export interface TestAttemptDetail {
  error_message?: string | null
  status_code?: number | null
  latency_ms?: number | null
+  request_url?: string | null
+  request_headers?: Record<string, unknown> | null
+  request_body?: unknown
+  response_headers?: Record<string, unknown> | null
+  response_body?: unknown
 }

 export interface TestModelFailoverResponse {
@@ -1,3 +1,5 @@
+import type { ProviderKeyStatusSnapshot } from './statusSnapshot'
+
 /**
 * 代理配置类型
 * 支持两种模式：
@@ -12,6 +14,13 @@ export interface ProxyConfig {
  enabled?: boolean   // 是否启用代理（false 时保留配置但不使用）
 }

+export interface OAuthOrganizationInfo {
+  id?: string | null
+  title?: string | null
+  is_default?: boolean | null
+  role?: string | null
+}
+
 /**
 * 请求头规则类型
 * - set: 设置/覆盖请求头
@@ -35,8 +44,6 @@ export interface HeaderRuleRename {
  to: string
 }

-export type HeaderRule = HeaderRuleSet | HeaderRuleDrop | HeaderRuleRename
-
 /**
 * 请求体规则类型
 * - set: 设置/覆盖字段
@@ -129,10 +136,28 @@ export type BodyRuleConditionOp =
  | 'exists' | 'not_exists'
  | 'in' | 'type_is'

-export interface BodyRuleCondition {
+export interface BodyRuleConditionLeaf {
  path: string
  op: BodyRuleConditionOp
  value?: unknown  // exists / not_exists 不需要 value
+  source?: 'original' | 'current'
+}
+
+export interface BodyRuleConditionAll {
+  all: BodyRuleCondition[]
+}
+
+export interface BodyRuleConditionAny {
+  any: BodyRuleCondition[]
+}
+
+export type BodyRuleCondition =
+  | BodyRuleConditionLeaf
+  | BodyRuleConditionAll
+  | BodyRuleConditionAny
+
+export type HeaderRule = (HeaderRuleSet | HeaderRuleDrop | HeaderRuleRename) & {
+  condition?: BodyRuleCondition
 }

 /**
@@ -264,8 +289,13 @@ export interface EndpointAPIKey {
  oauth_email?: string | null  // OAuth 授权的邮箱
  oauth_plan_type?: string | null  // Codex 订阅类型: plus/free/team/enterprise
  oauth_account_id?: string | null  // Codex ChatGPT 账号 ID
-  oauth_invalid_at?: number | null  // OAuth Token 失效时间（Unix 时间戳）
-  oauth_invalid_reason?: string | null  // OAuth Token 失效原因
+  oauth_account_user_id?: string | null  // Codex ChatGPT account-user 联合 ID
+  oauth_account_name?: string | null
+  oauth_organizations?: OAuthOrganizationInfo[] | null  // OAuth 关联组织/工作区摘要
+  oauth_temporary?: boolean | null  // 是否为仅 Access Token 导入的临时 OAuth 账号
+  oauth_invalid_at?: number | null  // 兼容字段；优先使用 status_snapshot.oauth
+  oauth_invalid_reason?: string | null  // 兼容字段；优先使用 status_snapshot.oauth
+  status_snapshot?: ProviderKeyStatusSnapshot | null
  // 上游元数据（由上游响应采集，如 Codex 额度信息 / Antigravity 配额信息）
  upstream_metadata?: UpstreamMetadata | null
  // Key 级别代理配置（覆盖 Provider 级别代理）
@@ -565,6 +595,7 @@ export interface HealthSummary {
    total: number
    active: number
    unhealthy: number
+    circuit_open?: number
  }
 }

@@ -0,0 +1,36 @@
+export interface OAuthStatusSnapshot {
+  code: 'none' | 'valid' | 'expiring' | 'expired' | 'invalid' | 'check_failed'
+  label?: string | null
+  reason?: string | null
+  expires_at?: number | null
+  invalid_at?: number | null
+  source?: string | null
+  requires_reauth?: boolean
+  expiring_soon?: boolean
+}
+
+export interface AccountStatusSnapshot {
+  code: string
+  label?: string | null
+  reason?: string | null
+  blocked: boolean
+  source?: string | null
+  recoverable?: boolean
+}
+
+export interface QuotaStatusSnapshot {
+  code: 'unknown' | 'ok' | 'exhausted'
+  label?: string | null
+  reason?: string | null
+  exhausted: boolean
+  usage_ratio?: number | null
+  updated_at?: number | null
+  reset_seconds?: number | null
+  plan_type?: string | null
+}
+
+export interface ProviderKeyStatusSnapshot {
+  oauth: OAuthStatusSnapshot
+  account: AccountStatusSnapshot
+  quota: QuotaStatusSnapshot
+}
@@ -1,8 +1,9 @@
 import apiClient from './client'
-import type { ActivityHeatmap } from '@/types/activity'
 import type { TieredPricingConfig } from './endpoints/types'
-import { cachedRequest, buildCacheKey } from '@/utils/cache'
 import type { BillingSummary } from './auth'
+import type { UserSession } from '@/types/session'
+
+export type { UserSession }

 export interface Profile {
  id: string // UUID
@@ -22,8 +23,6 @@ export interface Profile {
 export interface UserPreferences {
  avatar_url?: string
  bio?: string
-  default_provider_id?: string // UUID
-  default_provider?: Record<string, unknown>
  theme: string
  language: string
  timezone?: string
@@ -34,87 +33,6 @@ export interface UserPreferences {
  }
 }

-// 提供商配置接口
-export interface ProviderConfig {
-  provider_id: string
-  priority: number  // 优先级（越高越优先）
-  weight: number    // 负载均衡权重
-  enabled: boolean  // 是否启用
-}
-
-// 使用记录接口
-export interface UsageRecordDetail {
-  id: string
-  provider: string
-  model: string
-  input_tokens: number
-  output_tokens: number
-  total_tokens: number
-  cost: number  // 官方费率
-  actual_cost?: number  // 倍率消耗（仅管理员可见）
-  rate_multiplier?: number  // 成本倍率（仅管理员可见）
-  response_time_ms?: number
-  is_stream: boolean
-  created_at: string
-  cache_creation_input_tokens?: number
-  cache_read_input_tokens?: number
-  status_code: number
-  error_message?: string
-  input_price_per_1m: number
-  output_price_per_1m: number
-  cache_creation_price_per_1m?: number
-  cache_read_price_per_1m?: number
-  price_per_request?: number  // 按次计费价格
-  api_key?: {
-    id: string
-    name: string
-    display: string
-  }
-}
-
-// 模型统计接口
-export interface ModelSummary {
-  model: string
-  requests: number
-  input_tokens: number
-  output_tokens: number
-  total_tokens: number
-  total_cost_usd: number
-  actual_total_cost_usd?: number  // 倍率消耗（仅管理员可见）
-}
-
-// 提供商统计接口
-export interface ProviderSummary {
-  provider: string
-  requests: number
-  total_tokens: number
-  total_cost_usd: number
-  success_rate: number | null
-  avg_response_time_ms: number | null
-}
-
-// 使用统计响应接口
-export interface UsageResponse {
-  total_requests: number
-  total_input_tokens: number
-  total_output_tokens: number
-  total_tokens: number
-  total_cost: number  // 官方费率
-  total_actual_cost?: number  // 倍率消耗（仅管理员可见）
-  avg_response_time: number
-  billing: BillingSummary
-  summary_by_model: ModelSummary[]
-  summary_by_provider?: ProviderSummary[]
-  pagination?: {
-    total: number
-    limit: number
-    offset: number
-    has_more: boolean
-  }
-  records: UsageRecordDetail[]
-  activity_heatmap?: ActivityHeatmap | null
-}
-
 export interface ApiKey {
  id: string // UUID
  name: string
@@ -126,7 +44,7 @@ export interface ApiKey {
  created_at: string
  total_requests?: number
  total_cost_usd?: number
-  allowed_providers?: ProviderConfig[]
+  rate_limit?: number | null
  force_capabilities?: Record<string, boolean> | null  // 强制能力配置
 }

@@ -159,14 +77,36 @@ export const meApi = {
    return response.data
  },

+  async listSessions(): Promise<UserSession[]> {
+    const response = await apiClient.get<UserSession[]>('/api/users/me/sessions')
+    return response.data
+  },
+
+  async updateSessionLabel(sessionId: string, deviceLabel: string): Promise<UserSession> {
+    const response = await apiClient.patch<UserSession>(`/api/users/me/sessions/${sessionId}`, {
+      device_label: deviceLabel,
+    })
+    return response.data
+  },
+
+  async revokeSession(sessionId: string): Promise<{ message: string }> {
+    const response = await apiClient.delete(`/api/users/me/sessions/${sessionId}`)
+    return response.data
+  },
+
+  async revokeOtherSessions(): Promise<{ message: string; revoked_count: number }> {
+    const response = await apiClient.delete('/api/users/me/sessions/others')
+    return response.data
+  },
+
  // API密钥管理
  async getApiKeys(): Promise<ApiKey[]> {
    const response = await apiClient.get<ApiKey[]>('/api/users/me/api-keys')
    return response.data
  },

-  async createApiKey(name: string): Promise<ApiKey> {
-    const response = await apiClient.post<ApiKey>('/api/users/me/api-keys', { name })
+  async createApiKey(data: { name: string; rate_limit?: number }): Promise<ApiKey> {
+    const response = await apiClient.post<ApiKey>('/api/users/me/api-keys', data)
    return response.data
  },

@@ -196,48 +136,14 @@ export const meApi = {
    return response.data
  },

-  // 使用统计
-  async getUsage(params?: {
-    start_date?: string
-    end_date?: string
-    preset?: string
-    timezone?: string
-    tz_offset_minutes?: number
-    search?: string  // 通用搜索：密钥名、模型名
-    limit?: number
-    offset?: number
-  }): Promise<UsageResponse> {
-    const response = await apiClient.get<UsageResponse>('/api/users/me/usage', { params })
-    return response.data
-  },
-
-  // 获取活跃请求状态（用于轮询更新）
-  async getActiveRequests(ids?: string): Promise<{
-    requests: Array<{
-      id: string
-      status: 'pending' | 'streaming' | 'completed' | 'failed' | 'cancelled'
-      input_tokens: number
-      output_tokens: number
-      cache_creation_input_tokens?: number | null
-      cache_read_input_tokens?: number | null
-      cost: number
-      actual_cost?: number | null
-      rate_multiplier?: number | null
-      response_time_ms: number | null
-      first_byte_time_ms: number | null
-      api_format?: string | null
-      endpoint_api_format?: string | null
-      has_format_conversion?: boolean | null
-    }>
-  }> {
-    const params = ids ? { ids } : {}
-    const response = await apiClient.get('/api/users/me/usage/active', { params })
-    return response.data
-  },
-
-  // 获取可用的提供商
-  async getAvailableProviders(): Promise<Array<Record<string, unknown>>> {
-    const response = await apiClient.get('/api/users/me/providers')
+  async updateApiKey(
+    keyId: string,
+    data: { name?: string; rate_limit?: number | null }
+  ): Promise<ApiKey & { message: string }> {
+    const response = await apiClient.put<ApiKey & { message: string }>(
+      `/api/users/me/api-keys/${keyId}`,
+      data
+    )
    return response.data
  },

@@ -281,16 +187,6 @@ export const meApi = {
    return response.data
  },

-  // 提供商绑定管理相关方法已移除，改为直接从可用提供商中选择
-
-  // API密钥提供商关联
-  async updateApiKeyProviders(keyId: string, data: {
-    allowed_providers?: ProviderConfig[]
-  }): Promise<{ message: string }> {
-    const response = await apiClient.put(`/api/users/me/api-keys/${keyId}/providers`, data)
-    return response.data
-  },
-
  // API密钥能力配置
  async updateApiKeyCapabilities(keyId: string, data: {
    force_capabilities?: Record<string, boolean> | null
@@ -317,39 +213,4 @@ export const meApi = {
    return response.data
  },

-  // 获取请求间隔时间线（用于散点图）
-  async getIntervalTimeline(params?: {
-    hours?: number
-    limit?: number
-  }): Promise<{
-    analysis_period_hours: number
-    total_points: number
-    points: Array<{ x: string; y: number; model?: string }>
-    models?: string[]
-  }> {
-    const cacheKey = buildCacheKey('me:interval-timeline', params as Record<string, unknown> | undefined)
-    return cachedRequest(
-      cacheKey,
-      async () => {
-        const response = await apiClient.get('/api/users/me/usage/interval-timeline', { params })
-        return response.data
-      },
-      30000
-    )
-  },
-
-  /**
-   * 获取活跃度热力图数据（用户）
-   * 后端已缓存5分钟
-   */
-  async getActivityHeatmap(): Promise<ActivityHeatmap> {
-    return cachedRequest(
-      'me-activity-heatmap',
-      async () => {
-        const response = await apiClient.get<ActivityHeatmap>('/api/users/me/usage/heatmap')
-        return response.data
-      },
-      60000
-    )
-  }
 }
@@ -0,0 +1,161 @@
+import apiClient from './client'
+
+export interface VideoBilling {
+  task_type: 'video' | 'image' | 'audio'
+  duration_seconds?: number
+  resolution?: string
+  video_price_per_second?: number
+  video_cost?: number
+  cost?: number
+  rule_name?: string
+  expression?: string
+  status?: string
+}
+
+export interface RequestDetail {
+  id: string
+  request_id: string
+  user: {
+    id: string | null
+    username: string | null
+    email: string | null
+  }
+  api_key: {
+    id: string | null
+    name: string | null
+    display: string | null
+  }
+  provider_api_key: {
+    id: string | null
+    name: string | null
+  }
+  provider: string
+  api_format?: string
+  model: string
+  target_model?: string | null
+  tokens: {
+    input: number
+    output: number
+    total: number
+  }
+  cost: {
+    input: number
+    output: number
+    total: number
+  }
+  input_tokens?: number
+  output_tokens?: number
+  total_tokens?: number
+  cache_creation_input_tokens?: number
+  cache_creation_input_tokens_5m?: number
+  cache_creation_input_tokens_1h?: number
+  cache_read_input_tokens?: number
+  input_cost?: number
+  output_cost?: number
+  total_cost?: number
+  cache_creation_cost?: number
+  cache_read_cost?: number
+  request_cost?: number
+  input_price_per_1m?: number
+  output_price_per_1m?: number
+  cache_creation_price_per_1m?: number
+  cache_read_price_per_1m?: number
+  price_per_request?: number
+  request_type: string
+  is_stream: boolean
+  status_code: number
+  status?: string
+  error_message?: string
+  response_time_ms: number
+  created_at: string
+  request_headers?: Record<string, unknown>
+  request_body?: Record<string, unknown>
+  provider_request_headers?: Record<string, unknown>
+  provider_request_body?: Record<string, unknown>
+  response_headers?: Record<string, unknown>
+  client_response_headers?: Record<string, unknown>
+  response_body?: Record<string, unknown>
+  client_response_body?: Record<string, unknown>
+  has_request_body?: boolean
+  has_provider_request_body?: boolean
+  has_response_body?: boolean
+  has_client_response_body?: boolean
+  metadata?: Record<string, unknown>
+  tiered_pricing?: {
+    total_input_context: number
+    tier_index: number
+    tier_count: number
+    source?: 'provider' | 'global'
+    current_tier: {
+      up_to?: number | null
+      input_price_per_1m: number
+      output_price_per_1m: number
+      cache_creation_price_per_1m?: number
+      cache_read_price_per_1m?: number
+      cache_ttl_pricing?: Array<{
+        ttl_minutes: number
+        cache_creation_price_per_1m?: number
+        cache_read_price_per_1m?: number
+      }>
+    }
+    tiers: Array<{
+      up_to?: number | null
+      input_price_per_1m: number
+      output_price_per_1m: number
+      cache_creation_price_per_1m?: number
+      cache_read_price_per_1m?: number
+      cache_ttl_pricing?: Array<{
+        ttl_minutes: number
+        cache_creation_price_per_1m?: number
+        cache_read_price_per_1m?: number
+      }>
+    }>
+  } | null
+  video_billing?: VideoBilling | null
+}
+
+export interface CurlData {
+  url: string
+  method: string
+  headers: Record<string, string>
+  body: Record<string, unknown>
+  curl: string
+}
+
+export interface ReplayRequest {
+  provider_id?: string
+  endpoint_id?: string
+  api_key_id?: string
+  body_override?: Record<string, unknown>
+}
+
+export interface ReplayResponse {
+  url: string
+  provider: string
+  status_code: number
+  response_headers: Record<string, string>
+  response_body: Record<string, unknown>
+  response_time_ms: number
+}
+
+export const requestDetailsApi = {
+  async getRequestDetail(requestId: string, options: { includeBodies?: boolean } = {}): Promise<RequestDetail> {
+    const response = await apiClient.get<RequestDetail>(`/api/admin/usage/${requestId}`, {
+      params: { include_bodies: options.includeBodies ?? true },
+    })
+    return response.data
+  },
+
+  async getCurlData(requestId: string): Promise<CurlData> {
+    const response = await apiClient.get<CurlData>(`/api/admin/usage/${requestId}/curl`)
+    return response.data
+  },
+
+  async replayRequest(requestId: string, params?: ReplayRequest): Promise<ReplayResponse> {
+    const response = await apiClient.post<ReplayResponse>(
+      `/api/admin/usage/${requestId}/replay`,
+      params || {},
+    )
+    return response.data
+  },
+}
@@ -1,241 +0,0 @@
-import apiClient from './client'
-import { cachedRequest, dedupedRequest, buildCacheKey } from '@/utils/cache'
-import type { ActivityHeatmap } from '@/types/activity'
-
-export interface UsageRecord {
-  id: string // UUID
-  user_id: string // UUID
-  username?: string
-  provider_id?: string // UUID
-  provider_name?: string
-  model: string
-  input_tokens: number
-  output_tokens: number
-  cache_creation_input_tokens?: number
-  cache_read_input_tokens?: number
-  total_tokens: number
-  cost?: number
-  response_time?: number
-  created_at: string
-  has_fallback?: boolean // 🆕 是否发生了 fallback
-}
-
-export interface UsageStats {
-  total_requests: number
-  total_tokens: number
-  total_cost: number
-  total_actual_cost?: number
-  avg_response_time: number
-  today?: {
-    requests: number
-    tokens: number
-    cost: number
-  }
-  activity_heatmap?: ActivityHeatmap | null
-}
-
-export interface UsageByModel {
-  model: string
-  request_count: number
-  total_tokens: number
-  total_cost: number
-  avg_response_time?: number
-}
-
-export interface UsageByUser {
-  user_id: string // UUID
-  email: string
-  username: string
-  request_count: number
-  total_tokens: number
-  total_cost: number
-}
-
-export interface UsageByProvider {
-  provider_id: string
-  provider: string
-  request_count: number
-  total_tokens: number
-  total_cost: number
-  actual_cost: number
-  avg_response_time_ms: number
-  success_rate: number
-  error_count: number
-}
-
-export interface UsageByApiFormat {
-  api_format: string
-  request_count: number
-  total_tokens: number
-  total_cost: number
-  actual_cost: number
-  avg_response_time_ms: number
-}
-
-export interface UsageFilters {
-  user_id?: string // UUID
-  provider_id?: string // UUID
-  model?: string
-  start_date?: string
-  end_date?: string
-  preset?: string
-  granularity?: 'hour' | 'day' | 'week' | 'month'
-  timezone?: string
-  tz_offset_minutes?: number
-  page?: number
-  page_size?: number
-}
-
-export const usageApi = {
-  async getUsageRecords(filters?: UsageFilters): Promise<{
-    records: UsageRecord[]
-    total: number
-    page: number
-    page_size: number
-  }> {
-    const response = await apiClient.get('/api/usage', { params: filters })
-    return response.data
-  },
-
-  async getUsageStats(filters?: UsageFilters): Promise<UsageStats> {
-    // 为统计数据添加30秒缓存
-    const cacheKey = `usage-stats-${JSON.stringify(filters || {})}`
-    return cachedRequest(
-      cacheKey,
-      async () => {
-        const response = await apiClient.get<UsageStats>('/api/admin/usage/stats', { params: filters })
-        return response.data
-      },
-      30000 // 30秒缓存
-    )
-  },
-
-  /**
-   * Get usage aggregation by dimension (RESTful API)
-   * @param groupBy Aggregation dimension: 'model', 'user', 'provider', or 'api_format'
-   * @param filters Optional filters
-   */
-  async getUsageAggregation<T = UsageByModel[] | UsageByUser[] | UsageByProvider[] | UsageByApiFormat[]>(
-    groupBy: 'model' | 'user' | 'provider' | 'api_format',
-    filters?: UsageFilters & { limit?: number }
-  ): Promise<T> {
-    const cacheKey = `usage-aggregation-${groupBy}-${JSON.stringify(filters || {})}`
-    return cachedRequest(
-      cacheKey,
-      async () => {
-        const response = await apiClient.get<T>('/api/admin/usage/aggregation/stats', {
-          params: { group_by: groupBy, ...filters }
-        })
-        return response.data
-      },
-      30000 // 30秒缓存
-    )
-  },
-
-  // Shorthand methods using getUsageAggregation
-  async getUsageByModel(filters?: UsageFilters & { limit?: number }): Promise<UsageByModel[]> {
-    return this.getUsageAggregation<UsageByModel[]>('model', filters)
-  },
-
-  async getUsageByUser(filters?: UsageFilters & { limit?: number }): Promise<UsageByUser[]> {
-    return this.getUsageAggregation<UsageByUser[]>('user', filters)
-  },
-
-  async getUsageByProvider(filters?: UsageFilters & { limit?: number }): Promise<UsageByProvider[]> {
-    return this.getUsageAggregation<UsageByProvider[]>('provider', filters)
-  },
-
-  async getUsageByApiFormat(filters?: UsageFilters & { limit?: number }): Promise<UsageByApiFormat[]> {
-    return this.getUsageAggregation<UsageByApiFormat[]>('api_format', filters)
-  },
-
-  async getUserUsage(userId: string, filters?: UsageFilters): Promise<{
-    records: UsageRecord[]
-    stats: UsageStats
-  }> {
-    const response = await apiClient.get(`/api/users/${userId}/usage`, { params: filters })
-    return response.data
-  },
-
-  async exportUsage(format: 'csv' | 'json', filters?: UsageFilters): Promise<Blob> {
-    const response = await apiClient.get('/api/usage/export', {
-      params: { ...filters, format },
-      responseType: 'blob'
-    })
-    return response.data
-  },
-
-  async getAllUsageRecords(params?: {
-    start_date?: string
-    end_date?: string
-    preset?: string
-    granularity?: 'hour' | 'day' | 'week' | 'month'
-    timezone?: string
-    tz_offset_minutes?: number
-    search?: string  // 通用搜索：用户名、密钥名、模型名、提供商名
-    user_id?: string // UUID
-    username?: string
-    model?: string
-    provider?: string
-    api_format?: string  // API 格式筛选（如 openai:chat, claude:chat）
-    status?: string // 'stream' | 'standard' | 'error'
-    limit?: number
-    offset?: number
-  }): Promise<{
-    records: Array<Record<string, unknown>>
-    total: number
-    limit: number
-    offset: number
-  }> {
-    const key = buildCacheKey('usage:records', params as Record<string, unknown> | undefined)
-    return dedupedRequest(key, async () => {
-      const response = await apiClient.get('/api/admin/usage/records', { params })
-      return response.data
-    })
-  },
-
-  /**
-   * 获取活跃请求的状态（轻量级接口，用于轮询更新）
-   * @param ids 可选，逗号分隔的请求 ID 列表
-   */
-  async getActiveRequests(ids?: string[]): Promise<{
-    requests: Array<{
-      id: string
-      status: 'pending' | 'streaming' | 'completed' | 'failed' | 'cancelled'
-      input_tokens: number
-      output_tokens: number
-      cache_creation_input_tokens?: number | null
-      cache_read_input_tokens?: number | null
-      cost: number
-      actual_cost?: number | null
-      rate_multiplier?: number | null
-      response_time_ms: number | null
-      first_byte_time_ms: number | null
-      provider?: string | null
-      api_key_name?: string | null
-      api_format?: string | null
-      endpoint_api_format?: string | null
-      has_format_conversion?: boolean | null
-      target_model?: string | null
-    }>
-  }> {
-    const params = ids?.length ? { ids: ids.join(',') } : {}
-    const response = await apiClient.get('/api/admin/usage/active', { params })
-    return response.data
-  },
-
-  /**
-   * 获取活跃度热力图数据（管理员）
-   * 后端已缓存5分钟
-   */
-  async getActivityHeatmap(): Promise<ActivityHeatmap> {
-    return cachedRequest(
-      'admin-usage-activity-heatmap',
-      async () => {
-        const response = await apiClient.get<ActivityHeatmap>('/api/admin/usage/heatmap')
-        return response.data
-      },
-      60000
-    )
-  }
-}
@@ -10,6 +10,7 @@ export interface User {
  allowed_providers: string[] | null  // 允许使用的提供商 ID 列表
  allowed_api_formats: string[] | null  // 允许使用的 API 格式列表
  allowed_models: string[] | null  // 允许使用的模型名称列表
+  rate_limit?: number | null  // null = 跟随系统默认，0 = 不限制
  created_at: string
  updated_at?: string
  last_login_at?: string | null
@@ -25,6 +26,7 @@ export interface CreateUserRequest {
  allowed_providers?: string[] | null
  allowed_api_formats?: string[] | null
  allowed_models?: string[] | null
+  rate_limit?: number | null
 }

 export interface UpdateUserRequest {
@@ -36,6 +38,7 @@ export interface UpdateUserRequest {
  allowed_providers?: string[] | null
  allowed_api_formats?: string[] | null
  allowed_models?: string[] | null
+  rate_limit?: number | null
 }

 export interface ApiKey {
@@ -49,11 +52,18 @@ export interface ApiKey {
  is_active: boolean
  is_locked: boolean  // 管理员锁定标志
  is_standalone: boolean  // 是否为独立余额Key
-  rate_limit?: number  // 速率限制（请求/分钟）
+  rate_limit?: number | null  // 普通Key: 0 = 不限制，历史 null 视为跟随系统默认
  total_requests?: number  // 总请求数
  total_cost_usd?: number  // 总费用
 }

+export interface UpsertUserApiKeyRequest {
+  name?: string
+  rate_limit?: number | null
+}
+
+export type { UserSession } from '@/types/session'
+
 export const usersApi = {
  async getAllUsers(): Promise<User[]> {
    const response = await apiClient.get<User[]>('/api/admin/users')
@@ -84,8 +94,41 @@ export const usersApi = {
    return response.data.api_keys
  },

-  async createApiKey(userId: string, name?: string): Promise<ApiKey & { key: string }> {
-    const response = await apiClient.post<ApiKey & { key: string }>(`/api/admin/users/${userId}/api-keys`, { name })
+  async getUserSessions(userId: string): Promise<UserSession[]> {
+    const response = await apiClient.get<UserSession[]>(`/api/admin/users/${userId}/sessions`)
+    return response.data
+  },
+
+  async revokeUserSession(userId: string, sessionId: string): Promise<{ message: string }> {
+    const response = await apiClient.delete(`/api/admin/users/${userId}/sessions/${sessionId}`)
+    return response.data
+  },
+
+  async revokeAllUserSessions(userId: string): Promise<{ message: string; revoked_count: number }> {
+    const response = await apiClient.delete(`/api/admin/users/${userId}/sessions`)
+    return response.data
+  },
+
+  async createApiKey(
+    userId: string,
+    data: UpsertUserApiKeyRequest
+  ): Promise<ApiKey & { key: string }> {
+    const response = await apiClient.post<ApiKey & { key: string }>(
+      `/api/admin/users/${userId}/api-keys`,
+      data
+    )
+    return response.data
+  },
+
+  async updateApiKey(
+    userId: string,
+    keyId: string,
+    data: UpsertUserApiKeyRequest
+  ): Promise<ApiKey & { message: string }> {
+    const response = await apiClient.put<ApiKey & { message: string }>(
+      `/api/admin/users/${userId}/api-keys/${keyId}`,
+      data
+    )
    return response.data
  },

@@ -98,10 +141,5 @@ export const usersApi = {
      `/api/admin/users/${userId}/api-keys/${keyId}/full-key`
    )
    return response.data
-  },
-  // 管理员统计
-  async getUsageStats(): Promise<Record<string, unknown>> {
-    const response = await apiClient.get('/api/admin/usage/stats')
-    return response.data
  }
 }
@@ -60,6 +60,36 @@ export interface WalletTransactionsResponse extends WalletBalanceResponse {
  offset: number
 }

+export interface DailyUsageRecord {
+  id?: string | null
+  date: string | null
+  timezone?: string | null
+  total_cost: number
+  total_requests: number
+  input_tokens: number
+  output_tokens: number
+  cache_creation_tokens: number
+  cache_read_tokens: number
+  first_finalized_at?: string | null
+  last_finalized_at?: string | null
+  aggregated_at?: string | null
+  is_today: boolean
+}
+
+export type FlowItem =
+  | { type: 'transaction'; data: WalletTransaction }
+  | { type: 'daily_usage'; data: DailyUsageRecord }
+
+export interface WalletFlowResponse extends WalletBalanceResponse {
+  today_entry: DailyUsageRecord | null
+  items: FlowItem[]
+  total: number
+  limit: number
+  offset: number
+}
+
+export type TodayCostResponse = DailyUsageRecord
+
 export interface PaymentOrder {
  id: string
  order_no: string
@@ -131,6 +161,16 @@ export const walletApi = {
    return response.data
  },

+  async getFlow(params?: { limit?: number; offset?: number }): Promise<WalletFlowResponse> {
+    const response = await apiClient.get<WalletFlowResponse>('/api/wallet/flow', { params })
+    return response.data
+  },
+
+  async getTodayCost(): Promise<TodayCostResponse> {
+    const response = await apiClient.get<TodayCostResponse>('/api/wallet/today-cost')
+    return response.data
+  },
+
  async createRechargeOrder(payload: WalletRechargeCreateRequest): Promise<{
    order: PaymentOrder
    payment_instructions: Record<string, unknown>
@@ -1,5 +1,8 @@
 <template>
-  <div class="w-full h-full">
+  <div
+    class="w-full"
+    :style="{ height: `${props.height}px` }"
+  >
    <canvas ref="chartRef" />
  </div>
 </template>
@@ -18,6 +21,7 @@ import {
  type ChartData,
  type ChartOptions
 } from 'chart.js'
+import { observeChartThemeChanges, resolveChartTheme } from '@/utils/chartTheme'

 const props = withDefaults(defineProps<Props>(), {
  height: 300,
@@ -44,6 +48,7 @@ interface Props {

 const chartRef = ref<HTMLCanvasElement>()
 let chart: ChartJS<'bar'> | null = null
+let stopThemeObserver: (() => void) | null = null

 const defaultOptions: ChartOptions<'bar'> = {
  responsive: true,
@@ -91,9 +96,7 @@ const defaultOptions: ChartOptions<'bar'> = {
  }
 }

-function createChart() {
-  if (!chartRef.value) return
-
+function buildChartOptions(): ChartOptions<'bar'> {
  const stackedOptions = props.stacked ? {
    scales: {
      x: { ...defaultOptions.scales?.x, stacked: true },
@@ -106,20 +109,26 @@ function createChart() {
    }
  }

+  return resolveChartTheme({
+    ...defaultOptions,
+    ...stackedOptions,
+    ...props.options
+  })
+}
+
+function createChart() {
+  if (!chartRef.value) return
+
  chart = new ChartJS(chartRef.value, {
    type: 'bar',
-    data: props.data,
-    options: {
-      ...defaultOptions,
-      ...stackedOptions,
-      ...props.options
-    }
+    data: resolveChartTheme(props.data),
+    options: buildChartOptions()
  })
 }

 function updateChart() {
  if (chart) {
-    chart.data = props.data
+    chart.data = resolveChartTheme(props.data)
    chart.update('none')
  }
 }
@@ -127,9 +136,17 @@ function updateChart() {
 onMounted(async () => {
  await nextTick()
  createChart()
+  stopThemeObserver = observeChartThemeChanges(() => {
+    if (!chart) return
+    chart.data = resolveChartTheme(props.data)
+    chart.options = buildChartOptions()
+    chart.update('none')
+  })
 })

 onUnmounted(() => {
+  stopThemeObserver?.()
+  stopThemeObserver = null
  if (chart) {
    chart.destroy()
    chart = null
@@ -139,11 +156,8 @@ onUnmounted(() => {
 watch(() => props.data, updateChart, { deep: true })
 watch(() => props.options, () => {
  if (chart) {
-    chart.options = {
-      ...defaultOptions,
-      ...props.options
-    }
-    chart.update()
+    chart.options = buildChartOptions()
+    chart.update('none')
  }
 }, { deep: true })
 </script>
@@ -1,5 +1,8 @@
 <template>
-  <div class="w-full h-full">
+  <div
+    class="w-full"
+    :style="{ height: `${props.height}px` }"
+  >
    <canvas ref="chartRef" />
  </div>
 </template>
@@ -16,10 +19,12 @@ import {
  type ChartData,
  type ChartOptions
 } from 'chart.js'
+import { observeChartThemeChanges, resolveChartTheme } from '@/utils/chartTheme'

 const props = withDefaults(defineProps<Props>(), {
  height: 300,
-  options: undefined
+  options: undefined,
+  showLegend: true,
 })

 ChartJS.register(
@@ -34,15 +39,22 @@ interface Props {
  data: ChartData<'doughnut'>
  options?: ChartOptions<'doughnut'>
  height?: number
+  showLegend?: boolean
 }

 const chartRef = ref<HTMLCanvasElement>()
 let chart: ChartJS<'doughnut'> | null = null
+let stopThemeObserver: (() => void) | null = null

 const defaultOptions: ChartOptions<'doughnut'> = {
  responsive: true,
  maintainAspectRatio: false,
  cutout: '60%',
+  events: ['mousemove', 'mouseout', 'click', 'touchstart', 'touchmove'],
+  interaction: {
+    mode: 'nearest',
+    intersect: true,
+  },
  plugins: {
    legend: {
      position: 'right',
@@ -71,22 +83,59 @@ const defaultOptions: ChartOptions<'doughnut'> = {
  }
 }

+function resolvedOptions(): ChartOptions<'doughnut'> {
+  const incomingOptions = props.options ?? {}
+  const incomingPlugins = (incomingOptions.plugins ?? {}) as Record<string, unknown>
+  const incomingLegend = (incomingPlugins.legend ?? {}) as Record<string, unknown>
+  const incomingTooltip = (incomingPlugins.tooltip ?? {}) as Record<string, unknown>
+  const defaultPlugins = (defaultOptions.plugins ?? {}) as Record<string, unknown>
+  const defaultLegend = (defaultPlugins.legend ?? {}) as Record<string, unknown>
+  const defaultTooltip = (defaultPlugins.tooltip ?? {}) as Record<string, unknown>
+
+  return resolveChartTheme({
+    ...defaultOptions,
+    ...incomingOptions,
+    interaction: {
+      ...defaultOptions.interaction,
+      ...incomingOptions.interaction,
+    },
+    plugins: {
+      ...defaultPlugins,
+      ...incomingPlugins,
+      legend: {
+        ...defaultLegend,
+        ...incomingLegend,
+        display: incomingLegend.display ?? props.showLegend,
+        labels: {
+          ...(defaultLegend.labels as Record<string, unknown> | undefined),
+          ...(incomingLegend.labels as Record<string, unknown> | undefined),
+        },
+      },
+      tooltip: {
+        ...defaultTooltip,
+        ...incomingTooltip,
+        callbacks: {
+          ...(defaultTooltip.callbacks as Record<string, unknown> | undefined),
+          ...(incomingTooltip.callbacks as Record<string, unknown> | undefined),
+        },
+      },
+    },
+  })
+}
+
 function createChart() {
  if (!chartRef.value) return

  chart = new ChartJS(chartRef.value, {
    type: 'doughnut',
-    data: props.data,
-    options: {
-      ...defaultOptions,
-      ...props.options
-    }
+    data: resolveChartTheme(props.data),
+    options: resolvedOptions(),
  })
 }

 function updateChart() {
  if (chart) {
-    chart.data = props.data
+    chart.data = resolveChartTheme(props.data)
    chart.update('none')
  }
 }
@@ -94,9 +143,17 @@ function updateChart() {
 onMounted(async () => {
  await nextTick()
  createChart()
+  stopThemeObserver = observeChartThemeChanges(() => {
+    if (!chart) return
+    chart.data = resolveChartTheme(props.data)
+    chart.options = resolvedOptions()
+    chart.update('none')
+  })
 })

 onUnmounted(() => {
+  stopThemeObserver?.()
+  stopThemeObserver = null
  if (chart) {
    chart.destroy()
    chart = null
@@ -106,8 +163,15 @@ onUnmounted(() => {
 watch(() => props.data, updateChart, { deep: true })
 watch(() => props.options, () => {
  if (chart) {
-    chart.options = { ...defaultOptions, ...props.options }
-    chart.update()
+    chart.options = resolvedOptions()
+    chart.update('none')
  }
 }, { deep: true })
+
+watch(() => props.showLegend, () => {
+  if (chart) {
+    chart.options = resolvedOptions()
+    chart.update('none')
+  }
+})
 </script>
@@ -1,5 +1,8 @@
 <template>
-  <div class="w-full h-full">
+  <div
+    class="w-full"
+    :style="{ height: `${props.height}px` }"
+  >
    <canvas ref="chartRef" />
  </div>
 </template>
@@ -16,9 +19,11 @@ import {
  Title,
  Tooltip,
  Legend,
+  Filler,
  type ChartData,
  type ChartOptions
 } from 'chart.js'
+import { observeChartThemeChanges, resolveChartTheme } from '@/utils/chartTheme'

 const props = withDefaults(defineProps<Props>(), {
  height: 300,
@@ -34,7 +39,8 @@ ChartJS.register(
  LineController,
  Title,
  Tooltip,
-  Legend
+  Legend,
+  Filler
 )

 interface Props {
@@ -45,6 +51,18 @@ interface Props {

 const chartRef = ref<HTMLCanvasElement>()
 let chart: ChartJS<'line'> | null = null
+let stopThemeObserver: (() => void) | null = null
+
+function buildChartOptions(): ChartOptions<'line'> {
+  return resolveChartTheme({
+    ...defaultOptions,
+    ...props.options
+  })
+}
+
+function buildChartData(): ChartData<'line'> {
+  return resolveChartTheme(props.data)
+}

 const defaultOptions: ChartOptions<'line'> = {
  responsive: true,
@@ -88,17 +106,14 @@ function createChart() {

  chart = new ChartJS(chartRef.value, {
    type: 'line',
-    data: props.data,
-    options: {
-      ...defaultOptions,
-      ...props.options
-    }
+    data: buildChartData(),
+    options: buildChartOptions()
  })
 }

 function updateChart() {
  if (chart) {
-    chart.data = props.data
+    chart.data = buildChartData()
    chart.update('none') // 禁用动画以提高性能
  }
 }
@@ -106,24 +121,29 @@ function updateChart() {
 onMounted(async () => {
  await nextTick()
  createChart()
+  stopThemeObserver = observeChartThemeChanges(() => {
+    if (!chart) return
+    chart.data = buildChartData()
+    chart.options = buildChartOptions()
+    chart.update('none')
+  })
 })

 onUnmounted(() => {
+  stopThemeObserver?.()
+  stopThemeObserver = null
  if (chart) {
    chart.destroy()
    chart = null
  }
 })

-// 监听数据变化
-watch(() => props.data, updateChart, { deep: true })
+// 监听引用变化，避免深监听触发整图重算
+watch(() => props.data, updateChart)
 watch(() => props.options, () => {
  if (chart) {
-    chart.options = {
-      ...defaultOptions,
-      ...props.options
-    }
-    chart.update()
+    chart.options = buildChartOptions()
+    chart.update('none')
  }
-}, { deep: true })
-</script>
+})
+</script>
@@ -119,6 +119,11 @@ let chart: ChartJS<'scatter'> | null = null
 const crosshairY = ref<number | null>(null)
 const gapInfoList = ref<GapInfo[]>([])

+interface PreparedRenderData {
+  chartData: ChartData<'scatter'>
+  gaps: GapInfo[]
+}
+
 const crosshairStats = computed<CrosshairStats | null>(() => {
  if (crosshairY.value === null || !props.data.datasets) return null

@@ -294,6 +299,22 @@ function transformData(data: ChartData<'scatter'>): ChartData<'scatter'> {
  }
 }

+function prepareRenderData(): PreparedRenderData {
+  let dataToUse = props.data
+  let gaps: GapInfo[] = []
+
+  if (props.compressGaps) {
+    const compressedResult = compressTimeGaps(props.data)
+    dataToUse = compressedResult.data
+    gaps = compressedResult.gaps
+  }
+
+  return {
+    chartData: transformData(dataToUse),
+    gaps
+  }
+}
+
 // 格式化时长
 function formatDuration(ms: number): string {
  const hours = Math.floor(ms / (1000 * 60 * 60))
@@ -516,22 +537,12 @@ function handleMouseLeave() {
 function createChart() {
  if (!chartRef.value) return

-  let dataToUse = props.data
-  gapInfoList.value = []
-
-  // 如果启用间隙压缩
-  if (props.compressGaps) {
-    const { data: compressedData, gaps } = compressTimeGaps(props.data)
-    dataToUse = compressedData
-    gapInfoList.value = gaps
-  }
-
-  // 转换数据
-  const transformedData = transformData(dataToUse)
+  const { chartData, gaps } = prepareRenderData()
+  gapInfoList.value = gaps

  chart = new ChartJS(chartRef.value, {
    type: 'scatter',
-    data: transformedData,
+    data: chartData,
    options: {
      ...defaultOptions,
      ...props.options
@@ -544,16 +555,9 @@ function createChart() {

 function updateChart() {
  if (chart) {
-    let dataToUse = props.data
-    gapInfoList.value = []
-
-    if (props.compressGaps) {
-      const { data: compressedData, gaps } = compressTimeGaps(props.data)
-      dataToUse = compressedData
-      gapInfoList.value = gaps
-    }
-
-    chart.data = transformData(dataToUse)
+    const { chartData, gaps } = prepareRenderData()
+    gapInfoList.value = gaps
+    chart.data = chartData
    chart.update('none')
  }
 }
@@ -573,21 +577,22 @@ onUnmounted(() => {
  }
 })

-watch(() => props.data, updateChart, { deep: true })
-watch(() => props.compressGaps, () => {
-  if (chart) {
-    chart.destroy()
-    chart = null
-  }
-  createChart()
-})
+watch(
+  [
+    () => props.data,
+    () => props.compressGaps,
+    () => props.gapThreshold,
+    () => props.compressedGapSize
+  ],
+  updateChart
+)
 watch(() => props.options, () => {
  if (chart) {
    chart.options = {
      ...defaultOptions,
      ...props.options
    }
-    chart.update()
+    chart.update('none')
  }
-}, { deep: true })
+})
 </script>
@@ -4,7 +4,7 @@
      ref="fileInputRef"
      type="file"
      :accept="accept"
-      multiple
+      :multiple="multiple"
      class="hidden"
      @change="handleFileSelect"
    >
@@ -94,6 +94,7 @@ const props = withDefaults(defineProps<{
  disabled?: boolean
  resetKey?: string | number
  accept?: string
+  multiple?: boolean
  dropTitle?: string
  dropHint?: string
  manualLabel?: string
@@ -106,6 +107,7 @@ const props = withDefaults(defineProps<{
  disabled: false,
  resetKey: '',
  accept: '.json,.txt',
+  multiple: true,
  dropTitle: '拖入导入文件或点击选择',
  dropHint: '支持 .json / .txt，可多选',
  manualLabel: '',
@@ -218,13 +220,19 @@ function mergeFileContents(contents: string[]): string {
 }

 async function readFiles(files: File[]) {
-  const validFiles = files.filter(isValidFileType)
+  const sourceFiles = props.multiple ? files : files.slice(0, 1)
+
+  if (!props.multiple && files.length > 1) {
+    emitError('仅支持选择 1 个文件，已读取第一个文件', '提示')
+  }
+
+  const validFiles = sourceFiles.filter(isValidFileType)
  if (validFiles.length === 0) {
    emitError('仅支持 .json 或 .txt 文件', '格式错误')
    return
  }
-  if (validFiles.length < files.length) {
-    emitError(`已忽略 ${files.length - validFiles.length} 个不支持的文件`, '提示')
+  if (validFiles.length < sourceFiles.length) {
+    emitError(`已忽略 ${sourceFiles.length - validFiles.length} 个不支持的文件`, '提示')
  }

  try {
@@ -1,117 +0,0 @@
-<template>
-  <div class="space-y-2">
-    <Label class="text-sm font-medium">允许的模型</Label>
-    <div class="relative">
-      <button
-        type="button"
-        class="w-full h-10 px-3 border rounded-lg bg-background text-left flex items-center justify-between hover:bg-muted/50 transition-colors"
-        @click="isOpen = !isOpen"
-      >
-        <span :class="modelValue.length ? 'text-foreground' : 'text-muted-foreground'">
-          {{ modelValue.length ? `已选择 ${modelValue.length} 个` : '全部可用' }}
-          <span
-            v-if="invalidModels.length"
-            class="text-destructive"
-          >({{ invalidModels.length }} 个已失效)</span>
-        </span>
-        <ChevronDown
-          class="h-4 w-4 text-muted-foreground transition-transform"
-          :class="isOpen ? 'rotate-180' : ''"
-        />
-      </button>
-      <div
-        v-if="isOpen"
-        class="fixed inset-0 z-[80]"
-        @click.stop="isOpen = false"
-      />
-      <div
-        v-if="isOpen"
-        class="absolute z-[90] w-full mt-1 bg-popover border rounded-lg shadow-lg max-h-48 overflow-y-auto"
-      >
-        <!-- 失效模型（置顶显示，只能取消选择） -->
-        <div
-          v-for="modelName in invalidModels"
-          :key="modelName"
-          class="flex items-center gap-2 px-3 py-2 hover:bg-muted/50 cursor-pointer bg-destructive/5"
-          @click="removeModel(modelName)"
-        >
-          <input
-            type="checkbox"
-            :checked="true"
-            class="h-4 w-4 rounded border-gray-300 cursor-pointer"
-            @click.stop
-            @change="removeModel(modelName)"
-          >
-          <span class="text-sm text-destructive">{{ modelName }}</span>
-          <span class="text-xs text-destructive/70">(已失效)</span>
-        </div>
-        <!-- 有效模型 -->
-        <div
-          v-for="model in models"
-          :key="model.name"
-          class="flex items-center gap-2 px-3 py-2 hover:bg-muted/50 cursor-pointer"
-          @click="toggleModel(model.name)"
-        >
-          <input
-            type="checkbox"
-            :checked="modelValue.includes(model.name)"
-            class="h-4 w-4 rounded border-gray-300 cursor-pointer"
-            @click.stop
-            @change="toggleModel(model.name)"
-          >
-          <span class="text-sm">{{ model.name }}</span>
-        </div>
-        <div
-          v-if="models.length === 0 && invalidModels.length === 0"
-          class="px-3 py-2 text-sm text-muted-foreground"
-        >
-          暂无可用模型
-        </div>
-      </div>
-    </div>
-  </div>
-</template>
-
-<script setup lang="ts">
-import { ref, computed } from 'vue'
-import { Label } from '@/components/ui'
-import { ChevronDown } from 'lucide-vue-next'
-import { useInvalidModels } from '@/composables/useInvalidModels'
-
-export interface ModelWithName {
-  name: string
-}
-
-const props = defineProps<{
-  modelValue: string[]
-  models: ModelWithName[]
-}>()
-
-const emit = defineEmits<{
-  'update:modelValue': [value: string[]]
-}>()
-
-const isOpen = ref(false)
-
-// 检测失效模型
-const { invalidModels } = useInvalidModels(
-  computed(() => props.modelValue),
-  computed(() => props.models)
-)
-
-function toggleModel(name: string) {
-  const newValue = [...props.modelValue]
-  const index = newValue.indexOf(name)
-  if (index === -1) {
-    newValue.push(name)
-  } else {
-    newValue.splice(index, 1)
-  }
-  emit('update:modelValue', newValue)
-}
-
-function removeModel(name: string) {
-  const newValue = props.modelValue.filter(m => m !== name)
-  emit('update:modelValue', newValue)
-}
-</script>
@@ -2,21 +2,28 @@
  <div class="relative">
    <button
      type="button"
-      :class="cn(
-        'h-9 px-3 border rounded-lg bg-background text-left flex items-center justify-between hover:bg-muted/50 transition-colors gap-1',
-        triggerClass,
-      )"
+      :class="
+        cn(
+          'flex h-10 w-full items-center justify-between rounded-lg border bg-background px-3 text-left transition-colors',
+          disabled ? 'cursor-not-allowed opacity-60 hover:bg-background' : 'hover:bg-muted/50',
+          triggerClass,
+        )
+      "
      :disabled="disabled"
      @click="isOpen = !isOpen"
    >
      <span
        :class="modelValue.length ? 'text-foreground' : 'text-muted-foreground'"
-        class="text-xs truncate"
+        class="truncate text-sm"
      >
        {{ displayText }}
+        <span
+          v-if="invalidItems.length"
+          class="text-destructive"
+        >({{ invalidItems.length }} 个已失效)</span>
      </span>
      <ChevronDown
-        class="h-3.5 w-3.5 shrink-0 text-muted-foreground transition-transform"
+        class="h-4 w-4 shrink-0 text-muted-foreground transition-transform"
        :class="isOpen ? 'rotate-180' : ''"
      />
    </button>
@@ -27,76 +34,163 @@
    />
    <div
      v-if="isOpen"
-      class="absolute z-[90] w-full mt-1 bg-popover border rounded-lg shadow-lg max-h-48 overflow-y-auto"
+      class="absolute z-[90] w-full mt-1 rounded-lg border bg-popover shadow-lg"
      :style="dropdownMinWidth ? { minWidth: dropdownMinWidth } : undefined"
    >
      <div
-        v-for="item in options"
-        :key="item.value"
-        class="flex items-center gap-2 px-3 py-1.5 hover:bg-muted/50 cursor-pointer text-xs"
-        @click="toggle(item.value)"
+        v-if="showSearch"
+        class="sticky top-0 z-10 border-b bg-popover/95 p-1 backdrop-blur supports-[backdrop-filter]:bg-popover/85"
      >
-        <input
-          type="checkbox"
-          :checked="modelValue.includes(item.value)"
-          class="h-4 w-4 rounded border-border/60 bg-card/80 text-primary shadow-sm accent-primary cursor-pointer"
-          @click.stop
-          @change="toggle(item.value)"
-        >
-        <span class="text-sm">{{ item.label }}</span>
+        <div class="relative">
+          <Search
+            class="pointer-events-none absolute left-3 top-1/2 h-3.5 w-3.5 -translate-y-1/2 text-muted-foreground"
+          />
+          <Input
+            v-model="searchQuery"
+            :placeholder="searchPlaceholder"
+            class="h-8 rounded-md border-border/60 bg-popover pl-8 text-xs"
+            @keydown.stop
+          />
+        </div>
      </div>
-      <div
-        v-if="options.length === 0"
-        class="px-3 py-2 text-sm text-muted-foreground"
-      >
-        {{ emptyText }}
+
+      <div class="max-h-48 overflow-y-auto">
+        <div
+          v-for="item in filteredInvalidItems"
+          :key="'invalid-' + item"
+          class="flex cursor-pointer items-center gap-2 bg-destructive/5 px-3 py-2 hover:bg-muted/50"
+          @click="remove(item)"
+        >
+          <input
+            type="checkbox"
+            :checked="true"
+            class="h-4 w-4 shrink-0 cursor-pointer rounded border-gray-300"
+            @click.stop
+            @change="remove(item)"
+          >
+          <span class="min-w-0 truncate text-sm text-destructive">{{ item }}</span>
+          <span class="shrink-0 text-xs text-destructive/70">(已失效)</span>
+        </div>
+
+        <div
+          v-for="item in filteredOptions"
+          :key="item.value"
+          class="flex cursor-pointer items-center gap-2 px-3 py-2 hover:bg-muted/50"
+          @click="toggle(item.value)"
+        >
+          <input
+            type="checkbox"
+            :checked="modelValue.includes(item.value)"
+            class="h-4 w-4 shrink-0 cursor-pointer rounded border-gray-300"
+            @click.stop
+            @change="toggle(item.value)"
+          >
+          <span class="min-w-0 truncate text-sm">{{ item.label }}</span>
+        </div>
+        <div
+          v-if="filteredOptions.length === 0 && filteredInvalidItems.length === 0"
+          class="px-3 py-2 text-sm text-muted-foreground"
+        >
+          {{ searchQuery.trim() ? noResultsText : emptyText }}
+        </div>
      </div>
    </div>
  </div>
 </template>

 <script setup lang="ts">
-import { computed, ref } from 'vue'
-import { ChevronDown } from 'lucide-vue-next'
+import { computed, ref, watch } from 'vue'
+import { ChevronDown, Search } from 'lucide-vue-next'
+import { Input } from '@/components/ui'
 import { cn } from '@/lib/utils'
+import { matchesSearchQuery } from '@/utils/search'

 export interface MultiSelectOption {
  value: string
  label: string
 }

-const props = withDefaults(defineProps<{
-  modelValue: string[]
-  options: MultiSelectOption[]
-  placeholder?: string
-  emptyText?: string
-  triggerClass?: string
-  dropdownMinWidth?: string
-  disabled?: boolean
-}>(), {
-  placeholder: '请选择',
-  emptyText: '暂无选项',
-  triggerClass: '',
-  dropdownMinWidth: undefined,
-  disabled: false,
-})
+const props = withDefaults(
+  defineProps<{
+    modelValue: string[]
+    options: MultiSelectOption[]
+    placeholder?: string
+    emptyText?: string
+    noResultsText?: string
+    triggerClass?: string
+    dropdownMinWidth?: string
+    disabled?: boolean
+    searchable?: boolean
+    searchThreshold?: number
+    searchPlaceholder?: string
+  }>(),
+  {
+    placeholder: '请选择',
+    emptyText: '暂无选项',
+    noResultsText: '未找到匹配项',
+    triggerClass: '',
+    dropdownMinWidth: undefined,
+    disabled: false,
+    searchable: true,
+    searchThreshold: 8,
+    searchPlaceholder: '输入关键词搜索...',
+  },
+)

 const emit = defineEmits<{
  'update:modelValue': [value: string[]]
 }>()

 const isOpen = ref(false)
+const searchQuery = ref('')
+
+const validValues = computed(() => new Set(props.options.map(o => o.value)))
+
+const invalidItems = computed(() =>
+  props.modelValue.filter(v => !validValues.value.has(v)),
+)
+
+const totalCount = computed(() => props.options.length + invalidItems.value.length)
+
+const showSearch = computed(
+  () => props.searchable && totalCount.value >= props.searchThreshold,
+)
+
+const filteredInvalidItems = computed(() => {
+  if (!showSearch.value || !searchQuery.value.trim()) {
+    return invalidItems.value
+  }
+  return invalidItems.value.filter((item) =>
+    matchesSearchQuery(searchQuery.value, item),
+  )
+})
+
+const filteredOptions = computed(() => {
+  if (!showSearch.value || !searchQuery.value.trim()) {
+    return props.options
+  }
+
+  return props.options.filter((item) =>
+    matchesSearchQuery(searchQuery.value, item.label, item.value),
+  )
+})

 const displayText = computed(() => {
  if (props.modelValue.length === 0) return props.placeholder
  if (props.modelValue.length <= 2) {
    return props.modelValue
-      .map(v => props.options.find(o => o.value === v)?.label ?? v)
+      .map((v) => props.options.find((o) => o.value === v)?.label ?? v)
      .join(', ')
  }
  return `已选择 ${props.modelValue.length} 项`
 })

+watch(isOpen, (open) => {
+  if (!open) {
+    searchQuery.value = ''
+  }
+})
+
 function toggle(value: string) {
  const newValue = [...props.modelValue]
  const index = newValue.indexOf(value)
@@ -107,4 +201,8 @@ function toggle(value: string) {
  }
  emit('update:modelValue', newValue)
 }
+
+function remove(value: string) {
+  emit('update:modelValue', props.modelValue.filter(v => v !== value))
+}
 </script>
@@ -3,39 +3,24 @@
    <Select
      v-model="selectedPreset"
    >
-      <SelectTrigger class="h-8 w-32 text-xs border-border/60">
+      <SelectTrigger :class="presetTriggerClass">
        <SelectValue placeholder="选择时间段" />
      </SelectTrigger>
      <SelectContent>
        <SelectItem value="today">
          今天
        </SelectItem>
-        <SelectItem value="yesterday">
-          昨天
-        </SelectItem>
        <SelectItem value="last7days">
          最近7天
        </SelectItem>
        <SelectItem value="last30days">
          最近30天
        </SelectItem>
-        <SelectItem value="last90days">
-          最近90天
+        <SelectItem value="last180days">
+          最近180天
        </SelectItem>
-        <SelectItem value="this_week">
-          本周
-        </SelectItem>
-        <SelectItem value="last_week">
-          上周
-        </SelectItem>
-        <SelectItem value="this_month">
-          本月
-        </SelectItem>
-        <SelectItem value="last_month">
-          上月
-        </SelectItem>
-        <SelectItem value="this_year">
-          今年
+        <SelectItem value="last1year">
+          最近一年
        </SelectItem>
        <SelectItem value="custom">
          自定义
@@ -50,13 +35,13 @@
      <Input
        v-model="startDate"
        type="date"
-        class="h-8 w-36 text-xs border-border/60"
+        :class="dateInputClass"
      />
      <span class="text-xs text-muted-foreground">至</span>
      <Input
        v-model="endDate"
        type="date"
-        class="h-8 w-36 text-xs border-border/60"
+        :class="dateInputClass"
      />
    </div>

@@ -64,10 +49,16 @@
      v-if="showGranularity"
      v-model="selectedGranularity"
    >
-      <SelectTrigger class="h-8 w-24 text-xs border-border/60">
+      <SelectTrigger :class="granularityTriggerClass">
        <SelectValue placeholder="粒度" />
      </SelectTrigger>
      <SelectContent>
+        <SelectItem
+          v-if="includeAutoGranularity"
+          value="auto"
+        >
+          自动
+        </SelectItem>
        <SelectItem
          v-if="allowHourly && canUseHourly"
          value="hour"
@@ -100,26 +91,51 @@ import {
 } from '@/components/ui'
 import type { DateRangeParams } from '@/features/usage/types'

+type PickerGranularity = NonNullable<DateRangeParams['granularity']>
+
 const props = defineProps<{
  modelValue: DateRangeParams
  showGranularity?: boolean
  allowHourly?: boolean
+  includeAutoGranularity?: boolean
+  compact?: boolean
 }>()

 const emit = defineEmits<{
  'update:modelValue': [value: DateRangeParams]
 }>()

-const selectedPreset = ref(props.modelValue.preset || 'last7days')
+function getDefaultGranularity(): PickerGranularity {
+  return props.includeAutoGranularity ? 'auto' : 'day'
+}
+
+const selectedPreset = ref(props.modelValue.preset || 'today')
 const startDate = ref(props.modelValue.start_date || '')
 const endDate = ref(props.modelValue.end_date || '')
-const selectedGranularity = ref(props.modelValue.granularity || 'day')
+const selectedGranularity = ref<PickerGranularity>(props.modelValue.granularity || getDefaultGranularity())

 const showGranularity = computed(() => props.showGranularity !== false)
 const allowHourly = computed(() => props.allowHourly === true)
+const includeAutoGranularity = computed(() => props.includeAutoGranularity === true)
+const compact = computed(() => props.compact === true)
+const presetTriggerClass = computed(() =>
+  compact.value
+    ? 'h-8 w-28 text-xs border-border/60'
+    : 'h-8 w-32 text-xs border-border/60',
+)
+const dateInputClass = computed(() =>
+  compact.value
+    ? 'h-8 w-32 text-xs border-border/60'
+    : 'h-8 w-36 text-xs border-border/60',
+)
+const granularityTriggerClass = computed(() =>
+  compact.value
+    ? 'h-8 w-20 text-xs border-border/60'
+    : 'h-8 w-24 text-xs border-border/60',
+)

 const canUseHourly = computed(() => {
-  if (selectedPreset.value === 'today' || selectedPreset.value === 'yesterday') return true
+  if (selectedPreset.value === 'today') return true
  if (selectedPreset.value === 'custom' && startDate.value && endDate.value) {
    return startDate.value === endDate.value
  }
@@ -165,7 +181,7 @@ watch(() => props.modelValue, (value) => {
  if (value.preset) selectedPreset.value = value.preset
  if (value.start_date !== undefined) startDate.value = value.start_date || ''
  if (value.end_date !== undefined) endDate.value = value.end_date || ''
-  if (value.granularity) selectedGranularity.value = value.granularity
+  selectedGranularity.value = value.granularity || getDefaultGranularity()
  // 同步更新 lastEmittedValue，避免外部设置值后触发重复 emit
  lastEmittedValue = getValueKey(value)
 }, { deep: true })
@@ -173,7 +189,7 @@ watch(() => props.modelValue, (value) => {
 watch([selectedPreset, startDate, endDate, selectedGranularity], () => {
  if (!allowHourly.value || !canUseHourly.value) {
    if (selectedGranularity.value === 'hour') {
-      selectedGranularity.value = 'day'
+      selectedGranularity.value = includeAutoGranularity.value ? 'auto' : 'day'
    }
  }

@@ -9,6 +9,5 @@ export { default as AlertDialog } from './AlertDialog.vue'
 export { default as LoadingState } from './LoadingState.vue'

 // 表单组件
-export { default as ModelMultiSelect } from './ModelMultiSelect.vue'
 export { default as MultiSelect } from './MultiSelect.vue'
 export { default as TimeRangePicker } from './TimeRangePicker.vue'
@@ -13,8 +13,11 @@
        <p class="mt-0.5">
          {{ tooltip.day.requests }} 次请求 · {{ formatTokens(tooltip.day.total_tokens) }}
        </p>
-        <p class="text-[11px] text-muted-foreground">
-          成本 {{ formatCurrency(tooltip.day.total_cost) }}
+        <p class="text-[11px] text-muted-foreground whitespace-nowrap">
+          {{ valueLabel }} {{ formatCurrency(tooltip.day.total_cost) }}
+          <template v-if="actualValueLabel && tooltip.day.actual_total_cost !== undefined">
+            · {{ actualValueLabel }} {{ formatCurrency(tooltip.day.actual_total_cost) }}
+          </template>
        </p>
      </div>
    </Teleport>
@@ -90,7 +93,7 @@
          class="flex items-center invisible"
        >周六</span>
      </div>
-      <div class="flex-1 min-w-[200px]">
+      <div class="flex-1 min-w-0">
        <div
          ref="heatmapWrapper"
          class="relative block w-full"
@@ -100,12 +103,12 @@
            :style="horizontalGapStyle"
          >
            <div
-              v-for="(week, weekIndex) in weekColumns"
+              v-for="(week, weekIndex) in visibleWeekColumns"
              :key="`month-${weekIndex}`"
              :style="monthCellStyle"
              class="text-center"
            >
-              <span v-if="monthMarkers[weekIndex]">{{ monthMarkers[weekIndex] }}</span>
+              <span v-if="visibleMonthMarkers[weekIndex]">{{ visibleMonthMarkers[weekIndex] }}</span>
            </div>
          </div>
          <div
@@ -113,7 +116,7 @@
            :style="horizontalGapStyle"
          >
            <div
-              v-for="(week, weekIndex) in weekColumns"
+              v-for="(week, weekIndex) in visibleWeekColumns"
              :key="weekIndex"
              class="flex flex-col"
              :style="verticalGapStyle"
@@ -161,11 +164,15 @@ const props = withDefaults(defineProps<{
  title?: string
  subtitle?: string
  showHeader?: boolean
+  valueLabel?: string
+  actualValueLabel?: string
 }>(), {
  data: undefined,
  title: undefined,
  subtitle: undefined,
-  showHeader: true
+  showHeader: true,
+  valueLabel: '费用',
+  actualValueLabel: undefined,
 })

 const legendLevels = [0.08, 0.25, 0.45, 0.65, 0.85]
@@ -270,7 +277,47 @@ const monthMarkers = computed(() => {
    if (month === lastMonth) {
      return
    }
-    markers[index] = `${month + 1}月`
+    markers[index] = String(month + 1)
+    lastMonth = month
+  })
+
+  return markers
+})
+
+const MIN_CELL_SIZE = 6
+
+// 根据容器宽度计算最多能放多少周，然后从后面截取（保留最新数据）
+const maxVisibleWeeks = computed(() => {
+  const total = weekColumns.value.length
+  if (!heatmapWidth.value || !total) return total
+  // 先尝试用理想尺寸放全部
+  const totalGapAll = Math.max(total - 1, 0) * cellGap.value
+  const rawSizeAll = (heatmapWidth.value - totalGapAll) / total
+  if (rawSizeAll >= MIN_CELL_SIZE) return total
+  // 放不下，算容器最多能容纳几周（用最小尺寸）
+  // n * MIN_CELL_SIZE + (n-1) * gap <= width
+  const maxN = Math.floor((heatmapWidth.value + cellGap.value) / (MIN_CELL_SIZE + cellGap.value))
+  return Math.max(1, Math.min(maxN, total))
+})
+
+const visibleWeekColumns = computed(() => {
+  const all = weekColumns.value
+  const n = maxVisibleWeeks.value
+  if (n >= all.length) return all
+  return all.slice(all.length - n)
+})
+
+const visibleMonthMarkers = computed(() => {
+  const markers: Record<number, string> = {}
+  const columns = visibleWeekColumns.value
+  let lastMonth: number | null = null
+
+  columns.forEach((week, index) => {
+    const firstValid = week.find((day): day is DayWithMeta => day !== null)
+    if (!firstValid) return
+    const month = firstValid.dateObj.getUTCMonth()
+    if (month === lastMonth) return
+    markers[index] = String(month + 1)
    lastMonth = month
  })

@@ -282,7 +329,7 @@ let mediaQuery: MediaQueryList | null = null
 let mediaQueryHandler: ((event?: MediaQueryListEvent) => void) | null = null

 const recalcCellSize = () => {
-  const columnCount = weekColumns.value.length
+  const columnCount = visibleWeekColumns.value.length
  if (!columnCount || !heatmapWidth.value) {
    return
  }
@@ -290,12 +337,12 @@ const recalcCellSize = () => {
  const totalGap = Math.max(columnCount - 1, 0) * cellGap.value
  const availableSpace = Math.max(heatmapWidth.value - totalGap, 0)
  const rawSize = availableSpace / columnCount
-  // 自适应尺寸，最小 6px
-  cellSize.value = Math.max(6, rawSize)
+  // 当所有列都能放下时，让格子继续随容器放大，避免宽屏下出现大块留白。
+  cellSize.value = Math.max(MIN_CELL_SIZE, rawSize)
 }

 watch(
-  [() => heatmapWidth.value, () => weekColumns.value.length, () => cellGap.value],
+  [() => heatmapWidth.value, () => visibleWeekColumns.value.length, () => cellGap.value],
  () => {
    recalcCellSize()
  },
@@ -345,7 +392,7 @@ onBeforeUnmount(() => {

 function handleHover(day: ActivityHeatmapDay, event: MouseEvent) {
  const cellRect = (event.currentTarget as HTMLElement).getBoundingClientRect()
-  const tooltipWidth = 200
+  const tooltipWidth = 280
  const tooltipHeight = 72

  // Calculate horizontal position (centered on cell)
@@ -402,10 +449,10 @@ function getCellStyle(requests: number) {

 function buildTooltip(day: ActivityHeatmapDay): string {
  const dateLabel = day.date
-  const costLabel = formatCurrency(day.total_cost || 0)
+  const costLabel = `${props.valueLabel} ${formatCurrency(day.total_cost || 0)}`
  const parts = [`${dateLabel}`, `${day.requests} 次请求`, `${formatTokens(day.total_tokens)} tokens`, costLabel]
-  if (day.actual_total_cost !== undefined) {
-    parts.push(`倍率: ${formatCurrency(day.actual_total_cost)}`)
+  if (props.actualValueLabel && day.actual_total_cost !== undefined) {
+    parts.push(`${props.actualValueLabel} ${formatCurrency(day.actual_total_cost)}`)
  }
  return parts.join(' · ')
 }
@@ -1,38 +0,0 @@
-<template>
-  <Card class="p-4 space-y-2">
-    <div class="text-xs text-muted-foreground">
-      {{ label }}
-    </div>
-    <div class="text-lg font-semibold">
-      {{ value }}
-    </div>
-    <div
-      class="text-xs"
-      :class="changeClass"
-    >
-      <span v-if="changePercent !== null">{{ changePercent }}%</span>
-      <span v-else>--</span>
-      <span class="ml-1 text-muted-foreground">vs 对比期</span>
-    </div>
-  </Card>
-</template>
-
-<script setup lang="ts">
-import { computed } from 'vue'
-import { Card } from '@/components/ui'
-
-interface Props {
-  label: string
-  value: string
-  changePercent: number | null
-}
-
-const props = defineProps<Props>()
-
-const changeClass = computed(() => {
-  if (props.changePercent === null) return 'text-muted-foreground'
-  if (props.changePercent > 0) return 'text-emerald-500'
-  if (props.changePercent < 0) return 'text-rose-500'
-  return 'text-muted-foreground'
-})
-</script>
@@ -1,93 +0,0 @@
-<template>
-  <div class="space-y-3">
-    <div class="flex items-center justify-between">
-      <h3 class="text-sm font-semibold">
-        {{ title }}
-      </h3>
-      <span
-        v-if="subtitle"
-        class="text-xs text-muted-foreground"
-      >{{ subtitle }}</span>
-    </div>
-    <div
-      v-if="loading"
-      class="p-6"
-    >
-      <LoadingState />
-    </div>
-    <div
-      v-else
-      class="h-[280px]"
-    >
-      <LineChart
-        :data="chartData"
-        :options="chartOptions"
-      />
-    </div>
-  </div>
-</template>
-
-<script setup lang="ts">
-import { computed } from 'vue'
-import LineChart from '@/components/charts/LineChart.vue'
-import { LoadingState } from '@/components/common'
-import { formatCurrency } from '@/utils/format'
-
-interface Props {
-  title: string
-  subtitle?: string
-  history: Array<{ date: string; total_cost: number }>
-  forecast: Array<{ date: string; total_cost: number }>
-  loading?: boolean
-}
-
-const props = withDefaults(defineProps<Props>(), {
-  subtitle: undefined,
-  loading: false
-})
-
-const labels = computed(() => [
-  ...props.history.map(item => item.date),
-  ...props.forecast.map(item => item.date)
-])
-
-const chartData = computed(() => {
-  const historyValues = props.history.map(item => item.total_cost)
-  const forecastValues = props.forecast.map(item => item.total_cost)
-  return {
-    labels: labels.value,
-    datasets: [
-      {
-        label: '实际成本',
-        data: historyValues.concat(new Array(forecastValues.length).fill(null)),
-        borderColor: 'rgb(59, 130, 246)',
-        backgroundColor: 'rgba(59, 130, 246, 0.15)',
-        tension: 0.25,
-        pointRadius: 2
-      },
-      {
-        label: '预测成本',
-        data: new Array(historyValues.length).fill(null).concat(forecastValues),
-        borderColor: 'rgb(234, 179, 8)',
-        backgroundColor: 'rgba(234, 179, 8, 0.15)',
-        borderDash: [6, 4],
-        tension: 0.25,
-        pointRadius: 2
-      }
-    ]
-  }
-})
-
-const chartOptions = computed(() => ({
-  plugins: {
-    tooltip: {
-      callbacks: {
-        label: (context: { parsed?: { y?: number }; dataset: { label?: string } }) => {
-          const value = context.parsed?.y ?? 0
-          return `${context.dataset.label}: ${formatCurrency(value)}`
-        }
-      }
-    }
-  }
-}))
-</script>
@@ -1,14 +1,34 @@
 <template>
-  <div class="space-y-3">
-    <div class="flex items-center justify-between">
-      <h3 class="text-sm font-semibold">
-        {{ title }}
-      </h3>
-      <span
-        v-if="subtitle"
-        class="text-xs text-muted-foreground"
-      >{{ subtitle }}</span>
+  <div class="space-y-4">
+    <div class="flex flex-col gap-3 sm:flex-row sm:items-start sm:justify-between">
+      <div>
+        <h3 class="text-sm font-semibold">
+          {{ title }}
+        </h3>
+        <p
+          v-if="subtitle"
+          class="mt-1 text-[11px] text-muted-foreground"
+        >
+          {{ subtitle }}
+        </p>
+      </div>
+
+      <div
+        v-if="!unavailable"
+        class="flex flex-wrap gap-2"
+      >
+        <div class="rounded-full border border-rose-500/20 bg-rose-500/8 px-2.5 py-1 text-[10px] font-medium text-rose-700 dark:text-rose-300">
+          总错误 {{ totalErrors }}
+        </div>
+        <div
+          v-if="topCategory"
+          class="rounded-full border border-border/70 bg-muted/20 px-2.5 py-1 text-[10px] font-medium text-muted-foreground"
+        >
+          TOP {{ topCategory.label }} · {{ topCategory.count }}
+        </div>
+      </div>
    </div>
+
    <div
      v-if="loading"
      class="p-6"
@@ -16,57 +36,116 @@
      <LoadingState />
    </div>
    <div
-      v-else
-      class="h-[260px]"
+      v-else-if="unavailable"
+      class="flex h-[272px] items-center justify-center rounded-2xl border border-dashed border-border/70 bg-muted/10 text-xs text-muted-foreground"
    >
-      <DoughnutChart
-        :data="chartData"
-        :options="chartOptions"
-      />
+      {{ unavailableText }}
+    </div>
+    <div
+      v-else-if="!hasData"
+      class="flex h-[272px] items-center justify-center rounded-2xl border border-dashed border-border/70 bg-muted/10 text-xs text-muted-foreground"
+    >
+      暂无错误分布数据
+    </div>
+    <div
+      v-else
+      class="rounded-2xl border border-border/60 bg-background p-3"
+    >
+      <div class="mx-auto w-full max-w-[320px]">
+        <DoughnutChart
+          :data="chartData"
+          :options="chartOptions"
+          :show-legend="true"
+          :height="248"
+        />
+      </div>
    </div>
  </div>
 </template>

 <script setup lang="ts">
 import { computed } from 'vue'
+import type { ChartData, ChartOptions } from 'chart.js'
 import DoughnutChart from '@/components/charts/DoughnutChart.vue'
 import { LoadingState } from '@/components/common'
-import type { ErrorDistributionItem } from '@/api/admin'
+
+interface ErrorDistributionItem {
+  category: string
+  label: string
+  count: number
+}

 interface Props {
  title: string
  subtitle?: string
  distribution: ErrorDistributionItem[]
  loading?: boolean
+  unavailable?: boolean
+  unavailableText?: string
 }

 const props = withDefaults(defineProps<Props>(), {
  subtitle: undefined,
-  loading: false
+  loading: false,
+  unavailable: false,
+  unavailableText: '数据暂不可用',
 })

-const chartData = computed(() => ({
-  labels: props.distribution.map(item => item.category),
+const chartColors = [
+  'rgba(226, 85, 118, 0.88)',
+  'rgba(84, 135, 237, 0.86)',
+  'rgba(225, 165, 52, 0.84)',
+  'rgba(68, 187, 132, 0.82)',
+  'rgba(148, 163, 184, 0.8)',
+  'rgba(145, 118, 214, 0.82)',
+]
+
+const totalErrors = computed(() => props.distribution.reduce((sum, item) => sum + item.count, 0))
+const hasData = computed(() => totalErrors.value > 0)
+const distributionWithShare = computed(() => props.distribution.map(item => ({
+  ...item,
+  label: item.label?.trim() || item.category,
+  share: totalErrors.value > 0 ? `${((item.count / totalErrors.value) * 100).toFixed(1)}%` : '0.0%',
+})))
+const topCategory = computed(() => (
+  distributionWithShare.value.slice().sort((left, right) => right.count - left.count)[0] ?? null
+))
+
+const chartData = computed<ChartData<'doughnut'>>(() => ({
+  labels: distributionWithShare.value.map(item => item.label),
  datasets: [
    {
-      data: props.distribution.map(item => item.count),
-      backgroundColor: [
-        'rgba(239, 68, 68, 0.7)',
-        'rgba(59, 130, 246, 0.7)',
-        'rgba(234, 179, 8, 0.7)',
-        'rgba(34, 197, 94, 0.7)',
-        'rgba(148, 163, 184, 0.7)'
-      ],
-      borderWidth: 0
-    }
-  ]
+      data: distributionWithShare.value.map(item => item.count),
+      backgroundColor: distributionWithShare.value.map((_, index) => chartColors[index % chartColors.length]),
+      borderColor: 'rgba(255,255,255,0.92)',
+      borderWidth: 2,
+      hoverOffset: 6,
+    },
+  ],
 }))

-const chartOptions = computed(() => ({
+const chartOptions = computed<ChartOptions<'doughnut'>>(() => ({
+  cutout: '68%',
  plugins: {
    legend: {
-      position: 'bottom' as const
-    }
-  }
+      position: 'bottom',
+      align: 'center',
+      labels: {
+        usePointStyle: true,
+        pointStyle: 'circle',
+        boxWidth: 10,
+        padding: 14,
+      },
+    },
+    tooltip: {
+      callbacks: {
+        label: (context) => {
+          const value = Number(context.raw ?? 0)
+          const percentage = totalErrors.value > 0 ? ((value / totalErrors.value) * 100).toFixed(1) : '0.0'
+          return `${context.label ?? '错误'}: ${value} (${percentage}%)`
+        },
+      },
+    },
+  },
 }))
 </script>
@@ -1,34 +1,30 @@
 <template>
  <TableCard :title="title">
    <template #actions>
-      <Select
-        v-if="showMetricSelect"
-        :model-value="metric"
-        @update:model-value="emitMetric"
+      <div
+        v-if="isRefreshing"
+        class="inline-flex h-7 items-center rounded-lg border border-border/70 bg-background px-2.5 text-[10px] text-muted-foreground"
      >
-        <SelectTrigger class="h-8 text-xs w-28">
-          <SelectValue placeholder="指标" />
-        </SelectTrigger>
-        <SelectContent>
-          <SelectItem value="requests">
-            请求数
-          </SelectItem>
-          <SelectItem value="tokens">
-            Tokens
-          </SelectItem>
-          <SelectItem value="cost">
-            成本
-          </SelectItem>
-        </SelectContent>
-      </Select>
+        更新中
+      </div>
    </template>

    <div
-      v-if="loading"
+      v-if="showLoadingState"
      class="p-6"
    >
      <LoadingState />
    </div>
+    <div
+      v-else-if="showUnavailableState"
+      class="p-6"
+    >
+      <EmptyState
+        type="error"
+        :title="unavailableTitle"
+        :description="unavailableDescription"
+      />
+    </div>
    <div
      v-else-if="items.length === 0"
      class="p-6"
@@ -38,20 +34,29 @@
        description="当前时间范围内没有统计结果"
      />
    </div>
-    <Table v-else>
+    <Table
+      v-else
+      class="table-fixed"
+      :class="{ 'opacity-60 transition-opacity': isRefreshing }"
+    >
      <TableHeader>
        <TableRow>
-          <TableHead class="w-16">
+          <TableHead class="w-14 px-2.5">
            排名
          </TableHead>
-          <TableHead>名称</TableHead>
-          <TableHead class="text-right">
+          <TableHead class="w-[34%] px-2.5">
+            名称
+          </TableHead>
+          <TableHead class="w-[15%] px-2.5 text-right">
            请求数
          </TableHead>
-          <TableHead class="text-right">
+          <TableHead class="w-[18%] px-2.5 text-right">
            Tokens
          </TableHead>
-          <TableHead class="text-right">
+          <TableHead class="w-[16%] px-2.5 text-right">
+            费用
+          </TableHead>
+          <TableHead class="w-[17%] px-2.5 text-right">
            成本
          </TableHead>
        </TableRow>
@@ -60,68 +65,118 @@
        <TableRow
          v-for="item in items"
          :key="item.id"
+          class="cursor-pointer"
+          :class="item.id === selectedId
+            ? 'bg-[linear-gradient(90deg,rgba(232,145,89,0.16),rgba(232,145,89,0.05),transparent)] shadow-[inset_4px_0_0_0_rgba(232,145,89,0.9)] hover:bg-[linear-gradient(90deg,rgba(232,145,89,0.2),rgba(232,145,89,0.06),transparent)]'
+            : 'hover:bg-muted/25'"
+          @click="emit('select', item.id)"
        >
-          <TableCell class="font-medium">
-            {{ item.rank }}
+          <TableCell class="px-2.5 py-2.5">
+            <div
+              class="inline-flex h-7 min-w-7 items-center justify-center rounded-full border px-2 text-[11px] font-semibold tabular-nums transition-colors"
+              :class="item.id === selectedId
+                ? 'border-amber-500/70 bg-amber-500/15 text-amber-700 dark:text-amber-300'
+                : 'border-border/70 bg-muted/35 text-muted-foreground'"
+            >
+              {{ item.rank }}
+            </div>
          </TableCell>
-          <TableCell>{{ item.name }}</TableCell>
-          <TableCell class="text-right">
-            {{ item.requests }}
+          <TableCell class="px-2.5 py-2.5">
+            <div class="min-w-0">
+              <div
+                class="truncate text-sm font-medium"
+                :title="item.name"
+              >
+                {{ item.name }}
+              </div>
+              <div class="mt-0.5 truncate text-[10px] text-muted-foreground">
+                费用优先
+              </div>
+            </div>
          </TableCell>
-          <TableCell class="text-right">
+          <TableCell class="px-2.5 py-2.5 text-right text-[11px] tabular-nums">
+            {{ item.requests.toLocaleString() }}
+          </TableCell>
+          <TableCell class="px-2.5 py-2.5 text-right text-[11px] tabular-nums">
            {{ formatTokens(item.tokens) }}
          </TableCell>
-          <TableCell class="text-right">
+          <TableCell class="px-2.5 py-2.5 text-right text-[11px] font-medium tabular-nums text-amber-700 dark:text-amber-300">
            {{ formatCurrency(item.cost) }}
          </TableCell>
+          <TableCell class="px-2.5 py-2.5 text-right text-[11px] tabular-nums text-muted-foreground">
+            {{ formatCurrency(item.actualCost) }}
+          </TableCell>
        </TableRow>
      </TableBody>
    </Table>
+
+    <template v-if="showPagination" #pagination>
+      <Pagination
+        :current="currentPage"
+        :total="totalItems"
+        :page-size="pageSize"
+        :show-page-size-selector="false"
+        @update:current="emit('update:currentPage', $event)"
+      />
+    </template>
  </TableCard>
 </template>

 <script setup lang="ts">
 import { computed } from 'vue'
 import { EmptyState, LoadingState } from '@/components/common'
-import { TableCard } from '@/components/ui'
-import {
-  Select,
-  SelectContent,
-  SelectItem,
-  SelectTrigger,
-  SelectValue,
-  Table,
-  TableBody,
-  TableCell,
-  TableHead,
-  TableHeader,
-  TableRow
-} from '@/components/ui'
+import { Pagination, Table, TableBody, TableCell, TableHead, TableHeader, TableRow, TableCard } from '@/components/ui'
 import { formatCurrency, formatTokens } from '@/utils/format'
-import type { LeaderboardItem } from '@/api/admin'
+
+interface LeaderboardItem {
+  rank: number
+  id: string
+  name: string
+  requests: number
+  tokens: number
+  cost: number
+  actualCost: number
+}

 interface Props {
  title: string
  items: LeaderboardItem[]
-  metric: 'requests' | 'tokens' | 'cost'
+  selectedId?: string
  loading?: boolean
-  showMetricSelect?: boolean
+  hasLoaded?: boolean
+  currentPage?: number
+  totalItems?: number
+  pageSize?: number
+  showPagination?: boolean
+  unavailable?: boolean
+  unavailableTitle?: string
+  unavailableDescription?: string
 }

 const props = withDefaults(defineProps<Props>(), {
+  selectedId: undefined,
  loading: false,
-  showMetricSelect: true
+  hasLoaded: false,
+  currentPage: 1,
+  totalItems: 0,
+  pageSize: 5,
+  showPagination: false,
+  unavailable: false,
+  unavailableTitle: '排行榜暂不可用',
+  unavailableDescription: '接口未返回结果，请稍后重试',
 })

 const emit = defineEmits<{
-  (e: 'update:metric', value: 'requests' | 'tokens' | 'cost'): void
+  (e: 'select', value: string): void
+  (e: 'update:currentPage', value: number): void
 }>()

-const metric = computed(() => props.metric)
-
-function emitMetric(value: string) {
-  if (value === 'requests' || value === 'tokens' || value === 'cost') {
-    emit('update:metric', value)
-  }
-}
+const selectedId = computed(() => props.selectedId)
+const showLoadingState = computed(() => props.loading && (!props.hasLoaded || props.items.length === 0))
+const showUnavailableState = computed(() => props.unavailable && !props.loading && props.items.length === 0)
+const isRefreshing = computed(() => props.loading && props.hasLoaded && props.items.length > 0)
+const currentPage = computed(() => props.currentPage)
+const totalItems = computed(() => props.totalItems)
+const pageSize = computed(() => props.pageSize)
+const showPagination = computed(() => props.showPagination && props.totalItems > props.pageSize)
 </script>
@@ -1,27 +1,59 @@
 <template>
-  <div class="space-y-3">
-    <div class="flex items-center justify-between">
-      <h3 class="text-sm font-semibold">
-        {{ title }}
-      </h3>
-      <span
-        v-if="subtitle"
-        class="text-xs text-muted-foreground"
-      >{{ subtitle }}</span>
+  <div class="space-y-4">
+    <div class="flex flex-col gap-3 sm:flex-row sm:items-start sm:justify-between">
+      <div>
+        <h3 class="text-sm font-semibold">
+          {{ title }}
+        </h3>
+        <p
+          v-if="subtitle"
+          class="mt-1 text-[11px] text-muted-foreground"
+        >
+          {{ subtitle }}
+        </p>
+      </div>
+
+      <div
+        v-if="summaryMetrics.length > 0 && !unavailable"
+        class="flex flex-wrap gap-2"
+      >
+        <div
+          v-for="metric in summaryMetrics"
+          :key="metric.label"
+          class="rounded-full border px-2.5 py-1 text-[10px] font-medium"
+          :class="metric.className"
+        >
+          {{ metric.label }} {{ metric.value }}
+        </div>
+      </div>
    </div>
+
    <div
      v-if="loading"
      class="p-6"
    >
      <LoadingState />
    </div>
+    <div
+      v-else-if="unavailable"
+      class="flex h-[272px] items-center justify-center rounded-2xl border border-dashed border-border/70 bg-muted/10 text-xs text-muted-foreground"
+    >
+      {{ unavailableText }}
+    </div>
+    <div
+      v-else-if="!hasData"
+      class="flex h-[272px] items-center justify-center rounded-2xl border border-dashed border-border/70 bg-muted/10 text-xs text-muted-foreground"
+    >
+      暂无延迟数据
+    </div>
    <div
      v-else
-      class="h-[260px]"
+      class="rounded-2xl border border-border/60 bg-background p-3"
    >
      <LineChart
        :data="chartData"
        :options="chartOptions"
+        :height="248"
      />
    </div>
  </div>
@@ -29,9 +61,19 @@

 <script setup lang="ts">
 import { computed } from 'vue'
+import type { ChartData, ChartOptions } from 'chart.js'
 import LineChart from '@/components/charts/LineChart.vue'
 import { LoadingState } from '@/components/common'
-import type { PercentileItem } from '@/api/admin'
+
+interface PercentileItem {
+  date: string
+  p50_response_time_ms?: number | null
+  p90_response_time_ms?: number | null
+  p99_response_time_ms?: number | null
+  p50_first_byte_time_ms?: number | null
+  p90_first_byte_time_ms?: number | null
+  p99_first_byte_time_ms?: number | null
+}

 interface Props {
  title: string
@@ -39,61 +81,159 @@ interface Props {
  series: PercentileItem[]
  mode: 'response' | 'ttfb'
  loading?: boolean
+  unavailable?: boolean
+  unavailableText?: string
 }

 const props = withDefaults(defineProps<Props>(), {
  subtitle: undefined,
-  loading: false
+  loading: false,
+  unavailable: false,
+  unavailableText: '数据暂不可用',
 })

-const labels = computed(() => props.series.map(item => item.date))
-
-// 毫秒转秒
 function msToSeconds(ms: number | null | undefined): number | null {
  if (ms == null) return null
-  return ms / 1000
+  return Number((ms / 1000).toFixed(3))
 }

-const chartData = computed(() => {
-  const p50Key = props.mode === 'response' ? 'p50_response_time_ms' : 'p50_first_byte_time_ms'
-  const p90Key = props.mode === 'response' ? 'p90_response_time_ms' : 'p90_first_byte_time_ms'
-  const p99Key = props.mode === 'response' ? 'p99_response_time_ms' : 'p99_first_byte_time_ms'
-
-  return {
-    labels: labels.value,
-    datasets: [
-      {
-        label: 'P50',
-        data: props.series.map(item => msToSeconds(item[p50Key])),
-        borderColor: 'rgb(59, 130, 246)',
-        tension: 0.25,
-        pointRadius: 2
-      },
-      {
-        label: 'P90',
-        data: props.series.map(item => msToSeconds(item[p90Key])),
-        borderColor: 'rgb(234, 179, 8)',
-        tension: 0.25,
-        pointRadius: 2
-      },
-      {
-        label: 'P99',
-        data: props.series.map(item => msToSeconds(item[p99Key])),
-        borderColor: 'rgb(239, 68, 68)',
-        tension: 0.25,
-        pointRadius: 2
-      }
-    ]
+function getPercentileValue(item: PercentileItem, percentile: '50' | '90' | '99') {
+  if (props.mode === 'response') {
+    if (percentile === '50') return msToSeconds(item.p50_response_time_ms)
+    if (percentile === '90') return msToSeconds(item.p90_response_time_ms)
+    return msToSeconds(item.p99_response_time_ms)
  }
-})

-const chartOptions = computed(() => ({
+  if (percentile === '50') return msToSeconds(item.p50_first_byte_time_ms)
+  if (percentile === '90') return msToSeconds(item.p90_first_byte_time_ms)
+  return msToSeconds(item.p99_first_byte_time_ms)
+}
+
+function getLatestMetricValue(percentile: '50' | '90' | '99') {
+  for (let index = props.series.length - 1; index >= 0; index -= 1) {
+    const value = getPercentileValue(props.series[index], percentile)
+    if (value != null) return value
+  }
+  return null
+}
+
+function formatSeconds(value: number | null) {
+  if (value == null) return '--'
+  if (value < 1) return `${Math.round(value * 1000)}ms`
+  return `${value.toFixed(2)}s`
+}
+
+const labels = computed(() => props.series.map(item => item.date?.slice(5) ?? ''))
+const hasData = computed(() => props.series.some(item => (
+  getPercentileValue(item, '50') != null
+  || getPercentileValue(item, '90') != null
+  || getPercentileValue(item, '99') != null
+)))
+
+const summaryMetrics = computed(() => ([
+  {
+    label: 'P50',
+    value: formatSeconds(getLatestMetricValue('50')),
+    className: 'border-sky-500/20 bg-sky-500/8 text-sky-700 dark:text-sky-300',
+  },
+  {
+    label: 'P90',
+    value: formatSeconds(getLatestMetricValue('90')),
+    className: 'border-amber-500/20 bg-amber-500/8 text-amber-700 dark:text-amber-300',
+  },
+  {
+    label: 'P99',
+    value: formatSeconds(getLatestMetricValue('99')),
+    className: 'border-rose-500/20 bg-rose-500/8 text-rose-700 dark:text-rose-300',
+  },
+]))
+
+const chartData = computed<ChartData<'line'>>(() => ({
+  labels: labels.value,
+  datasets: [
+    {
+      label: 'P50',
+      data: props.series.map(item => getPercentileValue(item, '50')),
+      borderColor: 'rgb(70, 136, 240)',
+      backgroundColor: 'rgba(70, 136, 240, 0.10)',
+      borderWidth: 2.4,
+      fill: false,
+      tension: 0.32,
+      pointRadius: 0,
+      pointHoverRadius: 4,
+      pointHitRadius: 14,
+    },
+    {
+      label: 'P90',
+      data: props.series.map(item => getPercentileValue(item, '90')),
+      borderColor: 'rgb(223, 166, 55)',
+      backgroundColor: 'rgba(223, 166, 55, 0.08)',
+      borderWidth: 2.1,
+      fill: false,
+      tension: 0.3,
+      pointRadius: 0,
+      pointHoverRadius: 4,
+      pointHitRadius: 14,
+    },
+    {
+      label: 'P99',
+      data: props.series.map(item => getPercentileValue(item, '99')),
+      borderColor: 'rgb(214, 92, 111)',
+      backgroundColor: 'rgba(214, 92, 111, 0.08)',
+      borderWidth: 2,
+      borderDash: [6, 4],
+      fill: false,
+      tension: 0.28,
+      pointRadius: 0,
+      pointHoverRadius: 4,
+      pointHitRadius: 14,
+    },
+  ],
+}))
+
+const chartOptions = computed<ChartOptions<'line'>>(() => ({
+  responsive: true,
+  maintainAspectRatio: false,
+  interaction: {
+    mode: 'index',
+    intersect: false,
+  },
+  plugins: {
+    legend: {
+      position: 'bottom',
+      labels: {
+        usePointStyle: true,
+        boxWidth: 10,
+        padding: 14,
+      },
+    },
+    tooltip: {
+      callbacks: {
+        title: items => props.series[items[0]?.dataIndex ?? -1]?.date ?? '',
+        label: context => `${context.dataset.label}: ${formatSeconds(Number(context.raw ?? 0))}`,
+      },
+    },
+  },
  scales: {
-    y: {
+    x: {
+      grid: {
+        color: 'rgba(148, 163, 184, 0.06)',
+      },
      ticks: {
-        callback: (value: number) => `${value}s`
-      }
-    }
-  }
+        color: 'rgb(100, 116, 139)',
+        maxRotation: 0,
+      },
+    },
+    y: {
+      beginAtZero: true,
+      grid: {
+        color: 'rgba(148, 163, 184, 0.07)',
+      },
+      ticks: {
+        color: 'rgb(100, 116, 139)',
+        callback: value => formatSeconds(Number(value)),
+      },
+    },
+  },
 }))
 </script>
@@ -1,81 +0,0 @@
-<template>
-  <Card class="p-4 space-y-4">
-    <div class="flex items-center justify-between">
-      <h3 class="text-sm font-semibold">
-        {{ title }}
-      </h3>
-      <span
-        v-if="subtitle"
-        class="text-xs text-muted-foreground"
-      >{{ subtitle }}</span>
-    </div>
-
-    <div
-      v-if="loading"
-      class="p-4"
-    >
-      <LoadingState />
-    </div>
-    <div
-      v-else-if="providers.length === 0"
-      class="p-4"
-    >
-      <EmptyState
-        title="暂无数据"
-        description="暂无月卡配额数据"
-      />
-    </div>
-    <div
-      v-else
-      class="space-y-4"
-    >
-      <div
-        v-for="provider in providers"
-        :key="provider.id"
-        class="space-y-2"
-      >
-        <div class="flex items-center justify-between text-xs">
-          <span class="font-medium">{{ provider.name }}</span>
-          <span class="text-muted-foreground">
-            {{ formatCurrency(provider.used_usd) }} / {{ formatCurrency(provider.quota_usd) }}
-          </span>
-        </div>
-        <div class="h-2 rounded-full bg-muted">
-          <div
-            class="h-2 rounded-full bg-primary"
-            :style="{ width: `${Math.min(provider.usage_percent, 100)}%` }"
-          />
-        </div>
-        <div class="flex items-center justify-between text-[11px] text-muted-foreground">
-          <span>剩余 {{ formatCurrency(provider.remaining_usd) }}</span>
-          <span v-if="provider.estimated_exhaust_at">
-            预计耗尽 {{ formatDate(provider.estimated_exhaust_at) }}
-          </span>
-        </div>
-      </div>
-    </div>
-  </Card>
-</template>
-
-<script setup lang="ts">
-import { Card } from '@/components/ui'
-import { EmptyState, LoadingState } from '@/components/common'
-import { formatCurrency } from '@/utils/format'
-import type { QuotaUsageProvider } from '@/api/admin'
-
-interface Props {
-  title: string
-  subtitle?: string
-  providers: QuotaUsageProvider[]
-  loading?: boolean
-}
-
-withDefaults(defineProps<Props>(), {
-  subtitle: undefined,
-  loading: false
-})
-
-function formatDate(value: string) {
-  return new Date(value).toLocaleDateString()
-}
-</script>
@@ -1,7 +1,4 @@
 export { default as ActivityHeatmap } from './ActivityHeatmap.vue'
 export { default as LeaderboardTable } from './LeaderboardTable.vue'
-export { default as CostForecastChart } from './CostForecastChart.vue'
-export { default as QuotaProgressCard } from './QuotaProgressCard.vue'
 export { default as PercentileChart } from './PercentileChart.vue'
 export { default as ErrorDistributionChart } from './ErrorDistributionChart.vue'
-export { default as ComparisonCard } from './ComparisonCard.vue'
@@ -109,6 +109,7 @@ const props = defineProps<{
  zIndex?: number // Custom z-index for nested dialogs (default: 60)
  noPadding?: boolean // Disable default content padding
  persistent?: boolean // Prevent closing on backdrop click
+  closeOnBackdrop?: boolean // Allow closing on backdrop click (default: true)
 }>()

 // Emits 定义
@@ -145,7 +146,7 @@ function handleClose() {

 // 处理背景点击
 function handleBackdropClick() {
-  if (!props.persistent) {
+  if (!props.persistent && props.closeOnBackdrop !== false) {
    handleClose()
  }
 }
@@ -6,7 +6,7 @@
    <input
      ref="inputRef"
      :class="inputClass"
-      :style="inputStyle"
+      :style="maskStyle"
      :value="modelValue"
      :type="effectiveType"
      :autocomplete="autocompleteAttr"
@@ -42,7 +42,7 @@
    v-else
    ref="inputRef"
    :class="inputClass"
-    :style="inputStyle"
+    :style="maskStyle"
    :value="modelValue"
    :type="effectiveType"
    :autocomplete="autocompleteAttr"
@@ -69,24 +69,6 @@ const emit = defineEmits<{
  'update:modelValue': [value: string]
 }>()

-// 开发环境警告：type="password" 已被弃用
-const warnPasswordType = import.meta.env.DEV
-  ? (() => {
-      let warned = false
-      return () => {
-        if (!warned) {
-          warned = true
-          // eslint-disable-next-line no-console
-          console.warn(
-            '[Input] type="password" 已被弃用，请使用 masked 属性代替。\n' +
-              '示例：<Input v-model="apiKey" masked />\n' +
-              'masked 属性使用 CSS 遮蔽而非 password 类型，不会触发浏览器密码管理器。'
-          )
-        }
-      }
-    })()
-  : () => {}
-
 interface Props {
  modelValue?: string | number
  class?: string
@@ -99,9 +81,8 @@ interface Props {
  size?: 'default' | 'sm'
  /**
   * 遮蔽显示内容（用于 API Key 等敏感信息）
-   * 使用 CSS -webkit-text-security 实现，不会触发浏览器密码管理器
+   * 始终使用 text 输入框，隐藏态通过样式进行遮蔽
   * 同时会显示一个小眼睛按钮用于切换显示/隐藏
-   * 注意：Firefox 不支持 -webkit-text-security，会显示明文（但仍可通过按钮切换）
   */
  masked?: boolean
  /**
@@ -134,17 +115,23 @@ const shouldDisableAutofill = computed(() => {
  return props.disableAutofill ?? false
 })

-// 始终使用 text 类型，永远不用 password
 const effectiveType = computed(() => {
-  const attrType = attrs.type as string | undefined
-  // 如果传入 password，强制转为 text（配合 masked 使用）
-  if (attrType === 'password') {
-    warnPasswordType()
+  const attrType = (attrs.type as string | undefined) ?? 'text'
+  if (props.masked) {
    return 'text'
  }
  return attrType
 })

+const maskStyle = computed(() => {
+  if (!props.masked || isVisible.value) {
+    return undefined
+  }
+  return {
+    WebkitTextSecurity: 'disc'
+  }
+})
+
 // 过滤掉 type 和 class 属性，因为我们会单独处理
 const filteredAttrs = computed(() => {
  const { type: _type, class: _class, ...rest } = attrs
@@ -182,16 +169,6 @@ const inputClass = computed(() =>
  )
 )

-// 当 masked 为 true 且未显示时，用 CSS 遮蔽文字
-const inputStyle = computed(() => {
-  if (props.masked && !isVisible.value) {
-    // 使用 -webkit-text-security（Chrome, Safari, Edge 支持）
-    // Firefox 不支持此属性，会显示明文，但仍可通过小眼睛按钮切换
-    return { '-webkit-text-security': 'disc' }
-  }
-  return undefined
-})
-
 function handleInput(event: Event) {
  const target = event.target as HTMLInputElement
  emit('update:modelValue', target.value)
@@ -10,7 +10,32 @@
      :align-offset="alignOffset"
    >
      <SelectViewport :class="viewportClass">
+        <div
+          v-if="showSearchInput"
+          class="sticky top-0 z-10 bg-card/95 px-1 pt-1 pb-2 backdrop-blur supports-[backdrop-filter]:bg-card/85"
+        >
+          <div class="relative">
+            <Search
+              class="pointer-events-none absolute left-3 top-1/2 h-3.5 w-3.5 -translate-y-1/2 text-muted-foreground"
+            />
+            <Input
+              ref="searchInputRef"
+              v-model="searchQuery"
+              :placeholder="searchPlaceholder"
+              class="h-9 rounded-xl border-border/60 bg-background/80 pl-9 pr-3 text-sm"
+              @keydown.stop
+            />
+          </div>
+        </div>
+
        <slot />
+
+        <div
+          v-if="showEmptyState"
+          class="px-3 py-2 text-sm text-muted-foreground"
+        >
+          未找到匹配项
+        </div>
      </SelectViewport>
    </SelectContentPrimitive>
  </SelectPortal>
@@ -22,9 +47,25 @@ import {
  SelectPortal,
  SelectViewport,
 } from 'radix-vue'
+import { Search } from 'lucide-vue-next'
+import {
+  computed,
+  inject,
+  nextTick,
+  onMounted,
+  onUnmounted,
+  provide,
+  ref,
+  watch,
+} from 'vue'
+import Input from './input.vue'
 import { cn } from '@/lib/utils'
-import { computed, inject } from 'vue'
 import { DIALOG_CONTEXT_KEY } from './dialog/context'
+import {
+  SELECT_SEARCH_CONTEXT_KEY,
+  type RegisteredSelectItem,
+} from './select-search-context'
+import { matchesSearchQuery, preloadPinyin } from '@/utils/search'

 interface Props {
  class?: string
@@ -34,6 +75,9 @@ interface Props {
  align?: 'start' | 'center' | 'end'
  alignOffset?: number
  disablePortal?: boolean
+  searchable?: boolean
+  searchThreshold?: number
+  searchPlaceholder?: string
 }

 const props = withDefaults(defineProps<Props>(), {
@@ -44,19 +88,105 @@ const props = withDefaults(defineProps<Props>(), {
  align: undefined,
  alignOffset: undefined,
  disablePortal: false,
+  searchable: true,
+  searchThreshold: 8,
+  searchPlaceholder: '输入关键词搜索...',
 })

 const isInsideDialog = inject(DIALOG_CONTEXT_KEY, false)
-const shouldDisablePortal = computed(() => props.disablePortal || isInsideDialog)
+const shouldDisablePortal = computed(
+  () => props.disablePortal || isInsideDialog,
+)
+const searchQuery = ref('')
+const searchInputRef = ref<InstanceType<typeof Input> | null>(null)
+const registeredItems = ref<Record<string, RegisteredSelectItem>>({})
+
+const hiddenValues = computed(() => {
+  const query = searchQuery.value.trim()
+  if (!query) return new Set<string>()
+
+  const hidden = new Set<string>()
+  for (const item of Object.values(registeredItems.value)) {
+    if (!matchesSearchQuery(query, item.value, item.text)) {
+      hidden.add(item.value)
+    }
+  }
+  return hidden
+})
+
+provide(SELECT_SEARCH_CONTEXT_KEY, {
+  searchQuery,
+  hiddenValues,
+  registerItem(id, item) {
+    registeredItems.value = {
+      ...registeredItems.value,
+      [id]: item,
+    }
+  },
+  updateItem(id, item) {
+    const currentItem = registeredItems.value[id]
+    if (!currentItem) {
+      return
+    }
+
+    registeredItems.value = {
+      ...registeredItems.value,
+      [id]: {
+        ...currentItem,
+        ...item,
+      },
+    }
+  },
+  unregisterItem(id) {
+    if (!(id in registeredItems.value)) {
+      return
+    }
+
+    const nextItems = { ...registeredItems.value }
+    delete nextItems[id]
+    registeredItems.value = nextItems
+  },
+})
+
+const itemCount = computed(() => Object.keys(registeredItems.value).length)
+const showSearchInput = computed(
+  () => props.searchable && itemCount.value >= props.searchThreshold,
+)
+const showEmptyState = computed(
+  () =>
+    showSearchInput.value &&
+    searchQuery.value.trim().length > 0 &&
+    hiddenValues.value.size === itemCount.value,
+)
+
+onMounted(() => {
+  if (props.searchable) {
+    preloadPinyin()
+  }
+})
+
+watch(showSearchInput, async (visible) => {
+  if (!visible) {
+    searchQuery.value = ''
+    return
+  }
+
+  await nextTick()
+  searchInputRef.value?.inputRef?.focus()
+})

 const contentClass = computed(() =>
  cn(
    'z-[200] max-h-96 min-w-[8rem] overflow-hidden rounded-2xl border border-border bg-card text-foreground shadow-2xl backdrop-blur-xl pointer-events-auto',
    'data-[state=open]:animate-in data-[state=closed]:animate-out data-[state=closed]:fade-out-0 data-[state=open]:fade-in-0 data-[state=closed]:zoom-out-95 data-[state=open]:zoom-in-95',
    'data-[side=bottom]:slide-in-from-top-2 data-[side=left]:slide-in-from-right-2 data-[side=right]:slide-in-from-left-2 data-[side=top]:slide-in-from-bottom-2',
-    props.class
-  )
+    props.class,
+  ),
 )

 const viewportClass = 'p-1 max-h-[var(--radix-select-content-available-height)]'
+
+onUnmounted(() => {
+  searchQuery.value = ''
+})
 </script>
@@ -1,40 +1,116 @@
 <script setup lang="ts">
-import { SelectItem as SelectItemPrimitive, SelectItemIndicator, SelectItemText } from 'radix-vue'
+import {
+  SelectItem as SelectItemPrimitive,
+  SelectItemIndicator,
+  SelectItemText,
+} from 'radix-vue'
 import { Check } from 'lucide-vue-next'
 import { cn } from '@/lib/utils'
-import { computed } from 'vue'
+import {
+  computed,
+  getCurrentInstance,
+  inject,
+  onBeforeUnmount,
+  onMounted,
+  useSlots,
+  watch,
+} from 'vue'
+import { SELECT_SEARCH_CONTEXT_KEY } from './select-search-context'

 interface Props {
  class?: string
  value: string
  disabled?: boolean
+  textValue?: string
 }

 const props = defineProps<Props>()
+const slots = useSlots()
+const searchContext = inject(SELECT_SEARCH_CONTEXT_KEY, null)
+const instance = getCurrentInstance()
+const itemId = instance?.uid
+  ? `select-item-${instance.uid}`
+  : `select-item-${Math.random().toString(36).slice(2, 10)}`

+function extractText(node: unknown): string {
+  if (typeof node === 'string' || typeof node === 'number') {
+    return String(node)
+  }
+
+  if (Array.isArray(node)) {
+    return node.map(extractText).join(' ')
+  }
+
+  if (node && typeof node === 'object') {
+    const vnode = node as { children?: unknown }
+    return extractText(vnode.children)
+  }
+
+  return ''
+}
+
+const normalizedText = computed(() => {
+  const slotText = extractText(slots.default?.()).replace(/\s+/g, ' ').trim()
+  return (props.textValue ?? slotText ?? props.value).trim()
+})
+
+const isHidden = computed(
+  () => searchContext?.hiddenValues.value.has(props.value) ?? false,
+)
 const itemClass = computed(() =>
  cn(
-    'relative flex w-full cursor-pointer select-none items-center rounded-lg py-1.5 pl-8 pr-2 text-sm outline-none',
+    'relative flex min-w-0 w-full cursor-pointer select-none items-center rounded-lg py-1.5 pl-8 pr-2 text-sm outline-none',
    'data-[highlighted]:bg-accent focus:bg-accent text-foreground',
    'transition-colors data-[disabled]:pointer-events-none data-[disabled]:opacity-50',
-    props.class
-  )
+    isHidden.value && 'hidden',
+    props.class,
+  ),
 )
+
+onMounted(() => {
+  searchContext?.registerItem(itemId, {
+    value: props.value,
+    text: normalizedText.value,
+  })
+})
+
+watch(
+  () => [props.value, normalizedText.value] as const,
+  ([value, text]) => {
+    searchContext?.updateItem(itemId, {
+      value,
+      text,
+    })
+  },
+)
+
+onBeforeUnmount(() => {
+  searchContext?.unregisterItem(itemId)
+})
 </script>

 <template>
  <SelectItemPrimitive
    :class="itemClass"
    :value="value"
-    :disabled="disabled"
+    :disabled="disabled || isHidden"
+    :text-value="normalizedText"
  >
    <span class="absolute left-2 flex h-3.5 w-3.5 items-center justify-center">
      <SelectItemIndicator>
        <Check class="h-4 w-4" />
      </SelectItemIndicator>
    </span>
-    <SelectItemText>
-      <slot />
+    <SelectItemText
+      as="div"
+      class="min-w-0 flex-1 overflow-hidden"
+    >
+      <span
+        class="block w-full min-w-0 truncate"
+        :title="normalizedText"
+      >
+        <slot />
+      </span>
    </SelectItemText>
  </SelectItemPrimitive>
 </template>
@@ -0,0 +1,17 @@
+import type { InjectionKey, Ref } from 'vue'
+
+export interface RegisteredSelectItem {
+  value: string
+  text: string
+}
+
+export interface SelectSearchContext {
+  searchQuery: Ref<string>
+  hiddenValues: Ref<Set<string>>
+  registerItem: (id: string, item: RegisteredSelectItem) => void
+  updateItem: (id: string, item: Partial<RegisteredSelectItem>) => void
+  unregisterItem: (id: string) => void
+}
+
+export const SELECT_SEARCH_CONTEXT_KEY: InjectionKey<SelectSearchContext> =
+  Symbol('select-search-context')
@@ -25,7 +25,7 @@ const triggerClass = computed(() =>
    :class="triggerClass"
    :disabled="disabled"
  >
-    <span class="truncate">
+    <span class="min-w-0 flex-1 truncate">
      <slot />
    </span>
    <ChevronDown class="h-4 w-4 opacity-50 pointer-events-none flex-shrink-0" />
@@ -0,0 +1,3 @@
+export { useLeaderboardData } from './useLeaderboardData'
+export type { LeaderboardDimension } from './useLeaderboardData'
+export { usePerformanceData } from './usePerformanceData'
@@ -0,0 +1,23 @@
+import type { AnalyticsLeaderboardItem } from '@/api/analytics'
+
+export interface LeaderboardTableRow {
+  rank: number
+  id: string
+  name: string
+  requests: number
+  tokens: number
+  cost: number
+  actualCost: number
+}
+
+export function mapAnalyticsLeaderboardItem(item: AnalyticsLeaderboardItem): LeaderboardTableRow {
+  return {
+    rank: item.rank,
+    id: item.id,
+    name: item.label,
+    requests: item.requests_total,
+    tokens: item.total_tokens,
+    cost: item.total_cost_usd,
+    actualCost: item.actual_total_cost_usd,
+  }
+}
@@ -0,0 +1,272 @@
+import { computed, onMounted, onUnmounted, ref, watch, type Ref } from 'vue'
+import { analyticsApi, type AnalyticsFilterOption, type AnalyticsLeaderboardItem, type AnalyticsTimeseriesBucket } from '@/api/analytics'
+import { buildTimeRangeParams, createLoader, createRequestGuard, useInjectedAnalyticsFilters } from '@/composables/useAnalyticsFilters'
+import { fillMissingTimeseriesBuckets } from '@/utils/analyticsTimeseries'
+import { isAnalyticsSingleDayRange, resolveAnalyticsAutoGranularity, type AnalyticsGranularityOption } from '@/utils/analyticsGranularity'
+import { mapAnalyticsLeaderboardItem, type LeaderboardTableRow } from './mappers'
+
+export type LeaderboardDimension = 'users' | 'api_keys'
+
+export interface LeaderboardSummary {
+  total_requests: number
+  total_tokens: number
+  total_cost: number
+  total_actual_cost: number
+  avg_response_time: number
+}
+
+function dimensionToEntity(dimension: LeaderboardDimension): 'user' | 'api_key' {
+  return dimension === 'users' ? 'user' : 'api_key'
+}
+
+export function useLeaderboardData(dimension: Ref<LeaderboardDimension>) {
+  const filters = useInjectedAnalyticsFilters()
+
+  const leaderboardItems = ref<AnalyticsLeaderboardItem[]>([])
+  const latestUserOptions = ref<AnalyticsFilterOption[]>([])
+  const latestApiKeyOptions = ref<AnalyticsFilterOption[]>([])
+  const leaderboardLoading = ref(false)
+  const leaderboardError = ref(false)
+  const leaderboardHasLoaded = ref(false)
+  const selectedId = ref<string>('')
+
+  const selectedSummary = ref<LeaderboardSummary | null>(null)
+  const selectedTrendBuckets = ref<AnalyticsTimeseriesBucket[]>([])
+  const panelLoading = ref(false)
+  const panelError = ref(false)
+  const panelHasLoaded = ref(false)
+
+  const leaderboardGuard = createRequestGuard()
+  const filterOptionsGuard = createRequestGuard()
+  const panelGuard = createRequestGuard()
+
+  const entityOptions = computed(() => (
+    dimension.value === 'users' ? latestUserOptions.value : latestApiKeyOptions.value
+  ))
+  const entityLabelMap = computed(() => (
+    new Map(entityOptions.value.map(option => [option.value, option.label]))
+  ))
+
+  const leaderboard = computed<LeaderboardTableRow[]>(() => (
+    leaderboardItems.value.map(item => ({
+      ...mapAnalyticsLeaderboardItem(item),
+      name: entityLabelMap.value.get(item.id) ?? item.label,
+    }))
+  ))
+
+  const selectedItem = computed(() => (
+    leaderboard.value.find(item => item.id === selectedId.value) ?? null
+  ))
+
+  const selectedLabel = computed(() => (
+    selectedItem.value?.name ?? entityLabelMap.value.get(selectedId.value) ?? null
+  ))
+
+  const resolvedTrendGranularity = computed(() => resolveAnalyticsAutoGranularity(filters.timeRange.value))
+  const baseTimeRangePayload = computed(() => ({
+    start_date: filters.timeRange.value.start_date,
+    end_date: filters.timeRange.value.end_date,
+    preset: filters.timeRange.value.preset,
+    timezone: filters.timeRange.value.timezone,
+    tz_offset_minutes: filters.timeRange.value.tz_offset_minutes,
+  }))
+  const trendTimeRangePayload = computed(() => ({
+    ...baseTimeRangePayload.value,
+    granularity: resolvedTrendGranularity.value,
+  }))
+
+  const currentTrendGranularitySelection = computed<AnalyticsGranularityOption>({
+    get: () => filters.timeRange.value.granularity || 'auto',
+    set: (value) => {
+      if (value === 'hour' && !isAnalyticsSingleDayRange(filters.timeRange.value)) {
+        value = 'auto'
+      }
+      filters.timeRange.value = {
+        ...filters.timeRange.value,
+        granularity: value,
+      }
+    },
+  })
+
+  function resetPanelData(invalidate = false) {
+    if (invalidate) panelGuard.invalidate()
+    panelLoading.value = false
+    panelError.value = false
+    selectedSummary.value = null
+    selectedTrendBuckets.value = []
+  }
+
+  async function loadFilterOptions() {
+    const requestId = filterOptionsGuard.next()
+    const response = await analyticsApi.getFilterOptions({
+      scope: { kind: 'global' },
+      time_range: baseTimeRangePayload.value,
+      filters: {},
+    }).catch(() => null)
+
+    if (filterOptionsGuard.isStale(requestId) || !response) return
+    latestUserOptions.value = response.users ?? []
+    latestApiKeyOptions.value = response.api_keys ?? []
+  }
+
+  async function loadLeaderboard() {
+    const requestId = leaderboardGuard.next()
+    leaderboardLoading.value = true
+    leaderboardError.value = false
+    try {
+      const response = await analyticsApi.getLeaderboard({
+        scope: { kind: 'global' },
+        time_range: baseTimeRangePayload.value,
+        filters: {
+          statuses: ['completed', 'failed', 'cancelled'],
+        },
+        entity: dimensionToEntity(dimension.value),
+        metric: 'total_cost_usd',
+        limit: 20,
+      })
+
+      if (leaderboardGuard.isStale(requestId)) return
+      leaderboardItems.value = response.items
+
+      const nextSelectedId = response.items.some(item => item.id === selectedId.value)
+        ? selectedId.value
+        : (response.items[0]?.id ?? '')
+
+      if (selectedId.value !== nextSelectedId) {
+        selectedId.value = nextSelectedId
+      }
+      leaderboardHasLoaded.value = true
+    } catch {
+      if (leaderboardGuard.isStale(requestId)) return
+      leaderboardError.value = true
+    } finally {
+      if (leaderboardGuard.isCurrent(requestId)) {
+        leaderboardLoading.value = false
+      }
+    }
+  }
+
+  async function loadPanel() {
+    if (!selectedId.value) {
+      resetPanelData(true)
+      return
+    }
+
+    const requestId = panelGuard.next()
+    panelLoading.value = true
+    panelError.value = false
+    try {
+      const scope = dimension.value === 'users'
+        ? { kind: 'user' as const, user_id: selectedId.value }
+        : { kind: 'api_key' as const, api_key_id: selectedId.value }
+
+      const basePayload = {
+        scope,
+        time_range: baseTimeRangePayload.value,
+        filters: {
+          statuses: ['completed', 'failed', 'cancelled'],
+        },
+      }
+
+      const [overview, timeseries] = await Promise.all([
+        analyticsApi.getOverview(basePayload),
+        analyticsApi.getTimeseries({
+          ...basePayload,
+          time_range: trendTimeRangePayload.value,
+        }),
+      ])
+
+      if (panelGuard.isStale(requestId)) return
+      selectedSummary.value = {
+        total_requests: overview.summary.requests_total,
+        total_tokens: overview.summary.total_tokens,
+        total_cost: overview.summary.total_cost_usd,
+        total_actual_cost: overview.summary.actual_total_cost_usd,
+        avg_response_time: overview.summary.avg_response_time_ms / 1000,
+      }
+      selectedTrendBuckets.value = fillMissingTimeseriesBuckets(
+        timeseries.buckets,
+        trendTimeRangePayload.value,
+      )
+      panelHasLoaded.value = true
+    } catch {
+      if (panelGuard.isStale(requestId)) return
+      panelError.value = true
+    } finally {
+      if (panelGuard.isCurrent(requestId)) {
+        panelLoading.value = false
+      }
+    }
+  }
+
+  const filterOptionsLoader = createLoader(loadFilterOptions)
+  const leaderboardLoader = createLoader(loadLeaderboard)
+  const panelLoader = createLoader(loadPanel)
+
+  watch(
+    () => [filters.timeRange.value.preset, filters.timeRange.value.start_date, filters.timeRange.value.end_date],
+    () => {
+      if (filters.timeRange.value.granularity === 'hour' && !isAnalyticsSingleDayRange(filters.timeRange.value)) {
+        filters.timeRange.value = {
+          ...filters.timeRange.value,
+          granularity: 'auto',
+        }
+      }
+    },
+  )
+
+  watch(
+    () => [baseTimeRangePayload.value, dimension.value],
+    () => {
+      filterOptionsLoader.schedule()
+      leaderboardLoader.schedule()
+    },
+    { deep: true },
+  )
+
+  watch(
+    () => [
+      selectedId.value,
+      dimension.value,
+      baseTimeRangePayload.value,
+      trendTimeRangePayload.value,
+    ],
+    () => panelLoader.schedule(),
+    { deep: true },
+  )
+
+  onMounted(() => {
+    void (async () => {
+      await filterOptionsLoader.execute()
+      await leaderboardLoader.execute()
+      await panelLoader.execute()
+    })()
+  })
+
+  onUnmounted(() => {
+    filterOptionsLoader.cleanup()
+    leaderboardLoader.cleanup()
+    panelLoader.cleanup()
+    filterOptionsGuard.invalidate()
+    leaderboardGuard.invalidate()
+    panelGuard.invalidate()
+  })
+
+  return {
+    leaderboard,
+    leaderboardLoading,
+    leaderboardError,
+    leaderboardHasLoaded,
+    selectedId,
+    selectedItem,
+    selectedLabel,
+    selectedSummary,
+    selectedTrendBuckets,
+    panelLoading,
+    panelError,
+    panelHasLoaded,
+    resolvedTrendGranularity,
+    currentTrendGranularitySelection,
+    reloadLeaderboard: leaderboardLoader.schedule,
+  }
+}
@@ -0,0 +1,98 @@
+import { ref, watch, onMounted, onUnmounted } from 'vue'
+import { analyticsApi, type AnalyticsErrorCategory, type AnalyticsErrorTrendItem, type AnalyticsPercentilePoint, type AnalyticsPerformanceResponse, type AnalyticsProviderHealthItem } from '@/api/analytics'
+import { auditApi, type MonitoringSystemStatus } from '@/api/audit'
+import { getHealthSummary, type HealthSummary } from '@/api/endpoints'
+import { createLoader, createRequestGuard, useInjectedAnalyticsFilters } from '@/composables/useAnalyticsFilters'
+
+const EMPTY_LATENCY: AnalyticsPerformanceResponse['latency'] = {
+  response_time_ms: { avg: 0, p50: null, p90: null, p99: null },
+  first_byte_time_ms: { avg: 0, p50: null, p90: null, p99: null },
+}
+
+export function usePerformanceData() {
+  const filters = useInjectedAnalyticsFilters()
+
+  const latency = ref<AnalyticsPerformanceResponse['latency']>(EMPTY_LATENCY)
+  const percentiles = ref<AnalyticsPercentilePoint[]>([])
+  const errorDistribution = ref<AnalyticsErrorCategory[]>([])
+  const errorTrend = ref<AnalyticsErrorTrendItem[]>([])
+  const errorTotal = ref(0)
+  const errorRate = ref(0)
+  const providerStatus = ref<AnalyticsProviderHealthItem[]>([])
+  const healthSummary = ref<HealthSummary | null>(null)
+  const systemStatus = ref<MonitoringSystemStatus | null>(null)
+  const loading = ref(false)
+  const loadError = ref(false)
+  const hasLoaded = ref(false)
+  const guard = createRequestGuard()
+
+  async function load() {
+    const requestId = guard.next()
+    loading.value = true
+    loadError.value = false
+    try {
+      const [response, summaryResponse, systemResponse] = await Promise.all([
+        analyticsApi.getPerformance({
+          scope: { kind: 'global' },
+          time_range: filters.getTimeRangeParams(),
+          filters: {
+            statuses: ['completed', 'failed', 'cancelled'],
+          },
+        }),
+        getHealthSummary().catch(() => null),
+        auditApi.getSystemStatus().catch(() => null),
+      ])
+
+      if (guard.isStale(requestId)) return
+      latency.value = response.latency
+      percentiles.value = response.percentiles
+      errorTotal.value = response.errors.total
+      errorRate.value = response.errors.rate
+      errorDistribution.value = response.errors.categories
+      errorTrend.value = response.errors.trend
+      providerStatus.value = response.provider_health
+      if (summaryResponse) {
+        healthSummary.value = summaryResponse
+      }
+      if (systemResponse) {
+        systemStatus.value = systemResponse
+      }
+      hasLoaded.value = true
+    } catch {
+      if (guard.isStale(requestId)) return
+      loadError.value = true
+    } finally {
+      if (guard.isCurrent(requestId)) {
+        loading.value = false
+      }
+    }
+  }
+
+  const loader = createLoader(load)
+
+  watch(
+    () => [
+      filters.getTimeRangeParams(),
+    ],
+    () => loader.schedule(),
+    { deep: true },
+  )
+
+  onMounted(() => { void loader.execute() })
+  onUnmounted(() => { loader.cleanup(); guard.invalidate() })
+
+  return {
+    latency,
+    percentiles,
+    errorDistribution,
+    errorTrend,
+    errorTotal,
+    errorRate,
+    providerStatus,
+    healthSummary,
+    systemStatus,
+    loading,
+    loadError,
+    hasLoaded,
+  }
+}
@@ -0,0 +1,144 @@
+import { inject, ref, watch } from 'vue'
+import { useRoute, useRouter } from 'vue-router'
+import type { DateRangeParams, PeriodValue } from '@/features/usage/types'
+import { getDateRangeFromPeriod } from '@/features/usage/composables'
+
+/**
+ * 构建时间范围 API 参数（从各页面提取的公共逻辑）
+ */
+export function buildTimeRangeParams(timeRange: DateRangeParams) {
+  return {
+    start_date: timeRange.start_date,
+    end_date: timeRange.end_date,
+    preset: timeRange.preset,
+    timezone: timeRange.timezone,
+    tz_offset_minutes: timeRange.tz_offset_minutes,
+    granularity: timeRange.granularity && timeRange.granularity !== 'auto'
+      ? timeRange.granularity
+      : 'day',
+  }
+}
+
+/**
+ * 通用防抖加载器工厂
+ * 封装 120ms debounce + in-flight promise 去重 + requestId 过期检查
+ */
+export function createLoader(loadFn: () => Promise<void>) {
+  let loadPromise: Promise<void> | null = null
+  let hasPending = false
+  let debounceTimer: ReturnType<typeof setTimeout> | null = null
+
+  function schedule() {
+    if (debounceTimer) clearTimeout(debounceTimer)
+    debounceTimer = setTimeout(() => {
+      debounceTimer = null
+      void execute()
+    }, 120)
+  }
+
+  async function execute() {
+    if (loadPromise) {
+      hasPending = true
+      return loadPromise
+    }
+    loadPromise = loadFn()
+      .finally(() => {
+        loadPromise = null
+        if (hasPending) {
+          hasPending = false
+          void execute()
+        }
+      })
+    return loadPromise
+  }
+
+  function cleanup() {
+    if (debounceTimer) {
+      clearTimeout(debounceTimer)
+      debounceTimer = null
+    }
+    hasPending = false
+    loadPromise = null
+  }
+
+  return { schedule, execute, cleanup }
+}
+
+/**
+ * 请求 ID 守卫：用于确保异步回调中的数据不被过期请求覆盖
+ */
+export function createRequestGuard() {
+  let id = 0
+  function next() { return ++id }
+  function isStale(requestId: number) { return requestId !== id }
+  function isCurrent(requestId: number) { return requestId === id }
+  function invalidate() { id++ }
+  return { next, isStale, isCurrent, invalidate }
+}
+
+export interface AnalyticsFiltersOptions {
+  defaultPreset?: PeriodValue
+  syncToUrl?: boolean
+}
+
+/**
+ * 统一的分析页面筛选状态
+ * - timeRange: 时间范围
+ * - modelFilter / providerFilter / apiKeyFilter / userFilter: 多维度筛选
+ * - activeTab: 当前 Tab
+ * - URL query 双向同步
+ */
+export function useAnalyticsFilters(options: AnalyticsFiltersOptions = {}) {
+  const { defaultPreset = 'last30days', syncToUrl = true } = options
+  const route = useRoute()
+  const router = useRouter()
+
+  // 从 URL query 初始化，无则用默认值
+  const initialPreset = (syncToUrl && route.query.preset as PeriodValue) || defaultPreset
+  const initialTab = (syncToUrl && route.query.tab as string) || 'detail'
+
+  const timeRange = ref<DateRangeParams>(getDateRangeFromPeriod(initialPreset))
+  const activeTab = ref(initialTab)
+  const modelFilter = ref<string[]>([])
+  const providerFilter = ref<string[]>([])
+  const apiKeyFilter = ref<string[]>([])
+  const userFilter = ref<string[]>([])
+
+  // URL 同步：仅同步 preset 和 tab（各类多选筛选太冗长不入 URL）
+  if (syncToUrl) {
+    watch([() => timeRange.value.preset, activeTab], ([preset, tab]) => {
+      const query: Record<string, string> = {}
+      if (preset && preset !== defaultPreset) query.preset = preset
+      if (tab && tab !== 'detail') query.tab = tab
+      // 保留其他 query 参数
+      const currentQuery = { ...route.query }
+      delete currentQuery.preset
+      delete currentQuery.tab
+      router.replace({ query: { ...currentQuery, ...query } })
+    }, { flush: 'post' })
+  }
+
+  function getTimeRangeParams() {
+    return buildTimeRangeParams(timeRange.value)
+  }
+
+  return {
+    timeRange,
+    activeTab,
+    modelFilter,
+    providerFilter,
+    apiKeyFilter,
+    userFilter,
+    getTimeRangeParams,
+  }
+}
+
+export type AnalyticsFilters = ReturnType<typeof useAnalyticsFilters>
+
+export function useInjectedAnalyticsFilters(): AnalyticsFilters {
+  const filters = inject<AnalyticsFilters>('analyticsFilters', null)
+  if (!filters) {
+    throw new Error('analyticsFilters provider is missing')
+  }
+  return filters
+}
@@ -68,6 +68,86 @@ export function getProbeCountdown(nextProbeAt: string | null | undefined, _tick:
  return '探测中'
 }

+/**
+ * Codex 配额重置倒计时状态
+ */
+export interface CodexResetStatus {
+  text: string
+  isUrgent: boolean
+  isCritical: boolean
+  isExpired: boolean
+}
+
+/**
+ * 计算 Codex 配额重置倒计时
+ * @param resetAt 绝对重置时间（Unix 秒）
+ * @param resetSecs 相对剩余秒数（用于 fallback）
+ * @param updatedAt 元数据更新时间（Unix 秒）
+ * @param _tick 响应式触发器（传入 tick.value 以触发响应式更新）
+ * @param remainingPercent 当前窗口剩余额度百分比（0-100，100 表示满额不启动倒计时）
+ */
+export function getCodexResetCountdown(
+  resetAt: number | null | undefined,
+  resetSecs: number | null | undefined,
+  updatedAt: number | null | undefined,
+  _tick: number,
+  remainingPercent?: number | null
+): CodexResetStatus | null {
+  void _tick
+
+  if (remainingPercent != null) {
+    const normalizedRemaining = Number(remainingPercent)
+    if (Number.isFinite(normalizedRemaining) && normalizedRemaining >= 100) {
+      return null
+    }
+  }
+
+  const nowSec = Math.floor(Date.now() / 1000)
+  let remaining: number
+
+  if (resetAt != null && resetAt > 0) {
+    // 优先绝对时间戳，避免相对秒数快照漂移。
+    remaining = resetAt - nowSec
+  } else if (resetSecs != null && resetSecs >= 0) {
+    if (updatedAt != null && updatedAt > 0) {
+      // 时钟偏移下 updatedAt 可能晚于当前时间，elapsed 需要下限钳制到 0。
+      const elapsedSec = Math.max(nowSec - updatedAt, 0)
+      remaining = resetSecs - elapsedSec
+    } else {
+      remaining = resetSecs
+    }
+  } else {
+    return null
+  }
+
+  if (remaining <= 0) {
+    return { text: '已重置', isUrgent: false, isCritical: false, isExpired: true }
+  }
+
+  const total = Math.floor(remaining)
+  const days = Math.floor(total / 86400)
+  const hours = Math.floor((total % 86400) / 3600)
+  const minutes = Math.floor((total % 3600) / 60)
+  const seconds = total % 60
+  const pad = (n: number) => n.toString().padStart(2, '0')
+
+  let text: string
+  if (days > 0) {
+    text = `${days}天 ${hours}:${pad(minutes)}:${pad(seconds)}`
+  } else if (hours > 0) {
+    text = `${hours}:${pad(minutes)}:${pad(seconds)}`
+  } else {
+    text = `${minutes}:${pad(seconds)}`
+  }
+
+  return {
+    text,
+    isUrgent: total < 3600,
+    isCritical: total < 300,
+    isExpired: false,
+  }
+}
+
 /**
 * OAuth Token 状态信息
 */
@@ -94,15 +174,18 @@ export function getOAuthExpiresCountdown(
  invalidReason?: string | null
 ): OAuthStatusInfo | null {
  void _tick
+  const normalizedInvalidReason = typeof invalidReason === 'string'
+    ? invalidReason.trim()
+    : ''

  // 优先检查失效状态（失效比过期更严重）
-  if (invalidAt != null) {
+  if (invalidAt != null || normalizedInvalidReason) {
    return {
      text: '已失效',
      isExpired: false,
      isExpiringSoon: false,
      isInvalid: true,
-      invalidReason: invalidReason || undefined
+      invalidReason: normalizedInvalidReason || undefined
    }
  }

@@ -1,34 +0,0 @@
-import { computed, type Ref, type ComputedRef } from 'vue'
-
-/**
- * 检测失效模型的 composable
- *
- * 用于检测 allowed_models 中已不存在于 globalModels 的模型名称，
- * 这些模型可能已被删除但引用未清理。
- *
- * @example
- * ```typescript
- * const { invalidModels } = useInvalidModels(
- *   computed(() => form.value.allowed_models),
- *   globalModels
- * )
- * ```
- */
-export interface ModelWithName {
-  name: string
-}
-
-export function useInvalidModels<T extends ModelWithName>(
-  allowedModels: Ref<string[]> | ComputedRef<string[]>,
-  globalModels: Ref<T[]>
-): { invalidModels: ComputedRef<string[]> } {
-  const validModelNames = computed(() =>
-    new Set(globalModels.value.map(m => m.name))
-  )
-
-  const invalidModels = computed(() =>
-    allowedModels.value.filter(name => !validModelNames.value.has(name))
-  )
-
-  return { invalidModels }
-}
@@ -15,6 +15,8 @@ export interface StartTestParams {
  apiFormat?: string
  endpointId?: string
  message?: string
+  requestHeaders?: Record<string, unknown>
+  requestBody?: Record<string, unknown>
  concurrency?: number
  onSuccess?: (result: TestModelFailoverResponse) => void
  /** Return `true` to indicate the failure has been handled; otherwise the composable sets `testResult`. */
@@ -61,6 +63,16 @@ export function useModelTest(options: UseModelTestOptions) {
    }
  }

+  async function refreshTraceSnapshot(reqId: string) {
+    try {
+      const trace = await requestTraceApi.getRequestTrace(reqId, { attemptedOnly: false })
+      if (requestId.value !== reqId) return
+      testTrace.value = trace
+    } catch (err: unknown) {
+      if (isAxiosError(err) && err.response?.status === 404) return
+    }
+  }
+
  function stopPolling(opts: { clearState?: boolean } = {}) {
    tracePollToken += 1
    if (tracePollTimer) {
@@ -117,13 +129,19 @@ export function useModelTest(options: UseModelTestOptions) {
    startPolling(reqId)

    try {
+      const normalizedMessage = typeof params.message === 'string' && params.message.trim()
+        ? params.message.trim()
+        : undefined
+
      const result = await testModelFailover({
        provider_id: providerId(),
        mode: params.mode,
        model_name: params.modelName,
        api_format: params.apiFormat,
        endpoint_id: params.endpointId,
-        message: params.message ?? 'hello',
+        ...(normalizedMessage ? { message: normalizedMessage } : {}),
+        ...(params.requestHeaders ? { request_headers: params.requestHeaders } : {}),
+        ...(params.requestBody ? { request_body: params.requestBody } : {}),
        request_id: reqId,
        concurrency: params.concurrency,
      }, {
@@ -131,6 +149,9 @@ export function useModelTest(options: UseModelTestOptions) {
      })

      if (result.success) {
+        await refreshTraceSnapshot(reqId)
+        stopPolling({ clearState: false })
+        testResult.value = result
        const successAttempt = result.attempts.find(a => a.status === 'success')
        const latency = successAttempt?.latency_ms != null ? ` (${successAttempt.latency_ms}ms)` : ''
        const mapped = successAttempt?.effective_model && successAttempt.effective_model !== params.modelName
@@ -138,10 +159,10 @@ export function useModelTest(options: UseModelTestOptions) {
          : ''
        params.onSuccess?.(result)
        showSuccess(`${params.displayLabel}${mapped} 测试成功${latency}`)
-        resetState()
        return
      }

+      await refreshTraceSnapshot(reqId)
      stopPolling({ clearState: false })
      const handled = params.onFailure?.(result)
      if (!handled) {
@@ -0,0 +1,287 @@
+import { computed, ref, watch, onMounted, onUnmounted, type Ref } from 'vue'
+import {
+  analyticsApi,
+  type AnalyticsBreakdownRow,
+  type AnalyticsFilterOption,
+  type AnalyticsScope,
+  type AnalyticsSummary,
+  type AnalyticsTimeseriesBucket,
+} from '@/api/analytics'
+import type { DateRangeParams, PeriodValue } from '@/features/usage/types'
+import { getDateRangeFromPeriod } from '@/features/usage/composables'
+import { createLoader, createRequestGuard, buildTimeRangeParams } from '@/composables/useAnalyticsFilters'
+import { fillMissingTimeseriesBuckets } from '@/utils/analyticsTimeseries'
+import { getAnalyticsRangeDaysInclusive, resolveAnalyticsAutoGranularity, type ResolvedAnalyticsGranularity } from '@/utils/analyticsGranularity'
+import { type DailyUsageBreakdown } from '@/utils/usageBreakdown'
+
+export interface ModelSummary {
+  model: string
+  requests: number
+  tokens: number
+  cost: number
+  inputTokens: number
+  outputTokens: number
+  cacheCreationTokens: number
+  cacheReadTokens: number
+  cacheHitRate: number
+  avg_response_time?: number
+  avg_first_byte_time?: number
+}
+
+export interface ReportBucketStat {
+  bucket_start: string
+  bucket_end: string
+  requests: number
+  tokens: number
+  cost: number
+  avg_response_time: number
+  avg_first_byte_time: number
+  models_used_count: number
+}
+
+function mapBucketToReportStat(bucket: AnalyticsTimeseriesBucket): ReportBucketStat {
+  return {
+    bucket_start: bucket.bucket_start,
+    bucket_end: bucket.bucket_end,
+    requests: bucket.requests_total,
+    tokens: bucket.total_tokens,
+    cost: bucket.total_cost_usd,
+    avg_response_time: bucket.avg_response_time_ms / 1000,
+    avg_first_byte_time: bucket.avg_first_byte_time_ms / 1000,
+    models_used_count: bucket.models_used_count,
+  }
+}
+
+function mapBucketToBreakdown(bucket: AnalyticsTimeseriesBucket): DailyUsageBreakdown {
+  return {
+    date: bucket.bucket_start,
+    inputTokens: bucket.input_tokens,
+    outputTokens: bucket.output_tokens,
+    cacheCreationTokens: bucket.cache_creation_input_tokens,
+    cacheReadTokens: bucket.cache_read_input_tokens,
+    cacheCreationCost: bucket.cache_creation_cost_usd,
+    cacheReadCost: bucket.cache_read_cost_usd,
+    cacheHitRate: bucket.cache_hit_rate,
+    totalCacheCost: bucket.cache_cost_usd,
+    totalCacheTokens: bucket.cache_creation_input_tokens + bucket.cache_read_input_tokens,
+    totalTrackedTokens: bucket.total_tokens,
+    baseTokens: bucket.input_output_total_tokens,
+  }
+}
+
+function mapBreakdownRowToModelSummary(row: AnalyticsBreakdownRow): ModelSummary {
+  const requests = row.requests_total
+  const tokens = row.total_tokens
+  const cost = row.total_cost_usd
+  return {
+    model: row.label,
+    requests,
+    tokens,
+    cost,
+    inputTokens: row.input_tokens,
+    outputTokens: row.output_tokens,
+    cacheCreationTokens: row.cache_creation_input_tokens,
+    cacheReadTokens: row.cache_read_input_tokens,
+    cacheHitRate: row.cache_hit_rate,
+    avg_response_time: row.avg_response_time_ms / 1000,
+    avg_first_byte_time: row.avg_first_byte_time_ms / 1000,
+  }
+}
+
+type ReportGranularity = ResolvedAnalyticsGranularity
+
+function resolveReportGranularity(range: DateRangeParams): ReportGranularity {
+  return resolveAnalyticsAutoGranularity(range)
+}
+
+function areStringArraysEqual(left: string[], right: string[]): boolean {
+  if (left.length !== right.length) return false
+  return left.every((value, index) => value === right[index])
+}
+
+export interface UseReportsDataOptions {
+  scope?: AnalyticsScope
+  defaultPreset?: PeriodValue
+  timeRange?: Ref<DateRangeParams>
+  userFilter?: Ref<string[]>
+  apiKeyFilter?: Ref<string[]>
+  loadApiKeyOptions?: boolean
+}
+
+export function useReportsData(options: UseReportsDataOptions = {}) {
+  const {
+    scope = { kind: 'me' as const },
+    defaultPreset = 'today',
+    timeRange: externalTimeRange,
+    userFilter: externalUserFilter,
+    apiKeyFilter: externalApiKeyFilter,
+    loadApiKeyOptions = scope.kind === 'me',
+  } = options
+
+  const timeRange = externalTimeRange ?? ref<DateRangeParams>({
+    ...getDateRangeFromPeriod(defaultPreset),
+    granularity: 'auto',
+  })
+  const userFilter = externalUserFilter ?? ref<string[]>([])
+  const apiKeyFilter = externalApiKeyFilter ?? ref<string[]>([])
+  const apiKeyOptions = ref<AnalyticsFilterOption[]>([])
+
+  const summary = ref<AnalyticsSummary | null>(null)
+  const bucketStats = ref<ReportBucketStat[]>([])
+  const bucketBreakdowns = ref<Record<string, DailyUsageBreakdown>>({})
+  const modelSummary = ref<ModelSummary[]>([])
+  const loading = ref(false)
+  const loadError = ref(false)
+  const hasLoaded = ref(false)
+  const guard = createRequestGuard()
+  const apiKeysGuard = createRequestGuard()
+  let suppressNextApiKeyFilterLoad = false
+  const resolvedGranularity = computed(() => resolveReportGranularity(timeRange.value))
+  const timeseriesTimeRange = computed(() => ({
+    ...buildTimeRangeParams(timeRange.value),
+    granularity: resolvedGranularity.value,
+  }))
+
+  watch(
+    () => [timeRange.value.preset, timeRange.value.start_date, timeRange.value.end_date],
+    () => {
+      if (timeRange.value.granularity === 'hour' && getAnalyticsRangeDaysInclusive(timeRange.value) !== 1) {
+        timeRange.value = {
+          ...timeRange.value,
+          granularity: 'auto',
+        }
+      }
+    },
+  )
+
+  async function load() {
+    const requestId = guard.next()
+    loading.value = true
+    loadError.value = false
+    try {
+      const params = buildTimeRangeParams(timeRange.value)
+      const timeseriesParams = timeseriesTimeRange.value
+      const basePayload = {
+        scope,
+        time_range: params,
+        filters: {
+          user_ids: userFilter.value,
+          api_key_ids: apiKeyFilter.value,
+        },
+      }
+
+      const [overviewData, timeseriesData, modelBreakdownData] = await Promise.all([
+        analyticsApi.getOverview(basePayload),
+        analyticsApi.getTimeseries({
+          ...basePayload,
+          time_range: timeseriesParams,
+        }),
+        analyticsApi.getBreakdown({
+          ...basePayload,
+          dimension: 'model',
+          limit: 50,
+        }),
+      ])
+
+      if (guard.isStale(requestId)) return
+
+      const filledBuckets = fillMissingTimeseriesBuckets(timeseriesData.buckets, timeseriesParams)
+
+      summary.value = overviewData.summary
+      bucketStats.value = filledBuckets.map(mapBucketToReportStat)
+      bucketBreakdowns.value = Object.fromEntries(
+        filledBuckets.map(bucket => {
+          const breakdown = mapBucketToBreakdown(bucket)
+          return [breakdown.date, breakdown]
+        }),
+      )
+      modelSummary.value = modelBreakdownData.rows.map(mapBreakdownRowToModelSummary)
+      hasLoaded.value = true
+    } catch {
+      if (guard.isStale(requestId)) return
+      loadError.value = true
+    } finally {
+      if (guard.isCurrent(requestId)) {
+        loading.value = false
+      }
+    }
+  }
+
+  async function loadApiKeys() {
+    if (!loadApiKeyOptions) return
+    const requestId = apiKeysGuard.next()
+    const filterOptions = await analyticsApi.getFilterOptions({
+      scope,
+      time_range: buildTimeRangeParams(timeRange.value),
+      filters: {},
+    }).catch(() => null)
+    if (apiKeysGuard.isStale(requestId)) return
+
+    const nextOptions = filterOptions?.api_keys ?? []
+
+    apiKeyOptions.value = nextOptions
+    const validOptionValues = new Set(nextOptions.map(option => option.value))
+    const nextFilter = apiKeyFilter.value.filter(value => validOptionValues.has(value))
+    if (!areStringArraysEqual(nextFilter, apiKeyFilter.value)) {
+      suppressNextApiKeyFilterLoad = true
+      apiKeyFilter.value = nextFilter
+    }
+  }
+
+  const loader = createLoader(load)
+  const apiKeysLoader = createLoader(loadApiKeys)
+
+  watch(
+    timeRange,
+    () => {
+      void (async () => {
+        if (loadApiKeyOptions) {
+          await apiKeysLoader.execute()
+        }
+        loader.schedule()
+      })()
+    },
+    { deep: true },
+  )
+
+  watch(
+    [userFilter, apiKeyFilter],
+    () => {
+      if (suppressNextApiKeyFilterLoad) {
+        suppressNextApiKeyFilterLoad = false
+        return
+      }
+      loader.schedule()
+    },
+    { deep: true },
+  )
+
+  onMounted(() => {
+    void (async () => {
+      if (loadApiKeyOptions) {
+        await apiKeysLoader.execute()
+      }
+      await loader.execute()
+    })()
+  })
+  onUnmounted(() => {
+    loader.cleanup()
+    apiKeysLoader.cleanup()
+    guard.invalidate()
+    apiKeysGuard.invalidate()
+  })
+
+  return {
+    timeRange,
+    apiKeyFilter,
+    apiKeyOptions,
+    summary,
+    bucketStats,
+    bucketBreakdowns,
+    modelSummary,
+    loading,
+    loadError,
+    hasLoaded,
+    resolvedGranularity,
+  }
+}
@@ -0,0 +1,44 @@
+import { useRoute, useRouter, type LocationQueryValue, type LocationQuery } from 'vue-router'
+
+type QueryValue = LocationQueryValue | LocationQueryValue[]
+
+function normalizeQueryValue(value: QueryValue): string | undefined {
+  if (Array.isArray(value)) {
+    return value.length > 0 ? (value[value.length - 1] ?? undefined) : undefined
+  }
+  return typeof value === 'string' ? value : undefined
+}
+
+function queriesEqual(left: LocationQuery, right: LocationQuery): boolean {
+  const keys = new Set([...Object.keys(left), ...Object.keys(right)])
+  for (const key of keys) {
+    if (normalizeQueryValue(left[key]) !== normalizeQueryValue(right[key])) {
+      return false
+    }
+  }
+  return true
+}
+
+export function useRouteQuery() {
+  const route = useRoute()
+  const router = useRouter()
+
+  function getQueryValue(key: string): string | undefined {
+    return normalizeQueryValue(route.query[key])
+  }
+
+  function patchQuery(patch: Record<string, string | undefined | null>) {
+    const next: LocationQuery = { ...route.query }
+    for (const [key, value] of Object.entries(patch)) {
+      if (value == null || value.trim() === '') {
+        delete next[key]
+      } else {
+        next[key] = value
+      }
+    }
+    if (queriesEqual(route.query, next)) return
+    void router.replace({ query: next }).catch(() => {})
+  }
+
+  return { route, router, getQueryValue, patchQuery }
+}
--- a/Show More
+++ b/Show More